Objective Against the dual pressures of global climate change and rapid urbanization, the urban heat island effect is becoming increasingly severe. Urban green spaces offer a near-natural approach to mitigating the heat island effect. Their cooling effects are influenced by biotope characteristics and the differences among biotope types. Therefore, typologically revealing the impact patterns of urban green space biotope characteristics on cooling effects is crucial for improving the urban thermal environment and contributing to the construction of an ecologically livable environment.

Methods The study area is Around Taihu Lake region in the Yangtze River Delta, involving four cities (Suzhou, Wuxi, Changzhou, Huzhou) encompassing 16 county-level administrative units, covering approximately 13,000 km². This region, centered around Taihu Lake, features a dense river network, numerous lakes, and diverse types of urban green spaces containing various biotopes. However, the urban heat island problem has become increasingly severe in recent years. Consequently, optimizing biotope type configuration during urban ecological restoration to enhance the cooling effect of green spaces has become a significant focus for building an ecologically livable city in this region. This study first introduces a U-Net-ResNet50 backbone training network combined with high-resolution remote sensing images to build an intelligent identification model, overcoming the technical bottleneck of traditional methods struggling with large-scale, high-precision biotope surveys and classification. Building the intelligent identification model mainly involves four steps: 1) Using 25,983 satellite images from 508 green space samples as the model database. Subsequently, analyzing the composition characteristics of biotope elements in these 508 urban green space samples, the study identified 10 natural biotope elements as targets for manual annotation and semantic segmentation recognition. 2) Manually annotating local satellite images containing all biotope element types to create a foundational dataset for supervised learning model training. After model training and optimization, the supervised model was applied to annotate biotope elements in satellite images from other areas, ultimately producing the training and testing datasets required for the U-Net−ResNet50 network model. 3) A semantic segmentation platform was constructed using the PyTorch framework, with ResNet50 and U-Net as the backbone training networks. The prepared training set data was input for training, and a loss function was set to evaluate the model's robustness. Upon completion of training, the model achieved good stability and consistency. 4) Applying the trained and verified reliable model to Around Taihu Lake region. Subsequently, based on the intelligent identification results of urban green space biotope elements, this study used K-means clustering to derive subdivided biotope types. Next, using 30 m resolution land surface temperature retrieval data from a typical high-temperature day in 2022, this study employed mean and maximum land surface temperatures to characterize the cooling effect. Cross-analysis was used to reveal the differences in cooling effects among different biotope types and key influencing factors. Taking biotope types as the entry point, this research further refines the exploration of the cooling effects of urban green spaces. The study proposes differentiated preferential design guidance for biotopes, providing technical support for the precise ecological restoration of climate-resilient urban green spaces.

Results 1) A total of 6 primary-level and 39 secondary-level green space biotope types were identified through detailed classification. 2) The cooling effects of the primary biotope types were ranked in descending order: woodland biotope > waterfront composite biotope > vegetation composite biotope > farmland composite biotope > grassland biotope > hard-paved biotope. Urban-rural comparative results showed that the thermal environment stability of biotope types in suburban areas was significantly better than that in urban areas. 3) The cooling effects of various biotopes were influenced by the urban-rural gradient, with different key biotope characteristic factors identified as drivers. The cooling effect of woodland biotopes was mainly driven by topography and slope. For vegetation composite biotopes, vegetation structure was the core factor in urban areas, while a synergistic effect of topography and forest stand structure was observed in suburban areas. The cooling effect of waterfront composite biotopes was primarily governed by multiple factors, including water surface ratio, vegetation structure, and tree canopy coverage. Within farmland composite biotopes, agroforestry models with higher tree canopy coverage demonstrated significant cooling effects. 4) Based on the heterogeneity in the spatial distribution of cooling effects across different green space biotope types along the urban-rural spectrum, biotope types with high cooling efficacy were preferentially selected. For urban areas characterized by limited land resources, high development intensity, and reliance on single biotope element drivers, design strategies for specific biotope types were precisely targeted and regulated. For suburban areas, which benefit from a better ecological baseline and rely on the synergistic driving effects of multiple biotope elements, design strategies for specific biotope types were systematically enhanced.

Conclusion The U-Net-ResNet50 intelligent identification model facilitates the precise and batch-processing-enabled identification and classification of urban green space biotope elements. This approach overcomes the technical limitations of traditional biotope element identification methods, namely, low efficiency and restricted sample sizes, and provides a transferable intelligent methodology for precise biotope mapping and classification in related fields. The study elucidates the differential patterns in cooling effects and their driving factors across various urban green space biotope types, thereby robustly supporting the preferential design of green space biotopes. It offers a valuable reference for achieving climate-resilient, near-natural ecological restoration in urban green spaces.