Revealing the built environment impacts on truck emissions using interpretable machine learning

Understanding the factors influencing truck emissions remains critical for sustainable urban freight transport development. However, ignoring spatiotemporal and policy heterogeneity may lead to inaccurate predictions for specific regions and misinterpretation of outcomes. This study develops a comprehensive framework to analyze the nonlinear effects of the built environment on heavy-duty diesel truck emissions, utilizing large-scale GPS data from Shanghai, China. We introduce an interpretable predictive model that integrates random effects with a light gradient boosting machine to account for spatiotemporal and policy influences. The results show that proposed model outperforms baseline by 15 %–20 %, with an improvement exceeding 17 % in the more complex tasks of localized predictions in central urban areas. Land use and road design factors contribute 72.26 % to truck emissions, with industrial land density as the primary driver. Furthermore, the relationship between these factors and pollution emissions exhibits pronounced non-linearity, with threshold effects that vary under various policy restrictions.