Fire Scars Mapping Over Brazilian Amazon Forest by Exploiting Sentinel-2 Data and Deep Learning

Abstract

Wildfires in the Brazilian Amazon have raised significant concerns owing to the environmental, social, and global impacts associated with these events. They have led to habitat loss for various species and release of substantial amounts of carbon dioxide into the atmosphere. Thereby contributing to climate change and deterioration of air quality due to pollutants emission. The integration of advanced technologies, including high-spatial resolution satellite data and image processing algorithms, enables a more precise and comprehensive understanding of the wildfire scenario. This research introduces a model based on deep learning that can be applied over Sentinel-2 images to reliably detect fire scars with an accuracy above 90% (92% on training data and 82% on validation data). A SpectrumNet convolutional neural network was employed, incorporating features extracted from spectral bands at 10m and 20m.