Data C
Resume:
Remote Sensing for Science, Education, Rainer Reuter (Editor)
and Natural and Cultural Heritage EARSeL, 2010
MODIS Data Mining to Map Burned Areas
C. QUINTANOa,1, A. STEINb, W. BIJKERb, and A. FERN NDEZ-MANSOc
a
University of Valladolid, Electronic Technology Department, Spain
b
International Institute for Geo-Information Science and Earth Observation (ITC),
Earth Observation Science Department, The Netherlands
c
University of Le n, Agrarian Engineering Department, Spain
Abstract. Estimation of areas burned by forest fires in Mediterranean countries is achieved by ap-
plying an image mining method to MODerate-resolution Imaging Spectroradiometer (MODIS)
satellite data. The proposed image mining based method needs a post-fire image as input and in-
volves three steps: modelling as a sum of Gaussian functions, kernel based smoothing, and thres-
holding. The optimal stopping rule of the modelling iterative algorithm, the optimal smoothing pa-
rameter, and the optimal threshold have been identified by the statistic. This statistic measures
the accuracy of the burned areas estimation by relating the estimation area with the burned area pe-
rimeters measured by Global Positioning System (GPS). Two indexes, specifically designed for
burned area identification and adequate to distinguish burned areas have been used: the Burned
Area Index (BAI) and the Normalized Burn Ratio (NBR). A Z-test allowed us to compare the ac-
curacy of these estimates with the accuracy obtained by classifying directly the input images. Re-
sults show that image mining based methods allowed a higher accuracy of the burned areas estima-
tion. We conclude that MODIS data mining is valid to accurately map burned areas.
Keywords. Burned areas, MODIS, image mining
Introduction
The use of remote sensing data and methods for mapping burnt areas has a long history in the
European Mediterranean region [1]-[4]. It is able to provide temporal and spatial coverage of dam-
age assessment without costly and intense fieldwork. The resulting information is suitable for inte-
gration into a Geographic Information System (GIS) that allows the storage and processing of large
volumes of spatial data.
Data from the National Aeronautics and Space Administratiuon (NASA) MODerate resolution
Imaging Spectrometer (MODIS) has been broadly used to map burnt areas [4]-[9]. MODIS is a sen-
sor onboard of Terra (morning) and Acqua (afternoon) satellites with a daily coverage and more
than 30 narrow bands at wavelengths from the visible to the thermal infrared and at a variable spa-
tial resolution (250 m-1000 m).
Concerning the remote sensing methods dealing with fire scars mapping from reflectance satel-
lite data, three groups can be distinguished: 1) single image analysis (unitemporal perspective), in
which one image acquired after the fire event is available; 2) change detection techniques
(multitemporal perspective), in which pre-fire and post-fire conditions are combined to discriminate
changes; and 3) time series analysis, in which a fairly complete set of pre-fire and post-fire scenes
are used. Single image analysis is commonly used for the processing of medium to high or very
high spatial resolution satellite imagery, and any of the classification methods of remotely sensed
data can be used for mapping burnt areas [10].
1 Corresponding Author: University of Valladolid, Electronic Technology Department, Francisco Mendizabal, 1,
47014-Valladolid, Spain; E-mail: ******@****.***.**
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
Use of spectral indices is widespread among different inputs to the classifier algorithm. Vegeta-
tion indices such as the Normalized Difference Vegetation Index (NDVI) whose estimation typi-
cally involves data form red and near-infrared (NIR) bands have been commonly used to derive
vegetation properties but also to discriminate and map burned areas [2],[11]-[13]. According to [14],
burned vegetation results in a drastic reduction in NIR reflectance and an increase in shortwave in-
frared (SWIR) reflectance in most environment and fire regimes. Spectral indices that integrate data
from NIR and SWIR bands and are specifically designed for burned land discrimination have been
recently developed and tested. Novel spectral indices include the Burned Area Index for MODIS
data (BAIM) [15] and the Normalized Burn Ratio (NBR) [16]-[17]. Other studies have investigated
the utility of principal components analysis [18], texture analysis [19] and spectral mixture analysis
[20].
Image mining, and more specifically spatial image mining, is a relatively new field of study that
uncovers patterns or statistically significant structures in spatial data [21]. It allows for detecting
and mapping features that show a spatial pattern. Although there are almost no references relating
spatial image mining to burned areas by forest fires; in [22] the authors showed that the proposed
image mining algorithm successfully obtained accurate estimations of burned areas from satellite
imagery. Spatial mining algorithm comprises three steps: 1) modelling, expressed by applying
Gaussian distributions on individual grid cells with deviating values; 2) pre-classification consisting
of applying a smoothing kernel; and 3) thresholding with the objective to classify the image into
burned and unburned classes.
