From Occurrence to Impact: A Multi-Scale Machine Learning Pipeline for High-Resolution Wildfire Modelling

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 primary candidate for advanced machine learning (ML) approaches that integrate Earth Observation (EO) and climate data. In contrast to traditional operational risk systems commonly based only on weather conditions, these ML-EO systems can be trained on a much richer context, allowing the generation of not only more accurate wildfire occurrence risk maps but also the potential corresponding impacts at a much higher spatial resolution.

To demonstrate this paradigm shift, we present a multi-scale set of modeling developments that address the wildfire lifecycle from onset to impact. We begin at the landscape scale with a 100m calibrated daily probability of occurrence model that combines the Fire Weather Index (FWI), high-resolution land cover data, and historical fire event records. For all subsequent models, we deliberately neglect the highly unpredictable ignition component by focusing exclusively on burned areas to predict potential behavior in the event of a fire. Within this framework, we first explore the estimation of active fire intensity, evaluating how environmental drivers can enable the prediction of potential Fire Radiative Power (FRP). Moving further to evaluate the final physical consequences on the landscape, we introduce a 30m potential wildfire severity model for predicting vegetation damage in ecosystems; just from initial conditions, the model successfully identifies critical thresholds and skillfully predicts which specific areas within a fire's perimeter would be most severely burned. Finally, to capture the key role that spatial context plays in fire behavior, we explore Convolutional Neural Networks (CNNs) aiming to learn fire connectivity patterns directly from historical events and enabling the modelling of valuable variables for fire managers, such as the size of a potential wildfire event. Together, these developments mark a significant step toward an operational, high-resolution and comprehensive wildfire risk pipeline strengthening both early-warning capabilities and long-term resilience planning.