ABSTRACT 
 
 
 
 
Title of Dissertation: UNDERSTANDING ENVIRONMENTAL 
FACTORS DRIVING WILDLAND FIRE 
IGNITIONS IN ALASKAN TUNDRA   
  
 Jiaying He, Doctor of Philosophy, 2020 
  
Dissertation directed by: Professor, Tatiana V. Loboda, Department of 
Geographical Sciences 
 
 
Wildland fire is a dominant disturbance agent that drives ecosystem change, 
climate forcing, and carbon cycle in the boreal forest and tundra ecosystems of the 
High Northern Latitudes (HNL). Tundra fires can exert a considerable influence on 
the local ecosystem functioning and contribute to climate change through 
biogeochemical and biogeophysical effects. However, the drivers and mechanisms of 
tundra fires are still poorly understood. Research on modeling contemporary fire 
occurrence in the tundra is also lacking. This dissertation addresses the overarching 
scientific question of ?What environmental factors and mechanisms drive wildfire 
ignition in Alaskan tundra?? Environmental factors from multiple aspects are 
considered including fuel type and state, fire weather, topography, and ignition 
source. First, to understand the spatial distribution of fuel types in the tundra, multi-
year satellite observations and field data were used to develop the first fractional 
coverage product of major fuel type components across the entire Alaskan tundra at 
  
30 m resolution. Second, to account for the primary ignition source of fires in the 
HNL, an empirical-dynamical modeling framework was developed to predict the 
probability of cloud-to-ground (CG) lightning across Alaskan tundra, through the 
integration of Weather Research and Forecast (WRF) model and machine learning 
algorithm. Finally, environmental factors including fuel type distribution, fuel 
moisture state, WRF simulated ignition source and fire weather, and topographical 
features, were combined with empirical modeling methods to understand their roles in 
driving wildland fire ignitions across Alaskan tundra from 2001 to 2019. This work 
demonstrates the strong capability for accurate prediction of CG lightning and 
wildland fire probabilities, by incorporating dynamic weather models, empirical 
methods, and satellite observations in data-scarce regions like the HNL. The 
developed models present a novel component of fire danger modeling that can 
considerably strengthen the current capability to forecast fire occurrence and support 
operational fire management agencies in the HNL. In addition, the insights gained 
from this research will allow for more accurate representation of wildfire ignition 
probabilities in studies focused on assessing the impact of the projected climate 
change in HNL tundra which has largely absent in previous modeling efforts.  
 
 
 
 
 
 
  
 
 
 
 
 
 
 
 
 
 
UNDERSTANDING ENVIRONMENTAL FACTORS DRIVING WILDLAND 
FIRE IGNITIONS IN ALASKAN TUNDRA 
 
 
 
by 
 
 
Jiaying He 
 
 
 
 
 
Dissertation submitted to the Faculty of the Graduate School of the  
University of Maryland, College Park, in partial fulfillment 
of the requirements for the degree of 
Doctor of Philosophy 
2020 
 
 
 
 
 
 
 
 
 
 
Advisory Committee: 
Dr. Tatiana V. Loboda, Chair 
Dr. Christopher Justice 
Dr. Nancy French 
Dr. Evan Ellicott 
Dr. Rachel Pinker 
 
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
? Copyright by 
Jiaying He 
2020 
 
 
 
 
 
 
 
 
 
 
 
 
 
  
 
Dedication 
 
 
To my dear parents 
For their unconditional love and support 
 
To all women who contribute their life 
To the development of Science, Technology, Engineering, and Math (STEM) 
ii 
 
 
 
Acknowledgements 
Writing acknowledgements is never an easy task for someone like me who is not 
good at expressing his or her feelings. After submitting my dissertation draft to my 
dear dissertation committee, I finally get a chance to calm down and to think about 
who and how I would like to thank. Looking back at my six years as a member of 
University of Maryland and Department of Geographical Sciences, it has been such a 
precious journey in my life to learn how to grow up, to think about who I want to be, 
and even to reshape my personality.  
First and foremost, I would like to thank my advisor Dr. Tatiana Loboda for being 
such a great advisor, a scholar, and a role model as a female scientist. Thank you for 
all your support, guidance, and encouragement. Thank you for your trust and high 
standards that help me rebuild my confidence and push me to excel myself. Having 
you as my advisor is the luckiest thing that I have had during the past six years and 
encourages me to keep pursuing scientific research.  
I am also very grateful for the mentorship and encouragement from all the members 
of my dissertation advisory committee. To Dr. Chris Justice, thank you for your 
scientific mind and forward thinking that inspire me to think like a scientist. To Dr. 
Nancy French, thank you for your kind help and solid knowledge that guide my 
dissertation work and thank you for being a great role model of female scientist for 
me. To Dr. Evan Ellicott, thank you for all your knowledge and encouragement that 
support me throughout this journey. To Dr. Rachel Pinker, thank you for agreeing to 
serve as my dean?s representative.  
iii 
 
 
 
Special thanks to my lovely lab mates Varada, Tony, Joanne, Amanda and Allison! 
Thank you for always being so supportive and helpful during the past four years. I 
feel very grateful for having your company through this journey. I would like to thank 
Liza Jenkins as well, for all your knowledge and help during the most exciting field 
trip I have ever had at Seward Peninsula, Alaska in 2017.  
To Rachel Berndtson, thank you for your patience and kind help whenever I ran into 
problems and came to you for solutions. To Jack Ma, Jonathan Resop, Elle Lim and 
Kristen Bergery, thank you all for the best TA experiences I have ever had working 
with the MPGIS program. Also thanks to Varaprasad Bandaru for the great RA 
opportunity to work with and to learn from you. 
To all my cohort and colleagues at Department of Geographical Sciences, it?s a great 
opportunity to know and to work with all of you. Thank you for sharing the journey 
of grad school with me. Also special thanks to all the lab members of Room 2138 at 
Lefrak Hall. I really enjoy being a member of this lab because of you. 
To all the instructors of the RecWell group fitness classes at UMD, thank you all for 
your energetic and fun classes that always reboot my brain after working for an entire 
day.  
Last but not the least, I would like to express my thanks to my family and friends. 
Thank you for always being there for me. 
 
iv 
 
 
 
Table of Contents 
 
Dedication ..................................................................................................................... ii 
Acknowledgements ...................................................................................................... iii 
Table of Contents .......................................................................................................... v 
List of Tables .............................................................................................................. vii 
List of Figures ............................................................................................................ viii 
List of Abbreviations .................................................................................................. xii 
Chapter 1: Introduction ................................................................................................. 1 
1.1 Background and Motivation ............................................................................... 1 
1.2 Environmental Factors Related to Wildfire Behavior ........................................ 4 
1.2.1 Ignition source ............................................................................................. 5 
1.2.2 Fire weather conditions ................................................................................ 7 
1.2.3 Surface fuels ................................................................................................. 9 
1.2.4 Topography ................................................................................................ 10 
1.3 Fire modeling efforts ......................................................................................... 11 
1.4 Research Questions ........................................................................................... 12 
1.5 Dissertation Structure ........................................................................................ 14 
Chapter 2: Mapping Fractional Cover of Major Fuel Type Components across 
Alaskan Tundra ........................................................................................................... 16 
2.1 Summary ........................................................................................................... 16 
2.2 Introduction ....................................................................................................... 17 
2.3 Study Area ........................................................................................................ 23 
2.4 Data and Methods ............................................................................................. 24 
2.4.1 Data Sources .............................................................................................. 26 
2.4.2 Developing seasonal composites of spectral bands and indices ................ 30 
2.4.3 Generating a tundra vegetation mask for shrub or graminoid dominated 
tundra .................................................................................................................. 34 
2.4.4 RF modeling of fractional cover ................................................................ 35 
2.4.5 Fuel type component mapping ................................................................... 37 
2.5 Results ............................................................................................................... 38 
2.5.1 Accuracy assessment of tundra vegetation mask ....................................... 38 
2.5.2 Accuracy assessment of fractional cover estimation ................................. 39 
2.5.3 Fractional cover of major fuel type components across Alaskan tundra ... 42 
2.6 Discussions ....................................................................................................... 50 
2.7 Conclusions ....................................................................................................... 54 
Chapter 3: Modeling cloud-to-ground lightning probability in Alaskan tundra through 
the integration of Weather Research and Forecast (WRF) model and machine learning 
method ......................................................................................................................... 56 
3.1 Summary ........................................................................................................... 56 
3.2 Introduction ....................................................................................................... 57 
3.3 Study Area ........................................................................................................ 61 
3.4 Materials and Methods ...................................................................................... 62 
3.4.1 Cloud-to-ground lightning observations .................................................... 63 
3.4.2 WRF model setup and sensitivity analysis ................................................ 63 
3.4.3 RF modeling of cloud-to-ground lightning probability ............................. 67 
v 
 
 
 
3.4.4 Forecasting capability assessment ............................................................. 71 
3.5 Results ............................................................................................................... 71 
3.5.1 Sensitivity analysis of WRF simulation ..................................................... 71 
3.5.2 Accuracy assessment of RF models ........................................................... 74 
3.5.3 Forecasting performance ............................................................................ 77 
3.5.4 Evaluation of variable importance ............................................................. 78 
3.6 Discussions ....................................................................................................... 79 
3.7 Conclusions ....................................................................................................... 82 
Chapter 4: Exploring environmental factors driving wildland fire ignitions in Alaskan 
tundra .......................................................................................................................... 84 
4.1 Summary ........................................................................................................... 84 
4.2 Introduction ....................................................................................................... 85 
4.3 Study Area ........................................................................................................ 89 
4.4 Data and Methods ............................................................................................. 90 
4.4.1 Determining fire ignition locations ............................................................ 93 
4.4.2 Modeling cloud-to-ground lightning probability with WRF simulations .. 96 
4.4.3 Near surface weather conditions and calculation of the CFFWIS ............. 97 
4.4.4 Estimating fuel state with vegetation indices ........................................... 100 
4.4.5 Extracting topographic features ............................................................... 101 
4.4.6 Empirical modeling of fire ignition probability ....................................... 101 
4.5 Results ............................................................................................................. 102 
4.5.1 Variable selections based on correlations ................................................ 102 
4.5.2 Accuracy of empirical modeling .............................................................. 106 
4.5.3 Evaluation of variables driving tundra fire ignitions ............................... 107 
4.6 Discussions ..................................................................................................... 116 
4.7 Conclusions ..................................................................................................... 120 
Chapter 5: Conclusion ............................................................................................... 122 
5.1 Major research findings .................................................................................. 122 
5.2 Contribution to Broader Arctic Science Research .......................................... 126 
5.3 Implications for fire management efforts in the HNL .................................... 128 
5.4 Future research directions ............................................................................... 130 
Appendices ................................................................................................................ 133 
Bibliography ............................................................................................................. 137 
 
 
 
 
 
 
vi 
 
 
 
List of Tables 
 
Chapter 2: 
Table 2.1. LANDFIRE fuel products and their major fuel strata or types available for 
Alaskan tundra ............................................................................................................ 19 
Table 2.2. Summary of Landsat 8 OLI tiles processed in this study. WRS2 path/row 
overlaps exist between different regions. .................................................................... 27 
Table 2.3. Landsat 8 spectral metrics used for developing seasonal composites. ...... 32 
Table 2.4. Land cover classes and corresponding classification codes from 
LANDFIRE products. ................................................................................................. 35 
Table 2.5. Confusion matrix of land cover classification for developing tundra 
vegetation mask. ......................................................................................................... 39 
Table 2.6. Statistical summary of three RF regression models. .................................. 39 
 
Chapter 3: 
Table 3.1. WRF schemes used for the pan-Arctic region in existing literature .......... 65 
Table 3.2. Six candidates of parameterization combinations for sensitivity analysis 66 
Table 3.3. Summary of case studies for sensitivity analysis ....................................... 66 
Table 3.4. List of independent variables retrieved from WRF output for RF modeling
..................................................................................................................................... 68 
Table 3.5. Cases studies of WRF simulations for CG lightning modeling ................. 70 
Table 3.6. Contingency matrix of variables used to calculate statistical scores ......... 70 
Table 3.7. Statistical criteria used for assessing modeling performance .................... 71 
Table 3.8. Optimized combination of physical parameterizations ............................. 74 
Table 3.9. Confusion matrix for (a) ?24-hr model? and (b) ?48-hr model? ............... 74 
Table 3.10. Statistical criteria calculated using the validation data for: (a) 24-hr 
model; (b) 48-hr model. .............................................................................................. 75 
Table 3.11.  Statistical criteria calculated for the forecasting test .............................. 78 
 
Chapter 4: 
Table 4.1. Full list of fire ignition dates detected from sampled years ....................... 96 
Table 4.2. Vegetation indices for estimating live fuel moisture content .................. 100 
Table 4.3. RF modeling results of the two models for predicting fire ignitions ....... 106 
Table 4.4. Logistic regression results for the two models ........................................ 109 
Table 4.5. Welch t-test results across different tundra regions. ................................ 110 
 
vii 
 
 
 
List of Figures 
 
Chapter 1: 
Figure 1.1. Active fires detected by Moderate Resolution Imaging Spectroradiometer 
(MODIS) sensors in the pan-Arctic tundra ecosystems from 2001 to 2016. ................ 2 
Figure 1.2. Fire count and burned area in different regions of Alaskan tundra from 
2001 to 2017 according to Alaska Large Fire Database (ALFD). ................................ 2 
Figure 1.3. Factors controlling fires at different scales as an extension of the 
commonly used fire triangles. The left triangle represents the Fire Fundamentals 
Triangle and the middle one represents the Fire Environmental Triangle. This figure 
is from Moritz et al. (2005). .......................................................................................... 5 
Figure 1.4. Comparison of CFFWIS (left) and NFDRS (right), source: National 
Wildfire Coordinate Group (https://www.nwcg.gov/publications/pms437/cffdrs/fire-
weather-index-system). ................................................................................................. 8 
Figure 1.5. Framework of Research Questions and Objectives .................................. 12 
 
Chapter 2: 
Figure 2.1. Examples of highly mixed fuel type components in Alaskan tundra: (a) 
shrub and graminoid tussocks; (b) grass, lichen and moss. ........................................ 21 
Figure 2.2. Alaskan tundra region defined by CAVM and corresponding ecoregions 
identified by the Unified Ecoregions of Alaska: 2001. Each color represents one 
ecoregion unit. Ecoregion units within the same Level 2 ecoregion groups are 
separated by different line patterns. ............................................................................ 24 
Figure 2.3. Flowchart of fractional fuel component mapping. ................................... 26 
Figure 2.4. Alaskan tundra region as defined by CAVM (a) and three field campaign 
sites: (b) 2016 field plots in Noatak River National Preserve, (c) 2017 field plots in 
Seward Peninsula, (d) 2012 field plots in North Slope. .............................................. 28 
Figure 2.5. Averaged NDVI profiles extracted from MODIS product (MOD13Q1) on 
Google Earth Engine in four regions across Alaskan tundra (a). Ten pixels were 
randomly selected and plotted for each region as an example (b): (1) North Slope, (2) 
Noatak River National Preserve, (3) Seward Peninsula, (4) Southwest Alaska. ........ 31 
Figure 2.6. Unmasked cloud shadow pixels with high NDVI values from a Landsat 8 
scene (Landsat Scene Identifier: LC80690122017228LGN00). Examples are 
highlighted in yellow rectangles: (a) LandsatLook natural color image, (b) NDVI and 
cloud/shadow/water mask identified by level-2 QA data from the same Landsat scene.
..................................................................................................................................... 33 
Figure 2.7. Scatter plots comparing estimated and observed fractional cover using the 
validation data for (a) woody, (b) herbaceous, and (c) nonvascular components. 
RMSE and R-squared values between the estimations and observations are reported 
in the scatter plot of each component. RSMEs are within 2% of those from the 
internal model statistics. R-squared values reach 0.95 for all components. ............... 40 
Figure 2.8. Top 10 important independent variables and their MDA values from RF 
regression models for (a) woody, (b) herbaceous and (c) nonvascular components. 
viii 
 
 
 
Spectral bands and indices during pre- and peak-growing seasons play the most 
important role in determining the fractional cover of woody and herbaceous 
components (a-b), while spectral indices in the post-growing seasons contribute most 
to the cover estimation of nonvascular component (c). .............................................. 41 
Figure 2.9. Fractional cover of three major fuel type components across Alaskan 
tundra: (a) woody, (b) herbaceous, and (c) nonvascular components. ....................... 43 
Figure 2.10. Examples of fractional cover distributions in a tundra region near Lake 
Narvakrak in the Noatak River National Preserve: (a) VHR image from Google Earth; 
(b) woody component; (b) herbaceous component; (d) nonvascular component. ...... 44 
Figure 2.11. Boxplots of woody component cover against ecoregions in Alaskan 
tundra: (a) woody component cover by Level 1 scheme; (b) woody component cover 
by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion units are colored from 
light yellow to dark blue based on the latitude from north to south). Shrub coverage 
increases gradually among all the Level 2 groups and the ecoregion units, from north 
to south (b). ................................................................................................................. 45 
Figure 2.12. Boxplots of herbaceous component cover against ecoregions in Alaskan 
tundra: (a) herbaceous component cover by Level 1 scheme; (b) herbaceous 
component cover by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion units 
are colored from light yellow to dark blue based on the latitude from north to south). 
Herbaceous cover values among Level 2 groups are generally comparable, with 
relative higher values in the ?Brooks Foothills? and ?Bering Sea Islands? units (b). 46 
Figure 2.13. Boxplots of nonvascular component cover against ecoregions in Alaskan 
tundra: (a) nonvascular component cover by Level 1 scheme; (b) nonvascular 
component cover by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion units 
are colored from light yellow to dark blue based on the latitude from north to south). 
Nonvascular coverage shows a decreasing trend among the Level 2 groups and the 
ecoregion units from north to south (b). ..................................................................... 47 
Figure 2.14. Scatter plots comparing our field observations and estimations from 
other products: (a) 2000 shrub cover (Beck et al.,  2011); (b) PFT products (Macander 
et al., 2017). Compared to our field observations, the PFT product (Macander et al., 
2017) tends to overestimate the shrub cover, but underestimate the lichen/moss cover. 
The fractional cover of herbaceous species (sedge/grass) is relatively comparable. .. 48 
Figure 2.15. Density plots comparing my results and estimations from existing 
products: (a) woody component cover circa 2000 by Beck et al. (2011); (b) woody 
component cover circa 2015 by Macander et al. (2017); (c) herbaceous component 
cover circa 2015 by Macander et al. (2017); (d) nonvascular component cover circa 
2015 by Macander et al. (2017). Fractional cover values of woody component in both 
products are higher than the estimations in my results.  Nonvascular component tends 
to have much higher coverage than in the PFT product. ............................................ 50 
 
 
Chapter 3: 
Figure 3.1. Study area: (a) Alaskan tundra defined by CAVM; (b) WRF simulation 
domains. ...................................................................................................................... 62 
Figure 3.2. Overall workflow of Chapter 3 ................................................................. 63 
ix 
 
 
 
Figure 3.3. Illustration of the ?24-hr model? and ?48-hr model? components ........... 67 
Figure 3.4. Bar plots summarizing the ranks of six candidate schemes based on 
Pearson?s r correlation, MAE and RMSE for: (a) air temperature, (b) dew point 
temperature, (c) relative humidity, and (d) solar radiation. The largest rank sum value 
represents the best overall performance. ..................................................................... 72 
Figure 3.5. Comparison of CAPE simulation in Domain 2 between (a) 
Morrison_MYJ and (b) Thompson_MYJ for the 2010 case as an example. .............. 73 
Figure 3.6. ROC curves of validation results on lightning days with different severity 
levels: (a) 24-hr model, (b) 48-hr model. .................................................................... 76 
Figure 3.7. (a) Observed CG lightning strikes during 2010/07/02 00:00 and 
2010/07/03 00:00UTC and (b) modeled CG lightning probability with the ?24-hr 
model? in Alaskan tundra as an example. ................................................................... 77 
Figure 3.8. ROC curves of forecasting tests on lightning days with different severity 
levels ........................................................................................................................... 78 
Figure 3.9. Top 20 important variables ranked by MDA in (a) ?24-hr model? and (b) 
?48-hr model?. ............................................................................................................ 79 
 
Chapter 4: 
Figure 4.1. Methodology framework of modeling fire ignition probability in Alaskan 
tundra in Chapter 4. ..................................................................................................... 91 
Figure 4.2. Modeling workflow .................................................................................. 93 
Figure 4.3. Workflow for extracting fire ignition locations in Alaskan tundra .......... 94 
Figure 4.4. Number of fire ignition events detected from 2001 to 2019 with 
MCD14DL data. Sampled fire years are highlighted in orange boxes. ...................... 95 
Figure 4.5. Flow chart of Canadian FWI system, source: National Wildfire 
Coordinating Group (https://www.nwcg.gov/publications/pms437/cffdrs/fire-weather-
index-system) .............................................................................................................. 98 
Figure 4.6. Correlation matrix for static and dynamic variables extracted on the fire 
ignition days. ?LP?: CG lightning probability, ?Temp: air temperature, ?Elev?: 
elevation, ?Rough?: roughness, ?Herb?: herbaceous component, and ?Nonva?: 
nonvascular component. ........................................................................................... 104 
Figure 4.7. Correlation matrix for static and dynamic variables extracted on the dates 
before fire ignitions. ?LP?: CG lightning probability, ?Temp: air temperature, ?Elev?: 
elevation, ?Rough?: roughness, ?Herb?: herbaceous component, and ?Nonvas?: 
nonvascular component. ........................................................................................... 105 
Figure 4.8. ROC curves estimated based on validation datasets for the two models: (a) 
?Current model?; (b) ?Previous model?. .................................................................. 107 
Figure 4.9. Variable importance rank for models: (a) ?Fire ignition day?, (b) ?One 
day before ignition? .................................................................................................. 108 
Figure 4.10. Boxplots of CG lightning probability between ?Fire? and ?No fire? 
conditions across three major fuel regions: (a) CG lightning probability on fire 
ignition day, (b) CG lightning on the previous day. ................................................. 112 
Figure 4.11. Boxplots of (a) air temperature and (b) RH on fire ignition days between 
?Fire? and ?No fire? conditions across three major fuel regions. ............................. 113 
x 
 
 
 
Figure 4.12. Boxplots of (a) FFMC and (b) DC on fire ignition days between ?Fire? 
and ?No fire? conditions across three major fuel regions. ........................................ 113 
Figure 4.13. Boxplots of (a) ISI and (b) BUI on fire ignition days between ?Fire? and 
?No fire? conditions across three major fuel regions. ............................................... 114 
Figure 4.14. Boxplots of (a) NDII6 and (b) NDII7 on fire ignition days between ?Fire? 
and ?No fire? conditions across three major fuel regions. ........................................ 114 
Figure 4.15. Boxplots of mean fractional covers of (a) woody component, (b) 
herbaceous, and (c) nonvascular components between ?Fire? and ?No fire? events 
across three major fuel regions. ................................................................................ 115 
 
xi 
 
 
 
List of Abbreviations 
 
ABoVE ? Arctic-Boreal Vulnerability Experiment 
ALDN ? Alaskan Lightning Detection Network 
ALFD ? Alaskan Large Fire Database 
ALFRESCO ? Alaska Frame-based Ecosystem Code 
AUC ? Area Under Curve 
BLM ? Bureau of Land Management 
BUI ? Buildup Index 
CAPE ? Convective Available Potential Energy 
CAVM ? Circumpolar Arctic Vegetation Map 
CFFBPS ? Canadian Forest Fire Behavior Prediction System 
CFFDRS ? Canadian Forest Fire Danger Rating System 
CFFWIS ? Canadian Forest Fire Weather Index System 
CG ? Cloud-to-ground 
DBSCAN ? Density-Based Spatial Clustering of Applications with Noise 
DC ? Drought Code 
DEM ? Digital Elevation Model 
DMC ? Duff Moisture Code 
DOE ? Department of Energy 
DOI ? Department of Interior 
DOY ? Day of Year 
FCCS ? Fuel Characteristic Classification System 
FFMC ? Fine Fuel Moisture Code 
xii 
 
 
 
FIRMS ? Fire Information for Resource Management System 
FLM ? Fuel Loading Model 
FWI ? Fire Weather Index 
GVMI ? Global Vegetation Moisture Index 
HNL ? High Northern Latitudes 
IfSAR ? Interferometric Synthetic Aperture Radar 
IQR ? Interquartile Range 
ISI ? Initial Spread Index 
MAE ? Mean Absolute Error 
MDA ? Mean Decrease Accuracy 
MODIS ? Moderate Resolution Imaging Spectroradiometer 
MSE ? Mean Squared Error 
NARR ? North American Regional Reanalysis 
NASA ? National Aeronautics and Space Administration  
NBR ? Normalized Burn Ratio 
NCEP ? National Centers for Environmental Prediction 
NDII ? Normalized Difference Infrared Index 
NDVI ? Normalized Difference Vegetation Index 
NDWI ? Normalized Difference Water Index 
NFDRS ? National Forest Danger Rating System 
NGEE ? Next Generation Ecosystem Experiments  
NWP ? Numerical Weather Prediction 
OLI ? Operational Land Imager 
xiii 
 
 
 
OOB ? Out-of-bag 
ORNL DAAC ? Oak Ridge National Laboratory Distributed Active Archive Center  
PFT ? Plant Functional Type 
PLI ? Parcel Lifted Index 
QA ? Quality Assessment 
RAWS ? Remote Automated Weather Stations 
RF ? Random Forest 
RH ? Relative Humidity 
RMSE ? Root Mean Squared Error 
ROC ? Receiver Operating Characteristics 
TC ? Tasseled Cap  
UTC ? Coordinated Universal Time 
VHR ? Very High Resolution 
WRF ? Weather Research and Forecast 
xiv 
 
 
 
Chapter 1: Introduction 
 
1.1 Background and Motivation 
Wildland fire is a dominant disturbance agent that drives ecosystem change, 
climate forcing and carbon cycle in the boreal forest and tundra ecosystems of the 
High Northern Latitudes (HNL; Bond-Lamberty et al., 2007; Sulla-Menashe et al., 
2018). Satellite observations show that tundra fires are common throughout the pan-
Arctic region (Figure 1.1; Masrur et al., 2018). The Alaskan tundra has also witnessed 
several severe fires or fire seasons in recent years, namely the 2010 fire season in the 
Noatak River valley and the 2015 fire season in Southwest Alaska (Figure 1.2). 
Although Alaskan tundra only occupies approximately 10% of the world?s tundra 
area, it has contributed more than 50% of the global tundra burned area since 2001. 
Moreover, paleoecological data indicate that frequent burnings had occurred in 
Alaskan tundra under suitable climate and fuel scenarios (Higuera et al., 2011; Hu et 
al., 2015). Climate is now changing at a rapid rate in the Arctic (Larsen et al., 2014). 
As the only Arctic region in the United States, Alaska has experienced twice the 
warming rate of other areas in the US since the 1950s (Stewart et al., 2013). This 
warming is likely to drive an increase in fire activity and changes in fire regimes 
within the Alaskan tundra ecosystem in the future regardless of climate change 
mitigation strategies (French et al., 2015; Young et al., 2017). 
 
 
1 
 
 
 
 
Figure 1.1. Active fires detected by Moderate Resolution Imaging Spectroradiometer 
(MODIS) sensors in the pan-Arctic tundra ecosystems from 2001 to 2016. 
 
Figure 1.2. Fire count and burned area in different regions of Alaskan tundra from 
2001 to 2017 according to Alaska Large Fire Database (ALFD). 
2 
 
 
 
Tundra fire acts as a catalyst for biogeochemical processes and releases large 
amounts of ancient carbon stored in the organic soil and widespread underlying 
permafrost, thus affecting both the regional and global carbon cycle (Jones et al., 
2015; Mack et al., 2011). The extreme 2007 Anaktuvuk fire in Alaska burned an area 
of 1,039 km2 and released approximately 2.1 Tg of carbon stocked in the organic soil 
and permafrost into the atmosphere (Mack et al., 2011). Fire-induced changes in 
surface albedo and post-fire vegetation compositions alter the surface radiative 
forcing with subsequent feedbacks into the climate system (French et al., 2016; Frost 
et al., 2020; McManus et al., 2012; Pearson et al., 2013).  Climate change is regarded 
as the leading factor that alters ecosystem services and functions in the Arctic 
(Hinzman et al., 2005; Meredith et al., 2019; Post et al., 2009), however, the impact 
of climate change is likely to be strongly amplified by the rapid response of fire 
activities, which could eclipse the direct influences of climate change on the 
ecosystems (Flannigan et al., 2000).  
Despite the importance of tundra fires, current research is primarily limited to 
the evaluation of post-fire impacts such as fire severity (Loboda et al., 2013), 
ecosystem responses (Bret-Harte et al., 2013), and carbon budget change (Mack et al., 
2011) with comparatively little attention to modeling tundra fire occurrence.  
Previous studies have modeled historical or future tundra fire regimes with either 
empirical methods or ecosystem models, e.g. the Alaska Frame-Based Ecosystem 
Code (ALFRESCO; Higuera et al., 2011; Joly et al., 2012; Young et al., 2017). 
recently examined the climatic factors influencing tundra fire activities at a large 
spatial scale for the entire circumpolar region. However, research on modeling 
3 
 
 
 
contemporary fire occurrence is lacking in current English language peer-reviewed 
literature. Although numerous studies have modeled the interactions between 
environmental factors and wildfire occurrences in boreal forests, their results are not 
directly applicable to the tundra ecosystems due to their different ecosystem 
functioning and responses (French et al., 2015; Rocha et al., 2012). Studies of 
wildland fires have three major components: fire occurrence, fire spread and fire 
impacts. Considering the lack of efforts on understanding the driving mechanisms of 
wildland fire in the tundra and on modeling fire potential, this dissertation focuses on 
examining fire occurrence, and particularly how fire starts in the tundra 
1.2 Environmental Factors Related to Wildfire Behavior 
Wildfires are controlled by a variety of interacting factors (Pyne et al., 1996). 
Moritz et al. (2005) summarized the interactions between the influencing 
environmental factors and fire at multiple spatial and temporal scales (Figure 1.3). In 
particular, oxygen, heat and fuel are the three key components that support the 
combustion processes, which are typically known as the Fire Fundamentals Triangle. 
Beyond the combustion, fire behavior is a result of environmental conditions referred 
to as the Fire Environment Triangle, primarily including fuel, weather and 
topography. In the context of fire ecology, the life cycle of a wildland fire includes 
several stages: ignition, transition to spread, acceleration, and steady spread (Pyne et 
al., 1996). From the perspective of wildland fire behavior, the potential of fire 
occurrence refers to that of fire ignition. Although ignition source is not part of the 
Fire Environment Triangle, ignition source is a key factor that should be considered 
for understanding the fire ignition in addition to fuel, weather, and topography.  
4 
 
 
 
 
Figure 1.3. Factors controlling fires at different scales as an extension of the 
commonly used fire triangles. The left triangle represents the Fire Fundamentals 
Triangle and the middle one represents the Fire Environmental Triangle. This figure 
is from Moritz et al. (2005). 
1.2.1 Ignition source 
Wildfire ignition is primarily caused by cloud-to-ground (CG) lightning 
flashes and a variety of human factors including open flames, intentional and 
unintentional arson, equipment sparks, or power lines (Pyne et al., 1996). Majority of 
wildland fires in the remote and sparsely populated boreal forest and tundra 
ecosystems are ignited by lightning strikes (French et al., 2015; Veraverbeke et al., 
2017). Current studies have diverging conclusions regarding the impact of climate 
change on global lightning activity in the future with some predicting an increase 
(Krause et al., 2014; Price & Rind, 1994) while others indicating a decrease (Finney 
et al., 2018; Murray, 2018). Therefore, understanding the factors and mechanisms 
driving lightning activity plays a vital role in understanding current and modeling 
future tundra fire potential.  
5 
 
 
 
Lightning is typically monitored using satellite- or ground-based systems for 
forecasting severe weather conditions and ensuring population safety (Nag et al., 
2014). Satellite-based lightning monitoring primarily focuses on the tropical regions 
or examines the global patterns with geostationary satellites. Due to the limited 
coverage and coarse spatial resolution of satellite-based systems for the HNL, 
lightning monitoring in Alaska is primarily based on the ground-based Alaska 
Lightning Detection Network (ALDN) maintained by the Bureau of Land 
Management (BLM) since the 1980s (Fronterhouse, 2012). According to the ALDN 
observations, boreal forests of the Interior Alaska experience more frequent lightning 
strikes than the tundra (Dissing and Verbyla, 2003), which can be explained by the 
enhancement of local lightning storm development induced by air-mass 
thunderstorms and surface properties of the forests. The spatial distribution of 
lightning strikes throughout Alaska also varies seasonally and interannually. More 
than 90% of the lightning strikes in Alaska occur between June and August (Reap, 
1991). Under relatively warm and dry conditions in the summer, Alaska experiences 
an increased quantity of lightning strikes induced by amplified convective activities, 
thus facilitating fire ignition potential (Peterson et al., 2010; Wendler et al., 2011).  
However, current studies of lightning mechanisms and related impacts on fires are 
limited to the boreal forests and are not readily transferable to the treeless tundra 
ecosystems where known CG lightning and the surface meteorology related to it 
differs substantially (Dissing and Verbyla, 2003). 
6 
 
 
 
1.2.2 Fire weather conditions 
In addition to the weather processes that support lightning, weather elements 
such as air temperature, precipitation, and relative humidity play a dominant role in 
determining fire potential at the regional scale (Pyne et al., 1996). These elements that 
influence fire behavior are typically referred to as fire weather at an hourly or daily 
scale. Increased warming in the HNL regions is likely to lengthen the fire season and 
lead to an increase in burned area (Jolly et al., 2015). Specifically, temperature has 
been widely accepted as the most important factor that affects fire behavior primarily 
through an increase in evapotranspiration and associated decrease in surface moisture 
content (Flannigan et al., 2016; Williams et al., 2015). Additionally, dry conditions 
increase the probability of lighting-induced ignitions (Peterson et al., 2010) and 
enhance fuel flammability (Xiao and Zhuang, 2007) in the North American boreal 
forests.  
Fire danger rating systems have been utilized around the world to quantify 
wildfire potential with numeric indices that guide fire management activities by 
integrating weather elements and other factors (Pyne et al., 1996). Two systems have 
been developed for North America from the early 20th century and updated multiple 
times since then, including the U.S. National Fire Danger Rating System (NFDRS) 
and the Canadian Forest Fire Danger Rating System (CFFDRS). Combining both 
hourly and daily weather observations, the latest 2016 version NFDRS produced four 
primary indices including Ignition Component (IC), Spread Component (SC), Energy 
Release Component (ERC), and Burning Index (BI) (Figure 1.4).  
7 
 