In [22], we showed that image mining based on modelling and smoothing, produces more accu-
rate burned area estimations than conventional methods based on thresholding an NDVI image. The
accuracy level was measured by means of statistic. The objective of this research is to validate the
spatial image mining method to map burned areas from the BAIM and NBR indexes derived from
MODIS that are specifically designed to do so.
1. Study area and dataset
The study area is the Barbanza county in Galicia (Northwest Spain) with an area of 29 000 km2).
The Galician landscape mosaic is heterogeneous due to the high degree of fragmentation caused by
human activity (64% of the total area is forestry land). Most fires occur between June and Septem-
ber, being the dry summer period. During the fire season in 2006 in Galicia, 939 km2 (540 km2 for-
ested) was burned. A total of 1651 fires occurred from 4 to 15 August, mainly caused intentionally
and favored by strong winds and a dry, hot summer [23].
Two types of data were used: 1) the Global Positioning System (GPS) based perimeter of
burned areas in the Barbanza county; and 2) MODIS surface reflectance data. Both data suffered
some preprocessing. The contiguous burned areas on the GPS-based burned area perimeter vector
file were grouped and only burned areas larger than 0.5 km2 were considered. The preprocessed
GPS-based vector file consisted of 15 burned areas, and the total considered burned area was 150
km2.
Next, the MYD13Q1 (vegetation indexes, 16-day, 250 m) MODIS product from 21 August to 5
September was windowed to the study area, re-projected to the Universal Transverse Mercator
(UTM) system, WGS84-29n, and co-registered with the preprocessed GPS based burned area pe-
rimeter. This product has a spatial resolution of 250 m and includes four reflectance bands (blue,
red, NIR, SWIR) and two vegetation indexes (NDVI and Enhanced Vegetation Index EVI-).
440
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
2. Method
The proposed image mining algorithm consists of three independent modules: modelling (M),
smoothing (S), and thresholding (T) that are explained in detail in the next subsections. To obtain a
burnt area estimation from the input image, four alternatives have been considered:
1) Thresholding the input image (T). This estimation method does not include an image mining
based method. Therefore, it serves as a reference to be compared to other burnt area estima-
tions identified below.
2) Thresholding the modelled input image (M+T).
3) Smoothing the image, followed by thresholding the smoothed input (S+T).
4) Modelling the input, followed by smoothing and thresholding (M+S+T).
2.1. Calculation of BAIM and NBR
The BAIM index is estimated using the Equation (1):
1
BAIM = (1)
( CNIR NIR ) + ( CSWIR SWIR ) 2
2
where CNIR and CSWIR are the NIR and SWIR reference reflectance values, respectively, and NIR
and SWIR are the pixel reflectance values in the same bands. Following [15], the reference reflec-
tance values were: CNIR = 5 percentile of burned NIR distribution and CSWIR = 95 percentile of
burned SWIR distribution.
The NBR index is obtained by applying the Equation 2 to the preprocessed MODIS image:
SWIR
NBR = NIR (2)
NIR + SWIR
where NIR and SWIR are the NIR and SWIR reflectance values, respectively.
2.2. Modelling module (M)
The modelling module (M) models the input image (I) as a sum of N weighted Gaussian functions
(Equation 3), each of them representing a small part of the burned area. Their characteristics, num-
ber and weight are determined by using a image mining iterative algorithm based on the two coor-
dinates xi1 and xi2 of the ith observation location in two perpendicular directions, following [24].
N
I ( x1, x2 ) wi fi ( xi1, xi 2 ) (3)
i =1
2.3. Smoothing module (S)
A kernel smoother uses weights that smoothly decrease when moving away from the pixel under
consideration. The Gaussian kernel function used has a weight function based on the Gaussian den-
sity function (Equation 4) and assigns weights to points that decrease exponentially with the
squared Euclidian distance from the pixel under consideration, 0:
x x
2
1
K ( x0, x) = exp
0
(4)
2
The parameter corresponds to the variance of the Gaussian density, and controls the width of
the neighbourhood. A large value of implies a low variance as the kernel averages more observa-
tions, but it also implies a larger bias, whereas a small value of implies a relatively large variance
and a small bias. Its optimal value was defined experimentally at the accuracy assessment stage.
441
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
2.4. Thresholding module (T)
The classification of the input image into burnt and unburnt allow obtaining a burnt area esti-
mation. In this study, a pixel of the input image will be assigned to the burnt class if its value is
higher than a pre-defined threshold. Following [2], instead of absolutes values, relative thresholds
based on the mean value of the image to be thresholded and on its standard deviation have been ap-
plied. As happened before, the optimal threshold was defined at the accuracy assessment stage.
