Quantitative reconstruction of long-term spatiotemporal patterns of high-resolution ground-level NO2 concentrations in mainland China using fusion techniques and a machine learning framework

Nitrogen dioxide (NO₂), as a critical trace gas, plays multiple roles in the atmosphere and poses potential threats to human health. However, existing satellite monitoring methods face challenges, including limited satellite mission durations, poor data quality, and low spatial resolution, which hinder the ability to provide long-term, high-precision NO₂ information. To address these issues, this study uses the TROPOMI tropospheric NO₂ column concentration product as a baseline and employs partition and cumulative distribution function (CDF) techniques to generate a satellite fusion dataset with both long time spans and high consistency. Based on this dataset, a high-performance, high-spatial-resolution long-term surface NO₂ estimation model was developed using machine learning algorithms combined with multi-source geographic data. The model successfully estimates daily average near-surface NO₂ concentrations (1 km² resolution) for mainland China from 2014 to 2020. The results show that the proposed fusion method effectively integrates OMI and TROPOMI data, improves the spatial correlation between satellite products by 16.2 % (R = 0.74 → 0.86), significantly enhances the spatial coverage, and thus more accurately characterizes the spatial distribution characteristics of NO₂. The surface-level NO₂ estimates based on the LGBM model achieved an R² of 0.85 in ten-fold cross-validation, with corresponding root mean square error (RMSE) and mean absolute error (MAE) of 7.51 µg/m³ and 5.22 µg/m³, respectively, demonstrating good extrapolation ability for temporal variations. The long-time series results accurately reflect the temporal and spatial evolution of NO₂ in mainland China, while the high-precision estimates provide detailed pollution exposure information, revealing urban-scale pollution differences and seasonal variations.