Key points are not available for this paper at this time.
Urban parks are vital thermal regulators in rapidly urbanizing regions, while advances in spatial metrics and machine learning exist, the systematic identification of environmental thresholds across multiple cooling metrics and translating multidimensional cooling analyses into actionable design guidelines for medium-to-large urban parks remain challenging. This study applies an ensemble engineering pipeline by integrating Random Forest feature selection, LightGBM modeling, and SHAP interpretability to analyze cooling performance across 196 parks in Shenzhen. Our analysis reveals critical compound thresholds: Parks under 20 ha achieve optimal cooling effect when situated in high-vegetation zones, whereas larger parks maintain stable cooling efficiency. Forest area benefits for park cooling intensity plateau at approximately 5 ha but continue enhancing cooling area, indicating metric-specific optimization needs. Road density exceeding 15% significantly impairs cooling performance, with distinct thresholds per metric. We translate these findings into practical design rules: Maintaining tree heights at 10–15 m maximizes cooling efficiency for smaller parks, while limiting impervious surfaces below 40% enhances cooling extent. By systematically examining interactions between two-dimensional landscape metrics, three-dimensional vegetation characteristics, and urban context across multiple cooling indicators, this work demonstrates how ensemble machine learning techniques, when coupled with interpretability methods, can generate implementable thermal resilience strategies within existing urban design frameworks. The study bridges the gap between complex nonlinear analysis and actionable engineering guidance for subtropical megacities facing acute heat challenges.
Tang et al. (Thu,) studied this question.