 
 
 
Figure 1.4. Comparison of CFFWIS (left) and NFDRS (right), source: National 
Wildfire Coordinate Group (https://www.nwcg.gov/publications/pms437/cffdrs/fire-
weather-index-system). 
Although originally introduced in 1925, the current CFFDRS has been 
developed since 1968 and widely applied throughout Canada since 1970 (Stocks et 
al., 1989). CFFDRS is comprised of two major sub-systems, including the Canadian 
Forest Fire Weather Index System (CFFWIS) and the Canadian Forest Fire Behavior 
Prediction System (CFFBPS). In particular, CFFWIS developed three fuel moisture 
codes and three fire behavior indices to describe the fuel moisture conditions and 
relative potential of fire behavior using consecutive daily observations of near-surface 
weather (Figure 1.4). Despite the similarities that NFDRS and CFFWIS share in both 
input and output data, these two systems are different in the specific requirements of 
8 
 
 
 
weather observations and assessment methods of fire danger (Pyne et al., 1996). Due 
to its simplicity in data preparation and calculation, the CFFWIS has been widely 
adopted by researchers and fire management teams worldwide in different ecosystems 
(Taylor and Alexander, 2006; Wang et al., 2017).  
1.2.3 Surface fuels 
Fuel type and state are critically important factors that control fire-
environment interactions through altering fire characteristics and affecting ignition 
easiness. (Pyne et al., 1996). Fuel type represents properties of the fuel itself, such as 
fuel composition, continuity and loading. Fuel state is mainly related to moisture 
content primarily driven by the changing weather conditions at different temporal 
scales.   
Fuel classification schemes, such as fuel models, inventory and photo guides, 
are commonly used to describe fuel type variability and organize fuel information. 
LANDFIRE, a program sponsored by U.S. Department of the Interior (DOI) and 
Department of Agricultural Forest Service (FS), has developed consistent surface fuel 
products throughout the US including Alaska and Hawaii, to characterize fuel 
composition and properties based on existing fuel classification systems. In Alaska, 
LANDFIRE provides products for the 13 Anderson Fire Behavior Fuel Model 
(FBFM13), 40 Scott and Burgan Fire Behavior Fuel Model (FBFM40), Canadian 
Forest Fire Danger Rating System (CFFDRS), Fuel Characteristic Classification 
System (FCCS), and Fuel Loading Model (FLM). These categorical classification 
datasets are developed with Landsat imagery using a rule-based approach that 
considers the limited availability of field data from the LANDFIRE reference 
9 
 
 
 
database and the difficulty of describing fuel conditions with generalized methods 
(Rollins, 2009).  
The Alaskan tundra presents a very fine-scale mixture of shrub, herbaceous 
plants, moss and lichen components, which are inseparable even by very high 
resolution (VHR) images and thus is poorly represented by current categorical 
classifications at the Landsat scale. It is crucial to characterize the fractional 
representation of tundra fuel types to better understand the roles of individual 
component in driving tundra fire behavior.  
The cumulative influences of previous and present fire weather conditions can 
further affect the dynamic fuel state characteristics including fuel moisture content 
and fuel temperature (Flannigan et al., 2016). Satellite-based measurements are 
widely used with physical or empirical models to assess fuel moisture levels at a large 
spatiotemporal scale (Yebra et al., 2013). Significant statistical relationships have 
been found between satellite-derived vegetation indices and fuel moisture content in 
different ecosystems (Yebra et al., 2008). 
1.2.4 Topography 
Variations in topographic features including elevation, slope steepness, aspect 
and land configuration also have the potential to affect wildfire behavior (Pyne et al., 
1996). By controlling the exposure to sunlight and moisture pooling, topography can 
modify weather patterns and alter interactions of fuel and weather over time, 
therefore affecting fuel type and moisture content. Elevation above sea level can 
affect the fire environment by influencing climate conditions and fuel availability. 
With a direct impact on fire flame length and spread rate, slope is an important factor 
10 
 
 
 
affecting fire spread behavior. Representing the direction that a slope faces, aspect 
can affect fire behavior by altering the solar radiation amount and wind, therefore 
influencing fire weather and fuel conditions.  
1.3 Fire modeling efforts 
To understand the roles of these environmental factors on wildfire behaviors 
and further model the impacts of these factors on driving fire danger, two types of 
models have been developed in previous studies, including physical-based models 
that account for fire processes separately, and empirical models that utilize statistical 
methods to describe the patterns. Physical-based models that explicitly represent the 
wildfires processes are amongst the earliest attempts of fire occurrence modeling 
(Anderson, 2002; Anderson et al., 2000; Kourtz and Todd, 1991). Several major 
modules related to fire occurrence including lightning occurrence, fire ignition, fire 
smoldering and fire detection were described in these models. Subsequently, a series 
of equations based on experiments and assumptions were developed to incorporate 
the related environmental variables and to model the probability of fire occurrence.  
In recent years, empirical models have been commonly adopted for modeling 
wildfire ignitions in studies across different ecosystems (Prestemon et al., 2013). For 
example, Liu et al. (2012), Yang et al. (2015), and Woo et al. (2017) predicted the 
forest fire risk using spatial point modeling methods. Generalized Linear Models have 
also been utilized for modeling fire occurrences in a variety of landscapes (Ager et 
al., 2018; Barbero et al., 2014; Vilar et al., 2016). Logistic regression and random 
forest (RF) based algorithms are also among the most popular methods for 
establishing the relationships between environmental factors and fire occurrences 
11 
 
 
 
with strong predictive capability (Guo et al., 2016; Keyser and Leroy Westerling, 
2017; Van Beusekom et al., 2018; Vec?n-Arias et al., 2016; Viedma et al., 2018; 
Wotton and Martell, 2005). 
1.4 Research Questions 
Considering the research gaps discussed above, this dissertation aims to 
address the overarching scientific question: What factors and mechanisms drive 
wildland fire ignitions in Alaskan tundra biome? This dissertation examines the 
impacts of surface fuel distribution, ignition sources, and other environmental factors 
on fire ignition probability by integrating field observations, remote sensing products, 
numerical weather forecasting model, and machine learning algorithm. To address the 
overarching scientific question, three integrated studies were conducted in this 
dissertation (Figure 1.4). 
 
Figure 1.5. Framework of Research Questions and Objectives 
12 
 
 
 
Question 1: What are the spatial distribution patterns of major fuel types in Alaskan 
tundra? 
This question examines the spatial distribution of major fuel type components 
in the shrub and graminoid dominated tundra in Alaska. This question focuses on 
developing a fractional cover product of the key tundra fuel components using field 
observations and Landsat 8 imagery at 30m resolution.  
Question 2: What atmospheric factors and mechanisms drive CG lightning activity in 
Alaska tundra? 
This question identified significant atmospheric and cloud properties that 
drive the distribution of CG lightning strikes in Alaskan tundra. This question is 
centered on developing an empirical-dynamical modeling framework for predicting 
CG lightning distribution combining a commonly used Numeric Weather Prediction 
(NWP) model, Weather Research and Forecast (WRF), and RF algorithm, to support 
fire monitoring and modeling effects in the tundra. 
Question 3: How do environmental factors drive fire ignition probability in Alaskan 
tundra? 
This question develops an empirical model for predicting wildfire ignition 
probability in the tundra using Moderate Resolution Imaging Spectroradiometer 
(MODIS) derived fire ignition locations and related environmental factors obtained 
through the integration of remote sensing products and WRF-generated weather 
conditions. Furthermore, this question explores the important environmental factors 
that drive wildfire ignitions in Alaskan tundra through these modeling efforts. 
13 
 
 
 
1.5 Dissertation Structure 
Chapter 1 provides the overall background and motivation of this doctoral 
dissertation. It also introduces the major research questions and the organization of 
this study. The remainder of the dissertation is organized into three separate research 
chapters, a conclusion, supplementary materials and bibliography. 
Chapter 2 focuses on addressing the research question 1 through the 
development of a fractional cover map product of major fuel type components, 
including woody (shrub), herbaceous (sedge and grass) and nonvascular (lichen and 
moss), in Alaskan tundra. I used Landsat 8 Operational Land Imager (OLI) surface 
reflectance data collected from 2013 to 2018 to develop seasonal mosaics of surface 
reflectance and spectral indices across the entire tundra in Alaskan. I then developed 
the fractional cover maps circa 2015 with field observations from three field trips to 
the tundra and the seasonal mosaics. This product has shown strong capability of 
capturing both detailed distribution of vegetation fuel components and describing the 
general spatial patterns of vegetation communities across the ecoregions in Alaskan 
tundra. This work has been peer reviewed and published in Remote Sensing of 
Environment (He et al., 2019). The data product I developed has been archived on the 
Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC; 
He et al., 2020). 
Chapter 3 addresses the research question 2. This chapter develops an 
empirical-dynamical framework to model the CG lightning probability across 
Alaskan tundra based on WRF simulation and RF algorithm. I first conducted a 
sensitivity analysis to identify the optimal parameterization setting combination for 
14 
 
 
 
WRF simulation in the study area. Then four types of atmospheric variables were 
extracted from WRF simulation results to train an empirical model using RF 
algorithm, with CG lightning observations from the ALDN.  
Chapter 4 corresponds to the research question 3 and discusses the results of 
modeling fire ignition probability in Alaskan tundra from 2001 to 2019 with a variety 
of related environmental factors, based on the results from Chapters 2 and 3. In this 
chapter, I prepared environmental factors for the tundra using WRF simulated 
weather variables and remote sensing products. CG lightning probability, near surface 
weather variables and CFFWIS were obtained based on WRF simulations using the 
method developed in Chapter 3, due to the lack of weather stations in the remote 
tundra. Fuel type and fuel state were obtained from Landsat-based product from 
Chapter 2 and MODIS surface reflectance data. I was then able to identify the key 
factors driving tundra fire ignitions in Alaska through modeling efforts with both RF 
and logistic regression algorithms.  
Chapter 5 presents the primary conclusions of this doctoral dissertation. This 
chapter first summarizes the major research findings from Chapters 2 ? 4. It then 
discusses the overall contribution of this dissertation to the broader Arctic scientific 
research and the operational fire management efforts in the HNL. Future insights of 
this research are also discussed in this chapter.  
 
15 
 
 
 
Chapter 2: Mapping Fractional Cover of Major Fuel Type 
Components across Alaskan Tundra1 
 
2.1 Summary 
Wildland fire is common and widespread in Alaskan tundra. Tundra fires exert 
considerable influence on local ecosystem functioning and contribute to climate 
change through biogeochemical (e.g. carbon cycle) and biogeophysical (e.g. albedo) 
effects. These treeless landscapes are characterized by a high degree of variation in 
fuel loading at scales much finer than moderate (30 m) satellite observations. 
However, because of the remoteness of the tundra and its lower contribution to 
carbon release compared to boreal forests, most frequently tundra fuels are poorly 
characterized, limiting the effective development of tundra-specific fire occurrence 
and behavior models. This study presents an approach to mapping the fractional 
coverages of major fuel type components in Alaskan tundra circa 2015 combining 
field data and Landsat 8 Operational Land Imager (OLI) observations. I adopt a 
multi-step method based on random forest (RF) algorithm to estimate the fractional 
vegetation cover of woody, herbaceous, and nonvascular components at subpixel 
level. I demonstrate the strong capability of exploiting multi-seasonal spectral 
information to identify these component types, with R-squared values around 0.9 and 
root mean squared errors below 10% for predicting their fractional cover. The 
 
1 This chapter has been published in Remote Sensing of Environment as He, J., Loboda, T. V, Jenkins, 
L., Chen, D., 2019. Mapping fractional cover of major fuel type components across Alaskan tundra. 
Remote Sensing of Environment, 232, 111324. The output maps of this chapter have been documented 
on Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC) as He, J., T.V. 
Loboda, L. Jenkins, and D. Chen. 2019. ABoVE: Distribution Maps of Wildland Fire Fuel 
Components across Alaskan Tundra, 2015. ORNL DAAC, Oak Ridge, Tennessee, USA. 
16 
 
 
 
mapping products depict the spatial distribution of woody, herbaceous, and 
nonvascular components at subpixel resolution across Alaskan tundra, which can 
function as a critical input for studying wildland fire risk and behavior in the tundra. 
The distributions of these fuel components align well with climate-based tundra 
ecoregions although climate variables are not included in my models. 
2.2 Introduction 
Wildland fire is common across tundra, the coldest vegetated land ecosystem 
on Earth. Active fire products derived from satellite data identify a widespread 
distribution of fire across the pan-Arctic tundra (Masrur et al., 2018). Out of the 
10,260 km2 global burned area in the tundra between 2001 and 2015, 54% was 
concentrated in Alaska as estimated by satellite-based burned area data (Loboda et al., 
2017). Paleoecological and historical records also reveal frequent fire occurrence in 
Alaskan tundra (French et al., 2015; Higuera et al., 2011). 
As a major disturbance in the tundra, wildfire exerts strong influence on the 
ecosystem state and functioning, including deepening of the active layer (Jones et al., 
2015), release of ancient carbon to the atmosphere (Mack et al., 2011), decrease of 
land surface albedo (French et al., 2016) and shift in vegetation communities (Racine 
et al., 2004). These impacts further contribute to climate change through the alteration 
of surface energy budget and global carbon cycle (French et al., 2016; Mack et al., 
2011; Pearson et al., 2013). Rapid climate warming in the Arctic observed during 
recent decades and projected under various climate change scenarios is likely to 
increase tundra fire occurrence in future (French et al., 2015; Young et al., 2017). 
17 
 
 
 
Despite its importance for global biogeochemical and biogeophysical 
processes, tundra fire receives much less attention compared to fire in other 
ecosystems. Although the number of studies on tundra fire has grown considerably in 
recent years, they mainly focus on quantifying post-fire impacts (French et al., 2016; 
Loboda et al., 2013) or examining fire regimes (French et al., 2015; Rocha et al., 
2012). Critical for accurate monitoring of fire potential and for assessing its 
ecological and climatic impacts, in-depth knowledge of fire ecology and improved 
modeling capability of fire occurrence are still lacking for the tundra. Current 
approaches primarily developed upon boreal forest fire studies are thus insufficient to 
establish improved modeling and predictive capability to assess the present and future 
tundra fire potential.  
Wildfire occurrence is controlled through the interaction of fuel, weather, and 
topography (Pyne et al., 1996). Effective modeling of fire occurrence requires 
accurate characterization of these environmental factors. This is, however, difficult 
for the tundra under current efforts due to the lack of in situ data for all three factors. 
Compared to other ecosystems, very limited fuel inventory plots have been visited in 
Alaskan tundra because of its remoteness. Existing fuel type products provided by 
LANDFIRE (https://landfire.gov) for Alaskan tundra (Table 2.1) are primarily 
developed based on the generic fuel classification schemes designed for other 
ecosystems. In particular, 13 Anderson Fire Behavior Fuel Model (FBFM13), 40 
Scott and Burgan Fire Behavior Fuel Model (FBFM40), and Canadian Forest Fire 
Danger Rating System (CFFDRS) include logging slash in their schemes and thus 
cannot describe the surface fuel compositions in the tundra. Limited fire weather 
18 
 
 
 
stations and little to none geodetic surveys within Alaskan tundra also lead to the lack 
of accurate weather and topographical measurements. Primarily built upon remotely 
sensed observations, models and interpolations, existing strategies and datasets 
developed for wildfire monitoring thus simply provide a generic estimation of the 
tundra environment in Alaska. 
Table 2.1. LANDFIRE fuel products and their major fuel strata or types available for 
Alaskan tundra 
Fuel classification system Major fuel strata or types 
13 Anderson Fire Behavior Fuel Model 
(FBFM13; Anderson, 1982) Grass, shrub, timber, logging slash 
40 Scott and Burgan Fire Behavior Fuel Nonburnable, grass, grass-shrub, 
Model (FBFM40; Scott & Burgan, 2005) shrub, timber-understory, timber litter, slash-blowdown 
Canadian Forest Fire Danger Rating Coniferous, deciduous, mixedwood, 
System (CFFDRS; Hirsch, 1996) slash, open 
Fuel Characteristic Classification System Canopy, shrub, nonwoody fuels, 
(FCCS; Ottmar et al., 2007) woody fuels, litter-lichen-moss, ground fuels 
 
Whereas, tundra environment varies at a very fine spatial scale with extremely 
high local heterogeneity (Davidson et al., 2016; Frost et al., 2014; Lara et al., 2018; 
McManus et al., 2012), which makes existing products poorly suited for capturing its 
detailed spatial variability. Specifically, Alaskan tundra presents a highly variable and 
mixed combination of burnable vegetation materials, including shrubs, herbaceous 
plants, mosses and lichens. These fuel type components can affect fire occurrence and 
behavior in multiple ways (Innes, 2013; Rocha et al., 2012). First, fuel flammability 
varies by plant types in Alaskan tundra (Sylvester and Wein, 1981). Live leaves from 
shrub species tend to have higher fire-potential ratings than herbaceous plants. 
Second, the variability of post-fire fuel recovery and accumulation rates among 
19 
 
 
 
vegetation types can lead to different fire frequency levels across the tundra (Innes, 
2013). Tundra regions dominated by grasses or sedges tend to burn more frequently 
than those with shrubs or mosses do because of their faster recovery rates. Third, 
tundra vegetation composition also impacts burned area extent (Rocha et al., 2012).  
For example, larger burned areas are more frequently found within moist shrub and 
tussock tundra compared to non-acidic tundra. Thus, identifying the detailed 
composition and distribution of fuel type components is critical for understanding 
tundra fire occurrence and behavior. 
However, the fine-scale fuel components in the tundra can only be separated 
at centimeter-level resolution with drone data (Figure 2.1) and are not directly 
classifiable even in very high resolution (VHR; 1?5 m) imagery. Categorical 
LANDFIRE products developed at 30-m resolution poorly capture the actual fuel 
distribution in the tundra. Recent efforts, although limited, have been conducted to 
characterize the fractional vegetation properties at subpixel level for Alaskan tundra 
with Landsat imagery. For example, Macander et al., (2017) developed fractional 
coverage maps of plant functional types (PFTs) across the North Slope region of 
Alaska. Berner et al. (2018) also mapped the fractional dominance and aboveground 
biomass (AGB) of shrubs on the North Slope. However, these products are not 
directly transferable to fuel classification schemes in the tundra because of the 
different vegetation cover definitions adopted. They also fail to capture the 
information of other tundra regions like Seward Peninsular and Noatak, which have 
more frequent and intense fire activities than the North Slope based on historical 
records (French et al., 2015; Rocha et al., 2012). 
20 
 
 
 
 
Figure 2.1. Examples of highly mixed fuel type components in Alaskan tundra: (a) 
shrub and graminoid tussocks; (b) grass, lichen and moss. 
While fractional cover products for tundra are only at their early stages of 
emergence, a variety of approaches has been developed to unmix land cover fractions 
with multi-source remote sensing data across different (most frequently tree 
dominated) ecosystems. Supervised regression algorithms with multi-temporal 
spectral metrics are among the most commonly adopted methods for fractional cover 
mapping (Hansen et al., 2013; Olthof and Fraser, 2007; Selkowitz, 2010). RF 
regression, an ensemble learning method based on decision tree regression, has been 
found to have strong capability in distinguishing vegetation fractions (Gessner et al., 
2013; Liu et al., 2017; Marino et al., 2016). A second type of algorithms employs 
spectral mixture analysis to decompose sub-pixel fractional coverages (Guan et al., 
2012; Ma et al., 2015; Mu et al., 2018; Somers et al., 2011). They primarily rely on 
spectral indices to determine the fractions of endmembers. However, the variability 
among endmembers is typically ignored in the modeling procedure (Somers et al., 
2011). A third type of geometric-optical models has also been developed to derive 
21 
 
 
 
vegetation fractions with multi-angular remote sensing data (Chopping et al., 2008), 
though their applications are limited by the spatial data coverage. 
The Arctic-Boreal Vulnerability Experiment (ABoVE) proposed by the 
National Aeronautics and Space Administration (NASA) has provided a new 
opportunity for conducting in-depth research on wildland fire in the tundra. Through 
the support of field campaigns by this program, researchers have been able to enlarge 
the spatial coverage and environmental conditions of field observations. Coupling 
with remote sensing datasets and existing algorithms, these field datasets make it 
possible to develop broad-scale mapping products for the tundra. In this study, I 
present an RF-based approach to mapping the fractional distributions of wildland fuel 
components in Alaskan tundra using multi-spectral and multi-temporal Landsat data 
circa 2015 and a suite of field observations collected across a large span of tussock 
and shrub tundra sites. Specifically, I focus on three fire-carrying fuel type 
components for the shrub or graminoid dominated tundra in the study area: (1) woody 
(shrub) component, (2) herbaceous (primarily sedge and grass) component, and (3) 
nonvascular (lichen and moss) component, and further develop separate maps for 
describing their fractional coverages. I determine these components in a qualitative 
way considering their differences in vegetation genera, fuel characteristics and 
spectral properties. This scheme also corresponds to fuel strata in the Fuel 
Characteristic Classification System (FCCS) and major fuel components in other 
existing systems (Table 2.1).  
22 
 
 
 
2.3 Study Area 
A commonly used dataset, the Circumpolar Arctic Vegetation Map (CAVM; 
Walker et al., 2005; Figure 2.2), was used to determine the study area of Alaskan 
tundra (Beck et al., 2011; Bhatt et al., 2013; French et al., 2015; Raynolds et al., 
2008). Burning in Alaskan tundra is primarily supported by surface vegetation fuels, 
including evergreen or deciduous shrubs, herbaceous species (sedges and grasses), 
mosses and lichens (Hu et al., 2015; Rocha et al., 2012). Dry sites in the tundra tend 
to be dominated by dwarf shrubs with some grasses, mosses, and lichens, while wet 
sites have more sedges and mosses (Sylvester and Wein, 1981). 
Alaskan tundra shares similar vegetation communities and species across 
ecoregions (Alaska Department of Fish and Game, 2006; Viereck et al., 1992). The 
Unified Ecoregions of Alaska identifies four ecoregion groups and eleven ecoregion 
units within Alaskan tundra based on their climate, vegetation, geology, and 
topography (Nowacki et al., 2003; Figure 2.2). With a dry, polar climate, the Arctic 
Tundra ecoregion group is dominated by shrub and sedge tundra underlain by 
continuous permafrost. The Bering Tundra group shows a moist polar or maritime 
climate with principally Dryas-lichen and moist sedge-tussock tundra. Bering Taiga 
group, having a moist polar climate with relatively thin permafrost, is generally 
covered by Dryas-lichen, sedge-shrub tundra and mixed forests. Dominated by a dry 
continental climate, the Intermontane Boreal group are primarily covered by 
shrublands and forests. 
23 
 
 
 
 
Figure 2.2. Alaskan tundra region defined by CAVM and corresponding ecoregions 
identified by the Unified Ecoregions of Alaska: 2001. Each color represents one 
ecoregion unit. Ecoregion units within the same Level 2 ecoregion groups are 
separated by different line patterns. 
2.4 Data and Methods 
The variability of the tundra vegetation types in surface reflectance and 
phenology makes it possible to capture the subpixel compositions using multi-spectral 
and multi-temporal remote sensing data. Dwarf shrub, sedge, and moss/lichen mix 
show distinguishable spectra according to the ground-based hyperspectral profiles 
24 
 
 
 
measured in the North Slope (Buchhorn et al., 2013; Davidson et al., 2016). Both 
field observations and satellite-derived spectral indices also identify considerable 
variations in the phenology patterns of tundra vegetation species during their short 
growing seasons (Beamish et al., 2017; Shaver and Kummerow, 1991; Stow et al., 
1993). For example, graminoid species such as tussock cottongrass (Eriphorum 
vaginatum) and Bigelow's sedge (Carex bigelowii) tend to have an earlier onset of 
leaf expansion compared to shrub species like tealeaf willow (Salix pulchra) and 
dwarf birch (Betula nana; Shaver and Kummerow, 1991). By summarizing the fuel 
properties, spectral characteristics and phenology patterns of these vegetation types 
described above, I identified the following components of surface fuel types in 
Alaskan tundra for large-scale mapping in this study: (1) woody (shrub), (2) 
herbaceous (primarily sedge and grass), and (3) nonvascular (lichen and moss). 
I then designed a four-step method to map the spatial distributions of these 
fuel components (Figure 2.3): (1) developing seasonal composites of spectral bands 
and indices with multi-temporal Landsat imagery for the entire Alaskan tundra; (2) 
generating a ?tundra vegetation mask? to identify the shrub or graminoid dominated 
tundra with land cover classification; (3) RF modeling of fractional fuel component 
cover; (4) mapping fractional cover of major fuel type components across the shrub 
or graminoid dominated tundra in Alaska. 
25 
 
 
 
 
Figure 2.3. Flowchart of fractional fuel component mapping. 
2.4.1 Data Sources 
Landsat observations 
Landsat 8 OLI imagery acquired from 2013 to 2017 was used to develop 
spatially continuous mosaics of Alaskan tundra. I downloaded the Level-2 surface 
reflectance data generated with Landsat 8 Surface Reflectance Code (LaSRC; 
Vermote et al., 2016) from the U.S. Geological Survey Earth Resources Observation 
and Science Center (Table 2.2). Since the growing seasons of tundra vegetation 
generally start from late-April or May and end in October, for each year, I 
26 
 
 
 
downloaded images acquired during late-April through October with land cloud 
coverage smaller than 70%, as estimated by the product metadata. Additionally, I 
excluded images with 80 ? 90% snow coverage over vegetated land surface in late 
April, May and October based on my visual interpretation during the data download. 
In total, I obtained and processed 1837 Landsat 8 scenes covering 113 Worldwide 
Reference System-2 (WRS2) path/rows as summarized in Table 2.2. 
Table 2.2. Summary of Landsat 8 OLI tiles processed in this study. WRS2 path/row 
overlaps exist between different regions. 
WRS2 path WRS2 row Total Total Major regions range range number of number of path/rows scenes 
North Slope and 
Noatak 64 ~ 85 10 ~ 13 70 932 
Seward Peninsula 75 ~ 84 13 ~ 15 25 358 
Southwest Alaska 73 ~ 80 15 ~ 19 29 752 
 
Field observations 
I collected fractional cover observations of the three fuel components from 
222 10?10 m plots during three field campaigns in the tussock and shrub tundra of 
Alaska (Figure 2.4): 2012 campaign in the North Slope (NASA Terrestrial Ecology 
Grant NNX10AF41G), 2016 campaign in the Noatak River National Preserve and 
2017 campaign in the Seward Peninsula (NASA Terrestrial Ecology Grant 
NNX15AT79A). Fractional coverages of fuel components within each plot were 
determined using ocular assessment. All three campaigns were conducted during late-
July to mid-August. Considering the limited amount of field sites that I can visit 
during one single field trip, I combined field data from all three campaigns in this 
27 
 
 
 
study to provide a robust sample collection for driving an ecosystem-wide mapping 
method. 
 
Figure 2.4. Alaskan tundra region as defined by CAVM (a) and three field campaign 
sites: (b) 2016 field plots in Noatak River National Preserve, (c) 2017 field plots in 
Seward Peninsula, (d) 2012 field plots in North Slope. 
Field plots were established following a generally accepted scheme (Dyrness 
and Norum, 1983; Viereck, 1979). During each campaign, the data were collected 
within recovering burns as well as within areas that have not been identified as 
burned by either management records or satellite observations. This data collection 
design ensured that field data observations sampled a broad range of fire history and 
topographical features with varying compositions of woody, herbaceous, and 
nonvascular vegetation cover. The 2016 and 2017 data collection campaigns also 
ensured that a variety of drainage conditions was incorporated into the stratified 
sampling scheme. Specifically I identified four drainage categories based on slope 
28 
 
 
 
and flow accumulation using a method proposed by Kasischke and Hoy (2012). I then 
randomly selected South-East corner point for the 10?10m plots considering these 
factors prior to field visits. Although assessment of fractional cover during the 2012 
field visit was conducted using the similar protocol, the site set up was more 
systematic in order to support a proper characterization of field sites for radar 
observations. In this case, corner points of 10?10 m plots were set up in two parallel 
transect lines 100m apart within an area of visibly uniform conditions following a 
previously established field protocol in existing studies (Bourgeau-Chavez et al., 
2013; Bourgeau-Chavez et al., 2007). 
Auxiliary data 
To assist the development of seasonal composites, I utilized the 16-day 
Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Indices 
product (MOD13Q1 Version 6) available on Google Earth Engine from 2013 to 2017 
to examine the phenology of tundra vegetation and to determine the growing season 
periods for Landsat data collection and processing. The Normalized Difference 
Vegetation Index (NDVI) data layer in MOD13Q1 was used for deriving phenology 
trends. 
Although the quality assessment (QA) band provided by the Level-2 LaSRC 
product identifies water pixels for each scene, I adopted the 30-m ABoVE Surface 
Water Extent data (Carroll et al., 2016) centered in 2011 to identify the representative 
extent of water bodies across the study area for spatial and temporal consistency. 
Using Landsat imagery centered on 1991, 2001 and 2011, this product maps the 
29 
 
 
 
distribution of surface water across the boreal and tundra regions in North America in 
these epochs. 
To develop the tundra vegetation mask in step 3, I utilized a set of 30-m 
LANDFIRE products and VHR imagery freely available on Google Earth to assist the 
sampling of training and validation datasets for land cover classification. I acquired 
the FCCS, CFFDRS and FBFM40 products for Alaska (Table 2.1) in the latest 
available version 1.4.0, which incorporates Landsat imagery through 2014. VHR time 
series imagery provided by Google Earth was also accessed to help determine training 
and validation data through visual interpretation. 
2.4.2 Developing seasonal composites of spectral bands and indices 
First, I developed continuous composites of surface reflectance bands and 
spectral indices for pre-growing (late-April to early-June), peak-growing (mid-July to 
mid-August) and post-growing season (end-August to end-September). Since the 
specific timing of growing events varies by year and latitude, I extracted vegetation 
phenology patterns using NDVI data from MOD13Q1 to determine the specific dates 
of three growing seasons. I particularly examined four graminoid or shrub tundra 
regions as identified by CAVM across Alaska from north to south and compared the 
NDVI profiles of ten randomly sampled pixels in each region (Figure 2.5). According 
to the vegetation growing patterns shown from the profiles, I assigned the Landsat 
data acquired from April 21 to June 10 as the pre-growing season, the data acquired 
from July 1 to August 20 as the peak-growing season, and the data acquired from 
August 30 to September 30 as the post-growing season. 
30 
 
 
 
 
Figure 2.5. Averaged NDVI profiles extracted from MODIS product (MOD13Q1) on 
Google Earth Engine in four regions across Alaskan tundra (a). Ten pixels were 
randomly selected and plotted for each region as an example (b): (1) North Slope, (2) 
Noatak River National Preserve, (3) Seward Peninsula, (4) Southwest Alaska. 
I then downloaded and organized the Landsat 8 data by growing seasons and 
extracted all surface reflectance and QA bands. Four ratio-based spectral indices and 
three Tasseled Cap (TC) components (Brightness, Greenness, and Wetness; Baig et 
al., 2014) were calculated for each scene to capture the detailed variability of 
vegetation spectral characteristics (Table 2.3). Particularly, I considered four ratio-
based indices here including NDVI (Tucker, 1979), Normalized Burn Ratio (NBR; 
Garc?a and Caselles, 1991), Normalized Difference Water Index with near-infrared 
(NIR) and shortwave-infrared (SWIR) bands (NDWI1; Gao, 1996), and Normalized 
Difference Water Index with green and NIR bands (NDWI2; McFeeters, 1996). 
 
 
31 
 
 
 
Table 2.3. Landsat 8 spectral metrics used for developing seasonal composites. 
Categories Properties Details 
Band 2 Blue: 0.45 ? 0.51 ?m 
Band 3 Green: 0.53 ? 0.59 ?m 
Surface Band 4 Red: 0.64 ? 0.67 ?m 
Reflectance Band 5 NIR: 0.85 ? 0.88 ?m 
Band 6 SWIR1: 1.57 ? 1.65 ?m 
Band 7 SWIR2: 2.11 ? 2.29 ?m 
NDVI (?!"# ? ?#$%)/(?!"# + ?#$%) 
Spectral NBR (?!"# ? ?&'"#()/(?!"# + ?&'"#() 
Indices NDWI1 (?!"# ? ?&'"#))/(?!"# + ?&'"#)) 
NDWI2 (?*+$$, ? ?!"#)/(?*+$$, + ?!"#) 
TC Brightness 0.3029?-./$ + 0.2786?*+$$, + 0.4733?#$%
(TCB) + 0.5599?!"# + 0.508?&'"#)+ 0.1872?&'"#( 
Tasseled Cap TC Greenness ?0.2941?-./$ ? 0.243?*+$$, ? 0.5424?#$%
Components (TCG) + 0.7276?!"# + 0.0713?&'"#)? 0.1608?&'"#( 
TC Wetness 0.1511?-./$ + 0.1973?*+$$, + 0.3283?#$%
(TCW) + 0.3407?!"# ? 0.7117?&'"#)? 0.4559?&'"#( 
 
Since Maximum Value Compositing (MVC) approach is effective in 
minimizing the impacts of cloud contamination, off-nadir viewing, aerosol and water 
vapor, I adopted this widely used method for developing seasonal mosaics across 
Alaskan tundra (Holben, 1986; Potapov et al., 2008; Roy et al., 2010; Stow et al., 
2004). Before mosaicking, I masked out the cloud, cloud shadow, and snow pixels 
detected by the CFMask algorithm (Zhu et al., 2015). I further conducted 
morphological dilation for the masked pixels using a disk-shaped structuring element 
with five as the radius to remove the undetected cloud and shadow pixels. Although 
cloud/shadow pixels are generally thought to have lower NDVI than clear-sky pixels, 
pixels along the cloud/shadow edges mixed by shadow and vegetation signals could 
have higher NDVI than the clear-sky ones (Figure 2.6). Since the dilated mask could 
32 
 
 
 
still omit these edge pixels with high NDVI values that can affect MVC results, I 
filtered them out using the following criteria based on empirical values identified in 
the study area: 
?!"# < 	0.2	???	???? > 0.6, 
where ?!"# represents the surface reflectance of NIR band. For each growing season, 
I identified the Landsat pixels with the maximum NDVI and then generated the 
mosaics for each band using the values from these pixels. The three output mosaics 
each include six spectral bands and seven indices as listed in Table 2.3. They then 
served as the input data for the following steps. 
 
