Full text of DOI:10.1080/01431160120888

This article was downloaded by: [Erciyes University]
On: 03 January 2015, At: 15:30
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
DMSP-OLS estimation of
tropical forest area impacted
by surface fires in Roraima,
Brazil: 1995 versus 1998
C. D. Elvidge , V. R. Hobson , K. E. Baugh , J. B.
Dietz , Y. E. Shimabukuro , T. Krug , E. M. L. M.
Novo & F. R. Echavarria
Published online: 25 Nov 2010.
To cite this article: C. D. Elvidge , V. R. Hobson , K. E. Baugh , J. B. Dietz , Y.
E. Shimabukuro , T. Krug , E. M. L. M. Novo & F. R. Echavarria (2001) DMSP-OLS
estimation of tropical forest area impacted by surface fires in Roraima, Brazil:
1995 versus 1998, International Journal of Remote Sensing, 22:14, 2661-2673
To link to this article: http://dx.doi.org/10.1080/01431160120888
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever

or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions

int. j. remote sensing, 2001, vol. 22, no. 14, 2661–2673
DMSP-OLS estimation of tropical forest area impacted by surface
Ž res in Roraima, Brazil: 1995 versus 1998
C. D. ELVIDGE
NOAA National Geophysical Data Center, 325 Broadway, Boulder, Colorado
80303, USA; e-mail: cde@ngdc.noaa.gov
V. R. HOBSON, K. E. BAUGH
Cooperative Institute for Research in Environmental Sciences, University of
Colorado, Boulder, Colorado 80303, USA
J. B. DIETZ
Cooperative Institute for Research on Atmosphere, Colorado State University,
Ft. Collins, Colorado, USA
Y. E. SHIMABUKURO, T. KRUG, E. M. L. M. NOVO
Instituto Nacional de Pesquisas Espaciais, Sa˜o Jose´ dos Campos, Sa˜o Paulo,
Brazil
and F. R. ECHAVARRIA
OYce of Ecology, US Department of State, Washington, DC 20520, USA
(Received 30 July 1999; in Ž nal form 18 August 2000)
Abstract. A procedure has been developed to locate and estimate the area of heavy
forest burning based on the frequency of DMSP-OLS (US Air Force Defense
Meteorological Satellite Program Operational Linescan System) Ž re detection from
time series of observations across the Ž re season. A calibration was developed for
Roraima, Brazil, using Landsat Thematic Mapper (TM) data acquired near the end
of the 1998 burn season and analysed to identify unburnt, partially burnt and heavily
burnt forest areas. A Ž re detection frequency threshold of Ž ve nights was used to
map heavily burnt forest using the 3 months of DMSP-OLS observations. The
threshold of Ž ve Ž re detections, which could occur anytime during the 3-month time
period, was selected to constrain errors of commission involving unburnt forest to
10% of the total area for unburnt forest in the calibration area. At this threshold
setting the DMSP-OLS estimate of heavily burnt forest area covered 79% of the
Landsat measured area. It was found that 77% of the 1998 heavily burnt forest area
was outside of the state’s protected areas (national parks, reserves, indigenous areas).
Two of the protected areas sustained a substantial increase in heavily burnt forest in
1998 relative to 1995 (Reserva Biologica MucajaÌ and Parque IndÌgena Yanomami).
´
´
´
The 1998 forest burning in these two areas was concentrated in their eastern-most
sections. The core of the Yanomami area did not sustain extensive burning in 1998.
Protected areas in the north-eastern section of the state, where forests are mixed with
cerrado, had moderate increases in heavily burnt forest in 1998. Other protected
areas were largely free of the heavy forest burning, which was concentrated to the
west of the state’s primary cerrado zone.
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2001 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160010025961

2662
C. D. Elvidge et al.
1. Introduction
During the Ž rst 3 months of 1998, drought in Brazil’s northern state of Roraima
resulted in major Ž res in tropical forest areas. The extreme Ž re potential had been
predicted by Brazil’s Instituto Nacional de Pesquisas Espaciais (INPE). The state
has a number of national parks, biological reserves, forest reserves, and indigenous
areas that have high priority for conservation (Ž gure 1). Tropical forest is the domin-
ant landcover of the state. There is a large area of cerrado in the north-east quarter
of the state. The cerrado is characterized by the presence of widely distributed small
trees with shrubs and grass, similar to tropical savanna. Portions of this cerrado
area are burnt each year during the dry season, which is typically centred on the
December through March time period.
The Roraima Ž res demonstrated that under severe conditions, Ž res can move
through Amazon tropical forests, causing considerable damage. This possibility
raised new relevance to the remote sensing of Ž res in forests of the Amazon and
other tropical areas.
/
Figure 1. Forest non-forest land cover for the State of Roraima. Overlain are the outlines of
parks, reserves and indigenous areas.

