Development of innovative Earth Observation data applications to support forest fire management

This doctoral research focuses on the development of machine learning methodologies applied to satellite Earth Observation data for the integrated management of wildfires. The study addresses the three main phases of a wildfire event: pre-fire, active fire, and post-fire by designing and evaluating tailored data-driven approaches for each stage. In the pre-fire phase, an automatic mapping system for fuel types is developed, following the Scott/Burgan classification model. The method leverages Sentinel-2 multispectral imagery and convolutional neural network to classify vegetation and surface types relevant to fire risk assessment. During the active fire phase, the study investigates the feasibility of on-board wildfire monitoring by embedding U-Net deep learning models directly on satellite edge-computing hardware. The approach jointly performs cloud filtering and fire detection in near-real time, prioritizing high-quality acquisitions and reducing data latency. The models, trained on Landsat-8 imagery and tested on PRISMA hyperspectral data, are deployed on hardware accelerators, demonstrating the practical applicability of real-time on-orbit processing for emergency response. In the post-fire phase, GEDI LiDAR observations are used as reference data to train convolutional neural network with multi-source predictors, including Sentinel-1, Sentinel-2, and topographic variables. The most relevant GEDI metrics are selected to capture key aspects of forest vertical structure, such as canopy height, vertical heterogeneity, vegetation cover, and plant surface area. This approach enabled the generation of spatially continuous pre- and post-fire maps of structural attributes, overcoming the sparse sampling of GEDI data. The resulting maps are used to quantify fire-induced structural changes and support the assessment of ecosystem recovery trajectories. The results demonstrate the potential of combining satellite remote sensing with advanced machine learning to support comprehensive, timely, and scalable wildfire management strategies.