Automatic Burned Land Mapping From MODIS Time Series Images: Assessment in Mediterranean Ecosystems

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 9, SEPTEMBER 2011 3401

Automatic Burned Land Mapping From MODISTime Series Images: Assessment in

Mediterranean EcosystemsAitor Bastarrika, Emilio Chuvieco, and M. Pilar Martín

Abstract—A novel automatic burned area mapping algorithmfor Mediterranean ecosystems based on Moderate-ResolutionImaging Spectroradiometer (MODIS) time series data is presentedin this paper. This algorithm is based on a two-phase approach.The first phase detects the most severely burned areas, usingspectral/temporal rules computed from dynamic temporal win-dows. The second phase improves the discrimination of burnedareas around those “seed” burned pixels using contextual algo-rithms based on edge detectors. The use of those filters improvedthe performance of the contextual algorithm, by refining the dis-crimination of fire perimeters while restricting the segmentationprocess. The algorithm was validated over six Mediterraneanregions during the fire season of 2003, where reference datawas generated using Landsat TM/ETM+ images. Omission andcommission errors were below 20%, with an overall Kappa valueof 0.846. The validation based on regression scattergraphs of5 × 5 km grids showed good agreement as well (R2 = 0.972).The standard MODIS burned area product (MCD45A1) showedlower accuracy than the proposed algorithm, with higher omissionerrors (38.6%) and lower Kappa (0.704) and R2 (0.838) values.

Index Terms—Burned areas, contextual algorithms, logis-tic regression, Moderate-Resolution Imaging Spectroradiometer(MODIS).

I. INTRODUCTION

MAPPING burned areas at the global scale is critical,since fire is one of the principal disturbance factors that

affects the structure of the ecosystems, carbon budgets, andnutrients cycles, as well as one of the significant causes ofgreenhouse emissions [1], [2]. At a regional scale, burned areamaps provide a spatial assessment of economic and environ-mental effects of the fires while facilitating improvements ofprefire planning [3] and post-fire rehabilitation efforts [4].

A large number of burned area mapping studies have beenpublished in the last two decades, based on both high and lowspatial resolution sensors [5]. At global scale, those studies canbe classified in two broad groups, namely, those that use the

Manuscript received June 2, 2010; revised November 3, 2010 andJanuary 14, 2011; accepted March 5, 2011. Date of publication April 29, 2011;date of current version August 26, 2011.

A. Bastarrika is with the Department of Surveying Engineering, Universityof the Basque Country, 01006 Vitoria-Gasteiz, Spain (e-mail: [email protected]).

E. Chuvieco is with the Department of Geography, University of Alcalá,28801 Alcalá de Henares, Spain (e-mail: [email protected]).

M. P. Martín is with the Center of Human and Social Sciences, NationalResearch Council, Spain (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2011.2128327

thermal contrast of active fires with the background and thosebased on post-fire reflectance changes. The former approach ismore reliable because spectral contrast related to the tempera-ture increase of the burning areas is higher than the temporaldifference between pre- and post-fire reflectance. However,mapping burned areas from active fire detection will only beaccurate when the fire is observed in short intervals, which islimited from current polar orbiting satellites. In addition, theclouds or the dense smoke caused by the fire can complicate thedetection of hot spots and only those fires warm and/or largeenough are confidently detected [6]–[8]. It is also difficult toestimate the actual proportion of a hot spot pixel that it is infact being burned.

The detection of burned areas based on pre- and post-firereflectance aims to identify changes caused by char and ashdeposition, as well as the reduction or elimination of the prefirevegetation. Both of these signals are weaker but last muchlonger than the thermal signal of active fires, thus providinga more reliable discrimination of the total area burned. Severalpapers have explored the combined use of the two methodolo-gies [9]–[13].

The Moderate-Resolution Imaging Spectroradiometer(MODIS), onboard of satellites Terra and Aqua, has beenextensively used in recent years for burned area mapping[12], [14]–[16]. MODIS provides a suitable balance betweenspatial (250, 500, and 1000 m), spectral [visible, near infrared(NIR), short-wave infrared (SWIR), and thermal infraredbands], and temporal resolution (up to two images per day,combining Terra and Aqua satellites) for burned land mappingat regional and global scales. Other significant advantagesare the geolocation accuracy of images and their free accessvia the servers enabled by National Aeronautics and SpaceAdministration. Data acquired by the MODIS sensor haveserved to develop a set of products [17] that include, amongothers, active fires at 1 km spatial resolution [6], [7] and, morerecently, a burned land product at 500 m spatial resolution(MCD45A1) [2], [14], [18]. The latter product is not yet fullyvalidated, but preliminary validation studies have shown goodresults in several ecosystems [19], [20].

The MCD45A1 product makes use of a set of fixedand dynamic thresholds that identify significant changes inreflectance. While this approach may be suitable at the globalscale, we hypothesise that it will not adapt well to areas withsmall fires and complex spatial patterns that are commonto Mediterranean ecosystems, in which locally adaptive

0196-2892/$26.00 © 2011 IEEE

3402 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 9, SEPTEMBER 2011

Fig. 1. MODIS tiles where the algorithm was calibrated and tested (h17v04, h17v05, h18v04, h18v05, h19v04, h19v05) and Landsat validation scenes (for thefire season of 2003).

algorithms might be more precise. To test our hypothesis,we assessed various approaches for mapping burned areas inMediterranean ecosystems using a two-phase algorithm, bothusing coarse- and high-resolution data [21]–[24]. This paperpresents the methods and the results obtained from applyingthis algorithm to several Mediterranean areas affected byfires. The input data are the MODIS bidirectional reflectancedistribution function (BRDF)-corrected images (MCD43A4,https://lpdaac.usgs.gov/lpdaac/products/modis_products_table/nadir_brdf _adjusted_reflectance/16_day_l3_global_500m/mcd43a4). The algorithm described in this paper includestwo phases. The first one aims to detect as many burnedpatches as possible while keeping a low commission error(false detection of burned areas). The second phase appliescontextual algorithms to refine the delimitation of those burnedpatches by decreasing the omission errors (missing real burnedareas). We hypothesised that using a two-phase algorithmwould reduce both commission and omission errors of the finalresult by focusing the spatial search of burned pixels (secondphase) only on those areas with a high probability of beingburned (results of the first phase). The validation was done withreference data obtained from Landsat Thematic Mapper (TM)and Enhanced Thematic Mapper Plus (ETM+), in six areasaffected by typical Mediterranean fires in Portugal, Spain,France, Croatia, and Bosnia.

