FIRE REFERENCE PERIMETERS EXTRACTED FROM SENTINEL-2 DATA FOR VALIDATION OF BURNED AREA PRODUCTS IN AFRICA BIOMES

In this work we present a procedure for building a dataset of fire reference perimeters over the African continent from Sentinel-2 (S2) time series. The strategy relies on a sampling scheme designed on the characteristics of the S2 tiling system to provide units suitable for statistical sampling of validation units. The S2 archive is searched to extract cloud free images for building time series that are classified with a Random Forest algorithm to provide fire reference perimeters. A test S2 tile over Tropical savanna (35 LMD) is used to assess the accuracy of S2 reference perimeters by comparison with polygons of burned areas extracted from high resolution Planetscope data.