IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1

Transcription

1 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1 Estimating Burned Area in Mato Grosso, Brazil, Using an Object-Based Classification Method on a Systematic Sample of Medium Resolution Satellite Images Yosio Edemir Shimabukuro, Jukka Miettinen, René Beuchle, Rosana Cristina Grecchi, Dario Simonetti, and Frédéric Achard Abstract This paper presents a new approach for estimating burned areas at a regional scale, using a systematic sample of medium spatial resolution satellite images. This approach is based on a pan-tropical deforestation survey developed by the Joint Research Centre. We developed and tested our approach over Mato Grosso State, located in the Brazilian Legal Amazon region, with a total area of km 2. We analyze Landsat-5 TM imagery over 77 sample sites (20 km 20 km in size) located at each full degree confluence of latitude and longitude. Our new approach leads to an estimate of burned area for year 2010 at km 2, representing approximately 7.3% of the Mato Grosso area. This estimate is compared to estimates from two different approaches: 1) from a method developed by the Brazilian Institute for Space Research, applied to a wall-to-wall coverage of Landsat-5 TM imagery and 2) from a method using MODIS MCD64A1 products of the University of Maryland, resulting in and km 2 of burned area, respectively (representing 7.8% or 6.1% of Mato Grosso area). Our method produces statistically valid estimates of burned areas for the Brazilian State of Mato Grosso in a more efficient manner than previous methods and enables the inclusion of small burn scars typically missed by coarse resolution satellites. This approach can be applied for regional and global assessments as well as for refining and evaluating burned area products based on coarse spatial resolution imagery like MODIS or SPOT-VEGETATION. Index Terms Image classification, Landsat, MODIS, regional and global assessment, sampling, segmentation, tropical forest. Manuscript received December 24, 2014; revised July 17, 2015; accepted July 23, Y. E. Shimabukuro is with the European Commission, Joint Research Centre (JRC), Institute for Environment and Sustainability (IES), Forest Resource, and Climate Unit, Ispra, Italy, and also with the Brazilian National Institute for Space Research (INPE), São José dos Campos, Brazil ( yosio@dsr.inpe.br). J. Miettinen was with the European Commission, Joint Research Centre (JRC), Institute for Environment and Sustainability (IES), Forest Resource, and Climate Unit, Ispra, Italy. He is now with the Centre for Remote Imaging, Sensing, and Processing, National University of Singapore, Singapore , Singapore ( jimietti@yahoo.com). R. Beuchle, R. C. Grecchi, D. Simonetti, and F. Achard are with the European Commission, Joint Research Centre (JRC), Institute for Environment and Sustainability (IES), Forest Resource, and Climate Unit, Ispra, Italy ( rene.beuchle@jrc.ec.europa.eu; rosana.grecchi@jrc.ec.europa.eu; dario.simonetti@jrc.ec.europa.eu; frederic.achard@jrc.ec.europa.eu). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /JSTARS I. INTRODUCTION F IRES in vegetation cover play an important role in the emission of greenhouse gases and aerosols affecting the Earth s radiation balance [1]. Information on the location and extent of the areas affected by fire is necessary to assess the effects of biomass burning on atmospheric chemistry, ecosystem functioning, and human health. Satellite sensor data provide a unique source of spatial information in detecting, monitoring, and characterizing fires for global environmental change research [2]. Remote sensing applications for the detection and monitoring of fires from local to continental and global scales have been developed over many years, using a number of different sensors and systems such as Landsat thematic mapper (TM) [3], moderate resolution imaging spectroradiometer (MODIS) [4], National Oceanic and Atmospheric Administration (NOAA) advanced very high resolution radiometer (AVHRR) [5], SPOT-VEGETATION [6], the meteorological satellite program (DMSP) satellite [7], and the geostationary operational environmental satellite (GOES) [8] (for a review at global scale see [9]). Vegetation fires are expected to negatively impact carbon stocks, biological diversity, and human health. Moreover, fires can potentially compromise the efficacy of emission reduction policies, such as activities related to Reducing Emissions from Deforestation and Degradation (REDD+) in framework of the United Nations Framework Convention on Climate Change (UNFCCC) [10]. Vegetation fires are becoming increasingly important worldwide, not only for understanding and predicting their future environmental impacts but also for implementing efficient climate change mitigation policies. Fire in the Brazilian tropical forests is the principal source of greenhouse gas emissions, representing about 75% of the total volume of CO 2 released in the country [11]. Deforestation in the Brazilian Amazon is considered to release over 200 million tons of CO 2 -equivalent carbon per year, contributing significantly for the global amounts of greenhouse gas emissions and, consequently, for the planet climate changes [12]. In the Amazon, fire is widely used for the initial conversion of extensive areas of natural vegetation into agricultural fields and pasture areas, and for the subsequent maintenance of deforested areas [13], [14]. In contrast, the Cerrado biome is a fire-dependent ecosystems with fire adapted IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See for more information.