II. STUDY AREAS AND DATA SET DESCRIPTION

A. Input Data Set

The input data set to develop the algorithm comprised fiveMODIS tiles for the 2000 and 2005 fire seasons (h17v04,h17v05, h18v04, h19v04, and h19v05) and one (h19v04) forthe 2007 fire season. These tiles cover almost fully the mostaffected countries by fires in Europe and provide a represen-tative set of burned and unburned land cover samples in theMediterranean region (Fig. 1).

Three basic MODIS products were used, namely, theNadir BRDF-Adjusted Reflectance (MCD43A4), the ther-mal anomalies (MOD14A2 and MYD14A2), and the stan-dard burned area product (MCD45A1) (https://lpdaac.usgs.gov/lpdaac/products/modis_products_table). The main inputdata for our algorithm was the MCD43A4, taken from Collec-tion 5 product. It included the first seven reflectance MODISchannels at 500 m spatial resolution and spanned 16 daysof Terra and Aqua combined data. This product is adjustedusing a BRDF that models the reflectance values as if theywere taken from nadir view and shown to be adequate toburn scar mapping in South America [25]—low commissionerrors—with the previous collection product (Collection 4) at1 km resolution. The MCD43A2 quality product was usedas well to mask areas without full inversion of the BRDFmodel in all bands and also to ensure processing of data fromland areas.

The MOD14A2 and MYD14A2, which include the thermalanomalies/fire detected by Terra and Aqua instruments, respec-tively, were used to extract temporal reflectance burned samplesand train the algorithm. Corine Land Cover 2000 was used toextract the land cover of the reference samples [26].

B. Test Data Set

The developed algorithm was tested using independent dataset of MODIS images acquired during the 2003 fire season.Previously mentioned products (MCD43A4 and MCD43A2)were downloaded for the h17v04, h17v05, h18v04, h18v05,h19v04, and h19v05 MODIS tiles which the algorithm de-veloped in this paper was applied. Moreover, the active fireproducts (MOD14A2 and MYD14A2) were used to comparetheir capability in the detection of fires against the methodologydeveloped in this paper. The standard MCD45A1 product wasused as well to compare its performance with the results of ouralgorithm.

BASTARRIKA et al.: AUTOMATIC BURNED LAND MAPPING FROM MODIS TIME SERIES IMAGES 3403

TABLE ILANDSAT TM/ETM+ ACQUISITIONS USED TO PRODUCE REFERENCE DATA SETS

TABLE IIFIRE OCCURRENCE ON THE REFERENCE SITES

The MODIS products (both reference and test data) weredownloaded from the Web Warehouse Inventory Search Tool(WIST) (https://wist.echo.nasa.gov/~wist/api/imswelcome/).The temporal series used for our algorithm spanned the mainfire season in Mediterranean areas, from the 1st of May tothe 31st of October: 24 8-day composites for each year (theMCD43A4 product has a temporal overlap of 8 days), productand tile, with the exception of the MCD45A1 that is a monthlyproduct (six composites per each fire season).

C. Burned Area Reference Data

The six sites from the 2003 fire season used to validatethe algorithm are shown in Table I and Fig. 1. All burnedarea perimeters were generated using Landsat TM/ETM+ timeseries. All the scenes available between the 1st May and 31stOctober were downloaded from the U.S. Geological Survey(USGS) Glovis site (http://glovis.usgs.gov/). The time spanbetween scenes varies with the site between 3 and 5 months.

The reference burned area perimeters were obtained with anautomatic burned area mapping algorithm that uses a similartwo-phase methodology to the one presented in this paper, butadapted to the spectral and spatial characteristics of LandsatTM/ETM+ sensors [24]. The results were checked visually,

confirming each burned patch, reshaping them to the realboundaries in case of errors, and adding those patches omittedin the automatic process. Perimeters of areas smaller than 5 hawere rejected. This size threshold was considered appropriateto validate MODIS input data with a pixel size of 25 ha. Theareas containing clouds, or cloud shadows, and croplands weremasked out for the validation exercise.

The reference study sites are diverse (Table II) con-taining areas with high fire incidence, like Croatia–Bosnia189/029 and Portugal–Spain 203/034 with about 80 000 haand Portugal–Spain 203/032 with approximately 190 000 ha.The rest of the studied areas had less than 21 000 ha.In Croatia–Bosnia 189/029, France–Italy 193/031, France195/30, and Portugal–Spain 203/034, fire affected areas cov-ered mostly by shrubs between 44.5 and 75.9%, while inother areas, forest was the most affected land cover (between53.8 and 59.1%). The higher proportion of burned areas incroplands were found in Portugal–Spain 203/032 (15.5%) andin Croatia–Bosnia 189/029 (33.4%).

III. METHODS

Fig. 2 shows a flowchart summarizing the methodologyproposed in this paper.


Fig. 2. Flowchart of the burned area algorithm development and validation.

A. Data Preprocessing

The downloaded MODIS products were already geomet-rically corrected. This paper was intended for the standardspatial resolution of MCD43A4 (500 m); as a consequence,the MOD14A2 and MYD14A2 products were resampled from1 km to 500 m. The original reference and coordinate sys-tem were kept (sinusoidal cartographic projection and WGS84datum).

Landsat images downloaded were already geometricallycorrected, so they were only converted to top-of-the theatmosphere reflectance using calibration coefficients includedin [27].

B. Variable Selection

To carry out the two phases of our algorithm, different setsof input bands were considered, both including the originalbands and several spectral indexes that have been previouslyused in the literature to discriminate burned areas [16], [23],[28]–[34]. Three spectral domains were considered: Red–NIR,NIR–SWIR, and SWIR–SWIR.

At the Red–NIR spectral space, Normalized DifferenceVegetation Index (NDVI) [35] and Global Environmental

Monitoring Index (GEMI) [36] were computed as follows:

NDVI =ρNIR − ρRρNIR + ρR

GEMI =η(1− 0.25η)− (ρR − 0.125)

(1− ρR)

η =2(ρ2NIR − ρ2R

)+ 1.5ρNIR + 0.5ρR

(ρR + ρNIR + 0.5)

where ρNIR is the reflectance in the NIR (MODIS Band 2); andρR is the reflectance in the Red (MODIS Band 1).