Surface Ž res in Roraima, Brazil
2663
The vast majority of the cleared forest lands in the Amazon region are used for
pasture, which is burnt periodically to clear shrub, vine and renewed tree growth to
enhance grazing conditions and to reduce the numbers of pests such as rattlesnakes.
Forest clearing typically involves the clear cutting of trees and drying for an extended
period of time prior to burning. The drying period is a full year in many instances.
Under normal circumstances forests are too wet to burn and can be used as Ž re
breaks to control the burning used to clear pasture lands (Uhl and KauVman 1990,
Nepstad et al. 1999 ).
˜
The extreme Ž re season in Roraima during 1998 was associated with El Nino
induced drought. Cochrane and Schulze (1998) and Cochrane et al. (1999) provide
excellent descriptions of drought-induced Ž res in tropical forests. The stress induced by
prolonged drought causes trees to drop their leaves, creating a dry fuel layer on the
forest  oor increasing the forest’s  ammability. Fires typically enter the forest initially
from adjacent pasture burning. Surface Ž res move across the forest  oor in thin lines,
consuming the dry leaf litter. Most of the burning is restricted to the forest  oor and
are only able to move into the forest canopy in isolated patches. Unlike Ž res in other
ecosystems, surface Ž res in tropical forests create new fuel for future Ž res by drying
(but not burning) portions of the canopies, inducing the loss of leaf and stem materials.
Trees in tropical forests have thin bark and are not well adapted to Ž res. Small diameter
trees are more readily damaged. However, even larger trees can be impacted if a surface
Ž re moves through an area multiple times. The multiple burn pattern can emerge
within a single year or can be carried over into Ž res in succeeding years.
The description provided by Cochrane and Schulze (1998) and Cochrane et al.
(1999) indicate that the detection of forest Ž res in areas like Roraima using coarse
resolution meteorological satellite imagery is not a trivial matter. The  ames tend
to be organized in narrow ribbons, moving 100–150 m per day. The Ž reline intensity
is reported to be generally less than 50 kW m and is under a forest canopy. The
Õ1
Ž res tend to die down at night and restart from burning embers the next day and
generally reaching a peak in early afternoon. The forest canopy is left largely intact,
with evidence of burning limited to loss in green leaves and occasional consumption
of canopies in isolated pockets. This raises questions whether these Ž res can be
detected with satellite imagery having ground resolutions measured in kilometres.
The night-time Ž re detection capabilities of US Air Force Defense Meteorological
Satellite Program (DMSP) Operational Linescan System (OLS) have been known
since the early 1970s (Croft 1973). However, due to limited access to the digital data,
very few studies documenting the OLS Ž re detection capabilities have been per-
formed. Brazilian Institute for Space Research (INPE) was asked by the Government
of Brazil through the Ministry of Science and Technology to perform an assessment
of the Roraima Ž res to answer questions regarding the location and spatial extent of
the tropical forest burning. In response to a request from INPE, NOAA processed
nightly DMSP-OLS Ž re products for Roraima from 1 January through the end of
March for two years: 1995 and 1998. We report here on the results comparing the
Ž re events from the two years and results obtained with March 1998 Landsat data.
2. Fire detection with DMSP-OLS data
DMSP has operated low light imaging sensors since the early 1970s. However, digital
data was not readily available until 1992, when the US Air Force and NOAA established
the DMSP digital archive at the NOAA National Geophysical Data Center (NGDC)
in Boulder, Colorado. NGDC receives raw DMSP data via an electronic connection to

