Ensemble Machine Learning for Vertical Profiling of Particulate Mass Concentration Using LiDAR and Ground-Based Observations in a Coastal City of Southeast China

Fine particulate matter (PM_2.5) adversely affects human health and climate, yet its vertical distribution—critical for understanding transport mechanisms, radiative forcing, and altitude-resolved exposure—remains poorly characterized due to the spatial limitations of ground-based monitoring. This study develops an ensemble machine learning framework to estimate vertical profiles of PM_2.5 mass concentration and chemical composition by integrating Light Detection and Ranging (LiDAR)-derived aerosol extinction coefficients with ground-based speciation data from Xiamen, a coastal urban location in southeastern China. A Mie scattering LiDAR system provided high-resolution vertical profiles of aerosol optical properties, while co-located measurement supplied hourly concentrations of PM_2.5, sulfate (SO₄²⁻), nitrate (NO₃⁻), ammonium (NH₄⁺), organic carbon (OC), and elemental carbon (EC). Meteorological variables from surface stations and ERA5 reanalysis were incorporated as predictors. We systematically compared linear regression, support vector regression (SVR), artificial neural networks (ANN), XGBoost, and an ensemble model combining all three ML algorithms. The ensemble model outperformed traditional approaches, achieving a slope of 0.64 and bin-MAE of 5.56 µg m⁻³ for total PM_2.5 mass on an independent test dataset, with SHAP analysis identifying the extinction coefficient as the dominant predictor. Reconstructed vertical profiles revealed distinct seasonal patterns: maximum concentrations near the surface in spring and winter (18.1 ± 8.09 µg m⁻³ at 0.4 km), suggesting the combined influence of low-level regional transport and local emissions, broadly consistent with trajectory clustering results showing predominant transport from northern regions. In contrast, elevated maxima above 1000 m in summer suggest enhanced marine-influenced regional transport aloft.