首页|An interpretable spatially weighted machine learning approach for revealing spatial nonstationarity impacts of the built environment on air pollution

An interpretable spatially weighted machine learning approach for revealing spatial nonstationarity impacts of the built environment on air pollution

扫码查看
原文链接
  • NETL
  • NSTL
  • Elsevier
With urbanization acceleration, air pollution has become increasingly severe, and the built environment plays a crucial role in pollutant dispersion and accumulation. In this study, the central urban area of Nanjing is analysed in a case study, focusing on the plot scale to systematically explore spatial heterogeneity and nonlinear interactive mechanisms of built environment impacts. To accurately capture the complex spatial nonstationarity effects of the built environment on air pollution, a geographically weighted machine learning (GeoML) approach driven by an adjacency-based spatial mechanism integrated with the SHapley Additive exPlanations (SHAP) framework for interpretability is established. The findings reveal that the GeoML model outperforms other models (the ordinary least squares (OLS), random forest (RF), geographically weighted regression (GWR), and geographically weighted random forest (GWRF) models), with an R2 value of 0.74, effectively capturing spatial heterogeneity and nonlinearity. Further analysis reveals significant variations in the impact of spatial weighting on the contributions of different feature variables: the explanatory power of most variables increases with the inclusion of spatial effects, whereas the contributions of variables such as population density, road network density, and parcel compactness are reduced. The SHAP framework is employed to provide interpretable analysis of the patterns in the impacts of various feature variables on air pollution. This study emphasizes the importance of spatial effects at the microscale plot level in understanding the mechanisms of air pollution and highlights their importance for developing urban environmental policies, thus promoting the transformation of air pollution control towards spatially refined management.

Air pollutionBuilt environmentGeographically weighted machine learningSpatial nonstationaritySHapley additive Explanations methodYANGTZE-RIVER DELTASOCIOECONOMIC-FACTORSPM2.5 CONCENTRATIONSCROSS-VALIDATIONCANYON GEOMETRYURBAN FORMLAND-USEQUALITYDISPERSIONGUIDELINES

Wang, Shumin、Hu, Mingxing、Li, Jianyu、Liu, Guoao、Hu, Wanqi、Qi, Junheng、Huang, Jingyue、Liu, Ziye、Wang, Hui、Han, Bing

展开 >

Southeast University School of Architecture

Univ Penn

Nanjing Forestry Univ

Jinling Inst Technol

展开 >

2025

10.1016/j.buildenv.2025.113150

Building and environment

Building and environment

SCI
ISSN：0360-1323
年,卷(期)：2025.280(Jul.)
  • 97