Figure 2.6. Unmasked cloud shadow pixels with high NDVI values from a Landsat 8 
scene (Landsat Scene Identifier: LC80690122017228LGN00). Examples are 
33 
 
 
 
highlighted in yellow rectangles: (a) LandsatLook natural color image, (b) NDVI and 
cloud/shadow/water mask identified by level-2 QA data from the same Landsat scene. 
2.4.3 Generating a tundra vegetation mask for shrub or graminoid dominated tundra 
Since CAVM simply defines an approximate tundra boundary based on 1-km 
Advanced Very Higher Resolution Radiometer (AVHRR) data (Walker et al., 2005),  
I further developed a ?tundra vegetation mask? layer to refine the shrub or graminoid 
dominated tundra region using RF classification. I have removed water body pixels in 
the previous step using the nominal water mask (Carroll et al., 2016). Here I 
identified three land cover classes including (1) tall shrub or tree, (2) built-up or 
barren land, (3) shrub or graminoid dominated tundra. 
I first used the three acquired LANDFIRE products (CFFDRS, FBFM40 and 
FCCS) to define the general regions of the three classes based on their classification 
codes (Table 2.4). I then examined the random points by visually interpreting the 
moderate resolution and VHR time series imagery available on Google Earth based 
on expert knowledge. Points that are not representative of the classes they are 
supposed to represent were removed from the sampling dataset. For each class, I 
extracted the intersections of each product as its boundary for generating stratified 
sample points by area. In total, I acquired 436 points for the tall shrub or tree, 1176 
points for the built-up or barren land, and 5238 points for shrub and graminoid tundra. 
 
 
 
34 
 
 
 
Table 2.4. Land cover classes and corresponding classification codes from 
LANDFIRE products. 
Land cover LANDFIRE 
classes products Fuel class or identification code 
Tall shrub or FCCS 85, 87, 88, 89, 92, 93, 94, 101, 103, 105, 322, 332 
tree CFFDRS C1, C2, C3, C4, C5, C6, C7 
Built-up or FCCS 0 
barren land CFFDRS NB1, NB9 
FBFM40 NB1, NB9 
Shrub or 95, 97, 98, 99, 100, 318, 323, 324, 326, 327, 330, 
graminoid FCCS 331, 333, 334, 336, 337, 338, 339, 601, 602, 603, 
dominated 604, 610, 611, 614, 615, 616, 617, 620, 623, 624, 
tundra 625, 627, 629, 630, 632, 635, 637, 638 CFFDRS D1, S2, S3, O1A, M1, M2A, M2B, M2C, M3 
 
I used all bands from the seasonal mosaics as the input features for training 
and mapping the tundra vegetation mask. For each class, I randomly selected 70% of 
the sampled points for training the RF classification algorithm and reserved the 
remaining 30% for assessing the classification accuracy. I also reported the overall 
out-of-bag (OOB) error rate to estimate the classifier error based on the training data. 
Specifically, the OOB error rate is generated by estimating the ratio of 
misclassification among all bootstrap iterations (Breiman, 2001).  
2.4.4 RF modeling of fractional cover 
Before modeling, I compared my designed sampling plots with full fire 
records, including both MODIS Active Fire Product and Alaska Large Fire Database, 
to make sure the plots were not impacted by fire and thus representative of 
undisturbed conditions since the time of the measurement. To predict the fractional 
cover of surface fuel components at the large scale, I trained individual RF regression 
models for the three components separately with the field observations and seasonal 
35 
 
 
 
composites developed in Section 2.4.2. In particular, the spectral properties from the 
three growing seasons, including all six surface reflectance bands, four ratio-based 
spectral indices, and three TC components, were used as the input parameters to train 
the RF models for estimating fractional cover. For each RF regression model, I set the 
number of input variables at each split to 14, the number of trees to grow as 500, and 
the node size to 5.  
To assess the modeling accuracy, I not only examined the internal metrics 
provided by the RF regression algorithm, but also conducted cross-validation with our 
field observations. Specifically, I assessed the internal metrics including OOB, 
percent of variance explained, mean squared error (MSE) and root mean squared 
error (RMSE). For the cross-validation, I reserved 70% field data for model training 
and kept the remaining 30% for validation using a stratified random sampling 
strategy. I then reported the RMSE and R-squared values between the observed and 
estimated fraction values with the 30% validation dataset. For each fuel type 
component, I ran RF regression multiple times and selected an optimal one as the 
final model for estimating fractional cover. 
I also assessed the importance of predictors in determining the fractional 
cover in each model. Although Mean Decrease Gini and Mean Decrease Accuracy 
(MDA; Breiman, 2001) are the most commonly used indicators to assess the 
contributions of independent variables, there is an on-going debate about their 
comparative robustness (Louppe et al., 2013). Here I chose to report the MDA, which 
evaluates the variable importance by estimating the mean decreased MSE with 
permutated variable values. 
36 
 
 
 
2.4.5 Fuel type component mapping 
In the final step, I combined the results generated from the previous sections 
to develop fractional cover maps for the three fuel components. Only shrub or 
graminoid dominated tundra pixels identified by the ?tundra vegetation mask? in 
Section 2.4.3 were subsequently used for mapping. I used the RF regression models 
developed in Section 2.4.4 to estimate the fractional cover of fuel components in 
Alaskan tundra. In addition to the statistical metrics adopted for assessing the RF 
modeling accuracy in Section 2.4.4, I further assessed the mapping results through 
comparisons with existing data products at similar spatial scales considering the 
limited coverage of our field observations. 
I first examined the distribution of fuel component fractions against the 
ecoregions based on the expert knowledge. The variability of climatic and 
topographic conditions across ecoregions in the tundra affects the actual distribution 
of the burnable vegetation materials. Here I utilized the Unified Ecoregions of Alaska 
(Nowacki et al., 2003) product to define the ecoregions within the tundra (Figure 2.2). 
For each fuel component, I generated boxplots to summarize the fraction distributions 
within the two Level 1 ecoregion groups (?Boreal? and ?Polar?). Since most 
ecoregions in Alaskan tundra are elements of the ?Polar? group, I further examined 
the distributions of each ecoregion units within the ?Polar? Level 1 group. 
I then compared my results with existing vegetation cover products developed 
for the North Slope of Alaska to examine the differences caused by fractional cover 
definitions, field sampling strategies, and modeling methods. Although no fractional 
fuel type products have been developed, Beck et al. (2011) mapped the shrub cover of 
37 
 
 
 
deciduous species circa 2000 and Macander et al. (2017) quantified the fractional 
distributions of 16 PFTs across the North Slope of Alaska. I examined these products 
against the field observations. I also compared the overall results between the output 
maps and these two products with randomly selected sample points in the overlapping 
regions of all products. Since the definition of PFTs is not directly transferable to my 
fuel component scheme, here I used the ?Total shrub? of PFT to compare with the 
woody component, the ?Total herbaceous? of PFT to compare with the herbaceous 
component, and the ?Total nonvascular plants? to compare with the nonvascular 
component. Although Berner et al. (2018) also mapped the dominance of shrub in the 
North Slope, they defined the shrub dominance differently as the percentage of shrub 
AGB over the total plant AGB and thus was not considered for comparison. 
2.5 Results 
2.5.1 Accuracy assessment of tundra vegetation mask 
Both internal metrics from the RF algorithm and accuracy assessment were 
used to evaluate the performance of the tundra vegetation mask. The RF algorithm 
showed an overall OOB error rate of 2.21%. I then generated the confusion matrix 
using the reserved 30% data to validate the classification result (Table 2.5). The 
overall classification accuracy reaches 97.91%, although the producer's accuracy and 
user's accuracy vary among each land cover class. The class of tall shrub or tree 
receives the lowest producer's accuracy (77.86%) and user's accuracy (90.27%), while 
the other two classes have consistently high accuracy values. Across the 494,971.4 
38 
 
 
 
km2 Alaskan tundra as estimated by CAVM, I identified 380,755.2 km2 as shrub or 
graminoid dominated tundra. 
Table 2.5. Confusion matrix of land cover classification for developing tundra 
vegetation mask. 
Reference 
 Tall Built-up/ User?s 
shrub/ tree barren Tundra Total accuracy 
Tall shrub/ tree 102 0 11 113 90.27% 
Built-up/ barren 0 351 1 352 99.72% 
Tundra 29 2 1560 1591 98.05% 
Map Total 131 353 1572 2056 1 
Producer?s Overall 
accuracy 77.86% 99.43% 99.24% 1 Accuracy: 97.91% 
2.5.2 Accuracy assessment of fractional cover estimation 
To assess the modeling results, I first examined the internal metrics generated 
from RF regression (Table 2.6). The RF models developed for all three fuel 
components perform well in explaining the percent of the variance, with 79.83%, 
80.76%, and 80.02% for the woody (shrub), herbaceous (sedge and grass), and 
nonvascular (lichen and moss) components respectively. They also report high 
predictive power for fractional cover mapping with low MSE and RMSE values. The 
woody, herbaceous and nonvascular models show MSE values of 0.008852, 0.008117 
and 0.007234, and RMSE values of 9.41%, 9.01%, and 8.51%, respectively. 
Table 2.6. Statistical summary of three RF regression models. 
Fuel component % variance MSE RMSE 
Woody component 79.83 0.008852 9.41% 
Herbaceous component 80.76 0.008117 9.01% 
Nonvascular component 80.02 0.007234 8.51% 
 
39 
 
 
 
I then evaluated the predicted fractional cover using the reserved 30% field 
samples. The comparisons between the modeled and observed values show strong 
agreement for the three fuel components (Figure 2.7). The RSME values for the 
validation samples are within 2% of those generated from the internal model statistics 
with 8.04%, 9.57%, and 10.11% for woody, herbaceous and nonvascular components, 
respectively. The R-squared values between observations and estimations for the 
validation data are approximately 0.95 for all fuel components, with 0.9717, 0.9633, 
and 0.9395 for the woody, herbaceous and non- vascular components respectively. 
 
Figure 2.7. Scatter plots comparing estimated and observed fractional cover using 
the validation data for (a) woody, (b) herbaceous, and (c) nonvascular components. 
RMSE and R-squared values between the estimations and observations are reported 
in the scatter plot of each component. RSMEs are within 2% of those from the 
internal model statistics. R-squared values reach 0.95 for all components. 
Spectral bands and indices during pre- and peak-growing seasons play the 
most important role in determining the fractional cover of woody and herbaceous 
components in Alaskan tundra (Figure 2.8 a-b), while spectral indices during post-
growing seasons contribute most to the cover estimation of nonvascular component 
(Figure 2.8 c). For the woody component, the spectral reflectance of the red band 
40 
 
 
 
during the peak-growing season shows the highest MDA value of 14.73% in the RF 
regression model, followed by that of the blue band during the pre-growing season 
(13.53%) and NDVI during the peak-growing season (11.78%). For the herbaceous 
component, the spectral reflectance of the red band during the pre-growing season 
shows the highest MDA value of 16.66%, followed by NBR during the peak-growing 
season (15.79%) and the spectral reflectance of the green band during the pre-
growing season (13.83%). As for the nonvascular component, NBR during the post-
growing season is the most important independent variable in the regression model 
with an MDA value of 19.15%. NDWI2 and NDVI during the post-growing season 
also show high MDA values (13.77% and 11.75% respectively) in determining the 
fractional cover. Although RF regression can alleviate the multicollinearity issue 
through bootstrap aggregation, biases in variable importance can still exist among 
correlated features. 
 
Figure 2.8. Top 10 important independent variables and their MDA values from RF 
regression models for (a) woody, (b) herbaceous and (c) nonvascular components. 
Spectral bands and indices during pre- and peak-growing seasons play the most 
important role in determining the fractional cover of woody and herbaceous 
41 
 
 
 
components (a-b), while spectral indices in the post-growing seasons contribute most 
to the cover estimation of nonvascular component (c). 
2.5.3 Fractional cover of major fuel type components across Alaskan tundra 
The fractional distributions of three fuel components were mapped across the 
shrub and graminoid tundra in Alaska circa 2015 (Figure 2.9). From the south to the 
north of Alaska, shrub cover shows a slightly decreasing trend as the temperature falls 
(Figure 2.9 a). The Southwest Alaska shows a larger portion of the area with high 
shrub fraction compared to the North Slope and the Seward Peninsula. For the North 
Slope and the Seward Peninsula, shrub cover is higher along the rivers (Figure 2.9). 
The herbaceous component of sedge and grass is dominant and widely distributed 
across the entire study area (Figure 2.9 b). In particular, the central North Slope has a 
high fractional cover of sedge and grass. As can be expected, the nonvascular 
component is highly concentrated in the northern part of the North Slope (Figure 2.9 
c). From the north to the south across the entire tundra region, a general decreasing 
trend in the distribution of lichen and moss is clearly observed. 
A closer examination of my mapping results in the Noatak River National 
Preserve shows an increase in the shrub fraction along the drainages (Figure 2.10 b), 
while other regions have higher coverages of the herbaceous fuels including sedge 
and grass (Figure 2.10 c). The amount of the nonvascular component is generally low 
in this example and is most frequently observed in high concentrations only close to 
the barren land along the river or the mountains (Figure 2.10 d). 
42 
 
 
 
 
Figure 2.9. Fractional cover of three major fuel type components across Alaskan 
tundra: (a) woody, (b) herbaceous, and (c) nonvascular components. 
43 
 
 
 
 
Figure 2.10. Examples of fractional cover distributions in a tundra region near Lake 
Narvakrak in the Noatak River National Preserve: (a) VHR image from Google 
Earth; (b) woody component; (b) herbaceous component; (d) nonvascular 
component. 
I then summarized the spatial distributions of fractional fuel cover by the 
unified ecoregions in Alaskan tundra to examine their patterns. As expected, the 
comparison between the two Level 1 schemes (Boreal and Polar) shows a higher 
coverage of the woody component within the ?Boreal? scheme (Figure 2.11 a). The 
?Boreal? scheme has a mean shrub coverage of 37.96% while that of the ?Polar? 
scheme has a lower mean value of 31.57%. The ?Boreal? scheme has an interquartile 
range (IQR) of 33.23% to 43.47%, while the IQR of ?Polar? ranges from 24.23% to 
37.82%. Since the ?Polar? scheme includes three Level 2 groups and nine ecoregion 
units, I further examined the distribution of shrub coverage within this scheme. From 
44 
 
 
 
north to south, the shrub coverage increases gradually among all the Level 2 groups 
and the ecoregion units (Figure 2.11 b). The woody component cover of units in 
?Arctic Tundra? and ?Bering Tundra? is comparable but lower than that in the 
?Bering Taiga? in general. Specifically, the ?Beaufort Coastal Plain? located in the 
northernmost part of Alaska has the lowest shrub cover on average of 27.61%, while 
the southernmost ?Ahklun Mountains? unit has the highest mean cover of 39.04%.  
 
Figure 2.11. Boxplots of woody component cover against ecoregions in Alaskan 
tundra: (a) woody component cover by Level 1 scheme; (b) woody component cover 
by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion units are colored 
from light yellow to dark blue based on the latitude from north to south). Shrub 
coverage increases gradually among all the Level 2 groups and the ecoregion units, 
from north to south (b). 
For the herbaceous component (sedge and grass), the comparison between the 
two Level 1 schemes (Boreal and Polar) suggests no obvious differences (Figure 2.12 
a). The ?Boreal? scheme shows a mean herbaceous coverage of 38.76% while that of 
45 
 
 
 
the ?Polar? scheme has a higher value of 39.58%. The IQR of ?Boreal? and ?Polar? 
schemes are about 33.77% ? 43.23% and 33.28% ? 45.68% respectively. The mean 
fractional cover values of all Level 2 groups are also comparable in general, with 
values of about 38% to 40%. The sedge/grass distribution of units within the ?Bering 
Taiga? group is consistent (Figure 2.12 b). However, the ?Brooks Foothills? in 
?Arctic Tundra? and the ?Bering Sea Islands? in ?Bering Tundra? tend to have a 
higher herbaceous cover than other units within groups do. 
 
Figure 2.12. Boxplots of herbaceous component cover against ecoregions in Alaskan 
tundra: (a) herbaceous component cover by Level 1 scheme; (b) herbaceous 
component cover by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion 
units are colored from light yellow to dark blue based on the latitude from north to 
south). Herbaceous cover values among Level 2 groups are generally comparable, 
with relative higher values in the ?Brooks Foothills? and ?Bering Sea Islands? units 
(b). 
46 
 
 
 
For the nonvascular (lichen and moss) component, the comparison between 
the Level 1 schemes shows a higher cover of nonvascular vegetation within the 
?Polar? scheme (Figure 2.13 a). The ?Boreal? scheme shows an average coverage of 
33.95% while that of the ?Polar? scheme has a higher value of 41.32%. The ?Boreal? 
scheme has an IQR of 15.24% to 40.11%, while that of the ?Polar? scheme ranges 
from 23.41% to 56.99%. From north to south, the lichen/moss cover shows a slightly 
decreasing trend among the Level 2 groups and the ecoregion units (Figure 2.13 b). 
The mean values of lichen/moss cover in the ?Arctic Tundra?, ?Bering Tundra? and 
?Bering Taiga? are about 47.14%, 36.90%, and 30.04% respectively. Specifically, the 
?Beaufort Coastal Plain? and ?Brooks Range? show the highest lichen/moss covers 
compared to other units. Within the ?Bering Tundra? and ?Bering Taiga? groups, the 
nonvascular fractional cover also decreases gradually as the latitudes become lower. 
 
Figure 2.13. Boxplots of nonvascular component cover against ecoregions in Alaskan 
tundra: (a) nonvascular component cover by Level 1 scheme; (b) nonvascular 
component cover by Level 2 ecoregion groups in the ?Polar? scheme (ecoregion 
47 
 
 
 
units are colored from light yellow to dark blue based on the latitude from north to 
south). Nonvascular coverage shows a decreasing trend among the Level 2 groups 
and the ecoregion units from north to south (b). 
I further compared my outputs with existing fractional vegetation cover 
products developed for the tundra. A scatter plot comparison between our field 
observations and the shrub cover circa 2000 (Beck et al., 2011) suggests that the 2000 
shrub cover product overestimates the actual fractional cover of shrub in the North 
Slope (Figure 2.14 a). Compared to the field observations of fractional cover 
collected in the North Slope, the PFT maps circa 2015 (Macander et al., 2017) show 
an overestimation of shrub cover and an underestimation of lichen/moss cover, while 
the fractional cover values of herbaceous species (sedge/ grass) are relatively 
comparable (Figure 2.14 b). 
 
Figure 2.14. Scatter plots comparing our field observations and estimations from 
other products: (a) 2000 shrub cover (Beck et al.,  2011); (b) PFT products 
(Macander et al., 2017). Compared to our field observations, the PFT product 
(Macander et al., 2017) tends to overestimate the shrub cover, but underestimate the 
48 
 
 
 
lichen/moss cover. The fractional cover of herbaceous species (sedge/grass) is 
relatively comparable. 
The comparison between my maps and the existing products using 
approximately 20,000 randomly selected pixels in the North Slope (Figure 2.15) 
indicates similar patterns to those observed in comparing other products to field data 
directly. Fractional cover values of woody component (shrub) in both products are 
higher than the estimations in my results (Figure 2.15 a?b). Specifically, the majority 
of values were identified as 100% or 0% in Beck et al. (2011) in the North Slope. The 
fractional cover values of herbaceous species (sedge/grass) are comparable between 
the result and the product developed by Macander et al. (2017), ranging from about 
20% to 75% (Figure 2.15 c). However, the coverage values of nonvascular 
component (lichen/moss) are much higher in my result when compared to those in the 
PFT product (Figure 2.15 d). The PFT product suggests that the nonvascular species 
including lichen and moss have a general coverage of 0% ? 25% in the North Slope, 
while my result indicates that lichen/moss can cover from 0% to 80%, mostly 
concentrated between 10% and 50% (Figure 2.15 d). 
49 
 
 
 
 
Figure 2.15. Density plots comparing my results and estimations from existing 
products: (a) woody component cover circa 2000 by Beck et al. (2011); (b) woody 
component cover circa 2015 by Macander et al. (2017); (c) herbaceous component 
cover circa 2015 by Macander et al. (2017); (d) nonvascular component cover circa 
2015 by Macander et al. (2017). Fractional cover values of woody component in both 
products are higher than the estimations in my results.  Nonvascular component tends 
to have much higher coverage than in the PFT product. 
2.6 Discussions 
Woody species in Alaskan tundra usually have higher surface reflectance in 
the NIR to SWIR spectrum but lower reflectance in the visible spectral regions 
50 
 
 
 
compared to herbaceous species such as sedge and grass (Buchhorn et al., 2013; 
Strauss et al., 2012). Previous studies have demonstrated the importance of multi-
spectral bands and NDVI in estimating the fractional coverage of shrub (Kushida et 
al., 2009; Selkowitz, 2010; Vierling et al., 1997). Multi-seasonal information and 
middle-infrared portion of the spectrum also contribute strongly to the accuracy of 
shrub cover prediction (Selkowitz, 2010). Both woody and herbaceous models 
demonstrate the importance of pre-growing season spectral information in 
distinguishing the woody and herbaceous components. This can be explained by their 
different phenological stages (Shaver and Kummerow, 1991). The in-situ 
measurements of surface reflectance have shown that the lichen/moss component has 
different spectrum patterns when compared to woody or herbaceous species 
(Buchhorn et al., 2013; Huemmrich et al., 2013; Strauss et al., 2012). My modeling 
result suggests that spectral indices integrating these bands in the post-growing 
season are effective in separating lichen/moss from woody and herbaceous 
components. This is likely to be driven by the exposure of spectral signals by 
lichen/moss due to the senescence of vascular species in the tundra during September.  
In addition to spectral bands provided by 30-m Landsat imagery, existing 
research has suggested that higher spatial resolution and red-edge (RE) spectral bands 
have the potential to improve fractional vegetation mapping efforts in the tundra 
(Davidson et al., 2016; Selkowitz, 2010; Stow et al., 1993). The launch of Sentinel-2 
Multi- Spectral Instrument incorporating three RE bands provides the opportunity to 
improve the estimation accuracy of surface vegetation distribution at a broad spatial 
scale. 
51 
 
 
 
Developed only using spectral signatures, my mapping results also show 
effectiveness in representing the fuel component distribution across ecoregions with 
different bioclimatic conditions. The ?Boreal? ecoregion scheme shows a higher 
cover of the woody component but a lower cover of the nonvascular component than 
the ?Polar? ecoregion scheme in general. Mostly located in interior Alaska with a dry 
continental climate, ecoregions in the ?Boreal? scheme tend to be dominated by 
shrubs and even boreal forests in these mountainous regions. Within the ?Polar? 
scheme, the fractional cover of fuel components identified in this study also 
corresponds to the bioclimatic environment based on expert knowledge (Gallant et 
al., 1995; Nowacki et al., 2003). The low fractional cover values of the woody 
component in the Beaufort Coastal Plain and Brooks Foothills ecoregions are likely to 
be driven by their polar climate and poor soil drainage conditions. With higher 
summer temperature and more annual precipitation then other tundra areas, 
ecoregions as part of the ?Bering Taiga? group (Nulato Hills, Yukon-Kuskokwim 
Delta and Ahklun Mountains) are covered with more woody fuels such as dwarf or 
tall scrub communities with wet graminoid species dominating in some regions, 
which is also evident in my maps. 
Although the accuracy assessment shows strong predictive capability for 
fractional fuel mapping in Alaskan tundra, uncertainties still exist in the modeling and 
mapping results due to the limited spatial coverage of field observations, partially 
caused by the difficulty of obtaining field observations in the tundra. Regions with no 
available field data such as the Southwest Alaska and the northern North Slope could 
have higher uncertainties in my fractional cover maps. The mismatch of spatial 
52 
 
 
 
resolutions between the 10?10 m field plots and the 30?30 m Landsat 8 pixels could 
also introduce errors in the modeling and validation processes. During field 
campaigns, we have made our efforts on establishing sample plots in areas with 
visibly homogeneous distributions of vegetation species. However, it is possible that 
the actual vegetation coverages across one Landsat pixel may not be fully represented 
by the smaller than pixel field plots. Since the impacts of sample plot size on 
modeling and validation of fractional coverages is still unknown, future studies can 
elaborate on this and provide insights. 
In addition, the results largely rely on the quality of the seasonal mosaics 
developed across the entire study area. First, maintaining the spatial consistency of 
these mosaics in Alaskan tundra is hindered by the pervasive cloud coverage and 
incomplete masking of cloud and cloud shadow pixels in the Landsat data. Although 
the CFMask algorithm delineates the major regions of cloud and shadow, it fails to 
identify all cloud- and shadow-impacted pixels. Second, matching multiple 
phenological stages of vegetation across different regions is challenging because of 
the differences in illumination properties and vegetation growing states under 
different climatic conditions (Muller et al., 1999; Selkowitz, 2010). Therefore, in 
order to maintain spatially and temporally consistent mosaics, I adopted the MVC 
method based on NDVI values with carefully defined growing periods using multi-
year Landsat imagery. 
It is also worth mentioning that the mismatches between my results and 
existing products can be caused by the differences in the cover definition and field 
sampling strategy. Beck et al. (2011) defined two shrub types ? tall (> 1 m) and short 
53 
 
 
 
? and mapped the total and tall shrub coverages. Here I chose the total shrub cover for 
comparison, which can explain why my estimates are lower. Macander et al. (2017) 
developed the fractional distribution maps of detailed PFTs. I used their integrated 
coverages of total shrubs, total herbaceous and top nonvascular plants for 
comparison, which can lead to the differences in my result comparisons. In addition, 
instead of collecting field measurements, Beck et al. (2011) extracted sample data by 
aggregating classified pixels from VHR imagery for training their regression 
algorithm, which could lead to the overestimation the actual shrub coverage in their 
mapping output. Moreover, Macander et al. (2017) collected field data from sites 
widely distributed across the North Slope, while our field data had no coverage within 
the Beaufort Coastal Plain ecoregion and only limited sites in the Brooks Foothills. 
This could explain the differences in the estimated cover of the nonvascular 
component in those regions between my result and that of Macander et al. (2017). 
2.7 Conclusions 
In this study, I deliver the first fractional cover maps of major fuel type 
components across Alaskan tundra circa 2015. I map the spatial distributions of 
woody, herbaceous and nonvascular components in this highly heterogeneous 
landscape. Although our field observations are comparatively limited in quantity and 
spatial coverage, the findings are broadly consistent with expected distribution 
according to bioclimatological conditions.  
I present a method of using multi-spectral and multi-seasonal observations in 
the differentiation of fuel components. The results show that this combination offers 
strong predictability from moderate resolution data and thus is critical for mapping 
54 
 
 
 
efforts at a broad spatial scale. This method can be adopted in monitoring other 
vegetation properties such as vegetation dominance and biomass. 
In addition, the long-term archive of moderate resolution data in Alaska offers 
an opportunity to examine the fuel composition change in Alaskan tundra as a result 
of both climate change and fire occurrence. The outputs and spectral indicators 
identified in this paper can assist long-term monitoring of fuel type components in the 
tundra. These fuel maps and their periodic updates can create a strong basis for 
enhancing modeling capabilities for both assessing fire behavior and post-fire 
ecological impacts.   
55 
 
 
 
Chapter 3: Modeling cloud-to-ground lightning probability in 
Alaskan tundra through the integration of Weather Research and 
Forecast (WRF) model and machine learning method2 
 
3.1 Summary 
Wildland fires exert substantial impacts on tundra ecosystems ranging from 
biogeochemical impact on climate system to habitat suitability for various species. 
Lightning is the primary ignition source of wildfires and it is critical to understand 
mechanisms and factors driving cloud-to-ground (CG) lightning strikes in this cold 
treeless environment to support operational forecasting and future modeling of fire 
activity. The existing CG lightning studies primarily focus on Alaskan and Canadian 
boreal forests where land-atmospheric interactions are different and, thus, not likely 
to be representative of tundra conditions. In this study, I designed an empirical-
dynamical method integrating Weather Research and Forecast (WRF) simulation and 
machine learning algorithm to model the probability of CG lightning strikes across 
Alaskan tundra between 2001 and 2017. This study recommended using Thompson 2-
moment and Mellor-Yamada-Janjic (MYJ) schemes as microphysics and planetary 
boundary layer parameterizations for WRF simulations in the tundra. The modeling 
and forecasting test results have shown strong capability of predicting CG lightning 
probability in Alaskan tundra, with Area Under the Curve (AUC) values above 0.9. I 
found that parcel lifted index (PLI) and vertical profiles of atmospheric variables, 
including geopotential height, dew point temperature, relative humidity (RH) and 
 
2 This chapter is under review in Environmental Research Letters, the special issue of ?Resiliency and 
Vulnerability of Arctic and Boreal Ecosystems to Environmental Change: Advances and Outcomes of 
ABoVE (the Arctic Boreal Vulnerability Experiment)?. 
56 
 
 
 
velocity speed, important in predicting lightning occurrence, suggesting the key role 
of convection in lightning formation in the tundra. This method can be applied to 
data-scarce regions and support future studies of fire potential in the High Northern 
Latitudes (HNL). 
3.2 Introduction 
Wildfire is a primary disturbance across boreal forest and tundra ecosystems 
in the pan-Arctic region (Bond-Lamberty et al., 2007; French et al., 2015; Goetz et 
al., 2005). Although much more rare and generally less severe than boreal fires, 
tundra fires can alter ecosystem functioning through biogeophysical and 
biogeochemical processes and drive environmental changes in carbon cycling and 
energy budget (Bret-Harte et al., 2013; French et al., 2016; Jones et al., 2015; Mack et 
al., 2011; Pearson et al., 2013). Specifically, they release large stores of carbon 
generally locked in organic soil and permafrost: for example, the 2007 Anaktuvuk 
River fire on the North Slope of Alaska burned 1,039 km2 and released ~2.1 Tg 
carbon into the atmosphere (Mack et al., 2011). Although most tundra fires are not 
nearly as severe as the Anaktuvuk River fire event, they are common, particularly in 
Alaska. Alaskan tundra burns more than any other tundra regions worldwide, 
accounting for 54% of the 10,260 km2 burned area in the tundra worldwide between 
2001 and 2015, as estimated by satellite observations (He et al., 2019). With very 
short growing periods in the HNL, post-fire recovery for critical components of 
tundra ecosystem, e.g. lichen, can last for several decades, which is comparable to 
recovery rate of the boreal forests (Jandt et al., 2008; Racine et al., 1987). As a result, 
57 
 
 
 
tundra fires impact long-term winter forage availability for caribou and subsequently 
influence the subsistence resources of local communities (Gustine et al., 2014). 
Fires in the remote and generally inaccessible HNL are primarily ignited by 
CG lightning (French et al., 2015; Veraverbeke et al., 2017). Future climate 
projections indicate potential increase of CG lightning under warming conditions 
(Price and Rind, 1994a; Romps et al., 2014), which will subsequently lead to more 
fire occurrence and larger burned area in both boreal forest and tundra ecosystems  
(French et al., 2015; Krause et al., 2014; Veraverbeke et al., 2017; Wotton et al., 
2010). Numerous studies have examined characteristics (Dissing and Verbyla 2003; 
Kochtubajda et al. 2019; Reap 1991; Farukh et al. 2011), explored driving factors and 
developed predictive models (Blouin et al., 2016; Burrows, 2008; Burrows et al., 
2005) for CG lightning activity in Alaskan and Canadian boreal forests. However, 
considerably less is known about factors driving CG lightning for the treeless tundra. 
The substantial differences in surface layer conditions between tree-dominated and 
treeless landscapes (Beringer et al., 2005; Dissing and Verbyla, 2003; Rivas Soriano 
et al., 2019; Van Heerwaarden and Teuling, 2014) imply that understanding of 
lightning processes in the boreal forests is not necessarily readily transferable to the 
tundra. Therefore, tundra-focused studies are critical for enhancing the modeling 
capability of lightning and fire potential and assisting wildfire management efforts in 
future.  
Typically, lightning formation is associated with atmospheric convection in 
cumulonimbus clouds (Anderson, 1992). Lightning flashes are generated through the 
buildup, separation, and transfer procedures of electric charges between cloud 
58 
 
 
 
particles (Saunders, 2008; Yair, 2008). The occurrence and intensity of lightning 
activity are generally related to factors such as convective cloud development, cloud 
structure, and hydrometeor attributes (Baker et al., 1999; Buiat et al., 2017; Price and 
Rind, 1994b). However, explicit simulation and prediction of the electrification 
processes can be computationally expensive (Zepka et al., 2014). Further efforts are 
also required  to comprehensively understand the detailed microphysical procedures 
contributing to charge accumulation (Rakov and Uman, 2003; Saunders, 2008). CG 
lightning modeling thus primarily rely on developing its relationships with observed 
or model-resolved parameters related to convective activities and cloud microphysical 
properties. 
Classificatory schemes or simple regression methods developed with 
convective indices or weather conditions from observations or weather model outputs 
were among the early attempts for lightning modeling (Anderson, 1991; Andersson et 
al., 1989; Fuquay, 1980; Reap and Foster, 1979; Reap and MacGorman, 1989; Sly, 
1965). With the development of General Circulation Models (GCMs) and Numerical 
Weather Prediction (NWP) models, lightning schemes based on microphysics 
principles have been parameterized upon these models from regional to global scales 
(Barthe et al., 2010; Lynn et al., 2012; Price and Rind, 1994b; Wong et al., 2013; Yair 
et al., 2010). Simple strategies were also adopted to improve the modeling capability 
with WRF simulations (Giannaros et al., 2015; Zepka et al., 2014). 
In addition to physical parameterizations, empirical-based methods such as 
logistic regression (Bates et al., 2018; Shafer and Fuelberg, 2006) and random forest 
(RF; Blouin et al., 2016; Sch?n et al., 2018) were developed in recent studies to 
59 
 
