Examining temporally varying nonlinear effects of urban form on urban heat island using explainable machine learning: A case of Seoul

Many empirical studies have examined the relationship between urban form and canopy layer urban heat island (CUHI). However, these studies mostly have used small sample sizes and statistical methods that assume CUHI responds linearly to urban form features. Additionally, the differences in the effect of urban form on CUHI at different time instances during the diurnal cycle have not received sufficient attention. To address these issues, this study employs a relatively large sample of 1,058 microclimate sensors in Seoul, Korea, and analyzes CUHI at 9 a.m., 3 p.m., and 9 p.m. on a typical summer day. Gradient boosting decision tree (GBDT) based regression model and linear regression model and their variants with spatial information were constructed and compared for each time instance. Results show that spatially explicit GBDT models had the highest accuracy. Feature importance analysis shows that built form factors such as mean building height were more important during the afternoon, while surface fractions such as road surface fraction had greater influences during the morning and nighttime. Partial dependence plots (PDPs) show that urban form features influence CUHI in a complex nonlinear manner that varies for each time instance. PDPs were further scrutinized based on their activation and threshold effects. GBDT model findings directionally aligned with linear regression, but they provided nuanced insights into the form-CUHI relationship. These findings help planners to better understand the complexity of urban climate and intervention required to reduce CUHI magnitude across the diurnal cycle.

Automated summary

This study investigates the relationship between urban form and canopy layer urban heat island (CUHI) using a relatively large sample of microclimate sensors in Seoul, Korea. The study compares different statistical models and finds that the spatially explicit gradient boosting decision tree (GBDT) model has the highest accuracy. The study also shows that the effect of urban form on CUHI varies at different time instances during the day. These findings provide valuable insights for planners to understand the complexity of urban climate and reduce CUHI magnitude.