A CNN - SVR model for NO2 profile Prediction based on MAX-DOAS observation

Xin Tian,Yifeng Pan,Pinhua Xie,Jin Xu,Ang Li,Zijie Wang,Zhaokun Hu,Jiangyi Zheng
DOI: https://doi.org/10.5194/egusphere-egu23-6091
2023-01-01
Abstract:In the Multi-Axis Differential Optical Absorption Spectroscopy (MAX-DOAS) trace gases profile retrieval, it needs to obtain the vertical profile of aerosols as the a priori, and depends on the atmospheric radiative transfer model (RTM). Therefore, a data mining method named CNN-SVR was adopted to achieve the prediction of NO2 profile, which combines the advantages of convolutional neural network (CNN), support vector regression (SVR) and MAX-DOAS. The optimization core of the hybrid model is embodied in three aspects. (1) CNN extracting the effective features of MAX-DOAS spectral data. The input data are MAX-DOAS spectrum, wind direction, wind speed, season, temperature, relative humidity, aerosol optical depth (AOD), cloud cover. Feature variables of MAX-DOAS spectra were extracted by CNN. The output data set is the NO2 profile retrieved by MAX-DOAS profile inversion algorithm PriAM. The data set is processed by normalization to unify the dimensions to ensure the accelerated convergence of the program. (2) The mean impact value (MIV) method selecting the input variables sensitive to NO2 profile forecasting. The MAX-DOAS spectral data, temperature, AOD and low cloud cover are finally determined as the best input parameters of the prediction model. (3) The hybrid forecasting method. Combined with the advantages that CNN can reduce the amount of data processing and retain useful information, and SVR does not depend on the dimension of input space, a CNN-SVR hybrid prediction model is proposed. The average percentage error (MAPE) of the CNN-SVR model is 9.14%. Compared with the separately constructed CNN, SVR and backpropagation models, the MAPE of CNN-SVR is reduced by 8.22%, 6.00% and 32.28% respectively. Therefore, CNN-SVR can effectively predict tropospheric NO2 profiles by using MAX-DOAS observation.
What problem does this paper attempt to address?