Leveraging multi-source observations for vertical PM2.5 estimation: A deep learning fusion model.

Abstract

Vertical information on PM2.5 is important for understanding aerosol evolution and supporting air pollution management, yet most machine-learning studies focus mainly on near-surface PM2.5 and insufficiently incorporate vertically resolved aerosol and meteorological observations. In this study, we developed a multi-source fusion framework based on an AM-CNN-BiLSTM, to estimate PM2.5 concentrations at 200 and 300 m at the Shenzhen Meteorological Gradient Tower from July 2022 to June 2023. The model integrated lidar-derived aerosol extinction coefficient (EC), microwave radiometer (MWR) thermodynamic profiles, meteorological variables, and tower-based PM2.5 observations. To reduce temporal information leakage, model performance is evaluated using a chronological test set and an expanding window with validation strategy. AM-CNN-BiLSTM achieved R2 values of 0.87 and 0.83 at 200 and 300 m, respectively, outperforming the baseline models. SHAP analysis, ablation experiments, and sensitivity analysis consistently identify EC as the dominant constraint on vertical PM2.5 estimation. In the meantime, RH, temperature, and wind variables provide complementary thermodynamic and transport information. The results demonstrate the feasibility of multisource observations for vertical PM2.5 estimation at a coastal urban site. However, the learned EC-PM2.5 relationship and humidity response should be regarded as site-specific and require further multi-year and cross-site validation.