2664
C. D. Elvidge et al.
Air Force Global Weather Central (AFGWC), OVutt Air Force Base, Omaha, Nebraska.
The DMSP data arrive at NGDC within 1–2h after data are initially acquired. It is
anticipated that DMSP-OLS will continue to  y until approximately 2010. The night-
time low light imaging capabilities of the OLS will be continued under the U.S. National
Polar-orbiting Operational Environmental Satellite System (NPOESS).
The DMSP generally operates two satellites in sun-synchronous orbits, one in a
dawn–dusk orbit, the other in a day–night orbit. The OLS is an oscillating scan
radiometer designed for cloud imaging with two spectral bands (VIS and TIR) and
a swath of 3000 km. The ‘VIS’ bandpass straddles the visible and near-infrared
~
(VNIR) portion of the spectrum (0.5–0.9 mm). Satellite attitude is stabilized using
four gyroscopes (three axis stabilization), a star mapper, Earth limb sensor and a
solar detector. The OLS VIS band signal is intensiŽ ed at night using a photomulti-
plier tube (PMT), for the detection of moonlit clouds. The low light sensing capabilit-
ies of the OLS at night permit the measurement of radiances down to 10 W cm
Õ9
Õ2
sr mm . This is more than four orders of magnitude lower than the OLS daytime
Õ1 Õ1
VIS band or the VNIR bands of other sensors, such as the NOAA AVHRR or the
Landsat Thematic Mapper (TM).
There are two spatial resolution modes in which data can be acquired. The full
resolution data, having nominal spatial resolution of 0.5 km, is referred to as ‘Ž ne’.
On board averaging of Ž ve by Ž ve blocks of Ž ne data produces ‘smoothed’ data
with a nominal spatial resolution of 2.7 km. Most of the data received by NOAA-
NGDC is in the smoothed spatial resolution mode.
The potential use of night-time OLS data for the detection of Ž res was Ž rst noted
by Croft (1973, 1978, 1979). The Ž rst systematic inventory of Ž res with OLS data
was reported by Cahoon et al. (1992), who worked with OLS Ž lm. More recently,
Elvidge et al. (1996 and 1997) developed algorithms to identify and geolocate Ž res
and city lights in digital OLS imagery. Time series analysis is used to distinguish
persistent nocturnal emissions generated by villages, towns and cities from ephemeral
emissions arising from Ž res.
3. Data sources and preparation
3.1. Fire products
DMSP-OLS data was extracted from the archive for each night from the begin-
ning of January to the end of March for 1998 and 1995. The 1995 data was processed
to provide a comparison between the extreme Ž re events of 1998. Meteorological
records indicate 1995 was a typical year in terms of precipitation. The data were
processed using the OLS algorithms described in Elvidge et al. (1996, 1997, 1998)
to generate 30 arc second Ž re and cloud images for each night where data was
available in the archive.
3.2. L and cover
A 30 m resolution forest/non-forest map of the Amazon, prepared from Landsat
Thematic Mapper data by the NASA Landsat PathŽ nder project (Chomentowski et al.
1994) was used as the primary land cover source. This dataset was assembled using
Landsat TM data from the early 1990s. The high resolution coverage was regridded to
the 30 arc second grid being used for the DMSP Ž re products. In locations where there
was no land cover data available from the Landsat data, land cover was drawn from
the global 1 km Advanced Very High Resolution Radiometer (AVHRR) land cover
product prepared by the US Geological Survey, EROS Data Center (Brown et al. 1993).
The resulting land cover map is shown in Ž gure 1. Forest is the primary land

Surface Ž res in Roraima, Brazil
2665
cover of the state. There is a large cerrado zone in the north-east section of the state,
indicated as non-forest. Settlement areas in forest regions, indicated by trellis like
road patterns mapped as non-forest, are present to the south of the cerrado zone.
3.3. Conservation areas
The outline of parks, reserves, and indigenous areas were manually digitized from
´
a one to million scale map. A total of 10 areas were digitized: Area Ind´Ìgena Waimiri
´
Atroari, Area IndÌ´gena San Marcos, Area Ind´Ìgena Raposa do Sol, Area Ind´Ìgena
´
´
´
Manoa Pium, Area IndÌgena Araca, Parque IndÌgena Yanomami, Parque Nacional
´
do Rio Branco, Estacao Ecolo´gica Caracara´Ì, Reserva Biolo´gica Mucaja´Ì and Reserva
´
¸ ˜
Florestal Parima. The location of the 10 areas are shown on Ž gure 1.
˜
3.4. L andsat burn analysis
INPE obtained and analysed low cloud cover Landsat TM data of Path 232
Row 58 acquired for 1997 and on 11 March 1998. The two TM scenes were
georeferenced using ground control points extracted from topographic maps. The
TM image of 1997 was analysed to establish the locations of intact forests prior to
the Ž re events of 1998. The 1998 data was analysed to identify three classes of
forests: unburnt, partially burnt, and heavily burnt.
The analysis of the 1997 TM image was based on the methodology proposed by
Shimabukuro et al. (1997) for the PRODES (Estimation of Gross Deforestation of
Amazoˆnia) project. Bands 3, 4 and 5 were used as input data for a spectral linear
mixing model (Shimabukuro and Smith 1991, Shimabukuro et al. 1998), generating
three fraction images obtained for land surface pixels: green vegetation, soil and shade.
The segmented shade fraction image (similarity and area thresholds of 8 and 25,
respectively) was submitted to a non-supervised classiŽ cation process. The result of
image segmentation was inspected visually against a TM colour composite (R5 G4
B3) image and it showed good discrimination of forest and non-forest areas. After the
segmentation a non-supervised classiŽ cation was performed using ISOSEG algorithm
with a probability threshold of 95%. The end result was classiŽ cation map indicating
areas with forest and non-forest areas. This was again compared against a colour
composite image to visually conŽ rm the accuracy of the forest/non-forest discrimination.
The 1998 TM data were processed with the same spectral mixture analysis tools
as the 1997 data, identifying the proportion of green vegetation, soil and shade present
in each land surface pixel. The analysis proceeded by inspecting the fraction images
for the areas determined to be forest in 1997 against colour composites of the 1997
and 1998 scenes. Of the three spectral mixture components the vegetation component
was the most sensitive to the detection of areas aVected by the Ž re in the forest areas.
Based on this observation, the vegetation fraction image was segmented, using a
combination of similarity and area thresholds of 7 and 25, respectively. The segmented
image was then classiŽ ed using a supervised Bhattacharrya algorithm to identify three
forest classes: unburnt, heavily burnt and partially burnt. Heavily burnt forests were
identiŽ ed as areas that were mapped as forests in the 1997 data that had lost nearly
all of their green vegetation spectral component in the 1998 data. The implication is
that these are forest areas where repeated surface Ž res have resulted in a nearly
complete loss of canopy foliage. Unburnt forests were identiŽ ed as areas mapped as
forest in 1997 that retained high proportions of green vegetation in the 1998 data.
Partially burnt forests were deŽ ned as an intermediate class, between the heavily burnt
and unburnt forest classes. The resulting Ž re intensity classiŽ cation derived from the
Landsat data was aggregated to a 3 arc second grid ( 100 m pixels).
~

