A practical two-step framework for all-sky land surface temperature estimation

Land surface temperature (LST) is a key parameter in global ecological and climate system, while its accurate estimation under cloudy-sky conditions remains a challenge. In this paper, a practical two-step framework has been proposed for all-sky LST estimation, which effectively combined physical mechanisms and machine learning algorithms. The first step was the estimation of hypothetical clear-sky LST. The Bayesian optimizationbased Extreme Gradient Boosting (BO-XGB) model was developed to establish the implicit and complex relationship between hypothetical clear-sky LST and corresponding independent variables, including hypothetical clear-sky radiation energy input, surface biophysical parameters, atmospheric conditions, as well as spatial and temporal features. The second step was the cloud radiative forcing effect (CRFE) correction for hypothetical clear-sky LST under cloudy-sky conditions. Based on the surface energy balance (SEB) equation and conventional force-restore method, the analytical expressions of CRFE correction terms were derived, which speeded up the calculation and facilitated error analysis. Based on several remote sensing products, including Advanced Baseline Imager (ABI) LST and auxiliary data from Moderate Resolution Imaging Spectroradiometer (MODIS) and Global LAnd Surface Satellite (GLASS), as well as ERA5 reanalysis data, the proposed framework was applied to recovered spatially continuous hourly all-sky LST information over the contiguous US (CONUS) in 2018. The accuracy of the proposed method has been validated by ground measurements from 60 representative sites operated by the Surface Radiation Budget (SURFRAD) and AmeriFlux. For clear-sky conditions, the overall root mean square error (RMSE) was 2.67 K, with a bias of -0.47 K and R2 of 0.96. For cloudy-sky conditions, the corresponding accuracy metrics were 2.60 K, -0.49 K, and 0.96, respectively, and RMSEs for all sites were below 4 K, which indicated the stability of the proposed method. Besides, daily average LST has also been calculated based on the estimated hourly LSTs with RMSE of 1.68 K and bias of -0.48 K.