Anais do XIX Simpósio Brasileiro de Sensoriamento Remoto

The use of fire for land management is one of the main anthropic activities that have led to the impoverishment of tropical forests. Therefore, mapping these areas is paramount for public policies implementation. Currently, machine learning techniques have shown very effective results in the classification of land cover on extensive areas. This paper aims to compare the Random Forest (RF) and Support Vector Machine (SVM) algorithms performance on burned areas mapping in Amazon. Using a multiresolution segmentation algorithm applied to a Landsat image, the training dataset included 300 objects of burned and non-burned areas. Additionally, 24 attributes were tested in both RF and SVM approaches. An overall classification accuracy of 91% was achieved by RF and SVM models using spectral and geometric attributes. Nonetheless, regarding the omissions and inclusion errors, SVM models had the best performance on burned areas mapping.