Learning Fire Connectivity: A Convolutional Neural Network for assessing wildfire risk

Accurate wildfire prediction is becoming increasingly critical as climate change drives warmer and drier conditions worldwide. The complex, non-linear interactions among meteorological factors, fuel characteristics, and landscape structure make wildfire risk a strong candidate for advanced machine learning (ML) approaches that integrate Earth Observation (EO) and climate data. Recent progress on this front has already led to significant improvement on operational systems, such as the ECMWF wildfire forecast, demonstrating clear advantages over traditional, meteorology-only indicators. However, most current ML models are based on single pixel predictions that lack essential spatial context. This limits their ability to capture how static forest connectivity interacts with dynamic fire processes, including spread, intensity, and likelihood of occurrence. To overcome these constraints, we propose a Convolutional Neural Network (CNN) architecture designed to explicitly learn and exploit the additional predictability from these complex spatial relationships. The model fuses multiscale inputs by processing high-resolution landscape variables (e.g., above-ground biomass, land cover, soil moisture, topography) alongside coarse-resolution meteorological fields. To represent the full spectrum of wildfire risk, we experiment with multiple target variables including probability of burn, fire severity, and fire extent. Through these experiments, the CNN is forced to learn connectivity patterns directly from historical wildfire events. The successful implementation of this approach would constitute a major step toward operational, high-resolution, context-aware wildfire risk mapping, strengthening both early-warning capabilities and long-term resilience planning.