A multimodal transformer neural network framework for fine-scale wildfire occurrence forecasting

Abstract Wildfires have emerged as a major natural hazard under ongoing climate change, posing significant risks to ecosystems, infrastructure, and human life. While ignition events are often anthropogenic, wildfire propagation is predominantly governed by environmental drivers. These include (i) atmospheric conditions (e.g., temperature, wind velocity and direction, and relative humidity), (ii) fuel characteristics such as the spatial distribution, load, and type of combustible vegetation, and (iii) topographic features that modulate microclimate, precipitation patterns, and fire spread dynamics. Consequently, the development of accurate, high-resolution wildfire forecasting models constitutes a critical and challenging problem in environmental science. In this study, we propose a multimodal Transformer-based deep learning framework for fine-scale spatiotemporal wildfire occurrence prediction. The model integrates heterogeneous data streams across multiple scales, combining coarse-grained meteorological forecasts (e.g., hourly weather predictions) with fine-grained geospatial information, including terrain morphology and vegetation structure derived from high-resolution Google Earth imagery. The architecture leverages attention mechanisms to capture complex nonlinear dependencies and cross-modal interactions between atmospheric, ecological, and topographic variables. The proposed model is trained using historical wildfire records in the United States from 1992 to 2015, enabling it to learn probabilistic mappings between environmental conditions and ignition likelihood. Given operational weather forecast data and the most recent available pre-fire satellite imagery, the trained model produces spatially explicit predictions of wildfire occurrence probability at a resolution of approximately $$100~ m^2$$ , with a forecasting horizon of up to 24 h. Because inference is computationally inexpensive once the model is trained, the framework provides a scalable, data-driven approach for rapid, fine-scale wildfire risk assessment, although the present implementation is an offline-trained forecasting model rather than a continuously updated operational monitoring system.