2 2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING plant species [15]. However, natural fire occurrences and accidental burning are extremely rare, with the vast majority of burning events resulting from anthropogenic, i.e., deliberate, use of fire [13]. Fire detection and burned area mapping are most typically performed with coarse spatial resolution (>250 m) satellite data which offer high temporal frequency enabling near real-time monitoring of the development of the burning season [4] [6]. However, depending on the fire regimes and the burned scar size distribution in the area of interest, burned area monitoring from coarse spatial resolution data may only seriously distort both the overall estimation and the spatial distribution of burned areas [16], [17]. Furthermore, developing automatic burned area mapping with coarse resolution data is very challenging due to the variability and temporal development of the burn scar s spectral characteristics. Due to its higher spatial detail, medium spatial resolution (30 m) satellite data are more suitable for burned area assessment and offer means to analyze the extent, distribution, and characteristics of burned areas with higher reliability [18]. However, it is difficult to obtain full coverage of medium resolution satellite images suitable for burned area mapping over large areas in tropical regions, due to the combination of a relatively low image repetition rate (16 days for Landsat TM) and the potential cloud cover during the burning season [19]. At the same time, it is generally very time consuming to conduct wall-to-wall image analysis over large areas with acceptable accuracy. Sampling approaches have been proposed as a cost-effective solution to assess land cover and land cover changes from remote sensing data with sufficient accuracy [20], [21]. With a sample-based approach, statistical estimates can be derived while retaining the advantages of fine spatial detail in the data. The fine spatial detail allows implementing semiautomated image interpretation combined with visual classification or validation steps. The use of sample units reduces dramatically the workload and enables to derive estimates for large areas. Systematic sampling has been successfully used for regional, continental, and global tree cover change assessments [22] [25], but it has not been yet demonstrated for burned area monitoring over large areas. In this context, we propose a new method to assess burned areas with medium resolution satellite data acquired over a regular grid of samples (20 km 20 km in size) located at every full degree confluence of latitude and longitude, resulting in an approximate distance of 110 km between the sample site centers in north south and east west directions (this distance is only valid for low latitudes). Our objective is to perform a burned area assessment over the state of Mato Grosso, Brazil, for the burning season of year 2010 using Landsat satellite imagery at 30 m spatial resolution. Our results are compared to two different approaches: 1) wall-to-wall assessment derived from Landsat imagery and 2) from MODIS burned area product. This comparison allows evaluating the advantages and limitations of a sample-based approach using medium resolution imagery and its suitability for producing estimates of burned areas over large areas (e.g., at continental level). Fig. 1. Location of the study area: State of Mato Grosso with systematic sample plots. II. MATERIALS AND METHODS A. Study Area The study area corresponds to the Mato Grosso State, Brazil (Fig. 1), with a total area of km 2 [26]. Due to variable climate, terrain relief, precipitation patterns, and length of the dry season, the State of Mato Grosso comprises parts of three Brazilian biomes, the Amazon, the Cerrado, and a small portion of the Pantanal, and has a naturally very high biodiversity with many different vegetation types from dense evergreen forest to deciduous open forest, savannas, and natural grasslands. Furthermore, Mato Grosso State, located partly in the arc of deforestation area at the southern extent of the Brazilian Legal Amazon, has one of the highest annual deforestation rates in Brazil [27]. Forest clearance causes, among others, habitat fragmentation, which leaves the remaining forest more vulnerable to edge effects such as fire [28]. In consequence, the fire events are likely to happen more frequently at the border of deforestation fronts. Temporally, the fires in Mato Grosso State are highly concentrated in the dry season between June and October, with 80 95% of all yearly fire activity taking place during this time period [29]. B. Satellite Datasets 1) Landsat-5 TM Data: We use 42 Landsat-5 TM images to cover the 77 sample sites (20 km 20 km in size) of our study area, located at each full degree confluence of latitude and longitude. Regarding the image selection process, we aim to obtain cloud-free images acquired at the end of the burning season of year 2010 (i.e., during the months of September and October). However, due to increasing cloud cover at the end of the dry season, it was not always possible to find cloud-free Landsat data acquired during this period. The majority of the cloud free images selected for our study were acquired during the first half of September, as shown in Fig. 2. The same 42 Landsat- 5 TM images are then used for the wall-to-wall comparative assessment. 2) MOD/MYD14 Active Fire Dataset: The MODIS hotspots (MOD/MYD14 version 5.1) data were downloaded from the Fire Information for Resource Management System

