Hourly land surface temperature retrieval over the Tibetan Plateau using Geo-LightGBM framework: Fusion of Himawari-8 satellite, ERA5 and site observations

The Tibetan Plateau (TP) is highly sensitive to even minor fluctuations in land surface temperature (LST), which can result in permafrost melting and degradation of alpine grasslands, leading to serious ecological consequences. Therefore, it is crucial to have high-temporal-resolution and seamless hourly estimating and monitoring of LST for a better understanding of climate change on the TP. Here, we employed Himawari-8 satellite, Digital Elevation Model (DEM), ERA5 reanalysis and meteorological station observations data to develop a new LightGBM framework (called Geo-LightGBM) for estimating LST on the TP, and then analyzed the spatiotemporal variations of those LST. Geo-LightGBM demonstrated excellent LST estimation accuracy, with an R² (coefficient of determination) of 0.971, RMSE (root-mean-square error) of 2.479 °C, and MAE (mean absolute error) of 1.510 °C. The estimated LST values for the year 2020 agreed well with observed values, with remarkable differences in hourly LST variations. Meanwhile, the estimated LST was more accurate than that from FY-4A. Spatially, there were two high LST centers, located in the Yarlung Zangbo River Basin and the Qaidam Basin, and a low LST center located in the central TP. The SHAP (SHapley Additive exPlanations) and correlation analyses revealed DSCS (the mean ground downward shortwave radiation under clear-sky conditions) to be the most importantly input variable for estimating LST. Spatiotemporal dummy variables (e.g., longitude, latitude, DEM) were also found to be crucial for model accuracy improvement. Our findings indicate the potential for constructing a high-precision and seamless 24-h LST real-time retrieval and monitoring platform for the TP by combining satellite and China's independently developed CLDAS (China Land Data Assimilation System) data in future.