Machine Learning Approaches for Estimating Wildfire Propagation

Wildfires are a growing environmental and socio-economic concern, with climate change driving more frequent and severe events. Remote sensing has enabled accurate post-event burned area mapping, yet predictive approaches that anticipate fire impact before ignition remain limited and highly challenging. The complex interactions between environmental, climatic, and topographic factors hinder the development of robust pre-event models. This thesis investigates the feasibility of proactive wildfire prediction through a custom multimodal deep learning architecture designed to estimate burned area extent from pre-fire observations. The framework integrates diverse data sources, including Sentinel and Landsat optical imagery, Copernicus Digital Elevation Model, landcover information, and meteorological variables from ERA5, each processed through dedicated encoders tailored to the nature of the input. The feature representations are subsequently fused through specialized fusion blocks, enabling the model to capture complementary aspects of the wildfire dynamics within a unified predictive structure. Conducted as an experimental study, this work does not aim to deliver an operational system but to explore the opportunities and challenges of shifting from post-event segmentation to pre-event prediction. The best configuration achieved a maximum Intersection over Union (IoU) of 38.59%, demonstrating both the potential and the limitations of the approach. The results enhance the study of employing multimodal deep learning for forecasting wildfire risk.