Contextual multitemporal classification of burned areas in coarse resolution imagery

The objective of this chapter is to present a methodology that allows to exploit spatial and temporal information for extracting burned areas from time series of coarse resolution satellite images. The proposed methodology is based on the hierarchical use of the Multi-Layer Perceptron (MLP) neural network and allows to exploit dependencies of spectral information of the observed targets with spatial and temporal information of the phenomenon under study. These heterogeneous information are fused together and adaptively weighted within the neural classification procedure. The experimental work has been carried out in the framework of the Global Burnt Area-2000 (GBA2000) initiative whose aim is the mapping of burned areas at a global scale from SPOT-Vegetation for the year 2000. The study area corresponds to the Northern part of the African continent, entirely covered by a mosaic of daily SPOT-Vegetation images during the dry season 1999-2000. Results obtained from the application of this methodology confirm the importance of a pattern recognition approach able to exploit spatial and temporal dimensions within a unified scheme.