A convolutional neural networks method for tropospheric ozone vertical distribution retrieval from Multi-AXis...

Area of Science:

• Atmospheric Chemistry and Physics
• Remote Sensing
• Machine Learning Applications in Environmental Science

Background:

• Tropospheric ozone (O3) vertical distribution is critical for atmospheric processes.
• Multi-AXis Differential Optical Absorption Spectroscopy (MAX-DOAS) measurements face challenges with stratospheric O3 interference for tropospheric profiling.
• Accurate O3 profiles are essential for air quality monitoring and climate studies.

Purpose of the Study:

• To develop and validate a Convolutional Neural Network (CNN) model for retrieving tropospheric O3 vertical distribution from MAX-DOAS measurements.
• To address and overcome the interference from stratospheric O3 absorption in MAX-DOAS data.
• To enhance the accuracy and reliability of tropospheric O3 profile retrievals.

Main Methods:

• A hybrid feature selection model (PCA-F_Regression-SVR) was employed to identify sensitive factors for O3 inversion, including meteorological and trace gas profiles.
• A CNN model was constructed using preprocessed MAX-DOAS spectra and selected sensitive factors as input, with reanalysis O3 profiles as output.
• The model's performance was evaluated using Mean Absolute Percentage Error (MAPE) and compared against independent datasets.

Main Results:

• The CNN model successfully reproduced tropospheric O3 profiles, with Mean Absolute Percentage Error (MAPE) decreasing from 26% to approximately 19%.
• The retrieved O3 profiles exhibited a Gaussian-like distribution, peaking around 950 hPa (550 m).
• The model showed a tendency to overestimate surface O3 in summer due to temperature influences, while being insensitive to extreme values in smoothed reanalysis data.

Conclusions:

• The developed CNN model effectively leverages MAX-DOAS spectra for accurate tropospheric O3 vertical distribution retrieval.
• The method successfully mitigates stratospheric O3 interference, offering a more reliable approach to O3 profiling.
• This technique provides valuable data for atmospheric research, air quality management, and climate modeling.