Accelerating Computation for Estimating Land Surface Temperature: An Efficient Global–Local Regression (EGLR) Framework

Rapid urbanization elevates land surface temperature (LST) through complex urban spatial relationships, intensifying the urban heat island (UHI) effect. This necessitates efficient methods to analyze surface urban heat island (SUHI) factors to help develop mitigation strategies. In this study, we propose an efficient global–local regression (EGLR) framework by integrating XGBoost-SHAP with global–local regression (GLR), enabling accelerated estimation of LST. In a case study of Wuhan, the EGLR reduces the computation time of GLR by 44.21%. The main contribution of computational efficiency improvement lies in the procedure of Moran eigenvector selecting executed by XGBoost-SHAP. Results of validation experiments also show significant time decrease of the EGLR for a larger sample size; in addition, transplanting the framework of the EGLR to two machine learning models not only reduces the executing time, but also increases model fitting. Furthermore, the inherent merits of XGBoost-SHAP and GLR also enables the EGLR to simultaneously capture nonlinear causal relationships and decompose spatial effects. Results identify population density as the most sensitive LST-increasing factor. Impervious surface percentage, building height, elevation, and distance to the nearest water body are positively correlated with LST, while water area, normalized difference vegetation index, and the number of bus stops have significant negative relationships with LST. In contrast, the impact of the number of points of interest, gross domestic product, and road length on LST is not significant overall.