H. Chen
Colorado State University, Colorado, United States
Keywords: AI, satellite remote sensing, resilience, weather prediction
Accurate and timely monitoring of natural disasters such as wildfire and extreme storms is essential for decision-making regarding energy infrastructure and resilience. Recent advances in deep learning (DL) have shown promise for environmental applications using satellite data; successful examples include precipitation retrievals and wildfire mapping. Despite the availability of various satellite data, high-resolution near-real-time applications remain challenging due to the inherent limitations in the spatiotemporal resolution of current satellite observations. For example, the geostationary satellites offer frequent observations, yet its coarse spatial resolution limits the accuracy of quantitative applications. In contrast, low-Earth orbit (LEO) satellites provide high spatial resolution but significantly low temporal coverage. To overcome these limitations, this research presents a novel DL model to downscale satellite data, with an emphasis on precise and timely wildfire detection across different geographical regions. In addition, this study develops an AI forecasting model to improve multi-step wildfire and precipitation prediction. This forecasting model employs a patch-based Swin Transformer backbone with periodic convolutions to handle longitudinal continuity and integrates time and noise embeddings via conditional layer normalization. A dual-branch decoder separately predicts wildfire and total precipitation, enabling targeted freezing of the corresponding decoder–encoder pathways to facilitate specialized training for each task.