2666
C. D. Elvidge et al.
The Landsat analysis was assisted by the review of  ight lines of airborne video
data acquired by INPE in May of 1998. Interpretation of the airborne video data
was complicated by the regrowth of green leaves in many of the tree canopies in
forest areas that were heavily burnt. Evidence for the passage of surface Ž res in the
airborne video data frequently had to be implied by subtle losses of green leaf
spectral signature and density of burnt openings and burnt canopies in an otherwise
intact forest canopy.
4. Analysis and results
The nightly OLS Ž re and cloud images were analysed with the intention of
developing an estimate of the areal extent of tropical forest burning in 1998 relative
to 1995. The nightly Ž re and cloud images were used to make composites of cloud-
free Ž re detections for the Ž rst 3 months of 1995 and 1998 (Ž gure 2). A calibration
for estimation of the burnt area based on the frequency of Ž re observations was
developed using the Landsat results. This calibration was then applied to the entire
state to identify tropical forest areas that had been aVected by Ž re. These results
were used to estimate the cumulative area of tropical forest burning for the entire
state and each of the parks and reserves in each of the two years.
Because of the long wavelength of the OLS thermal band (10–12 mm) Ž res are
rarely detected in the thermal band data. With the high sensitivity of OLS low light
imaging, Ž res of subpixel size can saturate the visible band data and are yet undetect-
able in the thermal band. Due to the lack of multispectral information on Ž re
brightness, it is not possible to estimate the active burnt area on a single night. As
a result, indirect methods for estimating burnt area from time series of OLS
observations have been investigated. A calibration procedure has been developed for
identifying the location and estimating the area of forest burning based on the
frequency with which Ž res were detected during a 3-month time period.
The OLS Ž re pixel count from 1 January through 11 March 1998 was resampled
Figure 2. DMSP-OLS Ž re detection frequency for Roraima during the Ž rst 3 months of 1995
and 1998.

