Multitask segmentation from satellite imagery for burned area delineation and severity estimation

In recent years, the frequency and intensity of natural disasters has significantly and dangerously increased in Europe and in the world due to factors such as climate change, population growth and aggressive urbanization of rural areas. Every year, hundreds of wildfires destroy millions of hectares of forest. Rapidly delineating burned areas from satellite has become a crucial task for first responders and decision makers, to enhance the preparedness, response and recover phases during such crises. The European Union and the European Space Agency are intensifying their efforts to accumulate information on natural disasters. Data about past catastrophic events are collected by Copernicus Emergency Management System (CEMS) and categorized according to the type of event. Exploiting wildfire EMS activations, the first objective of this thesis was the generation of a large dataset focused mainly on the European soil, collecting satellite imagery from Sentinel-2. The dataset includes different maps, including delineation and grading masks provided by EMS, as well as a cloud cover label to mask clouds in the images, thus reducing possible errors during inference. Starting from this dataset, this thesis also proposes a multitask learning semantic segmentation approach for wildfire delineation and burn severity estimation. A multitasking scenario allows a model to jointly learn from both delineation and severity estimation of wildfires. Several state-of-the-art semantic segmentation models are tested to assess their performance in both burned area delineation and severity estimation, using post-wildfire images only. Experiments show that the combination of a large dataset and the multitask approach allows to reach robust results, achieving F1 scores over 0.9 considering the delineation, and RMSE scores lower than 0.5 for severity estimates.