 
 
model CG lightning occurrence based on dynamic meteorological conditions and 
thunderstorm characteristics. Opportunities for integrating dynamic NWP and 
statistical models have been explored to improve the modeling capability of lightning 
potential. Burrows et al. (2005) and Burrows (2008) trained tree-structured regression 
models to forecast lightning probability using predictors generated from the Global 
Environmental Multiscale (GEM) model. Sousa et al. (2013) and Gijben et al. (2017)  
also combined NWP models such as WRF with logistic regression to develop 
statistical-dynamical methods for lightning prediction in different regions.  
Due to the remoteness of the tundra, meteorological observations, including 
weather stations and atmospheric soundings, are sparsely distributed and thus 
unsuitable for describing the spatial variation of tundra conditions. Although 
reanalysis products provide atmospheric variables with spatial-temporal consistency, 
their performances are limited by the coarse spatial resolution, availability of 
observations, and uncertainty of diagnostic variables (Dee et al., 2016). Therefore, 
purely empirical models trained with observations or reanalysis data are unsuitable 
for lightning modeling in data-scarce regions like tundra.  
Although NWP has not been specifically applied in tundra studies, existing 
research has demonstrated its suitability and effectiveness for modeling lightning 
potential (Burrows et al., 2005; Reap, 1991) and fire danger (Di Giuseppe et al., 
2016; M?lders, 2010, 2008) in the boreal forests. Considering that tundra has 
different land-atmosphere interactions and fewer meteorological stations available for 
measuring near-surface and atmospheric conditions when compared to boreal forests, 
this study focuses on improving the modeling capacity for CG lightning probability 
60 
 
 
 
and understanding what atmospheric factors drive CG lightning in Alaskan tundra 
through the integration of WRF and RF, a commonly used machine learning 
algorithm. I first assessed WRF parameterization schemes to identify an optimal 
combination for reproducing the observed meteorological conditions in the tundra. I 
then examined data from 2001 to 2019 to understand what atmospheric factors drive 
CG lightning probability across Alaskan tundra.  
3.3 Study Area 
In Alaska, more than 99% of CG lightning occurs from May to August, 
particularly during June and July (Mcguiney et al., 2005; Reap, 1991). A typical 
diurnal pattern exists starting from noon and lasting until midnight with a peak 
between 4pm and 8pm. Elevation and forest cover can affect the spatial variation of 
lightning in Alaska by altering the convective activity (Dissing and Verbyla, 2003). In 
particular, large-scale atmospheric instability and local convergence are the major 
contributors to thunderstorm formation and lightning occurrence in Alaska (Reap, 
1991). Here I defined Alaskan tundra using the Circumpolar Arctic Vegetation Map 
(CAVM; Figure 3.1 a; Walker et al. 2005). Fire regimes vary by year and across 
different tundra regions (Figure 1.2). 
61 
 
 
 
 
Figure 3.1. Study area: (a) Alaskan tundra defined by CAVM; (b) WRF simulation 
domains. 
3.4 Materials and Methods 
I designed an empirical-dynamic method combining WRF-simulated 
atmospheric variables and ground-based lightning observations to model CG 
lightning probability in Alaskan tundra (Figure 3.2). I first conducted a sensitivity 
analysis to identify an optimal WRF parameterization scheme that best describes 
tundra meteorological conditions. I then ran WRF for selected case studies and 
trained RF models for predicting CG lightning probability. 
62 
 
 
 
 
Figure 3.2. Overall workflow of Chapter 3 
3.4.1 Cloud-to-ground lightning observations 
I obtained lightning data from the Alaska Lightning Detection Network 
(ALDN) maintained by the Bureau of Land Management (Fronterhouse, 2012). This 
system has a detection efficiency better than 5km and positional accuracy higher than 
70%, as estimated by early studies (Dissing and Verbyla, 2003; Reap, 1991). It has 
been updated multiple times to improve detection performance (Fronterhouse, 2012). 
Specifically, devices employed before 2012, developed by Vaisala, Inc., recorded 
lightning flash with multiplicity (i.e., lightning strikes per flash). The new system, 
provided by TOA Systems, Inc., records lightning strikes instead (Fronterhouse, 
2012). To ensure the consistency of lightning records between systems, I utilized 
lightning strikes rather than flashes in this study. 
3.4.2 WRF model setup and sensitivity analysis 
I adopted the Advanced Research WRF version 4.0 (Skamarock et al., 2019) to 
simulate weather conditions in the tundra. I used the National Centers for 
Environmental Prediction Final Operational Model Global Tropospheric Analysis 
63 
 
 
 
data (NCEP FNL) at 1??1? resolution and 6-hour interval for model initialization 
(NCEP, 2000). Developed with the Global Data Assimilation System (GDAS), this 
product provides a variety of parameters to describe initial and boundary 
meteorological conditions on the surface and at 26 mandatory pressure levels. I 
defined two domains with 25-km (Domain 1) and 5-km (Domain 2) grid spacing for 
two-way nested WRF simulation (Figure 3.1b). The vertical dimension was 
configured with 33 unevenly spaced full sigma levels with the model top at 50hPa. 
WRF provides multiple parameterization schemes for its major physics components 
(Table 3.1; Skamarock et al., 2019). The various assumptions and mechanisms 
adopted in these schemes can affect simulation results for a specific region. 
Since existing applications of WRF in the HNL primarily focus on the boreal 
forests or the pan-Arctic region, their schemes may not be suitable for tundra. 
Therefore, I employed a sensitivity analysis to determine WRF settings that achieve 
the closest description of tundra conditions compared to meteorological observations. 
I first reviewed existing literature to explore a list of WRF schemes as candidates 
(Table 3.1). The majority of the literature utilized the same schemes for land surface 
model, radiation, and cumulus components, namely Noah, Rapid Radiative Transfer 
Model (RRTM), and Grell-Devenyi, respectively (Table 3.1). These schemes were 
then applied with updated versions in WRF 4.0. Microphysics schemes explicitly 
resolve physical processes related to water vapor, cloud, and precipitation, with 
mixed-phase ones recommended for simulating icing or convective conditions at 
horizontal resolution finer than 10 km (Skamarock et al., 2019). I thus selected 
Morrison 2-moment, Thompson 2-moment, and WRF 2-moment 6-class (WRF6) for 
64 
 
 
 
microphysics in the sensitivity analysis. Mellor-Yamada-Janjic (MYJ)/ Monin-
Obukhov MM5, and Yonsei University (YSU) / Monin-Obukhov Eta combinations 
were also considered for Planetary Boundary Layer (PBL) / Surface Layer (SL) 
representation. All six candidates of schemes are summarized in Table 3.2.  
Table 3.1. WRF schemes used for the pan-Arctic region in existing literature 
Components Schemes References 
Morrison 2-moment (Bieniek et al., 2016) 
Thompson 2-moment (Kim et al., 2014; M?lders, 2010, 2008) 
Microphysics WRF single-moment 6- (Cai et al., 2016; Hines and 
class Bromwich, 2008) 
Goddard Cumulus (Cassano et al., 2011; Glisan et al., 
Ensemble Model 2013) 
Rapid Radiative Transfer (Bieniek et al., 2016; Cai et al., 2016; 
Longwave Model (RRTM) 
Glisan et al., 2013; Hines et al., 2011; 
Kim et al., 2014; M?lders, 2010, 
radiation Longwave 2008) 
Community Atmospheric (Cassano et al., 2011; Glisan et al., 
Model (CAM) Longwave 2013) 
RRTM Shortwave (Bieniek et al., 2016; Glisan et al., 2013; Kim et al., 2014) 
Shortwave Dubhia Shortwave (Cai et al., 2016; M?lders, 2010, 
radiation 2008) Goddard Shortwave (Hines et al., 2011) 
CAM Shortwave (Cassano et al., 2011; Glisan et al., 2013) 
Cumulus Grell-Devenyi ensemble 
(Bieniek et al., 2016; Cassano et al., 
2011; Glisan et al., 2013) 
Kain-Fritsch (Cai et al., 2016) 
Planetary Mellor-Yamada-Janjic (Bieniek et al., 2016; Cassano et al., 
Boundary (MYJ) / Monin-Obukhov 2011; Glisan et al., 2013; Hines et al., 
Layer (PBL) Eta (Eta) 2011) 
/ Surface Yonsei University (YSU) / Monin-Obukhov MM5 (Cai et al., 2016; Kim et al., 2014; layer (MM5) M?lders, 2010, 2008) 
(Bieniek et al., 2016; Cai et al., 2016; 
Land surface Noah Cassano et al., 2011; Glisan et al., 
model 2013; Hines et al., 2011) 
Rapid Update Cycle (M?lders, 2010, 2008) 
 
65 
 
 
 
Table 3.2. Six candidates of parameterization combinations for sensitivity analysis 
Combination notation Microphysics PBL + SL 
Morrison_MYJ Morrison 2-moment MYJ + Eta 
Thompson_MYJ Thompson 2-moment MYJ + Eta 
WRF6_MYJ WRF6 MYJ + Eta 
Morrison_YSU Morrison 2-moment YSU + MM5 
Thompson_YSU Thompson 2-moment YSU + MM5 
WRF6_YSU WRF6 YSU + MM5 
 
Near-surface weather observations collected by Remote Automated Weather 
Stations (RAWS; https://raws.nifc.gov) provide the densest weather network in 
Alaska to date. Four weather variables recorded by RAWS, including air temperature, 
dew point temperature, relative humidity, and solar radiation were utilized as ?ground 
truth? for sensitivity analysis. To consider different fire regimes (Figure 1.2), I 
selected four cases for 24-hr simulations in years of varying fire season intensity 
(2006 ? low, 2007 ? moderate, 2010 and 2015 ? intense, Figure 1.2; Table 3.3). Daily 
observations were obtained for all available RAWS stations in the tundra. To identify 
the optimal combination group, I calculated three statistical metrics, including root-
mean-square error (RMSE), mean absolute error (MAE) and Pearson?s r correlation 
between simulations and observations. For each weather variable, I ranked the three 
metrics from all candidates for each year from 1 (lowest) to 6 (highest). For each 
metric, the yearly rankings were then summed up to create a single rank sum (ranging 
from 4 to 24), with the larger rank sum representing better overall performance. 
Table 3.3. Summary of case studies for sensitivity analysis 
Year Number of RAWS stations  Simulation period (UTC time) 
2006 8 08/16 00:00 - 08/17 00:00 
2007 9 05/01 00:00 - 05/02 00:00 
2010 12 07/01 00:00 - 07/02 00:00 
2015 29 06/18 00:00 - 06/19 00:00 
66 
 
 
 
3.4.3 RF modeling of cloud-to-ground lightning probability 
RF classification algorithm (Breiman, 2002, 2001) was used for modeling CG 
lightning probability with the atmospheric variables simulated from WRF. As an 
ensemble method, RF generates a large number of individual decision trees through 
permutation and integrates its results for a more stable modeling performance. I 
designed two separate modeling experiments referred to as the ?24-hr model? and 
?48-hr model? in this study, to compare the consistency of modeling performance and 
variable importance with different WRF simulation periods (0-24hrs and 0-48hrs 
respectively; Figure 3.3). For each experiment, atmospheric variables of interest were 
extracted from WRF outputs simulated at 24hrs or 48hrs after initialization, 
respectively. Four groups of predictors were summarized from existing literature 
(Blouin et al., 2016; Burrows, 2008; Burrows et al., 2005; Reap, 1991; Sousa et al., 
2013) including atmospheric stability indices, cloud properties, weather conditions at 
multiple pressure levels (500, 700, 850, and 1000 hPa), and two lightning 
parameterizations from WRF (Table 3.4). Both the Price and Rind (PR92; Wong et 
al., 2013) and Lightning Potential Index (LPI; Yair et al., 2010) lightning schemes 
were included for modeling. 
 
Figure 3.3. Illustration of the ?24-hr model? and ?48-hr model? components 
67 
 
 
 
Table 3.4. List of independent variables retrieved from WRF output for RF modeling 
Categories Variables Description 
CAPE Convective Available Potential Energy 
CIN Convective Inhibition 
LCL Lifted Condensation Level 
Atmospheric LFC Level of Free Convection 
stability indices TT Total Totals 
KI K Index 
PLI Parcel Lifted Index (to 500 hPa) 
BI Boyden Index 
SHOW Showalter Index 
CFtotal Total cloud cover fraction 
CFhigh High-level cloud cover fraction 
CFmid Mid-level cloud cover fraction 
CF low Low-level cloud cover fraction 
CTT Cloud top temperature 
CTH Cloud top height 
CTP Cloud top pressure 
Cloud properties IWP Ice water path 
LWP Liquid water path 
ERIce Effective radius of cloud ice 
ERWater Effective radius of cloud water 
QCloud Cloud water mixing ratio 
QIce Ice mixing ratio 
QRain Rain mixing ratio 
BT Brightness temperature 
T Air temperature at surface & multiple pressure levels 
T  Dewpoint temperature at surface and multiple d pressure levels 
T.T  Temperature-dewpoint spread at multiple pressure d levels 
RH Relative humidity at surface and multiple pressure 
Weather levels 
variables GPZ Geopotential height at multiple pressure levels 
DZ Thickness between any two pressure layers 
W Vertical velocity at multiple pressure levels 
Helicity Helicity 
UH Updraft helicity 
Rain Total precipitation 
PW Precipitable water 
SLP Sea level pressure 
68 
 
 
 
Lightning PR92 Flash distribution of CG lightning with PR92 
parameterizations LPI Lightning probability index 
 
All predictors were retrieved with WRF outputs of Domain 2 using the NCAR 
Command Language (NCL) and the NCEP Unified Post Processor System (UPP). 
Then I used these atmospheric factors to model the presence and absence of CG 
lightning strikes during the following day (24hrs) after the timing of variable 
extraction, which were considered as dependent variables in my RF models. CG 
lightning points extracted from the ALDN dataset during the corresponding period 
were labeled as the presence of lightning, while sample points with no lightning 
occurred (labeled as absence) were randomly generated within Alaskan tundra region 
during the same period.  
To ensure the representative of our models in describing tundra lightning 
conditions, three lightning severity levels were identified based on the total number of 
daily CG lightning strikes for model training, following a similar method by Farukh 
et al. (2011): > 2000 strikes per day as severe, 500 ? 2000 strikes per day as 
moderate, and 0 ? 500 strikes per day as low. For each level, I selected five cases for 
WRF simulation using a stratified random sampling strategy (Table 5). I then 
randomly selected 70% of presence and absence points for model training and 
reserved the rest 30% for validation. Training and testing points from all levels were 
combined for model development and accuracy assessment, respectively. I set the 
number of trees to 500 and the number of variables at each split as 8 in the RF 
algorithm. 
 
69 
 
 
 
Table 3.5. Cases studies of WRF simulations for CG lightning modeling 
Initialization of  Total CG lightning strokes  Total CG lightning strokes 
simulation (UTC) during 0~24hrs during 24~48hrs  
2003/06/23 00:00 1305 1560 
2005/06/11 00:00 732 1027 
2005/06/29 00:00 3449 2151 
2005/08/16 00:00  5 15 
2007/07/04 00:00 4280 479 
2007/07/11 00:00 2321 2247 
2008/06/25 00:00 2030 295 
2009/06/09 00:00 335 64 
2010/07/01 00:00 1015 1991 
2013/06/20 00:00 3073 3641 
2013/08/16 00:00 60 10 
2015/07/14 00:00 847 4320 
2015/06/21 00:00 974 905 
2015/07/23 00:00 175 80 
2016/07/11 00:00 93 109 
 
I reported the out-of-bag (OOB) error rate provided by the RF classification 
algorithm to evaluate the overall accuracy of the trained models. I further calculated 
the commonly used statistical criteria using the prediction results of the validation 
dataset for accuracy assessment (Tables 3.6 and 3.7). Receiver Operating 
Characteristics (ROC) curves and area under curve (AUC) were also generated based 
on the validation dataset. Additionally, I examined the contribution of predictors in 
determining CG lightning potential in the tundra with variable importance quantified 
by Mean Decrease in Accuracy (MDA;  Breiman, 2002, 2001). 
 
Table 3.6. Contingency matrix of variables used to calculate statistical scores 
 CG lightning event observed Presence Absence 
CG lightning Presence ? (hit) ? (false alarm) 
event predicted Absence ? (miss) ? (correct non-event) 
70 
 
 
 
Table 3.7. Statistical criteria used for assessing modeling performance 
Statistical scores Abbreviation Formula 
Probability of Detection POD ?/(? + ?) 
Critical Success Index CSI ?/(? + ? + ?) 
False Alarm Ratio FAR ?/(? + ?) 
False Alarm Rate F ?/(? + ?) 
 
3.4.4 Forecasting capability assessment 
In addition to empirical modeling, I further tested the capability of our 
empirical-dynamic method in forward forecasting of CG lightning probability at a 
future timing, by applying the RF model developed for a previous period. Here I 
utilized the ?24-hr model? to forecast the CG lightning probability with atmospheric 
conditions simulated 48hrs after initialization. Statistical criteria listed in Table 7 and 
ROC curves were also generated for quantifying the forecasting capability of CG 
lightning potential.  
3.5 Results 
3.5.1 Sensitivity analysis of WRF simulation 
The results of the sensitivity analysis indicate a substantial variability in 
performance of different WRF parameterizations by individual meteorological 
variable and across different case study periods (Figure 3.4; Tables S1 ? S4). The 
combined statistical ranking for each variable can range between 4 (the lowest rank 
across all four sensitivity cases, see Table 3.3) and 24 (the highest rank across all four 
cases). Based on these results, Thompson_MYJ and Morrison_MYJ emerged as 
strongest performing settings for tundra meteorology simulations (Figure 3.4). 
Although majority of the correlation values for all metrological variables are around 
71 
 
 
 
0.7 ~ 0.8, the overall simulation results for air temperature and dew point temperature 
across all six candidates outperform those for relative humidity and solar radiation 
according to the RMSE and MAE values (Tables S1 ? S4).  
 
Figure 3.4. Bar plots summarizing the ranks of six candidate schemes based on 
Pearson?s r correlation, MAE and RMSE for: (a) air temperature, (b) dew point 
temperature, (c) relative humidity, and (d) solar radiation. The largest rank sum 
value represents the best overall performance. 
Specifically, Thompson_MYJ shows the best results when compared to other 
schemes for air temperature and dew point temperature (Figure 3.4 a-b; Tables S1 ? 
72 
 
 
 
S2). For RH, using MYJ for the PBL scheme show superior results than adopting 
YSU, particularly when combined with the Morrison and Thompson schemes (Figure 
3.4 c). While for solar radiation, the Morrison scheme outperforms other 
microphysics schemes, followed by Thompson. Since lightning activity is largely 
related to convection activity, I further examined the spatial distribution of 
Convective Available Potential Energy (CAPE), as a representative of convection, 
generated by both Thompson_MYJ and Morrison_MYJ schemes for comparison. The 
Thompson_MYJ scheme shows a more detailed distribution of CAPE values for 
describing convective activities (Figure 3.5). I therefore chose to adopt the 
Thompson_MYJ combination and summarized the optimized WRF schemes for CG 
lighting modeling experiments in Table 3.8. 
 
 
Figure 3.5. Comparison of CAPE simulation in Domain 2 between (a) Morrison_MYJ 
and (b) Thompson_MYJ for the 2010 case as an example. 
73 
 
 
 
Table 3.8. Optimized combination of physical parameterizations 
Physical component Parameterization scheme Setting option in WRF 
Microphysics Thompson 2-moment mp_physics = 8 
Cumulus Grell-Freitas ensemble cu_physics = 3 
PBL MYJ bl_pbl_physics = 2 
SL Eta sf_sfclay_physics = 2 
Land surface model Noah sf_surface_physics = 2 
Longwave radiation RRTMG ra_lw_physics = 4 
Shortwave radiation RRTMG ra_sw_physics = 4 
 
3.5.2 Accuracy assessment of RF models 
The overall OOB estimate of error rate and the class errors have shown high 
accuracy in predicting CG lightning strikes in Alaskan tundra. For the ?24-hr model?, 
the OOB is 4.64% and the overall accuracy reaches 95.36%. Specifically, the absence 
of CG lightning has a class error of 7.26%, while that of the presence reaches 2.54% 
(Table 3.9 a). For the ?48-hr model?, the OOB is 6.81% with an overall accuracy of 
about 93.19%. The class errors of the absence and presence of lightning events are 
11.15% and 3.43%, respectively (Table 3.9 b).  
Table 3.9. Confusion matrix for (a) ?24-hr model? and (b) ?48-hr model? 
(a) 24-hr model 
Confusion matrix CG lightning predictions Absence Presence Class error 
CG lightning  Absence 7013 549 0.0726 
observations Presence 241 9239 0.0254 
(b) 48-hr model 
Confusion matrix CG lightning predictions Absence Present Class error 
CG lightning Absence 4872 612 0.1115 
observations Presence 243 6837 0.0343 
 
The validation results against the reserved dataset show that both models have 
overall high accuracy in predicting CG lightning strikes. I assessed the overall 
74 
 
 
 
statistical metrics (Table 3.10) and the ROC curves (Figure 3.6). I also reported them 
separately by severity level and found a notably stronger performance during severe 
and moderate lightning days when compared to the low severity level. For the severe 
and moderate levels, the ?24-hr model? produces a Probability of Detection (POD) 
value of 0.96 and a Critical Success Index (CSI) of 0.91, while the two error metrics 
False Alarm Ratio (FAR) and False Alarm Rate (F) are below 0.1 (see Table 3.7). 
While the POD and CSI values are below 0.8 for the validation data from the low-
level lightning cases. The AUC values estimated from the ROC curves are above 0.95 
across lightning days in all severity levels.  
Table 3.10. Statistical criteria calculated using the validation data for: (a) 24-hr 
model; (b) 48-hr model. 
Models Metrics All Severe level Moderate level Low level 
POD 0.9628 0.9821 0.9456 0.7885 
CSI 0.9065 0.9365 0.8569 0.7454 
24-hr 
model FAR 0.0606 0.0472 0.0987 0.0682 
F 0.0858 0.1167 0.0794 0.0277 
AUC 0.9869 0.9898 0.9818 0.9743 
POD 0.9668 0.9701 0.9643 0.2273 
48-hr CSI 0.8815 0.8808 0.8833 0.1829 
model FAR 0.0909 0.0946 0.0868 0.5161 F 0.1247 0.1291 0.1196 0.0366 
AUC 0.9810 0.9810 0.9809 0.8751 
 
The ?48-hr model? shows an overall POD value of 0.97 and a CSI of 0.88, 
while the FAR is below 0.1 and the F metric is about 0.12. Similar to the ?24-hr 
model?, it also has the best performance for the validation records collected during 
the severe lightning days, followed by those collected during moderate lightning 
days. However, the POD and CSI values are below 0.5 for the low severity cases. 
75 
 
 
 
Similar patterns can be observed from the ROC curves (Figure 3.6 b). The AUC 
reported from validation data collected during severe and moderate lightning days are 
higher than 0.95, while its value drops to 0.87 for the low severity level days (Table 
3.10).  
 
Figure 3.6. ROC curves of validation results on lightning days with different severity 
levels: (a) 24-hr model, (b) 48-hr model. 
In addition to the purely statistical accuracy assessment, I generated CG 
lightning probability maps across the entire Alaskan tundra to visually compare the 
predicted spatial patterns to observed patterns of lightning strike distribution (see an 
example for a severe lightning day in Figure 3.7. 
 
76 
 
 
 
 
Figure 3.7. (a) Observed CG lightning strikes during 2010/07/02 00:00 and 
2010/07/03 00:00UTC and (b) modeled CG lightning probability with the ?24-hr 
model? in Alaskan tundra as an example. 
3.5.3 Forecasting performance 
The forecasting test using the ?24-hr model? for the 48-hr simulation 
demonstrates strong statistical performance of my method in forecasting forward, 
with POD as 0.71 and CSI as 0.67 (Table 3.11). The two false ratios FAR and F are 
below 0.08. When separated by different severity levels of lightning days, the 
forecasting performance is consistent with previously reported results: the accuracy 
appears to be substantially higher for severe and moderate lightning conditions 
compared to low severity days. Similar patterns can be found from the ROC curves 
(Figure 3.8), with the AUC values around 0.9 for the entire data and those from the 
moderate level cases (Table 3.11). 
 
77 
 
 
 
Table 3.11.  Statistical criteria calculated for the forecasting test 
Metrics All Severe level Moderate level Low level 
POD 0.7063 0.6372 0.7707 0.1515 
CSI 0.6660 0.5898 0.7397 0.1397 
FAR 0.0788 0.1119 0.0516 0.2 
F 0.0780 0.1024 0.0547 0.0132 
AUC 0.9203 0.8766 0.9519 0.769 
 
 
Figure 3.8. ROC curves of forecasting tests on lightning days with different severity 
levels 
3.5.4 Evaluation of variable importance 
To understand the roles of the predictors in determining the CG lightning 
potential in the tundra, I examined the top 20 important variables ranked according to 
MDA from both the ?24-hr model? and the ?48-hr model? for comparison (Figure 
3.9). According to both models, Parcel Lifted Index (PLI) is found to be the most 
important variable in determining the accumulated lightning strikes among all the 
predictors. For the ?24-hr model?, weather variables at multiple pressure levels, 
including geopotential height, dew temperature, relative humidity, and velocity speed, 
78 
 
 
 
also show their importance in determining CG lightning occurrence in the tundra 
(Figure 3.9 a). In addition, cloud fraction, sea level pressure, helicity, lifted 
condensation level, and atmospheric stability indices like the Total Totals and the 
Showalter Index (SHOW) are also among the top 20 important indicators. Although 
the variable ranking order of the ?48-hr model? differs from that of the ?24-hr 
model?, the majority of the top 20 important variable remains the same. Only layer 
thickness between 700 and 850 hPa levels (DZ700-850) and brightness temperature 
appear to play an important role in determining the lightning potential for the ?48-hr 
model? but not in the ?24-hr model?. 
 
Figure 3.9. Top 20 important variables ranked by MDA in (a) ?24-hr model? and (b) 
?48-hr model?. 
3.6 Discussions 
In general, my results successfully demonstrate the strong capability of the 
empirical-dynamic method in representing meteorological conditions that support CG 
79 
 
 
 
lightning well in Alaskan tundra. As integration of WRF (using the selected 
parameterization schemes) and RF algorithm, this method shows excellent 
performance in both modeling CG lightning strikes with 24hr and 48hr WRF 
simulations, and forward forecasting of lightning probability. This supports the 
effectiveness of the empirical-dynamical framework in accurate prediction of future 
CG lightning potential in the data-scarce regions like the HNL. Considering the 
primary role of lightning in igniting wildland fire in the boreal forest and tundra 
ecosystems (French et al., 2015; Veraverbeke et al., 2017), this study can further 
support the monitoring of fire weather conditions and probability as well.  
The results indicate that atmospheric conditions supporting CG lightning 
activity in Alaskan tundra may differ from those in the boreal. Although instability 
indices SHOW and CAPE were recognized as the top-ranking indicators for lightning 
potential in boreal forests (Blouin et al., 2016; Burrows et al., 2005), I found PLI as 
the most important factor in modeling lightning activity than any other indices in 
Alaskan tundra. This highlights the key role of the lift potential in providing sufficient 
convection to support lightning formation in the tundra than in the boreal forests in 
general, which is consistent with the modeling results of Burrows et al. (2005) in the 
far west and north region of the North America dominated by tundra. While Burrows 
et al. (2005) suggested that the occurrence of lightning was influenced by the 
interaction between strong convection and precipitable water (PW) in the cloud, here 
I found much lower rankings of PW than instability indices like PLI and SHOW in 
both models. This indicates that convection plays a more critical role than PW in 
determining the lightning potential in the tundra, which is consistent with the findings 
80 
 
 
 
of Reap (1991). Besides, I also found the vertical profiles of variables like 
geopotential height, relative humidity, dew point temperature, and layer thickness 
important in modeling lightning activity (Figure 9). However, neither of the two 
lighting parameterizations (PR92 and LPI) reflect lightning potential in the tundra at 
the regional scale well. This is not surprising because these parameterizations were 
not specifically developed for the HNL.  
Additionally, the justification of using Thompson 2-moment and MYJ for 
WRF simulation in the tundra is consistent with existing findings of WRF application 
in lightning modeling. With a more detailed distribution of CAPE values for 
describing convective activities (Figure 4), the Thompson scheme is also 
recommended in existing studies (Giannaros et al., 2015; Zepka et al., 2014), given a 
detailed representation of ice-phase processes and improved simulation performance 
for convection related events like precipitation. For lightning modeling purpose, the 
MYJ scheme is also suitable for describing PBL conditions considering its optimal 
representation of atmospheric conditions for triggering convection activities 
(Giannaros et al., 2015; Sousa et al., 2013).  
Despite WRF?s strong capabilities in regional modeling and dynamic 
downscaling of atmospheric conditions, its performance of describing physical 
processes is limited by the assumptions and mechanisms employed for developing the 
parameterization schemes. Polar WRF is now under development for an improved 
description of near-surface and atmospheric conditions for the Arctic regions with 
improved WRF schemes, which has the potential for improving lightning modeling 
for Alaska in future (Cai et al., 2018; Hines et al., 2011; Wilson et al., 2012). 
81 
 
 
 
Although our modeling results show very impressive prediction capabilities, 
they are subject to uncertainties inherent in CG lightning observations detected from 
either the ground-based networks. Satellite-based sensors such as Lightning Imaging 
Sensor or Geostationary Lightning Mapper monitor lightning at large spatial scale. 
Their datasets, however, are not available for the HNL due to the limited spatial 
coverage and resolution (Matsangouras et al., 2016; Nag et al., 2014). Separating the 
CG and intra-cloud lightning from satellite observations can introduce errors as well 
(Nag et al., 2014). Although ground-based systems are constrained by position 
accuracy and detection efficiency, I found reporting probability a more effective way 
to describe CG lightning activity.  
Although atmospheric factors function as the key predictors for lightning, 
synoptic-scale dynamic forcings control meteorological mechanisms driving lightning 
and fire weather (Flannigan and Wotton, 2001; Reap, 1994; Santos et al., 2013). For 
example, Kochtubajda et al. (2019) found more frequent ridging and ridge 
displacements during the 2014 wildfire season in the Northwest Territories of 
Canada. Synoptic weather conditions should be explored and incorporated in future 
modeling efforts to improve our understanding of lightning and fire regimes in the 
tundra. 
3.7 Conclusions 
Wildfire dominates the disturbance regimes in the Arctic and boreal 
ecosystems in North America. In recent decades, these regimes have been notably 
changing and, thus, necessitating a better understanding of the current regimes and 
their feedbacks to the climate system (Fisher et al., 2018). This study makes the first 
82 
 
 
 
effort in examining the factors driving lightning activity, the primary ignition source 
for wildfire, in Alaska tundra with my statistical-dynamical modeling method. This 
study demonstrated the effectiveness of integrating WRF and machine learning for 
lightning modeling in Alaskan tundra at 5km spatial resolution. The results provide 
insights on understanding the mechanisms of lightning-ignited fires in the tundra. PLI 
and weather variables at multiple pressure levels were found to be the most important 
predictors for modeling lightning potential in the tundra, indicating the primary role 
of convection in the formation of thunderstorms and CG lightning. Moreover, 
applicable for other data-scare regions, this method can further support the lightning 
and fire prediction as well as management efforts in the HNL. 
83 
 
 
 
Chapter 4: Exploring environmental factors driving wildland 
fire ignitions in Alaskan tundra 
 
4.1 Summary 
Tundra fires are common across the pan-Arctic region, particularly in Alaska. 
Fires lead to significant impacts on terrestrial carbon balance and ecosystem 
functioning in the tundra. They can even affect the forage availability of herbivorous 
wildlife and living resources of local human communities. Also, interactions between 
fire and climate change can enhance the fire impacts on the Arctic ecosystems. 
However, the drivers and mechanisms of wildland fire ignitions in Alaskan tundra are 
still poorly understood. Research on modeling contemporary fire probability in the 
tundra is also lacking. This study focuses on exploring the critical environmental 
factors controlling wildfire ignitions in Alaskan tundra and modeled the fire ignition 
probability, accounting for ignition source, fuel types, fire weather conditions, and 
topography. The fractional cover maps of fuel type components developed Chapter 2 
serve as input data for fuel type distribution. The probability of cloud-to-ground (CG) 
lightning and fire weather conditions are simulated using WRF. Topographic features 
are also calculated from the Digital Elevation Model (DEM) data. Additionally, fire 
ignition locations are extracted from Moderate Resolution Imaging Spectroradiometer 
(MODIS) active fire product for Alaskan tundra from 2001 to 2019. Empirical 
modeling methods, including RF and logistic regression, are then utilized to model 
the relationships between environmental factors and wildfire ignitions in the tundra 
and to evaluate the roles of these factors. The results suggest that CG lightning is the 
primary driver controlling fire ignitions in the tundra, while warmer and drier weather 
84 
 
 
 
conditions also support fires. The results of this study highlight the important role of 
CG lightning in driving tundra fires and that incorporating CG lightning modeling is 
necessary and essential for fire monitoring and management efforts in the High 
Northern Latitudes (HNL).  
4.2 Introduction 
Wildland fires have played an important role in altering ecosystem 
functioning, driving land cover change and affecting carbon balance in the boreal 
forest and tundra ecosystems (Gustine et al., 2014; Kasischke et al., 2010; Mack et 
al., 2011; Randerson et al., 2006; Rocha and Shaver, 2011; Turetsky et al., 2011). 
Although tundra fires are typically less severe than the fires in the boreal forests, they 
are common and widespread across the pan-Arctic region according to satellite based 
observations (He et al., 2019). In recent years, several large fire seasons have 
occurred in Alaskan tundra, such as the 2010 fire season in the Noatak River Valley 
and the 2015 fire season in the Southwest Alaska. Tundra fires also have the potential 
to release the ancient carbon stored in the permafrost underneath the land surface and 
lead to widespread development of thermokarst (Jones et al., 2015; Mack et al., 
2011). Moreover, fires in the tundra can influence the habitats and forage availability 
of wildlife like caribou, which would further affect the resources of local human 
communities (Gustine et al., 2014; Jandt et al., 2008; Rupp et al., 2006). Furthermore, 
fire activities are likely to increase in the tundra under climate warming in the future 
(French et al., 2015).  
Despite the importance of tundra fires, they are less studied compared to fires 
in other ecosystems. Current research primarily focuses on evaluating post-fire 
85 
 
