Ensemble fire probability prediction by integrating meteorological AI models with lightweight prediction networks

Abstract

Wildfires pose persistent threats to ecosystems and human safety, making timely fire-risk assessment essential for disaster prevention and resource management. However, conventional wildfire early-warning approaches are often constrained by static fire-weather indices or machine-learning models driven by reanalysis data, which cannot exploit future atmospheric fields. We therefore propose and evaluate a decoupled cascade framework that decomposes the problem into (i) an upstream AI weather model (Pangu-Weather) to generate future atmospheric states and (ii) downstream probabilistic fire classifiers. To the best of our knowledge, this is the first study to systematically compare the performance of three operational forecasting strategies (reanalysis, one-step forecast, and rolling forecast) for fire probability estimation. Using Africa as the study area, with multi-source data from 2014 to 2025, we find that one-step AI forecasts preserve the spatial fire-risk structure nearly as well as reanalysis data. Based on this dataset, we evaluated 11 downstream classifiers within a two-stage fire-occurrence and fire-intensity prediction design, while rolling forecasts suffer from degraded detection due to accumulated meteorological errors, with persistent reliability failure occurring after approximately six weeks. Cross-model consensus based on SHapley Additive exPlanations (SHAP) identifies vegetation, land-surface conditions, and meteorological variables as robust contributors. Our results demonstrate that meteorological foundation models can serve as modular and interchangeable upstream modules for operational wildfire hazard estimation, offering a scalable alternative to end-to-end black-box fire foundation models.