Deep-learning deforestation detection in the Legal Amazon area based on Sentinel-1 data

Abstract

The Amazon rainforest, the largest in the world, has been in the global spotlight for decades, given its size (more than half of the Brazilian territory), biodiversity and impact on global weather, economy, politics and on other ecosystems. Deforestation monitoring of the Amazon area can be a herculean task for governmental agencies, non-governmental organizations and other interested parties, especially during the region’s rainy season, roughly from October to May. During the drier season, it is possible to monitor large areas using optical satellite data to calculate temporal changes in the Normalized Difference Vegetation Index (NDVI), but during the rainy season the extremely clouded images render this method impractical. Synthetic Aperture Radar (SAR) imagery such as those from the Sentinel-1 mission, on the other hand, is insensitive to weather conditions, becoming a great candidate for deforestation monitoring, even though there is no NDVI equivalent for radar data. In this work, we propose a deep-learning method to detect new deforestation events in the Legal Amazon area using bi-temporal Sentinel-1 data, by segmenting former images for forest detection and latter images for deforestation. Results show that our solution is effective at drawing attention to areas that have undergone some sort of change consistent with deforestation patterns.