Causal Machine Learning Uncovers Thresholds and Seasonal Reversal in Landscape-Meteorology-Driven Urban Heat-PM₂.₅ Coupling

In the context of rapid urbanization, the mutual amplification between the urban heat island (UHI) effect and air pollution has emerged as a severe composite environmental challenge. Most existing studies focus on the linear driving mechanisms of individual environmental issues, leaving a systematic understanding of their nonlinear interactions, seasonal dynamics, and causal pathways largely unexplored. To address this gap, this study focuses on the Yangtze River Delta urban agglomeration in China as a case study. It bridges interpretable machine learning (XGBoost-SHAP) with structural causal modeling (SCM) to systematically quantify the nonlinear impacts and causal mechanisms of blue-green space (BGS) patterns and meteorological factors on land surface temperature (LST) and PM 2.5 concentrations. The results indicate that: (1) LST drivers exhibit a pronounced seasonal shift-BGS coverage dominates in summer, whereas meteorological factors prevail in winter-whereas PM 2.5 is consistently governed by wind speed across seasons; (2) key drivers exhibit threshold effects (e.g., in this study, 40-42% vegetation coverage for LST, 79-80% water aggregation for PM 2.5 ), beyond which they act as “effect-switching thresholds” that reverse the influence of other factors and form a nonlinear interaction network; and (3) SCM reveals the dual causal pathways through which BGS patterns influence the environment-via direct cooling and indirect modulation of wind fields-and quantitatively verifies the stable positive causal effect of PM 2.5 on LST, clarifying the pathway of “pollution exacerbating warming”. In summary, this study uncovers the nonlinear, threshold-dependent characteristics of the UHI-pollution complex system and elucidates their synergistic intensification from a causal perspective, providing a quantitative basis for seasonally adaptive strategies to synergistically mitigate urban heat and pollution.