Deep Learning for Near-Surface Air Temperature Estimation from FY-4A Satellite Data

Near-surface air temperature is a crucial weather parameter that significantly impacts human health and is widely utilized in numerical weather forecasting and climate prediction studies. However, the most common ground-based meteorological station observation and radar observation are often limited by geographic and natural constraints. With the advantages of global coverage and high spatiotemporal resolution, satellite remote sensing has become a valuable support in overcoming data scarcity issues related to ground-based station and radar observations in complex geographic and natural conditions. Although remote sensing indirectly reflects atmosphere variables (e.g., near-surface air temperature), accurately estimating the atmosphere variables through satellite remote sensing remains a significant challenge. This paper introduces a deep learning Transformer-based neural network (TaNet) for near-surface air temperature estimation. TaNet automatically extracts information from imageries captured by China's new-generation geostationary meteorological satellite FengYun-4A and generates grid near-surface air temperature data in near real-time. Extensive experiments conducted using the state-of-the-art operational reanalysis product ERA5 and meteorological station observations as benchmark standards demonstrate the effectiveness and superiority of TaNet. It achieves an impressive Pearson's correlation coefficient (CC) of 0.990 with ERA5 and 0.959 with station observations, outperforming the other products, such as CFSv2, CRA, and U-Net, on root mean square error (RMSE) and CC metrics. TaNet reduces the RMSE of CFSv2, CRA, and U-Net by a margin of 10.551% (2.594 ${}^{\circ }\mathrm{C}$ vs. 2.900 ${}^{\circ }\mathrm{C}$ ), 2.261% (2.594 ${}^{\circ }\mathrm{C}$ vs. 2.654 ${}^{\circ }\mathrm{C}$ ), and 5.535% (2.594 ${}^{\circ }\mathrm{C}$ vs. 2.746 ${}^{\circ }\mathrm{C}$ ), respectively, using station observations as the benchmark.