Analysis of land surface temperature drivers in Beijing’s central urban area across multiple spatial scales: An explainable ensemble learning approach

Urbanization has accelerated dramatically in recent years, making the urban heat island effect a critical factor that adversely impacts both urban environments and public health. Land surface temperature (LST), a key indicator of the urban heat island effect, is influenced by a complex interplay of factors from the built, natural, and socio-economic environments.However,most existing studies have predominantly focused on single-factor analyses, lacking a comprehensive multidimensional approach. To address this gap, this study investigates the influence mechanisms of various environmental factors on LST through multi-scale spatial analysis and proposes a machine learning model that explicitly incorporates spatial effects. An integrated dataset encompassing variables from the built, natural, and socio-economic environments was constructed. Spatial autocorrelation analysis confirmed that the spatial effects of LST follow a spatial error model (SEM). Based on this spatial effect pattern, a spatial machine learning model was developed that significantly outperformed traditional spatial learning models. By comparing seven commonly used machine learning models, we selected the four best-performing models—CatBoost, XGBoost, Gradient Boosting, and HistGradientBoosting—for ensemble learning, ultimately adopting the Stacking strategy over the Voting approach. Incorporating Bayesian optimization for hyperparameter tuning further enhanced the Stacking model’s predictive performance.Furthermore, interpretability tools including SHAP, Partial Dependence Plots (PDP), and Individual Conditional Expectation (ICE) were employed to systematically analyze the factors influencing LST. The analyses revealed that the Normalized Difference Vegetation Index (NDVI) and spatial error are the most critical determinants of LST, while built environment features such as building height, as well as natural environment factors like green spaces and water bodies, also have significant impacts. This study provides robust scientific evidence for urban planners and offers valuable guidance for developing effective strategies to mitigate extreme heat and alleviate the urban heat island effect.