BURNT AREAS SEMANTIC SEGMENTATION FROM SENTINEL DATA USING THE U-NET NETWORK TRAINED WITH SEMI-AUTOMATED ANNOTATIONS

Abstract

The Pantanal biome is one of the most important wetlands on the planet, harboring a rich biodiversity whilst being critical in maintaining hydrological cycles and climate regulation. However, the occurrence of fires in the biome has represented a significant threat to this unique ecosystem and its multiple functions. Understanding the extent, intensity and environmental impacts caused by fires in the Pantanal, is of unique importance for the preservation of the biome's biodiversity. Remote sensing techniques have played an important role in detecting and mapping burnt areas, especially SAR (Synthetic Aperture Radar) orbital systems, that are able to collect data in regions with frequent cloud cover or during extreme fire events. In this context, the objective of this study was to evaluate the potential of the U-Net semantic segmentation network applied to SAR data in the detection of burnt areas in the Brazilian Pantanal. For this, a semi-automatic annotated dataset was generated and considered as ground truth to evaluate the result obtained by the network. Two input datasets were evaluated in the detection of burnt areas, one containing optical and SAR data whereas the other containing only SAR data. The predictions of the two datasets were consistent with the semi-automatically generated annotation, showing similar spatial distribution but presenting a greater number of burnt areas. The model using both optical and SAR data achieved IoU (Intersection of Union) of 0.69 whereas the SAR only model had 0.60. Considering the amount of available data and the complexity of burnt area detection, the predictions achieved were adequate.