Surface Ž res in Roraima, Brazil
2667
to a 3 arc second grid and overlaid this on the Landsat Ž re intensity classiŽ cation.
A calibration area was selected in the south-west corner of the Landsat scene,
covering the core of forest burning observed in the state. Once this overlay was
complete the area of the three forest burn classes versus OLS Ž re detection frequency
was plotted. This graph, presented in Ž gure 3, shows the heavily burnt forest class
to have higher OLS Ž re detection frequency than the partially or unburnt forest
classes. The curves for partially and unburnt forests show a steady decline as the
OLS Ž re detection frequency increases. For both partially burnt and unburnt forest
the area drops approximately by half for each increased step in OLS Ž re detection
frequency. In contrast, the heavily burnt forest area declines gradually as the Ž re
detection frequency moves from one to three. For OLS Ž re detection frequencies of
Ž ve to nine the associated area of heavily burnt forests remains at a high and nearly
constant level of 145 km before entering a gradual decline from 9 to 18 Ž re detections.
2
Figure 3 demonstrates the potential for estimating the area of heavily burnt forest
based on OLS Ž re detection frequency. A calibration was developed to classify pixels
as heavily burnt forests if their landcover was mapped as forest and their OLS Ž re
detection frequency during a 3-month period exceeded a speciŽ ed value. Knowing
the total area of heavily burnt forest in the calibration area it would be possible to
select a Ž re frequency threshold that results in a burnt area estimate that closely
matches the burnt area derived from the Landsat data. But would there be a
correspondence in the identiŽ cation of speciŽ c areas impacted by surface Ž res between
the two systems? That is to say, can the OLS burn frequency be used as a classiŽ er
Figure 3. Area of heavily, partially and unburnt forest from 1998 Landsat TM analysis for
each OLS Ž re detection frequency level. Data was generated in a calibration area
using OLS Ž re detection frequency data from 1 January to 11 March 1998 that was
coregistered to the 11 March 1998 Landsat TM forest burning classiŽ cation. The
calibration area is in the south-west corner of the Landsat scene which covers the
core of the forest burning observable in the scene.

2668
C. D. Elvidge et al.
to distinguish between tropical forest areas that were free of surface Ž re during a
burn season and locations where surface Ž res passed?
As can be seen in Ž gure 3, any threshold that detects a large percentage of the
heavily burnt forest area will also detect partially burnt and unburnt forest areas
(errors of commission). Setting an OLS Ž re detection frequency threshold requires
the balancing of the correct identiŽ cation of heavily burnt forest against the errors
of omission and commission that would be present at any Ž xed threshold. It was
agreed that detection of partially burnt forest, while not desirable, was not as
deleterious as the detection of unburnt forest based on OLS Ž re detection frequency.
The eVect of the frequency threshold on the identiŽ cation of heavily burnt forests
was explored by overlaying images of the Landsat heavily burnt forest class and 10
threshold classes of OLS Ž re detection frequency in forest areas (Ž gure 4). The 10
OLS cumulative Ž re detection frequency classes are deŽ ned as the set of pixels which
had Ž re frequencies equal to or exceeding the speciŽ ed number. The images were
generated by assigning the heavily burnt forest class to cyan (blue plus green) and
the OLS Ž re detection cumulative frequency class to red. Where the two coincide,
the image area is white, indicating the correct identiŽ cation of heavily burnt forest.
Where the OLS Ž re detection frequency class does not coincide with heavily burnt
forest, the image shows red (errors of commission). Where the heavily burnt forest
class does not coincide with the OLS Ž re detection frequency class, the image shows
cyan (errors of omission). The Ž rst image in the series (upper left) shows the class of
reference grid pixels that had one or more OLS Ž re detections in 1998 up to 11 March.
This image shows a core of white in the upper left corner, but large areas of red
Figure 4. The eVect of the frequency threshold on the identiŽ cation of heavily burnt forest
is demonstrated by overlaying images of the Landsat-derived heavily burnt forest class
and 10 threshold classes of OLS Ž re detection frequency in forest areas. The OLS Ž re
frequency classes are deŽ ned as the set of pixels which had Ž re frequencies equal to
or exceeding the threshold value. The images were generated by assigning the heavily
burnt forest class to cyan (blue plus green) and the OLS Ž re detection cumulative
frequency class to red. Heavily burnt forest areas that were correctly identiŽ ed by an
OLS Ž re frequency class are white. Red areas are either partially or unburnt forest
areas covered by the OLS Ž re frequency class. Cyan areas are heavily burnt forests
that were not covered by the OLS Ž re frequency class.

