ABSTRACT 
 
 
 
 
Title of Document: TOWARDS AN INTEGRATED SYSTEM FOR 
VEGETATION FIRE MONITORING IN THE 
AMAZON BASIN   
  
 Wilfrid Schroeder, Doctor of Philosophy, 2008 
Directed By: Professor Christopher O. Justice 
Department of Geography 
 
 
 
Biomass burning is a major environmental problem in Amazonia. Satellite fire 
detections represent the primary source of information for fire alert systems, decision 
makers, emissions modeling groups and the scientific community in general. Those 
various users create a growing demand for good quality fire data of higher spatial and 
temporal resolution that can only be achieved via integration of multiple satellite fire 
detection products. The main objective of this dissertation was to develop an 
integrated fire product capable of improved monitoring and characterization of fire 
activity in Brazilian Amazonia.  
Two major active fire detection algorithms based on MODIS and GOES data 
were used to meet the users demand for fire information. Large differences involving 
the performance of the MODIS and GOES fire products required the quantification of 
omission and commission errors in order to allow for appropriate treatment of 
individual detections produced by each data set. 
  
Relatively small omission errors due to cloud obscuration were estimated for 
Brazilian Amazonia. Regional climate conditions result in reduced cloud coverage in 
areas of high fire activity during the peak of the dry season, therefore minimizing the 
effects of cloud obscuration on fire detection omission errors. 
Clear sky omission and commission errors were largely dependent on the 
vegetation and background conditions. Relatively large commission errors occurring 
in high percentage tree cover areas suggested that fire detection algorithms must 
either be regionalized or incorporate additional tests to provide more consistent fire 
information across a broader range of surface conditions. 
Integration of MODIS and GOES fire products using a physical parameter 
describing fire energy (i.e., fire radiative power) was proven difficult due to 
limitations involving the interplay between sensor characteristics and the types of 
fires that occur in Amazonia. As part of this research, a new integrated product was 
generated based on binary fire detection information derived from MODIS and GOES 
data, incorporating adjustments to reduce commission and omission errors and 
optimizing the complementarities among individual detections.  
These findings made a significant contribution to fire monitoring science in 
Amazonia and could play an important role in the development of future fire 
detection algorithms for tropical regions. 
 
  
 
 
 
 
 
 
 
 
TOWARDS AN INTEGRATED SYSTEM FOR VEGETATION FIRE 
MONITORING IN THE AMAZON BASIN 
 
 
 
By 
 
 
Wilfrid Schroeder 
 
 
 
 
 
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 
2008 
 
 
 
 
 
 
 
 
Advisory Committee: 
Professor Christopher O. Justice, Chair 
Dr. Ivan A. Csiszar (Co-advisor) 
Professor Ruth S. DeFries  
Professor Zhanqing Li  
Dr. Jeffrey T. Morisette  
Professor John R. J. Townshend  
 
 
 
 
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
? Copyright by 
Wilfrid Schroeder 
2008 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 ii 
 
Preface 
The material contained in Chapters 2 and 3 of this dissertation was published in a 
peer reviewed journal and is for exclusive use in this document in compliance with 
the copyright terms that apply. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 iii 
 
 Acknowledgements 
I thank Dr. Jeffrey Morisette, Douglas Morton, Dr. Ivan Csiszar, Prof. Chris Justice 
and Prof. Ruth DeFries for the unconditional support given to me from the very first 
moment I embarked in this project, and for all the fruitful discussions that followed. I 
thank Dr. Louis Giglio, Elaine Prins and Chris Schmidt for providing numerous 
positive feedbacks and for making large volumes of data easily accessible during 
various phases of my analyses.  
I thank the Brazilian counterparts at the National Institute for Space Research 
(INPE), at the Brazilian Institute for the Environment (IBAMA), and at the System 
for the Protection of Amazonia (SIPAM) and Brazilian Air Force for providing 
essential data and support for this project. I thank Dr. Alberto Setzer, Luis Maurano, 
Dr. Luiz Augusto Machado, Dr. Juan Ceballos, Dr. Carlos Frederico Angelis and Dr. 
Yosio Shimabukuro from INPE for helping with numerous data requests. I thank 
Geraldo Lucatelli, Jo?o Raposo Pereira, Alexandre Perotto, Joaquim Toledo, Antonio 
Cattaneo and so many others from the PROARCO and PREVFOGO forest fire 
monitoring and management programs at IBAMA for helping with the field 
campaigns conducted in the early stages of the research. I am grateful to the late Dr. 
Alexandre Santos for collecting valuable field data and for making many difficult 
moments during the field campaigns seem so fun. 
 I thank the Large Scale Biosphere-Atmosphere Experiment in Amazonia 
(LBA), the Earth System Science Interdisciplinary Center (ESSIC) at the University 
of Maryland, NASA?s Earth and Space Science Fellowship Program, and the 
 
 iv 
 
National Environmental Satellite, Data and Information Service (NESDIS) at NOAA 
Camp Springs for providing financial support for this research project. 
 I thank my wife Renata, my daughter Amanda and my son Daniel for their 
love and patience, and for adding so much joy to the entire process. I also thank all 
my relatives for their constant support despite the long geographic distances involved. 
 
 v 
 
Table of Contents 
 
 
Preface........................................................................................................................... ii 
Acknowledgements......................................................................................................iii 
Table of Contents.......................................................................................................... v 
List of Tables .............................................................................................................. vii 
List of Figures............................................................................................................viii 
 
Chapter 1: Introduction................................................................................................. 1 
1.1 BACKGROUND...................................................................................................... 1 
1.2 SATELLITE ACTIVE FIRE DETECTION IN BRAZILIAN AMAZONIA .......................... 4 
1.3 OBJECTIVES OF THE RESEARCH............................................................................ 6 
1.4 RESEARCH OUTLINE............................................................................................. 8 
 
Chapter 2: Quantifying the Impact of Cloud Obscuration on Remote Sensing of 
Active Fires in Brazilian Amazon............................................................................... 11 
2.1 INTRODUCTION................................................................................................... 11 
2.2 DATA ................................................................................................................. 14 
2.2.1 Active Fire Product.................................................................................... 14 
2.2.2 Precipitation Data ..................................................................................... 16 
2.2.3 Cloud Mask ................................................................................................ 19 
2.3 METHODS........................................................................................................... 23 
2.3.1 Fire Dynamics and Precipitation............................................................... 24 
2.3.2 Fire Spatial and Temporal Dynamics........................................................ 28 
2.3.3 Implementation and Accuracy Assessment ................................................ 31 
2.4 RESULTS AND DISCUSSION................................................................................. 36 
2.5 FINAL REMARKS ................................................................................................ 45 
 
Chapter 3: Validation of GOES and MODIS Active Fire Detection Products Using 
ASTER and ETM+ Data............................................................................................. 49 
3.1 INTRODUCTION................................................................................................... 49 
3.2 DATA ................................................................................................................. 51 
3.2.1 GOES ......................................................................................................... 51 
3.2.2 MODIS ....................................................................................................... 52 
3.2.3 ASTER........................................................................................................ 53 
3.2.4 ETM+......................................................................................................... 53 
3.2.5 Vegetation Data ......................................................................................... 54 
3.2.6 Ground Data .............................................................................................. 55 
3.2.7 Airborne Data ............................................................................................ 55 
3.3 METHODS........................................................................................................... 56 
3.4 RESULTS AND DISCUSSION................................................................................. 62 
3.4.1 Overall Detection Performance................................................................. 63 
3.4.2 Omission Errors......................................................................................... 66 
3.4.3 Commission Errors .................................................................................... 70 
 
 vi 
 
3.4.4 Reducing Commission Errors.................................................................... 79 
3.5 CONCLUSIONS .................................................................................................... 85 
 
Chapter 4: Integrating Geostationary and Polar Orbiting Data for Monitoring 
Vegetation Fires in Brazilian Amazonia..................................................................... 88 
4.1 INTRODUCTION................................................................................................... 88 
4.2 VEGETATION FIRES IN BRAZILIAN AMAZONIA................................................... 92 
4.3 METHODS........................................................................................................... 96 
4.3.1 MODIS FRP............................................................................................... 96 
4.3.2 GOES FRP................................................................................................. 97 
4.3.3 Simulation Data ......................................................................................... 99 
4.3.4 Comparing MODIS and GOES Fire Detections...................................... 101 
4.4 RESULTS AND DISCUSSION............................................................................... 102 
4.4.1 MODIS and GOES FRP........................................................................... 102 
4.4.2 MODIS and GOES Fire Detections......................................................... 107 
4.5 CONCLUSIONS .................................................................................................. 114 
 
Chapter 5:  Integrated Fire Product........................................................................... 118 
5.1 INTRODUCTION................................................................................................. 118 
5.2 METHODS......................................................................................................... 119 
5.3 RESULTS .......................................................................................................... 122 
5.4 QUALITY ASSESSMENT .................................................................................... 125 
5.5 CONCLUSIONS .................................................................................................. 128 
 
Chapter 6: Conclusions............................................................................................. 130 
6.1 SYNTHESIS OF RESEARCH ................................................................................ 130 
6.1.1 Omission Errors due to Cloud Obscuration ............................................ 130 
6.1.2 Clear Sky Omission and Commission Errors .......................................... 131 
6.1.3 Towards Integrating MODIS and GOES Fire Detection Data ............... 137 
6.1.4 Integrated Fire Product ........................................................................... 141 
6.2 IMPLICATIONS FOR SATELLITE-BASED FIRE MONITORING IN BRAZILIAN 
AMAZONIA............................................................................................................. 142 
6.3 FUTURE DEVELOPMENTS ................................................................................. 149 
 
Appendix A............................................................................................................... 152 
ETM+ Active Fire Mask........................................................................................... 152 
Bibliography ............................................................................................................. 154 
 
 
 
 
 
 
 vii 
 
List of Tables 
 
Table 2.1 Fire return rate based on GOES WF-ABBA active fire product  
using a 4?4 km grid covering Brazilian Amazonia??????..   30 
 
Table 2.2 Error matrices for evaluating the performance of the original  
GOES WF-ABBA active fire data (a), the cloud processed  
data (b), and the simple rule approach using 40 km and 120 km  
area sampling (c, d respectively) based on 20 m resolution CBERS  
data (our ?ground truth?) covering part of Acre state in western  
Brazilian Amazonia??????????????????.   44 
 
Table 3.1 Main characteristics of the satellite imagery used???????  56 
 
Table 3.2 Classification of MOD14 and WF-ABBA commission errors based  
on visual inspection of 30 m RGB composites of ASTER and  
ETM+ data??????????????????????. 72 
Table 5.1 Evaluation of WF-ABBA detections and the integrated fire  
product data using 30 m active fire data derived from  
Landsat 5 TM imagery?????????????????.. 128 
 
Table 6.1 Fire data requirements applicable for five major fire data user  
groups???????????????????????... 148 
 
 viii 
 
List of Figures 
Figure 1.1 Top 100 municipalities showing the largest numbers of fire  
detections in Brazilian Amazonia using MODIS Terra (A) and  
Aqua (B) Thermal Anomalies and the WF-ABBA GOES (C) data  
for 2003???????????????????????..  4 
 
Figure 1.2 Flow diagram of dissertation???????????????.  10 
 
Figure 2.1 Fire frequency (fires 10
?2
 km
?2
 year
?1
) across the nine Brazilian  
states in Amazonia. Values based on 3-year average (2003?2005)  
GOES WF-ABBA fire detection data using all observation hours  
available???????????????????????.  14 
 
Figure 2.2 Scatter plots of the satellite precipitation estimates and the surface 
precipitation observations using 1(A), 7(B) and 30 (C) day average  
values and of the difference between the satellite precipitation  
estimates and the surface precipitation data and the satellite  
variance (D) calculated for the 9 pixels centered at the  
Sao Gabriel weather station in Amazonas state????????.  18 
 
Figure 2.3 Brightness temperature (BT) image for 15 August 2004 1815 h  
UTC (A) and the 30 day (July 16? August 15 2004) temporal  
profile (B) for the pixel marked as ??? on the image. Solid line  
represents the observed BT values, the dotted line represents the  
30 day mean value, and the dashed lines the 30 day standard  
deviation???????????????????????.  21 
 
Figure 2.4 Mean precipitation values calculated for a single 40?40 km cell  
(approx. location: 60.10? W 12.50? S) based on N = 151 active fire  
detections observed during 2004?2005. R
2 
= 0.9974??????.  26 
 
Figure 2.5 Cloud obscuration processing diagram???????????..  32 
 
Figure 2.6 Stratification of the diurnal cycle of fire activity into three major  
areas, A1, A2, and A3??????????????????  34 
 
Figure 2.7 Relationship between fires produced by the cloud obscuration  
processing based on simulated data and observed fires for the same  
locations using 2005 data????????????????? 36 
 
 
 
 
 
 
 ix 
 
 
Figure 2.8 Fires omitted due to clouds (A) for year 2005 and the percentage 
contribution (B) of cloud adjusted fires in relation to the total  
number of GOES WF-ABBA fires detected in that same year for  
each 40?40 km grid cell covering Brazilian Amazonia?????. 37 
 
Figure 2.9 GOES WF-ABBA fire diurnal cycle distribution for Brazilian  
Amazonia and the corresponding adjustment produced using the  
cloud obscuration processing methodology and the simple rule  
approach using 40 km and 120 km sampling areas??????? 39 
 
Figure 2.10 Average maximum and minimum percentage cloud coverage (A  
and B respectively) and hours of maximum and minimum  
occurrence arranged in 3 hour bins (C and d respectively)???? 40 
 
Figure 3.1 Location of ASTER (shaded red areas) and ETM+ (red boxes)  
scenes used to validate the MOD14 and WF-ABBA fire detection 
products, over percentage tree cover data [Hansen et al., 2002]?? 54 
 
Figure 3.2 Prescribed burn of a deforestation plot located at 1?35?08? N  
60?57?18? W on January 28, 2003. (A) 1.5 m resolution data from  
AHS RGB combination of bands 9 (0.685 ?m), 5 (0.565 ?m) and  
1(0.445 ?m) acquired at 1438 UTC; (B) fire mask derived from  
AHS band 43 (4.95 ?m); (C) ASTER image of the same fire  
(RGB bands 8-3-1) acquired at 1436 UTC. The fire perimeter  
(white contour) derived from the ASTER fire mask is also  
shown in (C)?????????????????????... 58 
 
Figure 3.3 Typical temperature profiles associated with fires used for: forest 
conversion, burning of piled debris from deforestation, rejuvenating 
grassland, and clearing of secondary forest re-growth?????... 63 
 
Figure 3.4 Fire detection probability curves for MOD14 (A) and WF-ABBA  
(B) fire detection products derived for four percentage tree cover  
intervals???????????????????????..  65 
 
Figure 3.5 Omission error estimates produced for MODIS and GOES based  
on fires sampled at approximately 1030 local time??????? 67 
 
Figure 3.6 Commission error estimates for MOD14 and WF-ABBA. Mean  
percentage tree cover values (A) estimated for an area of 
approximately 20?20 km centered on the commission error  
pixel; standard deviation values (B) determined by sampling  
a 9?9 pixels window centered on the commission error pixel.??  71 
 
 
 
 x 
 
 
Figure 3.7 Daily number of hot spot detections from Terra MODIS (MOD14) 
and Aqua MODIS (MYD14) data coincident with 60 different  
locations where MOD14 false detections were identified via 30 m  
ASTER imagery. The shaded area describes the accumulated  
percentage of all 60 locations for which a first detection was  
assigned during the 30-day period before the confirmed false  
detection???????????????????????.  74 
 
Figure 3.8 Half-hourly distribution of WF-ABBA false detections identified  
in the validation data set. The absolute number of confirmed false 
detections identified via ASTER or ETM+ is shown on top of the  
vertical bars??????????????????????  76 
 
Figure 3.9 Diurnal cycle of brightness temperature derived from channel 2 of  
the GOES imager for two adjacent areas characterized by  
contrasting vegetation cover conditions. The images were acquired  
on 04 August 2002 and the forested and deforested areas were  
located at 11?59?29? 52?49?52? W and 12?06?20 S 52?49?09? W, 
respectively. The two curves describe the mean values based on 3 
and 4 pixels representing the deforested and forest areas,  
respectively. The occurrence of WF-ABBA false detections are  
presented along with the difference between the two curves???. 78 
 
Figure 3.10 Boxplots for channel 21 brightness temperature differences  
between the target pixel and its background for true and false  
MOD14 detections. Parameters depicted in the plot include:  
minimum, 25th percentile (lower quartile), mean, 75th percentile  
(upper quartile) and maximum value????????????.  79 
 
Figure 3.11 Brightness temperature (BT) profiles for true (filled circle) and  
false (filled square) detections using MODIS (A) and GOES (B)  
data. Open circles and squares represent the BT values during  
detection-free days. Detection information for the pixels marked in  
(C?F): false MOD14 detection (C): 55.1237? W 12.129? S, 23  
May 2002 at 1403 UTC; true MOD14 detection (D): 56.4828? W  
13.225? S, 05 October 2002 at 1409 UTC; false WF-ABBA  
detection (E): 50.879? W 8.4169? S, 29 August 2002 at 1315 UTC;  
true WF-ABBA detection (F): 50.711? W 10.0335? S, 29 August  
2002 at 1315 UTC. False color composites used: 30 m ASTER  
image (C?D); 30 m ETM+ image (E?F)........................................... 81 
 
 
 
 
 
 
 xi 
 
 
Figure 4.1 Cumulative frequency graphs of fire cluster size (A) and minimum 
distance to nearest fire (B)???????????????.. ..  94 
 
Figure 4.2 Burnt area size histogram derived using ETM+ data for Brazilian 
Amazonia. Data is representative of 2,785 individual burn sites 
extracted from 33 ETM+ scenes. A nominal MODIS pixel is equivalent  
to 100 ha, whereas the nominal GOES pixel covers an area of  
approximately 1600 ha.??????????????????  95 
 
Figure 4.3 3D scheme of the Point Spread Function of MODIS (A) and  
GOES (B) 4 ?m channel????????????????... 100 
 
Figure 4.4 Scatter plots of FRP estimates produced for spatially and  
temporally coincident MOD14-WF_ABBA (A) and  
MYD14-WF_ABBA (B) fire detections clusters. The one-to-one  
line is plotted for reference???????????????..  103 
 
Figure 4.5 Fire intensity (Wm
-2
) estimates for four percentage tree cover  
intervals [VCF, Hansen et al., 2002] calculated by dividing per  
pixel FRP estimates for GOES and MODIS by the approximate  
active fire area determined using coincident 30 m resolution  
ASTER and ETM+ imagery [Schroeder et al., 2008b]. Values  
plotted represent the median of the fire intensity data available for  
each VCF interval???????????????????. 105 
 
Figure 4.6 Simulated FRP for MODIS (A) and GOES (B) as a function of  
distance to the pixel?s center. FRP estimates are calculated for a  
single thermal component fire at 800 K covering an area of  
approximately 5.5 ha, with background temperature of 298 K?? 106 
 
Figure 4.7 Rates of coincident detection (gray bars) for MOD14-WF-ABBA 
instantaneous (A), MYD14-WF-ABBA instantaneous (B),  
MOD14-WF-ABBA daytime (C), and MYD14-WF-ABBA  
daytime (D) observations. Commission error rates for  
MOD14 and MYD14 are also shown (black bars)??????... 110 
 
Figure 4.8 Rates of coincident detection (gray bars) for  
WF-ABBA-MOD14 (A), WF-ABBA-MYD14 using  
instantaneous GOES data. Commission error rates for  
WF-ABBA are also shown (black bars)??????????... 111 
 
 
 
 
 
 
 
 xii 
 
 
Figure 4.9 Box plots for the relative contribution of MOD14 (A) and  
MYD14 (B) products to the sum of MODIS and GOES fire  
detections produced using an area of 200?300 km
2
 centered in  
Mato Grosso state, southern Brazilian Amazonia. Parameters  
depicted in the plot include: minimum, 25
th
 percentile, median,  
75
th
 percentile, and maximum values. Results were separated into  
10
o
 scan angle intervals of MODIS. Positive and negative scan  
angles describe the east and west parts of the MODIS swath,  
respectively?????????????????????... 113 
 
Figure 5.1 Flow diagram describing the processing scheme used to integrate  
MOD14, MYD14 and WF-ABBA fire detections. Ingestion of  
WF-ABBA and cloud obscuration omission error estimates occurs at  
30 min intervals during day and nighttime periods of every  
24 h cycle?????????????????.??????? 120 
 
Figure 5.2 Annual (2005) fire detection density maps for the original  
MOD14 (A), MYD14 (B), and WF-ABBA (C) products, and the 
integrated product (D) for Brazilian Amazonia. The scales  
represent the average number of days with detections calculated  
for individual 40 km cells????????????????. 123 
 
Figure 5.3 Net correction applied to the combined detections of MOD14,  
MYD14 and WF-ABBA based primarily on commission error  
rates and omission errors due to cloud obscuration. The scale  
represent the average number of days with detection that were  
added (positive values) or subtracted (negative values) from the  
integrated data for individual 40 km cells compared to the simple  
sum of the products above.????????????????.. 124 
 
 
 
 1 
 
Chapter 1: Introduction 
 
 
1.1 BACKGROUND 
Biomass burning is a major environmental phenomenon influencing the global 
climate, with important effects on the surface energy flux and atmospheric 
composition, and on the Earth?s radiation budget [IPCC, 1995]. Vegetation fires are a 
major source of greenhouse gases including CO
2
 and CH
4, 
and of chemically reactive 
constituents including CO and NO
x
 [Andreae and Merlet, 2001; van der Werf et al., 
2004; Crutzen, 1979; Delany et al., 1985]. Bond et al. [2004] estimated that total 
global emission of black carbon from biomass burning is comparable to that produced 
from the use of fossil fuel, whereas Penner et al. [1992] showed that carbonaceous 
aerosols produced during biomass burning could result in comparable radiative 
forcing to that of anthropogenic sulfates. 
Important feedbacks between forest fragmentation and the use of fire occur at 
the regional level, increasing the susceptibility of altered forests to larger and more 
destructive fires [Cochrane et al., 1999; Nepstad et al., 1999b]. Vegetation fires can 
impact biodiversity through large scale tree mortality, change forest composition and 
affect faunal populations and alter soil nutrient pools, thereby influencing secondary 
forest re-growth [Barlow and Peres, 2006; Barlow et al., 2003; Cochrane and 
Schulze, 1999; Hughes et al., 2000; Moran et al., 2000; Peres et al., 2003]. 
Vegetation fires are also found to have important social implications in tropical areas, 
 
 2 
 
including significant economic losses as a consequence of property damage and 
impacts on industry, and increased health problems among local populations 
[Mendon?a et al., 2004; Reinhardt et al., 2001] 
In the last two decades Brazilian Amazonia has been under significant 
pressure as a result of high annual rates of deforestation [Laurance et al., 2004; 
PRODES, 2008]. Large areas of exceptionally high fire activity resulted from the 
wide spread use of vegetation fires to convert evergreen tropical forests into pastures 
or croplands or to maintain previously deforested areas [Alencar et al., 1997; Nepstad 
et al., 1999a; Sorrensen, 2004]. Human activities associated with those land use 
processes are in turn largely influenced by the regional climate conditions, 
characterized by high rainfall rates occurring during a relatively long wet season 
when fire use is rarely possible. However, during the dry season months, when there 
is a noticeable reduction in precipitation fires are used extensively [Schroeder et al., 
2005]. 
Vegetation fires in Brazilian Amazonia are spatially concentrated in the major 
areas of deforestation [Cochrane and Laurance, 2002; Alencar et al., 2004]. Those 
areas often coincide with the agricultural frontiers, where investments in sustainable 
land use practices are scarce [Morton et al., 2006; Nepstad et al., 2001; Sorrensen, 
2004]. Conservation areas serve as an important mechanism to reduce fire occurrence 
across large areas of tropical forest in Brazilian Amazonia [Arima et al., 2007; 
Nepstad et al., 2006]. However, the increasing pressure from the surrounding areas 
and the limited control exerted by park administration and law enforcement groups 
 
 3 
 
can greatly reduce the efficacy of those areas to prevent fire spread [Ferreira et al., 
1999; Laurance and Williamson, 2001; Pedlowski et al., 2005]. 
Brazilian Amazonia covers an area of approximately 5 Million km
2
 
characterized by a sparse road network and limited infrastructure. In situ monitoring 
of fire activity is very limited and usually constrained to the immediate vicinity of a 
few environmental law enforcement offices [Ferreira et al., 2007]. Satellite active fire 
detection data represent the primary source of information on fire occurrence for 
county, state and federal environmental agencies, Non-Governmental Organizations 
(NGOs), civil society and for the scientific community. 
Fire detection data can be obtained from different satellite sensors covering 
the region [CPTEC, 2008]. However, most users only have access to limited 
information describing the location and timing of detections derived from individual 
products. Differences among fire products and the lack of information describing data 
quality create major difficulties for end users [Schroeder et al., 2005]. As a result, law 
enforcement activities and the decision making process are significantly 
compromised, state and federal strategic plans to assign resources to control fire 
activity at the county level are negatively affected, and scientific studies based on fire 
detection data become subject to large uncertainties (Figure 1.1).  
This dissertation investigates the performance of satellite active fire detection 
data for Brazilian Amazonia and analyzes the potential for integration of the different 
products. The main objective is to improve fire monitoring in the region by properly 
identifying and quantifying the major sources of error affecting individual products 
 
 4 
 
and by optimizing the use of multiple remote sensing fire products through data 
integration. 
 
 
 
 
Figure 1.1: Top 100 municipalities showing the largest numbers of fire detections in 
Brazilian Amazonia using MODIS Terra (A) and Aqua (B) Thermal Anomalies and the WF-
ABBA GOES (C) data for 2003. 
 
 
1.2 SATELLITE ACTIVE FIRE DETECTION IN BRAZILIAN AMAZONIA 
Satellite remote sensing data have been used extensively to detect vegetation fires in 
Brazilian Amazonia over the past two decades. The first program incorporating 
routine monitoring of vegetation fires in the region was implemented by the National 
Institute for Space Research (INPE) in Brazil in response to the Amazon Boundary 
Layer Experiment (ABLE-2A) which documented the occurrence of large smoke 
plumes spreading over forest areas [Setzer and Pereira, 1991]. Since then, awareness 
about the importance of biomass burning in the tropics has increased significantly 
among the scientific community [IPCC, 1995]. In response to the large demand for 
A B C
 
 5 
 
routine fire information for the region, the number of satellite fire detection products 
increased rapidly over the years [CPTEC, 2008]. 
Early satellite active fire detection methods were primarily based on the use of 
simple threshold tests applied to mid-infrared image data of the Advanced Very High 
Resolution Radiometer (AVHRR) aboard the NOAA series of polar orbiting 
environmental satellites [Matson and Holben, 1987; Setzer and Pereira, 1991]. In this 
spectral region, the fire radiative energy peaks above the regular surface background 
(e.g., green vegetation) making pixels with active fires distinguishable [Matson and 
Dozier, 1981]. However, solar contamination in the mid-infrared spectral interval 
often limits the utility of simple threshold algorithms to small geographic regions 
where problems such as false alarms can be managed with the use of conservative 
regionally adjusted tests [Li et al., 2000; Morisette et al., 2005b]. To overcome those 
limitations, alternative fire detection techniques were proposed for continental and 
global remote sensing fire products incorporating the use of contextual approaches 
applied to the mid-infrared channels, complemented by multi-channel filters to reduce 
in particular the commission errors in those products [Giglio et al., 1999; Kaufman et 
al., 1998b; Prins and Menzel, 1992]. 
The Wildfire Automated Biomass Burning Algorithm (WF-ABBA) applied to 
the Geostationary Operational Environmental Satellite (GOES) imager data has been 
generating 30 min active fire detections at 4 km resolution for Brazilian Amazonia 
since the mid 1990?s [Prins and Menzel, 1992, 1994; UW Madison CIMSS, 2008]. 
The WF-ABBA data have been applied to regional fire alert systems and biomass 
 
 6 
 
burning emissions models, with a large and diverse user community currently 
depending on that product. 
With the launch in late 1999 of the Moderate Resolution Imaging 
Spectroradiometer (MODIS) on board the Terra satellite, a new phase in active fire 
monitoring began. The addition of a fire detection channel enabled for the first time 
the systematic characterization of sub-pixel vegetation fires at 1 km spatial resolution 
[Giglio et al., 2003a; Justice et al., 2002]. The lower saturation of the sensor and 
higher geolocation accuracy of the Thermal Anomalies fire detection product helped 
the MODIS data gain popularity over data from previous instruments, including the 
AVHRR series. 
Previous case study analyses have assessed the detection performance of WF-
ABBA and MODIS Thermal Anomalies products over Brazilian Amazonia [Morisette 
et al., 2005b; Prins et al., 1998]. However, detailed characterization of those products 
was missing and differences in sensor characteristics and unexplained differences 
between WF-ABBA and MODIS Thermal Anomalies fire detections were common 
[Foster Brown, personal communication; Schroeder et al., 2005]. The lack of a large 
set of good quality ground truth data was one of the major obstacles limiting a 
comprehensive characterization of those products. 
 
