Objective The accelerated process of urbanization has intensified the urban heat island (UHI) effect, posing significant threats to public health and sustainable socio-ecological development. Urban fringe areas, characterized by their blue-green spaces (BGS), serve as critical ecological buffers.These spaces not only curb urban sprawl and mitigate the spread of the UHI effect but also provide essential peripheral cooling services to metropolitan cores. Consequently, enhancing the construction and functionality of BGS in these peripheral areas is of paramount importance for alleviating urban thermal environmental pressure and fostering coordinated ecological development between urban and rural landscapes. However, rapid urbanization has led to the increasing fragmentation of these vital BGS, which has severely impaired their ecosystem services, particularly their cooling capacity. While expanding green infrastructure is a direct approach, it is often infeasible under the constraints of limited land resources in densely populated regions. Therefore, a paradigm shift is required: instead of merely expanding the area of BGS, the focus should be on optimizing the connectivity among existing cooling sources to implement targeted and effective cooling strategies. Current research predominantly relies on qualitative descriptions or simplified linear models for thermal mitigation planning, resulting in a lack of robust and systematic analytical frameworks. This study aims to enhance cold source connectivity to improve the overall cooling efficiency of the urban fringe of megacities, thereby offering a novel, integrative analytical pathway for urban heat mitigation planning.

Methods Taking the Second Green Belt in Beijing as a case study, this study constructs an urban cooling network. Land surface temperature (LST) was retrieved from Landsat imagery to identify cold sources, and morphological spatial pattern analysis (MSPA) was applied to extract their core areas. Natural and socio-economic variables, together with landscape pattern indices, were incorporated as resistance factors. An XGBoost model combined with SHAP interpretation was employed to quantify the nonlinear relationships between these factors and LST, with absolute SHAP values used as weights to generate a resistance surface. Circuit theory was then applied to identify key cooling corridors and establish and optimize the cooling network. Complex network theory was subsequently employed to evaluate network performance before and after optimization.

Results LST retrieval revealed higher temperatures in the eastern, flatter, and more developed part of the study area, contrasting with the cooler, topographically complex western region with greater vegetation and water body coverage. The identified cool sources, predominantly located in the west, covered approximately 855.65 km². MSPA of these cool islands led to the identification of 53 core areas as major cool sources, primarily clustered in the northwestern and southwestern sectors, with smaller, fragmented sources in the east. The XGBoost-SHAP analysis revealed the relative importance of the resistance factors: the proportion of impervious surface (PLAND_IMP, 33.7%) was the most significant, followed by NDVI (19.0%), Nighttime light (15.1%), Building Height (9.8%), and the proportion of green space (PLAND_GRE, 9.5%). PLAND_IMP, NL, and BH contributed positively to LST (increasing resistance), while PLAND_GRE and NDVI had cooling effects (decreasing resistance). The resulting resistance surface exhibited a spatial pattern consistent with the LST distribution, with high values in impervious-dominated areas and low values in heavily vegetated northwestern zones. The initial cooling network comprised 117 corridors spanning 52.9 km. The optimization process proposed adding 11 new nodes from pinchpoint analysis and modifying 8 structurally weak nodes. The final optimized network incorporated these changes, resulting in 33 additional corridors, significantly enhancing connectivity, particularly in the eastern part of the study area. Stability tests demonstrated the optimized network's superior robustness.

Conclusion This study enhances the cooling effectiveness of blue-green spaces in the peripheral zones of megacities by integrating computer science with landscape ecology and strengthening the connectivity among fragmented cold sources, thereby providing a structural basis for near-natural restoration. The resistance-surface construction is advanced by incorporating landscape pattern indices and vertical-dimension factors, enabling a more accurate representation of surface heterogeneity. Explainable machine learning is employed to capture nonlinear interactions among resistance factors, allowing the cooling network to better reflect underlying ecological processes. By accounting for complex environmental conditions and disturbances, the study proposes a function–structure coordinated optimization framework that integrates cold-source pattern enhancement, land-use regulation, ecological corridor construction, and multi-sector governance to improve cooling efficiency and network stability.