Surface Ž res in Roraima, Brazil
2669
extending to other parts of the calibration area. The series of images show the area
of red decreasing rapidly as the Ž re detection frequency used to deŽ ne the class is
stepped to higher levels. The core of white also decreases, but less rapidly. The trade-
oV for lowered errors of commission is higher errors of omission, indicated by the
increase in cyan pixels as the image series progresses to the lower right-hand corner.
The selection of the OLS Ž re detection frequency threshold was made by con-
straining errors of commission involving unburnt forests to 10% or less of the total
area of unburnt forest in the calibration area, which covers the core of the 1998
forest burning. The cumulative area for each of the three forest Ž re classes versus
OLS Ž re frequency class is shown in Ž gure 5. By applying the selection criteria to
the data in Ž gure 5, an OLS Ž re detection frequency of Ž ve was selected as the
heavily burnt forest classiŽ cation threshold. At this threshold 8.7% of the unburnt
forest area and 11.5% of the partially burnt forest area are detected as errors of
commission in the calibration area. The area of heavily burnt forest correctly identi-
Ž ed with this threshold is 47.7% of the total. The estimated area of heavily burnt
forest observed with an OLS Ž re detection frequency threshold is the sum of the
detected areas for the three forest classes. With the selected threshold of Ž ve, the
estimated area of heavily burnt forest in the calibration area is 1299 km (948 from
2
heavily burnt forest, 254 from partially burnt forest, and 97 from unburnt forest).
This is 79% of the actual area of heavily burnt forest in the calibration area, as
mapped with Landsat data.
Figure 5. Plot showing the cumulative area of unburnt, partially burnt and heavily burnt
forests covered by each of the DMSP-OLS Ž re frequency classes. In order to satisfy
the requirement that the area of unburnt forest errors of commission be 10% or less
of the total area of unburnt forest, a Ž re frequency threshold of Ž ve was selected for
the identiŽ cation of heavily burnt forest.

2670
C. D. Elvidge et al.
Having developed a calibration between the Landsat and OLS Ž re detection
frequency capable of locating and estimating the cumulative area of heavily burnt
forests with a minimal detection of partially burnt and unburnt forests, this was then
applied to the entire state for 1995 and 1998 by identifying all forest areas where
Ž res were detected in the OLS grid Ž ve or more times. The OLS Ž re frequency
derived image maps of heavily burnt tropical forest for 1995 and 1998 are shown in
Ž gure 6. In 1995 the heavily burnt forest area was quite small (425 km ) and was
2
concentrated in settlement areas. There was a 20-fold increase in the extent of heavily
burnt forest in 1998, with an estimated area of 9038 km . The 1998 forest burning
2
was primarily concentrated along the fringes of the cerrado zone, with the greatest
extension to the west of the cerrado zone located in the north-eastern part of the state.
The estimated area of heavily burnt forest for parks, reserves and indigenous
areas for 1995 and 1998 are presented in Ž gure 7. No heavy forest burning was
´
´
observed in either year the Parque Nacional do Rio Branco and the Area IndÌgena
´
Waimiri Atroari. For Ž ve of the areas (Reserva Florestal Parima˜, Parque IndÌgena
´
Yanomami, Reserva Biologica Mucaja´Ì, Area Ind´Ìgena Manoa Pium and Estacao
¸˜
´
Ecologica Caracara´Ì) there was little or no heavy forest burning in 1995 but some
´
level of heavy forest burning in 1998. For three of these areas (Reserva Florestal
´
Parima, Area Ind´Ìgena Manoa Pium and Estacao Ecolo´gica Caracara´Ì) the increase
¸˜
˜
appears to have been minor. The most substantial increase in heavy forest burning
in 1998 relative to 1995 occurred in the Parque IndÌgena Yanomami and the Reserva
´
Biolo´gica Mucaja´Ì. In examining Ž gure 6 it can be seen that the increased burning
was concentrated in the eastern portion of these two areas. The indigenous areas in
the north-eastern part of the state, where tropical forests are mixed with cerrado
´
´
´
(Area Ind´Ìgena Araca, Area Ind´Ìgena San Marcos and Area Ind´Ìgena Raposa do
Sol) all sustained an increase in heavy forest burning in 1998 relative to 1995. These
three areas appear to be prone to sustain some level of forest burning each year due
to the mixture of forest with cerrado. On a statewide basis the Ž res in both years
were concentrated outside of the state’s parks, reserves and indigenous areas. In
1998, 77% of the heavy forest burning was outside of the protected areas. During
1995 63% of the heavy forest burning was outside of the protected areas.
Figure 6. Distribution of heavily burnt forest in Roraima from 1995 and 1998 based on a
DMSP-OLS Ž re detection frequency of Ž ve or greater during the January through
March time period.