In the NIR–SWIR spectral space, the Normalized Burn Ratio(NBR) [37] and the Modified Burned Area Index (BAIM) [16]were calculated as follows:

NBR =ρNIR − ρLSWIR

ρNIR + ρLSWIR

BAIM =1

(ρNIR − ρcNIR)2 + (ρLSWIR − ρcLSWIR)2

where ρNIR is the reflectance in the NIR; ρLSWIR is thereflectance in the long SWIR (MODIS Band 7); and ρcNIR

and ρcLSWIR are the convergence values (0.05 and 0.2,respectively).


Within the SWIR spectral space, Mid-Infrared Burned Index(MIRBI) [38] index was computed as follows:

MIRBI = 10ρLSWIR − 9.8ρSSWIR + 2

where ρLSWIR is the reflectance in the long SWIR (MODISband 7); and ρSSWIR is the reflectance in the short SWIR(MODIS band 6).

C. Phase 1: Determination of “Core Burned” Pixels

The methodology to determinate the seed pixels in the al-gorithm followed decision tree classification techniques, whichhave been widely used before for burned area mapping [39]–[41], as well as in land cover mapping [42], [43]. Decisiontrees were built iteratively by dividing the training samples(burned and unburned pixels in this case) using optimizeddiscrimination of the spectral space of the input bands/indexes.Various algorithms are available to create decision trees, differ-ing mostly on the strategies used to simplify the trees (pruning)and the rules to split the nodes. In this paper, we used the QuickUnbiased Efficient Statistical Trees (QUEST) algorithm [44],based on binary partitions, which proved to be very robust andcomputationally efficient [45]. The default options/values wereused as follows: significance level 0.05, five maximum partitionlevels, pruning off, and initial probability computed from thesame training pixels.

To take advantage of the high temporal resolution of theMODIS images in the detection of post-fire changes, the valueof the same pixel was extracted from the reference data set inseven consecutive MCD43A4 data sets. For burned samples,this seven consecutive composite temporal window was cen-tered in the composite where a hot spot had been identified withhigh confidence (in the MOD14A2 or MYD14A2 products).For unburned pixels, the seven consecutive composite temporalwindow was randomly centered in areas without active firedetection, and considering the 24 composites between Mayand October. For burned areas, 9000 pixels were sampled(same proportions in forest, scrub, and cropland), while thesample size was 510.000 pixels for unburned areas. The spectralindexes previously described were computed for each of thosepixels, along with some temporal statistics (Fig. 3):

1) The standard deviation for the central three composites(t− 1, t, and t+ 1 composite) named std_3comp, thestandard deviation for the central five (t− 2, t− 1, t,t+ 1, and t+ 2 composite) named std_5comp, and thestandard deviation for the seven values (t− 3, t− 2, t−1, t, t+ 1, t+ 2, and t+ 3 composite) named std_7comp.

2) The mean of the central and the next composites (t andt+ 1 composite) named mean_2comp, the mean of thecentral and the next two composites (t, t+ 1, and t+ 2composite) named mean_3comp, and the mean of thecentral and the next three composites (t, t+ 1, t+ 2, andt+ 3 composites) named mean_4comp.

The selection of the standard deviation as a discriminationvariable tried to improve sensitivity to detect post-fire re-flectance changes, as burned pixels should have higher temporalvariability than unburned samples. The post-fire mean statistic

Fig. 3. Seven consecutive composites and temporal statistics representation tocompute the explanatory variables introduced in the classification tree for thefirst phase of the algorithm.

aimed to assess the signal preservation after the fire. For eachband and all the spectral indexes, these six temporal statisticswere included to train the classification tree.

Since this paper aimed to classify burned pixels with thelowest possible commission error, only the terminal node withthe maximum discrimination power and the minimum errorwere selected from the decision tree. The criteria to set theseed phase combined the rules from the root of the tree untilthe selected final node.

D. Phase 2: Shaping the Burned Areas: Contextual Algorithm

The second phase of the burned land mapping algorithmaimed to reduce the omission errors and improve the delineationof fire patches, by analyzing the spectral contrast of pixels in thevicinity of those detected as burned in the first phase. For doingso, we had first to define which input variable was going toserve as input for the contextual analysis, and afterward, selecta suitable region growing algorithm to accept or not neighborpixels as actually burned.

Several variables were evaluated for the contextual analysis.The selected variable should have two properties, on one hand,a high separability between burned and unburned areas, onthe other, a low variance within the burned areas, to avoidpotential omission errors in the automatic growing process.Previous studies had identified variables that might complywith our selection criteria, such as spectral mixture analysis[46], [47], classification trees [48], logistic regression models[49], or spectral index-based thresholds, such as the BAI [21],GEMI [23], and BAIM [22]. We chose the logistic regressionapproach since it complies with other criteria and provided asound probabilistic framework for region growing algorithms.Logistic regression analysis has also been extensively used withLandsat TM images [50]–[52]. The samples used to computethe logistic regression were the same as those used for the firstphase of our algorithm. However, in this case, the number ofsamples for burned and unburned classes was the same, to keepthe discrimination threshold equilibrated. Temporal statisticswere not included in this computation to reduce spatial variabil-ity within the burned patches. For the same reason, only values


of the t+ 1 and t− 1 composite (the pre- and post-fire com-posites) were used. The model was computed with a stepwiseforward process using a significance level of 0.05 and 60% ofthe samples, whereas the remaining 40% were used for testing.

The contextual algorithm used was based on region grow-ing analysis [53]. This algorithm iteratively adds pixels inthe neighborhood of previously detected pixels. Instead ofsegmenting the whole image as the classical region grow-ing algorithms do, this algorithm segments only the burnedclass. In our iterative process, any of the 8 pixels (within a3 × 3 connectivity window) in the neighborhood of the seedsthat had a logistic probability value higher than 0.5 were labeledas burned. In a subsequent step, pixels previously selected wereconsidered as seeds for the next iteration. The segmentationprocess finished when no more candidate pixels were found. Inorder to improve the segmentation process in those areas withsharp changes in burned probability values, a border approachwas added, which does not take account for the candidatepixels (within the 9-pixel neighborhood of the seed or a pixellabeled as burned) if any of them is labeled as a border. Thisborder cartography is obtained over the logistic regressionoutput probability layer by applying first a Gaussian low-passfilter to reduce the local heterogeneity and an edge filter later(Sobel filter), binarizing it with an automatic threshold basedon the root mean square estimate of the image noise. Finally, amorphological operator is used to remove interior pixels leavingonly the boundary pixels.

