Enhancing Satellite-Based NO2 Monitoring with Machine Learning: From Near Surface Concentration Estimation to A-Priori Profile Development

Satellite observation plays an important role in air quality monitoring. Nitrogen dioxide (NO₂) is an important atmospheric trace gas that significantly impacts air quality, public health, and ecosystems. While satellite NO₂ observations have been widely used in the application of machine learning (ML) to estimate surface NO₂ distributions, this surface NO₂ modeling deserves further investigation, including examining satellite’s contributions to the estimation of NO₂ at different concentration levels. In addition, as satellite data evolve towards higher spatiotemporal resolution, the demand for high-resolution a-priori NO₂ profiles is increasing. Generating such profiles using traditional numerical models is computationally expensive, but ML offers an efficient solution to address this challenge. This presentation illustrates two developments based on ML technology: one focusing on the application of satellite observations to estimate surface NO₂ distributions, and the other on the development of high-resolution a-priori NO₂ profiles for satellite retrievals. Western Europe is used as the study area.

The first part addresses estimating high-resolution surface NO₂ concentrations (1 km, daily) using the Boosting Ensemble Conformal Quantile Estimator (BEnCQE). This model integrates diverse datasets, including TROPOMI NO₂ tropospheric vertical column densities (TVCDs), and demonstrates reliable performance validated against European Environmental Agency (EEA) surface observations (r = 0.80, R² = 0.64, RMSE = 8.08 µg/m³). Quantile regression in BEnCQE provides uncertainty estimates and feature importance analysis across different NO₂ levels. Results show that satellite observations significantly contribute to background NO₂ predictions but have less influence on high-concentration estimates, likely due to the relatively coarse spatial resolution of current satellite data. These findings highlight the need for higher-resolution satellite missions, such as CO2M (2 km resolution), to better capture localized pollution.

The second part focuses on generating high spatial resolution a-priori NO₂ profiles for satellite retrievals. We developed Deep Atmospheric Chemistry NO₂ (DACNO₂), a convolutional neural network framework, to produce 3D NO₂ distributions (8 levels from the surface to 5,000 m, 2 km spatial resolution, daily). Using a multi-constraint training approach that combines coarse-resolution CAMS-EU synthetic NO₂ data (10 km) and fine-scale EEA surface observations (2 km), DACNO₂ captures detailed spatial gradients near emission hotspots while maintaining broad physical consistency. The evaluation shows good performance aligned with EEA observations (r = 0.81, R² = 0.64, RMSE = 5.10 µg/m³) and CAMS-EU synthetic NO₂ (r = 0.94, R² = 0.89, RMSE = 1.11 µg/m³). The implementation of DACNO₂ is efficient, taking only minutes to compute one day's result using GPU acceleration.

Overall, this presentation introduces ML-based works on application and development aspects for satellite NO₂ observations to advance the coupling of ML technology and satellite observations of pollution.