Associational and Causal Effects of Urban Characteristics on the Block-Scale Thermal Environment in Beijing

With the increasing frequency of extreme-heat events, there is an urgent need to identify the key drivers of the urban thermal environment at fine spatial scales. Focusing on urban blocks within Beijing’s Fifth Ring Road, this study integrates Landsat 8 imagery acquired from March 2020 to February 2021 with multi-source data on land cover, buildings, population, pollution, and topography. LightGBM–SHAP, a theory-informed directed acyclic graph (DAG), CausalForestDML, and cross-fitted g-computation were employed to investigate predictive associations, Q25–Q75 average total treatment effects, and block-level responses to prespecified urban-morphology intervention scenarios for seasonal land surface temperature (LST). The within-season SHAP analyses consistently placed building height (BH) and building density (BD) among the relatively important predictors, whereas the predictive patterns of the normalized difference vegetation index (NDVI) and the proportion of impervious surfaces (ID) were more season-specific. The autumn and winter models also assigned relatively high within-season importance to the digital elevation model (DEM), PM2.5, and CO2 emission proxy. Because the four seasonal models differed in predictive performance and LST distributions, these cross-seasonal patterns were interpreted qualitatively rather than as direct comparisons of absolute SHAP values or rank positions. Causal-effect estimation further indicated that, under the primary DAG and identification assumptions, the Q25–Q75 point estimates were positive for BD and negative for BH in all four seasons. In summer, the Q25–Q75 effects of NDVI and ID were −0.772 and +1.548 °C, respectively. Intervention-scenario analysis further indicated that the estimated responses varied across blocks and seasons, emphasizing the importance of considering baseline urban conditions and common support when interpreting potential planning effects. Additional spatially blocked validation yielded lower predictive performance than random validation, while significant positive residual spatial autocorrelation remained in all four seasons. These findings may inform the local evaluation of season- and context-specific surface-temperature mitigation strategies within the observed-support range, but they should not be interpreted as universal planning prescriptions.