- 1Key Laboratory of Environmental Optical and Technology, Anhui Institute of Optics and Fine Mechanics, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China(phxie@aiofm.ac.cn)
- 2Information Materials and Intelligent Sensing Laboratory of Anhui Province, Institutes of Physical Science and Information Technology, Anhui University, Hefei, China(xtian@ahu.edu.cn)
Retrieving the vertical distribution of tropospheric Ozone (O3) based on ground-based Multi-AXis Differential Optical Absorption Spectroscopy (MAX-DOAS) observations presents challenges due to the interference of stratospheric O3 absorption. A Convolutional Neural Networks (CNN) method is proposed for retrieving the vertical distribution of tropospheric O3 based on ground-based MAX-DOAS observations. This method circumvents the issue of stratospheric O3 absorption interference when obtaining tropospheric O3 profiles by using CNN to extract features from MAX-DOAS spectral segments, enabling the retrieval of tropospheric O3 profiles. The core optimizations of this method are reflected in the following three aspects: (1) Enhancement of MAX-DOAS spectral features and establishment of a dataset with multiple features. To improve the feature extraction capability of the CNN model, mathematical methods are employed to enhance the features of the 320-340 nm spectral segments, which exhibit strong absorption characteristics for O₃. Additional datasets of various sensitive factors are incorporated to improve model inversion accuracy. The Z-Score normalization method is applied to unify dimensions and expedite model convergence, addressing inversion errors resulting from disparate dataset dimensions; (2) Constructing a PCA-F_Regression-SVR hybrid model to screen the optimal ancillary dataset for modeling. Principal Component Analysis (PCA) is utilized to reduce the dimensionality of all sensitive factors. A combination of Support Vector Regression (SVR) and the F_Regression function comprehensively evaluates and screens features sensitive to the tropospheric O₃ profiles retrieval. These features include profiles of temperature, specific humidity, fraction of cloud cover, eastward and northward winds, SO₂, NO₂, HCHO, as well as seasonal and temporal features; (3) The CNN inversion model is developed to extract the enhanced features from MAX-DOAS spectral segments and sensitive factors, enabling the retrieval of tropospheric O3 profiles. Aiming to minimize the loss function of the Mean Absolute Percentage Error (MAPE), the hyperparameters of the CNN inversion model are determined through cross-validation. The enhanced MAX-DOAS spectral features, along with sensitive factors, are used as the model inputs. The EAC4-CNEMC hybrid O3 profiles serve as the model outputs, resulting in a decrease in MAPE from 26% to 19%. The CNN inversion model is applied to independently retrieve tropospheric O3 profiles, and effectively reproduced the O3 profiles of the EAC4 dataset, exhibiting a Gaussian-like vertical distribution with peaks mainly around 950 hPa, and Absolute Percentage Errors (APEs) are generally controlled below 20%. In conclusion, leveraging MAX-DOAS spectra enables the retrieval of tropospheric O3 vertical distribution through the established CNN inversion model.
How to cite: Wang, Z., Tian, X., Xie, P., and Xu, J.: A Convolutional Neural Networks Method for Tropospheric Ozone Vertical Distribution Retrieval from Multi-AXis Differential Optical Absorption Spectroscopy Measurements, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3930, https://doi.org/10.5194/egusphere-egu25-3930, 2025.
