Towards a pan-European burnt scar mapping methodology based on single date medium resolution optical remote sensing data

A two stage approach for burnt scar detection from single date multispectral medium spatial resolution optical remote sensing data (AWIFS) has been developed. The approach includes first an identification of burnt scar seeds based on a learning algorithm followed by a region growing process. An Artificial Neural Network (ANN) and a Classification tree (CT) were tested as learning algorithms. Both learning algorithms were coupled with a bootstrap aggregation. Training data for the classifiers were obtained from MODIS-based polygons generated by the Rapid Damage Assessment (RDA) module of the European Forest Fire Information System (EFFIS), to which different levels of filtering were applied. The outcomes were validated against datasets generated from visual interpretation of ETM+ scenes. The method was tested in two locations of Portugal and Greece. Both ANN and CT alternatives produced similar results, with kappa coefficients close to 0.80 in the Greek location and higher than 0.70 in the Portuguese location. In test sites, more than 80% and 90% of burnt areas larger than 10 ha and 50 ha respectively were detected. The results show that filtering the training dataset reduces the overestimation of burnt areas and produce higher accuracies.