Pattern validation for MODIS image mining of burned area objects

An image mining method was applied to Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data to estimate the area burned by forest fires occurring in Galicia (Spain) between 4 August and 15 August 2006. Five different inputs were considered: post-fire near-infrared reflectance (NIR) band, post-fire Normalized Difference Vegetation Index (NDVI) image, pre-fire and post-fire NDVI difference image and 4-m and 11-m thermal bands. The proposed image mining method consists of three steps: a pre-classification step, applying kernel smoothing, based on the fast Fourier transform (FFT), a modelling step applying Gaussian distributions on individual grid cells with deviating values, and a thresholding step classifying the model into burned and unburned classes. Polygons collected in the field with Global Positioning System (GPS) measurements from a helicopter permitted validation of the burned area estimation. A Z-test based on the  statistic compared the accuracy of this estimation with the accuracies achieved by traditional methods based both on spectral changes in reflectance after the fire and active fire detection. The results showed a significant improvement (7.5%) in the accuracy of the burned area estimation after kernel smoothing. Burned area estimation based on the smoothed difference image between pre-fire and post-fire NDVI image had the highest accuracy ( = 0.72). We conclude that the image mining algorithm successfully extracted burned area objects and that extraction was best when smoothing was applied prior to classification. Image mining methods based on using the  statistic thus provide a valuable validation procedure when selecting the optimal MODIS input image for estimating burned area objects.