Heterogeneous bus ridership responses to micro-scale built environment: Evidence from explainable AI

Despite the inherent complexity of urban mobility, conventional methodologies often oversimplify the associations between transit demand and built environment, leaving nonlinearity and behavioral asymmetries obscured. This study integrates computer vision and explainable artificial intelligence (XGBoost-SHAP) to decode how the micro-scale built environment affects multidimensional bus ridership. Six environmental visual proxies were extracted through semantic segmentation from 3,736 street view imagery, and subsequently analyzed alongside 8.6 million ridership records over one month across 934 bus stops in Shenzhen, China. Four models distinguish temporal-behavioral patterns (weekday peak vs. weekend off-peak, boarding vs. alighting). The results reveal behavioral heterogeneity: stronger correlations were observed in boarding than alighting demand, while weekend discretionary trips exhibit greater sensitivity to micro-environmental quality than weekday mandatory trips. By quantifying nonlinear threshold effects and diminishing marginal utilities, this study develops a spatial screening matrix that supports planners in flagging environmental mismatches and advancing transit-demand-oriented, data-driven transportation planning.