E. Temporal Integration of the Two Phases

For each seven temporal composite windows (named t− 3,t− 2, t− 1, t, t+ 1, t+ 2, and t+ 3 composites), the coreburned (seeds) pixels were detected by applying the classifi-cation tree-based criteria described in 3.3 (first phase). Thelogistic regression model to be used by the border approachcontextual algorithm (second phase) was obtained using thet+ 1 and t− 1 temporal composites. The pixels detected asburned were labeled with the middle day of the central (t)temporal composite. In the next step, the temporal windowwas moved one composite ahead, applying the same two-phaseintegration and repeating the process until the t+ 3 compositewas the last temporal composite. In case that a pixel had alreadybeen detected as burned in previous temporal windows, the firstday of the detection was kept.

F. Application of the Algorithm and Validation

Once the algorithm was trained with our test sites, it wascomputed for the validation sites for the period between the 1stof May and the 31st of October, 2003. The burned patches de-tected by the algorithm were vectorized and compared againstthe reference data. The detection ability of the seeding phasewas evaluated by computing the percentage of fires correctlyidentified for three different sizes (smaller than 100 ha, from100 to 500 ha, and bigger than 500 ha). Commission errors(percentage of false burned area detected) and omission errors(percentage of true burned area undetected) were computedas well. Omission errors were expected to be high in the first

phase, because strict rules were applied to identify undoubtedlythe most clearly burned pixels, thus avoiding commission er-rors. Detection rate of our algorithm was also compared withthat of the active fire products (MOD14A2 and MYD14A2combined).

The validation of the second-phase (the final results) was as-sessed from two criteria. One was the omission and commissionerrors of the burned category and the traditional Kappa index[54]. The other was the linear regression between the resultsof the algorithm and the reference estimations of burned areapercentage in a 5 × 5 km grid. This criterion minimized theimpact of geometric misregistration between the input data andthe reference data [55]. In this part of the validation, only thegrids with at least 90% of valid area were considered. Individualvalidation for each study area was carried out, but also thejoined statistics were computed, considering the six study areasaltogether.

IV. RESULTS

The following set of rules, based on the seven consecutivecomposite windows, was extracted from the classification treeused to establish the burned seeds in the first phase:

(BAIM_Std_3comp > 10.06) AND

(NBR_Std_5comp > 0.13) AND

(MIRBI_Std_7comp > 0.30) AND

(NBR_Mean_4comp <= 0.30).

where BAIM_Std_3comp is the standard deviation of theBAIM for three temporal periods; NBR_Std_5comp is thestandard deviation of the NBR for five temporal periods;MIRBI_Std_7comp is the standard deviation of MIRBI forseven temporal periods; and NBR_Mean_4comp is the meanvalue of the NBR for four temporal periods.

Table III shows the results of this first phase for thedifferent validation sites Almost 100% of the fires biggerthan 500 ha were detected, and only one fire was missedin the Portugal–Spain 203/032 validation site. The detectionrate for 100 to 500 ha fires was lower, between 52.6% forPortugal–Spain 203/034 and 80% in Spain 197/31, but withrates around 60% were all the study sites considered altogether.Those fires smaller than 100 ha are clearly underdetected, witha maximum detection rate of 8.2% for Spain 197/31 consideringthe six individual study sites and 2.1% for all the study areas.The commission errors of the first phase were low, with amaximum error of 6.6% in Spain 197/31 and 5.4% for thejoined area.

For comparative purposes, Table IV shows the detectionability if the active fires had been applied as seeds in the firstphase of this algorithm. Active fires showed similar detectionrates than our results for fires larger than 500 ha, although,in Portugal–Spain 203/032, the active fires increased the de-tection rate over our results in almost 9%, and decreased inCroatia–Bosnia 189/029, France 195/30, and Spain 197/31between 11.5 and 40% for those burned areas between 100and 500 ha. The detection rate of active fires was significantlybetter than our results for smaller fires (< 100 ha), with anincrease up to 29% in three study areas that resulted in the


TABLE IIIVALIDATION RESULTS FOR THE FIRST PHASE USING THE CRITERIA EXTRACTED FROM THE CLASSIFICATION TREE

TABLE IVDETECTION RATE USING THE ACTIVE FIRES FROM THE UNION OF THE MOD14A2 AND MYD14A2 PRODUCTS

absolute detection rate of 24.1% for all study areas. However,the commission errors were considerably higher than our resultsin all sites, with a maximum error of 77.4% in France 193/31and 57.3% when considering all sites combined (Fig. 4).

A. Results of the Second Phase

The logistic regression model for the contextual analysis was(see Table V for details): (See equation later)

According to the Wald test, all explanatory variables in-troduced in the model were significant (p < 0.01, Table V).Following the −2 log likelihood change the main variables ofthe model were post-fire NBR, followed by prefire NDVI, post-fire MIRBI, and prefire NBR. The model correctly predictedmore than 90% of burned and unburned sampled pixels.

The outputs of the region growing algorithm showed sig-nificant improvements over the first phase in terms of bothomission errors and global accuracy (Table VI). The increasein Kappa index was significant in each of the six studysites (p < 0.05). Omission errors decreased in the range of19–34% for the six study sites (a decrease of 28% whenthe joined study area was considered), while commission er-rors increased 14% in total. The MCD45A1 product providedlower commission errors than our algorithm in almost all thestudy sites (Table VII), with the highest difference in theSpain 197/31 site, and a total decrease of 3% in the joinedstudy site. Conversely, the omission errors of the standardMODIS product were significantly higher than our results inall the study sites, ranging between 3 and 8%. Taking intoaccount all sites together the increase was of 22% (see Figs. 5and 6). Both Kappa and linear regression statistics showed a

P =1

1 + e−(16.691·NDV I(t−1)−25.595·NBR(t+1)+9.024∗NBR(t−1)+5.688∗MIRBI(t+1)−14.195)


Fig. 4. (Left) Seeds detected by the criterions extracted by the classification tree and (right) MOD14A2-MYD14A2 active fire detections product in the Croatiaand Bosnia (Croatia–Bosnia 189/029) during the summer of 2003.

