The effects of 3D architectural patterns on the urban surface temperature at a neighborhood scale: Relative contributions and marginal effects

Urban architecture is an important contributor to urban heat island (UHI) effects. Yet thorough investigations into how three-dimensional (3D) architectural patterns influence urban thermal conditions collectively and individually are limited. This study bridges this gap by adopting a machine learning method, boosted regression tree (BRT), to analyze the relative influences and marginal effects of 3D urban architecture on land surface temperature (LST). Ten architectural metrics are incorporated to describe the composition and configuration of 3D architectural patterns at a neighborhood scale in the typical megacity of Shanghai. The results show that in summer, the building coverage ratio (BCR), mean architecture height (MAH), mean architecture height standard deviation (AHSD) and mean architecture projection area (MAPA) are the most influential factors, with relative contributions of 39.3%, 16.5%, 12.3% and 10.4%, respectively. The regulation amplitudes (Delta T) of the dominant metrics for the neighborhood average LST temperature are 2.7 degrees C, 0.9 degrees C, 0.6 degrees C, and 0.6 degrees C, respectively. Moreover, with the LST, the BCR exhibits a monotonic positive correlation, the MAH and AHSD show a stepwise negative correlation, and the MAPA shows a combination of positive and negative correlation. It is generally recommended to decrease the development intensity and architectural base area, while increase the building height and roughness to improve the urban thermal conditions at the neighborhood scale. The dominant contributors and related marginal effects are generally consistent across different seasons. These findings can provide quantitative insights for mitigating the LST effects via rational design and management of 3D architectural patterns. Given the distinctive insights provided, the BRT method is recommended for disentangling the relationship between LST and environmental variables in upcoming studies. (C) 2020 Elsevier Ltd. All rights reserved.