SHAP-based comparisons between the street- and roof-level urban heat islands in Seoul, Republic of Korea

초록

Canopy urban heat island (CUHI) is shaped by both atmospheric conditions and urban form, and these effects on UHI can differ by observation height. This study investigates how meteorological and urban variables affect UHI intensity at street level (mean 3.6 m) and roof level (mean 34 m) in Seoul under four scenarios: summer daytime, summer nighttime, winter daytime, and winter nighttime. We developed eight Random Forest regression models and applied SHAP method to (1) identify the relative importance of predictors, (2) evaluate group-level (meteorology vs. urban factors) and individual-variable contributions, and (3) analyze nonlinear relationships and threshold effects of urban predictors. The results show that roof-level UHI was dominated by meteorological factors, which accounted for approximately 73% of total absolute SHAP values on average, with wind speed being the most influential variable. In contrast, street-level UHI was more influenced by urban factors (approximately 40%, compared with approximately 25% at roof level), with building coverage and greenspace ratios exerting the greatest influence among urban variables. These height-based contrasts were strongest in summer, and urban factors were more pronounced during nighttime. Land-cover variables exhibited similar nonlinear patterns with threshold ranges, whereas height-related morphological variables, such as building height and canyon ratio, showed opposing directional effects between street and roof levels. These findings suggest that variable importance and effect direction vary with observation height, and therefore UHI mitigation should be approached as a context-dependent strategy that selects the most effective combination of interventions suited to local morphological and land-cover conditions.

키워드