3 SHIMABUKURO et al.: ESTIMATING BURNED AREA IN MATO GROSSO, BRAZIL 3 Fig. 2. Image acquisition dates for the sampling boxes (77) over the Mato Grosso State and MCD64A1 burned area (in km 2 ) accumulated in the burning season of (FIRMS) website ( accessed 27/08/2014) for the period of June October The fire detection is based on a contextual fire detection algorithm, which utilizes the 1-km resolution of MODIS thermal bands [30]. The MODIS satellites (Aqua and Terra) pass over Mato Grosso State four times a day, thereby, providing good information on the temporal and spatial distribution of fire activity in the study area. C. Methods 1) Semiautomatic Method Applied on Landsat Imagery Over 77 Sample Units: The method developed in our study is adapted from the approach used in the TREES-3 pan-tropical deforestation survey [24]. The satellite data preprocessing includes radiometric calibration, dehazing, spectral normalization, and cloud-masking [31], followed by a multistage image segmentation to create spatially and spectrally consistent mapping units (polygons) with a nominal minimum mapping unit (MMU) of 5 ha. The object-based TREES-3 process allows a maximum of 5% of the objects being smaller than the nominal MMU, with an absolute minimum object size of 3 ha [32]. In our segmentation approach, the smallest objects are created where the spectral differences between an area and its surroundings are largest which is the very often the case for a burned area in comparison with its neighboring areas. In consequence, the smallest detected burned areas have a size of 3 ha (approximately 33 Landsat pixels). We evaluated this approach by assessing the burned area objects smaller than 3 and 5 ha, created from Landsat imagery without any MMU threshold (described below). The burned area objects below 3 and 5 ha represent 0.007% and 0.026% of the total detected burned areas for Mato Grosso State, thus, are statistically not relevant in the estimate of burned areas for the whole area of interest. In each of the sample sites (20 km 20 km size), objects representing burned areas are automatically classified and corrected through visual checking with a dedicated tool [33]. The MODIS active fire detections (i.e., hotspots) were overlaid on the Landsat scenes and used as supporting information during this visual interpretation step. In many cases, the MODIS hotspots helped to verify burning of a suspected area unit (polygon). In case, no MODIS hotspots were detected inside a suspected polygon and burning could not be confirmed through the visual analysis, the area is considered as unburned. Fig. 3. Part of Landsat TM 226/069 showing the steps to map burned areas. (a) TM R5 G4 B3 color composite. (b) Shade fraction image overlaid by segmented polygons. (c) Unsupervised classification. (d) Red colored objects were assigned to burned class (blue color) and manually edited to correct especially the commission errors. 2) Comparison to Wall-to-Wall Landsat Mapping: We also produce a wall-to-wall map of burned areas over the whole study area using the same Landsat TM dataset to compare with the sampling approach. For this, we applied the methodology developed by [34] using SPRING software [35] for each individual TM scene. The method consists basically in: 1) generation of vegetation, soil, and shade fraction images; 2) segmentation of shade fraction images; 3) classification using nonsupervised classifier; 4) visual labeling of the unsupervised classes (into burned and unburned classes); and 5) edition of the classification (Fig. 3). The fraction (vegetation, soil, and shade) images are generated using the linear spectral mixing model [36] with the purpose to enhance the targeted features. In this study, we used shade fraction images because they enhance the burned areas (represented by objects with low reflectance values) and reduce considerably the data volume to be analyzed (42 TM scenes). The shade fraction images are segmented using growing region segmentation technique. This segmentation step is performed using the following parameters in SPRING: similarity threshold at 8 (Euclidean distance between the mean shade fraction values of two regions under which there are grouped together) and area threshold at 50 (minimum size of a region in number of pixels). An unsupervised classification is then performed from the spectral information of objects extracted from TM bands 3 5. The resulting spectral classes are assigned to the burned and unburned classes through a visual assessment of Landsat TM color composites (bands 5, 4, and 3 displayed as red, green, and blue where burned areas appear as dark objects) and of the shade fraction images (burned areas appear as bright objects). The resulting assigned classes are edited visually for single polygons to correct, in particular, the commission errors. Finally, the results of each of the 42 TM scenes are mosaicked

