Mapping burned surfaces in sub-Saharan Africa based on multi-temporal neural classification

Biomass burning from vegetation fire is one of the major disturbances of natural terrestrial ecosystems and an increased understanding of its impact requires extensive documentation of the phenomenon at global scale. Satellite remote sensing has been demonstrated as the only tool able to cope with this global issue. Global daily coverage using the SPOT-VEGETATION instrument has been made for the year 2000 in the frame of the Global Burnt Area-2000 (GBA2000) initiative. This paper presents a new methodology for detecting burned areas from time series of coarse resolution satellite images. The proposed methodology is a supervised classification strategy based on hierarchical use of the Multi-Layer Perceptron neural network. The use of a neural network-based approach allows the exploitation of spatial and temporal dimensions within a unified scheme. The methodology has been applied to a series of daily SPOT-VEGETATION images covering the northern part of the African continent and daily burned area maps were produced for the dry season of the year 2000. Evaluation of the results has been done both in terms of the algorithm performance and of the final burned map accuracy using Landsat-7 Enhanced Thematic Mapper Plus (ETM+) data.