1.3 OBJECTIVES OF THE RESEARCH 
Vegetation fires in Brazilian Amazonia are largely caused by humans. The strong 
synergy between human activities and the fuel conditions determined by regional 
 
 7 
 
climate regimes create unique patterns of fire use across the region. As a result, the 
physical characteristics describing vegetation fires become highly influenced by the 
different land use and land cover types that occur in Brazilian Amazonia. The main 
objective of this dissertation is to quantify errors and evaluate the potential for 
integration of WF-ABBA and MODIS Thermal Anomalies fire detection products 
with attention to the regional conditions that lead to different fire regimes. The 
complementarities among those products are explored in order to generate improved 
fire detection rates with higher confidence levels. 
The main hypothesis developed for this research is: 
By integrating data from multiple sensors we can resolve differences between active 
fire detection products and increase the accuracy of vegetation fire spatial and 
temporal distribution 
Given the considerations listed above, this dissertation is focused on the 
effects of cloud obscuration on detection omission errors, on the commission and 
omission rates associated with different land cover conditions, and on the effects of 
sensor characteristics which can lead to variations in detection rates and fire 
characterization. The dissertation is composed of the following research themes: 
i) Quantify the impact of cloud coverage on fire detection and create 
mechanisms to adjust the detection numbers to reflect actual fire 
occurrence without creating spurious fires; 
ii) Quantify commission and omission errors of WF-ABBA and MODIS 
Thermal Anomalies data for fires with different vegetation-background 
conditions; 
 
 8 
 
iii) Evaluate fire characterization as defined by Fire Radiative Power (FRP)  
as a means to integrate the WF-ABBA and MODIS Thermal Anomalies 
products; 
iv) Quantify scan angle effects on fire detection rates of MODIS; 
v) Develop a strategy to integrate fire data from different sensors to generate 
an improved fire data record for Brazilian Amazonia. 
 
1.4 RESEARCH OUTLINE 
This dissertation is composed of five chapters. Chapter 1 covers the topic of biomass 
burning in Brazilian Amazonia and introduces the main scientific question, providing 
the rationale for this dissertation. In Chapter 2
1
, the effects of cloud obscuration on 
the detection rates of WF-ABBA over Brazilian Amazonia are quantified. A new 
cloud mask product is derived for the GOES imager data to identify pixels containing 
opaque clouds that could prevent fire detection. A pixel based approach is 
implemented using all daytime and nighttime GOES data at full temporal resolution. 
The cloud obscuration omission error estimates produced are compared to the 
approaches used by previous authors to account for the same problem [Cardoso et al., 
2003; Giglio et al., 2006; Roberts et al., 2005]. The comparative analysis is 
substantiated by 20 m resolution burnt area classification data from the China-Brazil 
Earth Resources Satellite (CBERS) used to validate the results. 
 Chapter 3
2
 addresses the detection performance of WF-ABBA and MODIS 
Thermal Anomalies products over Brazilian Amazonia. The two products are 
                                                 
1
 Chapter 2 is a replica of Schroeder et al. [2008a] with only small changes applied to conform with the 
format of this dissertation 
2
 Chapter 3 is a replica of Schroeder et al. [2008b] with only small changes applied to conform with 
the format of this dissertation 
 
 9 
 
validated using ?ground truth? information derived from 30 m resolution Advanced 
Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and Landsat 
Enhanced Thematic Mapper Plus (ETM+) data. In addition to those data sets, in situ 
and airborne data are used to provide information on the different types and dynamics 
of the fires sampled. Commission and omission errors are derived for four major 
regions characterized by distinct intervals of percentage tree cover. The primary 
factors influencing the commission error rates are identified, and an alternative 
method is proposed to reduce the high number of false detections produced by the 
WF-ABBA and MODIS Thermal Anomalies products near forested areas. 
 In Chapter 4 the potential for integrating WF-ABBA and MODIS Thermal 
Anomalies is investigated. Two different approaches are tested. First, FRP estimates 
generated from GOES and MODIS data are used. Major sources of error in FRP 
values are identified and quantified using simulated data. Secondly, the relationship 
between WF-ABBA and MODIS Thermal Anomalies fire detections is established for 
Brazilian Amazonia. Finally, the variation in detection performance of the MODIS 
fire product as a function of scan angle is quantified using area averaged daily sum 
statistics of fire detections for a subset region in southern Brazilian Amazonia. 
 A new integrated fire product based on the WF-ABBA and MODIS Thermal 
Anomalies products is presented in Chapter 5, incorporating adjustments to correct for 
detection errors in both products.  
 The synthesis of results is presented in Chapter 6, followed by a discussion of 
the implications for satellite fire monitoring in Brazilian Amazonia. The flow diagram 
summarizing the main components of this dissertation is shown in Figure 1.2. 
 
 10 
 
 
 
 
 
 
Figure 1.2: Flow diagram of dissertation. 
 
 
 11 
 
Chapter 2: Quantifying the Impact of Cloud Obscuration on 
Remote Sensing of Active Fires in Brazilian Amazon
1
 
 
 
2.1 INTRODUCTION 
Vegetation fires play a significant role in land and atmospheric processes globally. 
Their occurrence is particularly important in tropical regions where human activity is 
still heavily based on the use of fires for land use management and land cover change. 
In the latter case, deforestation and fires are found to be closely related eventually 
leading to important feedback processes that favor faster and more destructive 
depletion of the local forests [Cochrane et al., 1999; Nepstad et al., 1999b]. In 
contrast, at higher latitude regions a greater percentage of fires are caused by 
lightning or are accidental in nature and are highly influenced by the local weather 
and climate conditions [Kasischke et al., 2002]. Fires in the tropics are influenced by 
local conditions too but will also present an equally important component due to the 
influence of human activities which are reflected in the spatial and temporal 
distributions observed across regional to global scales [Barbosa et al., 1999; Scholes 
et al., 1996]. 
Correct quantification of fire events is needed primarily for understanding the 
dynamics of land use and land cover change and therefore subsidize regional 
environmental programs, as well as for providing information for modeling of 
                                                 
1
 The material presented in this Chapter is part of Schroeder et al. [2008a] 
 
 12 
 
emission estimates from biomass combustion [Korontzi et al., 2004; Van der Werf et 
al., 2003]. Consideration of the fire diurnal cycle is also required for biomass burning 
transport models, making high observation frequency a particularly important 
characteristic on active fire monitoring systems [Freitas et al., 2005; Giglio et al., 
2006].  
Satellite sensors have been used to monitor vegetation fire activity for many 
years now, providing greater insight on the processes associated with fire dynamics at 
different scales [Bucini and Lambin, 2002; Carmona-Moreno et al., 2005; Di Bella et 
al., 2006; Dwyer et al., 2000; Giglio et al., 2006]. In fact, the use of satellite data is 
the only way to assess fire activity at spatial and temporal scales required for land 
surface and atmospheric modeling studies. However, the remote sensing methods 
used for monitoring fires have limitations that tend to cause important biases in the 
final products [Boles and Verbyla, 2000; Cardoso et al., 2005; Eva and Lambin, 
1998; Kasischke et al., 2003; Schroeder et al., 2005].  
A major factor influencing fire numbers derived from remotely sensed data is 
the effect caused by cloud obscuration. Because fires have their highest spectral 
emission located in the mid-infrared band, active fire products exploit that part of the 
spectrum to distinguish biomass burning events from the surrounding background 
[Giglio et al., 1999; Justice et al., 2002]. The presence of optically thick clouds along 
the atmospheric path between the target (fire) and the satellite sensor will, however, 
greatly reduce the ability to detect a fire due to severe attenuation of the spectral 
signal emitted by either flaming or smoldering phases of biomass combustion.  
 
 13 
 
Current methods used to compensate satellite active fire detection to account 
for fires missed due to cloud obscuration tend to rely on the assumption that fires 
occur with the same frequency under cloud covered areas as they do in the open 
[Cardoso et al., 2003; Giglio et al., 2003b, 2006; Roberts et al., 2005]. Despite being 
an attractive approach for its simplistic assumption, the adoption of such procedures 
becomes problematic in areas where fires are unevenly distributed in space. Under 
such conditions, the resulting adjustment numbers will be potentially influenced by 
the cell size selected to extrapolate the applicable clear sky fire density to the 
complementary cloud covered area. The major implication of such an approach is 
associated with the assumption of fires in areas with no burning activity which would 
lead to an overestimation of fire numbers.  
Here we present an approach that uses precipitation data and land use 
information to more precisely quantify the potential omission error associated with 
the cloud obscuration affecting satellite active fire detection products. The proposed 
approach is applied to a geostationary satellite fire data set, in order to characterize 
the cloud effect on fire detection over the entire diurnal cycle. The analyses are 
focused on Brazilian Amazonia where intense fire activity and frequent cloud cover 
are prevalent (Figure 2.1). In the sections to follow we describe the data sets used and 
the method developed and present the results produced for 2005. 
 
 
 14 
 
 
Figure 2.1. Fire frequency (fires 10
?2
 km
?2
 year
?1
) across the nine Brazilian States in 
Amazonia. Values based on 3-year average (2003?2005) GOES WF-ABBA fire detection 
data using all observation hours available. 
 
 
2.2 DATA 
2.2.1 Active Fire Product 
In the past two decades multiple satellite-based active fire products have been 
designed using a variety of sensors [Elvidge et al., 1996; Giglio et al., 2000; Kaufman 
et al., 1990, 1998a; Menzel et al., 1991]. The performance of individual products is 
found to be strongly dependent on the sensor's spectral characteristics and on the 
algorithm used, as well as on the imaging characteristics (e.g., pixel size, observation 
geometry) provided by the instruments and the orbital platforms on which they are 
 
 15 
 
mounted [Morisette et al., 2005b; Schroeder et al., 2005]. Despite their usual coarser 
spatial resolution, geostationary satellites are extremely interesting for fire monitoring 
as they provide high observation frequency. High frequency observations are 
important for deriving the diurnal cycle of fire activity at the regional to basin-wide 
level as well as for modeling time dependent transport of biomass burning emissions 
[Freitas et al., 2005]. 
Data from the Geostationary Operational Environmental Satellite positioned at 
75? longitude (GOES-East) are used in this study. The GOES-East satellite is located 
over the study area selected, thereby offering near optimal observation conditions 
(i.e., near nadir viewing geometry) for fire detection. The nominal 4?4 km pixel 
resolution at sub-satellite position remains stable throughout most of the region with 
less than 25% increase in pixel area being observed near the far most corners of the 
study area. 
Two major GOES-based operational active fire products are available for the 
study area. The Wildfire Automated Biomass Burning Algorithm (WF-ABBA) is 
produced by the University of Wisconsin ? Madison and uses a contextual approach 
for detecting active fires [Prins and Menzel, 1992; Prins et al., 1998]. This product is 
available for the period of 2000?present and covers the entire study area. The other 
product is generated by the Weather Forecast and Climate Studies Center in Brazil 
(CPTEC) and combines fixed threshold and contextual tests for detecting fires 
[CPTEC, 2008]; data are available from July 2004 to present. Because of limited 
temporal coverage found with the latter, we used WF-ABBA's active fire product in 
this study. The data set selected for use covers the period of January 2003?December 
 
 16 
 
2005 and is composed of 17,520 half hourly files per year. Each file contains 
information on fire location and acquisition date along with the temperature and area 
estimates and the assigned confidence for each detection. The latter is divided into 
five categories ranging from high (0) to low (5) confidence. Category 5 was rejected 
here as in some cases it was found to be strongly associated with spurious detections 
composed of large clusters of pixels with no similar detections from other sensors. 
Previous validation of WF-ABBA's fire product performed during the Smoke, Clouds 
and Radiation-Brazil (SCAR-B) experiment, showed that forest conversion fires as 
small as 1 ha could be detected by that algorithm [Prins et al., 1998]. WF-ABBA's 
active fire detection product was selected as the source data to which the cloud 
obscuration adjustments were implemented based on the methodology described in 
Section 2.3. The necessary WF-ABBA product files were made available by the 
University of Wisconsin ? Madison. 
 
2.2.2 Precipitation Data 
The second set of data used in this study was the satellite derived precipitation 
estimates. Brazilian Amazonia is still only partially covered by meteorological and 
hydrological surface stations from which precipitation data can be extracted. Rainfall 
is generated primarily within localized convective systems leading to potential effects 
on the quality of the precipitation estimates based on spatially interpolated surface 
station data from the region [Costa and Foley, 1998]. Availability of meteorological 
radar data is also scarce due to limited spatial and temporal coverage. Consequently, 
the use of radar data was not considered as a viable alternative for this study. Because 
 
 17 
 
of these limitations, satellite derived precipitation products were used for assessing 
rainfall conditions throughout Brazilian Amazonia. As with the remote sensing of 
active fires, precipitation estimates are provided by different sensors using different 
algorithms that can be based on single instrument or on multi-sensor multi-data 
approaches [Bellerby et al., 2001; Kummerow et al., 1998; Rudolf et al., 1996]. 
Consideration of the spatial and temporal resolutions and the accuracy of the 
precipitation products that have achieved operational mode led to the use of GOES-
based precipitation estimates for this study. CPTEC currently generates daily GOES-
based precipitation estimates for most of South America at the nominal spatial 
resolution of 4?4 km. CPTEC's product is a revised version of the Hydro-estimator 
method [Vicente et al., 1998], which uses improved coefficients for the specific 
conditions observed over Brazil [Vila and Lima, 2004; Carlos Frederico Angelis ? 
personal communication, 2005]. CPTEC's GOES-based precipitation estimates data 
production was initiated in January 2004 providing daily files with total accumulated 
precipitation at nominal 4 km resolution. The data set used in this study covers 
January 2004?December 2005. 
Prior to the application of CPTEC's Hydro-estimator precipitation data to the 
proposed methodology, an evaluation was performed using point precipitation data 
obtained from 19 automated surface weather stations available across the basin [URL: 
http://tempo.cptec.inpe.br/PCD/]. Figures 2.2A?C show the results for all stations 
selected using three different time integration periods ? 1, 7, and 30 days, 
respectively. The correlation between the two data sets is seen to improve 
significantly with the total number of days used in the sampling process. 
 
 18 
 
 
 
Figure 2.2. Scatter plots of the satellite precipitation estimates and the surface precipitation 
observations using 1 (A), 7 (B) and 30 (C) day average values and of the difference between 
the satellite precipitation estimates and the surface precipitation data and the satellite variance 
(D) calculated for the 9 pixels centered at the Sao Gabriel weather station in Amazonas state. 
 
 
Two major factors were believed to contribute to the dispersion seen in all 
three graphs. First, there is the natural difficulty in relating the surface station's point 
data to the 4?4 km pixel footprint as obtained with the GOES data. The large rainfall 
spatial heterogeneity mentioned above will cause significant impact on the area 
averaged estimates produced from the satellite data. The effect of the spatial 
variability in precipitation patterns in the region is exemplified in Figure 2.2D, where 
the absolute difference calculated for the satellite and surface station precipitation 
A B 
C D 
 
 19 
 
values is plotted against the satellite precipitation variance (?
2
) based on 9 pixels 
centered at one surface station location. 
Another important factor contributing to the dispersion seen in Figures 2.2A?
C is related to the inherent deficiencies in the precipitation product that are associated 
with the use of passive remote sensing data from GOES and empirical relationships 
that tend to impact the accuracy of the precipitation estimates produced [Boi et al., 
2004; Ebert et al., 2007; Rozumalski, 2000; Vicente et al., 2002]. Despite these 
limitations, the high observation frequency of the GOES precipitation estimates 
allows for improved integration of the 24 h precipitation totals thereby providing 
valuable information for use in this study. Also, as will be described in Section 2.3.1, 
our approach will be based on relative differences between precipitation estimates 
making absolute accuracy of the product of less significance. 
 
2.2.3 Cloud Mask 
The WF-ABBA fire product consists of a list of fire detections, but no cloud mask is 
included. However, a cloud mask is needed for identifying cloud obscured pixels that 
require processing for potential fire omission. Therefore we developed a new cloud 
mask product at the 4 km spatial and 30 min temporal resolution of the original active 
fire data, which is the basis of the cloud obscuration processing scheme. The cloud 
mask was designed based on the use of the Global-Merged Infra-Red Brightness 
Temperature (from here on designated simply as BT) product generated by the 
Climate Prediction Center [CPC, 2008]. The BT data are a single band global product 
derived from multiple geostationary satellite imagery and gridded to nominal 4?4 km 
 
 20 
 
resolution. In the case of Brazilian Amazonia the BT product is primarily derived 
from GOES-East imagery.  
The unique conditions of the Amazon region in terms of daily BT trends were 
explored to derive a method that uses fixed and dynamic thresholds to distinguish 
between clouds and land surfaces in the BT images. Daily variation of BT is small for 
clear sky pixels as surface temperature remains much constant within a 30 day period 
[Alval? et al., 2002]. The stability of surface conditions is demonstrated in Figure 
2.3B in which a typical pattern of BT values extracted from a 30 day profile obtained 
from the same pixel and observation hour is presented. The set of empirically defined 
tests used in the production of the cloud masks is summarized bellow: 
 
.5.1278
/5.1278
278237
237
3030
3030
SkyClearBTBTandKBT
CloudSubpixelWarmBTBTandKBT
CloudColdKBTandKBT
CloudColdSuperKBT
nnii
nnii
ii
i
????>
???<>
??>
??
==
==
?
?
 
 
Where BT is evaluated for each individual pixel imaged during the i
th
 observation 
hour. Conditions (iii?iv) use information from the previous 30 days to calculate the 
mean ( BT ) and the standard deviation (?) for observation hour i. Only pixels having 
BT values greater than 278 K are considered in the calculation of BT and ?. The 
cloud mask processing also requires that a minimum of 10% of the BT data available 
during the 30 day sampling period be greater than 278 K. The threshold of 278 K was 
(i) 
(ii) 
(iii) 
(iv) 
 
 21 
 
found to work well for this specific data set as it provided the dividing line between 
unambiguous cloudy pixels and other surfaces. Having a minimum of 10% of all 
values observed within a 30 day period above 278 K was then required to estimate the 
approximate clear sky BT value and therefore allows for proper classification of the 
pixels. For those cases where this criterion is not met, the pixels are classified as 
undetermined and thereby not used in the cloud obscuration processing. 
Approximately 1.5% of the pixels analyzed were classified as undetermined. 
 
 
 
Figure 2.3. Brightness temperature (BT) image for 15 August 2004 1815 h UTC (A) and the 
30 day (July 16? August 15 2004) temporal profile (B) for the pixel marked as ??? on the 
image. Solid line represents the observed BT values, the dotted line represents the 30 day 
mean value, and the dashed lines the 30 day standard deviation. 
A 
 
 22 
 
 
Figure 2.3. continued 
 
 
The performance of the cloud mask product was tested against another source 
(CPTEC's GOES-based operational multi-band daytime-only cloud classification 
method; http://satelite.cptec.inpe.br) to check for consistency. Two 20 day periods 
were analyzed: one in the wet season (February 2005) and one in the dry season 
(September 2005). The results showed the two products to be in good agreement and 
proved the cloud mask to be effective in identifying cloudy pixels (>95% 
correspondence with CPTEC's classification during both wet and dry seasons). Pixels 
classified as cloud-free by the cloud mask product showed a less significant 
agreement with CPTEC's product (23% during the wet season and 56% during the dry 
season). To help explain the differences between the two products we randomly 
selected approximately 20 cases from the wet and dry seasons for which GOES 1 km 
B 
 
 23 
 
and MODIS 1 km and 500 m bands were visually inspected. The visual analysis 
suggested that sub-pixel cumulus clouds that are routinely observed in the Amazon 
basin as well as other sorts of contaminants (e.g., thin cirrus clouds, smoke plumes) 
are more frequently classified as clouds by the more conservative method used by 
CPTEC resulting in the differences found. 
Half hourly global BT files were obtained from the NASA's Goddard Space 
Flight Center Distributed Active Archive Center (DAAC) covering years 2004 and 
2005 (17,520 files per year) and the cloud mask algorithm was applied for Brazilian 
Amazonia spatial subset. Analysis of the BT data showed that pixels having 
confirmed measurable precipitation (based on precipitation observations derived from 
the same surface station network used to evaluate the hydro-estimator product above) 
had a mean value of 237 K (? = 22 K). This result corroborates the findings of Arkin 
[1979] who obtained a 0.75 correlation coefficient between surface precipitation data 
and three month-average satellite rainfall estimates for brightness temperature values 
lower than 235 K. Based on this finding, the cloud obscuration analysis was only 
carried for pixels showing BT values greater than 237 K during the 24 h period 
between the last satellite precipitation estimate available and the following one. 
 
2.3 METHODS 
In order to predict whether there are fires present under clouds, we need to determine 
the conditions under which fires are likely to occur. In the case of Brazilian 
Amazonia, both physical and social factors play important roles in defining where 
and when fires occur. The methodology presented here uses precipitation estimates to 
 
 24 
 
determine the physical conditions of the environment, along with previous active fire 
data to derive the spatial and temporal distributions of fires that are known to be 
influenced by specific land use patterns observed throughout the region. Given that 
the extent of Brazilian Amazonia is over 5 million km
2
, the conditions leading to 
vegetation fires can vary significantly across the region. The existence of vegetation 
cover types ranging from grasslands to closed canopy evergreen forests requires that 
the area be divided into smaller parts in order to represent local characteristics of fire 
occurrence. Consequently a 40?40 km grid covering the study area (3358 cells total) 
was used to stratify the entire basin into smaller subparts. The grid size selected was 
meant to preserve the small scale phenomena represented in the analyses while 
balancing the need for reasonable sample sizes required to produce the summary 
statistics described below. In addition, the selected grid size allowed for easier 
association with the nominal 4?4 km GOES pixel distribution. Analysis of the 
physical conditions and spatial and temporal dynamics was undertaken for each one 
of the 48 daily GOES observation hours (30 min interval). In the following sections 
we describe the approach used to determine potential fire omissions due to cloud. 
 
2.3.1 Fire Dynamics and Precipitation 
Precipitation will have a great impact on the conditions of the local environment, 
influencing the mechanisms that control vegetation moisture content and temperature 
and the latent and sensible heat fluxes that affect land surface and atmospheric 
humidity and temperature conditions [Betts et al., 2002; Bruno et al., 2006; Nepstad 
et al., 2002]. WF-ABBA active fire data from 2004?2005 were used in conjunction 
 
 25 
 
with the GOES precipitation estimates for the same period to derive the physical 
conditions that facilitate fire development. Precipitation amounts during the 30 day 
period preceding each fire detection were extracted for every observation hour and 
the information aggregated into the 40?40 km grid following: 
 
.,,,
30
1
2
1
1
1
,, ?
?
?
?
?
?
=
???
?=
?=
?=
?=
?=
?=
day
day
n
day
day
n
day
day
nzni
rrrRTK 
 
Where RT is a 30 element array containing the accumulated precipitation r during a 
30 day period preceding fire n [n = 1,?, N] detected during observation hour i [i = 
1,?, 48] and located in grid cell z [z = 1,?, 3358]. A third degree polynomial was 
adjusted to the mean values representing the group of data found in each one of the 
40?40 km grid cells for each observation hour using: 
 
[] [] []
()()(){}
() .
.]30[,30,,]2[,2,]1[,1
30
,,
2
,
1
3
3
2
210,
,,,
,
1
,,
1
,,
1
,,
,
xaxaxaaxf
RRRM
N
RT
N
RT
N
RT
R
zi
zizizi
zi
Nn
n
zni
Nn
n
zni
Nn
n
zni
zi
+++=
=
?
?
?
?
?
?
?
?
?
?
?
?
=
???
=
=
=
=
=
=
K
K
 
 
Where R  is the 30 element array with the mean accumulated precipitation for N fires 
detected during observation hour i at cell z, and M is the resulting array of data for 
which the coefficients in Eq. (4) are calculated for. Eq. (4) is the function defining the 
(1) 
(2) 
(3) 
(4) 
 
 26 
 
relationship between the number of days (x) preceding a fire and the respective mean 
accumulated precipitation for that period. Interestingly, M assumes a strong linear 
distribution when N grows larger (typically for N > 50) (Figure 2.4). The application 
of a third degree polynomial proved a good fit with most conditions observed. The 
resulting curve coefficients were stored in individual look up tables (LUT, one for 
each observation hour) and were to be used in the final code as a reference value 
describing the mean precipitation conditions preceding fires at each 40?40 km grid 
cell. 
 
 
Figure 2.4. Mean precipitation values calculated for a single 40?40 km cell (approx. 
location: 60.10? W 12.50? S) based on N=151 active fire detections observed during 2004?
2005. R
2
=0.9974. 
 
 
 
 27 
 
 The information provided by the polynomial curves was used to set the 
probability ranges that describe the potential occurrence of fires based on the physical 
condition of the environment as defined by the precipitation data. The criteria used to 
determine the probability of fire occurrence as a function of precipitation were the 
following: 
 
()
()
()
()
()
?
?
?
?
?
?
?
?
?
=
=
=
=
=
=
=
=
?=?
?
?
?
?
?
?
?
?
?
?
>
=??
30
1
30
1
,
30
1
,
30
1
30
1
,,
30
1
,
30
1
,
30
1
,
30
1
,
.
2
1
dxxf
RA
dxxP
dxxfRA
and
dxxfRA
PdxxfRA
zi
day
day
zi
R
day
day
zizi
zi
day
day
zi
Rzi
day
day
zi
 
 
Where RA is the actual precipitation data that is accumulated over the 30 day period 
for a particular pixel at observation hour i and grid cell z, ?dx is time integral (30 
days) calculated for the polynomial curve f(x), and P
R
 is the assigned probability of 
fire occurrence related to precipitation. According to Eqs. (5) and (6), RA must be less 
than two times the area defined by the LUT curve covering the 30 day period for a 
particular location and observation hour for the pixel to be considered for the cloud 
obscuration analysis. Additionally, a third test was implemented to avoid particular 
conditions when the actual precipitation is concentrated in the most recent period 
(day-1? day-7). This test is based on anecdotal evidence, which suggests that land 
owners in the region normally wait for approximately one week to burn following a 
(5) 
(6) 
 
 28 
 
rainfall episode. In this case, a more conservative approach is applied using the 
criterion: 
()
?
?
=
=
=??
7
1
30
1
,,
.0
day
day
Rzizi
PdxxfRA  
 
From the above test, for the pixel to be considered for the cloud obscuration 
processing, the actual precipitation recorded during the 7 day period preceding an 
observation cannot be greater than the 30 day mean precipitation associated with fires 
for that same location and observation hour. 
 
2.3.2 Fire Spatial and Temporal Dynamics 
Fire activity in tropical areas tends to follow very specific patterns as a function of 
land use and land cover change. In those areas, the systematic use of fires for land 
clearing and maintenance creates reasonably consistent spatial and temporal patterns 
in satellite active fire detection from year to year [Giglio et al., 2006]. To establish 
the mean annual fire activity in time and space over Brazilian Amazonia, we used 
active fire records derived from GOES during 2003?2004. The 40?40 km grid used 
to extract the fire and precipitation relationship was also applied in this analysis. 
In order to derive the temporal distribution of fires, the year was divided into 
26 14-day periods. We used 14 day period intervals to properly represent the seasonal 
curves of fire activity. Fire seasons were usually constrained to approximately 2?3 
months, thereby 14 day periods allowed for a good representation of the progress in 
fire activity throughout the season while keeping our samples for each cell to a 
(7) 
 
 29 
 
reasonable size. The probability distribution of fires detected during each one of the 
14-day periods in relation to the 2003?2004 totals was calculated for all 40?40 km 
grid cells according to: 
() .
2004
2003
?
?
=
z
t
z
Tz
fires
fires
tP  
 
Where P
T
 is the active fire probability distribution for grid cell z, the numerator 
represents the number of fires detected during a 14 day period t, and the denominator 
represents the total number of fires observed in grid cell z during the entire period 
analyzed (2003?2004). A LUT containing the results for each grid cell was produced 
for every observation hour to be used in the final code. To reduce the effects of the 
inter-annual variability in fire activity on the adjustment factors produced, we used 
information from the year for which the processing was implemented (in this case 
2005) to account for departs from the mean values described by the LUTs containing 
the P
T
 data described above. For this we used a metrics P
14
 describing the sum of fires 
detected during the preceding 14 day period for each 40 km grid cell that was 
systematically updated during processing. Therefore, any increase or decrease in fire 
activity for a particular grid cell would be considered in the calculation of the cloud 
obscuration adjustment produced by either increasing or decreasing the contribution 
from the term P
T
 in Eq. (8) relatively to P
14
. 
 In addition to the annual distribution of fires observed with the active fire data 
record from 2003?2004, evaluation of fire's return frequency was also conducted. 
Vegetation fires are usually associated with a high return frequency in tropical 
(8) 
 
 30 
 
regions [Cochrane and Schulze, 1999]. The analysis of GOES active fire data during 
the period of 2003?2005 including all observation hours showed that approximately 
66% of the areas under the influence of fires in one year also exhibited some activity 
in the following year (Table 2.1). However, the likelihood of observing a fire in a 
particular area during the same observation hour over sequential years becomes much 
less evident. Multiple factors were believed to contribute to the latter, including but 
not limited to changes in the hour of the day a fire is ignited, changes in the fuel load 
affecting fire intensity and duration therefore impacting detectability, changes in land 
use, etc. Modeling these factors and the interactions among them is difficult and was 
not attempted here. In order to account for this greater variability when analyzing 
observation hours individually a random function (?) was used with the cloud 
obscuration analysis to represent fire return probability at every hour. This random 
function was adjusted to uniformly represent the average ranges of return 
probabilities at the individual observation hour from Table 2.1. 
 
Table 2.1: Fire return rate based on GOES ABBA active fire product using a 4?4km grid 
covering Brazilian Amazonia. 
 
Years with Detection 1745UTC* All Hours Included* 
2003 & 2004 33% 68% 
2004 & 2005 40% 65% 
2003 & 2005 33% 66% 
2003 & 2004 & 2005 19% 52% 
* Values represent the percentage of all cells with fire activity in one year that also showed 
detections in subsequent years based on (i) the observation hour of 1745UTC and (ii) all 
observation hours. 
 