Surface Ž res in Roraima, Brazil
2671
Figure 7. DMSP-OLS estimated area of heavy forest burning statewide and for each of the
protected areas during 1995 and 1998.
5. Conclusion
A methodology has been developed to locate and estimate the area of heavy
forest burning based on the frequency of DMSP-OLS Ž re detection from a time
series of observations across the Ž re season. This calibration was developed using
Landsat TM data acquired near the end of the 1998 burn season and analysed to
identify unburnt, partially burnt and heavily burnt forest areas. A Ž re frequency
threshold of Ž ve nights or greater was used to map heavily burnt forest using the 3
months of DMSP-OLS Ž re observations from 1995 and 1998. The threshold of Ž ve
Ž re detections was selected to constrain errors of commission involving unburnt
forest to 10% of the total area for unburnt forest in the calibration area. Recognizing
that the Ž re frequency associated with heavy forest burning varies with forest type
and species composition, we have set the threshold high to avoid overestimation of
burn area. For the calibration area, centred on an area of intense forest burning in
1998, the method underestimates the area of heavy forest burning by 21%.
Based on DMSP-OLS Ž re frequency we estimate the area of heavily burnt forest
in Roraima during 1998 to be 9038 km . This represents a 20-fold increase over the
2
estimate of heavily burnt forest in the state for the year 1995. The DMSP estimate
of heavily burnt forest area is 77% of the statewide estimates made by INPE using
Landsat and airborne video data (11 730 km ). The Instituto Nacional de Pesquisas
2
da Amazoˆnia (INPA) estimated the area of heavily burnt forest in Roraima for 1998
to be 13 000 km based on airborne missions and intensive Ž eld work. The DMSP-
2
OLS estimate is intentionally set to yield conservative estimates of forest burn area
to avoid overestimation.

2672
C. D. Elvidge et al.
It was found that 77% of the 1998 heavily burnt forest area was outside of the
state’s protected areas (national parks, reserves, indigenous areas). Two of the pro-
tected areas sustained a substantial increase in heavily burnt forest in 1998 relative
to 1995 (Reserva Biolo´gica Mucaja´Ì and Parque Ind´Ìgena Yanomami). The 1998
forest burning in these two areas was concentrated in their eastern-most sections.
The core of the Yanomami area did not sustain extensive burning in 1998. Protected
areas in the north-eastern quarter of the state, where forests are mixed with cerrado,
experienced moderate increases in heavily burnt forest in 1998. Other protected areas
were largely free of the heavy forest burning, which was concentrated to the west of
the Roraima’s primary cerrado zone.
Errors of commission for unburnt forest being identiŽ ed as heavily burnt
amounted to 8.7% of the TM identiŽ ed unburnt forest in the calibration area. When
our DMSP Ž re frequency procedure was applied to the entire state, large areas of
forest were identiŽ ed with very low Ž re frequencies. For instance, only 0.2% of the
´
´
combined 45 370 km of forest in the Reserva Forestal Parima˜, Area IndÌgena
2
Araraca, Mandapium, Waimiri Atroari and Parque Nacional do Rio Branco were
´
identiŽ ed as heavily burnt in 1998 based on DMSP Ž re frequency. This suggests
that the 8.7% error of commission for unburnt forest in the calibration area is a
substantial overestimation of the techniques errors when applied across large regions.
The primary factor contributing to the misidentiŽ cation of unburnt forest as
heavily burnt in the calibration area is the large disparity in the size of the OLS and
Landsat TM pixel footprints on the Earth’s surface. The OLS pixel footprints are
approximately 1000 times larger than the TM pixels. Because OLS is capable of
detecting Ž res which are much smaller than the pixel size, there is an in-built tendency
to overestimate burn area. Based on this argument errors of commission on the
order of 5–10% can be expected when Ž re activity is high.
Despite these uncertainties, direct comparison of OLS Ž re detection frequency in
early 1998 with a map of burnt forest from Landsat TM data reveals a substantial
level of agreement. At the scale of a Landsat scene there is a close correspondence
between the cumulative area estimates of the area for unburnt, partially burnt, and
heavily burnt tropical forest from Landsat multispectral data and OLS Ž re frequency.
When used as a classiŽ er, OLS Ž re frequency produces forest burn classes that have
a general correspondence to Landsat results. The Landsat classes tend to be
fragmented while the OLS classes tend to be more contiguous. This can be partially
explained by the diVerent spatial resolution of the two sensors. Another factor which
probably contributed to the fragmented versus continuous nature of the classiŽ cation
results may be explained by the diVerences in the observations made by the two
systems. The OLS is making a direct observation of the radiant energy emitted from
the Ž res with more than 80 nights of observations contributing to the Ž re frequency
results. The Landsat classiŽ cation is based on the spectral evidence of the Ž re impacts,
expressed primarily as a decline in the green vegetation spectral signature of tropical
forest canopies.
Our results indicate that the DMSP-OLS is capable of detecting surface Ž res at
night in tropical forests. These surface Ž res rarely move in to the upper levels of the
forest canopy and as a result much of the radiant energy from the Ž res is absorbed
by the canopy and forest structural elements. Only a fraction of the Ž re’s total
radiance is transmitted up through the canopy towards space. Because of the obscura-
tion of the Ž res by the forest canopy, the satellite detection of surface Ž res in tropical
forests is a challenging observation. The detection of the surface Ž res with OLS data

