Estimation of area burned by forest fires in Mediterranean countries: A remote sensing data mining perspective

Forest fires throughout the world result in tree mortality that can cause substantial timber and carbon losses. There is a critical need to map the areas burned by such fires to guide forest management decisions. Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery provides inexpensive and frequent coverage over large areas, facilitating forest health monitoring. In this study a MODIS post-fire image at a spatial resolution of 250m serves as the starting point of an image mining based method. It involves three algorithms: modeling as a sum of Gaussian functions, kernel based smoothing, and adaptive thresholding. Adaptive thresholding serves as the reference to be compared to the image mining based method. Three spectral indices specifically designed for burned area identification have been used: the Burned Area Index (BAI), the Burned Area Index adapted to MODIS bands (BAIM), and the Normalized Burn Ratio (NBR). The κ statistic is applied to quantify the accuracy of the burned areas estimations by relating the estimated area with burned area perimeters measured on the ground by Global Positioning System (GPS). In addition, the κ statistic allows us to identify both the optimal spectral index and the optimal algorithms’ parameters. In this work, an accurate estimation (κ>0.8) of areas burned by forest fires in Mediterranean countries is achieved, in particular if the BAIM index is used. The accuracy of these estimates is compared with the accuracy obtained by using the reference method by a McNemar’s test. Results show that our image mining based method allows a higher accuracy (the average increase of κ equals to 16%) than the reference method. We conclude that this method adequately maps burned areas, and that it may help management agencies to better understand of landscape-scale burn patterns.