● Links transit oriented development patterns to land surface temperature in Shinjuku ● Uses street view imagery, remote sensing and population data in one framework ● Shows well designed transit corridors can cool land surface by about 0.9°C ● Reveals that excessive density without greenery and shading increases urban heat ● Applies explainable machine learning to guide climate sensitive station area design Extreme heat is an escalating urban health and equity risk. Transit-oriented development (TOD) is widely promoted for low-carbon mobility, yet its thermal effects remain contested: densification may intensify midday surface heat exposure, whereas corridor design and streetscape morphology may mitigate it. Using Shinjuku Ward, Tokyo, we integrate summer daytime Landsat Collection 2 Level-2 land surface temperature (LST) with multimodal urban-form indicators derived from street-view semantics, OpenStreetMap kernel densities, building footprints, and fine-scale population (n = 12,562 street points). An XGBoost model learns a nonlinear mapping to Landsat-referenced LST and is interpreted with SHAP and one- and two-dimensional partial dependence plots. The model explains 92.1% of observed LST variance (R² = 0.921) and reveals strong non-additive interactions: higher railway-station intensity is associated with lower model-estimated LST only when paired with high railway-line density, indicating a corridor-node synergy rather than a station-only effect. Across the Railways × Railway Stations interaction surface, the model-predicted LST range spans approximately 1.4°C within the Landsat-referenced temperature space, demonstrating substantial contrasts among TOD configurations without comparing independent temperature products. Additional interaction surfaces suggest nonlinear thresholds involving building intensity, road configuration, and streetscape sky-tree composition. We discuss limitations of daytime LST as a proxy for pedestrian thermal comfort and outline transferability to other cities by adapting locally available mobility and morphology indicators.
Xie et al. (Sun,) studied this question.