Explainable Transformer-Based Estimation of Surface NO2 Concentrations Across Fujian Province Using TROPOMI Observations

Abstract

Atmospheric nitrogen dioxide (NO2) is one of the major pollutants affecting environmental quality and human health, yet accurate surface NO2 (NO2Surface) estimation remains challenging due to complex nonlinear relationships between satellite-derived NO2 column density (NO2column) and near-surface concentrations. Here, we proposed an explainable transformer-based framework for daily NO2Surface estimation across Fujian Province, integrating Sentinel-5P TROPOspheric Monitoring Instrument (TROPOMI) NO2 observations with meteorological data, digital elevation model (DEM), normalized difference vegetation index (NDVI), nighttime light (NTL), and population density (PD). Using ground-based observations, Tabular Prior-data Fitted Network (TabPFN) was trained and benchmarked against five conventional machine learning (ML) algorithms. Under random five-fold cross-validation, TabPFN achieved the highest accuracy, yielding a coefficient of determination (R2) of 0.93, root mean square error (RMSE) of 2.56 μg m−3, mean absolute error (MAE) of 1.05 μg m−3, and mean absolute percentage error (MAPE) of 5.82%, with R2 improvements of 0.17–0.26 and RMSE reductions of 45.3–53.5% relative to CatBoost and RF, respectively. Additional station-grouped spatial validation and independent temporal validation yielded R2 values of 0.86 and 0.79, respectively. The reconstructed NO2Surface distributions revealed a clear coastal-to-inland gradient, with concentrations frequently exceeding 50 μg m−3 in urbanized coastal areas but generally below 15 μg m−3 in northwestern mountainous regions. Strong seasonality was evident, with the peak NO2Surface and widespread pollution in winter, contrasting with lower and more localized levels in summer. Temporal analyses revealed persistent declines in NO2Surface, with annual mean concentrations decreasing by 11.7% from 2019 to 2020 and a further decline of 1.6% in 2021, while descriptive daily slopes were approximately −0.003 μg m−3 day−1. SHapley Additive exPlanations (SHAP) analysis highlighted TROPOMI NO2column and BLH-corrected NO2 (NO2BLH) as the most influential predictors, accounting for 15.07% and 11.36% of the contribution, respectively, with temperature- and humidity-related variables exhibiting strong nonlinear responses. This study provides an accurate and interpretable framework for 1 km gridded NO2Surface estimation and characterizes the predictive relationships associated with regional NO2 variability.