Complementing the analysis of the land use influence in the establishment of 
fire patterns across the study area, the spatial distribution of fires was determined with 
 
 31 
 
the implementation of a 4?4 km grid nested in the previous 40?40 km grid. This finer 
grid was created to represent the GOES pixel map at nominal resolution. It was 
intended to distinguish between areas under the influence of fires where the cloud 
obscuration evaluation would be performed from areas with no fire activity recorded 
in the most recent years where no intervention was to be conducted. All 4?4 km cells 
showing fire activity during years 2003 and 2004 were identified and the information 
stored as binary values (0 ? no fire history; 1 ? fire prone) in individual LUTs for 
every observation hour. These LUTs were updated during the actual processing to 
account for areas of fire expansion by adding new fire prone pixels to the existing list. 
 
2.3.3 Implementation and Accuracy Assessment 
The methodology implemented during this study was designed to use the information 
derived from the previous steps for compensating year 2005 GOES-East fire data for 
the cloud obscuration of fires. GOES data were processed sequentially for every 
acquisition hour during year 2005 totaling 17,520 observations. For a particular 
observation date and time the code would initially search for all 40?40 km grid cells 
that, according to the temporal LUTs, are likely to show fire activity for that 
particular bi-week (Figure 2.5). Following this first selection, all 4?4 km areas 
contained by the grid cells (i.e., the nominal GOES pixels) with a history of fire 
activity are identified. The cloud mask product that was produced for the same year 
(2005) is then used to search for cloudy pixels among the previously selected ones. 
Two major scenarios were considered during processing, each representing distinct 
cloud coverage conditions. 
 
 32 
 
 
Figure 2.5. Cloud obscuration processing diagram. 
 
 33 
 
In scenario one, a cloud covered pixel is selected for observation hour T and 
both adjacent detections (T?30 min and T+30 min) are also obscured by clouds. In 
this case, the actual precipitation data is assessed and compared against the values 
obtained from the LUT defining the mean precipitation curve associated with fire 
occurrence for that area (i.e., the 40?40 km cell). Having cleared the tests established 
in. (5), (6) and (7), the probability of fire omission (P
F
) will be derived for that pixel 
using (5), (6) and (8) and the random function (?), assuming the form: 
 
.???=
TRF
PPP  
 
In scenario two, a cloudy pixel detected in time T has either the preceding 
(T?30 min) or the succeeding (T+30 min) observation hour cloud free. Under these 
conditions the code will use information from the current WF-ABBA fire product 
(searching for adjacent detections), the pixel's local time and from a general fire 
diurnal cycle curve. For this particular application, the curve chosen to represent the 
fire diurnal cycle in the region had its peak located at 1500 h local time in order to 
coincide with the normal hour of maximum air temperature ? or inversely, minimum 
air humidity ? [da Rocha et al., 2004] as temperature and humidity will be strongly 
related to fire spread conditions. The diurnal cycle curve was divided into three parts, 
each describing a particular trend in fire progress over time (Figure 2.6).  
 
 
 34 
 
 
Figure 2.6. Stratification of the diurnal cycle of fire activity into three major areas, A1, A2, 
and A3. 
 
 
In area A1 of Figure 2.6, a cloudy pixel observed at local time T
local
 will be 
likely to omit a fire if an actual detection exists for T
local?30 min
, since fire continuation 
will be favored by the usual increase in temperature followed by a decrease in relative 
humidity. The likelihood of a fire being omitted at local time T
local
, when detection is 
observed during hour T
local+ 30 min
 will depend on the potential fire occurrence during 
hour T
local
 for that time of the year. Based on these considerations, cloudy pixels 
observed during local hour T
local
 and falling within area A1 will be evaluated 
according to the following criteria: 
 
Actual fire detection at T
local-30 min
 ? P
F
 = 1 
Actual fire detection at T
local+30 min 
? P
F
 = P
T
. 
 
In this case, if both T
local ? 30 min
 and T
local + 30 min
 observations are cloud free, 
condition (i) will prevail. For area A2 in Figure 2.6, a cloudy pixel observed during 
local hour T
local
 will have the trend in fire progress being inversely described 
according to the following criteria: 
(i) 
(ii) 
 
 35 
 
 
Actual fire detection at T
local-30 min 
? P
F
 = P
T
 
Actual fire detection at T
local+30 min 
? P
F
 = 1. 
 
In this case, condition (ii) will prevail when both T
local ?30 min
 and T
local + 30 min
 
observations are cloud free. Lastly, for area A3 in Figure 2.6 a cloudy pixel detected 
during local hour T
local
 will receive a nominal probability P
T
 if either T
local?30 min
 or 
T
local + 30 min
 observations are cloud free and a fire detection exists for any of those two 
observations. 
To verify the consistency of the outputs produced we tested the cloud 
obscuration processing described above using simulated cloud coverage data in 
combination with the actual fire detection data from WF-ABBA. We used year 2005 
cloud mask product to simulate the presence of clouds over randomly selected clear 
sky pixels found across the study area. The cloud obscuration analysis was then 
performed for those pixels selected in order to obtain the predicted number of fires 
missed due to clouds. For each pixel analyzed the actual fires detected by WF-ABBA 
were identified. The results were aggregated using the 40?40 km grid and the 
relationship between predicted fires versus observed fires assessed. The simulations 
demonstrated the predicted fires to be highly correlated with the observed fires (R
2
 
=0.82) (Figure 2.7). The sum of all fires produced by the cloud obscuration 
processing was able to represent the actual fires missed due to clouds to within 1.5% 
on an annual basis and to within 5% on a monthly basis. 
(i) 
(ii) 
 
 36 
 
 
Figure 2.7. Relationship between fires produced by the cloud obscuration processing based 
on simulated data and observed fires for the same locations using 2005 data. 
 
2.4 RESULTS AND DISCUSSION 
The main processing code was used to process year 2005 data and the results are 
presented below. The numbers produced represent those cases where a cloud 
obscured pixel was identified and at least one of the criteria established above was 
met for any of the 17,520 annual GOES observations over an area of 5.4 million km
2
 
(335,800 pixels at 4?4 km spatial resolution) covering Brazilian Amazonia. The 
probability of fire omission derived for each pixel was accumulated over multiple 
hours and days of observation resulting in a total number of fires potentially missed 
due to the presence of clouds equivalent to 59,650, or approximately 11% of the 
actual GOES fires detected during 2005 (545,286 total ? representing WF-ABBA's 
fire types 0?4 only; category 5 associated with low probability fires not considered) 
(Figure 2.8). 
 
 37 
 
 
 
Figure 2.8. Fires omitted due to clouds (A) for year 2005 and the percentage contribution (B) 
of cloud adjusted fires in relation to the total number of GOESWF-ABBA fires detected in 
that same year for each 40?40 km grid cell covering Brazilian Amazonia. 
 
A 
B 
 
 38 
 
The 2005 data set was also processed using the simple approach for 
compensating for clouds, which is based on the assumption that fires occur at the 
same frequency under clouds as they do in cloudless areas [Cardoso et al., 2003; 
Giglio et al., 2003b, 2006; Roberts et al., 2005]. In order to evaluate the impact of 
spatial sampling, the simple approach was implemented using two different grids of 
40?40 km and 120?120 km resolution each. The average fire frequency was 
calculated for the cloud free fraction of each grid cell based on: 
iz
iz
iF
S
D
F
,
,
,
=  
Where D
z
 is the number of actual fires detected within grid cell z during 
observation hour i, and S
z
 is the number of 4?4 km cloud free pixels in the same area. 
The number of cloud obscured pixels in each cell (i.e., 100?S
z,i
 and 900?S
z,i
 for the 
40?40 km and 120?120 km grids, respectively) was then multiplied by Eq. (10) in 
order to derive the number of fires potentially missed. For both 40 km and 120 km 
grid resolutions used, the simple approach produced a larger number of potentially 
missed fires due to cloud obscuration (178,968 and 241,804 ? 33% and 44% 
increment ? respectively) as compared to the proposed methodology described above 
(Figure 2.9). The difference was particularly important in areas where fires were 
unequally distributed in space, causing fire free pixels (e.g., pixels located in forested 
areas and distant from human activities) to be erroneously classified as fire omission 
areas by the simple rule approach. The use of LUTs describing the fire activity in 
space helped reduce this effect with the methodology proposed by only selecting fire 
prone pixels during the cloud obscuration processing. 
 
(10) 
 
 39 
 
 
Figure 2.9. GOES WF-ABBA fire diurnal cycle distribution for Brazilian Amazonia and the 
corresponding adjustment produced using the cloud obscuration processing methodology and 
the simple rule approach using 40 km and 120 km sampling areas. 
 
 
The majority of the fire omissions identified by the proposed methodology 
were associated with the conditions described by scenario 1, where all three 
observation hours T?30 min, T, and T+30 min are obscured by clouds. Average cloud 
coverage in the region varied between a minimum of 22% and a maximum of 78% 
depending on the observation hour, with the lowest values found in the areas 
surrounding the evergreen tropical forests in the southern and eastern parts of the 
basin (Figure 2.10). These areas are generally composed of Cerrado (savanna) 
vegetation and transition forests. The highest values occurred in the western part of 
the basin where the hydrological cycle is enhanced by the prevalence of closed 
canopy forests resulting in high evapotranspiration rates and by westerly winds that 
bring supplemental moisture from the eastern parts of the basin. From the remote 
sensing perspective, this spatial distribution of fires and cloud coverage is beneficial 
 
 40 
 
as it minimizes the cloud obscuration problem in areas where fire activity is most 
pronounced (e.g., the states of Mato Grosso, Tocantins, Maranh?o, and eastern Par?). 
Minimum cloud coverage was found to coincide with the early afternoon observation 
hours (Figure 2.10), which has major implications for fire detection as it approaches 
the hour of maximum fire activity in the region [Giglio, 2007; Kaufman et al., 1998a; 
Prins et al., 1998]. This latter finding corroborates the discussion presented in 
Schroeder et al. [2005] where inter-comparison analyses using multi-sensor fire data 
and visual confirmation of smoke plumes demonstrated the strong diurnal signature 
of fire activity in the region. 
 
 
 
 
Figure 2.10. Average maximum and minimum percentage cloud coverage (A and B 
respectively) and hours of maximum and minimum occurrence arranged in 3 hour bins (C 
and d respectively). 
A 
 
 41 
 
 
 
Figure 2.10. continued 
B 
C 
 
 42 
 
 
Figure 2.10. continued 
 
 
In order to evaluate the implications of the method implemented above we 
used high resolution (20 m) China?Brazil Earth Resources Satellite (CBERS-2) burn 
scar maps to delineate areas of fire activity relating those areas to the fire omission 
pixels produced. The burn scar maps were generated for the state of Acre in the 
western Brazilian Amazonia covering the 2005 fire season in the region (July?
September). Using supervised classification we produced a burnt area mosaic based 
on 9 CBERS images corresponding to two individual scene locations (i.e., path 180 
rows 111 and 112) covering from early to late  fire season in 2005 (i.e., July to 
October respectively). Detailed mapping of burn scars resulting from maintenance 
D 
 
 43 
 
(e.g., pasture and agricultural) and conversion (e.g., slash and burn) fires was 
produced for an area of roughly 22,000 km
2
. The quality of the burnt area mosaic was 
assessed using GPS points and digital photos collected from 40 h of over-flights from 
which all burnt area sites inspected could be visually confirmed [Foster Brown ? 
personal communication 2005]. Cloud obscuration in the region is known to 
significantly influence active fire products, usually limiting detection capacity 
irrespective of the sensor used [Brown et al., 2006]. The GOES WF-ABBA fire 
counts (excluding fire category 5) were spatially aggregated into nominal 4?4 km 
cells along with the burn scar areas and the fire increments produced independently 
by the method proposed above and the simple rule approach (using 40 km and 120 
km grid resolutions) and error matrices were derived (Table 2.2a?d). 
The original WF-ABBA active fires detected during the same period covered 
by the burn scar map showed an omission error of 58.3% and a commission error of 
6.4%; multiple detections were frequently observed for individual 4?4 km cells 
across the area. Despite the reasonably high omission error, 4110 of a total of 4133 
fire detections observed during the period analyzed had a spatially coincident burn 
scar (i.e., within ? 2 km of WF-ABBA's fire pixel coordinate); whereas 15 of 23 fire 
coordinates associated with false detection had large burn scars located in the 
immediate vicinity of the pixel. Nevertheless, the fire pixels detected accounted for 
76% of the total area burned measured for the selected plot; with the remaining 24% 
associated with omission showing significantly different characteristics of mean burnt 
area size (4 times smaller than the confirmed detections) therefore potentially falling 
outside the detection envelope of the GOES imager. 
 
 
 44 
 
Table 2.2: Error matrices for evaluating the performance of the original GOES WF-ABBA 
active fire data (a), the cloud processed data (b), and the simple rule approach using 40 km 
and 120 km area sampling (c, d respectively) based on 20 m resolution CBERS data (our 
?ground truth?) covering part of Acre state in western Brazilian Amazonia. 
 
 
 
 
The application of the cloud obscuration processing method showed a 
reduction in the omission error to 43.7% along with a minor increment of the 
commission error (8.8%). The pixels selected in this latter case accounted for 88% of 
the total area burned measured. The mean burned area size resembled the mean area 
size associated with the original WF-ABBA fire product to within 15%.  
The application of the simple rule approach showed a major reduction of the 
omission errors for both 40 km and 120 km grids used (10% and 0%, respectively) at 
the cost of substantial commission errors being produced (63% and 100%, 
 
 45 
 
respectively). The mean burned area size described by the simple rule approach was 
considerably smaller (45%) as compared to the value observed with the original WF-
ABBA data, with the burned area frequency histogram having shifted towards smaller 
burn scars. Despite the improvements observed primarily in terms of the reduction of 
omission errors, the additional fires created during processing of the cloud 
obscuration did not have a significant impact on the relationship between the 
accumulated fire counts obtained for an individual pixel area and the associated size 
of the burn scar measured for that same plot. That relationship remained unresolved 
for all four data sets analyzed (i.e., the original GOES WF-ABBA active fire data, the 
cloud adjusted numbers derived with the above methodology, and with the simple 
approach at 40 and 120 km), with weak correlation being observed in all cases. 
 
2.5 FINAL REMARKS 
Fire detection omission due to cloud obscuration is a major problem affecting remote 
sensing of active fires. Here we described an approach designed to address the cloud 
obscuration problem using high frequency geostationary observations. The 
methodology takes advantage of three different input data, namely precipitation 
estimates, a cloud mask and active fire data that are derived from the same 
instrument. This consideration was especially important for facilitating data 
registration and also to reduce image navigation problems that can occasionally affect 
the GOES data. The use of active fire data from recent years along with precipitation 
data provided means to establish the general patterns of fire use in both space and 
 
 46 
 
time across the entire study area, at the same time preserving specific regional and 
local characteristics with the implementation of pixel based processing. 
The strong correlation between the predicted fires missed and the actual fires 
observed demonstrated the effectiveness of the approach. In terms of the overall 
trends caused by cloud obscuration across Brazilian Amazonia, we observed that the 
net effect of fire omission was partially minimized in areas of intense fire activity as 
these generally coincided with the areas of minimum cloud coverage. Nonetheless, in 
relation to the percentage contribution of fires missed due to clouds, we found areas 
where fire omission was significant (Figure 2.8) despite those being depicted as 
relatively low fire activity areas by the original WF-ABBA product (Figure 2.1). This 
could signal the need to more detailed analysis of fire dynamics in areas considered to 
be of low priority under current regional fire management programs (e.g., 
northeastern Roraima). 
Another important aspect presented was the partial overlap between the hours 
of minimum cloud coverage and of maximum fire activity as a result of a strong 
basin-wide fire diurnal cycle signature. The cloud adjusted numbers maintained the 
same fire diurnal cycle signal confirming our field observations that indicated 
systematic use of fires in the mid-afternoon hours as part of regional land 
management techniques. As compared to a more simplistic approach, the 
methodology was proven successful in reducing the omission errors while 
maintaining the commission errors nearly unchanged, and preserving the general 
quality of the fires described by the original fire product from WF-ABBA. 
 
 47 
 
By means of routine updating of the LUTs used during processing, this cloud 
obscuration modeling technique should be capable of consistently mapping fire 
omission in tropical areas such as Brazilian Amazonia. While the number of fire 
pixels missed due to clouds could be successfully defined, additional work is required 
to better describe the relationship between active fires and the total area burnt which 
modeling of biomass emissions so much depend on. 
We consider this methodology to be applicable to other geostationary systems 
covering different regions of the globe provided that similar data layers are available 
for use. The proposed cloud correction scheme can be included in the current effort 
by the Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD) Fire 
Mapping and Monitoring Theme (URL:http://gofc-fire.umd.edu) to establish a global 
fire monitoring network from geostationary satellites. This activity, also being 
coordinated using principles of the Committee on Earth Observation Satellites 
(CEOS; URL: http://www.ceos.org) constellation concept and incorporated into the 
Coordination Group for Meteorological Satellites (CGMS; URL: 
http://www.wmo.ch/web/sat/CGMShome.html) is also a contributor to Group on 
Earth Observations (GEO; URL: http://www.earthobservations.org/index.html) 
efforts. However, we must warn for the fact that sensor dependencies (e.g., detection 
omission and commission rates) need to be resolved before any global analysis of the 
effects of clouds on remote sensing fire products is attempted. 
In principle, an improved scheme to correct for cloud obscuration effects is 
also needed for active fire detections from polar orbiting satellites. Consistent 
correction is needed to integrate all geostationary and polar orbiting data into a long-
 
 48 
 
term active fire data record as part of the Fire Disturbance Essential Climate Variable 
(ECV) as defined in the Global Climate Observing System Implementation plan 
(GCOS; URL: http://www.wmo.ch/web/gcos/gcoshome.html). However, due to their 
limited observation frequency, polar orbiting satellites would require a different 
approach from the one presented here for the estimation of fire omission due to 
clouds. For instance, scenario 2 described in Section 2.3.3 which uses the information 
from adjacent observations of GOES (30 min before and after observation time t) 
would not be applicable to a polar orbiting system. Other issues such as variable 
imaging geometry would also impact the cloud analysis using sensors such as 
MODIS and AVHRR. 
 
 49 
 
Chapter 3: Validation of GOES and MODIS Active Fire 
Detection Products Using ASTER and ETM+ Data
1
 
 
 
3.1 INTRODUCTION 
Brazilian Amazonia currently represents one of the most active regions of 
deforestation and biomass burning in the world [Csiszar et al., 2005; Dwyer et al., 
2000; Giglio et al., 2006]. Widespread use of fires for land clearing and management 
necessitates consistent methods for monitoring and mapping biomass burning on a 
routine basis [Korontzi et al., 2004; Schroeder et al., 2005]. 
The number of operational or near-operational fire products serving Amazonia 
has increased considerably in the past decade as a result of growing demand for fire 
information from the regional and global scientific communities, environmental 
enforcement agencies, and other end users. Information on fire activity for Amazonia 
is currently available from geostationary and polar orbiting satellites [CPTEC, 2008; 
Govaerts, et al., 2007; UW Madison CIMSS, 2008]. 
While some studies have assessed the performance of the polar orbiting and 
geostationary active fire products through validation exercises [Menzel and Prins, 
1996; Prins et al., 1998; Morisette et al., 2005a; Morisette et al., 2005b; Csiszar et al., 
2006], detailed characterization of these products on a regional to sub-continental 
                                                 
1
 The material presented in this Chapter is part of Schroeder et al. [2008b] 
 
 50 
 
scale is still needed. The increased demand for these products requires that an effort 
be made to fully characterize the quality of their data. 
In this study we investigate the performance of two major active fire detection 
products available for Brazilian Amazonia derived from the polar orbiting Moderate 
Resolution Imaging Spectroradiometer (MODIS), on board the EOS-AM (Terra) and 
EOS-PM (Aqua) satellites, and the imager on board the Geostationary Operational 
Environmental Satellite, positioned at 75?W longitude along the equator (GOES-
East). We use a large selection of scenes from the Advanced Spaceborne Thermal 
Emission and Reflection Radiometer (ASTER) and the Landsat 7 Enhanced Thematic 
Mapper Plus (ETM+) at 30 m resolution to serve as our ?ground truth? data to 
validate the coarser resolution products. The scene selection covers a wide range of 
vegetation fire conditions that are found across the study region. 
We rigorously test the accuracy of the MODIS and GOES active fire products 
to quantify sources of commission and omission errors, and suggest a method to 
enhance the current contextual fire detection algorithms using brightness temperature 
temporal profiles to reduce the commission error rate over tropical forest regions. We 
apply a bottom-up approach by associating the fire statistics derived from the 30 m 
resolution data to the moderate and coarse resolution data of MODIS and GOES, 
respectively. Complementary in-situ measurements are also used in this study along 
with data from an airborne imaging instrument to independently analyze the 
properties of the fires mapped by the orbital systems. 
 
 
 51 
 
3.2 DATA 
The data used in this study cover nine states that form the Legal Brazilian Amazon, 
an area characterized by a large gradient of percentage tree cover (see Figure 3.1). In 
order to represent all of this variation in the validation process of GOES and MODIS, 
a large volume of data was processed. The details of each individual data set are 
given below: 
 
3.2.1 GOES 
The study region is centered within the regular scanning zone of the GOES East 
imager which produces 4?4 km resolution images at 30 min intervals. Centered in the 
fire sensitive mid-infrared region of 3.8?4.0 ?m, channel 2 provides the primary 
radiometric measurement for GOES imager fire detection. In order to increase our 
sample size, we used a combination of versions 5.9 (2000?2002) and 6.0 (2002?
2005) of the Wildfire Automated Biomass Burning Algorithm (WF-ABBA) 
generated by the Cooperative Institute for Meteorological Satellite Studies (CIMSS) 
at the University of Wisconsin, Madison. The two products are very similar; changes 
in version 6.0 were mainly associated with the implementation of more stringent tests 
to eliminate potential spurious fires from highly reflective clouds [Prins and Menzel, 
1994; Prins et al., 2003]. 
We used the full resolution active fire masks provided by CIMSS as well as 
the regular ASCII data files commonly available to end users. GOES channel 2 
Variable Format (GVAR) data were also obtained from NOAA's Comprehensive 
 
 52 
 
Large Array data Stewardship System (NOAA-CLASS; URL: 
http://www.class.noaa.gov/saa/products/welcome). The GVAR data were used for the 
purpose of image registration as well as to derive brightness temperature estimates 
used in Section 3.4.4. A total of 119 individual images were used for the validation 
analysis covering the acquisition dates and hours of ASTER or ETM+ data. An 
additional 446 images were used in the analyses described in Sections 3.4.3?3.4.4. 
 
3.2.2 MODIS 
The two MODIS instruments on board the polar orbiting Terra and Aqua satellites 
provide daily images of Amazonia at nominal 1?1 km spatial resolution. We used the 
Collection 4 MODIS Level 2 (un-projected swath) fire product (MOD14) [Giglio et 
al., 2003a], available from NASA's Land Processes Distributed Active Archive 
Center (LP-DAAC) (URL:http://edcimswww.cr.usgs.gov/pub/imswelcome/), in 
conjunction with the MODIS geolocation product (MOD03), available from NASA's 
Level 1 and Atmosphere Archive and Distribution System (URL: 
http://ladsweb.nascom.nasa.gov/data/). A total of 135 data granules (5-minute orbit 
segments) from Terra MODIS were used for the validation analyses coinciding with 
the acquisition dates of our 2001?2005 ASTER scenes. An additional 164 MOD14 
data granules were used along with the corresponding Level 1B Calibrated Radiance 
granules (MOD021km) to support the analyses described in Sections 3.4.3?3.4.4. 
 
 
 53 
 
3.2.3 ASTER 
ASTER is an on-demand radiometer flying on board the Terra satellite and therefore 
it collects data simultaneously with Terra MODIS [Yamaguchi et al., 1998]. 
Although the instrument does not possess a mid-infrared channel, active fire detection 
was proven successful in previous studies by combining the top of the atmosphere 
(TOA) reflectance estimates from channel 3 near infrared data resized from 15 to 30 
m resolution and channel 8 short wave infrared data at 30 m resolution (Table 3.1) 
[Morisette et al., 2005b; Giglio et al., 2008].We used a total of 162 ASTER Level 1B 
Registered Radiance at the Sensor scenes obtained from the LP-DAAC (Figure 3.1). 
 
3.2.4 ETM+ 
The ETM+ instrument flies on board Landsat-7, which has a similar orbital path as 
the Terra satellite, albeit with an earlier daytime equator crossing time of 1000 
[Goward et al., 2001]. The spatial and spectral specifications of ETM+ channels 4 
and 7 are similar to those of ASTER channels 3 and 8, respectively, making ETM+ 
equally suitable for active fire mapping (see Appendix A). In this study we used 123 
scenes available through the National Institute for Space Research (INPE, Brazil) and 
the Global Land Cover Facility (GLCF; URL: 
http://glcf.umiacs.umd.edu/index.shtml). The scenes covered the period from 2000?
2003. 
 
 
 54 
 
 
Figure 3.1: Location of ASTER (red shaded areas) and ETM+ (red open boxes) scenes used 
to validate the MOD14 and WF-ABBA fire detection products, over percentage tree cover 
data [Hansen et al., 2002]. 
 
3.2.5 Vegetation Data 
The 500 m percentage tree cover data derived from the MODIS Vegetation 
Continuous Fields (VCF) [Hansen et al., 2002; Hansen et al., 2003] Collection 4 
product was used in this study to stratify the study region into vegetation sub-groups. 
A total of 12 MODIS 10??10? Sinusoidal Projection tiles were required to produce a 
mosaic covering the entire study region. The annual VCF product for 2000?2005 was 
available through GLCF. 
 
 
 55 
 
3.2.6 Ground Data 
In-situ fire temperature measurements were obtained for seven different burning 
locations. These measurements were derived as part of validation campaigns 
implemented during 2003?2004 when prescribed burns were arranged in different 
parts of the study region. We used a thermocouple linked to a data logger [Campbell 
Scientific Inc. CR21X] which recorded the development of the fire fronts at 0.2 Hz 
for the duration of the burning (i.e., from fire onset through the smoldering phase). 
These measurements were mainly intended to characterize the differences in fire 
intensity and duration among the main fire types typically observed in the study 
region. 
 
3.2.7 Airborne Data 
Airborne data were acquired for Roraima state in northern Brazilian Amazonia using 
the Airborne Hyperspectral Scanner [AHS AA5201, Argon ST] over two prescribed 
burns and over 90 randomly chosen fires during a 7-day field campaign in January 
2003. The AHS sensor consists of a 50 channel scanner covering 0.445?12.08 ?m. 
The instrument is owned and operated by the System for the Protection of Amazonia 
(SIPAM) and it was first flown over vegetation fires during the field campaign in 
January 2003. We applied supervised tests to the data in order to produce fire masks 
at ground resolutions of 1?1.65 m. The fire masks were used to delineate the contour 
of the active flaming areas in each image. 
 
 
 56 
 
 
Table 3.1: Main characteristics of the satellite imagery used. 
 
Instrument Primary Bands Used for 
Fire Detection 
Spatial 
Resolution 
Temporal 
Resolution 
Number of 
images  
GOES Imager Channel 2: 
3.80-4.00?m 
4km 30min 565 
MODIS Channels 21 & 22: 
3.929-3.989?m 
1km 1-2days 299 
ASTER Channel 8: 
2.295-2.365 ?m 
30m 16days 162 
ETM+ Channel 7: 
2.09-2.35 ?m 
30m 16days 123 
 
 
3.3 METHODS 
Three previous studies relied on ASTER data to validate the Terra MODIS active fire 
detection product. First, Morisette et al. [2005b] used 18 ASTER scenes from 
August?October 2001 to validate the MOD14 collection 3 and 4 products over 
Southern Africa. Next, Morisette et al. [2005a] used 22 ASTER scenes to 
simultaneously validate two different active fire detection products based on Terra 
MODIS data for three sites in Brazilian Amazonia. Lastly, Csiszar et al. [2006] used 
131 ASTER scenes to validate the MOD14 collection 4 product over Siberia. All 
three studies used active fire masks derived from ASTER imagery to validate the 
coincident Terra MODIS active fire product. 
In order to validate and characterize the MOD14 fire product over Brazilian 
Amazonia, we used a similar approach as described in the studies listed above. We 
produced active fire masks for all 162 ASTER scenes using a contextual approach 
based on ASTER channels 3 (0.76?0.86 ?m) and 8 (2.295?2.365 ?m) [Giglio et al., 
 
 57 
 
2008]. The resulting 30-m mask indicates the presence (?1?) or absence (?0?) of 
active fires within each moderate-to-coarse resolution pixel, and 30-m pixel counts 
are used as surrogates for actively burning area. We recognize this is still an 
intermediate step towards deriving validation statistics based purely on ?true? 
physical quantities defining fires. 
We used information from the Terra MODIS geolocation product (MOD03) to 
overlay the MOD14 product on top of the ASTER fire masks. From the fire mask 
data we derived the sum of 30-m active fire pixels and the number of individual fire 
clusters within each MODIS pixel. These statistics were used to determine the 
detection performance of the MODIS instrument by means of a statistical logistic 
regression model defined as: 
 
() )1(
1
1
1
?
+
?
=
=
+
=
+
n
j
ij
x
j
n
j
ijj
e
e
xP
x
i
??
??
 
 
Here P(x
i
) is the probability of detection (0-1) of MODIS pixel i based on the 
linear combination of ASTER fire pixel summary statistics (j = 1, n), and ? and ?
j
 are 
parameters derived from the data [Agresti, 1990; Morisette et al., 2005b]. Error 
matrix analysis was also applied using information on the number of 30-m active fire 
pixels to produce empirical curves defining the omission error estimates for the 
MODIS fire product. 
 
