Objective With the acceleration of urbanization, the inter-regional transmission of urban heat island (UHI) effect has become increasingly prominent, posing an urgent environmental challenge. Traditional methods for constructing resistance surfaces in UHI studies are limited by their subjectivity and inability to fully account for spatial heterogeneity, restricting a deep understanding of thermal environment linkage mechanisms among urban agglomerations. This study focused on the Xiamen-Zhangzhou-Quanzhou metropolitan area to construct a cold and heat island spatial network, aiming to reveal the thermal environment linkage mechanisms between urban clusters, and propose multi-scale collaborative optimization strategies. These strategies provide a scientific basis for optimizing urban thermal environments, and support regional climate adaptation planning to effectively address the challenges posed by UHI effect.

Method To achieve the above objectives, this study innovatively integrated the XGBoost-GeoSHapley interpretable machine learning model with circuit theory. By employing the XGBoost-GeoSHapley additive explanation analysis method, combined with land surface temperature classification results, we accurately constructed resistance surfaces for cold and heat islands, and quantified the spatial effects, non-linear relationships, and interactions of resistance factors. Building on this foundation and using circuit theory, we precisely identified the cold and heat island networks, detecting key pinch points, obstacle points, and centrality areas. Finally, we proposed targeted optimization strategies informed by regional characteristics, offering specific guidance for optimizing the thermal environment in the Xiamen-Zhangzhou-Quanzhou metropolitan area.

Result The cold island network exhibits a pattern dominated by northwestern mountainous areas, supplemented by coastal wetlands, while the heat island network is primarily concentrated in the eastern urban areas. A total of 43 heat island corridors (total length of 412.99 km) and 85 cold island corridors (total length of 1 693.08 km) were identified, along with a heat island priority rectification area of 1 067.25 km² (4.4%) and a cold island protection core area of 7 980.71 km² (32.9%). Compared with traditional methods, the cold and heat island network structure constructed in this study was more rational, with significantly improved connectivity and integrity. Key indicators for cold islands, such as the closure value α, line-to-point ratio β, and connectivity rate γ, outperformed their heat island indicators, indicating superior connectivity and integrity for the cold island network. This provides strong support for optimizing urban thermal environments.

Conclusion By integrating the XGBoost-GeoSHapley model with circuit theory, this study proposes a method for constructing cold and heat island spatial networks at the urban agglomeration scale, effectively overcoming the issues of subjective assignment and spatial heterogeneity in traditional resistance surface construction methods. Through enhanced precision in identifying and regulating cold and heat island corridors, we have established the “one belt, three axes, three cores, and many pieces” interaction pattern of cold and heat islands in the Xiamen-Zhangzhou-Quanzhou metropolitan area. We also propose a governance framework of “vertical blocking + horizontal channeling”, offering significant theoretical and practical references for constructing cold and heat island networks in megacity agglomerations, and providing substantial support for regional climate adaptation planning.