Link-Level Mapping of Fugitive Road-Dust Emissions in Urban China Using Explainable Machine Learning and Open Traffic Data

Fugitive road dust (FRD) is a major contributor to urban particulate matter (PM) in Chinese cities, accounting for an estimated 25%–90%^1,2 of total PM emissions and creating substantial air-quality and health burdens. Despite this relevance, policy and research have focused primarily on exhaust emissions, while FRD remains comparatively under-characterized and weakly regulated³. Importantly, non-exhaust PM from FRD persists under fleet electrification; recent evidence indicates that the monetized impacts of PM associated with battery electric vehicles can be comparable to, or higher than, those from internal combustion engine vehicles due to continued non-exhaust sources⁴.

The U.S. EPA AP-42 approach is widely used to estimate road dust resuspension and is also embedded in China’s technical guidance (HJ/T 393-2007). Within this framework, the silt load (sL, mass of particles <75 μm per meter squre, g.m^-2) is the critical input governing emission intensity. However, conventional sL sampling (e.g., gravimetric sampling or mobile vacuum-based surveys) is labor-intensive and difficult to scale to national wide inventories with representative spatial coverage.

Here, we compile a national database of in situ sL measurements from 28 Chinese cities (>300 roads) and develop interpretable machine-learning models to predict sL by road class and city context. We fuse these predictions with open-source traffic data (200 cities; 20-min resolution; 8 months of records) and apply the AP-42 framework to construct a link-level FRD PM_2.5 emission inventory for urban China. Multiple algorithms (XGBoost, support vector regression, and ensembles) are evaluated, and SHAP (SHapley Additive exPlanations) is used to quantify feature contributions and diagnose non-linear effects.

The best-performing models achieve strong generalization (test R² > 0.7). SHAP results identify road class, precipitation, ambient PM₁₀ concentration, cleaning-vehicle density, longitude, traffic volume, and heavy-duty vehicle share as key drivers of sL, with pronounced non-linear decrease in sL as vehicle speed rises. FRD emission’s contribution to the traffic PM_2.5 emission were estimated in city-level, range from 25% to ~80%. Overall, this work first to infer silt load nationally using ML and translate it into a link-level inventory using open traffic data, provides a scalable pathway to high-resolution FRD emission estimation and supports targeted mitigation and urban transport planning.

• 1.Wang, L. et al.Environmental challenges in electrification: Traffic-induced non-exhaust PM2.5 emissions in Cangzhou, China. Transp. Res. Part Transp. Environ.151, 105137 (2026).
• 2.Chen, S. et al.Fugitive Road Dust PM2.5 Emissions and Their Potential Health Impacts. Environ. Sci. Technol.53, 8455–8465 (2019).
• 3.Harrison, R. M. et al.Non-exhaust vehicle emissions of particulate matter and VOC from road traffic: A review. Atmos. Environ.262, 118592 (2021).
• 4.Liu, Y. et al.Exhaust and non-exhaust emissions from conventional and electric vehicles: A comparison of monetary impact values. J. Clean. Prod.331, 129965 (2022).