4 4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING TABLE I COMPARISON OF BURNED AREA ESTIMATES FOR THE BURNING SEASON OF 2010 * Standard error = standard deviation divided by the square root of the number of samples. ** Estimated based on the same ratio (total burnt area/burnt area detected by Landsat acquisition dates) derived from the MCD64A1 results. for the whole state of Mato Grosso. The resulting map is taken as a reference for evaluating the state-wide burned area estimate from the proposed sampling method. 3) Comparison to the MODIS Burned Area Product MCD64A1: The moderate resolution imaging spectrometer (MODIS) burned area product, MCD64A1, [4] is based on an automated burned area detection algorithm which takes advantage of an index derived from two 500 m MODIS bands (5 and 7), coupled with 1 km MODIS active fire (i.e., hotspot) observations (MOD/MYD14). The combined use of active fire and reflectance data enables the algorithm to adapt regionally over a wide range of pre- and postburn conditions and across multiple ecosystems [4]. The MCD64A1 product provides a date of burned event for each pixel, allowing to derive estimates of burned areas within each sampling box for the date of Landsat image acquisition. For comparison to the Landsat sampling-based results, we only considered the MCD64A1- derived burned areas before the acquisition date of the Landsat images, assuring that the MCD64A1 and the Landsat burned area estimates coincide temporally. The MCD64A1 product is also used to estimate the proportion of total burned area (for full dry season) missed in Landsat-based mapping due to the Landsat image acquisition dates. A proportion of burned area accumulated before and after the Landsat acquisition dates is derived from the MODIS MCD64A1 burned area product. III. RESULTS A. Landsat Sampling-Based Burned Area Estimation On average, the Landsat sample units (each covering an area of 400 km 2 ) contain 29.4 km 2 or 7.3% of burned area. This leads to an estimate of km 2 of burned area for the entire State of Mato Grosso (Table I). However, the burned areas are not evenly distributed across the state (Fig. 4), ranging from 0 to 307 km 2 within the sample units with a standard error of the mean at 6.5 km 2. On the state level, this corresponds to a standard error at km 2 (or 22%) of the estimated state-wide Fig. 4. Wall-to-wall Landsat TM versus sample-based mapping on Landsat TM. burned area. The relatively large standard error of the samplebased estimate reflects the uneven spatial distribution of burned areas within the state (Fig. 4). In order to further investigate the sensitivity of the sample-based estimate to the variation in burned extent mapped in individual sample sites, we performed a leave-one-out sensitivity analysis. We run the statistical estimates repeatedly by leaving one of the sample sites out, until all plots had been left out once. The mean state-wide burned area estimates in this sensitivity analysis varied between a minimum of km 2 (6.4%) and a maximum of km 2 (7.4%), with respective standard errors of and km 2. B. Comparison of the Sampling-Based Results to Wall-to-Wall Landsat Mapping In order to investigate further how the uneven burned area distribution affects the state-wide estimate derived from the sample, we also compared our estimate with a wall-to-wall burned area mapping. This wall-to-wall analysis (Fig. 4) leaded to an estimate of burned area at km 2, which is 6% higher than the estimate from the systematic sample ( km 2 ), but well within the standard error. This difference should be at least partially attributed to the sampling frame, since very high correlation (r 2 =0.91) is found between the sampling-based and wall-to-wall mapping burned areas within the sample units (Fig. 5). C. Comparison of the Sampling-Based Results to MODIS MCD64A1 Product Overall, a strong correlation (r 2 =0.85) is found between the two datasets, indicating a good agreement between the two mapping methods (Fig. 6, left). However, if we focus on small burned areas, i.e., < 100 km 2 per unit (Fig. 6, right), an expected pattern emerges. It can be clearly seen that the MODIS-derived product generally underestimates or entirely misses burned areas in these sites. This is assumed to be due to the insufficient spatial resolution of the MODIS sensor (500 m), as opposed to the 30-m resolution of the Landsat data, which allows the detection of small burned areas and, therefore, makes a more complete assessment feasible.

