Deep spatiotemporal signal learning with transformers for multi-day wildfire forecasting

Abstract

Wildfire forecasting is a critical challenge in environmental signal processing and disaster response planning. The ability to interpret multimodal spatiotemporal signals is essential for early warning systems and resource deployment. This study addresses these limitations by proposing a unified prediction-to-action framework. We utilized four open-access datasets—wildland fire emissions database (WFED), fire information for resource management system (FIRMS), Sentinel Hub, and a custom moderate resolution imaging spectroradiometer+shuttle radar topography mission (ERA5+MODIS+SRTM) fusion—covering fire occurrences, vegetation indices, meteorological parameters, and topographic features. These heterogeneous signals were preprocessed, aligned, and transformed into structured tensors for model training and evaluation. We use a transformer-based system to understand long-term patterns in space and time, enhanced by a belief–desire–intention (BDI) reasoning module that connects our predictions to flexible wildfire response plans. The novelty lies in the integration of signal-aware attention mechanisms with symbolic decision modeling. Model performance was evaluated using F1-score, intersection over union (IoU), mean absolute error (MAE), and directional accuracy. The suggested framework did better than the basic convolutional neural network (CNN) models, reaching an F1-score of 0.74, a directional accuracy of 84.3%, and lowering the MAE to 7.6 km², while also providing clear and relevant action suggestions.