TABLE VESTIMATED PARAMETERS AND SIGNIFICANCE FOR THE EXPLANATORY VARIABLES SELECTED IN THE LOGISTIC

REGRESSION MODEL DEVELOPED FOR THE SECOND PHASE OF THE ALGORITHM

better agreement between our results and the reference datathan for the results of the MODIS product. Our algorithmhad an overall Kappa of 0.846, whereas it was 0.704 forMCD45A1, with determination coefficients of 0.972 and 0.838,respectively.

V. DISCUSSION

We have presented a two-phase automatic methodologyfor mapping burned areas using MODIS BRDF-correctedreflectance imagery (MCD43A4). The method was developedusing a reference data set and validated with six regionalsites, all located in Mediterranean environments affected byforest fires.

A two-phase approach algorithm aims to balance out omis-sion and commission errors, thus reducing problems associatedwith the over- and underestimation of burned areas. The firstphase tries to identify the most clearly burned pixels within eachburned patch, while the second phase improves the delimitationof the burned area by analyzing the spatial context around thosepixels established to be burned in the first phase. The two phases

have different requirements; the first one should avoid thosepixels that are partially or slightly burned, because they willprobably create confusions with some unburned land covers(cloud shadows, shades, water, and partially vegetated darksoils are very common ones to create confusions). The secondphase focuses on the vicinity of “core” burned pixels, usingregion growing and edge detector algorithms and improvesomission errors.

The first phase of our approach could be based on activefires, which is a currently derived from MODIS and othersatellite sensors [Along Track Scanning Radiometer (ATSR)and Spinning Enhanced Visible and Infrared Imager (SEVIRI)].In this case, we used the standard MOD14A2 and MYD14A2products. The paper has briefly compared this approach withours, based on decision tree classification. The classificationtrees provide numerous final nodes, but only the one presentingthe lowest commission error was selected. The final goal was toobtain a set of simple and robust rules, not overfitted to the ref-erence data and adequately adapted to the diverse conditions ofthe Mediterranean environments. The predictive variables usedin this paper considered spectral and temporal variation. The


TABLE VIOMISSION AND COMMISSION ERRORS, KAPPA VALUES AND LINEAR REGRESSION STATISTICS FOR THE SECOND-PHASE OF THE ALGORITHM

TABLE VIIOMISSION AND COMMISSION ERRORS, KAPPA VALUES AND LINEAR REGRESSION STATISTICS FOR THE MCD45A1 PRODUCT

former was tested through a set of spectral indexes, whereasthe latter depended on the variation/change of statistics ascomputed from the temporal standard deviation (consideringthree, five, and seven consecutive composites), as well as fromthe post-fire tendencies (considering the mean of two, three, andfour weeks). The set of rules extracted from the classificationtree included the NIR spectral region (Band 2) and SWIR(Band 6 and 7) within the variables of temporary change BAIM(in a window of three composites), NBR (in a window of fivecomposites), and MIRBI (in a window of seven composites).The BAIM index was the main rule in the tree, and though thecriterion established (> 10.06) was quite loose, it showed highperformance to eliminate confusions with soils and water bod-ies, although it did not solve discrimination problems with some

shrublands and croplands. In previous studies, we observed thata more strict threshold produced higher fire detection rates withvery low commission errors in Portugal and Spain [16]. How-ever, when applied to our test sites, a notable decrease in the firedetection rate has been observed for Eastern Europe areas. Thisfact might be related to the definition of the convergence pointfor the BAIM, which was originally derived from burned sitesin the Iberian Peninsula [16]. The temporary criterion based onNBR (NBR_Std_5comp > 0.13) was more strict that the onedefined for the BAIM and decreased the confusion with crop-lands and shrubs. The temporary criterion based on the MIRBIindex, calculated from a temporal window of seven periods(MIRBI_Std_7p > 0.30) was the most strict and determinant:it caused a reduction of the fire detection rates, while reducing


Fig. 5. (Left) Burned area cartography obtained by the algorithm proposed in this paper and (right) MCD45A1 product in Corsica island (France 193/31) duringthe summer of 2003.

Fig. 6. (Left) Burned area cartography obtained by the algorithm proposed in this paper and (right) MCD45A1 product in Portugal (Portugal–Spain 203/032)during the summer of 2003.

the confusion in all land covers. Previous works already hadobserved that MIRBI index increased the omission error inMediterranean areas [16], though this effect was not so impor-tant in the first phase of the algorithm. The last criterion basedon post-fire NBR average values turned out to be a soft criterionthat only filtered out only some residual errors in croplands.

Although these criteria extracted from the classification treeshowed very low commission errors in the reference data set(smaller than 2%) as much as in all the six sites validated(around 4%), it was remarkable to observe that most confusionoccurs with agricultural land covers. This was especially truefor the Eastern Europe sites, which have extensive arable areas.However, these confusions were not considered critical for theycan be masked out from land cover maps. In addition, these firesare not relevant in terms of economical and ecological impactsas those occurring in shrubs and forested areas.

The detection ability of our classification approach wasclearly lower than active fire products for fires smaller than100 ha, but it showed significantly less commission errors(> 40% less). Those commission errors of the active fires prod-uct may be related to some of the problems previously identifiedby other authors [7]. Also, the hotspot product showed lowerdetection ability than our approach for fires between 100 and500 ha. Therefore, it was considered not suitable for the seedingstage of the algorithm.

The logistic regression model developed as the pillar ofthe region growing process included the explanatory variablesNBR and post-fire MIRBI, as well as the prefire NDVI andNBR, which decreased confusions with soils, urban, and othernoncombustible areas. The output probability derived fromthis model showed a good visual separability between theburned and unburned categories, with the exception of some


Fig. 7. Scatter plot of the proportions of 5 × 5 km cells labeled as burned bythe algorithm developed in this paper plotted against the proportion labeled asburned by the Landsat-based reference data.

land covers, mostly arable croplands at the end of fire season(September–October) and dark bare soils. As the output ofthe logistic regression model is a continuous probability valuebetween 0 and 1, the optimum cutoff value for the regiongrowing process had to be defined. For this purpose, edgefilters showed to be quite effective, making the growing processless dependent of the threshold as they indentified pixels withsharp neighbor changes. Moreover, they facilitated a measure ofspatial texture that enhanced the performance of the algorithmin unburned areas that showed intermediate probability of beingburned and high texture.