2.5. Accuracy assessment
As recommended by [25], confusion matrices and the statistic were used to measure the accuracy
of the achieved burned area estimations. In addition the overall accuracy (OA), producer s accuracy
(PA) (omission error) and user s accuracy (UA) (commission error) were considered as well. As the
ground truth image, a rasterized map resulting from the preprocessed GPS-based vector file was
used. A stratified random disproportional sampling strategy was employed. All pixels correspond-
ing to the burned areas in the ground truth image were considered and the 50% of the pixels corre-
sponding to the unburned areas were randomly selected.
In addition, we used a leave-one-out cross-validation strategy [26]. In our study, this strategy
meant that of the 15 burnt areas, 14 were used as training data to identify the optimal smoothing
parameter and the optimal threshold that maximize the statistic, and one was retained as a valida-
tion area. This process was repeated 15 times, with each of the 15 burned areas used exactly once as
the validation area. Finally, the optimal pair of smoothing parameters and thresholds were defined
as the mode of the optimal values obtained for each of the validated burned areas ( statistic higher
or equal to 0.6); and the definitive estimation of the whole burned area in the Barbanza district was
calculated using this optimal pair of values [22].
A two-sided Z test was applied to the individual statistics associated to these definitive estima-
tions to know if there are significant differences between two estimations. Note that z = 1.96 at thec
95% confidence level, and that the null hypothesis H : = 0 is rejected when Z > z .
0 1 2 c
3. Results and discussion
Table 1 displays the highest statistics obtained by applying each alternative to the original NIR
band and to the spectral indexes. Note that the statistics obtained when (S+T) and (M+S+T) alter-
natives were applied are almost equal ((S+T) lightly superior) and higher than the statistics ob-
tained by using (T) or (M+T) alternatives. Regarding the inputs, NIR band, EVI and BAIM resulted
in more accurate estimations, whereas NDVI and NBR showed the lowest accuracy. The smoothing
parameters and threshold value used to obtain the more accurate burned area estimations for each
input are displayed in Table 2 (OA, PA and UA are also presented).
Table 1. The highest statistics obtained by applying each alternative to each input (bold values show the highest
statistic obtained for each input).
Alternative
Input
(T) (M+T) (S+T) (M+S+T)
NIR 0.66 0.64 0.68
0.69
NDVI 0.57 0.40 0.60
0.61
EVI 0.67 0.64 0.67
0.68
BAIM 0.64 0.66 0.67
0.67
NBR 0.57 0.51 0.58
0.61
The Z statistics obtained when comparing the burned area estimations achieved applying the
different alternatives to each input are shown in table 3. For the alternatives (T) and (S+T) (the ref-
442
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
erence alternative and the alternative with the highest statistics, respectively), significant differ-
ences were observed when NDVI or NBR were used as inputs, with the statistic being equal to
0.61 in both cases. Noticeably no significant differences were observed when BAIM was used as
input. All the inputs except BAIM showed significant differences between (M+T) and (S+T).
Table 2. Information about the highest accurate burned area estimation ( : mean, : standard deviation).
k
Input Optimal Optimal Optimal K statistic OA PA UA
alternative threshold
+ 0.25*
NIR (S+T) 1.50 0.69 0.0105 0.84 0.84 0.84
+ 0.25*
NDVI (S+T) 1.75 0.61 0.0115 0.80 0.80 0.80
+ 0.25*
EVI (S+T) 0.75 0.68 0.0107 0.84 0.84 0.84
+ 0.25*
BAIM (S+T) 1.25 0.67 0.0109 0.84 0.84 0.84
+ 0.25*
NBR (S+T) 2 0.61 0.0117 0.81 0.80 0.81
Table 3. Z statistics (bold values show Z > 1.96) comparing the four alternatives: (T): thresholding, (M+T): modeling
and thresholding, (S+T): smoothing and thresholding, and (M+S+T): modeling, smoothing and thresholding.