5 SHIMABUKURO et al.: ESTIMATING BURNED AREA IN MATO GROSSO, BRAZIL 5 Fig. 5. Correlation between Landsat sampling and wall-to-wall burned area mapping results (in km 2 ) within the sampling boxes. Fig. 7. Example of mapped burned area for four sample units (magenta) based on the sample-based mapping on Landsat TM and corresponding MODIS MCD64A1 product for the same sample units compared to the Landsat TM R5 G4 B3 color composite and wall-to-wall burned mapping (cyan). No burned areas were detected in the sample plot number 3 by both approaches. Fig. 6. Correlation between MCD64A1 and Landsat visually mapped burned area (in km 2 ) within the Landsat sampling boxes. Left image presents the full dataset (77 points) while the right image is zoomed into burned areas less than 100 km 2. For reference, a 1:1 line has been added to the right image. For better comparison, we aggregate the MCD64A1 burned areas product only up to the acquisition dates of the Landsat scenes in their respective coverage areas. As a result, we obtain a state-wide estimate of burned areas derived from MODIS at km 2, which is a significantly smaller area than the Landsat-derived sample-based estimate at km 2.Thisdifference is interpreted to be mainly caused by the small or fragmented burn scars which are undetectable at 500 m spatial resolution (Figs. 6 and 7). Although the different mapping approaches generally agree well on burned area distribution (Fig. 7), the difference in the spatial resolution leads to significant differences in state-wide estimates. The ratio between the MODIS-derived burned area for the full dry season and the MODIS-derived burned areas aggregated to the Landsat acquisition dates provides an indicative proportion of missed burned areas due to the acquisition dates of the Landsat images used. If we correct the Landsat samplebased estimate using this ratio (i.e., MODIS-derived burned area for the full dry season/modis-derived burned area up to the Landsat acquisition dates), we obtain a corrected burned area estimate for the entire burning season at km 2 (Table I). The difference between this corrected estimate and the Landsat sample-based burned area estimate highlights the potential effects of the Landsat images temporal distribution (Fig. 2). IV. CONCLUSION In this paper, we presented a new approach to map and assess burned areas from a systematic sample of medium resolution satellite data. The proposed sample-based method allows deriving statistically valid burned area estimates comparable to those of wall-to-wall classification of Landsat TM or MODIS satellite images. It provides fast and effective means to derive information on geographically large areas with Landsat-type medium spatial resolution images. The medium spatial resolution, the sampling approach, and the visual interpretation component in the classification process bring several advantages. There are higher chances to obtain cloud-free medium spatial resolution images over the sample units than to cover large regions in a wall-to-wall manner. The use of a sample enables to produce estimates from medium spatial resolution acquired over a short time-frame needed for burned area mapping [19], when it may not be possible to obtain a full coverage of satellite data. In addition, the processing and analysis of sample units is much less time consuming as compared to the use of a wall-to-wall approach for large areas. Moreover, a permanent sampling grid enables a comparison between successive assessments in the future. Existing samplebased schemes (e.g., the systematic sample of the global FAO Remote Sensing Survey) can potentially be used or adapted for burned area assessment. The medium spatial resolution allows including small burn scars in the estimates, which may constitute a significant proportion of the total area burned in some regions [17]. Furthermore, it is possible to visually analyze the attributes of each individual burned area with more thematic details (e.g., type of fire and potential prefire land cover). Such attributes can bring information on the fire regimes in the area of interest, which is essential for improved estimation of fire induced