 58 
 
In order to account for the effects of the MODIS triangular point spread 
function [Wolfe et al., 2002], the projected surface area for the MODIS pixel was 
considered equal to 2?1 km at the sub-satellite point. This procedure is applied for all 
MODIS pixels overlapping the ASTER image.  
We tested the consistency of the fire masks using 1.5 m resolution data from 
the AHS instrument over two prescribed burns conducted in January 2003 which 
were coincident with ASTER imaging (Figure 3.2). Additionally, hundreds of image 
quick looks were produced to verify the consistency of the MOD14 detections when 
overlaid on top of the ASTER fire masks and false color composites.  
 
 
Figure 3.2: Prescribed burn of a deforestation plot located at 1?35?08? N 60?57?18?W on 
January 28, 2003. (A) 1.5m resolution data from AHS RGB combination of bands 9 (0.685 
?m), 5 (0.565 ?m) and 1(0.445 ?m) acquired at 1438 UTC; (B) fire mask derived from AHS 
band 43 (4.95 ?m); (C) ASTER image of the same fire (RGB bands 8-3-1) acquired at 1436 
UTC. The fire perimeter (white contour) derived from the ASTER fire mask is also shown in 
(C). 
A 
 
 59 
 
 
 
 
 
 
Figure 3.2 continued 
 
B 
C 
 
 60 
 
Our validation analyses of the GOES active fire detection product followed a 
similar approach to the one described above for the MODIS data. However, three 
major issues necessitated that a slightly different processing scheme be used for the 
GOES data: 
(i) Temporal differences in image acquisition needed to be considered 
in order to avoid the effects of short-term changes in fire 
characteristics. As part of our field campaign in January 2003, 
multi-temporal data from the AHS instrument were acquired during 
a prescribed burn of a 75 ha deforestation plot in Roraima state. 
Approximately 1.5 hours of fire activity were monitored by the 
AHS instrument at 1.5 m resolution. Once the fire lines were well 
established and moving freely, four sequential AHS images were 
produced separated by 5 minute intervals from each other. We 
found the instantaneous fire-affected areas to be within 10% of the 
mean estimated area (35 ha) mapped during the 20 minute interval. 
In a separate study, Csiszar and Schroeder [2007] used ETM+ and 
ASTER images acquired 30 minutes apart to assess the impact of 
varying fire conditions on the performance of the MODIS active 
fire product. Their results suggested that short term changes in 
vegetation fire conditions have only a minor impact on moderate 
resolution active fire products. In this study we adopted a 
conservative approach limiting the use of the 119 GOES images 
 
 61 
 
selected for the validation analysis to within 15 minutes of the 
acquisition of the available higher resolution data. 
(ii) Navigation errors are known to affect the GOES images on a 
regular basis, and image geolocation problems can occur even 
between 30 minute acquisitions [Menzel and Purdom, 1994]. To 
eliminate the problem, all GOES images used were manually co-
registered to within 1 pixel. 
(iii) Due to the coarse resolution of the GOES imager, fewer pixels are 
sampled compared to MODIS for the same area covered by 
ASTER. Consequently, a larger number of ASTER scenes must be 
processed in order to create an equivalent sample size. 
Alternatively, one can use ETM+ data to increase the area covered 
by each scene processed (see Appendix A). Therefore our 
validation of the WF-ABBA product used ASTER and ETM+ fire 
mask data to produce 30 m active fire summary statistics for all co-
located GOES pixels within a 15 min window between acquisition 
times.  
For projecting the GOES pixel area on top of the higher resolution data we 
used a full nominal resolution pixel size in the along and across scan directions (i.e., 
4?4 km in size near nadir). We considered this area estimate to be a good balance 
between the pixel size resulting from the over-sampled image matrix and the pixel 
dimensions if accounting for the GOES point spread function.  
 
 
 62 
 
3.4 RESULTS AND DISCUSSION 
Two major types of fires are normally observed across our study region, namely 
conversion and maintenance fires. Conversion fires will normally be associated with 
deforestation activities occurring along the new frontiers of land development 
dominated by high percentage tree cover (i.e., > 60%). As they burn large biomass 
volumes, these fires will normally be characterized by high flame temperatures 
followed by a relatively hot smoldering stage which can last from a few hours to a 
few days (Figure 3.3). Maintenance fires are normally used to rejuvenate pastures or 
to clear crop residue and are particularly important in areas dominated by low 
percentage tree cover (i.e., < 20%). As such, these fires are characterized by relatively 
low flame temperatures followed by a shorter smoldering stage ranging from a few 
minutes to a few hours. Compared to conversion fires, maintenance fires show a 
narrower flaming front as well as a shorter smoldering stage (Figure 3.3). Between 
these two major classes we can also find intermediate fire types (e.g., fires resulting 
from the clearing of secondary forests and woody Cerrado (savannas)) with reduced 
biomass load compared to old growth or pristine tropical forests. These fires can be 
more variable in terms of their flame temperature and burning duration. Data in 
Figure 3.3 suggest that the intensity and duration of conversion fires may facilitate 
their detection as compared to other types of tropical vegetation fires. 
 
 
 
 63 
 
 
Figure 3.3: Typical temperature profiles associated with fires used for: forest conversion, 
burning of piled debris from deforestation, rejuvenating grassland, and clearing of secondary 
forest re-growth. 
 
 
3.4.1 Overall Detection Performance 
The processing of all pairs of MODIS-ASTER images resulted in approximately 
7,300 MODIS pixels containing at least one ASTER 30 m active fire pixel. A total of 
1,640 MOD14 pixels were flagged as ?fire? for that data set. The pairs of GOES-
ASTER images resulted in approximately 2,900 GOES pixels containing at least one 
ASTER active fire pixel; the pairs of GOES-ETM+ images covered another 14,500 
GOES pixels with at least one ETM+ 30 m active fire pixel. This entire dataset 
included a total of 560 WF-ABBA fire detections.  
Summary statistics were produced from the higher resolution fire masks for 
all MODIS and GOES pixels sampled. These statistics were used to fit the logistic 
 
 64 
 
regression model in (1), from which the general detection probability curves for 
MOD14 and WF-ABBA were derived. The differences in detection performance 
between the MOD14 and WF-ABBA fire products are shown in Figures 3.4A and 
3.4B, respectively. However, the distances that separate the curves and the associated 
performance of the two products are smaller than expected. For the MODIS-ASTER 
configuration used in this study, all MODIS pixels sampled were imaged close to 
nadir resulting in pixels with 1?1 km nominal spatial resolution. The pairs of GOES-
ASTER and GOES-ETM+ images used were acquired under slightly variable 
conditions in terms of the pixel size sampled, ranging from near nadir geometry to 
cases where the additional pixel area enlargement was equivalent to about 20%. 
Consequently, the spatial resolution ratio of MODIS to GOES was always greater 
than or equal to 16?, which should result in an equivalent difference between their 
detection performances. However, an approximate ratio equal to or less than 4? 
separates the pairs of detection curves for each tree-cover interval described by the 
MOD14 and WF-ABBA products in Figures 3.4A and 3.4B. 
 
 
 65 
 
 
 
Figure 3.4: Fire detection probability curves for MOD14 (A) and WF-ABBA (B) fire 
detection products derived for four percentage tree cover intervals. 
 
 
In the case of Figure 3.4, a one-to-one comparison of the two instruments is 
difficult even for a small case study analysis as differences between products will 
A 
B 
 
 66 
 
likely be a result of: (i) varying sensor-target-sun relative positions; (ii) atmospheric 
attenuation, including fire obscuration by heavy smoke released during burning; (iii) 
relative position of the actual pixel area in relation to the location of the active fire 
zone and the effects of the point spread function, (iv) timing of acquisition, etc. Since 
the actual high-resolution validation data sets for MODIS and GOES are different, 
some difference in fire types may add uncertainty despite the rigorous sample used 
for this analysis. 
Our field data indicated that conversion and maintenance fires can show 
temperature variations on the order of hundreds of Kelvins even over very small 
areas. Consequently, their peak in emittance is likely to spread over a broader spectral 
interval than most land surfaces (e.g., bare soils and vegetated areas). With the GOES 
imager having a broader mid-infrared channel compared to MODIS (Table 3.1), it is 
plausible that the former will sense the radiative signal from a wider range of fire 
temperatures. This increase in the fire signal compared to the background could 
produce a higher relative capacity of GOES to detect them. 
 
3.4.2 Omission Errors 
Omission estimates were empirically derived for WF-ABBA and MOD14 by 
assessing the number of 30 m active fire pixels corresponding to true fire detections 
using an error matrix analysis. We counted the number of WF-ABBA or MOD14 
detections for each subset of GOES and MODIS pixels, respectively, showing a 
number of ASTER or ETM+ fire pixels greater or equal to N. By varying N between 
 
 67 
 
1 and the maximum value in our data set (~1,200), we produced the omission error 
estimates for the two products (Figure 3.5).  
 
 
Figure 3.5: Omission error estimates produced for MODIS and GOES based on fires 
sampled at approximately 1030 local time. 
 
 
Two areas of relatively high omission error are found for WF-ABBA at high 
counts of 30-m active fire pixels, one around 770 counts and a second one near 350 
counts. Visual inspection of the areas affected by such large fires indicated that the 
vast majority of the events were indeed associated with long fire lines occurring over 
areas of savanna-like vegetation with associated percentage tree cover typically less 
than 20%. In this case, the fire characteristics are analogous to the maintenance fire 
 
 68 
 
depicted in Figure 3.3. Consequently, the total radiant energy released is small 
compared to other types of fires rendering the detection of such events less likely. In 
addition, low thermal contrast with the warmer background was also found to 
contribute to the reduction in detection performance of WF-ABBA. In fact, we can 
distinguish two areas in the MOD14 curve where comparable high count omission 
errors are found, one near 170 counts and the second one near 245. As with WF-
ABBA, the areas of high count omission in the MOD14 data were also found to be 
associated with fires occurring over low percentage tree cover areas. 
Another factor potentially leading to high count omission errors in the GOES 
data is related to a nuance of the WF-ABBA detection algorithm. The detection code 
includes a persistence check analysis, which takes advantage of the high observation 
frequency of the GOES imager. The test checks each candidate fire pixel for the 
presence of any coincident detection in the previous 12 hours, allowing some 
tolerance in the spatial search criterion for small navigation drifts. In doing so, 
potential spurious detections can be filtered out of the product, therefore reducing 
commission errors. However, some new or short lived fires may also be removed 
from the final product. In order to quantify the effects of the temporal filtering, we 
also plotted in Figure 3.5 the corresponding omission error curve derived from the 
GOES data prior to the application of the persistency test. The result shows a 
significant increase in the total number of true detections produced (from 458 to 600) 
and a reduction of the omission errors especially towards the very high counts on the 
graph. 
 
 69 
 
Although useful for eliminating spurious fires, the application of temporal 
filters could represent a problem if one considers the instantaneous impact on the 
omission of relatively large fires that were mistakenly discarded from the final 
product. However, if such fires persist and maintain or increase their size and 
intensity in the subsequent GOES observations hours, a second detection will likely 
occur over time, causing the pixel to pass the persistency test as the previous 
observations would serve to confirm it. 
As suggested by previous studies [Giglio 2007; Menzel and Prins, 1996; Prins 
and Menzel, 1992, 1994; Prins et al., 1998], a strong fire diurnal cycle exists for 
regions such as Brazilian Amazonia. Under this circumstance, it is expected that fires 
observed during the mid to late-morning hours imaged by the ASTER and ETM+ 
instruments will likely grow in size and intensity towards the peak of fire activity 
located in the early to mid-afternoon hours. To assess the potential effects of 
changing fires conditions, we derived the overall omission error estimate for the WF-
ABBA product data using all GOES acquisition hours following the time of 
observation of the fire event by the higher resolution imagery until about 20:00h local 
time. The resulting omission error of 38% indicates that a large fraction of the fires 
observed in the morning hours which have persisted over subsequent observation 
hours might have intensified or increased their areas therefore entering the detection 
envelope of the GOES imager WF-ABBA fire product. 
 
 
 70 
 
3.4.3 Commission Errors 
False detections were characterized in our analysis as those showing zero 30-m active 
fire pixels within their footprints. We identified a total of 245 false detections among 
the 1,640 detections produced by the MOD14 product, whereas 102 false detections 
were identified for WF-ABBA based on a total of 560 detections sampled by that 
product. Contrasting surface temperature between the target pixel and its background 
dominated the false detection occurrences of MOD14 and WF-ABBA representing 
99% and 100% of the cases, respectively. Detection errors caused by the presence of 
clouds, sun glint zones or anthropogenic sources could not be identified in the data 
set. 
The areas of high thermal contrast were particularly pronounced along the 
deforestation expansion regions throughout Brazilian Amazonia where relatively cool 
evergreen tropical forests are replaced by bare soil and senescent grasses -- surfaces 
with increased surface temperature. Despite the differences in instrument 
characteristics, spatial resolution and algorithm used, both products agree to a large 
extent in the commission errors produced for all four percentage tree cover intervals 
considered (Figure 3.6). Errors increase rapidly as a function of the percentage tree 
cover indicating a potentially larger thermal contrast between the target pixel - often 
characterized by little or no vegetation - and the gradually greener and cooler 
background. 
 
 
 71 
 
 
 
Figure 3.6: Commission error estimates for MOD14 and WF-ABBA. Mean tree cover values 
(A) estimated for an area of approximately 20?20 km centered on the commission error pixel; 
standard deviation values (B) determined by sampling a 9?9 pixels window centered on the 
commission error pixel. 
 
A 
B 
 
 72 
 
 Detailed inspection of each individual false detection occurrence using all 
available bands of ASTER and ETM+ at their highest spatial resolution enabled us to 
refine the classification of those areas and separate them into four different classes 
(Table 3.2). The data indicated that a large fraction of the false detections were indeed 
associated with recent burning activity as confirmed by the presence of fairly 
homogeneous areas of dark char covering part or all of the MODIS and GOES pixel 
footprints. Areas of smoldering were also confirmed over a smaller number of cases 
via the identification of smoke traces or the indication of warm pixels in the short 
wave infrared channels 8 and 7 of ASTER and ETM+, respectively. When summed, 
these false detections over areas of fire-related activity corresponded to 
approximately 87% and 81% of the commission errors of MOD14 and WF-ABBA, 
respectively. The remaining false detections were unrelated to fire and mainly the 
result of warm and reflective bare soils surrounded by contrasting green vegetation. 
These false detections corresponded to approximately 3% of all detections made by 
the MOD14 and the WF-ABBA products. This later result agrees with the findings of 
Csiszar et al. [2006] for the MOD14 product over Northern Eurasia. 
 
Table 3.2: Classification of MOD14 and WF-ABBA commission errors based on visual 
inspection of 30 m RGB composites of ASTER and ETM+ data. 
 
Product Recent Burning 
(scars visible) 
Bare Soil Smoldering Spurious 
detection 
MOD14 79% 12% 8% 1% 
WF-ABBA 75% 19% 6% - 
 
 
 
 73 
 
As most of the false detections were associated with areas of recent burning, it 
is possible that these locations show a strong contrast with their backgrounds for as 
long as the dark char and ashes remain evenly spread across the burning site. This 
could lead to repeated false detections affecting those locations over multiple days or 
even weeks or until the signs of burning are removed from the surface. To investigate 
this possibility, we checked every false detection location sampled by our data set for 
the presence of equivalent detections in the previous days. We restricted our analysis 
to the 30 day period occurring before the observation of the false detection based on 
the 30 m resolution imagery. The use of a 30 day sampling period was based on the 
assumption that char and ashes can be removed by wind and rain or by land use (e.g., 
plowing, seeding, and irrigation). Using spatially coincident 30 m resolution data 
acquired 16, 32 and 48 days apart we could confirm that scars remained visible for as 
long as 32 days. For periods greater than that, the burning imprint on the landscape 
becomes less evident.  
We searched for all fire detections from Terra (MOD14) and Aqua (MYD14) 
data co-located to within 1 km of the original false detection. Aqua data was used to 
improve observation frequency to identify the very first detection made in the 30 day 
period along with all subsequent ones. Lacking other forms of confirmation, we 
assigned the actual fire occurrence date to the first detection found for each location 
and labeled all subsequent detections as potential false detections (Figure 3.7). 
Approximately 60% of the false detections in our data set had co-located detections 
from MOD14 or MYD14 occurring in the previous 30 days. Of those, about 50% had 
 
 74 
 
the first detection occurring in the previous 14 days. The average number of 
subsequent false detections was 4.  
 
 
 
Figure 3.7: Daily number of hot spot detections from Terra MODIS (MOD14) and Aqua 
MODIS (MYD14) data coincident with 60 different locations where MOD14 false detections 
were identified via 30 m ASTER imagery. The shaded area describes the accumulated 
percentage of all 60 locations for which a first detection was assigned during the 30-day 
period before the confirmed false detection. 
 
 
The locations mapped with the analysis above showed that the Terra 
(MOD14) product produced almost twice as many false detections as Aqua (MYD14) 
(117 vs. 67). The two products are processed using identical routines, reducing or 
even eliminating the chance of algorithm artifacts affecting our results. To compare 
these MODIS results to the GOES data, we analyzed all WF-ABBA false detections 
 
 75 
 
found in our validation data set by monitoring their progress throughout the day. We 
searched for co-located WF-ABBA detections occurring during the GOES 
observation hours made after the false detection was observed. In this case, we 
applied a search radius of 0.1? (or the equivalent to about 10 km) to allow some 
flexibility due to potential navigation drift affecting some of the unregistered data 
used. The number of persisting detections was mapped along with the occurrence of 
clouds, sun glint areas and gaps in image coverage due to rapid scan operations which 
could prevent proper observation of the target pixel. The number of false detections 
persisting in the data gradually decreased with the hours of the day (Figure 3.8). At 
about the same overpass time as Aqua (i.e., 1715-1845 UTC) the number of false 
detections was reduced by approximately 50%. Our results for the commission errors 
of GOES (WF-ABBA) were similar to the results produced for MODIS (MOD14 and 
MYD14) suggesting a good agreement between the two fire products. Although the 
WF-ABBA detection algorithm includes an adjustment to one of the primary tests 
used (i.e., the brightness temperature (BT) test on channel 2) based on the solar zenith 
angle measured for every pixel and observation hour (BT
2
 > cosine (solar zenith 
angle) ? 15 K + 285 K), the resulting correction factors for the approximate overpass 
times of Terra and Aqua were found to be similar for the study region therefore 
reducing their influence on the result above. 
 
 
 76 
 
 
Figure 3.8: Half hourly distribution of WF-ABBA false detections identified in the validation 
data set. The absolute number of confirmed false detections identified via ASTER or ETM+ 
is shown on top of the vertical bars. 
 
 Two major factors can contribute to the reduction in commission errors 
towards the afternoon hours. First, cloud coverage in tropical areas such as Amazonia 
may have a pronounced effect limiting the imaging of the land surface by passive 
remote sensing instruments. The impact of cloud obscuration on the detection of 
active fires in Amazonia has already been documented and it is suggested to account 
for an omission rate of approximately 11% of all fires imaged by coarse resolution 
sensors [Schroeder et al., 2008a]. In the case of the commission errors reported above 
however, areas of cloud shadow and semitransparent clouds which may not interfere 
on the detection of active fires will affect the thermal properties of the surface thereby 
potentially enhancing their impact beyond the 11% estimate produced for the active 
fire data. 
 
 77 
 
 Changes in the thermal contrast between adjacent areas with different 
vegetation cover conditions could also influence the commission errors. To test this 
assumption, we randomly selected 20 areas of low vegetation cover surrounded by 
forests to evaluate the changes in thermal contrast measured between the two surfaces 
throughout the day. We used two sets of GOES imagery acquired on 19 September 
2001 and 4 August 2002 covering the entire day at 30 min observation frequency. 
The two sets of images were characterized by low cloud coverage over the areas of 
major deforestation activity in Brazilian Amazonia favoring the extraction of cloud 
free brightness temperature profiles for the selected areas.  
Overall, the results showed a decrease in the thermal contrast between 
forested and non-forested areas measured for the approximate overpass time of Aqua 
compared to Terra, with the thermal contrast for the latter being 1 K to 10 K greater 
than the former. Figure 3.9 shows the profiles obtained for two adjacent areas of 
forest and non-forest surfaces where false detections were produced by the WF-
ABBA product. The difference between the two curves is included along with the 
approximate overpass times for Terra and Aqua. For the particular location used in 
Figure 3.9, clear sky conditions prevailed during the majority of the time. We can see 
in Figure 3.9 that the increased thermal inertia of the forest causes the peak in contrast 
with the deforested area to move closer to the observation hour of the Terra satellite. 
The false detections are found equally distributed around the area of greater thermal 
contrast and therefore tend to become more evident near the observation hours close 
to the Terra overpass and less so near the observation hour of Aqua. In this case, the 
reduction in thermal contrast for the two surfaces between the Terra and Aqua 
 
 78 
 
overpasses was small (~ 2 K) and false detections could also be observed until about 
15 minutes after the Aqua overpass time. In other situations where the change in 
surface conditions between Terra and Aqua overpasses is greater, the peak in thermal 
contrast approaches the Terra overpass time leading to fewer false detects during the 
observation hours near and after the Aqua overpass time. 
In many cases the effects of clouds and changes in thermal contrast were seen 
to manifest concomitantly making it difficult to separate and properly quantify the 
exact contribution of each term to the reduction of the commission errors observed. 
 
 
Figure 3.9: Diurnal cycle of brightness temperature derived from channel 2 of the GOES 
imager for two adjacent areas characterized by contrasting vegetation cover conditions. The 
images were acquired on 04 August 2002 and the forested and deforested areas were located 
at 11?59?29? S 52?49?52? W and 12?06?20 S 52?49?09? W, respectively. The two curves 
describe the mean values based on 3 and 4 pixels representing the deforested and forest areas, 
respectively. The occurrence of WF-ABBA false detections are presented along with the 
difference between the two curves. 
 
 
 79 
 
3.4.4 Reducing Commission Errors 
From the results presented above it becomes clear that one of the major complications 
involved with the application of contextual fire detection methods over tropical forest 
regions is the effect of contrasting warm and bright surfaces generating false 
detections. The parameters available with the MOD14 product Scientific Data show 
significant overlap between true and false detections (Figure 3.10), which eliminates 
the potential for successful post-processing of false detections using any of the 
available fire pixel information.  
 
 
Figure 3.10: Boxplots for channel 21 brightness temperature differences between the target 
pixel and its background for true and false MOD14 detections. Parameters depicted in the 
plot include: minimum, 25
th
 percentile (lower quartile), mean, 75
th
 percentile (upper quartile) 
and maximum value. 
 
 
 80 
 
Another approach to resolve the commission errors affecting the two data sets 
is the use of temporal metrics to complement the spatial analysis provided by the 
contextual method. Change detection methods based on the application of temporal 
metrics to assess sudden variations in the pixel signature of moderate and coarse 
resolution sensors are growing in importance as satellite data sets with improved 
navigation quality become available [Calle et al., 2006; Koltunov and Ustin, 2007; 
Schaaf et al., 2002].  
We selected over 100 detections from each of our MODIS and GOES 
validation data sets including true and false detections. Brightness temperatures (BT) 
were derived from the mid-infrared channels 21 and 22 of MODIS and channel 2 of 
GOES using MOD02 and GVAR Man computer Interactive Data Access System 
(McIDAS) AREA file data, respectively. The BT data were used to construct 30-day 
profiles from which the mean and standard deviation were extracted for each pixel. 
Figures 3.11A and 3.11B show a plot of the BT temporal profiles for two pixels fixed 
in space representing different cases of true and false detections in the MODIS and 
GOES data, respectively. The plots describe two important aspects of the data. First, 
MODIS shows a greater variability of the BT values compared to GOES, likely a 
function of variable imaging geometry. Despite its lower navigation quality, the 
GOES brightness temperatures show less variation over time. Second, since MODIS 
pixels are smaller compared to GOES, fires naturally occupy larger fractions of them, 
resulting in a greater departure from the mean value observed for the 30-day period 
analyzed. As a result, the potential limitations caused by a variable pixel size could be 
balanced by a high sensitivity to the presence of active fires within the pixel footprint, 
 
 81 
 
rendering this kind of application useful to separate false detections from true 
detections. On average, true detections departed by more than 4 standard deviations 
from the 30 day mean BT values, whereas for false detections the departure was 
limited to about 2 standard deviations. These results were equally valid for MODIS 
and GOES.  
 
 
Figure 3.11: Brightness temperature (BT) profiles for true (filled circle) and false (filled 
square) detections using MODIS (A) and GOES (B) data. Open circles and squares represent 
the BT values during detection-free days. Detection information for the pixels marked in (C-
F): false MOD14 detection (C): 55.1237
o
 W 12.129
o
 S, 23 May 2002 at 1403 UTC; true 
MOD14 detection (D): 56.4828
o
 W 13.225
o
 S, 05 October 2002 at 1409 UTC; false WF-
ABBA detection (E): 50.879
o
 W 8.4169
o
 S, 29 August 2002 at 1315 UTC; true WF-ABBA 
detection (F): 50.711
o
 W 10.0335
o
 S, 29 August 2002 at 1315 UTC. False color composites 
used: 30 m ASTER image (C-D); 30 m ETM+ image (E-F). 
A 
 
 82 
 
 
 
Figure 3.11. continued 
B 
C 
 
 83 
 
 
 
Figure 3.11. continued 
D 
E 
 
 84 
 
 
 
Figure 3.11. continued 
 
 
 The graphs in Figure 3.11 did not receive any rigorous treatment to reduce 
spurious oscillations in the derived BT values. Appropriate techniques could, 
however, be used for instance to model the daily variation of the MODIS BT data. In 
addition, the BT profiles in Figure 3.11 benefit from the favorable conditions 
observed in Amazonia where stable surface temperature conditions prevail 
throughout most of the year [da Rocha et al., 2004]. Nevertheless, the commission 
errors described above appear to be mainly associated with land use regimes marked 
F 
 
 85 
 
by large deforestation rates typical of the tropical forest regions. In these regions, 
similar surface conditions should be expected. 
 
3.5 CONCLUSIONS 
The MODIS and GOES fire products assessed in the validation analysis described 
above utilize conceptually similar algorithms based on the use of contextual tests to 
detect pixels with active fires. However, these algorithms are applied to images of 
markedly different spatial and spectral resolutions. The resulting detection 
performance estimates obtained through the validation analysis indicated that an 
approximate factor of 4 separates the detection curves of MOD14 and WF-ABBA 
products, a relatively small difference compared to the 16? factor between their 
nominal spatial resolutions.  
Although MODIS and GOES showed relatively high overall omission errors, 
their interpretation should be oriented by the type of application pursued. In this case, 
fire data users interested in larger biomass burning events may find both MODIS and 
GOES to have reasonably high rates of successful detection as is evidenced by the 
sharp decrease in omission errors especially near the low end of fire clusters 
described in Figure 3.5. 
The background conditions associated with active deforestation regions in 
Amazonia had a major effect on the false detections produced with both products. 
False detections can span multiple days provided that the burning scars remain 
evident and similar thermal conditions are preserved between adjacent areas. The 
 
 86 
 
reduction in commission error rates observed towards the afternoon hours has 
important consequences for the derivation of the fire diurnal cycle using contextual 
algorithms. In this case, an enhanced contrast between morning and afternoon hours 
should be noticed if all potential false detections are removed from the data. 
Removal of false detections based on the metrics which are routinely 
produced by the contextual algorithms proved difficult. Alternatively, we successfully 
tested the application of temporal metrics of brightness temperature to isolate false 
detections in our data set. The use of a hybrid fire detection algorithm which 
implements change detection methods in addition to the contextual tests appears 
promising especially over tropical regions. In these areas, stable surface thermal 
conditions might facilitate the extraction of temporal metrics which could be used to 
separate false detections or increase our confidence about the true ones. However, we 
recognize that this is a topic for further investigation necessitating careful 
consideration of all aspects involved. 
The results described above are valid for open sky fires typical of forest 
conversion and agricultural maintenance in Amazonia, but do not apply to cases of 
understory burning. Although a few sub-canopy fires could be visually confirmed in 
our validation data, proper identification of those events via automated detection was 
difficult even at the 30 m resolution level. Consequently, detection of active 
understory fires using MODIS or GOES data appears to be significantly difficult 
using standard mid-infrared detection algorithms. 
Another limitation of our analyses involves the conversion of 30 m active fire 
pixels statistics derived from ASTER and ETM+ into fire area and temperature 
 
 87 
 
estimates. Although theoretical assessment of ASTER channel 8 saturation was 
attempted relating minimum fire area and temperature [Giglio et al., 2008], a more 
complete characterization of fire properties is still under investigation. Advancement 
in this area is limited primarily by inadequate specifications of existing orbital 
sensors. 
This study extends a series of validation analyses applied to the MODIS 
Thermal Anomalies product and is the first one of its kind applied to the GOES WF-
ABBA product. Our results corroborate previous findings and include new aspects 
involving the performance of the MODIS and GOES fire detection algorithms which 
can help the fire data user community optimize their use of these products. 
Lessons learned from the validation of MODIS and GOES imager will be 
used to develop validation procedures for the fire products from new generation of 
US polar orbiter (NPOESS Visible Infrared Imager Radiometer Suite -- VIIRS) and 
geostationary (GOES-R Advance Baseline Imager -- ABI) platforms. Of particular 
importance is the development of a multi-platform validation system and the 
quantification of validation errors stemming from the use of non-simultaneous space-
based or airborne reference data. In addition, the better understanding of the detection 
envelopes of MODIS and GOES imager derived from this study can be used to assess 
the expected performance of the higher resolution VIIRS and ABI fire products. 
 
