城市建成环境与跑步活动的非线性和空间异质性关系——基于可解释机器学习方法OA

This study aims to address limitations in existing healthy city research,such as the failure to comprehensively measure environmental features(including spatial quality)and the reliance on assumptions of linear relations.It intends to reveal nonlinear and spatially heterogeneous relations between built environmental attributes and running activities.The study area covers the main urban area of Fuzhou.Multisource data,including building profile data,road network data,points of interest,and street-view images,were used.These data were analyzed using spatial statistics and deep learning algorithms,such as semantic segmentation,object detection,and image regression,to develop a comprehensive framework for evaluating urban built environment attributes.This framework incorporates two-and three-dimensional environmental elements as well as spatial quality characteristics.We then employed the Extreme Gradient Boosting algorithm and SHAP explainers to summarize the types of nonlinear relations between built environment features and residents' running activities.Combined with K-means clustering analysis,we classified street types according to local and spatial heterogeneity in the built environment's impact on running activity.The results indicate that(1)there are limited differences in the influence of two-dimensional and three-dimensional physical environmental features,environmental subjective perception,and physical disorder of the built environment on running activity.This finding suggests that analyses confined to conventional physical environmental features alone are inadequate for examining environmental effects on health behavior.Notably,factors,such as building density,POI density,POI mixture,greenery visibility,sidewalk visibility,and safety perception,ranked highly in terms of influence intensity.(2)Six types of nonlinear relations emerged between built environment factors and running density.These include:(I-1)a positive relation with an increasing marginal effect(including sky openness and blue visibility);(I-2)positive relation with a decreasing marginal effect(including POI density and street furniture);(I-3)positive relation with a marginal effect that first increases and then decreases(including POI mixture and safety perception);(II-1)negative relation with a decreasing marginal effect(including garbage and distance to parks and green spaces);(II-2)negative relation with a marginal effect that first increases and then decreases(including building density and aesthetic perception);and(Ⅲ)U-shaped relation with an initially negative effect followed by a positive effect(including street aspect ratio and greenery visibility).Different types of nonlinear relations require different environmental optimization strategies.(3)There is also spatial heterogeneity in the influence of the built environment on running activities.Based on this spatial heterogeneity,street segments in the study area were classified into five types:low-frequency running streets driven by motorization-oriented design(17％),high-frequency running streets driven by lush greenery and human scale(10％),medium-frequency running streets driven by safety quality(12％),low-frequency running streets induced by functional diversity and physical disorder(7％),and low-frequency running streets with no significant influencing factors(51％).The dominant factors influencing or inhibiting running differed substantially across street types.Strategies targeting specific regions should be implemented based on the spatial heterogeneity of the environmental factors.