A Fast Machine Learning Method To Compute Atmospheric Level-to-space Transmittance

The infrared band contains abundant atmospheric and surface information and is an important research object in the field of meteorology and remote sensing.Atmospheric transmittance is the core to solve the infrared radiation transfer equation.Also,it is the key to accurately simulate the radiation at the top of the atmosphere.RTTOV(Radiative Transfer for TOVS)is the fast radiative transfer model used in China's GRAPES(Global/Regional Assimilation and Prediction Enhanced System)numerical weather prediction system,which is developed by ECMWF and used to simulate the radiation observed by meteorological satellite sensors.RTTOV can achieve high accuracy while calculating fast.However,in the infrared water vapor band,due to the complex absorption characteristics of wing absorption,continuous absorption,and other gas absorption,the accuracy of RTTOV is relatively low.In essence,RTTOV is a statistic-based fast radiation transfer model.The performance of machine learning is usually better than traditional statistical methods when it comes to nonlinear statistical problems.Therefore,this paper utilizes machine learning methods to establish models for more accurately calculating the level-to-space transmittance of water vapor channels.In this paper,the tropical area within 30 ° latitude is selected as the research area,where the typical clear-sky profiles are chosen from IFS-137(The Integrated Forecast System,137-level-profile)in NWP SAF(Satellite Application Facility for Numerical Weather Prediction).For the three water vapor channels,7.43?m,7.33?m,and 6.52?m,of FY3 A / IRAS(Infra Red Atmospheric Sounder)and the corresponding bands of IASI(Infrared Atmospheric Sounding Interferometer),the channel level-to-space transmittances and brightness temperatures of these profiles at the top of the atmosphere(TOA)are calculated with RTTOV.Transmittances convolved from bands of IASI are set as the ground-true of the transmittances.The profiles and the corresponding transmittances compose the samples for training machine learning models.There are three ensemble learning methods being studied in this paper: gradient boosting tree(GBT),e Xtreme Gradient Boosting(XGBoost),and random forest(RF).Based on these methods,different models for calculating level-to-space transmittances of infrared water vapor channels are built.With transmittances predicted by these models,brightness temperatures(BT)of the TOA are calculated.The mean bias,root mean squared error(RMSE)and standard deviation(std)of the transmittance of each isobar,the mean absolute error(MAE),RMSE and the mean squared logarithmic error(MSLE)of the whole layer's transmittances are used as the evaluation criteria for the transmittance.The MAE and RMSE are the evaluation criteria of the BT.The transmittance and BT predicted by the model are compared to those calculated by RTTOV.The results show that the transmission errors predicted by the three methods are all smaller than the errors of RTTOV.The maximum error of each channel is at the peak of the weight function.In 7.43?m,7.33?m and 6.52?m channels,the MAE of the transmission through the whole layer is about 0.0011 smaller than the MAE of RTTOV.The MSLEs are different among the methods,but they are all smaller than the error of RTTOV.And the largest error reduction is at 6.52?m,the strongest water vapor absorption channel.The errors of GBT and XGBoost differ a bit,while the errors of random forest are relatively larger.In the verification of the BT at TOA,the MAE and RMSE of the three methods are smaller than those calculated by RTTOV.Specifically,the RMSEs of GBT and XGBoost are 0.1152 K,0.0610 K,and 0.1270 K smaller than those of RTTOV at 7.43?m,7.33?m and 6.52?m respectively,while in comparison,the RMSE of random forest is larger.To achieve the similar performance,GBT and XGBoost take the same amount of time,but the fandom forest takes much longer when executing with one thread.Moreover,the BTs predicted by GBT and RTTOV are compared with those of satellite observation.The bias between the BTs predicted by GBT,RTTOV and the observations of satellites,as well as the mean bias and standard deviation,are calculated.The results show that the BTs of the three channels simulated by GBT and RTTOV are generally smaller than the satellite observations.The stronger the water vapor absorption,the smaller the simulated BT.The mean bias and standard deviation of the BTs predicted by GBT is about 0.5K and 0.2K smaller than that of RTTOV,respectively.Taken together,machine learning can improve the accuracy of transmittance calculation of water vapor channel.Among the three ensemble learning methods,the GBT shows the best performance,with the biggest improvement on 6.52?m channel,whose water vapor absorption is strongest.This research shows that the machine learning method has potential for improving the accuracy of forward model in the infrared water vapor band.And it might do some favor for the development of fast radiative transfer model.