Predicting Wildfires Before They Spread Using Multimodal Foundation Models

Accurate prediction of wildfire onset and early propagation behavior remains one of the most urgent challenges in environmental hazard science. This study proposes a multimodal foundation model (MFM) framework that integrates satellite remote sensing imagery, numerical weather prediction outputs, digital elevation models, and normalized difference vegetation index time series to forecast wildfire spread prior to its occurrence. The architecture incorporates a residual feature extraction backbone combined with a cascade classification mechanism and cross-modal transformer encoder, jointly trained on large-scale Earth observation corpora. Experiments conducted on California and Mediterranean fire records demonstrate that the proposed framework achieves an area under the receiver operating characteristic curve of 0.924 and an F1 score of 0.847, outperforming convolutional neural network baselines by 7.3%. Ablation studies confirm that the residual encoding module and the hierarchical patch-level classifier contribute most significantly to predictive performance. Temporal validation against multi-hour fire progression maps further demonstrates the spatial alignment between model predictions and observed hotspot distributions. These results establish MFMs as a transformative paradigm for proactive, multimodal wildfire risk management at regional scale.