OFML

Wildfires are an escalating global threat, exacerbated by climate change, necessitating innovative approaches to enhance firefighting strategies. The OFML project aims to address this urgent need by leveraging advanced machine learning (ML) techniques to optimize firefighting efforts post-ignition. While existing ML models have focused on predicting wildfire spread, they often neglect the critical role of human interventions, such as the application of fire retardants and strategic firebreak placements. The motivation behind this research lies in bridging this gap, recognizing that effective firefighting strategies must consider both natural fire dynamics and human actions. To achieve this, the OFML project aims to utilize cutting-edge methodologies, including reinforcement learning, GPU-based fire simulation and deep neural networks, to first build an efficient forward model for predicting wildfire propagation with human interactions and then optimize real-time decisions regarding firebreak and retardant deployment, delivering significantly faster and more effective firefighting strategies compared to traditional models. These learning and simulation algorithms will be trained and evaluated using both simulated and real observational data, with a particular focus on France and the Mediterranean region. In particular, OFML will exploit real-world firebreak data collected from recent wildfire events in Southern Europe, supported by satellite-derived products such as Sentinel-2 and MODIS imagery, Copernicus land cover datasets, and firefighter operational reports. This combination will enable model validation and learning in realistic and diverse scenarios. The OFML project also contributes to frugal AI by combining efficient wildfire models, hybrid learning to reduce data requirements, and a focus on model generalizability across domains to avoid retraining from scratch. This approach not only enhances the operational efficiency of wildfire management but also aims to protect human lives and property more effectively.