Quantifying the nonlinear relationship between block morphology and the surrounding thermal environment using random forest method

Previous studies have extensively revealed the essential impact of urban morphology on the local thermal environment, but the nonlinear relationship has not yet been fully understood, especially for the diurnal vari-ation. In this study, the fine-scale investigation of the thermal environment is comprehensively assessed by field measurement in five representative blocks in Xi'an, China. Twenty-two potential morphological variables from spatial composition, land-use features, and surrounding environment were obtained from the geographic in-formation system. Both the ordinary least squares (OLS) and random forest (RF) models were adopted to explore the relationship between the thermal parameters and the predictor variables. The modeling results demonstrate that the RF model, effectively capturing the nonlinear relationship, outperforms the OLS model in study areas with a higher R2, a lower MSE and MAE values. The RF-based results clearly disclose that surrounding variables, such as distance to industry area (DI) and water bodies (DW), are generally the top-3 variables with the highest contribution rate (mean SHAP value>0.1) during the different periods. Building density is also a stronger driving factor than building height. Moreover, all important variables have an obvious nonlinear relationship with the threshold effect for the diurnal thermal environment variation. The proper thresholds for DI and DW are 3200 m and 1500 m, respectively, indicating that controlling DI larger than 3200 m and DW less than 1500 m may contribute to mitigating the heat island in Xi'an. Our study provides insights into machine learning models for thermal environment assessment and quantitative recommendations for decision-makers and urban planners to develop heat resilience cities.

Automated summary

This study comprehensively assessed the thermal environment at a fine-scale level in five representative blocks in Xi'an, China through field measurements. The results showed that the random forest model outperformed the ordinary least squares model in capturing the nonlinear relationship between thermal parameters and predictor variables. The analysis of variable importance revealed that variables related to the surrounding environment, such as distance to industry area and water bodies, were the major factors influencing the thermal environment variation. Proper control of these variables can help mitigate the heat island effect in Xi'an.