Advancing Global Short-Term Wildfire Forecasting with Wildfire-Specific Spatiotemporal Deep Learning

Abstract Reliable daily wildfire forecasting is essential for early warning, air-quality management, and reducing smoke-related health exposure. However, global approaches generally emphasize either fire-weather danger or pixel-wise occurrence prediction and do not explicitly represent temporal persistence, neighboring-grid continuity, and new ignition. Here we developed the STA-Net, a wildfire-specific spatiotemporal deep learning framework using multisource global data from 2013 to 2024. The STA-Net predicts daily fire probability at 0.5° resolution and achieves an area under the ROC curve (AUC) of 0.97 and produces more localized fire risk estimates than the operational fire weather index (FWI). Baseline comparisons showed that spatiotemporal learning improves global fire prediction, while the wildfire-specific design of STA-Net provides further gains beyond a generic spatiotemporal model. The largest improvements occurred for neighboring-grid and new ignition fires, where prediction performance (F1 score) roughly doubles. Attribution analysis showed that recent fire history was most important for persistent fires, while vegetation, topography, and meteorological variables contributed more to new ignition fires. These findings demonstrate that wildfire-specific spatiotemporal modeling improves global short-term wildfire forecasting and can provide an interpretable basis for early warning, fire management, air-quality forecasting, and emission-related applications.