U-Net framework for short-term wildfire spread forecasting using multi-source earth observation data in Thailand

Wildfire spread in Thailand is governed by fine-scale interactions among fuel conditions, near-surface atmospheric forcing, and monsoon-driven variability, yet operational tools for short-term forecasting remain limited. This study presents a U-Net–based framework for predicting 1–3-day wildfire spread using multi-source Earth observation (EO) data standardized to a 375 m spatial resolution. Model inputs include static and dynamic variables representing fuel conditions, fire status, and next-day atmospheric drivers. VIIRS active-fire detections are clustered using a DBSCAN-based approach to identify ignition points and construct event-based input patches, while a rolling prediction strategy enables multi-day fire propagation. Across dry-season wildfire events in 2024, the proposed U-Net framework achieved strong predictive performance (average F1-score = 0.66; IoU = 0.60) and consistently outperformed a rule-based Cellular Automata (CA) baseline, which serves as an interpretable reference for neighborhood-driven fire spread. Additional model comparisons suggest that increased architectural complexity does not necessarily improve predictive skill under short-lived and spatially fragmented fire conditions. Feature-ablation analysis shows that vegetation condition, near-surface temperature, wind components, and evaporative-demand–related moisture stress dominate short-term fire spread, while static landscape factors play a secondary role. These findings highlight that wildfire dynamics in Thailand are primarily controlled by local interactions and rapid temporal variability rather than sustained fire-front propagation.