Semantic Segmentation of Burned Areas in Sentinel-2 Satellite Imagery Using Deep Learning Transformer and Convolutional Attention Networks

Wildfires continue to occur every year, resulting in millions of acres of burned areas. This makes the task of mapping the burned areas arduous, and it is exacerbated by the fact that many of these burned areas are in remote and hard-to-reach regions. Satellite imagery with the highest freely available resolution combined with the latest deep learning architectures offer an opportunity to map these burned areas using models trained for semantic segmentation. In this work, we evaluate the state-of-the-art deep learning transformer architecture SegFormer and convolutional attention architecture SegNext specifically trained for the task of burned area semantic segmentation. The trained models are compared against U-Net and DeepLab neural network models trained on the same dataset. SegNext achieved a precision of 88.79%, a 4.44% increase compared to the best precision achieved by DeepLabV3+, and an intersection-over-union score of 76.71%, a 4.71% increase compared to the best score achieved by DeepLabV3.