Unraveling the multi-scale drivers of the urban thermal environment through interpretable machine learning models

Unraveling the multi-scale drivers of the urban thermal environment is critical for developing effective heat mitigation strategies. While land surface temperature (LST) is a key indicator, existing studies often lack a comprehensive analysis of multiple factor types, spatial scales, and nonlinear relationships. To address this gap, this study develops a multi-scale, interpretable machine-learning framework based on XGBoost model to quantify the nonlinear relationships between LST and 23 potential driving factors across ten spatial scales and four seasons in Shenzhen. The results show that: (1) most factors exert the strongest influence in summer, with NDVI, green space aggregation index (AI_G), sky view factor (SVF), and floor area ratio (FAR) identified as the dominant drivers; (2) factor contributions are strongly scale-dependent, with NDVI and AI_G consistently contributing more than 10%, while NDBI and building density (BD) are more influential at finer scales and population density (POP) at broader scales; and (3) the 2100 m scale is the optimal analytical scale, at which higher vegetation coverage, SVF, and AI_G are associated with lower LST, whereas higher FAR, POP, and cultivated land patch density (PD_C) are associated with higher LST. By integrating multi-scale analysis with an interpretable machine-learning framework and threshold analysis at the optimal scale, this study provides both a methodological contribution to understanding scale-dependent thermal drivers and practical guidance for sustainable urban heat mitigation and planning.