 
 
 88 
 
Chapter 4: Integrating Geostationary and Polar Orbiting Data 
for Monitoring Vegetation Fires in Brazilian Amazonia 
 
 
4.1 INTRODUCTION 
Biomass burning is a major problem affecting in particular the tropical regions 
[Crutzen et al., 1979; Andreae, 1991; Hao and Liu, 1994]. Vegetation fires contribute 
to large amounts of trace gases and aerosols emitted during combustion [Crutzen and 
Andreae, 1990; Crutzen et al., 1979; Seiler and Crutzen, 1980]. The large rate of 
emissions from tropical fires was found to have important consequences to 
atmospheric composition and to the global radiation budget [Penner et al., 1992, 
Andreae et al., 2004]. Chemically reactive constituents produced during combustion 
such as CO and nitrogen oxides (NOx) may combine to form tropospheric O
3
 
[Crutzen, 1979; Delany et al., 1985; Rummel et al., 2007; Watson et al., 1990], which 
under high concentrations near the surface can have adverse effects on live vegetation 
and human health [Crutzen and Andreae, 1990; Hansen et al., 2000; Vitousek et al., 
1997]. The role of biomass burning aerosols as cloud condensation nuclei was also 
proven to alter the nature and distribution of clouds in tropical regions especially 
when smoke concentrations are high [Andreae et al., 2004; Koren et al., 2004]. 
Quantification of vegetation fire activity and the associated emissions from 
biomass burning in Amazonia has been approached via remote sensing and modeling 
techniques [Andreae et al., 2004; Freitas et al., 2005; Kaufman et al., 1990, 1992; 
 
 89 
 
Setzer and Pereira, 1991]. In both cases, satellite active fire detection data represent 
one of the most important parameter for calculating source emissions. Fire alert 
systems alike depend heavily on satellite data to monitor fire activity in Amazonia as 
this is the only form of systematic observation available for the region [CONAE, 
2008; CPTEC, 2008; DIMARENA, 2008]. 
Fire monitoring and emissions modeling applications usually require the best 
spatial and temporal resolution data available [GCOS, 2006]. Early detection is most 
important for fire alert systems demanding information at high observation frequency 
and fine spatial resolution to increase the probability of detecting a fire immediately 
after its initiation or when its active area is still small. Similar data needs are valid for 
emissions modeling applications which depend on high temporal and spatial 
resolution data to determine the lifetime of individual fires and to calculate the 
resulting emissions [Freitas et al., 2005]. Emissions modeling applications also 
require characterization of fire properties (e.g., total radiative energy, area, 
temperature) for calculating emissions, demanding quality pixel data which can only 
be achieved by non-saturating satellite sensors [Justice et al., 2002; Kaufman et al., 
1998b]. 
The lack of orbital systems capable of delivering fire detection data at both the 
spatial and temporal resolutions required by users is a major limitation affecting 
current fire products [Schroeder et al., 2005]. Variations in product performance as a 
result of sensor characteristics and the diurnal cycle of fires introduce additional 
complications for fire data users influencing the use of the data [Giglio, 2007; 
Schroeder et al., 2005, 2008b]. To cope with these limitations, fire data users must 
 
 90 
 
incorporate assumptions in their analyses (e.g., emissions modelers need to estimate 
the lifetime of a fire using hypothetical curves) or prioritize one data set over others 
thereby avoiding conflicts among different products. However, the tendency of most 
emissions models and fire alert systems to deliver increasingly detailed data on fire 
activity creates a great demand for an integrated satellite product which takes 
advantage of the complementarities among current fire detection data [GOFC, 2006]. 
In this study, we assess the potential for integrating the fire product 
information derived from the Moderate Resolution Imaging Spectroradiometer 
(MODIS) on board Terra and Aqua satellites and the Geostationary Operational 
Environmental Satellite (GOES) imager. We use in our analyses the Thermal 
Anomalies MOD14 and MYD14 Collection 4 active fire detection products for 
MODIS Terra and Aqua, respectively, and the Wildfire Automated Biomass Burning 
Algorithm (WF-ABBA) version 6.0 for GOES [Giglio et al., 2003a; UW Madison 
CIMSS, 2008]. The high temporal resolution of the GOES imager and the moderate 
spatial resolution of MODIS provide a unique opportunity to test how fire detection 
products from instruments with very different configurations complement each other. 
Our main objective is to investigate how the fire detections from MODIS and GOES 
can be used in concert to improve fire monitoring capability and to assess the 
feasibility of developing a hybrid product based on the two products. We focus our 
analyses on Brazilian Amazonia, a region routinely covered by MODIS Terra and 
Aqua and the GOES imager. In this region, significantly high annual fire activity 
spreading over rather remote areas requires that optimized monitoring systems be 
used. 
 
 91 
 
We tested the hypothesis of data complementarity using two different 
approaches. First, we analyzed the MODIS and GOES imager fire detection 
information by assessing the Fire Radiative Power (FRP) estimates produced from 
each data set. The reasons for using FRP estimates to relate MODIS and GOES fire 
information are threefold: (i) it provides a physical parameter which facilitates 
interpretation of the results, (ii) it provides quantitative characterization of fire 
activity which is more informative compared the to categorical data represented by 
simple fire detection and (iii) it has been demonstrated that FRP can be used to 
estimate biomass combustion rates, an important parameter in emissions modeling 
[Kaufman et al., 1998b; Wooster, 2002; Wooster et al., 2003]. Therefore FRP 
becomes a potentially valuable parameter for fire data users and emissions modelling 
applications. Secondly, we assess the relationship between spatially and temporally 
coincident MODIS and GOES active fire detections using the binary classification 
(i.e., 1 = fire detection; 0 = no fire detection). Although less informative, binary fire 
detections are by far the most commonly used information among satellite fire data 
users. The comparison of the binary data examines the factors that influence the rates 
of the spatially and temporally coincident detections produced by MODIS Thermal 
Anomalies and WF-ABBA products. We substantiate our analyses of the MODIS and 
GOES fire products using sub-pixel fire characterization derived from higher 
resolution imagery from the Advanced Spaceborne Thermal Emission and Reflection 
Radiometer (ASTER) and the Enhanced Thematic Mapper Plus (ETM+) instruments. 
 
 
 92 
 
4.2 VEGETATION FIRES IN BRAZILIAN AMAZONIA 
Fires are used in Amazonia primarily for land clearing and maintenance applications 
[Andreae, 1991; Cochrane, 2003]. Burning occurs in a somewhat controlled fashion 
and is used to eliminate forest debris, or crop residues, or to replace or rejuvenate 
grasslands. Under these circumstances, fires tend to be constrained to the perimeter of 
the land being used although escaped fires are also commonly reported [Alencar et 
al., 2006]. 
The humid climate of Amazonia limits fire use to a small window of 
opportunity centered in the peak months of the dry season [Schroeder et al., 2005, 
2008a]. The decision to burn is highly influenced by the development of the dry 
season and the ?signs? of the subsequent wet season. Correct identification of those 
?signs?, especially by small farmers, can represent the difference between a 
prosperous year or perhaps starvation. The strong bonds among individuals 
occupying adjacent areas limit the apparent window of opportunity for burning even 
further. In this case, initiation of a fire in one property can influence others to follow. 
Consequently, vegetation fires can become significantly concentrated in time. For 
instance, Schroeder et al. (2005) reported that approximately 2,000 MODIS Aqua 
(MYD14) fire detections were produced for the state of Acre in western Brazilian 
Amazonia during 1-3 September  2003, representing nearly 30% of all fire detections 
in that year. 
 We used 285 ASTER and ETM+ scenes to quantify active fire size in 
Brazilian Amazonia and the minimum distance between adjacent fire clusters. The 
scenes were distributed in the main areas of fire activity in the region and were 
 
 93 
 
acquired during 2001-2005. Active fire masks at 30 m resolution were produced for 
ASTER and ETM+ data following the approach described in Giglio et al. (2008) and 
Schroeder et al. (2008b), respectively. Contiguous 30 m active fire pixels were 
clustered together to produce individual fire size estimates. Figure 4.1A shows the 
distribution of fire cluster size for the region. Fire size statistics derived from the data 
above are valid for fires imaged at around 10:00-10:30 local time; it is plausible that 
larger fires might prevail during the early afternoon hours as a result of the diurnal 
cycle of fires [Giglio, 2007; Prins et al., 1998]. Also, fire cluster sizes depicted in 
Figure 4.1A are expressed in terms of 30 m active fire pixels derived from the fire 
masks and will likely differ from the actual area occupied by flaming fronts. 
Regardless of the approximations involved, the typical spatial scale of Amazonian 
fires is seen to represent only a small fraction of a nominal 1?1 km MODIS or 4?4 
km GOES pixel areas. 
 The minimum distance between adjacent fire clusters was determined using 
the same data set above (Figure 4.1B). Approximately 50% of the fire locations 
derived from the 30 m fire masks showed an adjacent fire cluster within 2 km or less 
from their outer limits. This proximity between clusters reinforces the concept that 
fires are used simultaneously at several properties in one particular region, making 
them concentrated in time as well as in space. 
 
 
 
 94 
 
 
 
 
Figure 4.1: Cumulative frequency graphs of fire cluster size (A) and minimum distance to 
nearest fire (B). 
 
 
Complementing the characterization of vegetation fire conditions in 
Amazonia, 2,785 burnt area polygons were digitized using a representative subset of 
the 30 m resolution data above composed of 33 ETM+ scenes. While active fires 
represented only a small fraction of a moderate-to-coarse resolution pixel footprint, 
the resulting burnt area polygons were large enough to approach or even exceed the 
A 
B 
 
 95 
 
size of a MODIS or GOES pixel (Figure 4.2). The occurrence of pixel-size burn scars 
in Amazonia has important implications for active fire detection in particular in areas 
of high percentage tree cover where there is a high contrast between forest and non-
forested pixels in the mid-infrared spectral region [Tucker et al., 1984]. The 
occurrence of burn scars occupying large fractions of non-forested pixels may further 
reinforce the contrast with surrounding forest areas causing contextual algorithms 
such as used by the MOD14, MYD14 and WF-ABBA products to generate false 
detections, therefore resulting in high commission error rates [Schroeder et al., 
2008b]. 
 
 
 
Figure 4.2: Burnt area size histogram derived using ETM+ data for Brazilian Amazonia. 
Data are representative of 2,785 individual burn sites extracted from 33 ETM+ scenes. A 
nominal MODIS pixel is equivalent to 100 ha, whereas the nominal GOES pixel covers an 
area of approximately 1600 ha. 
 
 
 96 
 
4.3 METHODS 
The use of pixel level analyses methods applied to the active fire detection data from 
MODIS and GOES was not implemented in this study primarily due to limitations 
involving the distinct image geometry of the two instruments. Geometric distortion of 
off-nadir pixels introduced important across-image mapping and resampling errors 
[Wolfe et al., 2002]. Also, orbit oscillations affecting in particular the GOES data 
resulted in different navigation offsets across sectors of the images often leading to 
poor overall navigation quality. Persistent cloud coverage across Amazonia created 
major limitations for the use of automated methods to register data sets and to reduce 
navigation differences. Lastly, the spatial characteristics of Amazonian fires 
described in Figure 4.1 suggest that moderate or coarse resolution fire products may 
likely contain multiple fires distributed over adjacent detections. To cope with the 
limitations involving instruments? characteristics and the spatial characteristics of 
fires in the region, we opted to aggregate the fire detection information into clusters 
by analyzing groups of contiguous detections. When analyzing fire detections from 
MODIS and GOES, we also used area averaged estimates of fire activity in order to 
minimize the effects of localized fire detection differences. 
 
4.3.1 MODIS FRP 
In our analyses we used the FRP estimates available as part of the MODIS Thermal 
Anomalies Collection 4 product Scientific Data Set [Giglio et al., 2003a]. The 
MODIS FRP derivation is based on the relationship between the emitted fire energy 
 
 97 
 
and the mid-infrared brightness temperature, and is expressed as [Kaufman et al., 
1998b]: 
 
( )
8
4
8
4
19
1034.4
b
TTFRP ??=
?
  
 
Where T
4
 and T
4b
 are the brightness temperature estimates in the 4 ?m region of band 
22 (or alternatively band 21 when band 22 becomes saturated) for the target and the 
background pixels, respectively. The constant in equation (1) is adjusted for the 
spectral interval of MODIS channel 21-22.  
 
4.3.2 GOES FRP 
Currently there is no FRP estimate available with the operational WF-ABBA version 
6.0 product. Alternatively, fire size and temperature estimates derived using Dozier?s 
(1981) approach are available with the product. However, problems inherent to 
Dozier?s technique are known to affect the estimates produced [Giglio and Kendall, 
2001], limiting their application for FRP retrievals. Nonetheless, Wooster et al. 
(2003) have demonstrated that it is possible to derive physically sound FRP by 
approximating Planck?s Radiation law using a fourth order power law applied to mid-
infrared remote sensing data (e.g., channel 2 on the GOES imager) over the interval 
of typical vegetation fire temperatures (i.e., 600 ? 1500 K). After relatively simple 
considerations are made following the approximation above, FRP estimates can be 
calculated through the application of the following equation [Wooster et al., 2005]: 
 
(1) 
 
 98 
 
()SLL
a
FRP
bf ,2,2
?=
?
 
 
Where L
2,f
 and  L
2,b
 are here represented by the GOES channel 2 radiances (W m
-2 
sr
-1 
?m
-1
) estimated for the target pixel containing the fire and for the background pixels, 
respectively, ? is the Stefan-Boltzmann constant (5.67?10
-8
 W m
-2 
K
-4
), a is a 
constant derived for the particular spectral response function of channel 2 (value for 
GOES-12 imager: 3.08?10
-9
 Wm
-2
 sr
-1
 ?m
-1
 K
-4
), and S is the ground equivalent pixel 
area (m
2
).  
Derivation of the GOES FRP must take into consideration the significant 
along scan pixel overlap which occurs during normal imager operation [Menzel and 
Purdom, 1994]. Under such conditions, every point on the surface is effectively 
imaged by two adjacent pixels. Consequently, two WF-ABBA detections will likely 
be produced representing an individual surface fire. This condition could lead to an 
overestimation of GOES FRP based on (2), as term S would be biased by a factor of 
two. To deal with the uncertainties in sub-pixel fire location and the effects of along 
scan pixel overlap, we calculated GOES FRP for fire clusters containing two or more 
adjacent pixels in the along scan direction by assigning variable weights to those 
pixels. The criterion for determining the weight of different cluster elements was 
based on the brightness temperature of each pixel; brighter pixels were assigned 
greater weight in the FRP calculations. For a fire detection cluster composed of three 
adjacent pixels in the along scan direction we calculated the GOES FRP using the 
following two steps: 
(i) Rank pixels according to their brightness temperatures (e.g., T
2
>T
1
>T
3
) 
(2) 
 
 99 
 
(ii) Assign greater weight to brighter pixels in the calculations by adjusting the pixel 
area in (2) using: 
()()
()()
()()
321
323,3,33
2,2,22
121,1,11
FRPFRPFRPFRP
SSLL
a
FRP
SLL
a
FRP
SSLL
a
FRP
cluster
bf
bf
bf
++=
???=
??=
???=
?
?
?
?
?
  
 
Where the symbols are similar to equation (2) except for S
i ? j
, which indicates 
the area of overlap between adjacent pixels i and j. 
 
4.3.3 Simulation Data 
As shown in Figure 4.1A, Amazonian fires are relatively small and therefore occupy 
only a reduced fraction of a moderate-to-coarse resolution pixel. Because of the Point 
Spread Function (PSF) characteristics of sensors such as MODIS and GOES, the 
radiance corresponding to each pixel will be influenced by the spatial arrangement of 
the sub-pixel features and might as well be affected by radiance coming from surface 
elements located just outside the pixel?s nominal area [Cahoon et al., 2000; Zhang et 
al., 2006]. Using simulation data, we investigated how the relative position of a fire 
line contained in a moderate-to-coarse resolution pixel affects the FRP estimates 
derived from MODIS and GOES. Target pixels of nominal spatial resolution (1?1 km 
for MODIS and 4?4 km for GOES) were reproduced containing two distinct thermal 
components, namely: (i) the fire background at 311 K, and (ii) a high temperature 
 
 100 
 
source at 800 K with a fixed area of approximately 5.5 ha (155?365 m). The 
background pixels necessary for the calculation of FRP were simulated as single 
thermal component pixels with their temperatures set at the same fire background 
temperature of 311 K. Those conditions were considered representative of typical 
Amazonian fires detected by MODIS and GOES. The location of the fire line relative 
to the pixel center varied in our simulations to the extent permitted by each 
instrument, i.e., 0.5 km along scan and 0.5 km along track for MODIS; 1.25 km along 
scan and 2 km across scan for GOES. The reduced along scan spatial tolerance in 
GOES was designed to reproduce the over-sampling characteristic of its imager. The 
formats of the PSF used for MODIS and GOES are shown in Figure 4.3A and 4.3B, 
respectively. 
 
 
 
Figure 4.3: 3D scheme of the Point Spread Function of MODIS (A) and GOES (B) 4 
?m channels. 
 
 
A B
 
 101 
 
4.3.4 Comparing MODIS and GOES Fire Detections 
The relationship between MODIS (MOD14 and MYD14) and GOES (WF-ABBA) 
fire detections was derived using three years (2003-2005) of daytime data for 
Amazonia. The main objective of this analysis was to determine the rates of spatially 
and temporally coincident detections produced by the two products above. The 
instantaneous rate of coincident detections was derived from pairs of MODIS and 
GOES acquisitions separated by less than 15 minutes. Contiguous fire detections 
were aggregated into clusters and a search radius of 8 km was used to locate WF-
ABBA detections corresponding to a MOD14 (or MYD14) cluster, and vice-versa. 
This larger search radius was intended to accommodate navigations errors in the 
GOES data. 
We also derived the rates of coincident detections between MOD14 (and 
MYD14) and WF-ABBA using all daytime observations from GOES. In this case, the 
objective was to determine how the increased observation frequency of GOES would 
influence the rates of coincident detections. Using the spatial criterion above, we 
searched for same day WF-ABBA detections produced during the period of 
approximately 7:00 to 19:00 local time. 
To facilitate the interpretation of the results, the rates of spatially and 
temporally coincident fire detections produced by MODIS and GOES were separated 
into four discrete intervals of percentage tree cover based on the 500 m Vegetation 
Continuous Fields (VCF) product [Hansen et al., 2002; Hansen et al., 2003]. 
Complementing the analysis above, we derived the relationship between 
MOD14 (and MYD14) and WF-ABBA detections for a rectangular area of 200?300 
 
 102 
 
km
2
 centered in Mato Grosso state, southern Brazilian Amazonia. The area was 
oriented following the orbit inclination of Terra and Aqua in order to sample portions 
of the daily MODIS images showing similar pixel geometry. We derived the sum of 
MOD14 (and MYD14) and WF-ABBA fire detections occurring within 15 minutes 
from each other for the spatial subset above. The results were summarized according 
to discrete MODIS scan angle intervals in order to evaluate the effects of variable 
imaging conditions on the MODIS product. 
 
4.4 RESULTS AND DISCUSSION 
4.4.1 MODIS and GOES FRP 
FRP estimates were produced for 12 pairs of near coincident MODIS Terra and 
GOES imagery and for 11 pairs of near coincident MODIS Aqua and GOES imagery, 
each containing several fire detection clusters. The GOES images were registered to 
MODIS in order to minimize navigation errors, and FRP values were derived for each 
detection cluster. Clusters with one or more saturated pixels (more commonly seen in 
the GOES data) were not used. To ensure that individual pairs of MODIS and GOES 
detection clusters were representative of the same surface fire, we discarded all cases 
having more than one detection cluster from each instrument within a 10 km radius. 
The spatially and temporally coincident FRP estimates for MODIS Terra and 
GOES and MODIS Aqua and GOES are shown in Figures 4.4A and 4.4B, 
respectively. The correlation between MODIS and GOES remained low despite all 
measures adopted to reduce artifact contamination of the pairs of FRP values 
 
 103 
 
produced. Visual inspection of a subset of 77 pairs of MODIS Terra and GOES fire 
detection clusters using coincident ASTER data helped us identify the major factors 
influencing our results. 
 
 
 
 
Figure 4.4: Scatter plots of FRP estimates produced for spatially and temporally coincident 
MOD14-WF-ABBA (A) and MYD14-WF-ABBA (B) fire detections clusters. The one-to-one 
line is plotted for reference. 
A 
 
 104 
 
 
Figure 4.4 continued 
 
 
First, differences in FRP were found to prevail over areas of high landscape 
heterogeneity. Fire detections located in areas of marked variation in vegetation cover 
occurring at the same spatial scale of the MODIS or GOES pixel (e.g., deforestation 
sites in high percentage tree cover regions) showed systematically larger differences 
in FRP compared to areas of homogeneous background (e.g., grasslands in low 
percentage tree cover regions) (Figure 4.5). Differences between the vegetation cover 
of the target pixels and their surrounding areas are expected to influence the 
background characterization affecting the derivation of FRP through (1) and (2) 
[Wooster et al., 2003]. Detection clusters covering areas of deforestation where the 
surrounding pixels were partially or entirely covered by evergreen tropical forests 
were particularly prone to produce large differences between MODIS and GOES FRP 
B 
 
 105 
 
values. In these cases, the brightness temperature of the forested pixels can be 15 K 
cooler than the fire pixel background, resulting in a potential overestimation of FRP > 
20% according to (1) (assuming an intermediate fire detection condition). 
Differences between FRP estimates from MODIS Terra and GOES were also 
associated with omission of secondary fire lines contained in adjacent pixels. The 
landscape heterogeneity was equally important in this case as it influenced the 
propagation of fire lines, resulting in fragmented fires composed of several small 
flaming areas. Omission pixels constitute a relatively small percentage of the total 
energy of a fire cluster; however, their occurrence can also influence the 
characterization of the background pixels therefore amplifying their net effect on the 
FRP calculation. 
 
 
Figure 4.5: Fire intensity (Wm
-2
) estimates for four percentage tree cover intervals [VCF, 
Hansen et al., 2002] calculated by dividing per pixel FRP estimates for GOES and MODIS by 
the approximate active fire area determined using coincident 30 m resolution ASTER and 
ETM+ imagery [Schroeder et al., 2008b]. Values plotted represent the median of the fire 
intensity data available for each VCF interval. 
 
 
 106 
 
 Finally, large differences between MODIS and GOES FRP values were found 
to result from the effects of the PSF of each instrument. The same fires occupying the 
center of a pixel from one sensor and the edge of a pixel from the other sensor 
resulted in significantly different FRP estimates. Using the simulation data described 
in Section 4.3.3, we evaluated the contribution of the PSF of MODIS and GOES to 
the FRP values calculated (Figure 4.6). The reduction in FRP observed when a fire 
cluster is located away from the more responsive area near the pixel?s center 
represented the single most important factor contributing to the differences in FRP 
estimates produced by MODIS and GOES. Due to the predominantly small size of 
Amazonian fires, the effects of PSF become particularly pronounced in the GOES 
and MODIS data. 
 
 
 
Figure 4.6: Simulated FRP for MODIS (A) and GOES (B) as a function of distance to the 
pixel?s center. FRP estimates are calculated for a single thermal component fire at 800 K 
covering an area of approximately 5.5 ha, with background temperature of 311 K. 
A 
 
 107 
 
 
Figure 4.6 continued 
 
 
4.4.2 MODIS and GOES Fire Detections 
Our analysis of correspondence between the MODIS Thermal Anomalies and GOES 
WF-ABBA products used approximately 115,000 MOD14 detections, 225,000 
MYD14 detections, and 25,000 and 93,000 WF-ABBA detections produced within 15 
minutes from the MODIS Terra and Aqua overpasses, respectively.  
Figures 4.7A and 4.7B show the rates of coincident detection between 
MOD14 and WF-ABBA and between MYD14 and WF-ABBA, respectively, using 
the instantaneous observation of GOES at near the same overpass time of Terra and 
Aqua. Commission error estimates derived in Chapter 3 are also shown in Figures 
4.7A and 4.7B for reference. The low overall rate of MOD14 and MYD14 detections 
having a coincident WF-ABBA match occurs primarily due to the greater detection 
performance of the MODIS product [Schroeder et al., 2008b]. The rate of coincident 
detections increases towards intermediate-to-high percentage tree cover intervals as a 
B 
 
 108 
 
result of longer lasting and higher intensity fires which are typically found in those 
regions. These fires have a greater probability of detection by WF-ABBA [Schroeder 
et al., 2008b], thereby contributing to the increase of coincident detection rates with 
MOD14 and MYD14. Figures 4.7A and 4.7B also show the effects of the fire diurnal 
cycle on the results. The systematically lower (higher) rates of coincident detections 
between MOD14 (MYD14) and WF-ABBA show the contribution of smaller (larger) 
fires which predominate in the morning (afternoon) hours in response to the local 
atmospheric and fuel conditions. Due to its coarser spatial resolution, the WF-ABBA 
product will likely omit a greater percentage of the small morning fires compared to 
MOD14 creating the differences seen between Figures 4.7A and 4.7B. 
The results describing the rates of coincident detection between MOD14 and 
WF-ABBA and between MYD14 and WF-ABBA using all daytime observations 
from GOES are shown in Figures 4.7C and 4.7D, respectively. Similar to Figures 
4.7A and 4.7B, commission error rates for MOD14 and MYD14 derived from 
Chapter 3 are displayed in Figures 4.7C and 4.7D for reference. The increased rates 
of coincident detection produced demonstrate how effective the high observation 
frequency of GOES is in reducing the overall differences in detection performance 
seen in Figures 4.7A and 4.7B. The higher rates of coincident detection of MOD14 
and WF-ABBA in Figure 4.7C show once again the effects of the diurnal cycle of 
fires. Compared to MYD14, morning MOD14 fires have a greater potential to benefit 
from the diurnal variation in atmospheric and fuel conditions as they enjoy an 
extended number of hours of increasing temperature and decreasing humidity of both 
fuel (i.e., vegetation) and the surface air layer. Therefore, small fires detected only by 
 
 109 
 
MOD14 in the morning hours can develop into larger and more intense burning 
events until they become active enough to trigger a WF-ABBA detection during the 
subsequent observation hours [Schroeder et al., 2008b]. 
The fraction of MOD14 and MYD14 detections in Figures 4.7C and 4.7D 
occurring over intermediate-to-high percentage tree cover areas (VCF >40%) without 
a coincident WF-ABBA detection was similar to the commission error rates estimated 
for the MODIS products. Visual interpretation of 30 m ASTER scenes covering 46 
MOD14 false detections coincidently imaged by GOES data showed that the majority 
of the cases (n = 44) did not produce a coincident WF-ABBA detection; whereas all 
coincident MOD14 and WF-ABBA detections examined (n = 73) showed visual 
confirmation of fire activity. This result suggests that: (i) coincident detections 
produced by different products can be classified as true positives; and (ii) isolated 
MOD14 detections occurring in densely vegetated areas could be associated with 
false detections (more so than MYD14 based on Figures 4.7C and 4.7D). 
Consideration of commission errors could not help explain a large fraction of MOD14 
and MYD14 detections without spatially coincident WF-ABBA detections over low 
percentage tree cover areas. Short duration and low intensity fires which typically 
occur in Cerrado (savanna) regions may be the primary cause for the lower rates of 
coincident detections found as MOD14, MYD14 and WF-ABBA will show only 
small probability of detecting these highly dynamic fires. 
 
 
 
 
 110 
 
 
 
 
 
 
Figure 4.7: Rates of coincident detection (gray bars) for MOD14-WF_ABBA instantaneous 
(A), MYD14-WF_ABBA instantaneous (B), MOD14-WF_ABBA day-time (C), and 
MYD14-WF_ABBA day-time (D) observations. Commission error rates for MOD14 and 
MYD14 are also shown (black bars). 
 
 
The results in Figures 4.7A-D describe the rates of MODIS (MOD14 or 
MYD14) fire detections having a spatially coincident WF-ABBA detection. We also 
assessed the opposite condition, i.e., we derived the rates of WF-ABBA fire 
detections having spatially coincident MODIS detections. The results were 
considerably less intuitive in this case. On average, only 50% of the WF-ABBA 
detections produced at the time of MODIS Terra overpass showed spatially 
coincident MOD14 detections (Figure 4.8A). An even lower percentage (40%) of 
WF-ABBA detections produced at the time of MODIS Aqua overpass had spatially 
A B
C 
D
 
 111 
 
coincident MYD14 detections (Figure 4.8B). The low rates of coincident detections 
were consistent across all four percentage tree cover intervals. Visual inspection of 30 
WF-ABBA morning fire detections without a coincident MOD14 detection using 
ASTER 30 m scenes revealed that the majority of the cases (75%) were associated 
with a false detection. The remaining 25% of the WF-ABBA detections inspected had 
significantly small active fires within the GOES pixel footprint, and could indicate 
potential false alarms induced by relatively cooler background conditions. 
 
 
 
Figure 4.8: Rates of coincident detection (gray bars) for WF-ABBA-MOD14 (A), WF-
ABBA-MYD14 using instantaneous GOES data. Commission error rates for WF-ABBA are 
also shown (black bars). 
 
 
The daily sums of paired MOD14-WF-ABBA and MYD14-WF-ABBA 
detections derived for a 200?300 km
2
 area centered in Mato Grosso state, southern 
Brazilian Amazonia are shown in Figure 4.9. We used three years of data (2003-
2005) totaling 15,000 MODIS and 8,000 GOES fire detections overlapping with the 
area above.  The fixed location of the spatial subset used resulted in constant GOES 
viewing conditions, while the observation geometry of MODIS varied on a daily 
A B
 
 112 
 
basis. Consequently, the performance of the MODIS fire product changed as a 
function of scan angle due to the pixel distortion towards the edges of the swath. 
The diurnal cycle signature was also found to be pronounced in Figure 4.9. 
Morning fire detections were predominantly originated from MODIS Terra data, with 
the contribution of GOES detections increasing towards the afternoon. Both MODIS 
Terra and Aqua data showed the effects of pixel distortion. Assuming the detection 
performance of WF-ABBA to be invariant across the subset area and period analyzed, 
we estimated the detection performance of MODIS near the edge of the swath to be 
on average approximately 40% lower compared to nadir viewing conditions. Greater 
variability in detection performance is seen to affect both MOD14 and MYD14 
products at higher scan angles as indicated by the boxes and lines in Figure 4.9. The 
longer atmospheric path described at extreme MODIS scan angles could help explain 
such variability as it will likely affect the detection probability of underlying active 
fires for heavily contaminated atmospheres. The bow tie effect which produces 
significant pixel overlap at large MODIS scan angles could also contribute to the 
detection variability observed. In this case, the pixel overlap can produce duplicate 
detections composed of single or multi-pixel clusters with consequences to the daily 
sum statistics of fire detections of MODIS Terra and Aqua. 
 
