Video Swin-Hybrid-U-Net predicts Canadian wildfire spread from satellite data

A team led by Maulik Srivastava, Esha Saha, and Hao Wang has introduced a deep learning framework called Video Swin-Hybrid-U-Net to forecast wildfire spread across Canada. Published on arXiv (submitted June 15, 2026), the model combines a Video Swin Transformer encoder—which captures both spatial and temporal attention over three-day sequences of environmental variables—with a convolutional decoder to predict next-day fire incidence maps. All data is sourced from public repositories via Google Earth Engine, including meteorological and satellite imagery covering major Canadian wildfire events from 2014 to 2023. This design ensures transparency, scalability, and reproducibility, addressing a key limitation of many existing models that rely on proprietary datasets.

The model achieved strong predictive performance by effectively modeling the complex, landscape-specific dynamics of Canadian wildfires. Unlike traditional approaches that struggle with temporal variability, the Video Swin Transformer’s attention mechanism allows the model to weigh relevant past conditions across both time and space. The researchers position this work as a foundation for advanced spatio-temporal forecasting and operational applications, such as real-time risk assessment and resource allocation for firefighting. With its open-data pipeline, the framework could be adapted to other regions or integrated into government early-warning systems. The preprint has been submitted to the International Journal of Wildland Fire.