Input
Alternative 1 Alternative 2
NIR NDVI EVI BAIM NBR
(T) (M+T) 1.02 1.60 1.21
9.51 3.43
(T) (S+T) 1.91 1.19 1.70
2.00 2.27
(T) (M+S+T) 1.56 1.65 0.12 1.12 0.48
(M+T) (S+T) 0.48
2.93 11.51 2.80 5.75
(M+T) (M+S+T) 1.73 0.08
2.58 11.14 3.92
(S+T) (M+S+T) 0.36 0.34 1.07 0.57 1.80
The simultaneous analysis of Tables 1 and 3 allows identifying the combination of input and al-
ternative with the highest statistic that showed significant differences compared to the reference
alternative (T). Such combination is obtained when using NDVI or NBR as inputs of the (S+T) al-
ternative (k = 0.61). Visualizing both burnt area estimations (see Figure 2) the estimation based on
NDVI is preferred because the NBR-based estimation presents an excessive smoothing.
c) Input = NDVI
a) Ground truth b) Input = NIR
d) Input = BAIM f) Input = NBR
e) Input = EVI
Figure 2. Burnt area estimations based on thresholding the smoothed input
(S+T alternative with the parameters showed in Table 2).
Regarding the alternative that allowed more accurate estimations our result is in accordance
with other authors (see [22] and [27]) that stated that smoothing used as a pre-classifier leads for
higher accurate burned area estimations. Concerning the inputs used NBR stands as the more ade-
quate index to estimate burnt areas, which is in accordance with several studies (e.g. [28]-[29]) that
show how NBR highlights the severity of burned areas. NDVI performance is, however, unusual.
443
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
.
4. Conclusions
This study assesses the utility of image mining algorithms to map burned areas in Mediterranean
countries from MODIS data. Specifically the smoothing algorithm has been used as a preclassifier.
The results show that image mining algorithms allow burnt areas estimation with higher accuracy
than conventional methods such as thresholding.
The burnt area estimation obtained by using the NBR index as input to the (S+T) alternative was
more accurate than the burnt areas estimation obtained by thresholding the input image (T). In addi-
tion, the Z test of the statistics showed significant differences between this image mining algo-
rithm and the reference procedure (just thresholding).
We conclude that MODIS data mining is a valid methodology to accurately map burned areas.
Acknowledgments
The first and last authors acknowledge the financial support of the Ministry of Education of the
Castilla y Le n regional government (Research project: VA098A08).
References
E. Chuvieco, R.G. Congalton, Application of remote sensing and geographic information systems to forest fire haz-
[1]
ard mapping, Remote Sensing of Environment 29 (1989) 147-159.
A. Fern ndez, P. Illera, J.L. Casanova, Automatic mapping of surfaces affected by forest fires in Spain using
[2]
AVHRR NDVI composite image data, Remote Sensing of Environment 60 (1997) 153-162
A. DeSantis, E. Chuvieco, P.J. Vaughan, Short-term assessment of burn severity using the inversion of PROSPECT
[3]
and GeoSail models, Remote Sensing of Environment 113 (2009) 126 136.
S. Merino-de-Miguel, M. Huesca, F. Gonz lez-Alonso, MODIS reflectance and active fire data for burn mapping
[4]
and assessment at regional level, Ecological Modelling 221 (2010) 67 74
E. A.C.L. S, J.M.C. Pereira, M.J.P. Vasconcelos, J.M.N. Silva, N. Ribeiro; A. Awasse, Assessing the feasibility of
[5]
sub-pixel burned area mapping in miombo woodlands of northern Mozambique using MODIS imagery, Interna-
tional Journal of Remote Sensing 24-200*-****-****.
Chuvieco, S. Opazo, W. Sione, H. DelValle, J. Anaya, C. DiBella, I. Cruz, L. Manzo, G. L pez, N. Mari, F. Gon-
[6]
z lez-Alonso, F. Morelli, A. Setzer, I. Csiszar, J. Ander-Kanpandegi, A. Bastarrika, R. Libonati, Global burned-
land estimation in Latin America using MODIS composite data, Ecological Applications, 18 (2008) 64 79.
X. Cao, J. Chen, B. Matsushita, H. Imura; L. Wang, An automatic method for burn scar mapping using support
[7]
vector machines, International Journal of Remote Sensing 30 (2009) 577 594.
L. Boschetti, D. Roy, P. Barbosa, R. Boca; C. Justice, A MODIS assessment of the summer 2007 extent burned in
[8]
Greece, International Journal of Remote Sensing 29-200*-****-****.
C.O. Justice, L. Giglio, S. Korontzi, J. Owens, J.T. Morisette, D. Roy, J. Descloitres, S. Alleaume, F. Petitcolin,
[9]
and Y. Kaufman, The MODIS fire products, Remote Sensing of Environment 83 (2002) 244 262.
J. San-Miguel- Ayanz, J.M.C. Pereira, R. Boca, P. Strobl, J. Kucera and A. Pekkarinen, Forest fires in the European
[10]
Mediterranean region: mapping and analysis of burned areas, in E. Chuvieco (ed.) Earth Observation of wildland
fires in Mediterranean Ecosystems (189-204), Springer-Verlang, Berlin Heidelberg, 2009.
