Estimation of daily NO2 with explainable machine learning model in China, 2007-2020

Surface nitrogen dioxide (NO₂) is an effective indicator of anthropogenic combustion and is associated with regional burden of disease. Though satellite-borne column NO₂ is widely used to acquire surface concentration through the integration of sophisticated models, long-term and full-coverage estimation is hindered by the incomplete retrieval of satellite data. Moreover, the mechanical relationship between surface and tropospheric NO₂ is often ignored in the context of machine learning (ML) approach. Here we develop a gap-filling method to obtain full-coverage column NO₂ by fusing satellite data from different sources. The surface NO₂ is then estimated during 2007-2020 in China using the XGBoost model, with daily out-of-sample cross-validation (CV) R² of 0.75 and root-mean-square error (RMSE) of 9.11 µg/m³. The back-extrapolation performance is verified through by-year CV (daily R² = 0.60 and RMSE = 11.46 µg/m³) and external estimations in Taiwan before 2013 (daily R² = 0.69 and RMSE = 8.59 µg/m³). We explore the variable impacts in three hotspots of eastern China through SHAP (Shapley additive explanation) values. We find the driving contributions of column NO₂ to the variation of ground pollution during 2007-2020 (average SHAP = 5.09 µg/m³ compared with the baseline concentration of 33.39 µg/m³). The estimated effect is also compared with ordinary least squares (OLS) model to provide a straightforward understanding. The related health burden is further calculated by using the annual NO₂. We demonstrate the employment of explainable ML model is beneficial for comprehend the coupled relationship in surface NO₂ change.