Physics-Informed Deep Learning for Wildfire Spread Simulation Using Geospatial Foundation Models

Expansion of the wildland–urban interface (WUI), fire suppression, and climate change have substantially increased human exposure to wildfire hazards. Although data-driven machine learning approaches have been widely applied to wildfire spread prediction, they often lack physical consistency and do not adequately represent the landscape and environmental conditions controlling fire propagation, including vegetation and fuel conditions, surface moisture, topography, and dynamic weather. Remote-sensing-based Geospatial Foundation Models (GeoFMs) pretrained on large-scale remote sensing datasets provide rich environmental representations (i.e., embeddings) for Earth observation, yet their potential for wildfire spread simulation remains underexplored. In this research, we propose a physics-informed GeoAI framework that couples GeoFM-derived environmental embeddings with downstream decoders tailored for wildfire spread prediction, with three complementary deep learning strategies, including a physics-informed parameter learning, a physics-informed loss function enforcing plausible fire-state transitions, and a physics-guided forward bias representing the slope and wind-driven propagation effects. In addition, the proposed GeoAI framework integrates multiple physical coupling mechanisms, including state transition, environmental propagation factors, and heat transfer. Specifically, an LSTM-based decoder is designed for mapping the sequential hidden states with physics constraints for the forward process and physics-informed loss optimization, while a CNN–based partial differential equation (PDE) decoder explicitly models wildfire propagation through a learnable PDE formulation and integrates it with physics-based parameter learning within a unified framework. Using 45 large wildfire events in California, the experiment results demonstrate that the proposed physics-informed GeoAI framework outperforms both purely data-driven deep learning and physical models. Specifically, the GeoFM-LSTM and GeoFM–PDE decoders achieve up to 15\% and 6\% improvements in average F1-score (classification performance), respectively. Furthermore, the proposed framework provides insights on process-level interpretability by characterizing diffusion, advection, and source-dominated wildfire spread patterns.