Research On Neural Network Inversion Of Temperature Profile Of FY-4A/GIIRS Based On Spectral Feature Selection

The atmospheric temperature profile plays an important role in understanding the complexity of the atmospheric system.Conventional atmospheric temperature profile data mainly come from sounding observation systems thus the resolution of sounding data in time and space is limited to a certain extent.The hyperspectral infrared detector carried on the satellite has high spectral resolution and wide observation range,and the detected data has more extensive application value.Retrieval of atmospheric temperature profile by hyperspectral infrared sounder can make up for the lack of space-time resolution of profile data obtained by the sounding system.The Geostationary Interferometric Infrared Sounder(GIIRS)carried on China's second-generation stationary meteorological satellite FY-4A is currently the only hyperspectral infrared detector located in a geostationary orbit with a spectral resolution of 0.625cm^-1in the infrared band and 1650 channels.The theory and practice show that the information carried between channels often contains many correlations.Therefore,it is the main research content of this paper to effectively select features from the GIIRS spectrum for temperature profile retrieval.In the study of spectral feature selection methods,this paper proposes two methods.One is based on the Singular Spectrum Analysis(SSA)channel selection method.By decomposing the GIIRS bright temperature spectrum,the spectrum in the temperature absorption band is obtained and a total of 89 temperature channels were selected.The second spectral feature selection method is based on the Autoencoder,and 89 encoding features were obtained by constructing different Autoencoder models.In the research of temperature profile retrieval,the result of two spectral feature selection methods by constructing a neural network model was validated and the retrieval accuracy by constructing three non-spectral feature selection methods was compared.The average retrieval accuracy on the selected 23 pressure layers of based on the SSA channels(89)are about 0.15 K(Bias),1.49 K(MAE),and 2.06 K(RMSE).Based on the Autoencoder features(89)are about 0.35 K(Bias),1.79 K(MAE),and 2.36 K(RMSE).The average retrieval accuracy of non-spectral feature selection models are not more than 2.01 K(Bias),2.33 K(MAE),and 2.96 K(RMSE).The retrieval accuracy of spectral feature selection model from 300 h Pa to 100 h Pa is significantly better than that of the non-spectral feature selection models.