Near real-time land surface temperature reconstruction from FY-4A satellite using spatio-temporal attention network

Land Surface Temperature (LST) is a critical parameter for climate studies and land surface process models as it indicates ground surface temperature variations across landscapes and timescales. However, satellite-based LST products derived from infrared sensors suffer from substantial missing values due to extensive cloud covers on the Earth's surface. Traditional methods rely heavily on numerical LST simulations for gap-filling, but the latency significantly limits the timeliness of gapless LST products. In this study, a novel deep learning method called the Spatio-Temporal Attention Network (STAN) was proposed, which was based on a U-Net architecture but enhanced with two unique feature extraction modules for capturing spatially and temporally dependent LST variations. Unlike many previous methods depending highly on numerical simulations, STAN reconstructs LST relying on spatiotemporal context information learned from historical memories, enabling more efficient LST reconstruction in a more timely manner. Ground validation results demonstrate better performance of STAN over other companion methods, with root-mean-square errors of 1.99 K and 2.89 K under clear and cloudy sky respectively, when reconstructing LST data collected from the Chinese Fengyun-4A geostationary satellite in the Yangtze River Delta. Intercomparison studies and error analysis also confirm the superiority of STAN, showing high LST reconstruction accuracy across different land covers and seasons. Overall, the proposed STAN method offers a much more efficient solution to facilitate timely LST reconstruction, and the method can also be easily transferred to other parameters with significant spatio-temporal variation context.