Modeling and optimization of heat island networks based on machine learning and the perspective of spatial heterogeneity in metropolitan areas

Global climate change has intensified the regional linkage of the urban heat island effect (UHI), posing major challenges to urban sustainability. This study examines three major metropolitan areas in China to understand how thermal environments interact under different geographical conditions. It aims to improve the heat island network structure and enhance governance strategies. The study combines XGBoost machine learning with the GeoSHapley method to build a high-precision thermal resistance surface, overcoming the bias of traditional methods and capturing nonlinear effects, spatial variation, and factor interactions. The heat island network is modeled using circuit theory and evaluated and verified using structural indicators such as α closure, β line-to-point ratio and γ connectivity. Key findings include: 1) Resistance factors vary significantly across regions. In Chongqing, topographic factors (DEM and SLOPE) account for 72 % of the resistance, dominating the network. In Xiamen-Zhangzhou-Quanzhou and Wuhan, NDBI contributes 0.37 and 0.38, respectively, as the main driver. 2)GeoSHapley analysis identified cooling thresholds for resistance factors. For example, NDVI thresholds are [0.75–0.85] in Xiamen-Zhangzhou-Quanzhou and [0.65–0.70] in Wuhan, offering a scientific basis for resistance classification. 3)The Chongqing network shows the highest connectivity (γ = 0.81) and integrity (α = 0.70), with a model fit of R² = 0.907. Xiamen-Zhangzhou-Quanzhou also performs well, proving the method works in complex terrain. This study improves upon past methods by introducing dynamic resistance classification and interaction analysis. It offers a scalable framework for managing urban thermal environments based on local geography, supporting ecological and sustainable urban development.