 
 
impacts such as fire severity (Loboda et al., 2013), ecosystem responses (Bret-Harte 
et al., 2013), and carbon budget change (Mack et al., 2011) within comparatively little 
attention to modeling tundra fire occurrence. Previous studies have modeled historical 
or future tundra fire regimes with either empirical methods or an ecosystem model 
Alaska Frame-Based Ecosystem Code (ALFRESCO; Higuera et al., 2011; Joly et al., 
2012; Young et al., 2017).  However, research on modeling contemporary fire 
occurrence is lacking in current English language peer-reviewed literature. Although 
numerous studies have modeled the interactions between environmental factors and 
wildfire occurrences in boreal forests, their results are not directly applicable to 
tundra ecosystems due to the different ecosystem functioning and responses (French 
et al., 2015; Van Heerwaarden and Teuling, 2014).  
Wildfire occurrences are primarily ignited by cloud-to-ground lightning 
strikes and human activities. In addition to the ignition sources, wildfire behaviors are 
typically controlled by three types of influencing forces including fuel, weather and 
topography, as summarized by the Fire Environment Triangle (Pyne et al., 1996). 
Fuel type and state are critically important factors that control fire-environment 
interactions through altering fire characteristics and affecting ignition easiness (Pyne 
et al., 1996). Fuel type represents properties of the fuel itself, such as fuel 
composition, continuity and loading. Fuel state is mainly related to moisture content 
primarily driven by the changing weather conditions at different temporal scales. 
Topography also has the potential to affect the fire behaviors through controlling 
exposure to sunlight and moisture pooling. Weather conditions can also function as 
the dominant contributors to the fire environment from hourly to daily scales, by 
86 
 
 
 
influencing the fuel conditions and ignition sources. A variety of rating systems have 
been developed to quantify the danger of fire weather conditions and to predict the 
fire potential, such as National Fire Danger Rating System (NFDRS) and Canadian 
Forest Fire Weather Index System (CFFWIS).  
Majority of the existing studies exploring the impacts of these environmental 
factors in driving wildland fires are focused on boreal forests in the HNL, from the 
perspectives of both fire occurrence and fire spread. Liu et al. (2012) studied both 
lightning and human ignited fires in the boreal forests of Northeast China and found 
out that lightning fires are mainly controlled by fuel moisture and vegetation type. 
Veraverbeke et al. (2017), however, suggested the primary role of lightning in driving 
burned areas in recent large fire years in the boreal forests of North American. 
Peterson et al. (2010) also demonstrated the effects of atmospheric stability, lightning 
strike counts, and dry weather conditions on fire activity in the boreal forests of North 
America. Though lightning characteristics like polarity and peak current were found 
to be significant predictors of fire occurrences (M?ller and Vacik, 2017; Vec?n-Arias 
et al., 2016), they did not function as major contributors that drive fire ignitions in 
some other studies (Ad?mek et al., 2018; Pineda et al., 2014). These differences could 
be caused by different matching methods between lightning strikes and fires (Moris et 
al., 2020). Recently, Masrur et al. (2018) found out that warm and dry conditions 
affect the spatiotemporal patterns across the entire circumpolar tundra regions. 
However, their analyses were conducted across the pan-Arctic region and may not 
explain the situation of fires in Alaskan tundra. 
87 
 
 
 
Due to the lack of weather stations in the remote tundra, datasets developed 
based on station observations with interpolation methods such as Daymet are not 
suitable for capturing these environmental factors, particularly fire weather 
conditions, in this data-scare region. Reanalysis data products also have their 
limitations regarding the coarse spatial resolution. Meteorological variables simulated 
with Numerical Weather Prediction (NWP) models like Weather Research and 
Forecast (WRF) model have demonstrated their suitability for describing fire weather 
conditions and capturing their spatial variations in regions with limited observations 
available (M?lders, 2010, 2008), although the simulated indices may not always be 
trustworthy in accurate characterization of fire risk likely caused by the inaccurate 
simulation of wind and precipitation (M?lders, 2010; Simpson et al., 2014). In 
addition, remote sensing observations have also been widely adopted for capturing 
the spatial distribution of fuel types, fuel moisture states and topographical features 
(He et al., 2019; Loboda, 2009; Loboda and Csiszar, 2007; Ottmar et al., 2007; Yebra 
et al., 2013, 2008).  
Two types of models have been developed to uncover the impacts of the 
environmental factors on driving lightning-ignited wildfire occurrences and to model 
their relationships. Earliest attempts of fire occurrence modeling developed physical-
based models with explicit representations of wildland fire processes(Anderson, 
2002; Anderson et al., 2000; Kourtz and Todd, 1991). Specifically, major steps 
related to fire occurrences including the start of lightning, fire ignition, fire 
smoldering and fire detection were accounted for and represented in these physical-
based models. Through a series of equations developed upon experiments and 
88 
 
 
 
assumptions, these models were able to incorporate the related environmental drivers 
of fires and to model the probability of fire occurrences.  
Empirical based methods have been commonly adopted for modeling wildfire 
ignitions in recent-year research (Prestemon et al., 2013). The spatial point modeling 
methods have been adopted by several studies on modeling fires in different 
ecosystems (Liu et al., 2012; Woo et al., 2017; Yang et al., 2015). Generalized Linear 
Models have also been commonly utilized to predict fire occurrences (Ager et al., 
2018; Barbero et al., 2014; Vilar et al., 2016). In addition, with strong predictive 
capability, logistic regression and random forest (RF) based algorithms have been 
frequently applied to understand the relationships between environmental factors and 
fire occurrences (Guo et al., 2016; Keyser and Leroy Westerling, 2017; Van 
Beusekom et al., 2018; Vec?n-Arias et al., 2016; Viedma et al., 2018; Wotton and 
Martell, 2005). 
Therefore, in this study, I aim to understand ?What environmental factors 
drive fire ignition probability across Alaskan tundra?? Here fire ignition is referred 
as the detectable start of fire occurrence using satellite sensors. To address this 
research question, I plan to develop an empirical model using RF algorithm for 
predicting fire ignition probability in Alaskan tundra with environmental drivers 
derived from WRF simulations and satellite observations from 2001 to 2019.  
4.3 Study Area 
This study covers the entire Alaskan tundra beyond the northern tree line, as 
defined by the Circumpolar Arctic Vegetation Map (CAVM) data (Walker et al., 
2005; Figure 3.1 a), focusing on the shrub or graminoid dominated tundra in 
89 
 
 
 
particular. Major vegetation fuels in this region include evergreen or deciduous shrub, 
sedge, grass, moss, and lichen (Bliss et al., 1973). This region is underlain by 
continuous and discontinuous permafrost. Though varying by ecoregions, the study 
area has an average temperature below 0?C throughout the year. Despite the low 
annual precipitation, tundra lands tend to be wet because of the low evaporation rates 
and poor drainage conditions. Wildland fire activities in Alaskan tundra also have a 
large variability across different ecoregions (French et al., 2015; Rocha et al., 2012). 
Over the past half-century, the mean fire return intervals for Alaskan tundra 
ecoregions vary from 13 to 22 years, and the fire rotation periods have a range from 
roughly 400 to thousands of years (Rocha et al., 2012). Though most Alaskan tundra 
fires are small in spatial scale, the large ones are not rare. Particularly, Kotzebue 
Lowlands and Seward Peninsula tend to have higher frequency and larger burned 
extent compared to the others (French et al., 2015). 
4.4 Data and Methods 
To achieve the research goals, I considered five groups of environmental 
factors related to wildfire ignition, including fuel type distribution, fuel state 
conditions, fire weather conditions, topographical features, and fire ignition source 
(Figure 4.1). Fire ignition locations in Alaskan tundra extracted from satellite 
observations were used as the dependent variable. I then developed an empirical 
model for predicting the fire ignition probability in the tundra using RF algorithm. In 
particular, the fractional cover maps developed in Chapter 2 were used to represent 
fuel type distribution in this study. The empirical-dynamical method developed in 
Chapter 3 was adopted to model the probability of CG lightning strikes, the primary 
90 
 
 
 
ignition source of wildfire in Alaskan tundra. Additional environmental variables, 
including satellite-derived fuel state characterization, surface topography and weather 
conditions, are described in this chapter.   
 
Figure 4.1. Methodology framework of modeling fire ignition probability in Alaskan 
tundra in Chapter 4. 
In addition to the overall framework, the modeling workflow of this chapter is 
further summarized in Figure 4.2. All input environmental factors were first 
categorized into variables of relatively high and low temporal variability referred to 
as dynamic and static, respectively. Static variables including both topographical 
features (ref?) and fractional coverages of major fuel type components (He et al., 
2020). Although vegetation shifts and fuel type transitions can occur from years to 
decades under disturbances or climatic variability and change, the vegetation 
compositions are relatively stable when compared to the seasonal and diurnal changes 
of meteorological variables and fire weather conditions. The fuel types were therefore 
considered as static variables in this dissertation. Weather related variables, including 
91 
 
 
 
near surface, CFFWIS, and CG lightning probability, are dynamic throughout a short 
time period. Compared to fuel type, fuel moisture state can change under the 
influence of meteorological conditions in a few days, thus considered as dynamic 
variables.  
To characterize the dynamic components of fire ignition, I extracted the earliest 
detections from the space-time contiguous point clouds of MODIS active fire data 
(sec section 4.4.1) to determine ignition locations and dates for multiple years. These 
points served as ?Fire? events in the fire ignition modeling. The dates of these 
ignition points were used for determining dynamic variables used for modeling. I 
used the leaf moisture related vegetation indices calculated from the closest day of 
year (DOY) before the ignition dates using MODIS 8-day surface reflectance 
composites (MOD09A1; Vermote et al., 2015). I then determined WRF simulation 
dates for modeling CG lightning probability and near-surface meteorological 
conditions. Since the calculation of some of the fire weather indices within the 
CFFWIS is based on long-term accumulation of weather variable starting from 
snowmelt dates, here I combined downscaled WRF output starting from a week since 
the ignition dates, and the relative coarse resolution reanalysis data from the mid-
April to calculate indices from CFFWIS. After preparing all independent variables, 
two models were developed with RF algorithms using dynamic variable acquired on 
the fire ignition day and one day before ignition, to, first, assess the ability of the 
model to represent actual fire ignition instances (the day of ignition model) and, 
second, assess the model?s forecasting capacity (one day before ignition).  
92 
 
 
 
 
Figure 4.2. Modeling workflow 
4.4.1 Determining fire ignition locations 
Ignition point locations of tundra fires from 2001 to 2019 were determined 
using the 1km MODIS Thermal Anomalies/Fire locations data product (MCD14DL 
V006; Giglio et al., 2003) obtained from NASA?s Fire Information for Resource 
Management System (FIRMS). Since fires have strong emission of mid-infrared 
radiation, this active fire product detects fires based on brightness temperatures from 
MODIS 4- and 11-?m bands using a contextual algorithm (Giglio et al., 2003). In 
addition to coordinates (latitude and longitude) of active fire points, attributes 
including brightness temperature, acquisition date, acquisition time, fire radiative 
power, and detection confidence are also recorded in the data product. Only fire 
points classified as nominal or high confidence levels (detection confidence above 
30%) were considered in the processing. 
MODIS active fire points in the tundra were first extracted using a 10km 
buffered boundary of the CAVM data. I designed a clustering method based on the 
93 
 
 
 
Density-Based Spatial Clustering of Applications with Noise (DBSCAN; Ester et al., 
1996) algorithm to identify the tundra fire clusters (Figure 4.3). The DBSCAN 
algorithm separates points of high density from those of low density and locates these 
regions as clusters. Here the ? parameter in DBSCAN, which represents the 
maximum distance between two neighboring points, was set to 2.5 km based on the 
study of Loboda and Csiszar (2007). The DBSCAN algorithm was applied to tundra 
fire points for each year. Since fires occurring at different times could be grouped into 
the same cluster, I further examined the temporal gap (Gaptime) among fire points in 
each cluster and. The temporal gap was determined as 4 days to separate different 
fires in a certain spatial cluster (Loboda and Csiszar, 2007). The active fire points 
with the earliest acquisition time were considered as the ignition locations. 
 
Figure 4.3. Workflow for extracting fire ignition locations in Alaskan tundra  
94 
 
 
 
Individual fire ignition events were then identified from the MODIS active 
fire records between 2001 and 2019 across Alaskan tundra using the algorithm 
described above, suggesting a variety of wildfire activity patterns in this region 
(Figure 4.4). Thirteen of the nineteen years in the available data record have a number 
of fire events between twenty (2001 ? 2004, 2006, 2008 ? 2009 , 2011 ? 2014, 2016 ? 
2018), four years can be characterized as a moderate fire activity season with the 
number of fire events between 20 and 30 (2005, 2007, 2010, 2019), while the 2015 
fire season has particularly high fire activity with 49 fire events detected within the 
boundary of Alaskan tundra defined by CAVM.  
 
Figure 4.4. Number of fire ignition events detected from 2001 to 2019 with 
MCD14DL data. Sampled fire years are highlighted in orange boxes.  
To cover a variety of fire seasons, I sampled five seasons of low fire activity 
(2002, 2006, 2008, 2013, 2017), two years with moderate fire season activity (2007, 
2010), and one year with very high fire activity (2015), as illustrated in Figure 4.4. I 
then extracted the ignition points for individual fire events and defined the dates of 
ignition based on the time stamp of the ignition points to improve the computing 
95 
 
 
 
efficiency of WRF simulation. Fire ignition dates detected from the selected eight 
years were summarized in Table 4.1.  
Table 4.1. Full list of fire ignition dates detected from sampled years 
Year Fire ignition dates 
2002 5/23/2002, 6/3/2002, 6/18/2002, 7/10/2002, 7/20/2002, 7/28/2002, 8/2/2002, 8/5/2002, 8/15/2002, 8/22/2002, 9/22/2002 
2006 5/31/2006, 7/7/2006 
6/4/2007, 6/10/2007, 7/5/2007, 7/6/2007, 7/7/2007, 7/12/2007, 
2007 7/14/2007, 7/16/2007, 7/17/2007, 8/2/2007, 8/29/2007, 9/8/2007, 
9/9/2007, 9/19/2007, 9/30/2007 
2008 6/4/2018, 6/5/2018, 6/7/2018, 6/8/2018, 6/13/2018 
6/1/2010, 6/8/2010, 6/9/2010, 6/21/2010, 6/22/2010, 6/23/2010, 
2010 6/24/2010, 7/1/2010, 7/2/2010, 7/3/2010, 7/4/2010, 7/8/2010, 
7/10/2010 
2013 5/31/2013, 6/5/2013, 6/20/2013, 6/21/2013, 7/11/2013, 7/30/2013 
6/1/2015, 6/2/2015, 6/16/2015, 6/20/2015, 6/21/2015, 6/22/2015, 
2015 6/23/2015, 6/24/2015, 6/25/2015, 6/29/2015, 6/30/2015, 7/2/2015, 7/3/2015, 7/4/2015, 7/6/2015, 7/7/2015, 7/13/2015, 7/14/2015, 
7/20/2015, 7/24/2015, 7/25/2015, 8/1/2015, 8/6/2015 
2017 6/6/2017, 6/7/2017, 6/8/2017, 6/16/2017, 6/19/2017, 6/28/2017 
 
4.4.2 Modeling cloud-to-ground lightning probability with WRF simulations 
The empirical-dynamical framework ?24-hr model? developed in Chapter 3 
was adopted here to describe the probability distribution of ignition source across 
Alaskan tundra. Since the National Centers for Environmental Prediction (NCEP) 
final analysis data used in Chapter 3 does not provide precipitation data, in this 
section I used the North American Regional Reanalysis (NARR) data instead to 
initialize the two-way nested WRF simulation, for the consistency of CFFWIS 
calculation in Section 4.4.3. Developed by the NCEP, the NARR dataset assimilates a 
large amount of observations to generate long term estimations of weather conditions 
across the entire North America at 32 km resolution and 3-hour time interval. All 
96 
 
 
 
observations used in NCEP/NCAR Global Reanalysis project, precipitation data, 
radiances from Television Infrared Observations Satellite (TIROS) Operational 
Vertical Sounder (TOVS), profiler data, and land surface and moisture data have been 
utilized to produce this dataset.  
I then applied the modeling approach developed in Chapter 3 to develop CG 
probability grids at 5km resolution for the entire tundra region in Alaska, using WRF 
simulated atmospheric factors as independent variables. CG lightning probability was 
modeled and included in this analysis for each of the ?fire ignition? dates extracted in 
section 4.4.1. and ?the day before ignition? dates. The output 5km lightning 
probability grids were then used as input data for representing ignition sources of 
wildland fires.  
4.4.3 Near surface weather conditions and calculation of the CFFWIS 
As a sub-system of the Canadian Forest Fire Danger Rating System 
(CFFDRS), the CFFWIS has been developed since 1970 to account for the weather 
impacts on forest fuels and fires across the boreal forests of Canada (Van Wagner, 
1987; Figure 4.5). Although not specifically designed for the tundra fuels and fires, 
this system is suitable for describing fire weather conditions in the ecosystems of the 
HNL and considers for more detailed fuel types than the NFDRS (French et al., 2015; 
M?lders, 2010). 
97 
 
 
 
 
Figure 4.5. Flow chart of Canadian FWI system, source: National Wildfire 
Coordinating Group (https://www.nwcg.gov/publications/pms437/cffdrs/fire-weather-
index-system) 
The entire CFFWIS is comprised of six standard components, including three 
fuel moisture codes and three fire behavior indices (Van Wagner, 1987; Figure 4.5). 
The calculation of CFFWIS purely requires weather readings of air temperature, RH 
and wind at noon, as well as daily accumulated precipitation, which actually 
represents the peak fire danger during the midafternoon (Van Wagner, 1987). The 
design of three fuel moisture codes accounts for both water capacity and drying speed 
of the fuels. Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC) and 
Drought Code (DC) represent the moisture content and flammability of litter and fine 
fuels, loosely compacted organic matter of moderate depth, and deep and compact 
organic layers, respectively. Both DMC and DC are slow-reacting codes that are 
largely affected by the changing daylength and long-term accumulation, while the 
fast-reacting FFMC is less responsive those effects. For all three codes, larger values 
98 
 
 
 
suggest drier conditions and higher flammability of the fuels. Calculated based on the 
moisture codes, the fire behavior indices, including Initial Spread Index (ISI), 
Buildup Index (BUI) and Fire Weather Index (FWI), are designed to describe the 
important variables related to fire behavior, including spread rate, fuel weight and fire 
intensity.  
The CFFWIS requires long-term records of weather observations starting 
from snowmelt date for each year. Although the exact date of snowmelt varies across 
the regions, I set the startup date to a common day, as is common for calculating 
CFFWIS (Lawson and Armitage, 2008), in mid-April (DOY 101) for all years. Fire 
weather variables from NARR reanalysis data and WRF outputs were combined to 
support the calculation of CFFWIS to reduce the computing load caused by multi-
year WRF simulation, while achieving the improved spatial resolution of the results. 
Specifically, WRF simulations were run for a week before the ignition dates to 
downscale the data to 5 km. Then the NARR data were resampled with the Nearest 
Neighbor method to describe the meteorological conditions from the startup date till 
the WRF simulations. In addition to the four weather observations (air temperature, 
RH, wind and precipitation), the daylength data of the selected years for the entire 
North America were obtained from Daymet (Thornton et al., 2017) provided by the 
Oak Ridge National Laboratory Distributed Active Archive Center (ORNL DAAC).  
To prepare the input data for CFFWIS, the daily maximum values during 
daytime from noon to midafternoon were calculated based on the 3-hour intervals for 
air temperature, RH and wind. Daily accumulated precipitation values were also 
summed up based on the 3-hourly data. The initial conditions of seasonal start-up 
99 
 
 
 
values for FFMC, DMC, and DC were set to 85, 6, and 15, respectively (Lawson and 
Armitage, 2008). The three fuel moisture codes (FFMC, DMC, and DC) and three 
fire behavior indices (ISI, BUI, and FWI) from CFFWIS were then calculated 
according to the equations provided in Van Wagner and Pickett (1985) to describe 
long-term weather conditions for all the sampled eight years across the entire tundra.  
4.4.4 Estimating fuel state with vegetation indices 
A variety of vegetation indices have been developed to estimate vegetation moisture 
content for large-scale monitoring. Yebra et al. (2008) reviewed commonly used 
vegetation indices and suggested that Normalized Difference Infrared Index (NDII), 
Normalized Difference Water Index (NDWI) and Global Vegetation Moisture Index 
(GVMI) are directly related to leaf water content. I, therefore, decided to adopt these 
indices as a reference for fuel moisture state in this chapter (Table 4.2). MODIS 8-day 
surface reflectance data product MOD09A1 was used to compute these vegetation 
indices across Alaskan tundra. Only clear pixels were adopted for the calculation 
according to the state Quality Assurance (QA) flags. Indices from the closest DOY of 
8-day composites were used for RF modeling.  
Table 4.2. Vegetation indices for estimating live fuel moisture content 
Vegetation index Formula with MODIS bands Reference 
Normalized Difference ?( ? ?0 
Infrared Index (NDII ) ? (Hardisky et al., 1983)  6 ( + ?0
Normalized Difference ?( ? ?1
Infrared Index (NDII ) ? + ?  (Hardisky et al., 1983) 7 ( 1
Normalized Difference ?( ? ?2
Water Index (NDWI) ? + ?  (Gao, 1996) ( 2
Global Vegetation (?( + 0.1) ? (?0 + 0.02)
Moisture Index (GVMI) (Ceccato et al., 2002) (?( + 0.1) + (?0 + 0.02)
 
 
100 
 
 
 
4.4.5 Extracting topographic features 
The 5m DEM product collected with airborne Interferometric Synthetic 
Aperture Radar (IfSAR) data was obtained for the entire Alaska from the Alaska 
Elevation Portal supported by the State of Alaska (https://elevation.alaska.gov). The 
IfSAR data has been collected by USGS a 5-m post spacing throughout Alaska with 
since 2010. This data product has a vertical accuracy of 3-m with 90% confidence 
level for regions with slope between 0 ? 10 degrees. It has provided a full coverage 
for the tundra regions in Alaska. Four topographical features were extracted based on 
the DEM, including elevation, slope, aspect, and roughness using the Geospatial Data 
Abstraction Library (GDAL).  For a certain elevation grid cell, the roughness is 
defined as the maximum inter-cell difference between its elevation value and those 
from all surrounding cells (Wilson et al., 2007). Due to the extremely large size of the 
entire dataset covering Alaskan tundra, I resampled the DEM data to 30m for the 
extraction of topographical features considering the computing efficiency. 
4.4.6 Empirical modeling of fire ignition probability 
Empirical models were developed with RF classification algorithm to identify 
the important factors driving fire ignitions in the tundra and to predict the ignition 
probability. Fire ignition points extracted from the Section 4.4.1 were used to 
represent the presence of ?Fire? in Alaskan tundra, which was coded as ?1? in the 
modeling. I then randomly sampled similar amounts of points across the tundra 
regions on the same fire ignition dates as the ?No Fire? conditions, which was coded 
as ?0?. For dynamic variables, I extracted the values acquired on fire ignition dates 
and the dates before ignition separately, to construct two different datasets of 
101 
 
 
 
predictors for the two models, referred as ?Current model? and ?Previous model? in 
this dissertation. For both models, 70% of the records were randomly selected for 
model training, and the rest 30% were reserved for validation.  
Although spatial autocorrelation does not impact the accuracy of prediction 
within the RF models, it can impact the assessment of relative impact of individual 
input variables on the outcome.  Since the ultimate goal of this study is to understand 
the relative importance of the drivers of fire ignitions in Alaskan tundra, I first tested 
the Pearson?s r correlation between each pair of the variables and removed those that 
are highly correlated from modeling. Although the algorithm has strong capability of 
accurate prediction, it only provides the relative rankings of variable importance 
while fails to show how the independent variables relate to the fire events (negatively 
or positively) and the strength of the relationship. To account for these limitations of 
the RF modeling algorithm, I visualized the distribution of the values for different 
variables within ?fire? and ?non-fire? conditions using boxplots, and  quantified the 
significance of the observed differences in the mean values of environmental factors 
within ?fire? and ?non-fire? cases using Welch?s t-tests. Finally, logistic regressions 
were also developed with the two datasets to further support the exploration of 
driving factors of tundra fire ignitions.  
4.5 Results 
4.5.1 Variable selections based on correlations 
The study design includes the development and testing of two models.  The 
?Current? model represents the condition observed and modeled for the day of fire 
102 
 
 
 
ignition to test our ability to accurately model the ignition events.  The ?Previous? 
model is developed to add an assessment of fire ignition forecasting capabilities using 
weather and other environmental condition from the day before the actual date of fire 
ignition event. Before executing these empirical models, I first explored the Pearson?s 
r correlations between independent variables to exclude the highly correlated ones. 
The correlation matrices of variables in both ?Current? and ?Previous? models are 
shown in Figures 4.6 and 4.7, respectively. For both matrices, three groups of 
variables exhibit very strong correlations with Pearson?s r values above 0.8, including 
vegetation indices, fire weather variables, and topographic features. Specifically, all 
four vegetation indices are highly correlated with each other, with correlation values 
above 0.95. I therefore selected to include only NDII6 as a reference for the fuel state 
conditions in the modeling efforts. For the CFFWIS, I also found strong correlations 
between several pairs of CFFWIS fire weather indices, including FFMC-ISI and 
DMC-BUI, with correlation values above 0.95. FWI was not included in this analysis 
and further modeling steps because the dynamic range of the values remained at 0 for 
the dates. Therefore, only three fuel moisture codes ? FFMC, DMC and DC ? were 
included in the empirical modeling. Although near surface meteorological variables 
also show moderate correlations with fire weather indices, they are included to 
account for meteorological conditions irrespective of fuels. Finally, among 
topographic features, slope and roughness are closely related to elevation values for 
Alaskan tundra, and hence, are excluded from modeling. 
103 
 
 
 
 
Figure 4.6. Correlation matrix for static and dynamic variables extracted on the fire 
ignition days. ?LP?: CG lightning probability, ?Temp: air temperature, ?Elev?: 
elevation, ?Rough?: roughness, ?Herb?: herbaceous component, and ?Nonva?: 
nonvascular component. 
104 
 
 
 
 
Figure 4.7. Correlation matrix for static and dynamic variables extracted on the 
dates before fire ignitions. ?LP?: CG lightning probability, ?Temp: air temperature, 
?Elev?: elevation, ?Rough?: roughness, ?Herb?: herbaceous component, and 
?Nonvas?: nonvascular component. 
105 
 
 
 
4.5.2 Accuracy of empirical modeling 
The RF models have shown strong capability and high accuracy in predicting 
wildfire ignition probability in Alaskan tundra, according to the overall OOB estimate 
of error rate and the class errors. The overall OOB error rate of the ?Current? model 
is 6.03% with the overall accuracy of 93.97%. Specifically, the ?Fire? condition has a 
class error of 5.05%, while that of the ?No fire? reaches 7.17% (Table 4.3 a). The 
overall performance of the ?Previous? model is slightly lower. Its overall OOB error 
rate is 8.75% with the overall accuracy of 91.25%. The class errors of the absence and 
presence of fire events are 6.86% and 10.97%, respectively (Table 4.3 b).  
Table 4.3. RF modeling results of the two models for predicting fire ignitions 
(a)  Current model 
Confusion matrix Predictions No fire Fire Class error 
Observations No fire 263 14 0.0505 Fire 17 220 0.0717 
(b)  Previous model 
Confusion matrix Predictions No fire Fire Class error 
Observations No fire 258 19 0.0686 Fire 26 211 0.1097 
 
Validation performed against the reserved dataset shows that both models 
have a strong capability to reflect (for ?Current?) and forecast (for ?Previous?) fire 
ignition probability, according to the ROC curves (Figure 4.8). The AUC values 
estimated from the ROC curves are above 0.95 in both models, with 0.9704 and 
0.9683 for the ?Current? for ?Previous? models, respectively. 
106 
 
 
 
 
Figure 4.8. ROC curves estimated based on validation datasets for the two models: 
(a) ?Current model?; (b) ?Previous model?. 
4.5.3 Evaluation of variables driving tundra fire ignitions 
RF Mean Decrease in Accuracy (MDA) metric identifies CG lightning 
probability as the most important independent variable, with MDA values of 50.06% 
and 34.58% for the ?Current? and ?Previous? models, respectively (Figure 4.9). Near 
surface weather variables and fuel moisture codes from CFFWIS are also found to be 
important in determining fire ignition events in the tundra, particularly air 
temperature, RH and DC. Compared to the weather-related variables, fuel types, 
satellite-based estimates of fuel state, and topographical features contribute but do not 
appear to drive fire ignitions in the tundra.  
107 
 
 
 
 
Figure 4.9. Variable importance rank for models: (a) ?Fire ignition day?, (b) ?One 
day before ignition? 
Although RF models provide relative ranking of variable importance in 
predicting the dependent variable, they are limited in showing the types and strengths 
of the relationships between each independent environmental factor and fire ignition 
probability. Logistic regression models were also developed to support analyzing the 
potential impacts of environmental factors on fire. I developed two logistic regression 
models for ?Current? and ?Previous? conditions similarly to the RF modeling 
framework (Table 4.4). CG lightning probability, air temperature and DC are 
confirmed as significant variables in the logistic regression models (p < 0.05). All 
three variables are positively correlated with fire ignitions, implying that the increase 
in the value of those variables (i.e. higher probability of CG lightning, higher air 
temperature and higher values of DC) ultimately leads to higher probability of fire 
ignitions. Although not ranked as the top predictor in RF results, NDII6 is found to 
108 
 
 
 
have a significantly negative relationship with fire ignitions in both logistic models (p 
< 0.05). Other variables, like RH, rain, FFMC and fuel type covers are found 
significant in one of the two models. In particular, the fractional covers of woody and 
herbaceous components are positively related to the fire ignitions, while that of the 
nonvascular component has a negative relationship. 
Table 4.4. Logistic regression results for the two models 
Current model Previous model 
Variables Estimate Std. Error Pr(>|z|) Estimate 
Std. 
Error Pr(>|z|) 
Intercept -4.550 2.607 0.08 -2.003 2.367 0.7964 
Lightning 5.428 0.591 < 2e-16 3.430 0.543 1.0e-10 *** *** 
NDII6 -12.69 2.028 3.8e-10 *** -18.360 5.581 6.4e-5 ** 
Rain -0.136 0.064 0.0327* -0.043 0.037 0.2083 
Temp 0.166 0.053 0.0017 ** 0.098 0.042 0.0211 * 
RH 0.005 0.019 0.7910 -0.057 0.016 0.0001 *** 
Wind 0.012 0.074 0.8657 -0.225 0.076 0.0029 ** 
FFMC -0.029 0.016 0.0541 -0.034 0.014 0.0155 * 
DMC 0.008 0.031 0.7814 0.0001 0.027 0.6910 
DC 0.006 0.002 0.0002 *** 0.005 0.002 0.0027 ** 
Region 
(Seward) -1.220 0.520 0.0190 * -1.176 0.432 0.0047 ** 
Region 
(SW) -0.192 0.793 0.8084 1.973 0.662 0.0071 ** 
Elevation -0.002 0.001 0.0077 ** -0.001 0.001 0.0685 
Aspect -0.002 0.003 0.5904 -0.0004 0.003 0.8908 
Woody 1.089 2.173 0.6163 0.073 1.793 0.7246 
Herbaceous 3.625 1.953 0.0635 5.542 1.727 0.0005*** 
Nonvascular -2.811 1.234 0.0223 * -0.223 1.058 0.9110 
Significance codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05  
 
A categorical variable ?Region? was also included in the models to account 
for the potential variability across three major tundra regions, including the North 
Slope and Brook Range (referred as ?North Slope?), Seward Peninsula and Kotzebue 
109 
 
 
 
Lowlands (referred as ?Seward?), and Southwest Alaska (referred as ?SW?) (Figure 
3.1). Welch t-test results were summarized for environmental variables between 
?Fire? and ?No fire? conditions for each tundra region in Table 4.5. The mean values 
of variables including CG lightning probability, air temperature, RH and vegetation 
indices for ?Fire? and ?No fire? conditions are consistently statistically separable 
across the three tundra regions with p-values below 0.05. In contrast, fire weather 
indices do not show significant differences of mean values between ?Fire? and ?No 
fire? conditions for the Seward Peninsula. Although topographical features and fuel 
type covers are less important in RF, their mean values of ?Fire? and ?No fire? events 
are significantly different, particularly slope, roughness, and woody component 
coverage. 
Table 4.5. Welch t-test results across different tundra regions. 
Variables North Slope Seward Peninsula Southwest Alaska t p-value t p-value t p-value 
Lightning 
(current) -6.603 1.3e-6*** -5.113 5.9e-6*** -16.18 
<2.2e-
16*** 
Lightning 
(previous) -2.337 0.0293* -5.161 5.0e-6*** -6.414 
6.9e-
10*** 
Temperature <2.2e-
(current) -4.789 5.8e-5*** -1.142 0.2585 -16.59 16*** 
Temperature -4.726 6.4e-5*** -2.410 0.0192* -10.79 <2.2e-(previous) 16*** 
RH (current) 2.985 0.0062** 0.819 0.4156 9.192 <2.2e-16*** 
RH (previous) 3.844 0.0007*** 2.549 0.0134* 7.451 4.4e-13*** 
Rain (current) -0.132 0.8961 4.487 8.9e-6*** 3.755 0.0002** 
Rain (previous) 2.215 0.0300* -0.098 0.9225 4.461 1.0e-5*** 
Wind (current) 0.836 0.4121 -2.941 0.0045** 4.198 3.5e-5*** 
Wind (previous) 2.385 0.0246* 1.989 0.0519 3.523 0.0005** 
FFMC (current) -4.321 2.2e-5*** -1.571 0.1217 -9.433 <2.2e-16*** 
110 
 
 
 
FFMC 
(previous) -9.820 8.8e-5*** -0.021 0.9835 -8.876 
<2.2e-
16*** 
DMC (current) - - -0.305 0.7616 -3.873 0.0001*** 
DMC (previous) - - -0.280 0.7804 -3.375 0.0008*** 
DC (current) -3.446 0.0007*** 1.705 0.0933 0.601 0.5484 
DC (previous) -3.345 0.0010** 1.711 0.0921 0.735 0.463 
ISI (current) -6.010 1.1e-8*** 0.162 0.8719 -9.587 <2.2e-16*** 
ISI (previous) -6.497 9.1e-10*** 0.341 0.7343 -8.502 
2.7e-
16*** 
BUI (current) -3.807 0.0002*** -0.138 0.8905 -3.471 0.0006*** 
BUI (previous) -3.664 0.0003*** -0.113 0.9107 -3.014 0.0027* 
GVMI 2.124 0.0347* 3.879 0.0003*** 8.981 <2.2e-16*** 
NDII6 4.043 7.9e-5*** 3.916 0.0003*** 8.002 9.6e-15*** 
NDII7 4.423 1.9e-5*** 3.444 0.0012** 4.552 7.8e-6*** 
NDWI 3.111 0.0021** 3.278 0.0019** 5.414 1.0e-7*** 
Elevation 2.633 0.0088** 0.749 0.4569 -0.711 0.4775 
Slope 2.200 0.0289* 3.679 0.0004*** 3.394 0.0007*** 
Aspect -2.903 0.0041** -1.017 0.3147 2.832 0.0049* 
Roughness 2.947 0.0035** 4.056 0.0001*** 3.3943 0.0007*** 
Woody -8.689 1.3e-15*** -5.661 4.9e-7*** 9.818 
<2.2e-
16*** 
Herbaceous -1.810 0.0717 -0.879 0.3835 -7.148 3.3e-12*** 
Nonvascular 9.525 <2.2e-16*** -0.441 0.6612 -13.08 
<2.2e-
16*** 
Significance codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05  
 
Boxplots across different regions help with visualizing the extent in difference 
in environmental conditions between ?Fire? and ?No fire? settings. CG lightning 
probability values on both fire ignition days and the days before ignitions are 
significantly higher (p-value < 0.05 for all three regions) for fire ignition events 
(Figure 4.10 and Table 4.5). In particular, for the fire ignition days, the majority of 
the ?No fire? events have CG lightning probabilities smaller than 0.25, while the 
111 
 
 
 
lightning probabilities for ?Fire? events have an Interquartile Range (IQR) between 0 
to 1, with mean values around 0.5.  
 
