Investigating heterogeneous non-linear effects of key driving factors on the urban thermal environment in urban centres and non-centres: A case study of Shanghai

Urban heat-island research has typically focused on city-wide polycentric spatial structure, leaving the spatially heterogeneous effects of driving factors on local thermal dynamics in centres and non-centres largely underexplored. To address this gap, we combined fine-scale functional zoning with interpretable machine learning to uncover microclimatic dynamics in Shanghai. We divided the city into 500 × 500 m grid cells and classified them into employment centres, residential centres, or non-centres using mobile-signalling data. An XGBoost model was trained to predict land surface temperature (LST), and SHAP analysis was used to quantify the nonlinear and threshold effects of nine key drivers. The results show that economic agglomeration represented by nighttime light and population density raises LST by 1–3 °C across all grid types except employment centres. In employment centres the floor area ratio slightly cools LST while in residential centres it increases LST, and road density warms LST in employment centres but cools it in residential centres, which demonstrates the different nature of employment centres from residential centres. Socio-economic drivers such as nighttime light intensity and population density boost LST by 0.5–3 °C before their effects saturate or reverse in core work districts. Vegetation exhibits a consistent threshold effect, warming LST below an NDVI of 0.35 and cooling it above this value. Notably, natural environment factors influence LST with remarkable consistency regardless of grid function. Study integrates functional zoning with interpretable modelling to reveal zone-specific thermal thresholds and interactions, guiding the design of targeted, function-aware interventions to mitigate urban overheating.