The outputs of this second phase improved considerably theburned area detected in the first phase, resulting in a closeagreement with the validation data. The overall regressionslope between our results and reference perimeters was 1.103,with R2 of 0.972 (see Fig. 7). A slight overestimation wasobserved, especially for grids with high portion of burned area.These effects might be related to the lower spatial resolution ofMODIS data (500 against 30 m of reference images), especiallyin the borders of the fire and also to internal islands which arenot burned and are not detectable with low-resolution data.

The MCD45A1 product showed a significant underestima-tion of the burned areas resulting in a slope of 0.793 and a R2

of 0.838 (Fig. 8) and a considerable number of highly burnedpercentage grids underdetected.

VI. CONCLUSION

This paper has aimed to develop an automatic method to mapburned areas from MODIS images in Mediterranean environ-ments. The method is based on a two-phase approach, whichtries to split the reduction of omission and commission errorsin different phases. Conceptually, the reduction of both typesof errors requires a different focus, since commission errors are

Fig. 8. Scatter plot of the proportions of 5 × 5 km cells labeled as burned bythe MCD45A1 burned-area product plotted against the proportion labeled asburned by the Landsat-based reference data.

related to the internal diversity within burned areas (differentsignals caused by different levels of fire damage), while theomission errors are more associated to the external contrastwith other land covers. Consequently, reducing commissionerrors requires techniques than focus on the most clearly burnedpixels, while reduction of omission errors implies to establish aproper border with unburned areas.

As for the methodology proposed in this paper, the firstphase established the burned seeds using spectral/temporalrules extracted by a large number of reference samples usinga classification tree, whereas region growing algorithms inthe second phase improved the delimitation of burned patchesfrom previously detected seeds. The final results showed goodagreement with reference data and can be considered suitableto be implemented operationally in Mediterranean regions. Theresults obtained were more accurate than those provided by theMCD45A1 product, with significantly lower omission errors.The limitations of our method were observed in cropland areas,particularly when they had dark soils, as well as for small fires(< 100 ha), where the detection was not shown reliable.

REFERENCES

[1] P. J. Crutzen and M. O. Andreae, “Biomass burning in the tropics: Impacton atmospheric chemistry and biogeochemical cycles,” Science, vol. 250,no. 4988, pp. 1669–1678, Dec. 1990.

[2] D. P. Roy, L. Boschetti, C. O. Justice, and J. Ju, “The collection 5 MODISburned area product—Global evaluation by comparison with the MODISactive fire product,” Remote Sens. Environ., vol. 112, no. 9, pp. 3690–3707, Sep. 2008.

[3] C. Giri and S. Shrestha, “Technical note: Forest fire mapping in HuayKha Khaeng Wildlife Sanctuary, Thailand,” Int. J. Remote Sens., vol. 21,no. 10, pp. 2023–2030, 2000.

[4] A. Vafeidis, N. A. Drake, and J. Wainwright, “Regional estimation of post-fire erosion using remote sensing and GIS: An example from Greece,”presented at the Remote Sensing Society (RSS) Adding Value RemotelySensed Data, Leicester, U.K., 2000.


[5] E. Chuvieco, “Satellite observation of biomass burning: Implicationsin global change research,” in Earth Observation and Global Change,E. Chuvieco, Ed. New York: Springer-Verlag, 2008, pp. 109–142.

[6] C. O. Justice, L. Giglio, S. Korontzi, J. Owens, J. T. Morisette, D. Roy,J. Descloitres, S. Alleaume, F. Petitcolin, and Y. Kaufman, “The MODISfire products,” Remote Sens. Environ., vol. 83, no. 1/2, pp. 244–262,Nov. 2002.

[7] L. Giglio, J. Descloitres, C. O. Justice, and Y. J. Kaufman, “An enhancedcontextual fire detection algorithm for MODIS,” Remote Sens. Environ.,vol. 87, no. 2/3, pp. 273–282, Oct. 2003.

[8] L. Giglio, “Characterization of the tropical diurnal fire cycle using VIRSand MODIS observations,” Remote Sens. Environ., vol. 108, no. 4,pp. 407–421, Jun. 2007.

[9] D. P. Roy, L. Giglio, J. D. Kendall, and C. O. Justice, “Multi-temporalactive-fire based burn scar detection algorithm,” Int. J. Remote Sens.,vol. 20, no. 5, pp. 1031–1038, 1999.

[10] R. H. Fraser, Z. Li, and R. Landry, “SPOT VEGETATION for character-izing boreal forest fires,” Int. J. Remote Sens., vol. 21, no. 18, pp. 3525–3532, 2000.

[11] R. H. Fraser, Z. Li, and J. Cihlar, “Hotspot and NDVI DifferencingSynergy (HANDS): A new technique for burned area mapping overBoreal forest,” Remote Sens. Environ., vol. 74, no. 3, pp. 362–376,Dec. 2000.

[12] L. Giglio, T. Loboda, D. P. Roy, B. Quayle, and C. O. Justice, “An active-fire based burned area mapping algorithm for the MODIS sensor,” RemoteSens. Environ., vol. 113, no. 2, pp. 408–420, Feb. 2009.

[13] S. Merino-de-Miguel, M. Huesca, and F. González-Alonso, “MODIS re-flectance and active fire data for burn mapping and assessment at regionallevel,” Ecol. Model., vol. 221, no. 1, pp. 67–74, Jan. 2010.

[14] JinD. Roy, Y. , P. E. Lewis, and C. O. Justice, “Prototyping a globalalgorithm for systematic fire-affected area mapping using MODIS timeseries data,” Remote Sens. Environ., vol. 97, no. 2, pp. 137–162, Jul. 2005.

[15] T. Loboda, K. J. O’Neal, and I. Csiszar, “Regionally adaptable dNBR-based algorithm for burned area mapping from MODIS data,” RemoteSens. Environ., vol. 109, no. 4, pp. 429–442, Aug. 2007.

[16] M. P. Martín, I. Gómez, and E. Chuvieco, “Burnt Area Index (BAIM) forburned area discrimination at regional scale using MODIS data,” ForestEcol. Manage., vol. 234, p. S221, 2006.