Figure 4.10. Boxplots of CG lightning probability between ?Fire? and ?No fire? 
conditions across three major fuel regions: (a) CG lightning probability on fire 
ignition day, (b) CG lightning on the previous day.  
For near surface weather conditions, mean air temperature values are 
significantly higher while mean RH values are lower for ?Fire? conditions for both 
the North Slope and Southwest Alaska regions (Figure 4.11 and Table 4.5). The 
values, however, do not differ significantly for the Seward Peninsula. Similarly, 
?Fire? conditions have significantly higher mean values of FFMC, ISI and BUI in the 
North Slope and Southwest Alaska (Figures 4.12 and 4.13). DMC and DC values of 
the ?Fire? conditions are also higher than those of the ?No fire? ones in the Southwest 
Alaska and North Slope. Mean values of the vegetation indices related fuel state are 
slightly but significantly lower within the ?Fire? conditions compared to ?No fire? 
(Figure 4.14). 
112 
 
 
 
 
Figure 4.11. Boxplots of (a) air temperature and (b) RH on fire ignition days between 
?Fire? and ?No fire? conditions across three major fuel regions.  
 
Figure 4.12. Boxplots of (a) FFMC and (b) DC on fire ignition days between ?Fire? 
and ?No fire? conditions across three major fuel regions. 
113 
 
 
 
 
Figure 4.13. Boxplots of (a) ISI and (b) BUI on fire ignition days between ?Fire? and 
?No fire? conditions across three major fuel regions. 
 
Figure 4.14. Boxplots of (a) NDII6 and (b) NDII7 on fire ignition days between 
?Fire? and ?No fire? conditions across three major fuel regions. 
Different from the variables described above, the overall patterns (higher or 
lower) of the mean fuel type coverages between ?Fire? and ?No fire? conditions vary 
across different regions (Figure 4.15), particularly for woody and nonvascular 
components. Although logistic regressions suggest positive relationship between 
114 
 
 
 
woody cover and fire ignition, ?Fire? conditions in the North Slope and Seward 
Peninsula are linked to higher fractions of the woody component of the fuels, while 
those in the Southwest Alaska linked to the lower woody fuel fraction. Similarly, the 
significantly higher fraction of nonvascular plants is found within ?Fire? conditions in 
the Southwest Alaska, but the relationship is reversed within the North Slope region. 
 
Figure 4.15. Boxplots of mean fractional covers of (a) woody component, (b) 
herbaceous, and (c) nonvascular components between ?Fire? and ?No fire? events 
across three major fuel regions. 
115 
 
 
 
4.6 Discussions 
CG lightning is commonly assumed to be the primary source of ignition in the 
tundra ecosystems due to the remoteness of the region and the limited human 
activities. This study provides the first quantitative piece of evidence that fully 
justified this assumption, as the results from all models explored in this study point to 
CG lightning probability as the most influential factor that predicts fire ignition. This 
finding is also consistent with previous research conducted in the boreal forests of 
North America (Veraverbeke et al., 2017), while the role of ignition source is not 
always highlighted in research conducted for the other ecosystems (D?az-avalos et al., 
2001; Liu et al., 2012; Vec?n-Arias et al., 2016). Previous studies have also 
established relationships between fires and lightning characteristics observed from 
ground-based detection networks, such as count of strikes, polarity, and peak current 
(Peterson et al., 2010). This study, whereas, suggest that the probability of CG 
lightning modeled purely with atmospheric variables is a powerful indicator of tundra 
fire potential. 
Tundra is a very unusual environment with generally low temperatures and 
high water table, largely due to widespread underlying permafrost (Bliss et al., 1973; 
Wielgolaski and Goodall, 1997). This generally means that the ecosystems are rarely 
moisture limited and are not highly flammable. Evidences from both modeling and 
statistical analyses in this study emphasize the controls of warm and dry weather 
conditions over tundra fire ignitions in Alaska, with near surface air temperature and 
RH significantly related to fires. Higher temperature and lower moisture conditions 
have the potential to increase the flammability of the environment in general. 
116 
 
 
 
Similarly, the study of Masrur et al. (2018), which was conducted with reanalysis data 
at a coarser resolution and a larger spatial scale, indicated the same information. In 
addition to the impacts of air temperature and RH on fuel flammability, they might 
also reflect the high likelihood of convective potential which in turn leads to 
atmospheric instability and ultimately CG lightning.   
In addition, the dominant fuels in the tundra are composed of herbaceous 
(grass and sedge) and dwarf shrub components (Innes, 2013), which are considered 
fine fuels as defined in the CFFWIS. FFMC is designed to describe the fine surface 
fuels in boreal forests (Lawson and Armitage, 2008). Larger FFMC values indicate 
higher flammability of the fine fuels. FFMC values with the ?Fire? conditions in this 
study are generally higher than 70, typically representing dry fuels (Lawson and 
Armitage, 2008). This also shows the suitability of FFMC in describing the moisture 
conditions of tundra fuels. Therefore, it is not surprising that FFMC is highlighted as 
a highly predictive variable and the most influential among all fire weather indices in 
assessing fire ignition probability for the tundra. It is also important, that DC ? a 
variable that tracks deeper drying of fuels and longer local ?drought? conditions ? has 
been shown to be responsive to changes in deep moisture levels in the tundra. Since 
DC is a slow-reacting and representative of the long-term dry conditions (Lawson and 
Armitage, 2008), its significance in the logistic regression models suggests that long-
term dry conditions accumulated for days are also driving the ignition probability. All 
fire weather indices have been developed for boreal forest ecosystems and their 
ability to forecast tundra conditions has been most generally assumed rather than 
117 
 
 
 
tested. This study shows that both FFMC and DC provide a reasonable approximation 
of changes in fuel state that can more readily support fire ignition.  
The inconsistency of the role of vegetation indices representing fuel moisture 
between different models indicates that remote sensing-based indices may not be a 
good representative of the fuel moisture state. Fuel type distributions also show 
different patterns between ?Fire? and ?No fire? conditions across the three major 
tundra regions. Though the flammability of vegetation composites varies according to 
experiment results (Sylvester and Wein, 1981), these patterns suggest that the 
interactions of fuel and weather can be more important in driving fire ignitions than 
the fuel type itself.  
More importantly, this study has demonstrated the effectiveness and the 
strong capability of modeling fire ignition probability with WRF-simulated weather 
conditions in the remote tundra regions across space and time. The key role of CG 
lightning probability identified from the models suggests that current fire 
management efforts are inadequate regarding the integration of CG lightning 
probability in fire danger monitoring and modeling for the ecosystems in the HNL. 
By monitoring lightning potential, temperature and dry conditions, it is promising that 
fire ignition probability can be predicted with high accuracy in remote regions at 5km 
resolution with WRF. Besides, modeling experiments with two types of dynamic 
variables ? on fire ignition days and one day before ignition ? show that using data 
simulated from one day earlier can still achieve acceptable prediction results for 
estimating fire ignition probability. This suggests that by using data acquired before 
fire starts, the method developed from this study can still capture the probability of 
118 
 
 
 
fire ignition very well, indicating the potential improvement of fire forecasting 
performance.  
Although existing efforts have incorporated lightning characteristics in fire 
modeling by matching recorded lightning strikes and fires (Peterson et al., 2010; 
Wotton and Martell, 2005), this research suggests the suitability of using simulated 
CG lightning probability for modeling fire ignitions for several reasons. First, modern 
ground-based lightning detection networks typically have a location accuracy ranging 
from below 1km to 4km and a detection efficiency of between 70% and over 90% 
(Biagi et al., 2007; Dissing and Verbyla, 2003; Koshak et al., 2015; Nag et al., 2014). 
This indicates that CG lightning strikes may not always be recorded by the system 
and that the strike location may not be accurately triangulated. Therefore, commonly 
used matching methods to link lightning characteristics and ignitions can miss the 
lightning strikes that actually ignite the fires, which would introduce errors and 
uncertainties when using detected lightning characteristics in the modeling. The 
selection of matching methods could also affect the results (Moris et al., 2020). Third, 
simulation of lightning characteristics has not been developed based on existing 
numerical weather prediction models, which limits the potential of integrating them 
for fire ignition modeling and forecasting. 
Despite the capability of WRF simulated weather conditions in modeling fire 
ignition in this study, previous research questioned the suitability of utilizing WRF 
for calculating fire weather indices due to its limitations in accurate simulation of 
wind and precipitation (M?lders, 2010; Simpson et al., 2014). However, simulating 
weather variables with NWPs is probably the only available method for describing 
119 
 
 
 
near-surface conditions in the data-scarce regions with very sparsely distributed 
stations like Alaskan tundra.  
Previous work has also found that lightning and fire weather are associated 
with synoptic-scale atmospheric circulation patterns in the HNL (Balzter et al., 2005; 
Fauria and Johnson, 2006; Hayasaka et al., 2016; Kim et al., 2020; Kochtubajda et al., 
2019). Since this study has shown the importance of meteorological variables in 
predicting fire ignitions in the tundra, future work needs to be done to understand the 
synoptic weather conditions that drive the dynamic changes of lightning and fire 
weather, to improve our understanding about the mechanisms of wildland fires in the 
pan-Arctic ecosystems. 
4.7 Conclusions 
This study explores the key drivers of wildland fire ignitions in Alaskan 
tundra from 2001 to 2019 by modeling the impacts of environmental factors on fire 
ignition probability. Specifically, I consider environmental variables related to fuel 
type distribution, fuel state, fire weather, topography, and ignition source. Among all 
environmental factors, CG lightning probability is found to be the most important 
driver of tundra fire ignitions in Alaskan from 2001 to 2019. Higher CG lightning 
probability shows a significant positive relationship with fire ignition probability. In 
addition, warmer and drier weather conditions are also found to be important in 
determining fire ignitions. In particular, air temperature, FFMC, and DC show 
significant positive relationships in modeling fire ignition probability, while RH is 
negatively related with the ignition. Moreover, the empirical RF models developed in 
this study have shown a strong capability in predicting fire ignition probability, given 
120 
 
 
 
the input static variables (fuel types and topographical features) and dynamic 
variables (CG lightning probability and weather) simulated with WRF on both fire 
ignition day and one day before. The findings of this study highlight the necessity of 
incorporating CG lightning modeling and emphasize the benefits of WRF simulation 
for wildland fire monitoring efforts in data-scarce regions like Alaskan tundra. Future 
efforts need to be undertaken to improve both the modeling capability and 
understanding of the synoptic-scale climatic patterns that drive tundra fire ignitions.   
 
121 
 
 
 
Chapter 5: Conclusion 
5.1 Major research findings 
The objective of this dissertation was to advance the scientific understanding 
of wildland fire ignition in the tundra ecosystems of the High Northern Latitudes 
(HNL). Specifically, this dissertation focused on addressing the overarching scientific 
question: ?What environmental factors and mechanisms drive wildland fire ignitions 
in Alaskan tundra?? To achieve my dissertation objective, I explored three integrated 
research questions that took advantage of a large data archive of satellite observations 
and Numerical Weather Prediction (NWP) modeling. The research questions were 
outlined in Section 1.4 of Chapter 1 and described in detail in Chapters 2 ? 4.  
The development of the fractional coverage product of fuel type components 
in Chapter 2 was driven by the lack of efforts in describing the detailed distribution of 
fuel type compositions in Alaskan tundra, where the surface vegetation fuels are 
highly mixed. Although categorical fuel classification systems have been developed 
for years at 30m resolution for the entire Alaska (Anderson, 1982; Hirsch, 1996; 
Ottmar et al., 2007; Scott and Burgan, 2005), their applications in the tundra are 
limited. This is mainly because the systems were designed to describe fuel 
components for the forests, and categorical values were not representative for highly 
mixed surface fuel distribution. Recent efforts have elaborated on developing 
fractional cover products of vegetation types (Beck et al., 2011; Berner et al., 2018; 
Macander et al., 2017). However, these maps were limited to certain regions (e.g., the 
122 
 
 
 
North Slope) and thus not helpful for answering of my overall research question, 
which relates to the entire Alaskan tundra ecosystems.  
To overcome these limitations, I combined field observations collected by our 
team during three fieldtrips in the tundra, and Landsat 8 surface reflectance data from 
2013 to 2018 to develop a fuel type distribution product that covers the entire Alaskan 
tundra circa 2015. I first generated mosaicked composites of the study area from more 
than 1800 Landsat 8 tiles for three seasons (pre-growing, peak-growing and post-
growing) extracted from phenology information based on Moderate Resolution 
Imaging Spectroradiometer (MODIS) observations. I was then able to map the 
fractional cover distributions of three major fuel components, including wood 
(evergreen and deciduous shrub), herbaceous (sedge and grass) and nonvascular 
(lichen and moss). My mapping results not only capture the detailed spatial variations 
of vegetation fuel type components, but also depict the overall distributions of the 
related vegetation communities along the latitudes. The published manuscript (He et 
al., 2019) and data product (He et al., 2020) can support other tundra fire studies in 
the future. 
In addition to fuel type, I explored the primary ignition source of wildfires, 
cloud-to-ground (CG) lightning strikes, in the HNL in Chapter 3 of this dissertation. 
Despite the important role of lightning in igniting wildfires in the tundra, CG 
lightning has never been examined specifically for this ecosystem. Although several 
previous studies have modeled CG lightning activity in the boreal forests of North 
America (Blouin et al., 2016; Burrows et al., 2005; Reap, 1991), their findings are not 
suitable for explaining lightning potential tundra, because of the different land-
123 
 
 
 
atmospheric interactions in the two landscapes (Rivas Soriano et al., 2019; Van 
Heerwaarden and Teuling, 2014). Considering the downscaling and forecasting 
capability of the Weather Research and Forecast (WRF) model, I designed an 
empirical-dynamical framework to model CG lightning probability in the tundra 
through the integration of WRF and random forest (RF) algorithm and to understand 
the atmospheric factors driving lightning occurrence.  
To achieve my goals, I first conducted a sensitivity analysis method to 
identify an optimized parameterization setting of WRF simulation for Alaskan tundra. 
After comparing the statistical metrics calculated between WRF simulated variables 
and observations from the Remote Automatic Weather Stations (RAWS) for several 
case studies, I decided to utilize the Thompson 2-moment and the Mellor-Yamada-
Janjic (MYJ) schemes to represent the microphysics and Planetary Boundary Layer 
(PBL) physics options, respectively. I then identified case studies from lightning days 
with three different severity levels. WRF simulations were run for these cases to 
extract four groups of predictors for modeling CG lightning probability: cloud 
properties, meteorological variables at multiple pressure levels, atmospheric stability 
indices, and existing lightning parameterization schemes in WRF. I developed two 
RF models using variables simulated at 24hrs and 48hrs after WRF initialization time 
for comparison. Both models have shown strong capability of capturing CG lightning 
distribution in the tundra, with the parcel lifted index as the most important factor 
driving lighting distribution.  
After achieving the first two objectives, in Chapter 4, I explored the 
environmental factors driving wildland fires ignitions in Alaskan tundra, which is also 
124 
 
 
 
the overarching research question. Although the factors driving boreal forest fires 
have been extensively studied around the globe at different scales (Balzter et al., 
2005; Kim et al., 2020; Liu et al., 2012; Veraverbeke et al., 2017), much less is 
known for tundra fires. Recently, Masrur et al. (2018) examined the climatic factors 
contributing to tundra fire activity across the circumpolar Arctic region and found that 
warm and dry conditions important through statistical analyses and RF modeling. 
Despite their efforts on understanding climatic conditions suitable for tundra burning, 
more work is required to understand the mechanisms behind the fires. Contemporary 
modeling studies are also needed for accurate characterization of fire ignition 
probability in the tundra and for improved understanding of the driving factors that 
control tundra fire ignitions.  
In this study, I considered four types of environmental factors related to fire 
ignitions: ignition source, fuel conditions (fuel type and moisture state), weather and 
topographical features. I utilized the empirical-dynamical framework developed in 
Chapter 3 to model CG lightning probability, as a representative of ignition source. 
Remote sensing based products were used to describe fuel conditions: fractional 
cover product from Chapter 2 to map fuel type distribution, and MODIS-derived 
vegetation indices to estimate fuel moisture state. Due to the lack of weather stations 
in the remote tundra, weather conditions, including both near-surface weather 
variables and the Canadian Forest Fire Weather Index System (CFFWIS), were 
extracted and calculated using the WRF simulated outputs at 5km resolution. I 
obtained the Interferometric Synthetic Aperture Radar (IfSAR) Digital Elevation 
Model (DEM) product to generate topographical features. After preparing all these 
125 
 
 
 
environmental factors, I utilized both RF and logistic regression methods to 
understand the roles of each factor in driving fire ignition probability. I identified the 
probability of CG lightning as the most important driver of fire ignitions, followed by 
warm and dry weather conditions. Both near-surface weather and fire weather indices 
have shown that fire ignitions tend to occur with higher air temperature, lower 
relative humidity (RH), higher Fine Fuel Moisture Code (FFMC), and higher Drought 
Code (DC). In addition to exploring the drivers, I developed two RF models using 
dynamic weather variables extracted on fire ignition day and one day before fire 
ignition, both of which have shown strong capability of predicting fire ignitions in the 
tundra. My results also highlight that incorporating the modeling of CG lightning 
potential is critical for fire monitoring and management efforts in the HNL.   
5.2 Contribution to Broader Arctic Science Research 
The Arctic region has experienced a higher rate of climate warming than the 
global average, as recorded by surface air temperature during the last two decades 
(Meredith et al., 2019). The warming in the Arctic has led to increasing loss of sea ice 
and glaciers, permafrost thawing, disturbances and carbon emissions, which can in 
return contribute to amplified warming and threaten the human communities and 
wildlife according to the Intergovernmental Panel on Climate Change (IPCC) report 
(Meredith et al., 2019). In particular, this report highlights the role of wildfire as a 
major driver among all disturbance types in regularly affecting the terrestrial 
ecosystems at the continental scale across the HNL, with increasing wildfire activity 
linked with climate warming (Randerson et al., 2006; Veraverbeke et al., 2017). To 
address the impacts of climate change on the Arctic, multiple agencies and 
126 
 
 
 
organizations, including the National Aeronautics and Space Administration (NASA), 
Department of Energy (DOE), National Research Council of the National Academies 
of Sciences and the Interagency Arctic Research Policy Committee (IARPC), have 
started specific scientific programs aimed at understanding the underlying processes 
and potential influences of ecosystem changes in the HNL. 
The Arctic-Boreal Vulnerability Experiment (ABoVE) program initiated by 
NASA focuses on studying the terrestrial ecosystems of tundra and boreal forests in 
the HNL of North America. The ultimate goal of this program is to improve our 
understanding of the complex ecosystem processes and to prepare the society with 
rapid response and adaptation to the environmental and social changes at present and 
in the future (ABoVE, 2014). In particular, understanding the driving mechanisms 
and potential influences of wildfire disturbance regimes in the HNL responds to one 
of the key science questions of ABoVE (ABoVE, 2014): ?What processes are 
contributing to changes in disturbance regimes and what are the impacts of these 
changes?? As part of the Next-Generation Ecosystem Experiments (NGEE) program 
funded by DOE, NGEE Arctic focuses on advancing the predictive understanding and 
the modeling capability of the climate and ecosystem on Earth. One of the key 
science questions of the NGEE Arctic also aims to address: ?What controls the 
vulnerability and resilience of Arctic ecosystems to disturbance, and how do 
disturbances alter the physical and ecological structure and function of these 
ecosystems??  
Both questions from the two programs align very well with the underlying 
theme of this dissertation on advancing the understanding of factors driving wildfires 
127 
 
 
 
in the HNL. Besides, both programs emphasize the need for advancing the integrated 
modeling capabilities of the Arctic ecosystems to support future projections of 
climate change (Fisher et al., 2018). This dissertation has made the first efforts of 
contemporary modeling for wildfire ignitions and examined the driving factors of 
fires in Alaskan tundra, which supports the goals of these programs. The research 
findings of this dissertation, including the developed methods, models, and data 
product, provides the groundwork for an improved understanding of the contributors 
of tundra fires in Alaska, as well as a systematic modeling framework of fire danger 
prediction for the future.  
5.3 Implications for fire management efforts in the HNL 
The results of this dissertation provide insights from multiple perspectives for 
wildfire management efforts in both the tundra and the forest forests of the HNL. 
First, the empirical-dynamical modeling framework developed based on WRF and 
machine learning method in this study has demonstrated a strong capability of 
predicting wildfires in the data-poor regions and can be adopted easily for fire danger 
monitoring. The models developed with variables acquired on different dates in this 
dissertation also show potential for earlier prediction of lightning and fire ignition 
probabilities. However, since tundra weather typically fluctuates on a weekly basis, 
the performance of the modeling can be primarily affected by the capability of 
accurate weather forecasting with WRF. Additionally, modeling fire ignition is an 
essential start for estimating fire spread and quantifying carbon emissions and air 
quality impacts on human communities.  
128 
 
 
 
Second, the critical drivers of wildfire ignitions identified in this dissertation 
highlight the crucial roles of simulated weather conditions for fire danger modeling 
and prediction. CG lightning has been identified as the primary control of tundra fires 
in this dissertation and of boreal forest fires in previous studies (Veraverbeke et al., 
2017). This dissertation makes the first attempt in contemporary modeling of tundra 
fire ignitions with CG lightning incorporated in the modeling system. However, the 
existing fire practices of CG lightning monitoring and modeling adopted by the fire 
management teams are not sufficient for supporting fire danger prediction in the 
HNL, where the fires are primarily ignited by lightning strikes. Current fire danger 
monitoring efforts in the HNL of North America mainly rely on the Canadian Forest 
Fire Danger Rating System (CFFDRS), in which lightning has been involved with the 
ground-based detection networks. Other attempts that integrated characteristics of 
lightning strikes have been made (Castedo-Dorado et al., 2011; Vec?n-Arias et al., 
2016; Wotton and Martell, 2005). This dissertation, nevertheless, suggests that CG 
lightning potential could be a better way of dealing with the ignition source in the 
models compared to the detected locations or characteristics of CG lightning strikes. 
Other factors like near-surface weather and commonly used CFFWIS simulated by 
WRF can greatly support fire danger monitoring. However, more work needs to be 
conducted to improve the WRF simulation for tundra environmental modeling and 
calibrating CFFWIS for representing the actual fire danger in the tundra.  
Moreover, the methodology and data product developed in this dissertation for 
mapping fuel type distribution can be applied for monitoring and updating fuel 
conditions regularly at a large spatial scale, which can benefit the fire monitoring 
129 
 
 
 
efforts as well. Integration of Landsat 8 and Sentinel-2A/B imagery can achieve an 
average revisit time of about three days (Li and Roy, 2017), which can provide more 
data resources for accurate characterization of fuel type distribution and even fuel 
moisture state at 10 ? 30m resolution. Fuel type system specifically developed for the 
tundra will also be necessary for determining the fuel characteristics in this region.  
5.4 Future research directions 
The key theme of this dissertation is to uncover the underlying processes and 
factors that control the ignition potential of wildland fires in the remote Alaskan 
tundra. According to the results of this dissertation, CG lightning is the most critical 
factor that leads to fire ignitions in the tundra. The ignition probability is then 
affected by warm and dry weather conditions, as indicated by higher temperature, 
lower RH, and higher fuel moisture codes (FFMC and DC) from CFFWIS. The 
primary control of these weather variables on the tundra fires in Alaskan strongly 
suggests that further research on synoptic weather and atmospheric circulation 
patterns driving these factors and processes is crucial for improving our 
understanding of underlying mechanisms behind these environmental factors. 
Previous studies have either explored the synoptic-scale weather patterns related to 
lighting or fire weather conditions in boreal forests (Flannigan and Wotton, 2001; 
Hayasaka et al., 2016; Kochtubajda et al., 2019) or explored the impacts of 
teleconnections on boreal forest fires (Balzter et al., 2005; Fauria and Johnson, 2008, 
2006; Kim et al., 2020). However, these patterns have never been explored for the 
tundra fires across the entire pan-Arctic region. Further efforts need to be made to 
130 
 
 
 
gain a thorough understanding of the processes contributing to fire disturbances in the 
tundra.  
Moreover, the interactions and feedbacks among different factors, including 
fires, surface fuels, weather, and climate, are worth of exploring to improve both our 
understanding and modeling capability of tundra fires in the future. In Chapter 4, the 
different responses of fire or no fire events to fuel type components in different tundra 
regions suggests the roles of the weather and fuel interactions in driving fire ignitions. 
French et al. (2015) have already shown that fire weather in Alaskan tundra will 
become more suitable for burnings under climate warming, regardless of the 
mitigation strategies. Vegetation communities could also shift in the tundra as a 
consequence of climate warming and fire disturbances (Kittel et al., 2000; Van Der 
Kolk et al., 2016), which could further affect fire activities by altering fuel type and 
load. Additionally, climate change can affect future CG lightning activity as well, 
although there is a lack of consensus on its future trends (Finney et al., 2018; Romps 
et al., 2014).  
In addition to further efforts on understanding science, more work needs to be 
accomplished to improve the modeling and predicting capability of fires for resource 
management. Due to the vegetation shifts in the tundra, fuel type distribution and fuel 
load products should be updated regularly for accurate characterization to support fire 
ignition and spread modeling and post-fire impact assessment. The combination of 
Landsat 8 and Sentinel 2 imagery can provide frequency observations in the HNL at 
10 ? 30m resolution, which will benefit the mapping of fuel type and load mapping 
across a large spatial scale. More field campaigns, if possible, will also be necessary 
131 
 
 
 
for collecting measurements in the remote tundra to support mapping and validation 
efforts. Calibration and modification of WRF-simulated CFFWIS for the tundra will 
also be necessary for representing fuel moisture conditions and fire danger potential. 
The methodology framework developed in this study provides opportunities for 
modeling fire potential in other data-scarce regions as well.  
 
  
132 
 
 
 
Appendices 
 
Table S1. Statistical metrics for sensitivity analysis for air temperature (?C) 
Metrics Schemes 2006 2007 2010 2015 
Morrison_MYJ 0.8516 0.8746 0.8188 0.8871 
Morrison_YSU 0.8177 0.8265 0.8125 0.8847 
Thompson_MYJ 0.853 0.8719 0.8198 0.8876 
Correlation 
Thompson_YSU 0.8223 0.8325 0.8152 0.8848 
WRF6_MYJ 0.7991 0.8566 0.8149 0.8827 
WRF6_YSU 0.7814 0.7953 0.8037 0.8773 
Morrison_MYJ 2.5009 3.4409 3.3134 2.3018 
Morrison_YSU 2.7685 3.9639 3.3512 2.2776 
Thompson_MYJ 2.5057 3.4206 3.3084 2.3047 
RMSE 
Thompson_YSU 2.7367 3.8569 3.3304 2.2789 
WRF6_MYJ 2.9804 4.0524 3.3782 2.2872 
WRF6_YSU 3.0685 4.6816 3.4462 2.2829 
Morrison_MYJ 1.9731 2.5993 2.5481 1.9007 
Morrison_YSU 2.2029 2.9851 2.6099 1.8639 
Thompson_MYJ 1.9625 2.6014 2.5394 1.9041 
MAE 
Thompson_YSU 2.191 2.9136 2.5855 1.8678 
WRF6_MYJ 2.3314 3.0725 2.5779 1.8877 
WRF6_YSU 2.4759 3.5405 2.6031 1.8653 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
133 
 
 
 
Table S2. Statistical metrics for sensitivity analysis for dew point temperature (?C) 
Metrics Schemes 2006 2007 2010 2015 
Morrison_MYJ 0.8473 0.8058 0.2955 0.7083 
Morrison_YSU 0.8294 0.7961 0.2756 0.6761 
Thompson_MYJ 0.8505 0.7973 0.3024 0.7104 
Correlation 
Thompson_YSU 0.8314 0.7986 0.2823 0.678 
WRF6_MYJ 0.86 0.7888 0.2678 0.7097 
WRF6_YSU 0.8411 0.7645 0.2714 0.6771 
Morrison_MYJ 2.1353 2.1616 3.5545 1.6217 
Morrison_YSU 2.3401 2.202 3.7679 1.7151 
Thompson_MYJ 2.108 2.1449 3.5528 1.6147 
RMSE 
Thompson_YSU 2.3463 2.1009 3.7681 1.709 
WRF6_MYJ 2.2026 2.5742 3.6068 1.6179 
WRF6_YSU 2.4036 2.7192 3.7989 1.7119 
Morrison_MYJ 1.6818 1.6496 2.405 1.257 
Morrison_YSU 1.7835 1.6777 2.5472 1.2803 
Thompson_MYJ 1.6072 1.6539 2.3997 1.2506 
MAE 
Thompson_YSU 1.763 1.6157 2.5433 1.2817 
WRF6_MYJ 1.7363 2.0049 2.3924 1.251 
WRF6_YSU 1.854 2.1117 2.5506 1.28 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
134 
 
 
 
Table S3. Statistical metrics for sensitivity analysis for relative humidity (%) 
Metrics Schemes 2006 2007 2010 2015 
Morrison_MYJ 0.7786 0.7184 0.8504 0.8217 
Morrison_YSU 0.7382 0.7037 0.8479 0.8212 
Thompson_MYJ 0.7567 0.7365 0.848 0.822 
Correlation 
Thompson_YSU 0.7436 0.7012 0.8407 0.8222 
WRF6_MYJ 0.7257 0.7545 0.8133 0.8232 
WRF6_YSU 0.7074 0.6949 0.8292 0.8217 
Morrison_MYJ 12.552 18.672 11.317 9.6938 
Morrison_YSU 13.737 18.823 12.093 9.8978 
Thompson_MYJ 13.173 18.507 11.42 9.6969 
RMSE 
Thompson_YSU 13.714 18.663 12.299 9.8849 
WRF6_MYJ 13.892 18.275 12.046 9.6362 
WRF6_YSU 14.503 18.724 12.147 9.8568 
Morrison_MYJ 9.1164 14.65 8.7566 7.4311 
Morrison_YSU 10.289 15.226 9.2098 7.5925 
Thompson_MYJ 9.5386 14.489 8.8432 7.446 
MAE 
Thompson_YSU 10.018 14.984 9.5095 7.5872 
WRF6_MYJ 10.328 14.559 9.6644 7.3945 
WRF6_YSU 10.648 15.093 9.2811 7.5548 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
135 
 
 
 
Table S4. Statistical metrics for sensitivity analysis for downward solar radiation 
(shortwave and longwave, W/m2) 
Metrics Schemes 2006 2007 2010 2015 
Morrison_MYJ 0.697 0.8255 0.6532 0.8368 
Morrison_YSU 0.6738 0.8203 0.6473 0.8415 
Thompson_MYJ 0.6804 0.8179 0.6642 0.8352 
Correlation 
Thompson_YSU 0.657 0.8066 0.6621 0.8405 
WRF6_MYJ 0.6447 0.807 0.6455 0.835 
WRF6_YSU 0.5894 0.7761 0.6343 0.8393 
Morrison_MYJ 153.06 152.03 187.82 158 
Morrison_YSU 158 154.05 187.07 155.81 
Thompson_MYJ 164.74 155.36 185.41 158.73 
RMSE 
Thompson_YSU 168.67 159.86 178.57 156.02 
WRF6_MYJ 161.14 162.61 190.53 158.9 
WRF6_YSU 178.66 173.71 189.71 156.99 
Morrison_MYJ 93.529 83.737 120.51 74.334 
Morrison_YSU 98.417 86.255 118.22 73.847 
Thompson_MYJ 101.92 85.681 116.75 75.247 
MAE 
Thompson_YSU 105.304 91.491 113.74 74.281 
WRF6_MYJ 98.185 93.542 123.87 75.081 
WRF6_YSU 108.03 99.16 121.95 74.061 
 
 
136 
 
 
 