 
 113 
 
 
 
 
 
Figure 4.9: Box plots for the relative contribution of MOD14 (A) and MYD14 (B) products 
to the sum of MODIS and GOES fire detections produced using an area of 200?300 km
2
 
centered in Mato Grosso state, southern Brazilian Amazonia. Parameters depicted in the plot 
include: minimum, 25
th
 percentile, median, 75
th
 percentile, and maximum values. Results 
were separated into 10
o
 scan angle intervals of MODIS. Positive and negative scan angles 
describe the east and west parts of the MODIS swath, respectively. 
 
A 
B 
 
 114 
 
4.5 CONCLUSIONS 
Each year earth remote sensing satellites detect hundreds of thousands of vegetation 
fires across Brazilian Amazonia and thereby represent the primary source of 
information for a large number of fire data users. In this study, we assessed how 
different satellite active fire detection data might be used synergistically to provide 
improved detection rates with higher confidence than can be obtained from individual 
products.  
Our study showed that fires of relatively small active area predominate in 
Brazilian Amazonia. Approximately 90% of all fires sampled had an area of less than 
3 hectares (assuming the ASTER and ETM+ 30 m pixels used to create the statistics 
were completely covered by active fires), therefore representing a significantly small 
fraction of a moderate-to-coarse resolution pixel. Amazonian fires were also found to 
be both spatially and temporally concentrated, with multiple fires occurring 
simultaneously within a few kilometers from each other. Consequently, single or 
multi-pixel fire detection clusters coincidently produced by MOD14 (MYD14) and 
WF-ABBA were prone to contain different amounts of fire activity within each pixel. 
The challenges for active fire detection products based on moderate-to-coarse 
resolution data are many in this case, regardless of the algorithm used. For instance, 
fire characterization via FRP may result in biased estimates depending in particular 
on the number of individual active fires contained within the fire detection pixel and 
on how the energy released by the fire is represented by the instrument?s PSF. 
Additionally, Amazonian fires tend to concentrate in areas of deforestation where the 
landscape heterogeneity is high making the derivation of FRP values even more 
 
 115 
 
difficult. As a result, the potential for creating a hybrid fire product combining 
MODIS and GOES FRP estimates is significantly limited. 
 The analyses of fire detections revealed that if one is interested in representing 
the total fire occurrence for high percentage tree cover areas (>40%), the better 
temporal resolution of the GOES data detecting the diurnal cycle of fires improves 
upon the twice daily detection of MODIS at a higher spatial resolution. Therefore, 
differences in daily fire detection in particular between MOD14 and WF-ABBA may 
be reduced or even eliminated. False detections produced by MOD14 were 
predominantly isolated from WF-ABBA detections, while 100% of the areas showing 
coincident detections from those two products had visual confirmation of fire activity. 
Consequently, MOD14 (and similarly MYD14) could be used to reinforce our 
confidence of coincident WF-ABBA detections, while the remaining non-coincident 
MOD14 detections (more so than MYD14 ones) may indicate areas of potential 
commission errors particularly in high percentage tree cover areas (>40%).  
For low percentage tree cover areas, omission errors are high and fires are 
very dynamic thereby reducing the number of coincident detection between MOD14 
(MYD14) and WF-ABBA. In this case, the low overall commission error rates 
suggest that combining MOD14, MYD14 and WF-ABBA fire detections could be the 
best approach to improve biomass burning monitoring in those regions. 
 Visual inspection of 30 m resolution ASTER imagery suggested that 
commission errors in WF-ABBA could be largely responsible for the low rates of 
coincident detections between WF-ABBA and MOD14. Large burnt areas occurring 
at the same spatial scale of the GOES pixels were seen to produce WF-ABBA fire 
 
 116 
 
detection clusters without a MOD14 match. The percentage of WF-ABBA fire pixels 
without a coincident MOD14 or MYD14 detection exceeded the commission errors 
estimated for the WF-ABBA by Schroeder et al. [2008], and further investigation is 
required. 
The detection performance of the MOD14 and MYD14 at extreme scan angles 
(55?) was found to be approximately 40% lower compared to nadir viewing 
conditions. The use of area averaged statistics of daily MOD14 (MYD14) and WF-
ABBA fire detections as a means to compare the two products resulted in poor 
correlation between those datasets across a range of MODIS observation conditions. 
The ratio between MOD14 (MYD14) and WF-ABBA detections were especially 
variable at high MODIS scan angles, where the bow tie effects and the occurrence of 
thick smoke plumes could influence the number of detections produced. 
The results above summarize some of the major limitations of current satellite 
active fire detection products. The nature of fire activity in Brazilian Amazonia and 
the image characteristics of MODIS and GOES were found to affect fire detection 
probability and the retrieval of physical parameters of a large percentage of 
vegetation fires in the region. 
Regardless of the limitations above, increasingly sophisticated fire alert 
systems and biomass burning emissions models create a large demand for higher 
temporal and spatial resolution fire data [Freitas et al., 2005; GCOS, 2006; GOFC, 
2006; Wooster et al., 2003]. The next generation suite of sensors to become 
operational on board the National Polar-orbiting Operational Environmental Satellite 
System (NPOESS) and the GOES-R will show improvements in terms of spatial (and 
 
 117 
 
temporal in the case of GOES-R) resolution which could result in greater fire 
detection and characterization capabilities compared to existing products. However, 
the primary channel for active fire detection located in the mid-infrared region in both 
systems might be affected by inadequate pixel saturation specifications and pixel 
resampling methods which could penalize the performance of future fire detection 
products derived from those data sets. Unless the quality and the integrity of the new 
data are preserved, the user community may continue to experience a shortage of 
detailed information about fire activity with important consequences to the 
understanding of regional and global climate system processes and their implications 
to society. 
 
 
 118 
 
Chapter 5:  Integrated Fire Product 
 
 
5.1 INTRODUCTION 
The results derived from Chapters 2-4 were used to guide the production of a new 
integrated fire product for Brazilian Amazonia based on MOD14, MYD14 and WF-
ABBA input data. The new product was aimed to generate the best possible fire 
detection data with reduced commission and omission errors. A 4 km grid was used 
to process daily MODIS and GOES data including all daytime and nighttime 
detections. A set of criteria was developed utilizing the complementarity of the above 
products to generate an integrated product of fire detection data, namely: 
 
i. Cloud obscuration omission errors were calculated for each 4?4  km cell using 
the methodology described in Chapter 2; 
ii. Commission errors were calculated for each 4?4 km cell as a function of 
percentage tree cover based on the results from Chapter 3; 
iii. False detections in MOD14 (MYD14) and WF-ABBA were considered non-
coincident. Based on the results from Chapter 4, spatially coincident 
detections (i.e., detections within the same 4?4 km cell) produced by more 
than one product were classified as true positives; 
 
 119 
 
iv. Daytime MOD14 fire detections without a spatially coincident MYD14 or 
WF-ABBA detection were discarded for high percentage tree cover areas 
(>40%) based on the results from Chapter 4. 
 
In addition, nighttime detections were all considered true positives based on 
the analysis of 20 ASTER scenes used to validate the MOD14 product derived from 
nighttime MODIS Terra overpasses using the same methodology described in 
Chapter 3. All nighttime detections validated (n = 62) were classified as true 
positives, suggesting that false detections are primarily a result of daytime land 
surface temperature dynamics. 
 
5.2 METHODS 
The integrated product differs from the simple sum of MODIS and GOES fire 
detection products as it is based on the use of multiple criteria to determine the 
consistency of fire detections produced for an individual 4?4 km area. Starting at 
00:30 h local time each day, individual 4?4 km cells contained within Brazilian 
Amazonia were monitored for the occurrence of fire detections produced by MOD14, 
MYD14 and WF-ABBA until 24:00 h local time (Figure 5.1); all hours reported 
hereafter are considered local time. 
 
 120 
 
 
 
 
 
Figure 5.1: Flow diagram describing the processing scheme used to integrate MOD14, 
MYD14 and WF-ABBA fire detections. Ingestion of WF-ABBA and cloud obscuration 
omission error estimates occurs at 30 min intervals during day and nighttime periods of every 
24 h cycle. 
 
 
Daytime fire detections were adjusted for commission errors based on the 
percentage tree cover values of the affected 4?4 km cell following the error matrix 
analysis presented in Chapter 3. Percentage tree cover estimates were derived using 
the 2005 500 m Vegetation Continuous Fields (VCF) product [Hansen et al., 2002; 
Hansen et al., 2003]. Commission error rates at the time of Aqua overpass were 
assumed 50% lower compared to the time of Terra overpass based on the results from 
Section 3.4.3. Commission error rates for the WF-ABBA product at observation 
 
 121 
 
hours different from the Terra and Aqua overpass times analyzed in Chapter 3 were 
estimated using information describing the approximate diurnal cycle of brightness 
temperature of forest and non-forested pixels in the mid-infrared spectral region 
(Section 3.4.3, Figure 3.9). Large differences in brightness temperature values 
between forest and non-forested pixels were found to be an important factor leading 
to false detections in the MODIS and GOES active fire detection products validated 
in Chapter 3. Using the false detection and brightness temperature data shown in 
Figure 3.9, a second order polynomial was fitted to the curve representing the 
brightness temperature contrast between forest and non-forested GOES pixels and 
normalized. This was done in order to have the commission error rate at 13:30 h 
being 50% lower than at 10:30 h, with a peak (value equal to 1) centered at 
approximately 11:45 h. The resulting curve intercepts the zero commission error mark 
at approximately 8:30 h and 14:30 h; i.e., commission errors are considered null for 
GOES images acquired earlier than 8:30 h or later than 14:30 h. This approximation 
may be revised in the future provided that improved commission error estimates 
become available. 
For deriving the integrated product, when multiple daytime WF-ABBA fire 
detections occur for an individual 4?4 km cell, we used the lowest commission error 
estimated from all the observation hours producing detections. Based on the 
assumption that false detections from different products do not overlap, grid cells 
showing multiple daytime fire detections originated from more than one fire detection 
product are confirmed to have vegetation fire activity for the calendar date in 
question. 
 
 122 
 
Omission errors due to cloud obscuration estimated at 30 min intervals for 
individual 4?4 km cells are accumulated for each 24 hour cycle, resulting in a 0-
100% chance of fire omission at the end of each calendar date. 
Lastly, the occurrence of any nighttime fire detection overrules any of the 
previous assumptions, leading to confirmed fire activity for a 4?4 km cell irrespective 
of the product which originated the entry. 
 
5.3 RESULTS 
The integration analysis used data from 2005 for which the cloud obscuration 
omission error estimates derived in Chapter 2 were available. The 2005 MOD14, 
MYD14, WF-ABBA, and integrated fire detection density maps re-sampled to a 40 
km grid are shown in Figure 5.2A-D, respectively. Although the spatial extent of 
most features depicted by the original MOD14, MYD14 and WF-ABBA fire 
detection maps are similar, the integrated product helped highlight areas of increased 
fire activity across the region. Active deforestation areas including southeastern Par?, 
central Mato Grosso, and northern Rond?nia resulted in the largest fire detection 
densities in Brazilian Amazonia. Areas typically associated with low percentage tree 
cover (east and south regions on the map) also showed large fire detection densities in 
the integrated product suggesting that fire activity in those areas could be 
underestimated by the original fire products. The relatively small fire detection 
densities depicted by the MOD14 product in Figure 5.2A show that a strong fire 
diurnal cycle could significantly limit the detection rates of sensors flying on morning 
orbits. 
 
 123 
 
 
 
 
 
Figure 5.2: Annual (2005) fire detection density maps for the original MOD14 (A), MYD14 
(B), and WF-ABBA (C) products, and the integrated product (D) for Brazilian Amazonia. 
The scales represent the average number of days with detections calculated for individual 40 
km cells. 
 
 
 The differences in detections between the integrated product and the simple 
sum of fire products are shown in Figure 5.3. Negative and positive values result 
primarily from corrections applied to the integrated product to reduce commission 
errors and omission errors due to cloud obscuration, respectively. The result suggests 
that large areas of Brazilian Amazonia showing small differences between the 
A B
C 
D
 
 124 
 
integrated product and the simple sum of fire products could have those errors 
balanced. However, small regions including central Mato Grosso state could be under 
the influence of a systematic bias resulting in an overestimation of fire activity. In 
fact, high fire detection rates were observed in the integrated product over certain 
locations described by individual 4?4 km sub-cells. In those areas, potentially 
unrealistic annual detection rates exceeding 30 days of co-located burning activity 
suggested that the bias due to commission errors could be even larger than is depicted 
in Figure 5.3. While the number of individual locations with high fire detection rates 
was relatively small (approximately 1,000 sites compared to 340,000 cells contained 
in region), the persistent burning signal which results from those areas could have 
important implications for both fire alert systems and biomass burning emissions 
modeling.  
 
Figure 5.3: Net correction applied to the combined detections of MOD14, MYD14 and WF-
ABBA based primarily on commission error rates and omission errors due to cloud 
obscuration. The scale represent the average number of days with detection that were added 
(positive values) or subtracted (negative values) from the integrated data for individual 40 km 
cells compared to the simple sum of the products above. 
 
 
 125 
 
5.4 QUALITY ASSESSMENT 
A quality assessment of the results produced by the integrated product in central Mato 
Grosso was performed using two sets of Landsat 5 Thematic Mapper (TM) scenes at 
30 m resolution acquired 16 days apart (row/path: 227/68; first set acquired on July 
01 and July 17 2005; second set acquired on August 02 and August 18 2005). The 
Landsat 5 TM data were complemented by two collocated 20 m resolution China-
Brazil Earth Resources Satellite (CBERS) images acquired on July 13 2005 and 
August 08 2005. The selection of the central Mato Grosso region to assess the quality 
of the integrated product was motivated by two primary reasons: (i) results from 
Chapter 3 and from Section 5.3 above suggest that this area could have high 
commission errors; and (ii) the availability of Landsat 5 TM and CBERS images 
acquired within only a few days from each other provided a unique opportunity to 
generate detailed fire information for the area.  
To derive the reference burn area maps, the image data covering the 0.77-0.90 
?m region in each sensor were inter-calibrated in order to provide comparable top of 
the atmosphere reflectance measurements. In this spectral region, burned areas can be 
easily distinguished from the surface background as a result of a sharp reduction in 
the reflectance values measured over ash and char covered pixels. After resizing the 
CBERS images to 30 m resolution and registering them to the Landsat 5 TM data, 
image differences were computed and an empirical threshold was applied to classify 
new burn scars occurring between two consecutive acquisition dates. 
The reference burn scar data produced using the approach described above, 
supported the assessment of the integrated fire product generated for the same time 
 
 126 
 
periods separating two consecutive Landsat 5 TM and CBERS images. The 4?4 km 
cells showing burn scars of any size were identified and compared to the integrated 
fire detections produced using the same grid cell resolution and time interval. 
Despite the large number of fire detections appearing in the integrated product 
during the periods analyzed, the number of new burn scars was relatively small 
suggesting that potentially large commission errors could remain in the integrated fire 
product. However, inspection of active fire pixels identified using an empirical 
threshold applied to the 2.09-2.35 ?m channel of Landsat 5 TM showed that the large 
majority of the fires (> 90%) occurring in that region and time period did not produce 
a detectable burn scar. Visual interpretation of the active fires mapped indicated the 
prevalence of small areas of burning often associated with closely aligned fire lines 
typical of piled debris. Therefore, the use of higher resolution imagery acquired only 
a few days apart did not provide appropriate means to assess the quality of the 
integrated product at the critical area of central Mato Grosso. 
Alternatively, evaluation of the integrated product was attempted using the 30 
m active fire pixels mapped with the Landsat 5 TM data. The integrated product was 
processed for four individual calendar dates corresponding to the same acquisition 
dates of the Landsat 5 TM images selected. The main purpose of this analysis was to 
compare the fire detections produced by the original input fire products to the fire 
information resulting from the integrated product using the 30 m active fire data to 
validate the results. Only fire detections initiated at the approximate overpass time of 
Landsat 5 TM were considered (? 30 min tolerance); detections produced for those 
same locations during subsequent hours were also used for processing the integrated 
 
 127 
 
product. The time difference constraint limited the analysis to the WF-ABBA data 
only, since the Landsat 5 TM images in Amazonia overlap with areas without 
daytime MOD14 coverage for the same acquisition date. This lack of coverage results 
from the MODIS ground sampling characteristics near the equator which creates 
large spatial gaps [Wolfe et al., 2002]. The use of MYD14 was equally compromised 
due to the large differences in overpass time separating the orbits of the Landsat 5 and 
Aqua satellites.  
The WF-ABBA detections were classified as true positives and false positives 
depending on the presence or absence of coincident 30 m active fires, respectively. 
Similarly, the locations classified as fire in the integrated product were validated 
using the 30 m active fire data. Table 5.1 shows the classification of all fire locations 
indicated by WF-ABBA and the integrated product. Fractional numbers appearing in 
the columns representing the integrated product in Table 5.1 resulted primarily from 
the application of corrections to account for commission error rates in the WF-ABBA 
data ingested into the integrated product processing routine. Compared to the original 
WF-ABBA fire detection data, the integrated product resulted in a small decrease in 
false positives while preserving true positive detections nearly unchanged. This first 
evaluation suggests that the integrated product could improve the fire detection 
information of WF-ABBA although we recognize that due to the small sample used, a 
more comprehensive quality assessment of the integrated product should be in the 
subject for future study. 
 
 
 
 128 
 
Table 5.1: Evaluation of WF-ABBA detections and the integrated fire product data using 30 
m active fire data derived from Landsat 5 TM imagery. 
 
 True Positives  
(number of detections) 
False Positives 
(number of detections) 
WF-ABBA 28 16 
Integrated Product 27.8 14.1 
 
 
5.5 CONCLUSIONS 
A new integrated fire detection product based on the MOD14, MYD14 and WF-
ABBA products was presented in this Chapter. A set of judicious criteria was used to 
exploit the complementarity among those products and to increase the consistency of 
the detections represented in the integrated product. Areas of high detection 
frequency were highlighted in the integrated product. Those areas coincided with 
active deforestation regions reinforcing the concept that conversion fires are 
responsible for the largest detection frequencies in Brazilian Amazonia [Eva and 
Fritz, 2003; Morton et al., 2008]. The large differences among the detection rates 
derived using MODIS Terra and Aqua and GOES data reflected the importance of the 
fire diurnal cycle in the region and the need for more systematic high temporal 
resolution imaging of fire activity in Brazilian Amazonia. 
A complete quality assessment of the integrated fire product over areas 
including central Mato Grosso state could not be accomplished primarily due to the 
scarce availability of validation data and the absence of burn scars from these fires. 
 
 129 
 
Those two aspects limited the evaluation of the integrated fire product to the times 
near the acquisition hours of Landsat-equivalent data when fire activity is reduced.  
The occurrence of fires without a distinguishable burn scar in areas such as 
central Mato Grosso state raise some of the potential limitations involving burn scar 
mapping products using classification schemes based on visible and near-infrared 
bands (e.g., the MODIS burned area product [Roy et al., 2005]). It also highlights the 
importance of complementing those burn scar products with active fire data in order 
to achieve a more complete representation of biomass burning in the region.  
Although some problems undoubtedly remain in the integrated fire product 
described above, it improves upon single source and multisource products based on 
the simple accumulation of individual fire detections. The quantification of some of 
the primary sources of errors influencing the MOD14, MYD14 and WF-ABBA 
products and the relationships between them could provide invaluable information to 
assess the main implications for biomass burning emissions modeling in South 
America as well as to guide future improvements of the fire detection products. 
Consideration of the full diurnal cycle during processing of multiple fire 
detections required that improved validation data sets be used to fully assess the 
qualities and limitations of the new integrated product. The use of airborne systems 
providing higher frequency image acquisition may be necessary to comprehensively 
assess the diurnal cycle of commission errors and the performance of the integrated 
product.  
 
 130 
 
Chapter 6: Conclusions 
 
 
6.1 SYNTHESIS OF RESEARCH 
The demand for information on fire activity in Amazonia has significantly increased 
in the past two decades with the concern about the fate of the tropical rainforest 
[Nepstad et al., 1999b; Nepstad et al., 2001; Morton et al., 2006; Shukla et al., 1990]. 
Satellite active fire detection products represent the primary data source for 
Amazonian fire alert systems. These data are also of interest to the scientific 
community. However, little information is available about the performance of the 
current products and differences among data sets create additional problems for the 
user community. This dissertation was designed to quantify the errors and to assess 
the complementarity amongst different products. The main objective of the research 
was to improve the quantification of vegetation fire activity in Brazilian Amazonia 
through the use of multiple satellite remote sensing fire data. 
 
6.1.1 Omission Errors due to Cloud Obscuration 
Fire detection omission errors due to cloud obscuration in the WF-ABBA 
product derived from GOES-East imager data were quantified in Chapter 2. A pixel 
based probabilistic approach was proposed, using information on previous fire 
occurrence and precipitation. The methodology was implemented using data covering 
 
 131 
 
the entire diurnal cycle of fire activity and cloud occurrence. The results suggested 
that by ignoring land use patterns and rainfall regimes and by adopting generalized 
grid based analyses, previous studies using more simplistic approaches have 
overestimated the effects of cloud obscuration on fire detections. 
Fire occurrence was found to be influenced by regional precipitation regimes 
which limit fire use to a short period of time in the dry season, and was largely related 
to areas of rapid land cover change found in particular along the Arc of Deforestation. 
The assessment of the omission errors due to cloud obscuration used satellite 
precipitation data and WF-ABBA fire detection records from previous years to 
delineate the spatial and temporal characteristics of fire activity in the region. A cloud 
mask was generated for the GOES data and omission errors were estimated using a 4 
km grid to reproduce the nominal spatial resolution of the GOES imager. The 
methodology was proven successful, resulting in consistent cloud obscuration 
omission error estimates for the WF-ABBA fire product. 
 
6.1.2 Clear Sky Omission and Commission Errors 
Following the analyses of fire omission errors due to clouds, a comprehensive 
validation study was implemented to quantify clear sky omission and commission 
errors for the WF-ABBA and the MOD14 products over Amazonia. The analyses 
discussed in Chapter 3 represent the first in depth assessment of the fire detection 
performance of the WF-ABBA product using real time validation data. The validation 
approach built on the work of Morisette et al. (2005a) who first tested the application 
of 30 m resolution ASTER imagery to assess the detection performance of the 
 
 132 
 
MODIS Fire product (MOD14) for Southern Africa. That methodology was expanded 
to incorporate the GOES data into the analyses and a new active fire detection 
product was generated using Landsat ETM+ imagery. As with the ASTER data, the 
30 m resolution Landsat ETM+ fire product was used as ?ground truth? data to 
validate the WF-ABBA fire detections. This validation approach fulfills the 
requirements for independent data usage to assess the quality of remote sensing 
satellite products defined by Morisette et al. [2002]. Compared to ground validation, 
higher resolution imagery provide more appropriate means to tackle related scaling 
issues of fine-scale processes in moderate resolution data, and in this respect the use 
of 30 m Landsat ETM+ and ASTER data to validate MODIS and GOES products is 
recognized as a viable alternative [Morisette et al., 2002]. To assure the quality of fire 
masks derived from the Landsat ETM+ and ASTER, an extensive quality control was 
performed including detailed visual inspection of all images used. Errors in the 
Landsat ETM+ and ASTER fire mask data were therefore considered negligible. 
Also, comparison of near coincident Landsat ETM+ and ASTER fire masks showed 
that the two data sets produced similar results and therefore could be used 
interchangeably. Sum statistics of 30 m active fire pixels derived from ASTER and 
ETM+ were produced for the footprint of individual GOES pixels and used to fit a 
logistic regression model describing the detection probability curve for the WF-
ABBA product. 
The differences in detection performance for MOD14 and the instantaneous 
WF-ABBA product were systematically validated for different surface conditions. On 
average, WF-ABBA required four times more active fire area than MOD14 to 
 
 133 
 
achieve equivalent detection probability rates for similar fire conditions. The distance 
separating the detection probability curves of MOD14 and WF-ABBA derived from 
the logistic regression analysis was proportionally less than the difference in spatial 
resolution of the two instruments. In this case, other sensor characteristics including 
the spectral location and bandwidth of the mid-infrared GOES channel could have 
contributed to enhance the response of the WF-ABBA product to sub-pixel fires, 
thereby reducing the difference in detection performance relative to the MOD14 
product. 
Omission error rates estimated for the approximate time of the ASTER and 
ETM+ overpasses (i.e., 10:00-10:30 local time) were high for both WF-ABBA and 
MOD14 products when fires of all sizes detected with the 30 m resolution fire masks 
were considered. In particular, high omission errors were seen to affect the MOD14 
and the WF-ABBA products in areas of low percentage tree cover. Major differences 
in fire conditions across different land uses and land cover types were found through 
the analysis of data collected from multiple prescribed burns during three different 
field campaigns in Brazilian Amazonia. Maintenance fires (e.g., pasture and crop 
residue burning) predominating in areas of low percentage tree cover were 
characterized by lower temperature (~ 700 K) and narrow flaming fronts, followed by 
a relatively cool smoldering phase. These highly dynamic conditions result in a 
significantly small window of time during which fire detection is possible. The 
surrounding background conditions associated with those fires was also found to be 
an important factor influencing detection probability of contextual algorithms based 
on moderate or coarse resolution sensor data. The occurrence of warm and bright 
 
 134 
 
background pixels in low percentage tree cover regions resulted in reduced contrast 
with the fire affected pixel, therefore causing the latter to fail the contextual tests. 
Consequently, detection performance of moderate (500 -1 km) and coarse resolution 
(> 1 km) fire products over low percentage tree cover conditions could be considered 
limited, being subjected to potentially large omission errors. 
Vegetation fires used for clearing forest debris along biomass-rich 
deforestation areas showed different characteristics compared to the fires occurring in 
low percentage tree cover regions and omission error rates were consequently 
smaller. Forest conversion fires burned at very high temperatures (~ 1100 K) and for 
a prolonged time (hours as opposed to minutes), significantly increasing the chances 
of detection by one or more instruments. 
The high temporal resolution of the GOES imager proved very useful to 
reduce the instantaneous omission error rates derived for the WF-ABBA product 
using coincident ASTER and ETM+ data. In fact, when all daytime GOES imager 
data were considered the estimated WF-ABBA omission error was half of that found 
for the instantaneous MOD14 product based on a similar fire sample. This finding 
reinforces the importance of fire diurnal cycle in the region, suggesting that fire 
activity tends to intensify towards the afternoon hours enabling the detection of an 
increased number of fire pixels by the WF-ABBA product. 
Similarities among the commission error rates estimated for WF-ABBA and 
MOD14 showed a tendency of contextual algorithms to fail over areas characterized 
by large gradient in percentage tree cover. False detections were particularly prone to 
occur along areas of active deforestation in Brazilian Amazonia, resulting from the 
 
 135 
 
strong brightness temperature contrast in the mid-infrared region established between 
recently burned deforestation sites and the surrounding forests. Multiple false 
detections persisting over the same location suggested that the radiometric signature 
of some burn scars could remain unaltered for as many as 32 days, the maximum 
separation found for spatially coincident false detections in the MOD14 data. 
Smaller differences in thermal inertia between forested and deforested areas 
were largely responsible for a 50% reduction in commission errors during the early 
afternoon hours when fire activity approaches its maximum. This decrease in 
commission error rates should enhance the character of the diurnal cycle of fire 
detections even further provided that false alarms occurring during mid-to-late 
morning hours are accounted for. Fire-unrelated commission errors (i.e., pixels with 
no indication of active burning or burn scars) accounted for only 3% of all fire 
detections produced by WF-ABBA and MOD14. Therefore, active fire detection 
products could provide useful information for burned area mapping algorithms by 
identifying areas of fire occurrence with low commission error rates. 
Commission errors in MOD14 could not be reduced using the additional 
scientific data available for the detection pixel and its surroundings provided in the 
product. The data include the brightness temperature in the mid-infrared and thermal 
infrared channels for the target pixel and its background, detection confidence 
estimate, and fire radiative power derived for the fire pixel. Due primarily to the 
ambiguity between the conditions describing true and false detections based on the 
parameters above, post-processing of commission errors was not possible. However, 
refinement of both MODIS Thermal Anomalies and WF-ABBA products could be 
 
 136 
 
accomplished through the adoption of additional tests to help differentiate true and 
false detections. For instance, a method to reduce commission error rates was tested 
using temporal metrics based on the brightness temperature calculated for the mid-
infrared channels of GOES and MODIS. The application of such method to a subset 
of the validation data proved successful in separating true and false detections and 
thereby represents a potential area for future development of active fire detection 
products. Regions of tropical deforestation where high commission error rates (> 
30%) predominate are particularly suited to the application of a fire detection 
algorithm which incorporates the use of temporal metrics in the mid-infrared region. 
In those areas, surface temperature conditions remain mostly invariant for prolonged 
times favoring the use of change detection schemes to capture anomalous variations 
in pixel brightness temperature values between images that can be caused by 
vegetation fires. 
The clear sky omission and commission errors reported in Chapter 3 were 
largely a function of the fire use and land cover types found in Brazilian Amazonia. 
High commission error rates associated with recently burned forest clearings of 
approximately the same size of a MODIS or GOES pixel footprint could be specific 
to the Arc of Deforestation and less common elsewhere.  Despite the commission 
problem identified in this study, MODIS and GOES fire detection data provide 
invaluable information for the Amazon fire data user communities. Basin wide 
commission errors were approximately 15% for both MODIS and GOES fire 
products, with commission problem almost exclusively (> 80%) associated with 
recent fire activity.  The estimated fire-unrelated commission error rates of 3% agreed 
 
 137 
 
with the results of Csiszar et al. [2006] and Schroeder et al. [2008c] who derived 
clear sky commission and omission errors for the MODIS Terra Thermal Anomalies 
product over Northern Eurasia and the conterminous United States, respectively. This 
result suggests that a global commission error rate for the MODIS fire product may 
be ~3%.  
Additional fire product error assessment remains an important topic for 
regional studies. Priority regions for validation analyses include parts of Africa and 
southeast Asia where high fire detection frequency and intensive land use and land 
cover change similar to Amazonia may result in comparable commission issues. 
 