[17] C. O. Justice, J. R. G. Townshend, E. F. Vermote, E. Masuoka,R. E. Wolfe, N. Saleous, D. P. Roy, and J. T. Morisette, “An overview ofMODIS land data processing and product status,” Remote Sens. Environ.,vol. 83, no. 1/2, pp. 3–15, Nov. 2002.

[18] D. Roy, P. E. Lewis, and C. O. Justice, “Burned area mapping using multi-temporal moderate spatial resolution data—A bi-directional reflectancemodel-based expectation approach,” Remote Sens. Environ., vol. 83,no. 1/2, pp. 263–286, Nov. 2002.

[19] D. Roy, L. Boschetti, and O. Kelley, MODIS Collection 5 Burned AreaProduct MCD45 User’s Guide. College Park, MD: Univ. Maryland,2006.

[20] L. Boschetti, D. P. Roy, P. Barbosa, R. Boca, and C. O. Justice, “A MODISassessment of the summer 2007 extent burned in Greece,” Int. J. RemoteSens., vol. 29, no. 8, pp. 2433–2436, Apr. 2008.

[21] M. Garcia and E. Chuvieco, “Assessment of the potential ofSAC-C/MMRS imagery for mapping burned areas in Spain,” RemoteSens. Environ., vol. 92, no. 3, pp. 414–423, Aug. 2004.

[22] E. Chuvieco, S. Opazo, W. Sione, H. Del Valle, J. Anaya, C. Di Bella,I. Cruz, L. Manzo, G. López, N. Mari, F. González-Alonso, F. Morelli,A. Setzer, I. Csiszar, J. A. Kanpandegi, A. Bastarrika, and R. Libonati,“Global burned-land estimation in Latin America using MODIS compos-ite data,” Ecol. Appl., vol. 18, no. 1, pp. 64–79, Jan. 2008.

[23] E. Chuvieco, P. Englefield, A. P. Trishchenko, and Y. Luo, “Genera-tion of long time series of burn area maps of the boreal forest fromNOAA–AVHRR composite data,” Remote Sens. Environ., vol. 112, no. 5,pp. 2381–2396, May 2008.

[24] A. Bastarrika, E. Chuvieco, and M. P. Martín, “Mapping burned areasfrom Landsat TM/ETM+ data with a two-phase algorithm: Balancingomission and commission errors,” Remote Sens. Environ., vol. 115, no. 4,pp. 1003–1012, Apr. 2011.

[25] S. Opazo and E. Chuvieco, “Utilización de productos MODISpara la cartografía de áreas quemadas,” Revista de Teledetección, vol. 27,pp. 27–42, 2007.

[26] JRC-EEA, CORINE Land Cover Updating for the Year 2000: Image 2000and CLC2000, 2005.

[27] G. Chander, B. L. Markham, and D. L. Helder, “Summary of currentradiometric calibration coefficients for Landsat MSS, TM, ETM+, andEO-1 ALI sensors,” Remote Sens. Environ., vol. 113, no. 5, pp. 893–903,May 2009.

[28] A. Fernandez, P. Illera, and J. L. Casanova, “Automatic mapping ofsurfaces affected by forest fires in spain using AVHRR NDVI compos-ite image data,” Remote Sens. Environ., vol. 60, no. 2, pp. 153–162,May 1997.

[29] E. S. Kasischke and N. H. F. French, “Locating and estimating theareal extent of wildfires in Alaskan Boreal forests using multiple-seasonAVHRR NDVI composite data,” Remote Sens. Environ., vol. 51, no. 2,pp. 263–275, Feb. 1995.

[30] P. M. Barbosa, J.-M. Gregoire, and J. M. C. Pereira, “An algorithm forextracting burned areas from time series of AVHRR GAC data applied ata continental scale—An overview,” Remote Sens. Environ., vol. 69, no. 3,pp. 253–263, Sep. 1999.

[31] M. P. Martín, “Cartografía e inventario de incendios forestales en laPenínsula Ibérica a partir de imágenes NOAA-AVHRR,” Ph.D. disserta-tion, Universidad de Alcalá, Alcalá de Henares, Spain, 1998.

[32] E. Chuvieco, M. P. Martín, and A. Palacios, “Assessment of differentspectral indices in the red-near-infrared spectral domain for burned landdiscriminations,” Int. J. Remote Sens., vol. 23, no. 23, pp. 5103–5110,2002.

[33] J. M. C. Pereira, “A comparative evaluation of NOAA/AVHRR vegetationindexes for burned surface detection and mapping,” IEEE Trans. Geosci.Remote Sens., vol. 37, no. 1, pp. 217–226, Jan. 1999.

[34] A. M. S. Smith, N. A. Drake, M. J. Wooster, A. T. Hudak, Z. A. Holden,and C. J. Gibbons, “Production of Landsat ETM+ reference imagery ofburned areas within Southern African savannahs: Comparison of meth-ods and application to MODIS,” Int. J. Remote Sens., vol. 28, no. 12,pp. 2753–2775, Jan. 2007.

[35] J. W. Rouse, R. W. Haas, J. A. Schell, D. H. Deering, and J. C. Harlan,“Monitoring the vernal advancement and retrogradation (Greenwaveeffect) of natural vegetation,” NASA/GSFC, Greenbelt, MD, 1974,Type III Final Report.

[36] B. Pinty and M. M. Verstraete, “GEMI: A non-linear index to monitorglobal vegetation from satellites,” Vegetatio, vol. 101, no. 1, pp. 15–20,Jul. 1992.

[37] C. H. Key and N. C. Benson, “The Normalized Burn Ratio (NBR): ALandsat TM radiometric measure of burn severity,” U.S. Dept. Interior,Northern Rocky Mountain Sci. Center, Bozeman, MT, 1999.

[38] S. Trigg and S. Flasse, “An evaluation of different bi-spectral spacesfor discriminating burned shrub-savannah,” Int. J. Remote Sens., vol. 22,no. 13, pp. 2641–2647, 2001.

[39] D. Stroppiana, J. M. Gregoire, and J. M. C. Pereira, “The use of SPOTVEGETATION data in a classification tree approach for burnt areamapping in Australian savanna,” Int. J. Remote Sens., vol. 24, no. 10,pp. 2131–2151, 2003.