Bibliography 
 
ABoVE, 2014. A Concise Experiment Plan for the Arctic-Boreal Vulnerability 
Experiment. http://above.nasa.gov/acep/acep_final_pdf.pdf. Accessed 
1/25/2015. 
Ad?mek, M., Jankovsk?, Z., Hadincov?, V., Kula, E., Wild, J., 2018. Drivers of forest 
fire occurrence in the cultural landscape of Central Europe. Landsc. Ecol. 33, 
2031?2045. https://doi.org/10.1007/s10980-018-0712-2 
Ager, A.A., Barros, A.M.G., Day, M.A., Preisler, H.K., Spies, T.A., Bolte, J., 2018. 
Analyzing fine-scale spatiotemporal drivers of wildfire in a forest landscape 
model. Ecol. Modell. 384, 87?102. 
https://doi.org/https://doi.org/10.1016/j.ecolmodel.2018.06.018 
Alaska Department of Fish and Game, 2006. Our Wealth Maintained: A Strategy for 
Conserving Alaska?s Diverse Wildlife and Fish Resources 1?25. 
Anderson, H.E., 1982. Fire Behavior Fuel Model Descriptions B1?B7. 
Anderson, K., 2002. A model to predict lightning-caused fire occurrences. Int. J. 
Wildl. Fire 11, 163?172. 
Anderson, K., 1992. Forecasting Lightning Occurrence and Frequency 1?18. 
Anderson, K., Martell, D.L., Flannigan, M.D., Wang, D., 2000. Modeling of Fire 
Occurrence in the Boreal Forest Region of Canada, in: Fire, Climate Change, 
and Carbon Cycling in the Boreal Forest. pp. 357?367. 
Anderson, K.R., 1991. Models to predict lightning occurrence and frequency over 
Alberta. M.Sc. thesis. University of Alberta. 91 pp. 
Andersson, T., Andersson, M., Jacobsson, C., Nilsson, S., 1989. Thermodynamic 
indices for forecasting thunderstorms in southern Sweden. Meteorol. Mag. 118, 
141?146. 
Baig, M.H.A., Zhang, L., Shuai, T., Tong, Q., 2014. Derivation of a tasselled cap 
transformation based on Landsat 8 at-satellite reflectance. Remote Sens. Lett. 5, 
423?431. https://doi.org/10.1080/2150704X.2014.915434 
Baker, M.B., Blyth, A.M., Christian, H.J., Latham, J., Miller, K.L., Gadian, A.M., 
1999. Relationships between lightning activity and various thundercloud 
parameters: Satellite and modelling studies. Atmos. Res. 51, 221?236. 
https://doi.org/10.1016/S0169-8095(99)00009-5 
Balzter, H., Gerard, F.F., George, C.T., Rowland, C.S., Jupp, T.E., McCallum, I., 
Shvidenko, A., Nilsson, S., Sukhinin, A., Onuchin, A., Schmullius, C., 2005. 
Impact of the Arctic Oscillation pattern on interannual forest fire variability in 
Central Siberia. Geophys. Res. Lett. 32, 1?4. 
https://doi.org/10.1029/2005GL022526 
Barbero, R., Abatzoglou, J.T., Steel, E.A., Larkin, N.K., 2014. Modeling very large-
fire occurrences over the continental United States from weather and climate 
137 
 
 
 
forcing. Environ. Res. Lett. 9. https://doi.org/10.1088/1748-9326/9/12/124009 
Barthe, C., Deierling, W., Barth, M.C., 2010. Estimation of total lightning from 
various storm parameters: A cloud-resolving model study. J. Geophys. Res. 
Atmos. 115, 1?17. https://doi.org/10.1029/2010JD014405 
Bates, B.C., Dowdy, A.J., Chandler, R.E., 2018. Lightning prediction for Australia 
using multivariate analyses of large-scale atmospheric variables. J. Appl. 
Meteorol. Climatol. 57, 525?534. https://doi.org/10.1175/JAMC-D-17-0214.1 
Beamish, A., Coops, N., Chabrillat, S., Heim, B., 2017. A Phenological Approach to 
Spectral Differentiation of Low-Arctic Tundra Vegetation Communities, North 
Slope, Alaska. Remote Sens. 9, 1200. https://doi.org/10.3390/rs9111200 
Beck, P.S.A., Horning, N., Goetz, S.J., Loranty, M.M., Tape, K.D., 2011. Shrub 
Cover on the North Slope of Alaska: a circa 2000 Baseline Map. Arctic, Antarct. 
Alp. Res. 43, 355?363. https://doi.org/10.1657/1938-4246-43.3.355 
Beringer, J., Chapin, F.S., Thompson, C.C., McGuire, A.D., 2005. Surface energy 
exchanges along a tundra-forest transition and feedbacks to climate. Agric. For. 
Meteorol. 131, 143?161. https://doi.org/10.1016/j.agrformet.2005.05.006 
Berner, L.T., Jantz, P., Tape, K.D., Goetz, S., 2018. Tundra plant aboveground 
biomass and shrub dominance mapped across the North Slope of Alaska. 
Environ. Res. Lett. 13, 035002. https://doi.org/10.1088/1748-9326/aaaa9a 
Bhatt, U.S., Walker, D.A., Raynolds, M.K., Bieniek, P.A., Epstein, H.E., Comiso, 
J.C., Pinzon, J.E., Tucker, C.J., Polyakov, I. V., 2013. Recent declines in 
warming and vegetation greening trends over pan-arctic tundra. Remote Sens. 5, 
4229?4254. https://doi.org/10.3390/rs5094229 
Biagi, C.J., Cummins, K.L., Kehoe, K.E., Krider, E.P., 2007. National Lightning 
Detection Network (NLDN) performance in southern Arizona, Texas, and 
Oklahoma in 2003-2004. J. Geophys. Res. Atmos. 112, 1?17. 
https://doi.org/10.1029/2006JD007341 
Bieniek, P.A., Bhatt, U.S., Walsh, J.E., Rupp, T.S., Zhang, J., Krieger, J.R., Lader, 
R., 2016. Dynamical Downscaling of ERA-Interim Temperature and 
Precipitation for Alaska. J. Appl. Meteorol. Climatol. 55, 635?654. 
https://doi.org/10.1175/JAMC-D-15-0153.1 
Bliss, L.C., Courtin, G.M., Pattie, D.L., Riewe, R.R., Whitfield, D.W.A., Widden, P., 
1973. Arctic Tundra Ecosystems, in: Annual Review of Ecology and 
Systematics. pp. 359?399. 
Blouin, K.D., Flannigan, M.D., Wang, X., Kochtubajda, B., 2016. Ensemble lightning 
prediction models for the province of Alberta, Canada. Int. J. Wildl. Fire 25, 
421?432. https://doi.org/10.1071/WF15111 
Bond-Lamberty, B., Peckham, S.D., Ahl, D.E., Gower, S.T., 2007. Fire as the 
dominant driver of central Canadian boreal forest carbon balance. Nature 450, 
89?92. https://doi.org/10.1038/nature06272 
138 
 
 
 
Bourgeau-Chavez, L.L., Leblon, B., Charbonneau, F., Buckley, J.R., 2013. 
Evaluation of polarimetric Radarsat-2 SAR data for development of soil 
moisture retrieval algorithms over a chronosequence of black spruce boreal 
forests. Remote Sens. Environ. 132, 71?85. 
https://doi.org/https://doi.org/10.1016/j.rse.2013.01.006 
Bourgeau?Chavez, L.L., Kasischke, E.S., Riordan, K., Brunzell, S., Nolan, M., Hyer, 
E., Slawski, J., Medvecz, M., Walters, T., Ames, S., 2007. Remote monitoring of 
spatial and temporal surface soil moisture in fire disturbed boreal forest 
ecosystems with ERS SAR imagery. Int. J. Remote Sens. 28, 2133?2162. 
https://doi.org/10.1080/01431160600976061 
Breiman, L., 2002. Manual on setting up, using, and understanding random forests 
v3. 1. Tech. Report, http//oz.berkeley.edu/users/breiman, Stat. Dep. Univ. Calif. 
Berkeley, ? 29. https://doi.org/10.2776/85168 
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5?32. 
Bret-Harte, M.S., Mack, M.C., Shaver, G.R., Huebner, D.C., Johnston, M., Mojica, C. 
a, Pizano, C., Reiskind, J. a, 2013. The response of Arctic vegetation and soils 
following an unusually severe tundra fire. Philos. Trans. R. Soc. B 368, 
20120490. https://doi.org/10.1098/rstb.2012.0490 
Buchhorn, M., Walker, D.A., Heim, B., Raynolds, M.K., Epstein, H.E., Schwieder, 
M., 2013. Ground-based hyperspectral characterization of Alaska tundra 
vegetation along environmental gradients. Remote Sens. 5, 3971?4005. 
https://doi.org/10.3390/rs5083971 
Buiat, M., Porc?, F., Dietrich, S., 2017. Observing relationships between lightning 
and cloud profiles by means of a satellite-borne cloud radar. Atmos. Meas. Tech. 
10, 221?230. https://doi.org/10.5194/amt-10-221-2017 
Burrows, W.R., 2008. Dynamical-Statistical Models for Lightning Prediction to 48-
hr, in: Third Conference on Meteorological Applications of Lightning Data. 
Burrows, W.R., Price, C., Wilson, L.J., 2005. Warm Season Lightning Probability 
Prediction for Canada and the Northern United States. Weather Forecast. 20, 
971?988. https://doi.org/10.1175/WAF895.1 
Cai, L., Alexeev, V.A., Arp, C.D., Jones, B.M., Liljedahl, A., G?deke, A., 2016. 
Dynamical Downscaling Data for Studying Climatic Impacts on Hydrology, 
Permafrost, and Ecosystems in Arctic Alaska. Earth Syst. Sci. Data Discuss. 1?
39. https://doi.org/10.5194/essd-2016-31 
Cai, L., Alexeev, V.A., Arp, C.D., Jones, B.M., Liljedahl, A.K., G?deke, A., 2018. 
The Polar WRF Downscaled Historical and Projected Twenty-First Century 
Climate for the Coast and Foothills of Arctic Alaska. Front. Earth Sci. 5, 1?15. 
https://doi.org/10.3389/feart.2017.00111 
Carroll, M.L., Wooten, M.R., Dimiceli, C., Sohlberg, R.A., Townshend, J.R.G., 2016. 
ABoVE: Surface Water Extent, Boreal and Tundra Regions, North America, 
1991-2011. https://doi.org/10.3334/ornldaac/1324 
139 
 
 
 
Cassano, J.J., Higgins, M.E., Seefeldt, M.W., 2011. Performance of the Weather 
Research and Forecasting Model for Month-Long Pan-Arctic Simulations. Mon. 
Weather Rev. 139, 3469?3488. https://doi.org/10.1175/MWR-D-10-05065.1 
Castedo-Dorado, F., Rodriguez-Perez, J.R., Marcos-Menendez, J.L., Alvarez-
Taboada, M.., 2011. Modelling the probability of lightning-induced forest fire 
occurrence in the province of Le?n ( NW Spain ). For. Syst. 20, 95?107. 
Ceccato, P., Flasse, S., Gr?goire, J.M., 2002. Designing a spectral index to estimate 
vegetation water content from remote sensing data: Part 1: Theoretical approach. 
Remote Sens. Environ. 82, 188?197. https://doi.org/10.1016/S0034-
4257(02)00037-8 
Chopping, M., Su, L., Rango, A., Martonchik, J. V, Peters, D.P.C., Laliberte, A., 
2008. Remote sensing of woody shrub cover in desert grasslands using MISR 
with a geometric-optical canopy reflectance model. Remote Sens. Environ. 112, 
19?34. https://doi.org/https://doi.org/10.1016/j.rse.2006.04.023 
Davidson, S., Santos, M., Sloan, V., Watts, J., Phoenix, G., Oechel, W., Zona, D., 
2016. Mapping Arctic Tundra Vegetation Communities Using Field 
Spectroscopy and Multispectral Satellite Data in North Alaska, USA. Remote 
Sens. 8, 978. https://doi.org/10.3390/rs8120978 
Dee, D., Fasullo, J., Shea, D., Walsh, J., (Eds)., N.C. for A.R.S., 2016. The Climate 
Data Guide: Atmospheric Reanalysis: Overview & Comparison Tables. 
Di Giuseppe, F., Pappenberger, F., Wetterhall, F., Krzeminski, B., Camia, A., 
Libert???, G., Miguel, J.S., 2016. The potential predictability of fire danger 
provided by numerical weather prediction. J. Appl. Meteorol. Climatol. 55, 
2469?2491. https://doi.org/10.1175/JAMC-D-15-0297.1 
D?az-avalos, C., Peterson, D.L., Alvarado, E., Ferguson, S.A., Besag, J.E., 2001. 
Space-time modelling of lightning-caused ignitions in the Blue Mountains , 
Oregon. Can. J. For. Res. 31, 1579?1593. https://doi.org/10.1139/cjfr-31-9-1579 
Dissing, D., Verbyla, D.L., 2003. Spatial patterns of lightning strikes in interior 
Alaska and their relations to elevation and vegetation. Can. J. For. Res. 782, 
770?782. https://doi.org/10.1139/X02-214 
Dyrness, C.T., Norum, R.A., 1983. The effects of experimental fires on black spruce 
forest floors in interior Alaska. Can. J. For. Res. 13, 879?893. 
https://doi.org/10.1139/x83-118 
Ester, M., Kriegel, H.-P., Sander, J., Xu, X., 1996. A density-based algorithm for 
discovering clusters in large spatial databases with noise., in: Kdd. pp. 226?231. 
Farukh, M.A., Hayasaka, H., Kimura, K., 2011. Characterization of lightning 
occurrence in Alaska using various weather indices for Lightning Forecasting. J. 
Disaster Res. 6, 343?355. 
Fauria, M.M., Johnson, E.A., 2006. Large-scale climatic patterns control large 
lightning fire occurrence in Canada and Alaska forest regions. J. Geophys. Res. 
Biogeosciences 111. https://doi.org/10.1029/2006JG000181 
140 
 
 
 
Finney, D.L., Doherty, R.M., Wild, O., Stevenson, D.S., MacKenzie, I.A., Blyth, 
A.M., 2018. A projected decrease in lightning under climate change. Nat. Clim. 
Chang. 8, 1?4. https://doi.org/10.1038/s41558-018-0072-6 
Fisher, J.B., Hayes, D.J., Schwalm, C.R., Huntzinger, D.N., Stofferahn, E., Schaefer, 
K., Luo, Y., Wullschleger, S.D., Goetz, S., Miller, C.E., Griffith, P., Chadburn, 
S., Chatterjee, A., Ciais, P., Douglas, T.A., Genet, H., Ito, A., Neigh, C.S.R., 
Poulter, B., Rogers, B.M., Sonnentag, O., Tian, H., Wang, W., Xue, Y., Yang, 
Z.L., Zeng, N., Zhang, Z., 2018. Missing pieces to modeling the Arctic-Boreal 
puzzle. Environ. Res. Lett. 13. https://doi.org/10.1088/1748-9326/aa9d9a 
Flannigan, M.., Stocks, B.., Wotton, B.., 2000. Climate change and forest fires. Sci. 
Total Environ. 262, 221?229. https://doi.org/10.1016/S0048-9697(00)00524-6 
Flannigan, M.D., Wotton, B.M., 2001. Climate, weather, and area burned, in: Forest 
Fires. Elsevier, pp. 351?373. 
Flannigan, M.D., Wotton, B.M., Marshall, G.A., de Groot, W.J., Johnston, J., Jurko, 
N., Cantin, A.S., 2016. Fuel moisture sensitivity to temperature and 
precipitation: climate change implications. Clim. Change 134, 59?71. 
https://doi.org/10.1007/s10584-015-1521-0 
French, N.H.F., Jenkins, L.K., Loboda, T. V., Flannigan, M., Jandt, R., Bourgeau-
Chavez, L.L., Whitley, M., 2015. Fire in arctic tundra of Alaska: Past fire 
activity, future fire potential, and significance for land management and ecology. 
Int. J. Wildl. Fire 24, 1045?1061. https://doi.org/10.1071/WF14167 
French, N.H.F., Whitley, M.A., Jenkins, L.K., 2016. Fire disturbance effects on land 
surface albedo in Alaskan tundra. J. Geophys. Res. Biogeosciences 121, 841?
854. https://doi.org/10.1002/2015JG003177 
Fronterhouse, B.A., 2012. Alaska Lightning Detection Network (ALDN) Briefing 
Document. Ft. Wainwright, AK Bur. L. Manag. Alaska Fire Serv. 1?4. 
Frost, G. V., Epstein, H.E., Walker, D.A., 2014. Regional and landscape-scale 
variability of Landsat-observed vegetation dynamics in northwest Siberian 
tundra. Environ. Res. Lett. 9, 1?11. https://doi.org/10.1088/1748-
9326/9/2/025004 
Frost, G. V, Loehman, R.A., Saperstein, L.B., Macander, M.J., Nelson, P.R., Paradis, 
D.P., Natali, S.M., 2020. Multi-decadal patterns of vegetation succession after 
tundra fire on the Yukon-Kuskokwim Delta, Alaska. Environ. Res. Lett. 15, 
25003. https://doi.org/10.1088/1748-9326/ab5f49 
Fuquay, D.M., 1980. Forecasting lightning activity level and associated weather. 
USDA For. Serv. Res. Pap. INT-244. 
Gallant, A.L., Binnian, E.F., Omernik, J.M., Shasby, M.B., 1995. Ecoregions of 
Alaska - U.S. Geological Survey Professional Paper 1567 73. 
https://doi.org/0160482909\r0607010010 
Gao, B., 1996. NDWI ? A Normalized Difference Water Index for Remote Sensing 
of Vegetation Liquid Water from Space. Remote Sens. Environ. 58, 257. 
141 
 
 
 
https://doi.org/10.1016/S0034-4257(96)00067-3 
Garc?a, M.J.L., Caselles, V., 1991. Mapping burns and natural reforestation using 
thematic Mapper data. Geocarto Int. 6, 31?37. 
https://doi.org/10.1080/10106049109354290 
Gessner, U., Machwitz, M., Conrad, C., Dech, S., 2013. Estimating the fractional 
cover of growth forms and bare surface in savannas. A multi-resolution approach 
based on regression tree ensembles. Remote Sens. Environ. 129, 90?102. 
https://doi.org/10.1016/j.rse.2012.10.026 
Giannaros, T.M., Kotroni, V., Lagouvardos, K., 2015. Predicting lightning activity in 
Greece with the Weather Research and Forecasting (WRF) model. Atmos. Res. 
156, 1?13. https://doi.org/10.1016/j.atmosres.2014.12.009 
Giglio, L., Descloitres, J., Justice, C.O., Kaufman, Y.J., 2003. An Enhanced 
Contextual Fire Detection Algorithm for MODIS. Remote Sens. Environ. 87, 
273?282. https://doi.org/https://doi.org/10.1016/S0034-4257(03)00184-6 
Gijben, M., Dyson, L.L., Loots, M.T., 2017. A statistical scheme to forecast the daily 
lightning threat over southern Africa using the Unified Model. Atmos. Res. 194, 
78?88. https://doi.org/10.1016/j.atmosres.2017.04.022 
Glisan, J.M., Gutowski, W.J., Cassano, J.J., Higgins, M.E., 2013. Effects of spectral 
nudging in WRF on arctic temperature and precipitation simulations. J. Clim. 26, 
3985?3999. https://doi.org/10.1175/JCLI-D-12-00318.1 
Goetz, S.J., Bunn, A.G., Fiske, G.J., Houghton, R. a, 2005. Satellite-observed 
photosynthetic trends across boreal North America associated with climate and 
fire disturbance. Proc. Natl. Acad. Sci. U. S. A. 102, 13521?5. 
https://doi.org/10.1073/pnas.0506179102 
Guan, K., Wood, E.F., Caylor, K.K., 2012. Multi-sensor derivation of regional 
vegetation fractional cover in Africa. Remote Sens. Environ. 124, 653?665. 
https://doi.org/10.1016/j.rse.2012.06.005 
Guo, F., Wang, G., Su, Z., Liang, H., Wang, W., Lin, F., Liu, A., 2016. What drives 
forest fire in Fujian, China? Evidence from logistic regression and Random 
Forests. Int. J. Wildl. Fire 25, 505?519. 
Gustine, D.D., Brinkman, T.J., Lindgren, M.A., Schmidt, J.I., Rupp, T.S., Adams, 
L.G., 2014. Climate-driven effects of fire on winter habitat for Caribou in the 
Alaskan-Yukon Arctic. PLoS One 9. 
https://doi.org/10.1371/journal.pone.0100588 
Hansen, M.C.C., Potapov, P. V, Moore, R., Hancher, M., Turubanova, S.A. a, 
Tyukavina, A., Thau, D., Stehman, S.V. V, Goetz, S.J.J., Loveland, T.R.R., 
Kommareddy, A., Egorov, A., Chini, L., Justice, C.O.O., Townshend, 
J.R.G.R.G., Patapov, P.V., Moore, R., Hancher, M., Turubanova, S.A. a, 
Tyukavina, A., Thau, D., Stehman, S.V. V, Goetz, S.J.J., Loveland, T.R.R., 
Kommaredy, A., Egorov, A., Chini, L., Justice, C.O.O., Townshend, J.R.G.R.G., 
2013. High-Resolution Global Maps of 21st-Century Forest Cover Change. 
142 
 
 
 
Science (80-. ). 342, 850?854. https://doi.org/10.1126/science.1244693 
Hardisky, M.A., Klemas, V., Smart, R.M., 1983. The influence of soil salinity, 
growth form, and leaf moisture on the spectral radiance of Spartina alterniflora 
canopies. Photogramm. Eng. Remote Sens. 49, 77?83. 
Hayasaka, H., Tanaka, H.L., Bieniek, P.A., 2016. Synoptic-scale fire weather 
conditions in Alaska. Polar Sci. 10, 217?226. 
https://doi.org/10.1016/j.polar.2016.05.001 
He, J., Loboda, T. V, Jenkins, L., Chen, D., 2020. ABoVE: Distribution Maps of 
Wildland Fire Fuel Components across Alaskan Tundra, 2015. 
https://doi.org/10.3334/ORNLDAAC/1761 
He, J., Loboda, T. V, Jenkins, L., Chen, D., 2019. Mapping fractional cover of major 
fuel type components across Alaskan tundra. Remote Sens. Environ. 232, 
111324. https://doi.org/https://doi.org/10.1016/j.rse.2019.111324 
Higuera, P.E., Chipman, M.L., Barnes, J.L., Urban, M.A., Hu, F.S., 2011. Variability 
of tundra fire regimes in Arctic Alaska: Millennial-scale patterns and ecological 
implications. Ecol. Appl. 21, 3211?3226. https://doi.org/10.1890/11-0387.1 
Hines, K.M., Bromwich, D.H., 2008. Development and Testing of Polar Weather 
Research and Forecasting (WRF) Model. Part I: Greenland Ice Sheet 
Meteorology*. Mon. Weather Rev. 136, 1971?1989. 
https://doi.org/10.1175/2007MWR2112.1 
Hines, K.M., Bromwich, D.H., Bai, L.-S., Barlage, M., Slater, A.G., 2011. 
Development and Testing of Polar WRF. Part III: Arctic Land. J. Clim. 24, 26?
48. https://doi.org/10.1175/2010JCLI3460.1 
Hinzman, L.D., Bettez, N.D., Bolton, W.R., Chapin, F.S., Dyurgerov, M.B., Fastie, 
C.L., Griffith, B., Hollister, R.D., Hope, A., Huntington, H.P., Jensen, A.M., Jia, 
G.J., Jorgenson, T., Kane, D.L., Klein, D.R., Kofinas, G., Lynch, A.H., Lloyd, 
A.H., McGuire, A.D., Nelson, F.E., Oechel, W.C., Osterkamp, T.E., Racine, 
C.H., Romanovsky, V.E., Stone, R.S., Stow, D.A., Sturm, M., Tweedie, C.E., 
Vourlitis, G.L., Walker, M.D., Walker, D.A., Webber, P.J., Welker, J.M., 
Winker, K.S., Yoshikawa, K., 2005. Evidence and implications of recent climate 
change in Northern Alaska and other Arctic regions. Clim. Change 72, 251?298. 
https://doi.org/10.1007/s10584-005-5352-2 
Hirsch, K.G., 1996. Canadian Forest Fire Behavior Prediction (FBP) System?: user?s 
guide, Special Report 7. Northern Forestry Centre, Edmonton. 
Holben, B.N., 1986. Characteristics of maximum-value composite images from 
temporal AVHRR data. Int. J. Remote Sens. 7, 1417?1434. 
https://doi.org/10.1080/01431168608948945 
Hu, F.S., Higuera, P.E., Duffy, P., Chipman, M.L., Rocha, A. V., Young, A.M., 
Kelly, R., Dietze, M.C., 2015. Arctic tundra fires: Natural variability and 
responses to climate change. Front. Ecol. Environ. 
https://doi.org/10.1890/150063 
143 
 
 
 
Huemmrich, K.F., Gamon, J.A., Tweedie, C.E., Campbell, P.K.E., Landis, D.R., 
Middleton, E.M., 2013. Arctic tundra vegetation functional types based on 
photosynthetic physiology and optical properties. IEEE J. Sel. Top. Appl. Earth 
Obs. Remote Sens. 6, 265?275. https://doi.org/10.1109/JSTARS.2013.2253446 
Innes, R.J., 2013. Fire regimes of Alaskan tundra communities, Fire Effects 
Information System, [Online]. 
Jandt, R., Joly, K., Meyers, C.R., Racine, C., 2008. Slow Recovery of Lichen on 
Burned Caribou Winter Range in Alaska Tundra?: Potential Influences of 
Climate Warming and Other Disturbance Factors 40, 89?95. 
https://doi.org/10.1657/1523-0430(06-122) 
Jenkins, L., Bourgeau-Chavez, L., French, N., Loboda, T., Thelen, B., 2014. 
Development of methods for detection and monitoring of fire disturbance in the 
Alaskan tundra using a two-decade long record of synthetic aperture radar 
satellite images. Remote Sens. 6, 6347?6364. 
Jolly, W.M., Cochrane, M. a, Freeborn, P.H., Holden, Z. a, Brown, T.J., Williamson, 
G.J., Bowman, D.M.J.S., 2015. Climate-induced variations in global wildfire 
danger from 1979 to 2013. Nat. Commun. 6, 7537. 
https://doi.org/10.1038/ncomms8537 
Joly, K., Duffy, P. a., Rupp, T.S., 2012. Simulating the effects of climate change on 
fire regimes in Arctic biomes: implications for caribou and moose habitat. 
Ecosphere 3, art36. https://doi.org/10.1890/ES12-00012.1 
Jones, B.M., Grosse, G., Arp, C.D., Miller, E., Liu, L., Hayes, D.J., Larsen, C.F., 
2015. Recent Arctic tundra fire initiates widespread thermokarst development. 
Sci. Rep. 5, 15865. https://doi.org/10.1038/srep15865 
Kasischke, E.S., Hoy, E.E., 2012. Controls on carbon consumption during Alaskan 
wildland fires. Glob. Chang. Biol. 18, 685?699. https://doi.org/10.1111/j.1365-
2486.2011.02573.x 
Kasischke, E.S., Verbyla, D.L., Rupp, T.S., McGuire, A.D., Murphy, K.A., Jandt, R., 
Barnes, J.L., Hoy, E.E., Duffy, P.A., Calef, M., Turetsky, M.R., 2010. Alaska?s 
changing fire regime ? implications for the vulnerability of its boreal 
forestsThis article is one of a selection of papers from The Dynamics of Change 
in Alaska?s Boreal Forests: Resilience and Vulnerability in Response to Climate 
Warming. Can. J. For. Res. 40, 1313?1324. https://doi.org/10.1139/X10-098 
Keyser, A., Leroy Westerling, A., 2017. Climate drives inter-annual variability in 
probability of high severity fire occurrence in the western United States. 
Environ. Res. Lett. 12. https://doi.org/10.1088/1748-9326/aa6b10 
Kim, C.K., Stuefer, M., Schmitt, C.G., Heymsfield, A., Thompson, G., 2014. 
Numerical Modeling of Ice Fog in Interior Alaska Using the Weather Research 
and Forecasting Model. Pure Appl. Geophys. 171, 1963?1982. 
https://doi.org/10.1007/s00024-013-0766-7 
Kim, J.S., Kug, J.S., Jeong, S.J., Park, H., Schaepman-Strub, G., 2020. Extensive 
144 
 
 
 
fires in southeastern Siberian permafrost linked to preceding Arctic Oscillation. 
Sci. Adv. 6, 1?8. https://doi.org/10.1126/sciadv.aax3308 
Kittel, T. G. F., Steffen, W. L., & Chapin, F.S., 2000. Global and regional modelling 
of Arctic-boreal vegetation distribution and its sensitivity to altered forcing. 
Glob. Chang. Biol. 6, 1?18. https://doi.org/10.1046/j.1365-2486.2000.06011.x 
Kochtubajda, B., Stewart, R.E., Flannigan, M.D., Bonsal, B.R., Cuell, C., Mooney, 
C.J., 2019. An Assessment of Surface and Atmospheric Conditions Associated 
with the Extreme 2014 Wildfire Season in Canada?s Northwest Territories. 
Atmosphere-Ocean 57, 73?90. https://doi.org/10.1080/07055900.2019.1576023 
Koshak, W.J., Cummins, K.L., Buechler, D.E., Vant-Hull, B., Blakeslee, R.J., 
Williams, E.R., Peterson, H.S., 2015. Variability of CONUS lightning in 2003-
12 and associated impacts. J. Appl. Meteorol. Climatol. 54, 15?41. 
https://doi.org/10.1175/JAMC-D-14-0072.1 
Kourtz, P., Todd, B., 1991. Predicting the daily occurrence of lightning-caused forest 
fires. Forestry Canada, Petawawa National Forestry Institute. 
Krause, A., Kloster, S., Wilkenskjeld, S., Paeth, H., 2014. The sensitivity of global 
wildfires to simulated past, present, and future lightning frequency. J. Geophys. 
Res. Biogeosciences 119, 312?322. https://doi.org/10.1002/2013JG002502 
Kushida, K., Kim, Y., Tsuyuzaki, S., Fukuda, M., 2009. Spectral vegetation indices 
for estimating shrub cover, green phytomass and leaf turnover in a sedge?shrub 
tundra. Int. J. Remote Sens. 30, 1651?1658. 
https://doi.org/10.1080/01431160802502632 
Lara, M.J., Nitze, I., Grosse, G., McGuire, A.D., 2018. Tundra landform and 
vegetation productivity trend maps for the Arctic Coastal Plain of northern 
Alaska. Sci. Data 5, 180058. 
Larsen, J.N., Anisimov, O.A., Constable, A., Hollowed, A.B., Maynard, N., Prestrud, 
P., Prowse, T.D., Stone, J.M.R., n.d. Polar regions, in: Barros, V.R., Field, C.B., 
Dokken, D.J., Mastrandrea, M.D., Mach, K.J., Bilir, T.E., Chatterjee, M., Ebi, 
K.L., Estrada, Y.O., Genova, R.C., Girma, B., Kissel, E.S., Levy, A.N., 
MacCracken, S., Mastrandrea, P.R., White, L.L. (Eds.), Climate Change 2014: 
Impacts, Adaptation, and Vulnerability. Part B: Regional Aspects. Contribution 
of Working Group II to the Fifth Assessment Report of the Intergovernmental 
Panel of Climate Change. Cambridge University Press, Cambridge, United 
Kingdom and New York, NY, USA, pp. 1567?1612. 
Lawson, B.D., Armitage, O.B., 2008. Weather Guide Canadian Forest Fire Danger 
Rating System, Canadian Forest Service Northern Forestry Centre. 
Li, J., Roy, D.P., 2017. A Global Analysis of Sentinel-2A, Sentinel-2B and Landsat-8 
Data Revisit Intervals and Implications for Terrestrial Monitoring. Remote Sens. 
9. https://doi.org/10.3390/rs9090902 
Liu, N., Budkewitsch, P., Treitz, P., 2017. Examining spectral reflectance features 
related to Arctic percent vegetation cover: Implications for hyperspectral remote 
145 
 
 
 
sensing of Arctic tundra. Remote Sens. Environ. 192, 58?72. 
https://doi.org/10.1016/j.rse.2017.02.002 
Liu, Z., Yang, J., Change, Y., Weisberg, P.J., He, H.S., 2012. Spatial patterns and 
drivers of fire occurrence and its future trend under climate change in a boreal 
forest of Northeast China. Glob. Chang. Biol. 18, 2041?2056. 
https://doi.org/10.1111/j.1365-2486.2012.02649.x 
Loboda, T.V., Hall, J.V., Hall, A.H., Shevade, V.S., 2017. ABoVE: Cumulative 
Annual Burned Area, Circumpolar High Northern Latitudes, 2001-2015. 
https://doi.org/10.3334/ornldaac/1526 
Loboda, T. V., 2009. Modeling fire danger in data-poor regions: A case study from 
the Russian Far East. Int. J. Wildl. Fire 18, 19?35. 
https://doi.org/10.1071/WF07094 
Loboda, T. V., Csiszar, I.A., 2007. Assessing the risk of ignition in the Russian Far 
East within a modeling framework of fire threat. Ecol. Appl. 17, 791?805. 
https://doi.org/10.1890/05-1476 
Loboda, T. V., French, N.H.F., Hight-Harf, C., Jenkins, L., Miller, M.E., 2013. 
Mapping fire extent and burn severity in Alaskan tussock tundra: An analysis of 
the spectral response of tundra vegetation to wildland fire. Remote Sens. 
Environ. 134, 194?209. https://doi.org/10.1016/j.rse.2013.03.003 
Loboda, T. V, Csiszar, I.A., 2007. Reconstruction of fire spread within wildland fire 
events in Northern Eurasia from the MODIS active fire product. Glob. Planet. 
Change 56, 258?273. https://doi.org/10.1016/j.gloplacha.2006.07.015 
Louppe, G., Wehenkel, L., Sutera, A., Geurts, P., 2013. Understanding variable 
importances in forests of randomized trees. Adv. Neural Inf. Process. Syst. 26 
431?439. https://doi.org/NIPS2013_4928 
Lynn, B.H., Yair, Y., Price, C., Kelman, G., Clark, A.J., 2012. Predicting Cloud-to-
Ground and Intracloud Lightning in Weather Forecast Models. Weather 
Forecast. 1470?1478. https://doi.org/10.1175/WAF-D-11-00144.1 
Ma, L., Zhou, Y., Chen, J., Cao, X., Chen, X., 2015. Estimation of Fractional 
Vegetation Cover in Semiarid Areas by Integrating Endmember Reflectance 
Purification Into Nonlinear Spectral Mixture Analysis. IEEE Geosci. Remote 
Sens. Lett. 12, 1175?1179. https://doi.org/10.1109/LGRS.2014.2385816 
Macander, M.J., Frost, G. V., Nelson, P.R., Swingley, C.S., 2017. Regional 
quantitative cover mapping of tundra plant functional types in Arctic Alaska. 
Remote Sens. 9, 1?26. https://doi.org/10.3390/rs9101024 
Macias Fauria, M., Johnson, E.A., 2008. Climate and wildfires in the North American 
boreal forest. Philos. Trans. R. Soc. B Biol. Sci. 363, 2317?2329. 
https://doi.org/10.1098/rstb.2007.2202 
Mack, M.C., Bret-Harte, M.S., Hollingsworth, T.N., Jandt, R.R., Schuur, E.A.G., 
Shaver, G.R., Verbyla, D.L., 2011. Carbon loss from an unprecedented Arctic 
tundra wildfire. Nature 475, 489?92. https://doi.org/10.1038/nature10283 
146 
 