6 6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING carbon emissions in REDD+ context. The proposed approach allows combining burned area estimates derived from moderate and coarse spatial resolution imagery such as MODIS or SPOT- VEGETATION in order to refine the estimate of total extent of burned area derived from moderate spatial resolution sensors. Such an approach combines the advantages of the medium spatial resolution provided by Landsat imagery over the sample sites and the wall-to-wall coverage provided by the coarse resolution mapping. The sampling design may cause bias in the statistical estimates depending on the spatial distribution of burned areas in the area of interest. A systematic sampling design, although easy to implement and beneficial for comparison of repeated analysis, may not be optimal for estimation of burned areas which may be unevenly spread over the study zone. Although in this study, only a 6% difference was noticed between the sample-based and wall-to-wall estimates, the relative high standard error (22% of the mean) indicates a potential vulnerability of the systematic sampling approach. However, a larger sample size would decrease significantly the standard error (see e.g., [31]). For follow up studies, the sampling grid could be intensified for state-level assessments. In this context, it would be worth to investigate stratified sampling to maximize the efficiency of the sampling approach, if the general distribution of burned areas is previously known (e.g., from a MODIS fire distribution). To increase the accuracy of burnt area estimates, the temporal distribution of fires also needs to be considered. The acquisition dates of the images should match well the end of the burning season. In our study, the Landsat sample-based approach leads to an underestimation of the total burned area for year 2010 due to fires that had taken place after the Landsat TM image acquisition dates. In this context, based on the assessment through the 2010 MODIS burned area product, we estimate that approximately 15% of the total burned areas were missed in the Landsat-based sample approach. To complicate the matter of image selection, the optimal imaging window at the end of the burning season may be very short due to the oncoming rainy season and fast disappearance of the signs of burning. However, these restrictions are less critical for a samplingbased approach than for a wall-to-wall approach using medium spatial resolution images. Overall, the results of our study demonstrate that medium spatial resolution satellites data are important sources of timely information for mapping burned areas. The proposed samplebased method allows to derive statistically valid burned area estimates comparable to those of wall-to-wall maps while being fast to implement over large regions. Our approach has been tested at a state-wide scale, but the results merit further investigation at continental and pan-tropical scales. Information on the extent and distribution of burned areas continues to be critical for regional and global environmental studies and for efforts to control such burning in the future. The ever increasing availability of medium spatial resolution satellite data offers novel opportunities to enhance regional level burned area monitoring, either alone or as supporting data for coarse resolution monitoring. These opportunities should be fully utilized to improve estimates of burned area extent, distribution and characteristics in the coming years. REFERENCES [1] G. R. van der Werf et al., CO 2 emissions from forest loss, Nat. Geosci., vol. 2, no. 11, pp , [2] C. O. Justice and S. Korontzi, A review of the status of satellite fire monitoring and the requirements for global environmental change research, in Global and Regional Vegetation Fire Monitoring from Space: Planning a Coordinated International Effort, F. J. Ahern, J. G. Goldammer, and C. O. Justice, Eds. The Hague, The Netherlands: SPB Academic Publishing, 2001, pp [3] J. M. Brustet, J. B. Vickos, J. Fontan, A. Podaire, and F. Lavenu, Characterization of active fires in West African savannas by analysis of satellite data: Landsat Thematic Mapper, in Global Biomass Burning: Atmospheric, Climatic, and Biospheric Implications, J. Levine, Ed. Cambridge, MA, USA: MIT Press, 1992, pp [4] L. Giglio, T. Loboda, D. P. Roy, B. Quayle, and C. O. Justice, An activefire based burned area mapping algorithm for the MODIS sensor, Remote Sens. Environ., vol. 113, pp , [5] A. W. Setzer and M. C. Pereira, Amazonia biomass burnings in 1987 and an estimate of their tropospheric emissions, Ambio, vol. 20, pp , [6] K. Tansey et al., A new, global, multi-annual ( ) burned area product at 1 km resolution, Geophys. Res. Lett., vol. 35, no. 1, p. L01401, [7] D. R. Cahoon, B. J. Stocks, J. S. Levine, W. R. Cofer, and K. P. Oneill, Seasonal distribution of African savanna fires, Nature, vol. 359, no. 6398, pp , [8] E. M. Prins and W. P. Menzel, Trends in South American biomass burning detected with the GOES visible infrared spin scan radiometer atmospheric sounder from 1983 to 1991, J. Geophys. Res., vol. 99, pp , [9] F. Mouillot et al., Ten years of global burned area products from spaceborne remote sensing A review: Analysis of user needs and recommendations for future developments, Int. J. Appl. Earth Observ. Geoinf., vol. 26, pp , [10] L. E. O. C. Aragão and Y. E. Shimabukuro, The incidence of fire in Amazonian forests with implications for REDD, Science, vol. 328, no. 5983, pp , [11] MCT Ministry of Science and Technology, Comunicação Inicial do Brasil à Convenção-Quadro das Nações Unidas sobre Mudanças do Clima. Brasília, Brazil: Ministério da Ciência e Tecnologia, [12] P. M. Fearnside, Uso da terra na Amazônia e as mudanças climáticas globais, Braz.J.Ecol., vol. 10, no. 2, pp , [13] M. A. Cochrane et al., Positive feedbacks in the fire dynamics of closed canopy tropical forests, Science, vol. 284, pp , [14] L. Giglio, I. Csiszar, and C. Justice, Global distribution and seasonality of active fires as observed with the Terra and Aqua moderate resolution imaging spectroradiometer (MODIS) sensors, J. Geophys. Res., vol. 111(G2), p. G02016, [15] V. R. Pivello, The use of fire in the Cerrado and Amazonian rainforests of Brazil: Past and present, Fire Ecol., vol. 7, no. 1, pp , [16] P. S. Laris, Spatiotemporal problems with detecting and mapping mosaic fire regimes with coarse-resolution satellite data in savanna environments, Remote Sens. Environ., vol. 99, pp , doi: /J.RSE [17] J. Miettinen and S. C. Liew, Burn scar patterns and their effect on regional burnt area mapping in insular Southeast Asia, Int. J. Wildland Fire, vol. 18, pp , doi: /WF [18] S. Bartalev, V. Egorov, V. Efremov, E. Flitman, E. Loupian, and F. Stytsenko, Assessment of burned forest areas over the Russian federation from MODIS and Landsat-TM/ETM+ imagery, in Global Forest Monitoring from Earth Observation, F. Achard and M. Hansen, Eds. Boca Raton, FL, USA: CRC Press, 2012, pp [19] J. Miettinen, E. Hyer, A. S. Chia, L. K. Kwoh, and S. C. Liew, Detection of vegetation fires and burnt areas by remote sensing in insular Southeast Asian conditions: Current status of knowledge and future challenges, Int. J. Remote Sens., vol. 34, pp , doi: / [20] C. E. Särndal, B. Swensson, and J. Wretman, Model Assisted Survey Sampling. Berlin, Germany: Springer-Verlag, [21] P. Mayaux, P. Holmgren, F. Achard, H. D. Eva, H.-J. Stibig, and A. Branthomme, Tropical forest cover change in the 1990s and options for future monitoring, Philosoph. Trans. Roy. Soc. London Ser. B Biol. Sci., vol. 360, pp , [22] E. J. Lindquist et al., Global forest land-use change , Food Agric. Org. United Nations and Eur. Commission Joint Res. Centre, Rome, Italy, FAO Forestry Paper 169, 2012, 53 p.

7 SHIMABUKURO et al.: ESTIMATING BURNED AREA IN MATO GROSSO, BRAZIL 7 [23] H.-J. Stibig, F. Achard, S. Carboni, R. Raši, and J. Miettinen, Change in tropical forest cover of Southeast Asia from 1990 to 2010, Biogeosciences, vol. 11, pp , [24] F. Achard et al., Determination of tropical deforestation rates and related carbon losses from 1990 to 2010, Global Change Biol., vol. 20, pp , [25] R. Beuchle et al., Land cover changes in the Brazilian Cerrado and Caatinga biomes from 1990 to 2010 based on a systematic remote sensing sampling approach, Appl. Geogr., vol. 58, pp , [26] IBGE, Estados, 2013 [Online]. Available: estadosat/perfil.php?sigla=mt, accessed on Aug. 7, [27] INPE, Monitoramento da cobertura florestal da Amazônia por Satélites. Sistemas PRODES, DETER, DEGRAD e QUEIMADAS , INPE, 2008, p. 47 [Online]. Available: Relatorio_Prodes2008.pdf, accessed on Aug. 7, 2015 [28] M. A. Cochrane and W. F. Laurance, Fire as a large-scale edge effect in Amazonian forests, J. Trop. Ecol., vol. 18, no. 3, pp , [29] L. O. Anderson et al., Disentangling the contribution of multiple land covers to fires-mediated carbon emissions in Amazonia during the 2010 drought, Global Biogeochem. Cycles, to be published. [30] L. Giglio, J. Descloitres, C. O. Justice, and Y. J. Kaufman, An enhanced contextual fire detection algorithm for MODIS, Remote Sens. Environ., vol. 87, pp , [31] C. Bodart et al., Pre-processing of a sample of multi-scene and multidate Landsat imagery used to monitor forest cover changes over the tropics, ISPRS J. Photogramm. Remote Sens., vol. 66, pp , [32] R. Raši et al., An automated approach for segmenting and classifying a large sample of multi-date Landsat imagery for pan-tropical forest monitoring, Remote Sens. Environ., vol. 115, pp , [33] D. Simonetti, R. Beuchle, and H. D. Eva, User manual for the JRC land cover/use change validation tool, Eur. Commission, Joint Res. Centre, Inst. Environ. Sustain., Luxembourg, Europe, EUR EN, 2011, p. 21. [34] Y. E. Shimabukuro et al., Fraction images derived from Terra MODIS data for mapping burned areas in Brazilian Amazonia, Int. J. Remote Sens., vol. 30, no. 6, pp , [35] G. Câmara, U. Freitas, R. C. M. Souza, and J. Garrido, SPRING: Integrating remote sensing and GIS by object-oriented data modeling, Comput. Graph., vol. 20, no. 3, pp , [36] Y. E. Shimabukuro and J. A. Smith, The