Vision Transformers for burned area detection

The automatic identification of burned areas is an important task that was mainly managed manually or semi-automatically in the past. In the last years, thanks to the availability of novel deep neural network architectures, automatic semantic segmentation solutions have been proposed also in the emergency management domain. The most recent works in burned area delineation make use of Convolutional Neural Networks (CNNs) to automatically identify regions that were previously affected by forest wildfires. A largely adopted segmentation model, U-Net, demonstrated good performances for the task under analysis, but in some cases a high overestimation of burned areas is given, leading to low precision scores. Given the recent advances in the field of NLP and the first successes also in the vision domain, in this thesis, we investigate the adoption of vision transformers for semantic segmentation to address the burned area identification task. In particular, we explore the SegFormer architecture with two of its variants: the smallest SegFormer-B0 and the intermediate one, SegFormer-B3. We exploited different loss functions to deal with the complex structures that can be found in satellite imagery. The experimental results show that SegFormer provides better predictions, with higher precision and F1 score, but also better performance in terms of the number of parameters with respect to CNNs.