 
 
Marino, E., Ranz, P., Tom??, J.L., Noriega, M.???ngel, Esteban, J., Madrigal, J., 2016. 
Generation of high-resolution fuel model maps from discrete airborne laser 
scanner and Landsat-8 OLI: A low-cost and highly updated methodology for 
large areas. Remote Sens. Environ. 187, 267?280. 
https://doi.org/10.1016/j.rse.2016.10.020 
Masrur, A., Petrov, A.N., DeGroote, J., 2018. Circumpolar spatio-temporal patterns 
and contributing climatic factors of wildfire activity in the Arctic tundra from 
2001-2015. Environ. Res. Lett. 13. https://doi.org/10.1088/1748-9326/aa9a76 
Matsangouras, I.T., Nastos, P.T., Kapsomenakis, J., 2016. Cloud-to-ground lightning 
activity over Greece: Spatio-temporal analysis and impacts. Atmos. Res. 169, 
485?496. https://doi.org/10.1016/j.atmosres.2015.08.004 
McFeeters, S.K., 1996. The use of the Normalized Difference Water Index (NDWI) 
in the delineation of open water features. Int. J. Remote Sens. 17, 1425?1432. 
Mcguiney, E., Shulski, M., Wendler, G., 2005. Alaska lightning climatology and 
application to wildfire science, in: Proceedings of the 85th American 
Meteorological Society Annual Meeting. 
McManus, K.M., Morton, D.C., Masek, J.G., Wang, D., Sexton, J.O., Nagol, J.R., 
Ropars, P., Boudreau, S., 2012. Satellite-based evidence for shrub and graminoid 
tundra expansion in northern Quebec from 1986 to 2010. Glob. Chang. Biol. 18, 
2313?2323. https://doi.org/10.1111/j.1365-2486.2012.02708.x 
Meredith, M., Sommerkorn, M., Cassotta, S., Derksen, C., Ekaykin, A., Hollowed, 
A., Kofinas, G., Mackintosh, A., Melbourne-Thomas, J., Muelbert, M.M.C., 
Ottersen, G., Pritchard, H., Schuur, E.A.G., 2019. Polar regions. In: IPCC 
Special Report on the Ocean and Cryosphere in a Changing Climate [H.-O. 
P?rtner, D.C. Roberts, V. Masson-Delmotte, P. Zhai, M. Tignor, E. Poloczanska, 
K. Mintenbeck, A. Alegr?a, M. Nicolai, A. Okem, J. Petzold, B. Rama, N.M. 
Weyer. https://doi.org/10.1016/S1366-7017(01)00066-6 
M?lders, N., 2010. Comparison of Canadian Forest Fire Danger Rating System and 
National Fire Danger Rating System fire indices derived from Weather Research 
and Forecasting (WRF) model data for the June 2005 Interior Alaska wildfires. 
Atmos. Res. 95, 290?306. https://doi.org/10.1016/j.atmosres.2009.03.010 
M?lders, N., 2008. Suitability of the Weather Research and Forecasting (WRF) 
Model to Predict the June 2005 Fire Weather for Interior Alaska. Weather 
Forecast. 23, 953?973. https://doi.org/10.1175/2008WAF2007062.1 
Moris, J. V, Conedera, M., Nisi, L., Bernardi, M., Cesti, G., Pezzatti, G.B., 2020. 
Lightning-caused fires in the Alps: Identifying the igniting strokes. Agric. For. 
Meteorol. 290, 107990. 
https://doi.org/https://doi.org/10.1016/j.agrformet.2020.107990 
Moritz, M.A., Morais, M.E., Summerell, L.A., Carlson, J.M., Doyle, J., 2005. 
Wildfires, complexity, and highly optimized tolerance. Proc. Natl. Acad. Sci. 
102, 17912?17917. https://doi.org/10.1073/pnas.0508985102 
147 
 
 
 
Mu, X., Song, W., Gao, Z., McVicar, T.R., Donohue, R.J., Yan, G., 2018. Fractional 
vegetation cover estimation by using multi-angle vegetation index. Remote Sens. 
Environ. 216, 44?56. https://doi.org/https://doi.org/10.1016/j.rse.2018.06.022 
M?ller, M.M., Vacik, H., 2017. Characteristics of lightnings igniting forest fires in 
Austria. Agric. For. Meteorol. 240?241, 26?34. 
https://doi.org/https://doi.org/10.1016/j.agrformet.2017.03.020 
Muller, S. V., Racoviteanu, A.E., Walker, D.A., 1999. Landsat MSS-derived land-
cover map of northern Alaska: Extrapolation methods and a comparison with 
photo-interpreted and AVHRR-derived maps. Int. J. Remote Sens. 20, 2921?
2946. https://doi.org/10.1080/014311699211543 
Murray, L.T., 2018. An uncertain future for lightning. Nat. Clim. Chang. 8, 191?192. 
https://doi.org/10.1038/s41558-018-0094-0 
Nag, A., Murphy, M.J., Schulz, W., Cummins, K.L., 2014. Lightning locating 
systems: Characteristics and validation techniques. 2014 Int. Conf. Light. Prot. 
ICLP 2014 1070?1082. https://doi.org/10.1109/ICLP.2014.6973283 
National Centers for Environmental Prediction/National Weather 
Service/NOAA/U.S. Department of Commerce, 2000. NCEP FNL Operational 
Model Global Tropospheric Analyses, continuing from July 1999. 
https://doi.org/10.5065/D6M043C6 
Nowacki, G.J., Spencer, P., Fleming, M., Brock, T., Jorgenson, T., 2003. Unified 
Ecoregions of Alaska: 2001, Open-File Report. 
https://doi.org/10.3133/ofr2002297 
Olthof, I., Fraser, R.H., 2007. Mapping northern land cover fractions using Landsat 
ETM+. Remote Sens. Environ. 107, 496?509. 
https://doi.org/10.1016/j.rse.2006.10.009 
Ottmar, R.D., Sandberg, D. V, Riccardi, C.L., Prichard, S.J., 2007. An overview of 
the Fuel Characteristic Classification System ? Quantifying, classifying, and 
creating fuelbeds for resource planningThis article is one of a selection of papers 
published in the Special Forum on the Fuel Characteristic Classification Syste. 
Can. J. For. Res. 37, 2383?2393. https://doi.org/10.1139/X07-077 
Pearson, R.G., Phillips, S.J., Loranty, M.M., Beck, P.S.A., Damoulas, T., Knight, 
S.J., Goetz, S.J., 2013. Shifts in Arctic vegetation and associated feedbacks 
under climate change. Nat. Clim. Chang. 3, 673?677. 
https://doi.org/10.1038/nclimate1858 
Peterson, D., Wang, J., Ichoku, C., Remer, L.A., 2010. Effects of lightning and other 
meteorological factors on fire activity in the North American boreal forest: 
implications for fire weather forecasting. Atmos. Chem. Phys. 10, 6873?6888. 
https://doi.org/10.5194/acp-10-6873-2010 
Pineda, N., Montany?, J., van der Velde, O.A., 2014. Characteristics of lightning 
related to wildfire ignitions in Catalonia. Atmos. Res. 135?136, 380?387. 
https://doi.org/10.1016/j.atmosres.2012.07.011 
148 
 
 
 
Post, E., Forchhammer, M.C., Bret-Harte, M.S., Callaghan, T. V, Christensen, T.R., 
Elberling, B., Fox, A.D., Gilg, O., Hik, D.S., H?ye, T.T., Ims, R.A., Jeppesen, 
E., Klein, D.R., Madsen, J., McGuire,  a D., Rysgaard, S., Schindler, D.E., 
Stirling, I., Tamstorf, M.P., Tyler, N.J.C., van der Wal, R., Welker, J., Wookey, 
P.A., Schmidt, N.M., Aastrup, P., 2009. Ecological dynamics across the Arctic 
associated with recent climate change. Science 325, 1355?1358. 
https://doi.org/10.1126/science.1173113 
Potapov, P., Hansen, M.C., Stehman, S. V., Loveland, T.R., Pittman, K., 2008. 
Combining MODIS and Landsat imagery to estimate and map boreal forest 
cover loss. Remote Sens. Environ. 112, 3708?3719. 
https://doi.org/10.1016/J.RSE.2008.05.006 
Prestemon, J.P., Hawbaker, T.J., Bowden, M., Carpenter, J., Brooks, M.T., Abt, K.L., 
Sutphen, R., Scranton, S., 2013. Wildfire ignitions?: A review of the science and 
recommendations for empirical modeling 20. 
Price, C., Rind, D., 1994a. Possible implications of global climate change on global 
lightning distributions and frequencies. J. Geophys. Res. 99, 10823?10831. 
Price, C., Rind, D., 1994b. Modeling Global Lightning Distributions in a General 
Circulation Model. Mon. Weather Rev. https://doi.org/10.1175/1520-
0493(1994)122<1930:MGLDIA>2.0.CO;2 
Pyne, S.J., Andrews, P.L., Laven, R.D., 1996. Introduction to Wildland Fire, Second 
Edi. ed. John Wiley & Sons, Inc. 
Racine, C., Jandt, R., Meyers, C., Dennis, J., 2004. Tundra Fire and Vegetation 
Change along a Hillslope on the Seward Peninsula, Alaska, U.S.A. Arctic, 
Antarct. Alp. Res. 36, 1?10. https://doi.org/10.1657/1523-
0430(2004)036[0001:TFAVCA]2.0.CO;2 
Racine, C.H., Johnson, L.A., Viereck, L.A., 1987. Patterns of Vegetation Recovery 
after Tundra Fires in Northwestern Alaska, U.S.A. Arct. Alp. Res. 19, 461?469. 
Rakov, V.A., Uman, M.A., 2003. Lightning: physics and effects. Cambridge 
University Press. 
Randerson, J.T., Liu, H., Flanner, M.G., Chambers, S.D., Jin, Y., Hess, P.G., Pfister, 
G., Mack, M.C., Treseder, K.K., Welp, L.R., Chapin, F.S., Harden, J.W., 
Goulden, M.L., Lyons, E., Neff, J.C., Schuur, E. a G., Zender, C.S., 2006. The 
impact of boreal forest fire on climate warming. Science 314, 1130?1132. 
https://doi.org/10.1126/science.1132075 
Raynolds, M.K., Comiso, J.C., Walker, D.A., Verbyla, D., 2008. Relationship 
between satellite-derived land surface temperatures, arctic vegetation types, and 
NDVI. Remote Sens. Environ. 112, 1884?1894. 
https://doi.org/10.1016/j.rse.2007.09.008 
Reap, R., 1994. Analysis and prediction of lightning strike distributions associated 
with synoptic map types over Florida. Mon. Weather Rev. 
https://doi.org/10.1175/1520-0493(1994)122<1698:AAPOLS>2.0.CO;2 
149 
 
 
 
Reap, R.M., 1991. Climatological characteristics and objective prediction of 
thunderstorms over Alaska. Weather Forecast. https://doi.org/10.1175/1520-
0434(1991)0062.0.CO;2 
Reap, R.M., Foster, D.S., 1979. Automated 12?36 Hour Probability Forecasts of 
Thunderstorms and Severe Local Storms. J. Appl. Meteorol. 18, 1304?1315. 
https://doi.org/10.1175/1520-0450(1979)018<1304:AHPFOT>2.0.CO;2 
Reap, R.M., MacGorman, D.R., 1989. Cloud-to-Ground Lightning: Climatological 
Characteristics and Relationships to Model Fields, Radar Observations, and 
Severe Local Storms. Mon. Weather Rev. 117, 518?535. 
https://doi.org/10.1175/1520-0493(1989)117<0518:CTGLCC>2.0.CO;2 
Rivas Soriano, L., S?nchez Llorente, J.M., Gonz?lez Zamora, A., de Pablo D?vila, F., 
2019. Influence of land cover on lightning and convective precipitation over the 
European continent. Prog. Phys. Geogr. 43, 352?364. 
https://doi.org/10.1177/0309133318825285 
Rocha, A. V., Shaver, G.R., 2011. Postfire energy exchange in arctic tundra: The 
importance and climatic implications of burn severity. Glob. Chang. Biol. 17, 
2831?2841. https://doi.org/10.1111/j.1365-2486.2011.02441.x 
Rocha, A. V, Loranty, M.M., Higuera, P.E., Mack, M.C., Hu, F.S., Jones, B.M., 
Breen, A.L., Rastetter, E.B., Goetz, S.J., Shaver, G.R., 2012. The footprint of 
Alaskan tundra fires during the past half-century: implications for surface 
properties and radiative forcing. Environ. Res. Lett. 7, 044039. 
https://doi.org/10.1088/1748-9326/7/4/044039 
Rollins, M.G., 2009. LANDFIRE: A nationally consistent vegetation, wildland fire, 
and fuel assessment. Int. J. Wildl. Fire 18, 235?249. 
https://doi.org/10.1071/WF08088 
Romps, D.M., Seeley, J.T., Vollaro, D., Molinari, J., Schumann, U., Huntrieser, H., 
Levy, H., Moxim, W.J., Kasibhatla, P.S., Grewe, V., Bond, W.J., Keeley, J.E., 
Pausas, J.G., Keeley, J.E., Price, C., Rind, D., Michalon, N., Nassif, A., Saouri, 
T., Royer, J.F., Pontikis, C.A., Williams, E.R., Price, C., Rind, D., Price, C., 
Rind, D., Banerjee, A., Archibald, A.T., Maycock, A., Telford, P., Abraham, 
N.L., Yang, X., Braesicke, P., Pyle, J., Krause, A., Kloster, S., Wilkenskjeld, S., 
Paeth, H., Williams, E.R., Reeve, N., Toumi, R., Christian, H.J., Christian, H.J., 
Blakeslee, R.J., Boccippio, D.J., Boeck, W.L., Buechler, D.E., Driscoll, K.T., 
Goodman, S.J., Hall, J.M., Koshak, W.J., Mach, D.M., Stewart, M.F., Battan, 
L.J., Tapia, A., Smith, J.A., Dixon, M., Petersen, W.A., Rutledge, S.A., Meijer, 
E.W., Velthoven, P.F.J. van, Brunner, D.W., Huntrieser, H., Kelder, H., 
Williams, E.R., Geotis, S.G., Renno, N., Rutledge, S.A., Rasmussen, E., 
Rickenbach, T., Williams, E., Williams, E., Rosenfeld, D., Madden, N., Gerlach, 
J., Gears, N., Atkinson, L., Dunnemann, N., Frostrom, G., Antonio, M., Biazon, 
B., Camargo, R., Franca, H., Gomes, A., Lima, M., Machado, R., Manhaes, S., 
Nachtigall, L., Piva, H., Quintiliano, W., Machado, L., Artaxo, P., Roberts, G., 
Renno, N., Blakeslee, R., Bailey, J., Boccippio, D., Betts, A., Wolff, D., Roy, B., 
Halverson, J., Rickenbach, T., Fuentes, J., Avelino, E., Pawar, S.D., Lal, D.M., 
150 
 
 
 
Murugavel, P., Murugavel, P., Pawar, S.D., Gopalakrishan, V., Kitzmiller, D., 
Miller, D., Fulton, R., Ding, F., Wacker, R.S., Orville, R.E., Orville, R.E., 
Huffines, G.R., Holle, R.L., Ye, B., Genio, A.D. Del, Lo, K.K.W., Sobel, A.H., 
Camargo, S.J., Trapp, R.J., Diffenbaugh, N.S., Brooks, H.E., Baldwin, M.E., 
Robinson, E.D., Pal, J.S., Trapp, R.J., Diffenbaugh, N.S., Gluhovsky, A., 
Romps, D.M., Muller, C.J., O?Gorman, P.A., Back, L.E., Singh, M.S., 
O?Gorman, P.A., Held, I.M., Soden, B.J., Taylor, K.E., Stouffer, R.J., Meehl, 
G.A., 2014. Projected increase in lightning strikes in the United States due to 
global warming. Science (80-. ). 346, 851?854. 
https://doi.org/10.1126/science.1259100 
Roy, D.P., Ju, J., Kline, K., Scaramuzza, P.L., Kovalskyy, V., Hansen, M., Loveland, 
T.R., Vermote, E., Zhang, C., 2010. Web-enabled Landsat Data (WELD): 
Landsat ETM+ composited mosaics of the conterminous United States. Remote 
Sens. Environ. 114, 35?49. https://doi.org/10.1016/j.rse.2009.08.011 
Rupp, T.S., Olson, M., Adams, L.G., Dale, B.W., Joly, K., Henkelman, J., Collins, 
W.B., Starfield, A.M., 2006. Simulating the influence of various fire regimens 
on caribou winter habitat. Ecol. Appl. 16, 1730?1743. 
https://doi.org/10.1890/1051-0761(2006)016[1730:stiovf]2.0.co;2 
Santos, J.A., Reis, M.A., De Pablo, F., Rivas-Soriano, L., Leite, S.M., 2013. Forcing 
factors of cloud-to-ground lightning over Iberia: Regional-scale assessments. 
Nat. Hazards Earth Syst. Sci. 13, 1745?1758. https://doi.org/10.5194/nhess-13-
1745-2013 
Saunders, C., 2008. Charge separation mechanisms in clouds. Space Sci. Rev. 137, 
335?353. https://doi.org/10.1007/s11214-008-9345-0 
Sch?n, C., Dittrich, J., M?ller, R., 2018. The Error is the Feature: how to Forecast 
Lightning using a Model Prediction Error 2979?2988. 
https://doi.org/10.1145/3292500.3330682 
Scott, J.H., Burgan, R.E., 2005. Standard fire behavior fuel models: a comprehensive 
set for use with Rothermel?s surface fire spread model. 
https://doi.org/10.2737/RMRS-GTR-153 
Selkowitz, D.J., 2010. A comparison of multi-spectral, multi-angular, and multi-
temporal remote sensing datasets for fractional shrub canopy mapping in Arctic 
Alaska. Remote Sens. Environ. 114, 1338?1352. 
https://doi.org/10.1016/j.rse.2010.01.012 
Shafer, P.E., Fuelberg, H.E., 2006. A statistical procedure to forecast warm season 
lightning over portions of the Florida peninsula. Weather Forecast. 21, 851?868. 
https://doi.org/10.1175/WAF954.1 
Shaver, G.R., Kummerow, J., 1991. Phenology, resource allocation, and growth of 
arctic vascular plants, in: Arctic Ecosystems in a Changing Climate. Elsevier, 
pp. 193?211. 
Simpson, C.C., Pearce, H.G., Sturman, A.P., Zawar-Reza, P., 2014. Verification of 
WRF modelled fire weather in the 2009?10 New Zealand fire season. Int. J. 
151 
 
 
 
Wildl. Fire 23, 34?45. 
Skamarock, W.C., Klemp, J.B., Dudhia, J., Gill, D.O., Zhiquan, L., Berner, J., Wang, 
W., Powers, J.G., Duda, M.G., Barker, D.M., Huang, X.-Y., 2019. A Description 
of the Advanced Research WRF Model Version 4 NCAR Technical Note. Natl. 
Cent. Atmos. Res. 145. https://doi.org/10.5065/1dfh-6p97 
Sly, W.K., 1965. A convective index in relation to lightning strikes in Northwestern 
Alberta, CIR 4220 TEC 566. Dept. of Transport, Canada. 11 pp. 
Somers, B., Asner, G.P., Tits, L., Coppin, P., 2011. Endmember variability in 
Spectral Mixture Analysis: A review. Remote Sens. Environ. 115, 1603?1616. 
https://doi.org/https://doi.org/10.1016/j.rse.2011.03.003 
Sousa, J.F., Fragoso, M., Mendes, S., Corte-Real, J., Santos, J.A., 2013. Statistical-
dynamical modeling of the cloud-to-ground lightning activity in Portugal. 
Atmos. Res. 132?133, 46?64. https://doi.org/10.1016/j.atmosres.2013.04.010 
Stewart, B.C., Kunkel, K.E., Steven, L.E., Sun, L., Walsh, J.E., 2013. Regional 
Climate Trends and Scenarios?: Alaska 1?2. 
Stocks, B.J., Lynham, T.J., Lawson, B.D., Alexander, M.E., Wagner, C.E. Van, 
McAlpine, R.S., Dub?, D.E., 1989. The Canadian Forest Fire Danger Rating 
System: An Overview. For. Chron. 65, 450?457. 
https://doi.org/10.5558/tfc65450-6 
Stow, D.A., Burns, B.H., Hope, A.S., 1993. Spectral, spatial and temporal 
characteristics of Arctic tundra reflectance. Int. J. Remote Sens. 14, 2445?2462. 
https://doi.org/10.1080/01431169308904285 
Stow, D.A., Hope, A., McGuire, D., Verbyla, D., Gamon, J., Huemmrich, F., 
Houston, S., Racine, C., Sturm, M., Tape, K., Hinzman, L., Yoshikawa, K., 
Tweedie, C., Noyle, B., Silapaswan, C., Douglas, D., Griffith, B., Jia, G., 
Epstein, H., Walker, D., Daeschner, S., Petersen, A., Zhou, L., Myneni, R., 
2004. Remote sensing of vegetation and land-cover change in Arctic Tundra 
Ecosystems. Remote Sens. Environ. 89, 281?308. 
https://doi.org/10.1016/j.rse.2003.10.018 
Strauss, J., Ulrich, M., Buchhorn, M., 2012. Expeditions to permafrost 2012:" 
Alaskan North Slope/Itkillik"," Thermokarst in Central Yakutia"," EyeSight-
NAAT-Alaska". Berichte zur Polar-und Meeresforschung= Reports polar Mar. 
Res. 655. 
Sulla-Menashe, D., Woodcock, C.E., Friedl, M.A., 2018. Canadian boreal forest 
greening and browning trends: An analysis of biogeographic patterns and the 
relative roles of disturbance versus climate drivers. Environ. Res. Lett. 13. 
https://doi.org/10.1088/1748-9326/aa9b88 
Sylvester, T.W., Wein, R.W., 1981. Fuel characteristics of arctic plant species and 
simulated plant community flammability by Rothermel?s model. Can. J. Bot. 59, 
898?907. https://doi.org/10.1139/b81-125 
Taylor, S.W., Alexander, M.E., 2006. Science, technology, and human factors in fire 
152 
 
 
 
danger rating: The Canadian experience. Int. J. Wildl. Fire 15, 121?135. 
https://doi.org/10.1071/WF05021 
Thornton, P.E., Thornton, M.M., Mayer, B.W., Wilhelmi, N., Wei, Y., Devarakonda, 
R., Cook, R.B., 2017. Daymet: daily surface weather data on a 1-km grid for 
North America, Version 3. ORNL DAAC, Oak Ridge, Tennessee, USA. ORNL 
DAAC, Oak Ridge, Tennessee, USA. 
Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring 
vegetation. Remote Sens. Environ. 8, 127?150. 
https://doi.org/https://doi.org/10.1016/0034-4257(79)90013-0 
Turetsky, M.R., Kane, E.S., Harden, J.W., Ottmar, R.D., Manies, K.L., Hoy, E., 
Kasischke, E.S., 2011. Recent acceleration of biomass burning and carbon losses 
in Alaskan forests and peatlands. Nat. Geosci. 4, 27?31. 
https://doi.org/10.1038/ngeo1027 
Van Beusekom, A.E., Gould, W.A., Monmany, A.C., Khalyani, A.H., Qui?ones, M., 
Fain, S.J., Andrade-N??ez, M.J., Gonz?lez, G., 2018. Fire weather and 
likelihood: characterizing climate space for fire occurrence and extent in Puerto 
Rico. Clim. Change 146, 117?131. https://doi.org/10.1007/s10584-017-2045-6 
Van Der Kolk, H.J., Heijmans, M.M.P.D., Van Huissteden, J., Pullens, J.W.M., 
Berendse, F., 2016. Potential Arctic tundra vegetation shifts in response to 
changing temperature, precipitation and permafrost thaw. Biogeosciences 13, 
6229?6245. https://doi.org/10.5194/bg-13-6229-2016 
Van Heerwaarden, C.C., Teuling, A.J., 2014. Disentangling the response of forest and 
grassland energy exchange to heatwaves under idealized land-atmosphere 
coupling. Biogeosciences 11, 6159?6171. https://doi.org/10.5194/bg-11-6159-
2014 
Van Wagner, C.E., 1987. Development and Structure of the Canadian Forest 
FireWeather Index System, in: Canadian Forestry Service. Forestry Technical 
Report 35. 
Van Wagner, C.E., Pickett, T.L., 1985. Equations and FORTRAN program for the 
Canadian forest fire weather index system. Canadian Forestry Service, 
Petawawa National Forestry Institute, Chalk River, Ontario. Forestry Technical 
Report 33. 18 p. 
Vec?n-Arias, D., Castedo-Dorado, F., Ord?nez, C., Rodr?guez-P?rez, J.R., 2016. 
Biophysical and lightning characteristics drive lightning-induced fire occurrence 
in the central plateau of the Iberian Peninsula. Agric. For. Meteorol. 225, 36?47. 
https://doi.org/10.1016/j.agrformet.2016.05.003 
Veraverbeke, S., Rogers, B.M., Goulden, M.L., Jandt, R.R., Miller, C.E., Wiggins, 
E.B., Randerson, J.T., 2017. Lightning as a major driver of recent large fire 
years in North American boreal forests. Nat. Clim. Chang. 7, 529?534. 
https://doi.org/10.1038/nclimate3329 
Vermote, E., Justice, C., Claverie, M., Franch, B., 2016. Preliminary analysis of the 
153 
 
 
 
performance of the Landsat 8/OLI land surface reflectance product. Remote 
Sens. Environ. 185, 46?56. https://doi.org/10.1016/j.rse.2016.04.008 
Vermote, E.F., Roger, J.C., Ray, J.P., 2015. MODIS Surface Reflectance User?s 
Guide Correspondence. Http://Modis-Sr.Ltdri.Org 35. 
Viedma, O., Urbieta, I.R., Moreno, J.M., 2018. Wildfires and the role of their drivers 
are changing over time in a large rural area of west-central Spain. Sci. Rep. 8, 
17797. https://doi.org/10.1038/s41598-018-36134-4 
Viereck, L.A., 1979. Preliminary results of experimental fires in the black spruce type 
of interior Alaska. Dept. of Agriculture, Forest Service, Pacific Northwest Forest 
and Range~?. 
Viereck, L.A., Dyrness, C.T., Batten, A.R., Wenzlick, K.J., 1992. The Alaska 
Vegetation Classification. USDA Gen. Tech. Rep. 284. 
Vierling, L.A., Deering, D.W., Eck, T.F., 1997. Differences in arctic tundra 
vegetation type and phenology as seen using bidirectional radiometry in the 
early growing season. Remote Sens. Environ. 60, 71?82. 
https://doi.org/10.1016/S0034-4257(96)00139-3 
Vilar, L., G?mez, I., Mart?nez-Vega, J., Echavarr?a, P., Ria?o, D., Mart?n, M.P., 2016. 
Multitemporal Modelling of Socio-Economic Wildfire Drivers in Central Spain 
between the 1980s and the 2000s: Comparing Generalized Linear Models to 
Machine Learning Algorithms. PLoS One 11, e0161344. 
Walker, D.A., Raynolds, M.K., Dani?ls, F.J.A., Einarsson, E., Elvebakk, A., Gould, 
W.A., Katenin, A.E., Kholod, S.S., Markon, C.J., Melnikov, E.S., Moskalenko, 
N.G., Talbot, S.S., Yurtsev, B.A., Team,  the other members of the C., 2005. The 
Circumpolar Arctic vegetation map. J. Veg. Sci. 16, 267?282. 
https://doi.org/10.1111/j.1654-1103.2005.tb02365.x 
Wang, X., Wotton, B.M., Cantin, A.S., Parisien, M.A., Anderson, K., Moore, B., 
Flannigan, M.D., 2017. cffdrs: an R package for the Canadian Forest Fire 
Danger Rating System. Ecol. Process. 6. https://doi.org/10.1186/s13717-017-
0070-z 
Wendler, G., Conner, J., Moore, B., Shulski, M., Stuefer, M., 2011. Climatology of 
Alaskan wildfires with special emphasis on the extreme year of 2004. Theor. 
Appl. Climatol. 104, 459?472. https://doi.org/10.1007/s00704-010-0357-9 
Wielgolaski, F.E., Goodall, D.W., 1997. Polar and alpine tundra. Elsevier. 
Williams, A.P., Seager, R., MacAlady, A.K., Berkelhammer, M., Crimmins, M.A., 
Swetnam, T.W., Trugman, A.T., Buenning, N., Noone, D., McDowell, N.G., 
Hryniw, N., Mora, C.I., Rahn, T., 2015. Correlations between components of the 
water balance and burned area reveal new insights for predicting forest fire area 
in the southwest United States. Int. J. Wildl. Fire 24, 14?26. 
https://doi.org/10.1071/WF14023 
Wilson, A.B., Bromwich, D.H., Hines, K.M., 2012. Evaluation of polar WRF 
forecasts on the arctic system reanalysis domain: 2. Atmospheric hydrologic 
154 
 
 
 
cycle. J. Geophys. Res. Atmos. 117, 1?18. 
https://doi.org/10.1029/2011JD016765 
Wilson, M.F.J., O?Connell, B., Brown, C., Guinan, J.C., Grehan, A.J., 2007. 
Multiscale terrain analysis of multibeam bathymetry data for habitat mapping on 
the continental slope, Marine Geodesy. 
https://doi.org/10.1080/01490410701295962 
Wong, J., Barth, M.C., Noone, D., 2013. Evaluating a lightning parameterization 
based on cloud-top height for mesoscale numerical model simulations. Geosci. 
Model Dev 429?443. https://doi.org/10.5194/gmd-6-429-2013 
Woo, H., Chung, W., Graham, J.M., Lee, B., 2017. Forest fire risk assessment using 
point process modelling of fire occurrence and Monte Carlo fire simulation. Int. 
J. Wildl. Fire 26, 789?805. 
Wotton, B.M., Martell, D.L., 2005. A lightning fire occurrence model for Ontario. 
Can. J. For. Res. 35, 1389?1401. https://doi.org/10.1139/X05-071 
Wotton, B.M., Nock, C.A., Flannigan, M.D., 2010. Forest fire occurrence and climate 
change in Canada. Int. J. Wildl. Fire 19, 253?271. 
https://doi.org/10.1071/WF09002 
Xiao, J., Zhuang, Q., 2007. Drought effects on large fire activity in Canadian and 
Alaskan forests. Environ. Res. Lett. 2. https://doi.org/10.1088/1748-
9326/2/4/044003 
Yair, Y., 2008. Charge generation and separation processes. Space Sci. Rev. 137, 
119?131. https://doi.org/10.1007/s11214-008-9348-x 
Yair, Y., Lynn, B., Price, C., Kotroni, V., Lagouvardos, K., Morin, E., Mugnai, A., 
Del Carmen Llasat, M., 2010. Predicting the potential for lightning activity in 
Mediterranean storms based on the Weather Research and Forecasting (WRF) 
model dynamic and microphysical fields. J. Geophys. Res. Atmos. 115, 1?13. 
https://doi.org/10.1029/2008JD010868 
Yang, J., Weisberg, P.J., Dilts, T.E., Loudermilk, E.L., Scheller, R.M., Stanton, A., 
Skinner, C., 2015. Predicting wildfire occurrence distribution with spatial point 
process models and its uncertainty assessment: a case study in the Lake Tahoe 
Basin, USA. Int. J. Wildl. Fire 24, 380?390. https://doi.org/10.1071/wf14001 
Yebra, M., Chuvieco, E., Ria?o, D., 2008. Estimation of live fuel moisture content 
from MODIS images for fire risk assessment. Agric. For. Meteorol. 148, 523?
536. https://doi.org/10.1016/j.agrformet.2007.12.005 
Yebra, M., Dennison, P.E., Chuvieco, E., Ria?o, D., Zylstra, P., Hunt, E.R., Danson, 
F.M., Qi, Y., Jurdao, S., 2013. A global review of remote sensing of live fuel 
moisture content for fire danger assessment: Moving towards operational 
products. Remote Sens. Environ. 136, 455?468. 
https://doi.org/10.1016/j.rse.2013.05.029 
Young, A.M., Higuera, P.E., Duffy, P.A., Hu, F.S., 2017. Climatic thresholds shape 
northern high-latitude fire regimes and imply vulnerability to future climate 
155 
 
 
 
change. Ecography (Cop.). 40, 606?617. https://doi.org/10.1111/ecog.02205 
Zepka, G.S., Pinto, O., Saraiva, A.C. V, 2014. Lightning forecasting in southeastern 
Brazil using the WRF model. Atmos. Res. 135?136, 344?362. 
https://doi.org/10.1016/j.atmosres.2013.01.008 
Zhu, Z., Wang, S., Woodcock, C.E., 2015. Improvement and expansion of the Fmask 
algorithm: cloud, cloud shadow, and snow detection for Landsats 4?7, 8, and 
Sentinel 2 images. Remote Sens. Environ. 159, 269?277. 
https://doi.org/https://doi.org/10.1016/j.rse.2014.12.014 
 
156