least squares mixing models to generate fraction images derived from remote sensing multispectral data, IEEE Trans. Geosci. Remote Sens., vol. 29, no. 1, pp , Jan Yosio Edemir Shimabukuro received the B.S. degree in forestry engineering from the Federal Rural University of Rio de Janeiro (UFRRJ), Seropédica, Brazil, the M.S. degree in remote sensing from the Brazilian National Institute for Space Research (INPE), São José dos Campos, Brazil, and Ph.D. degree in forest and wood science from the Colorado State University, Fort Collins, CO, USA, in 1972, 1977, and 1987, respectively. From January 1992 to March 1994, he was a Visiting Scientist at the Biospheric Sciences Branch, NASA Goddard Space Flight Center, Greenbelt, MD, USA. From June 2013 to May 2015, he was a Visiting Scientist at the European Commission Joint Research Centre (EC-JRC), Ispra, Italy. Since 1973, he has been with INPE, studying satellite and ground-based remote sensing data for studying vegetation cover. His research interests include remote sensing and geographic information system techniques and environmental change detection models. Jukka Miettinen received the Ph.D. degree in agriculture and forestry from the University of Helsinki, Helsinki, Finland, in Between 2012 and 2015, he worked with the European Commission Joint Research Centre (EC- JRC), Ispra, Italy, concentrating on methodological development for tropical deforestation, forest degradation, and burned area monitoring. He has recently returned to the Centre for Remote Imaging, Sensing, and Processing (CRISP) at the National University of Singapore (NUS), where he coordinates the activities of the Terrestrial Ecosystem Monitoring Research Group. He has authored over 30 scientific peer-reviewed journal articles. His research interests include remote sensing based tropical forest, land cover and burned area monitoring. René Beuchle received the M.Sc. degree in cartography from Karlsruhe University of Applied Sciences, Karlsruhe, Germany, in In 2003, he joined the Joint Research Centre (JRC), Ispra, Italy, where he currently works in the Forest Resource and Climate Unit. His research interests include land cover mapping in the tropics and the assessment of forest cover changes with medium and high resolution optical remote sensing imagery and GIS, geographical focus on South and Central America, and remote sensing-based detection and monitoring of tropical forest degradation. Rosana Cristina Grecchi received the M.Sc. degree in geoenvironmental engineering from the University of São Paulo, São Paulo, Brazil, and the Ph.D. degree in remote sensing from the University of Sherbrooke, QC, Canada, in Between 2011 and 2013, she was a Postdoc Researcher with the Brazilian National Institute for Space Research (INPE), São José dos Campos, Brazil. Since 2013, she has been with the Joint Research Centre (JRC), European Commission, Ispra, Italy, as Postdoc Researcher with the Global Forest Resource Monitoring Project (ForObs). Her research interests include the analysis of forest cover changes (deforestation and forest degradation) in South America through the development of remote sensing-based monitoring techniques in combination with ancillary data. Dario Simonetti received the M.Sc. degree in informatics from Varese University, Como, Italy, in Since then, he joined the Institute for Environment and Sustainability, Joint Research Centre, Ispra, Italy, as a GIS Analyst Programmer. He is currently working with the Forest Resources and Climate Unit. His interests include development and assessment of processing chains for calibration, coregistration, topographic correction, and classification of remotesensed earth observation data; development of ad hoc geographic information systems application for interpretation/validation of satellite imagery classification across different epochs. Frédéric Achard received the Ph.D. degree in tropical ecology and remote sensing from Toulouse University, Toulouse, France, in He is a Senior Scientist with the Joint Research Centre (JRC), Ispra, Italy. He first worked in Department of Optical Remote Sensing, Institute for the International Vegetation Map (CNRS/University), Toulouse, France. In 1992, he joined the JRC and started a research activity in the framework of the TREES project. He has coauthored over 80 scientific peer-reviewed papers in leading scientific journals including Nature, Science, Global Change Biology, Global Biogeochemical Cycles, Remote Sensing of Environment, and the IEEE JSTARS. His current research interests include the development of Earth observation techniques for global and regional forest monitoring, and the assessment of the implications of forest cover changes in the Tropics and Boreal Eurasia on the global carbon budget.