6.1.3 Towards Integrating MODIS and GOES Fire Detection Data 
With the characterization of omission and commission errors completed, Chapter 4 
discussed the potential for integrating MODIS and GOES active fire data. The use of 
high temporal resolution GOES data combined with the moderate spatial resolution of 
MODIS created an optimum environment to test how fire products generated from 
sensors with different characteristics complement each other. 
The spatial and temporal characteristics of Amazonian fires were first 
considered in order to determine the most appropriate approach to use for the 
integration analyses. Fire size statistics derived from 30 m resolution ASTER and 
ETM+ data suggested that nearly 90% of the fires actively burning at 10:00 local time 
were composed of fire clusters with less than 30 pixels, or the equivalent to 
approximately 3 ha of a variable mix of flaming and smoldering fire components and 
unburned background vegetation. Fire size statistics demonstrated how vegetation 
 
 138 
 
fires in Amazonia are small in size. In most cases fires are used in a controlled 
fashion therefore being constrained to small subsets of individual properties where 
the land is being managed. Consequently, Amazonian fires usually occupy only a 
reduced fraction of a MODIS (<3%) or GOES (<0.2%) pixel. 
The communal use of fires involving simultaneous burning in neighboring 
properties which predominates in various parts of Amazonia was clearly depicted by 
the statistics describing the distance separating adjacent active fires. Approximately 
80% of the fires showed a second area of burning within less than 8 km. This small 
separation is equivalent to just about two adjacent GOES pixels, and is similar to the 
navigation error of the GOES data. 
 The small size of vegetation fires in Amazonia and the proximity between 
areas of simultaneous burning created major limitations for the use of any kind of 
pixel level analysis to integrate the MODIS and GOES fire detection products since 
overlapping pixels would likely contain different amounts of fire activity. To 
overcome this constraint, the effects associated with a high fire spatial frequency 
were minimized by aggregating adjacent MODIS or GOES fire detections into 
clusters. 
 The integration analysis tested the use of Fire Radiative Power (FRP) to 
integrate the MODIS and GOES fire data. FRP is a physical parameter describing the 
energy released primarily by the flaming components of vegetation fires; therefore it 
facilitates the comparison between different products while also being useful to 
estimate biomass burning combustion rates [Kaufman et al., 1998b; Wooster, 2002; 
Wooster et al., 2003]. The analysis involving MODIS and GOES FRP estimates was 
 
 139 
 
implemented using a conservative approach to reduce artifact contamination of the 
results, including the adoption of strict spatial criteria and removal of data points 
containing saturated pixels. Spatially and temporally coincident MODIS and GOES 
detection clusters produced significantly different FRP estimates for a large fraction 
of the cases analyzed. The results suggested that due to the conditions describing 
Amazonian fires, FRP estimates derived from moderate and coarse resolution data 
could be systematically subjected to large and variable errors, making it difficult to 
relate MODIS Thermal Anomalies and GOES WF-ABBA data via this parameter. 
The divergence between MODIS and GOES FRP estimates was particularly large in 
areas of high percentage tree cover, providing indications that landscape 
heterogeneity is an important factor limiting accurate FRP retrievals in Amazonia. 
Previous studies have almost entirely neglected the effects of Point Spread 
Function (PSF) on the calculation of FRP of MODIS and GOES. However, the 
simulation analysis implemented in Chapter 4 suggested that PSF could represent the 
largest single source of error on FRP estimates derived from those sensors. The 
results revealed that fires situated away from the more responsive area of the PSF 
located near the pixel center can have their energy underestimated by the FRP 
estimates produced. The simulations of MODIS and GOES pixels containing a 
typically small Amazonian fire were prone to significantly underestimate true FRP 
values due to the effects of PSF of those instruments. This result suggests that the 
potential for using FRP for biomass emissions modelling applications in Amazonia 
using MODIS or GOES data might be limited, since emission rates calculated from 
underestimated FRP values could be negatively biased. 
 
 140 
 
The results produced by the analyses using MOD14 (MYD14) and WF-
ABBA fire detection clusters suggested a strong influence of the diurnal cycle of fires 
on the rates of spatially and temporally coincident detections. Relative to MOD14, a 
larger number of MYD14 fire detections showed coincident WF-ABBA detections 
indicating that greater fire activity predominates during the afternoon hours. Inclusion 
of the daytime GOES data in the analysis significantly increased the rate of MOD14 
(MYD14) and WF-ABBA coincident detections. This result suggests that an 
important fraction of the smaller fires only detected by MOD14 and MYD14 could 
show variations in size and intensity during their life cycle, which evidently based on 
the temperature profiles shown in Chapter 3 (Figure 3.3) extends beyond the 
acquisition times of MODIS Terra and Aqua. 
The fraction of MOD14 and MYD14 detections without a corresponding WF-
ABBA detection occurring over high percentage tree cover areas was similar to the 
commission error rates derived for those MODIS products. This result suggests that 
MOD14 and MYD14 detections without a WF-ABBA match occurring in those areas 
could be primarily associated with false detections. Therefore, use of WF-ABBA 
detections to describe fire activity in high percentage tree cover areas (i.e., > 40%) 
could provide similar detection rates for most (> 90%) fires observed by MOD14 and 
MYD14 without the limitations of spatial coverage which create gaps in the MODIS 
data along the equator. Nonetheless, false detections in the WF-ABBA data need to 
be corrected for in order to reduce the commission errors observed over those high 
tree cover areas. 
 
 141 
 
In areas characterized by low percentage tree cover (i.e., < 40%), small rates 
of coincident detections for MOD14 (MYD14) and WF-ABBA resulted as a 
consequence of lower temperature and short duration fires influencing the detection 
probability of those products. Nonetheless, commission error rates were significantly 
reduced in low percentage tree cover areas therefore the combined use of MOD14, 
MYD14 and WF-ABBA should provide the highest detection rates for those areas. 
Large daily variations in detection performance of MOD14 (MYD14) and 
WF-ABBA could be verified using daily fire detection sum statistics from those 
products for a fixed area centered in southern Brazilian Amazonia. The variable pixel 
geometry of MODIS as a function of scan angle resulted on average in a 40% 
reduction of fire detection rates at the edge of the swath (i.e., ?55?) compared to nadir 
conditions. Differences between area averaged MOD14 (MYD14) and WF-ABBA 
fire detection statistics were particularly large over extreme MODIS scan angles 
where the bow tie effect in the MODIS imagery and the longer atmospheric path 
which cause greater attenuation of the fire emitted radiation could introduce 
additional variations in detection rates. Consequently, the use of area averaged 
statistics to integrate WF-ABBA, MOD14 and MYD14 detections was considered 
inadequate due to the large uncertainties involved. 
 
6.1.4 Integrated Fire Product 
In Chapter 5 a new integrated fire detection product was derived based on MOD14, 
MYD14 and WFABBA 2005 data. The results from Chapters 2-4 were used to 
generate a set of judicious criteria which takes advantage of multiple detections 
 
 142 
 
produced from daytime and nighttime MODIS and GOES imagery. These criteria 
were applied to the data to generate improved daily fire information for Brazilian 
Amazonia. The integrated data product was generated using a 4?4 km grid at 30 min 
temporal resolution. 
 Areas of high detection frequency were highlighted in the integrated product 
reinforcing the strong association between forest conversion and areas of high fire 
activity in Brazilian Amazonia. The low overall fire detection densities generated 
from MODIS Terra data indicated that a strong fire diurnal cycle prevails in most of 
the region, whereas the high fire detection frequencies produced with the GOES data 
suggested that higher temporal resolution instruments may offer great advantage for 
Amazonian fire monitoring. 
 A validation exercise designed to assess the quality of the integrated product 
near central Mato Grosso state, an area under the influence of potentially large 
commission errors, revealed that numerous active fires in that region could result 
from the burning of crop residue or piled forest debris which leave no distinguishable 
burn scars on the surface. Consequently, future validation of the integrated detections 
must rely on coincident ground reference data to permit proper characterization of the 
product?s quality. 
 
6.2 IMPLICATIONS FOR SATELLITE-BASED FIRE MONITORING IN 
BRAZILIAN AMAZONIA 
Amazonian fire alert systems and biomass burning emissions models demand high 
quality remote sensing data at the highest temporal and spatial resolution available, 
 
 143 
 
requiring that active fire detection products from different sensors be combined. In 
order to integrate different satellite active fire detection data sets, the performance of 
individual products must be fully characterized and omission and commission errors 
quantified. 
 The main objective of this dissertation was to assess the potential for 
integrating the MOD14, MYD14 and WF-ABBA fire detections products in order to 
improve fire monitoring in Amazonia. Those products are routinely generated for the 
region providing important fire information for a large number of users. The WF-
ABBA product derived from the GOES data has been running operationally since 
1995, representing one of the longest time series of fire detection data available for 
Amazonia. The WF-ABBA product generates the highest temporal resolution fire 
detection data available for the region, serving rapid response fire alert systems as 
well as biomass burning emissions models [CPTEC, 2008; Freitas et al., 2005]. 
Despite their shorter time series, the MOD14 and MYD14 products quickly gained a 
large number of users in Amazonia as they provide moderate resolution fire detection 
data with lower saturation levels and improved navigation quality compared to other 
polar orbiting sensors such as the Advanced Very High Resolution Radiometer 
(AVHRR) series. 
For the first time ever cloud obscuration omission errors and clear sky 
omission and commission errors were quantified in detail for the MOD14, MYD14 
and WF-ABBA products over Amazonia. The results suggested that additional data 
layers such as vegetation, precipitation and cloud mask products must be used in 
conjunction with the active fire data in order to reduce errors in those products. 
 
 144 
 
One of the big advantages of the method developed to quantify the cloud 
obscuration omission errors consisted in using GOES imager data to derive all input 
layers required to calculate the omission errors, including: a cloud mask, precipitation 
estimates and the land use data. In doing so, pixel level processing could be 
implemented and data registration errors and differences in data acquisition could be 
minimized. The use of pixel based processing is also advantageous as it facilitates 
aggregation of the omission error estimates for any particular grid cell resolution of 
current biomass burning emissions models. In fact, the methodology described in 
Chapter 2 will be used to generate fire detection omission error estimates due to cloud 
obscuration for the Coupled Aerosol and Tracer model to the Brazilian developments 
on the Regional Atmospheric Modelling System (CATT-BRAMS) dedicated to 
model biomass burning emissions in South America. 
The large number of false detections found in areas of high percentage tree 
cover demonstrated the importance of using land cover information to estimate 
commission error rates. The use of the 500 m Vegetation Continuous Fields data 
[Hansen et al., 2002; Hansen et al., 2003] in combination with the fire detection data 
proved particularly successful in this study as it allowed easy determination of the 
dependency of commission error rates on percentage tree cover. Based on the 
commission error rates derived in Chapter 3, the approximate number of MOD14 
(MYD14) and WF-ABBA false detections can be determined by weighting the 
distribution of fire detections as a function of percentage tree cover for any given 
area. The use of alternative land cover maps is contingent on users? preferences and 
 
 145 
 
on the potential of those maps to reproduce a continuum of vegetation classes that can 
be easily related to commission error rates. 
The potential for improving current contextual fire detection algorithms with 
the use of additional tests based on mid-infrared brightness temperature data to reduce 
commission errors was also demonstrated in Chapter 3. The results suggested that the 
stable surface thermodynamic conditions of Amazonia could facilitate the use of pixel 
based temporal metrics of brightness temperature estimates derived from the mid-
infrared channels of MODIS and GOES to identify false detections or increase our 
confidence about the true ones. In Section 3.4.4, 30-day averaged brightness 
temperature estimates derived for the pixel were used to detect changes in 
instantaneous brightness temperature measurements that were typically associated 
with true or false detections. 
The high clear sky omission error rates found for MOD14 and WF-ABBA 
have important implications for early warning fire alert systems, as the low detection 
probabilities of small fires will likely increase the latency of those systems. For other 
applications such as biomass burning emissions modeling which prioritize the 
detection information for larger and potentially more important fires with respect to 
their emissions, the relatively rapid decrease in omission errors with the increase in 
active fire area results in improved detection performance. 
Errors in the FRP estimates from MODIS and GOES suggest that the use of 
such parameter for modeling of biomass burning emissions in Amazonia could be 
significantly compromised. The increased uncertainty found over higher percentage 
tree cover areas is a major limitation to the use of FRP. Because FRP is being 
 
 146 
 
considered as a means to estimate biomass combustion rates [Kaufman et al., 1998b; 
Wooster, 2002; Wooster et al., 2003], even small errors in the total FRP value 
estimated for a fire detection cluster in high percentage tree cover areas will likely 
introduce large errors in the emission rates produced due to the high rates of biomass 
combustion involved. 
Consideration of land cover conditions was proven essential for the 
integration of fire detection data from MOD14 (MYD14) and WF-ABBA. Regions 
characterized by high percentage tree cover including the core of Amazonian forests 
and the surrounding areas of active deforestation must be treated separately from 
other regions where low percentage tree cover vegetation types such as the Cerrado 
(savanna) predominate. By doing so, the complementarities among MOD (MYD14) 
and WF-ABBA fire detections can be better assessed and commission errors properly 
accounted for. 
Finally, the close proximity between simultaneous active fires and 
consequently between burnt areas suggest that any validation analysis of active fire 
detection in Amazonia must use coincident ground truth information such as provided 
by ASTER and ETM+ data. This consideration significantly reduces the potential for 
validation of fire detection products derived from sensors flying on afternoon orbits. 
In this case, higher resolution data from sensors such as ASTER or ETM+ are not 
available and the potential for airborne and field data collection is limited as costs 
tend to be prohibitive. Greater coordination of the scientific community will be 
required to enable proper assessment of fire detection data sets without coincident 
 
 147 
 
higher resolution spaceborne imaging in order to leverage costs and optimize the 
results of future validation campaigns. 
 This study targeted two main Amazonian fire data user communities: 1) 
regional fire monitoring and management, and 2) regional biomass burning emissions 
and smoke transport modeling. Both groups require accurate and timely fire detection 
information.  Different tolerance for omission and commission errors between 
communities suggests that the results of this study apply differently to the policy, 
management, and science objectives users.  Key limitations involving primarily the 
spatial resolution of GOES and MODIS and the geolocation accuracy of the GOES 
imager can significantly reduce the capacity of the related fire detection products to 
resolve individual fire clusters and to quantify their properties. Nonetheless, other fire 
data user groups including decision makers and climate modelers usually show less 
stringent data requirements. In Table 6.1 user requirements derived from the 
implementation plan for the Global Climate Observing System [GCOS, 2006], from 
the report on the Committee on Earth Observation Satellites (CEOS) Land Product 
Validation  (LPV) working group workshop on global geostationary fire monitoring 
applications [GOFC 2004], from current regional and global emissions models 
[Freitas et al., 2005; van der Werf et al., 2006], and from personal communication 
with Amazonian fire managers are presented. While most users may benefit from the 
integrated fire product presented in Chapter 5, fire alert systems in Amazonia usually 
have low tolerance to commission errors as resources designated to law enforcement 
and field inspection tend to be limited. Consequently, fire alert systems might gain 
additional benefits from using regionally adjusted fire detection algorithms that lower 
 
 148 
 
false detection rates to complement the information provided by the hemispheric- to 
global-scale GOES and MODIS algorithms used in this study. 
 
 
Table 6.1: Fire data requirements applicable for five major fire data user groups. 
 
User Group & 
Source 
Spatial 
Resolution 
Temporal 
Resolution 
Geolocation 
Accuracy 
Tolerance to 
omission & 
commission 
errors  
Fire Alert Systems 
GOFC [2004] 
500 m 5 min 
Not specified; 
assumed equal or 
better than spatial 
resolution  
(500 m) 
Low 
    
Resource Managers 
Personal 
communication 
County level Daily ~ 500 m Low 
   
Regional emissions 
modeling 
Freitas et al. [2005] 
Limited to a 
few tens of 
kilometers 
Daily 
Dependent on spatial 
resolution of land 
cover classification 
used 
(500 m - 1 km) 
Moderate 
     
Global scale 
emissions modeling 
van der Werf et 
al.[2006] 
Gridded data 
of 
0.25??0.25? 
to 
1??1? 
resolution 
Weekly to 
monthly 
Dependent on spatial 
resolution of land 
cover classification 
used  
(500 m < grid size) 
High 
     
Global Climate 
change community 
GCOS [2006] 
250 m Daily 
Not specified; 
assumed equal or 
better than spatial 
resolution  
(250 m) 
Moderate 
 
 
 
 149 
 
6.3 FUTURE DEVELOPMENTS 
The analyses discussed in this dissertation represent the first detailed assessment of 
two of the major active fire detection products currently available for Brazilian 
Amazonia. Consideration of commission and omission errors and product inter-
comparison discussed in Chapters 2-4 should provide the basis for future 
development of fire detection algorithms and integration of different products in 
tropical areas. 
The quantification of omission errors due to cloud obscuration will be used to 
create a 10+ years time series of WF-ABBA fire detections for Brazilian Amazonia as 
part of the Large Scale Biosphere-Atmosphere Experiment in Amazonia (LBA). 
Improvements in WF-ABBA fire data information are already being implemented for 
the new version 6.5 product, including an opaque cloud mask layer which could allow 
direct derivation of cloud obscuration omission errors using the methodology 
described in Chapter 2. 
The next generation of sensors which will replace the existing GOES imager 
series and the MODIS sensors should provide enhanced capabilities for active fire 
monitoring. For instance, the improved geolocation accuracy and higher spatial 
resolution of the Advanced Baseline Imager (ABI) which will fly on board the 
GOES-R series and the reduced along scan pixel size variation of the Visible Infrared 
Imaging Radiometer Suite (VIIRS) which will fly on the upcoming National Polar-
orbiting Operational Environmental Satellite System (NPOESS) could facilitate the 
application of tests based on temporal change detection in addition to the spatial 
contextual tests to identify potential false detections in the fire products derived from 
 
 150 
 
those data. Consequently, problem areas such as high percentage tree cover regions 
where high commission errors are currently observed could benefit from the use of 
more comprehensive set of detection tests. Similarly, derivation of FRP should 
benefit from higher spatial resolution imagery as average Amazonian fires will tend 
to occupy a considerably larger fraction of the higher spatial resolution pixel 
footprints reducing the PSF effects on those estimates. The creation of temporal 
metrics aimed for active fire detection may also be advantageous to estimate the fire 
background conditions necessary to derive FRP values, since average brightness 
temperature conditions may be estimated for the fire affected pixel reducing the need 
to use information from radiometric distinct adjacent pixels. 
The need for continued validation of current and future remote sensing active 
fire detection products requires that higher resolution data be available for the 
scientific community. In this respect, changes in data sharing policy to be 
implemented by the United States Geological Survey (USGS) should facilitate 
significantly the access to higher resolution imagery from the Landsat mission and the 
upcoming Landsat Data Continuity Mission (LDCM). The availability of cost free 
data should make validation exercises more frequent contributing to our 
understanding of the qualities and limitations of fire detection products over broader 
spatial scales and fire regimes. Thanks to this new data policy, the use of Landsat data 
to validate the GOES fire product will be implemented at the Satellite Analysis 
Branch (SAB) of the National Environmental Satellite, Data and Information Service 
(NOAA/NESDIS) in Camp Springs, Maryland. Similar plans are being proposed to 
validate MODIS Terra Thermal Anomalies product at far off-nadir viewing angles 
 
 151 
 
using Landsat 5 data for the Conterminous United States. The use of additional 
validation data could help identify the major factors leading to the poor overall rate of 
coincident detections between WF-ABBA and MODIS Thermal Anomalies as 
presented in Section 4.4.2. 
Finally, the use of the data integration scheme presented in Chapter 5 should 
provide a valuable test-bed to assess the implications of fire detection errors for 
biomass burning emissions modeling as well as for operational fire monitoring in 
Brazilian Amazonia and elsewhere. Assimilation of the results presented in Chapter 5 
into the CATT-BRAMS model will be addressed and is expected to complement the 
10+ years fire detection data series being created for Brazilian Amazonia with a 
biomass burning emissions inventory. 
 
 
 152 
 
Appendix A 
 
ETM+ Active Fire Mask 
 
In order to improve our sampling capacity of the GOES data, a new 30 m fire mask 
product was created using Landsat 7 Enhanced Thematic Mapper Plus (ETM+) 
imagery. Due to similarities in spectral and spatial resolution between the ASTER 
and ETM+ data, the new ETM+ fire mask product builds on the Giglio et al. [2008] 
algorithm used for producing the ASTER product. Denoting the ETM+ channel 4 and 
7 top-of-atmosphere reflectance as ?
4
 and ?
7
, respectively, unambiguous fire pixels 
are classified using the criteria: 
3.0
5.2
4774
4
7
74
>?=
>=
??
?
?
D
R
 
 
Next, candidate pixels are selected based on the following tests: 
 
17.0
8.1
74
74
>
>
D
R
 
 
We apply a fixed 61?61 window centered on the candidate pixel to calculate the 
background statistics which will be used for the contextual part of the detection code. 
The mean and standard deviation are calculated for the reflectance ratio (
74
R  and 
74R
? , respectively) and for channel 7 reflectance (
7
? and 
7?
? , respectively). 
 
 153 
 
Unambiguous fire pixels and water pixels (pixels for which 04.0
7
<? ) are excluded 
from the computation. Candidate pixels which satisfy the following conditions are 
classified as fires: 
 
)08.0,3(
)8.0,3(
777
747474
?
???
?
Max
MaxRR
R
+>
+>
 
 
where the Max indicates that the maximum value is to be used. The numerical 
constants appearing in the tests were empirically derived for the scenes used in this 
study, and the resulting fire masks were visually inspected for consistency. 
If needed.  
 
 
 154 
 
 Bibliography 
 
Agresti, A. (1990). Categorical data analysis. John Wiley and Sons, 558 pp. 
 
Alencar, A., Nepstad, D., Silva, E., Brown, F. , Lefebvre, P., Mendosa, E.,  Almeida, 
D., and Carvalho Jr., O. (1997). O uso do fogo na Amaz?nia: Estudos de caso ao 
longo do arco de desmatamento. World Bank, Bras?lia, Brazil. 
 
Alencar, A., Nepstad, D., and Vera Diaz, M. C. (2006), Forest understory fire in the 
Brazilian Amazon in Enso and non-Enso years: Area burned and committed 
carbon emissions, Earth Interactions, 10, Paper 6. 
 
Alencar, A., Solorzano, L., and Nepstad, D. (2004). Modeling forest understory fires 
in an eastern Amazonian landscape. Ecological Application, 14 (4), S139-S149. 
 
Alval?, R. C. S., Gielow, R., da Rocha, H. R., Freitas, H. C., Lopes, J. M., Manzi, A. 
O., et al. (2002). Intradiurnal and seasonal variability of soil temperature, heat 
flux, soil moisture content, and thermal properties under forest and pasture in 
Rond?nia. Journal of Geophysical Research, 107, doi:10.1029/2001JD000599 
 
Andreae, M.O. (1991). Biomass burning: Its history, use and distribution and its 
impact on Environmental Quality and Global Climate. In J. Levine (Ed.), Global 
biomass burning, atmospheric, climatic, and biospheric implications, MIT press, 
3-21 pp. 
 
Andreae, M. O., and Merlet, P. (2001). Emission of trace gases and aerosols from 
biomass burning. Global Biogeochemical Cycles, 15 (4), 955?966. 
 
Andreae, M. O., Rosenfeld, D., Artaxo, P., Costa, A. A., Frank, G. P., Longo, K. M., 
and Silva-Dias, M. A. F. (2004). Smoking rain clouds over the Amazon. Science, 
303, 1337-1342, doi: 10.1126/science.1092779. 
 
Arima, E., Simmons, C. S., Walker, R., and Cochrane, M. A. (2007). Fire in the 
Brazilian Amazon: A spatially explicit model for policy impact analysis. Journal 
of Regional Sciences, 43 (3), 541-567. 
 
Arkin, P. A. (1979). The relationship between the fractional coverage of high cloud 
and rainfall accumulations during GATE over the B-array. Monthly Weather 
Review, 107, 1382?1387. 
 
Barbosa, P. M., Stroppiana, D., Gr?goire, J. -M., and Cardoso Pereira, J. M. (1999). 
An assessment of vegetation fire in Africa (1981?1991): Burned areas, burned 
biomass and atmospheric emissions. Global Biogeochemical Cycles, 13, 933?950. 
 
 
 155 
 
Barlow, J., and Peres, C. A. (2006). Effects of single and recurrent wildfires on fruit 
production and large vertebrate abundance in a central Amazonian forest. 
Biodiversity and Conservation, 15 (3), 985-1012. 
 
Barlow, J., Peres, C. A., Langan, B. O., and Haugaasen, T. (2003). Large tree 
mortality and the decline of forest biomass following Amazonian wildfires. 
Ecology Letters, 6 (1), 6-8. 
 
Bellerby, T., Todd,  M., Kniveton, D., and Kidd, C. (2001). Rainfall estimation from 
a combination of TRMM precipitation radar and GOES multispectral satellite 
imagery through the use of an artificial neural network. Journal of Applied 
Meteorology, 39, 2115?2128. 
 
Betts, A. K., Fuentes, J. D., Garstang, M., and Ball, J. H. (2002). Surface diurnal 
cycle and boundary layer structure over Rond?nia during the rainy season. Journal 
of Geophysical Research, 107 (D20), doi:10.1029/2001JD000356 
 
Boi, P., Marrocu, M., and Giachetti, A. (2004). Rainfall estimation from infrared data 
using an improved version of the Auto-Estimator Technique. International Journal 
of Remote Sensing, 10, 4657?4673. 
 
Boles, S. H., and Verbyla, D. L. (2000). Comparison of three AVHRR-based fire 
detection algorithms for interior Alaska. Remote Sensing of Environment, 72, 
1?16. 
 
Bond T. C., Streets, D. G., Yarber, K. F., Nelson, S. M., Woo, J.-H., and Klimont, Z. 
(2004). A technology-based global inventory of black and organic carbon 
emissions from combustion. Journal of Geophysical Research, 109, D14203, 
doi:10.1029/2003JD003697. 
 
Brown, I. F., Schroeder, W., Setzer, A., Maldonado, M. L. R., Pantoja, N., Duarte, A., 
et al. (2006). Monitoring fires in Southwestern Amazonia rain forests. Eos 
Transactions of the American Geophysical Union, 87(26), 253. 
 
Bruno, R. D., Rocha, H. R., de Freitas, H. C., Goulden, M. L., and Miller, S. (2006). 
Soil moisture dynamics in an eastern Amazonian tropical forest. Hydrological 
Processes, 20, 2477?2489. 
 
Bucini, G., and Lambin, E. (2002). Fire impacts on vegetation in Central Africa: A 
remote-sensing-based statistical analysis. Applied Geography, 22, 27?48. 
 
Cahoon, D. R. Jr., Stocks, B. J., Alexander, M. E., Baum, B. A., and Goldammer, J. 
G. (2000). Wildland fire detection from space: Theory and application. In Innes, J. 
L., Beniston, M., and Verstraete, M. M. (Eds.), Biomass burning and its inter-
relationships with the climate system. Kluwer Academic Publishers, Dordrecht and 
Boston, 151-169pp. 
 
 156 
 
 
Calle, A., Casanova, J. L., and Romo, A. (2006). Fire detection and monitoring using 
MSG Spinning Enhanced Visible and Infrared Imager (SEVIRI) data. Journal of 
Geophysical Research, 111, G04S06, doi:10.1029/2005JG000116. 
 
Cardoso, M. F., Hurtt, G. C., Moore, B., III, Nobre, C. A., and Bain, H. (2005). Field 
work and statistical analyses for enhanced interpretation of satellite data. Remote 
Sensing of Environment, 96, 212?227. 
 
Cardoso, M. F., Hurtt, G. C., Moore, B., III, Nobre, C. A., and Prins, E. M. (2003). 
Projecting future fire activity in Amazonia. Global Change Biology, 9, 656?669. 
 
Carmona-Moreno, C., Belward, A., Malingreau, J. -P., Hartley, A., Garcia-Alegre, 
M., Antonovskiy, M., et al. (2005). Characterizing interannual variations in global 
fire calendar using data from Earth observing satellites. Global Change Biology, 
11, 1537?1555. 
 
Cochrane, M., (2003), Fire science for rainforest. Nature, 421, 913-919. 
 
Cochrane, M., Alencar, A., Schulze, M., Souza Jr, C., Nepstad, D. C., Lefebvre, P., 
and Davidson, E. (1999). Positive feedbacks in the fire dynamic of closed canopy 
tropical forest. Science, 284, 1832-1835. 
 
