Investigating the effect of urban form on land surface temperature at block and grid scales based on XGBoost-SHAP

Abstract

The urban thermal environment is becoming increasingly severe. In this study, we integrated eXtreme Gradient Boosting with the SHapley Additive exPlanations method to investigate the effects of various urban factor indexes (UFIs) on land surface temperature (LST) at both block and grid scales. Additionally, we examined the differences in LST and its driving factors across local climate zones (LCZs) at the grid scale. The results show that LST is higher in central areas than in peripheral ones during summer and autumn, but this pattern is reversed in spring and winter. LST varies significantly across LCZs, with the normalized difference built-up index, normalized difference vegetation index (NDVI), and Shannon's diversity index (SHDI) identified as the main contributors. The sky view factor inhibits LST at the block scale but promotes it at the grid scale. The impacts of UFIs follow the seasonal trend: summer > spring > autumn > winter. LST responses to UFIs exhibit similar trends across scales, showing specific warming or cooling thresholds—for example, a cooling effect when SHDI exceeds 0.65, and a warming effect when building density exceeds 20 % (summer and autumn) or 40 % (spring and winter). Significant cooling occurs only when NDVI exceeds 0.4; however, NDVI generally remains low in all seasons except summer. High-contribution UFIs typically exhibit the strongest interaction effects with artificial factor indicators.