Generating 1 km Seamless Land Surface Temperature from China FY3C Satellite Data Using Machine Learning
Xinhan Liu, Weiwei Zhu, Qifeng Zhuang, Tao Sun, Ziliang Chen
Land Surface Temperature (LST), as a core variable in the coupling of land–atmosphere energy transfers and ecological responses, relies heavily on the global coverage capacity of thermal infrared remote sensing (TIR-LST) for dynamic monitoring. Currently, the time reconstruction method of the TIR-LST products from China’s Fengyun polar-orbiting satellite under dynamic cloud interference remains under exploration. This study focuses on the Heihe River Basin in western China, and addresses the issue of cloud coverage in relation to the Fengyun-3C (FY-3C) satellite TIR-LST. An innovative spatiotemporal reconstruction framework based on multi-source data collaboration was developed. Using a hybrid ensemble learning framework of random forest and ridge regression, environmental parameters such as vegetation index (NDVI), land cover type (LC), digital elevation model (DEM), and terrain slope were integrated. A downscaling and multi-factor collaborative representation model for land surface temperature was constructed, thereby integrating the passive microwave LST and thermal infrared VIRR-LST from the FY-3C satellite. This produced a seamless LST dataset with 1 km resolution for the period of 2017–2019, with temporal continuity across space. The validation results show that the reconstructed data significantly improves accuracy compared to the original VIRR-LST and demonstrates notable spatiotemporal consistency with MODIS LST at the daily scale (annual R2 ≥ 0.88, RMSE < 2.3 K). This method successfully reconstructed the FY-3C satellite’s 1 km level all-weather LST time series, providing reliable technical support for the use of domestic satellite data in remote sensing applications such as ecological drought monitoring and urban heat island tracking.