J.M.C, Pereira, A comparative evaluation of NOAA/AVHRR vegetation indexes for burned surface detection and
[11]
mapping, IEEE Transactions on Geoscience and Remote Sensing 37 (1999) 217-226.
E. Chuvieco, M.P. Mart n, A. Palacios, Assessment of different spectral indices in the red near-infrared spectral
[12]
domain for burned land discrimination, International Journal of Remote Sensing 23-200*-****-****.
C. Domenikiotis, N.R. Dalezios, A. Loukas, M. Karteris, Agreement assessment of NOAA/AVHRR NDVI with
[13]
Landsat TM NDVI for mapping burned forested areas, International Journal of Remote Sensing 23 (2002) 4235-
4246.
444
C. QUINTANO, et al.: MODIS Data Mining to Map Burned Areas
L.B. Lentile, Z.A. Holden, A.M.S. Smith, M.J. Falkowski, A.T. Hudak, P. Morgan, S.A. Lewis, P.E. Gessler, N.C.
[14]
Benson, Remote sensing techniques to assess active fire characteristics and post-fire effects. International Journal
of Wildland Fire 15 (2006) 319 345.
P. Mart n, I. G mez, E. Chuvieco, Performance of a burned area index (BAIM) for mapping Mediterranean
[15]
burned scars from MODIS data, Proceedings of the 5th International Workshop on Remote Sensing and GIS Appli-
cations to forest fire management: fire effects assessment, Zaragoza Spain- (2005) 193-197.
C.H. Key, N.C. Benson, The Normalized Burn Ratio (NBR): A Landsat TM Radiometric Measure of Burn severity.
[16]
USDA (Bozeman, Mont), (1999) (available at http://nrmsc.usgs.gov/research/ndbr.htm).
D.P. Roy, L. Boschetti, S.N. Trigg, Remote sensing of fire severity: assessing the performance of the Normalized
[17]
Burn Ratio, IEEE Transactions on Geoscience and Remote Sensing Letters 3 (2006) 112-116.
F.J. Garc a-Haro, M.A Gilabert, J. Meli, Monitoring fire-affected areas using Thematic Mapper data. International
[18]
Journal of Remote Sensing 22 (2001) 533 549.
A.M.S. Smith, M.J. Wooster, A.K. Powell, D. Usher, Texture-based feature extraction: application to burn scar
[19]
detection in Earth Observation satellite imagery. International Journal of Remote 23-200*-****-****.
C. Quintano, A. Fern ndez-Manso, O. Fern ndez-Manso, Y.E Shimabukuro, Mapping burned areas in Mediterra-
[20]
nean countries using Spectral Mixture Analysis from a unitemporal perspective. International Journal of Remote
Sensing 27 (2006) 645 662.
J. Shawe.Taylor, N. Cristianini, Kernel Methods for Pattern Analysis, Cambridge University Press, Cambridge,
[21]
2004.
C. Quintano, A. Stein, W. Bijker, A. Fern ndez-Manso, Pattern validation for data mining of burned area objects
[22]
from MODIS satellite images. International Journal of Remote Sensing (in press)
Xunta de Galicia, Informe sobre a vaga de incendios forestais do mes de Agosto de 2006, Conseller a do medio
[23]
rural -Direcci n Xeral do Montes e Industrias Forestais-, Galicia Spain-, 2006.
U. Rajasekar, W. Bijker, A. Stein, A., Image mining for modelling of forest fires from Meteosat images. IEEE
[24]
Transactions on Geoscience and Remote Sensing 45 (2007) 246-253.
R.G. Congalton, K. Green, Assessing the Accuracy of Remotely Sensed Data. Principles and Practices, CRC Press,
[25]
Taylor & Francis Group, Boca Raton, 2009
C. Goute, Note on free lunches and cross-validation, Neural Computation 9-199*-****-****.
[26]
C. Quintano, E. Cuesta, Improving satellite image classification by using fractional type convolution filtering, In-
[27]
ternational Journal of Applied Earth Observation and Geoinformation, (in press).
A.E. Cocke, P.Z. Fule, J.E. Crouse, Comparison of burn severity assessments using Differenced Normalized Burn
[28]
Ratio and ground data, International Journal of Wildland Fire 14 (2005) 189 198.
D.R. Roy, L. Boscheti, S.N. Trigg, Remote sensing of fire severity: assessing the performance of the normalized
[29]
burn ratio, IEEE Geoscience and Remote Sensing Letters 3 (2006) 112 116.
445
446
Contact this candidate