Development Of A Clear Sky Hyperspectral Fast Radiative Transfer Algorithm Based On Principal Component Analysis And Machine Learnin

Forward radiative transfer models（RTM）are essential for atmospheric applications such as remote sensing and weather research models,but as infrared radiation（IR）detection methods continuously upgraded and improved,the requirement of the forward hyperspectral radiative transfer models（HRTM）is also increasing.However,the rigorous radiative transfer simulations are very time-consuming and takes up a lot of storage resources.Currently,most RTMs are focused on the radiation domain while the monochromatic gas absorption are less discussed,which are also taking time because of the obvious change of the gas absorption coefficients.Therefore,based on the information redundancy of multiple domains,a fast and accurate HRTM based on principal component analysis（PCA）and machine learning（i.e.neural network,NN）for clear sky conditions is developed.Geosynchronous Interferometric Infrared Sounder（GIIRS）,the first infrared hyperspectral sounder on FY-4 satellites,is used to verify the accuracy and efficiency of our model.Our model uses either PCA and NN twice for the atmospheric transmittance domain and radiance domain to reduce the number of independent monochromatic simulations and accelerate the calculations.The fast transmittance model in our paper uses the PCA and NN method first in transmittance domain for each gas independently（3 main gases include H₂O,O₃ and CO₂）.Then we use the PCA and NN method secondly in the radiance domain as the traditional models.Meanwhile,a new method is introduced to choose representative variables for the PCA and NN which used the correlation between the influence factors（such as spectrum,temperature,etc.）of each domain（transmittance and radiation）.The information compression based on PCA and NN has improved the efficiency and accuracy of calculations.The validation results on GIIRS show that the average brightness temperature difference（BTD）is less than 0.1K between our model and the standard line-by-line（LBL）model,and a speedup of three orders is obtained.Both methods have their own advantages,but the error and the efficiency are relatively similar.Thus,it can be chosen according to the situations.Moreover,our model not only avoids the extra complicated transmittance model but also is highly flexible for other hyperspectral instruments with similar spectral ranges by only updating corresponding spectral response functions.The model can also give hyperspectral transmittance and radiance datasets,which can be used as separate submodules for other models.