Cochrane, M., and Laurance, W. F. (2002). Fire as a large-scale edge effect in 
Amazonia forests. Journal of Tropical Ecology, 18, 311-325. 
 
Cochrane, M. A., and Schulze, M. D. (1999). Fire as a recurrent event in tropical 
forests of the Eastern Amazon: Effects on forest structure, biomass, and species 
composition. Biotropica, 31 (1), 2-16. 
 
Comision Nacional de Actividades Espaciales (CONAE), Geston de Emergencias, 
Argentina, URL (last visited 24 April 2008): 
http://www.conae.gov.ar/WEB_Emergencias/Links_de_la_Izquierda/Incendios/inc
endios.html  
 
Costa, M. H., and Foley, J. A. (1998). A comparison of precipitation datasets for the 
Amazon basin. Geophysical Research Letters, 25, 155?158. 
 
CPC?Climate Prediction Center?National Center for Environmental Prediction 
(NCEP) ? National Weather Service (NWS), Global-merged Infra-red Brightness 
Temperature data set. URL (last visited 24 April 2008): 
http://disc.sci.gsfc.nasa.gov/data/datapool/TRMM/01_Data_Products/06_Ancillary
/01_Global_MERG_IR/index.html  
 
Centro de Previs?o do Tempo e Estudos Clim?ticos (CPTEC) Projeto Queimadas. 
URL (last visited 24 April 2008): http://cptec.inpe.br/queimadas 
 
 157 
 
 
Crutzen, P.J., (1979). The role of NO and NO2 in the chemistry of the troposphere 
and stratosphere. Annual Review of Earth and Planetary Science, 7, 443-472, 
doi:10.1146/annurev.ea.07.050179.002303. 
 
Crutzen, P. J., and Andreae, M. O. (1990). Biomass burning in the tropics: Impact on 
atmospheric chemistry and biogeochemical cycles. Science, 250, 1669-1678. 
 
Crutzen, P. J., Heidt, L. E., Krasnec, J. P., Pollock, W. H., and Seiler ,W. (1979). 
Biomass burning as a source of atmospherics gases: CO, H2, N2O, NO, CH3Cl, 
COS. Nature, 282, 253-256. 
 
Csiszar, I., Denis, L., Giglio, L., Justice, C. O., and Hewson, J. (2005). Global fire 
distribution from MODIS. International Journal of Wildland Fire, 14, 117?130. 
 
Csiszar, I., Morisette, J. T., and Giglio, L. (2006). Validation of active fire detection 
from moderate-resolution satellite sensors: the MODIS example in Northern 
Eurasia. IEEE Transactions on Geoscience and Remote Sensing, 44(7), 
1757?1764. 
 
Csiszar, I., and Schroeder, W. (2007). Fire observations from ETM+ and ASTER 
imagery and implications for active fire product validation from coarse resolution 
sensors. Proceedings of the 6th International Workshop of the EARSeL Special 
Interest Group on Forest Fires, 27?29 September 2007 Thessaloniki, Greece, 155-
159pp. 
 
da Rocha, H. R., Goulden, M. L.,Miller, S. D., Menton, M. C., Pinto, L. D. V. O., de 
Freitas, H. C., et al. (2004). Seasonality of water and heat fluxes over a tropical 
forest in eastern Amazonia. Ecological Applications, 14(4), S22?S32. 
 
Delany, A. C., Haagensen, P., Walters, S., Wartburg, A. F., and Crutzen, P.J. (1985). 
Photochemically produced ozone in the emission from large-scale tropical 
vegetation fires. Journal of Geophysical Research, 90, 2425-2429. 
 
Di Bella, C. M., Jobb?gy, E. G., Paruelo, J. M., and Pinnock, S. (2006). Continental 
fire density patterns in South America. Global Ecology and Biogeography, 15, 
192?199. 
 
Direccion de Manejo de Recusos Naturales (DIMARENA), Gobierno Departamental 
Prefectura Santa Cruz, Bolivia, URL (last visited 24 April 2008): 
http://servicios.santacruz.gov.bo/forestal/index.php?option=com_mamboezine&Ite
mid=36 
 
Dozier, J. (1981). A method for satellite identification of surface temperature fields of 
subpixel resolution. Remote Sensing of Environment, 11, 221-229. 
 
 
 158 
 
Dwyer, E., Pinnock, S., Gregoire, J. -M., and Pereira, J.M. C. (2000). Global spatial 
and temporal distribution of vegetation fire as determined from satellite 
observations. International Journal of Remote Sensing, 21, 1289?1302. 
 
Ebert, E. E., Janowiak, J. E., and Kidd, C. (2007). Comparison of near-real-time 
precipitation estimates from satellite observations and numerical models. Bulletin 
of the American Meteorological Society, 88, doi:10.1175/BAMS-88-1-47. 
 
Elvidge, C. D., Kroehl, H. W., Kihn, E. A., Baugh, K. E., Davis, E. R., and Hao, W. 
M. (1996). Algorithm for the retrieval of fire pixels from DMSP Operational 
Linescan System Data. In J. S. Levine (Ed.), Remote Sensing, Modeling and 
Inventory Development, and Biomass Burning in Africa - Biomass Burning and 
Global Change, Vol. 1, MIT Press, pp. 73?85. 
 
Eva, H., and Fritz, S. (2003). Examining the potential of using remotely sensed fire 
data to predict areas of rapid forest change in South America. Applied Geography, 
23, 189-204. 
 
Eva, H., and Lambin, E. F. (1998). Remote sensing of biomass burning in tropical 
regions: Sampling issues and multisensor approach. Remote Sensing of 
Environment, 64, 292?315. 
 
Ferreira, N. C., Ferreira, L. G., and Miziara, F. (2007). Deforestation hotspots in the 
Brazilian Amazon: Evidence causes of as assessed from remote sensing and census 
data. Earth Interactions, 11 (1), 1-16. 
 
Ferreira, L. V., de S?, R. M. L., Buschbacher, R., Batmanian, G., Bensuran, B. R., 
and Costa, K. L. (1999). ?reas protegidas ou espa?os amea?ados? World Wide 
Fund for Nature, Brasilia. 
 
Freitas, S. R., Longo, K. M., Siva-Dias, M. A. F., Silva-Dias, P. L., Chatfield, R., 
Prins, E., Artaxo, P., Grell, G. A., and Recuero, F. S. (2005). Monitoring the 
transport of biomass burning emissions in South America. Environmental Fluid 
Mechanics, 5, 135-167, doi: 10.1007/s10652-005-0243-7. 
 
Giglio, L. (2007). Characterization of the tropical diurnal fire cycle using VIRS and 
MODIS observations. Remote Sensing of Environment, 108(4), 407-421, 
doi:10.1016/j.rse.2006.11.018 
 
Giglio, L., Csiszar, I., and Justice, C. (2006). Global distribution and seasonality of 
active fires as observed with the Terra and Aqua Moderate Resolution Imaging  
Spectroradiometer (MODIS) sensors. Journal of Geophysical Research, 111, 
G02016, doi:10.1029/2005JG000142. 
 
Giglio, L., Csiszar, I., Restas, A., Morisette, J.T., Schroeder, W., Morton, D., et al. 
(2008). Active fire detection and characterization with the Advanced Spaceborne 
 
 159 
 
Thermal Emission and Reflection Radiometer (ASTER). Remote Sensing of 
Environment, in press. 
 
Giglio, L., Descloiters, J., Justice, C. O., and Kaufman, Y. (2003a). An enhanced 
contextual fire detection algorithm for MODIS. Remote Sensing of Environment, 
87, 273?282. 
 
Giglio, L., and Kendall, J., (2001). Application of the Dozier retrieval to wildfire 
characterization: A sensitivity analysis. Remote Sensing of Environment, 77, 34-
49. 
 
Giglio, L., Kendall, J. D., and Justice, C. O. (1999). Evaluation of global fire 
detection algorithms using simulated AVHRR infrared data. International Journal 
of Remote Sensing, 20, 1947?1985. 
 
Giglio, L., Kendall, J. D., and Mack, R. (2003b). A multi-year active fire dataset for 
the tropics derived from the TRMM VIRS. International Journal of Remote 
Sensing, 24, 4505?4525. 
 
Giglio, L., Kendall, J. D., and Tucker, C. J. (2000). Remote sensing of fires with the 
TRMM VIRS. International Journal of Remote Sensing, 21, 203?207. 
 
Global Climate Observing System (GCOS) (2006). World Meteorological 
Organization, Systematic Observation Requirements for Satellite-Based Products 
for Climate. GCOS-107, WMO/TD No. 1338. September 2006. 
 
Global Observation of Forest Cover ? Global Observation for Forest and Land Cover 
Dynamics (GOFC/GOLD) (2006). Fire Monitoring and Mapping Implementation 
Team. 2
nd
 Workshop on Geostationary Fire Monitoring and Applications, 
Darmstadt Germany, 4-6 December 2006. URL (last visited 24 April 2008): 
http://gofc-fire.umd.edu/products/pdfs/Events/2nd_GOFC_Geo_Workshop_ 
Report%20final.pdf  
 
Global Observation of Forest Cover ? Global Observation for Forest and Land Cover 
Dynamics (GOFC/GOLD) (2004). Report on the Joint GOFC/GOLD Fire and 
CEOS LPV Working Group Workshop on Global Geostationary Fire Monitoring 
Applications, Darmstadt Germany, 23-25 March 2004. URL (last visited 24 April 
2008): http://www.fao.org/gtos/gofc-gold/docs/GOLD_19.pdf  
 
Govaerts, Y. M., Wooster, M., Lattanzio, A., and Roberts, G. (2007). MSG/SEVIRI 
fire thermal anomaly (FTA) characterization ATBD, EUMETSAT, Report 
EUM/MET/SPE/06/0398 (pp. 32). 
 
Goward, S. N., Masek, J. G., Williams, D. L., Irons, J. R., and Thompson, R. J. 
(2001). The Landsat 7 mission terrestrial research and applications for the 21st 
century. Remote Sensing of Environment, 78, 3?12. 
 
 160 
 
 
Hansen, J., Sato, M., Ruedy, R., Lacis, A., and Oinas, V. (2000). Global warming in 
the twenty-first century: An alternative scenario. Proceedings of the National 
Academy of Sciences, 97, 9875-9880, doi:10.1073/pnas.170278997. 
 
Hansen, M. C., DeFries, R. S., Townshend, J. R., Carrol, M., Dimiceli, C., and 
Sohlberg, R. A. (2003). Global percent tree cover at a spatial resolution of 500 
meters: First results of the MODIS Vegetation Continuous Fields Algorithm. 
Earth Interactions, 7, paper 10. 
 
Hansen, M. C., DeFries, R. S., Townshend, J. R., Sohlberg, R., Dimiceli, C., and 
Carroll, M. (2002). Towards an operational MODIS continuous fields of percent 
tree cover algorithm: examples using AVHRR and MODIS data. Remote Sensing 
of Environment, 83, 303?319. 
 
Hao, W.M., and M.-H. Liu (1994). Spatial and temporal distribution of tropical 
biomass burning. Global Biogeochemical Cycles, 8 (4), 495-503. 
 
Hughes, R. F., Kauffman, J. B., and Cummings, D. L. (2000). Fire in the Brazilian 
Amazon: Dynamics of biomass, C and nutrient pools in regenerating forests. 
Oecologia, 124, 574-588. 
 
Intergovernamental Panel on Climate Change (IPCC) (1995). Climate Change 1995: 
The science of climate change. Contribution of Working Group I to the Second 
Assessment of the Intergovernmental Panel on Climate Change, JT Houghton, LG 
Meira Filho, BA Callender, N Harris, A Kattenberg and K Maskell (Eds), 
Cambridge University Press, UK, 572 pp. 
 
Justice, C., Giglio, L., Korontzi, S., Owens, J., Morisette, J. T., Roy, D., et al. (2002). 
The MODIS fire products. Remote Sensing of Environment, 83, 244?262. 
 
Kasischke, E. S., Hewson, J. H., Stocks, B., Van der Werf, G., and Randerson, J. 
(2003). The use of ATSR active fire counts for estimating relative patterns of 
biomass burning - A study from the boreal forest region. Geophysical Research 
Letters, 30(18), 1969, doi:10.1029/2003GL017859 
 
Kasischke, E. S., Williams, D., and Barry, D. (2002). Analysis of the patterns of large 
fires in the boreal forest region of Alaska. International Journal of Wildland Fire, 
11, 131?144. 
 
Kaufman, Y. J., Hobbs, P. V., Kirchhoff, V. W., Artaxo, P., Remer, L. A., Holben, B. 
N., et al. (1998a). Smoke, Clouds, and Radiation-Brazil (SCAR-B) experiment. 
Journal of Geophysical Research, 103, 31,783?31,808. 
 
 
 161 
 
Kaufman, Y. J., Justice, C., Flynn, L., Kendall, J., Prince, E. M., Ward, D. E., et al. 
(1998b). Potential global fire monitoring from EOS-MODIS. Journal of 
Geophysical Research, 103, 32215?32238. 
 
Kaufman, Y. J., Setzer, A., Ward, D., Tanre, D., Holben, B. N., Menzel, P., Pereira, 
M. C., and Rasmussen, R. (1992). Biomass burning airborne and spaceborne 
experiment in the Amazonas (BASE-A). Journal of Geophysical Research, 
97(D13), 14,581-14,599. 
 
Kaufman, Y. J., Tucker, C. J., and Fung, I. (1990). Remote sensing of biomass 
burning in the tropics. Journal of Geophysical Research, 95, 9927?9939. 
 
Koltunov, A., and Ustin, S. L. (2007). Early fire detection using non-linear 
multitemporal prediction of thermal imagery. Remote Sensing of Environment, 
110, 18?28. 
 
Koren, I., Kaufman, Y. J., Remer, L. A., Martins, J. V. (2004). Measurement of the 
effect of Amazon Smoke on Inhibition of Cloud Formation. Science, (303), 
doi:10.1126/science.1089424, 1342-1345. 
 
Korontzi, S., Roy, D. P., Justice, C. O., and Ward, D. E. (2004). Modeling and 
sensitivity analysis of fire emissions in southern Africa during SAFARI 2000. 
Remote Sensing of Environment, 92, 376?396. 
 
Kummerow, C., Barnes, W., Kozu, T., Shiue, J., and Simpson, J. (1998). The 
Tropical Rainfall Measuring Mission (TRMM) sensor package. Journal of 
Atmospheric and Oceanic Technology, 15, 809?817. 
 
Laurance, W. F., Albernaz, A. K. M., Fearnside, P., Vasconcelos, H. L., and Ferreira, 
L. V. (2004). Deforestation in Amazonia. Science, 304, 1109-1109. 
 
Laurance, W. F., and Williamson, G. B. (2001). Positive feedbacks among forest 
fragmentation, drought and climate change in the Amazon. Conservation Biology, 
15 (6), 1529-1535. 
 
Li, Z., Nadon, S., and Cihlar, J. (2000). Satellite-based detection of Canadian boreal 
forest fires: development and application of the algorithm. International Journal of 
Remote Sensing, 21 (16), 3057-3069. 
 
Matson, M., and Dozier, J. (1981). Identification of subresolution high temperature 
sources using a thermal IR sensor. Photogrammetric Engineering and Remote 
Sensing, 47, 1311-1318. 
 
Matson, M., and Holben, B. (1987). Satellite detection of tropical burning in Brazil. 
International Journal of Remote Sensing, 8 (3), 509-516. 
 
 
 162 
 
Mendon?a, M. J. C., Vera-Diaz, M. C., Nepstad, D., Seroa da Motta, R., Alencar, A., 
Gomes, J. C., and Ortiz, R. A. (2004). The economic cost of the use of fire in the 
Amazon. Ecological Economics, 49, 89- 105. 
 
Menzel, W. P., Cutrim, E. C., and Prins, E. M. (1991). Geostationary satellite 
estimation of biomass burning in Amazonia during BASE-A. In J. S. Levine (Ed.), 
Global Biomass Burning: Atmospheric, Climatic, and Biospheric Implications, 
MIT Press, 41?46pp. 
 
Menzel, W. P., and Prins, E. M. (1996). Monitoring biomass burning with the new 
generation of geostationary satellites. In J. S. Levine (Ed.), Biomass burning and 
global change, MIT Press, 56?64pp.. 
 
Menzel, W. P., and Purdom, J. F. W. (1994). Introducing GOES-I: the first of a new 
generation of geostationary operational environmental satellites. Bulletin of the 
American Meteorological Society, 75(5), 757?781. 
 
Moran, E., Brondizio, E. S., Tucker, J. M.,  Silva-Forsberg, M. C., McCracken, S., 
and Falesi, I. (2000). Effects of soil fertility and land-use on forest succession in 
Amaz?nia. Forest Ecology and Management, 139, 93-108. 
 
Morisette, J. T., Giglio, L., Csiszar, I., and Justice, C. O. (2005a). Validation of the 
MODIS active fire product over Southern Africa with ASTER data. International 
Journal of Remote Sensing, 26(19), 4239?4264. 
 
Morisette, J. T., Giglio, L., Csiszar, I., Setzer, A., Schroeder, W., Morton, W., et al. 
(2005b). Validation of MODIS active fire detection products derived from two 
algorithms. Earth Interactions, 9, 1?25 (paper no. 9). 
 
Morisette, J.T., Privette, J. L., and Justice, C. O. (2002). A framework for the 
validation of MODIS Land products. Remote Sensing of Environment, 83, 77-96. 
 
Morton, D. C., DeFries, R. S., Randerson, J. T., Giglio, L. Schroeder, W., and van der 
Werf, G. (2008). Agricultural intensification increases deforestation fire activity in 
Amazonia. Global Change Biology, in press. 
 
Morton, D. C., DeFries, R. S., Shimabukuro, Y. E., Anderson, L.O., Arai, E., del bon 
Espirito-Santo, F., Freitas, R., Freitas, R., and Morisette, J. (2006). Cropland 
expansion changes deforestation dynamics in the southern Brazilian Amazon. 
Proceedings of the National Academy of Science, 103 (39), 14637-14641. 
 
Nepstad, D., Carvalho, G., Barros, A. C., Alencar, A., Capobianco, J. P., Bishop, J., 
Moutinho, P., Lefebvre, P., Silva Jr, U. L., and Prins, E. (2001). Road paving, fire 
regime feedbacks, and the future of Amazon forests. Forest Ecology and 
Management, 154, 395-407.  
 
 
 163 
 
Nepstad, D. C., Moreira, A. G., and Alencar, A. (1999a), Flames in the rainforest: 
Origins, impacts and alternatives to Amazonian fires. The Pilot Program to 
Conserve of the Brazilian Rainforest, World Bank, Brasilia, Brazil. 
 
Nepstad, D.,  Moutinho, P., Dias-Filho, M. B., Davidson, E., Cardinot, G., 
Markewitz, D., et al. (2002). The effects of partial throughfall exclusion on canopy 
processes, aboveground production, and biogeochemistry of anAmazon forest. 
Journal of Geophysical Research, 107, doi:10.1029/2001JD000360. 
 
Nepstad, D. C., Schwartzman, S., Bamberger, B., Santilli, M., Ray, D., Schlesinger, 
P., Lefebvre, P., Alencar, A., Prinz, E., Fiske, G., and Rolla, A. (2006). Inhibition 
of Amazon deforestation and fire by parks and indigenous lands. Conservation 
Biology, 20(1), 65-73. 
 
Nepstad, D. C., Ver?ssimo, A., Alencar, A., Nobre, C., Lima, E., Lefebvre, P., 
Schlesinger, P., Potter, C., Moutinho, P., Mendonza, E., Cochrane, M., and 
Brooks, V. (1999b). Large scale impoverishment of Amazonian forests by logging 
and fire. Nature, 398, 505-508. 
 
Pedlowski, M. A., Matricardi, E. A. T., Skole, D., Cameron, S. R., Chomentowski, 
W., Fernandes, C., and Lisboa, A. (2005). Conservation units: A new deforestation 
frontier in the Amazonian state of Rondonia, Brazil. Environmental Conservation, 
32(2), 149-155. 
 
Penner, J. E., Dickenson, R. E., and O?Neill, C.A. (1992). Effects of aerosol from 
biomass burning on the global radiation budget. Science, 256, 1432-1434. 
 
Peres, C. A., Barlow, J., and Haugaasen, T. (2003). Vertebrate responses to surface 
fires in a Central Amazonian forest. Oryx, 37, 97-109. 
 
Prins, E. M., Feltz, J. M., Menzel, W. P., and Ward, D. E. (1998). An overview of 
GOES-8 diurnal fire and smoke results for SCAR-B and 1995 fire season in South 
America. Journal of Geophysical Research, 103(D24), 31821?31835. 
 
Prins, E. M., and Menzel, W. P. (1992). Geostationary satellite detection of biomass 
burning in South America. International Journal of Remote Sensing, 13, 
2783?2799.  
 
Prins, E. M., and Menzel,W. P. (1994). Trends in South American biomass burning 
detected with the GOES VAS from 1983?1991. Journal of Geophysical Research, 
99(D8), 16719?16735. 
 
Prins, E. M., Schmidt, C. C., Feltz, J. M., Reid, J. S., Wesphal, D. L., and Richardson, 
K. (2003). A two year analysis of fire activity in the Western Hemisphere as 
observed with the GOES Wildfire Automated Biomass Burning Algorithm. 
 
 164 
 
Preprints, 12th Conf. on Satellite Meteorology and Oceanography, Long Beach, 
CA, Amer. Meteor. Soc., CD-ROM, P2.28. 
 
PRODES Monitoramento da floresta Amaz?nica brasileira por sat?lite. URL (last 
visited 24 April 2008): http://www.obt.inpe.br/prodes/ 
 
Reinhardt, T. E., Ottmar, R. D., and Castilla, C. (2001). Smoke impacts from 
agricultural burning in rural Brazilian Town. Journal of Air and Waste Manage, 
51, 443-450. 
 
Roberts, G., Wooster, M. J., Perry, G. L. W., Drake, N., Rebelo, L. -M., and Dipotso, 
F. (2005). Retrieval of biomass combustion rates and totals from fire radiative 
power observations: Application to southern Africa using geostationary SEVIRI 
imagery. Journal of Geophysical Research, 110, D21111, 
doi:10.1029/2005JD006018. 
 
Roy, D. P., Jin, Y., Lewis, P. E., and Justice, C. O. (2005). Prototyping a global 
algorithm for systematic fire-affected area mapping using MODIS time series data. 
Remote Sensing of Environment, 97, 137-162. 
 
Rozumalski, R. A. (2000). A quantitative assessment of the NESDIS Auto-Estimator. 
Weather Forecasting, 15, 397?415.  
 
Rudolf, B., Hauschild, H., R?th, W., and Schneider, U. (1996). Comparison of 
raingauge analyses, satellite-based precipitation estimates and forecast model 
results. Advances in Space Research, 18, 53?62. 
 
Rummel, U., Ammann, C., Kirkman, G. A., Moura, M. A. L., Foken, T., Andreae, M. 
O., and Meixner, F. X. (2007). Seasonal variation of ozone deposition to a tropical 
rain forest in southwest Amazonia. Atmospheric Chemistry and Physics, 7, 5,415-
5,435. 
 
Schaaf, C. B., Gao, F., Strahler, A. H., Lucht, W., Li, X., Tsang, T., et al. (2002). 
First operational BRDF, albedo nadir reflectance products from MODIS. Remote 
Sensing of Environment, 83, 135?148. 
 
Scholes, R. J., Ward, D. E., and Justice, C. O. (1996). Emissions of trace gases and 
aerosol particles due to vegetation burning in southern hemisphere Africa. Journal 
of Geophysical Research, 101, 23677?23682. 
 
Schroeder, W., Csiszar, I., and Morisette, J. (2008a). Quantifying the impact of cloud 
obscuration on remote sensing of active fires in the Brazilian Amazon. Remote 
Sensing of Environment, 112, 456-470, doi:10.1016/j.rse.2007.05.004. 
 
 
 165 
 
Schroeder, W., Morisette, J. T., Csiszar, I., Giglio, L., Morton, D., and Justice, C. 
(2005). Characterizing vegetation fire dynamics in Brazil through multisatellite 
data: Common trends and practical issues. Earth Interactions, 9, Paper 13. 
 
Schroeder, W., Prins, E., Giglio, L., Csiszar, I., Schmidt, C., Morisette, J., and D. 
Morton (2008b). Validation of GOES and MODIS active fire detection products 
using ASTER and ETM+ data. Remote Sensing of Environment, 112 (5), 2711-
2726, doi:10.1016/j.rse.2008.01.005. 
 
Schroeder, W., Ruminski, M., Csiszar, I., Giglio, L., Prins, E., Schmidt, C., and 
Morisette, J. (2008c). Validation analyses of an operational fire monitoring 
product: The Hazard Mapping System. International Journal of Remote Sensing, 
submitted. 
 
Seiler, W., and Crutzen, P.J. (1980). Estimates of gross and net fluxes of carbon 
between the biosphere and the atmosphere from biomass burning. Climatic 
Change, 2(3), 207-247, doi:10.1007/BF00137988, 207-247. 
 
Setzer, A.W., and Pereira, M. C. (1991). Amazonia biomass burnings in 1987 and an 
estimate of their tropospheric emissions. Ambio, 20(1), 19-22. 
 
Shukla, J., Nobre, C., and Sellers, P. (1990). Amazon deforestation and climate 
change. Science, 247, 1322-1325. 
 
Sorrensen, C. (2004). Contributions of fire use study to land use/cover change 
framework: Understanding landscape change in agricultural frontiers. Human 
Ecology, 32(4), 395-420. 
 
Tucker, C. J., Holben, B. N., and Golf, T. E. (1984). Intensive forest clearing in 
Rondonia, Brazil, as detected by satellite remote sensing. Remote Sensing of 
Environment, 15 (3), 255-261. 
 
UW Madison Cooperative Institute for Meteorological Satellite Studies (CIMSS) 
Biomass Burning Monitoring Program, URL (last visited 24 April 2008): 
http://cimss.ssec.wisc.edu/goes/burn/WF-ABBA.html. 
 
van der Werf, G. R., Randerson, J. T., Collatz, G. J., and Giglio, L. (2003). Carbon 
emissions from fires in tropical and subtropical ecosystems. Global Change 
Biology, 9, 547?562. 
 
van der Werf, G. R., Randerson, J. T., Collatz, G. J., Giglio, L., Kasibhatla, P. S., 
Arellano Jr., A. F.,  Olsen, S. C., and Kasischke, E. S. (2004). Continental-scale 
partitioning of fire emissions during the 1997 to 2001 El Ni?o/La Ni?a Period. 
Science, 303, 73-76. 
 
 
 166 
 
van der Werf, G. R., Randerson, J. T., Giglio, L., Collatz, G. J., Kasibhatla, P. S., and 
Arellano Jr., A. F. (2006). Interannual variability in global biomass burning 
emissions from 1997 to 2004. Atmospheric Chemistry and Physics, 6, 3423-3441. 
 
Vicente, G. A., Davenport, J. C., and Scofield, R. A. (2002). The role of orographic 
and parallax corrections on real time high resolution satellite rainfall rate 
distribution. International Journal of Remote Sensing, 23, 221?230. 
 
Vicente, G. A., Scofield, R. A., and Menzel, W. P. (1998). The operational GOES 
infrared rainfall estimation technique. Bulletin of the American Meteorological 
Society, 79, 1883?1898. 
 
Vila, D., and Lima, A. (2004). Satellite rainfall estimation over South America: The 
hydroestimator technique. 14th International Conference on Clouds and 
Precipitation, Bologna, Italy, 8?23 July, 2004. 
 
Vitousek, P. M., Aber, J. D., Howarth, R. W., Likens, G. E., Matson, P. A., Schindler, 
D. W., Schlesinger, W. H., and Tilman, D. G. (1997). Technical report: Human 
alteration of the global nitrogen cycle: Sources and consequences. Ecological 
Applications, 7(3), 737-750. 
 
Watson, C. E., Fishman, J., and Reichle, H. G. Jr. (1990). The significance of biomass 
burning as a source of carbon monoxide and ozone in the southern hemisphere 
tropics: A satellite analysis. Journal of Geophysical Research, 95(D10), 16,443-
16,450. 
 
Wolfe, R. E., Nishihama, M., Fleig, A. J., Kuyper, J. A., Roy, D. P., Storey, J. C., et 
al. (2002). Achieving sub-pixel geolocation accuracy in support of MODIS land 
science. Remote Sensing of Environment, 83, 31?49. 
 
Wooster, M. J. (2002). Small-scale experimental testing of fire radiative energy for 
quantifying mass combusted in natural vegetation fires. Geophysical Research 
Letters, 29(21), doi:10.1029/2002GL015487. 
 
Wooster, M. J., Roberts, G., and Perry, G. L. W. (2005). Retrieval of biomass 
combustion rates and totals from fire radiative power observations: FRP derivation 
and calibration relationships between biomass consumption and fire radiative 
energy release. Journal of Geophysical Research, 110(D24311), 
doi:10.1029/2005JD006318. 
 
Wooster, M. J., Zhukov, B., and Oertel, D. (2003). Fire radiative energy for 
quantitative study of biomass burning: Derivation from the BIRD experimental 
satellite and comparison to MODIS fire products. Remote Sensing of Environment, 
86, 83-107. 
 
 
 167 
 
Yamaguchi, Y., Kahle, A. B., Tsu, H., Kawakami, T., and Pinel, M. (1998). 
Overview of advanced spaceborne thermal emission and reflection radiometer 
(ASTER). IEEE Transactions on Geoscience and Remote Sensing, 36, 1062?1071. 
 
Zhang, P., Li, J., Olson, E., Schmidt, T. J., Li, J., and Menzel, W. P.(2006). Impact of 
point spread function on infrared radiances from geostationary satellites. IEEE 
Transactions on Geoscience and Remote Sensing, 44 (8), 2176-2183.