Surface Ž res in Roraima, Brazil
2673
demonstrates the value of DMSP data for Ž re detection in tropical forests. This is
particularly signiŽ cant because night-time Ž res are those that are out of control and
most likely to persist over many days.
Acknowledgment
The authors appreciate the DMSP program oYce and the US Air Force Weather
Agency for providing NOAA-NGDC with DMSP data used in this research. This
research was sponsored in part by the US Agency for International Development.
The views and opinions expressed are solely the author’s and do not necessarily
represent those of the US Department of State.
References
Brown, J. F., Loveland, T. R., Merchant, J. W., Reed, B. C., and Ohlen, D. O., 1993,
Using multisource data in global land-cover characterization: concepts, requirements,
and methods. Photogrammetric Engineering and Remote Sensing, 59, 977–987.
Cahoon, D. R. Jr, Stocks, B. J., Levine, J. S., Cofer, W. S. III, and O’Neill, K. P., 1992,
Seasonal distribution of African savanna Ž res. Nature, 359, 812–815.
Chomentowski, W., Salas, B., and Skole, D. L., 1994. Landsat PathŽ nder project advances
deforestation mapping. GIS World, 7, 34–38.
Cochrane, M. A., and Schulze, M. D., 1998, Forest Ž res in the Brazilian Amazon.
Conservation Biology, 12, 948–950.
Cochrane, M. A., Alencar, A., Schultze, M. D., Souza, C. M. Jr, Nepstad, D. C.,
Lefebvre, P., and Davidson, E. A., 1999, Positive feedbacks in the Ž re dynamic of
closed canopy tropical forests. Science, 284, 1832–1835.
Croft, T. A., 1973, Burning waste gas in oil Ž elds. Nature, 245, 375–376.
Croft, T. A., 1978, Nighttime images of the earth from space. ScientiŽ c American, 239, 68–79.
Croft, T. A., 1979, The brightness of lights on Earth at night, digitally recorded by DMSP
satellite. Stanford Research Institute Final Report prepared for the US Geological
Survey.
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 Ž re pixels from DMSP Operational Linescan
System. In Global Biomass Burning, edited by Joel S. Levine (Cambridge, MA: MIT
Press), pp. 73–85.
Elvidge, C. D., Baugh, K. B., Kihn, E. A., Kroehl, H. W., and Davis, E. R., 1997, Mapping
city lights with nighttime data from the DMSP Operational Linescan System.
Photogrammetric Engineering and Remote Sensing, 63, 727–734.
Elvidge, C. D., Baugh, K. E., Hobson, V. R., Kihn, E. A., and Kroehl, H. W., 1998,
Detection of Ž res and power outages using DMSP-OLS data. In Remote Sensing
Change Detection: Environmental Monitoring Methods and Applications, edited by R. S.
Lunetta and C. D. Elvidge (Chelsea, Michigan: Ann Arbor Press), pp. 123–135.
Nepstad, D. C., Verissimo, A., Alencar, A., Nobre, C., Lima, E., Lefebvre, P., Schlesinger,
P., Potter, C., Moutinho, P., Mendoza, E., Cochrane, M., and Brooks, V., 1999,
Large-scale impoverishment of Amazon forests by logging and Ž re. Nature, 398,
505–508.
Shimabukuro, Y. E., and Smith, J. A., 1991, The least-squares mixing models to generate
fraction images derived from remote sensing multispectral data. IEEE T ransactions
on Geoscience and Remote Sensing, 29, 16–20.
Shimabukuro, Y. E., Mello, E. M. K., Moreira, J. C., and Duarte, V., 1997, Segmentaca˜o
¸
e classiŽ caca˜o da imagem sombra do modelo de mistura para mapear des orestamento
¸
na Amazonia. Instituto Nacional de Pesquisas Espaciais, (INPE-6147-PUD 029), Sa˜o
/
ˆ
Jose dos Campos, Brazil.
´
Shimabukuro, Y. E., Batista, G. T., Mello, E. M. K., Moreira, J. C., and Duarte, V.,
1998, Using shade fraction image segmentation to evaluate deforestation in Landsat
Thematic Mapper images of the Amazon region. International Journal of Remote
Sensing, 19, 535–541.
Uhl, C., and Kauffman, J. B., 1990, Deforestation, Ž re susceptibility and potential tree
responses to Ž re in the Eastern Amazon. Ecology, 71, 437–449.