A Deep Learning Retrieval for Tropospheric Ozone Profiles from High-Resolution Satellite Data

Tropospheric ozone is an important atmospheric trace gas that affects air quality, human health, and climate. However, its accurate retrieval from satellite observations remains challenging while in situ vertical profile measurements are sparse. The retrieval of tropospheric ozone is impeded by its weak signal compared to the dominant stratospheric ozone column, and current full physics retrievals often suffer from limited vertical resolution and show insufficient agreement to in situ observations. Recent advances in satellite instrumentation and machine learning provide an opportunity to overcome these limitations. In particular, the Tropospheric Monitoring Instrument (TROPOMI) offers high spatial resolution and signal-to-noise ratio, enabling more detailed daily observations of atmospheric ozone variability with global coverage.

We explore the feasibility of a deep-learning-based technique for retrieving high-resolution tropospheric ozone profiles from TROPOMI spectral measurements combined with auxiliary meteorological information, while avoiding some of the simplifying assumptions made in existing full physics retrieval approaches. Artificial neural networks are well-suited for this task, as they can learn complex, nonlinear relationships between ozone absorption features, surface and cloud properties, observation geometry, and atmospheric state variables.

The proposed methodology integrates TROPOMI spectral radiance/irradiance data (L1B) and the satellite position with meteorological information from ERA5 reanalysis data. The meteorological data includes boundary layer height and dissipation and surface pressure alongside temperature, humidity and wind speed profiles. The ground truth for the supervised training is comprised of co-located ozone profile measurements from ozone sondes (TOAR), aircraft measurements (IAGOS), lidar observations (TOLNet), and satellite microwave limb sounder (MLS).

Initial retrieval model is based on feed-forward fully connected neural network (multilayer perceptron), with planned extensions to convolutional architectures and dimensionality-reduction techniques. Using data from 2021 alone (approximately 1.2×10⁴ independent ozone profiles corresponding to ~1.5×10⁷ concentration measurements) our preliminary results demonstrate strong performance. The model’s ozone profile predictions are evaluated against the ground truth observations on an independent test set (including unseen time periods and locations). In this preliminary evaluation, the model achieves a coefficient of determination (R²) of 0.841 between retrieved and observed ozone concentrations, indicating the model’s ability to capture both vertical ozone structure and spatial ozone variability.