[40] J. M. C. Pereira, M. J. P. Vasconcelos, and A. M. O. Sousa, “A ruled-based system for burned area mapping in temperate and tropical regionsusing NOAA/AVHRR imagery,” in Biomass Burning and its Inter-Reationships With the Climate System, J. L. Innes, M. Beniston, andM. M. Verstraete, Eds. The Hague, The Netherlands: Kluwer, 2000.

[41] J. L. McCarty, C. O. Justice, and S. Korontzi, “Agricultural burning in theSoutheastern United States detected by MODIS,” Remote Sens. Environ.,vol. 108, no. 2, pp. 151–162, May 2007.

[42] M. Hansen, R. Dubayah, and R. Defries, “Classification trees: An alter-native to traditional land cover classifiers,” Int. J. Remote Sens., vol. 17,no. 5, pp. 1075–1081, Mar. 1996.

[43] M. A. Friedl, C. E. Brodley, and H. Strahler, “Maximizing land coverclassification accuracies produced by decision trees at continental toglobal scales,” IEEE Trans. Geosci. Remote Sens., vol. 37, no. 2, pp. 969–977, Mar. 1999.

[44] W.-Y. Loh and Y.-S. Shih, “Split selection methods for classificationtrees,” Statistica Sinica, vol. 7, pp. 815–840, 1997.

[45] T.-S. Lim, W.-Y. Loh, and Y.-S. Shih, “A comparison of predictionaccuracy, complexity, and training time of thirty-three old and newclassification algorithm,” Mach. Learn., vol. 40, no. 3, pp. 203–228,Sep. 2000.

[46] J. M. C. Pereira, A. C. L. Sa, A. M. O. Sousa, M. P. Martín, andE. Chuvieco, “Regional-scale burnt area mapping in Southern Europeusing NOAA-AVHRR 1 km data,” in Remote Sensing of Large Wildfires inthe European Mediterranean Basin, E. Chuvieco, Ed. Berlin, Germany:Springer-Verlag, 1999, pp. 139–155.

[47] A. T. Vafeidis and N. A. Drake, “A two-step method for estimating theextent of burnt areas with the use of coarse-resolution data,” Int. J. RemoteSens., vol. 26, no. 11, pp. 2441–2459, Jun. 2005.

[48] M. Maggi and D. Stroppiana, “Advantages and drawbacks of NOAA-AVHRR and SPOT-VGT for burnt area mapping in a tropical sa-vanna ecosystem,” Can. J. Remote Sens., vol. 28, no. 2, pp. 231–245,2002.


[49] R. H. Fraser, R. Fernandes, and R. Latifovic, “Multi-temporal burned areamapping using logistic regression analysis and change metrics,” in Proc.IEEE IGARSS, 2002, pp. 1486–1488.

[50] N. Koutsias and M. Karteris, “Logistic regression modelling of multi-temporal Thematic Mapper data for burned area mapping,” Int. J. RemoteSens., vol. 19, no. 18, pp. 3499–3514, 1998.

[51] N. Koutsias and M. Karteris, “Burned area mapping using logistic regres-sion modeling of a single post-fire Landsat-5 Thematic Mapper image,”Int. J. Remote Sens., vol. 21, no. 4, pp. 673–687, 2000.

[52] R. L. Pu and P. Gong, “Determination of burnt scars using logisticregression and neural network techniques from a single post-fire Landsat7 ETM+ image,” Photogramm. Eng. Remote Sens., vol. 70, no. 7, pp. 841–850, Jul. 2004.

[53] R. Adams and L. Bischof, “Seeded region growing,” IEEE Trans. PatternAnal. Mach. Intell., vol. 16, no. 6, pp. 641–647, Jun. 1994.

[54] J. Cohen, “A coefficient of agreement for nominal scales,” Educ. Psychol.Meas., vol. 20, no. 1, pp. 37–46, Apr. 1960.

[55] D. P. Roy, P. G. H. Frost, C. O. Justice, T. Landmann, J. L. Le Roux,K. Gumbo, S. Makungwa, K. Dunham, R. Du Toit, K. Mhwandagara,A. Zacarias, B. Tacheba, O. P. Dube, J. M. C. Pereira, P. Mushove,J. T. Morisette, S. K. Santhana Vannan, and D. Davies, “The SouthernAfrica Fire Network (SAFNet) regional burned area product validationprotocol,” Int. J. Remote Sens., vol. 26, no. 19, pp. 4265–4292, Oct. 2005.

Aitor Bastarrika received the B.Sc. degree insurveying engineering from the University of theBasque Country, Vitoria-Gasteiz, Spain, in 1998, ahigher degree in geodesic and cartography engineer-ing in 2000, and a Ph.D. degree in geography fromthe University of Alcalá, Alcalá de Henares, Madrid,Spain, in 2009.

Since 2000, he has collaborated with the Envi-ronmental Remote Sensing Research Group led byProf. Emilio Chuvieco. He is currently an AssistantLecturer with the University of Basque Country, and

his research interests include image processing and automatic cartographyextraction using remote sensing imagery and lidar data.

Emilio Chuvieco is currently a Professor of ge-ography with the University of Alcalá, Alcalá deHenares, Madrid, Spain, where he coordinates theMaster and Ph.D. programs in remote sensing andgeographic information systems, and CorrespondentMember of the Spanish Academy of Sciences. Hehas taught courses in 12 different countries. He wasa Visiting Scholar at the Universities of Berkeley,Nottingham, Clark, Cambridge, Santa Barbara, MD,and the Canada Center for Remote Sensing. He hascoordinated 26 research projects and 20 contracts.

He has supervised 27 Ph.D. dissertations and is the coauthor of 270 scientificpapers and 22 books.

His main research interests include environmental remote sensing (involvingforest fires, global change, deforestation, and natural hazards) and environmen-tal ethics and religious attitudes towards environmental conservation.

M. Pilar Martín received the Ph.D. degree in ge-ography from the University of Alcalá, Alcalá deHenares, Madrid, Spain, in 1998.

She is currently a Scientific Researcher with theCenter for Human and Social Science (SpanishCouncil for Scientific Research). Her research workhas been focused in the field of remote sensing andgeographic information system technologies appliedto environmental monitoring. The main applicationsubject is wildfire management including fire preven-tion, detection, and damage assessment.

Automatic Burned Land Mapping From MODIS Time Series Images: Assessment in Mediterranean Ecosystems

Documents