Machine Learning Based Tm Refinement Modeling And Its Study On PWV Retrieval

Precipitable water vapor(PWV)is an important parameter for measuring atmospheric water vapor content,and it is crucial for numerical weather prediction and climate research.Atmospheric weighted mean temperature(T_m)is a key parameter for converting GNSS zenith wet delay(ZWD)to PWV,and its precision can improve the conversion efficiency of ZWD to PWV.The current mainstream T_m models still have shortcomings,such as a general lack of universality,significant local accuracy loss,and inability to reflect the nonlinear relationship between T_m and meteorological factors and spatiotemporal information.This paper uses machine learning algorithms to construct a T_m combined model based on machine learning methods in China and a T_m combined model driven by unmeasured meteorological data in China.The main research content and work of this paper are as follows:1.The accuracy of the GPT3 model at two resolution versions for temperature,pressure,and T_m was evaluated by data from 46 radiosonde stations in China from 2017 to 2018.The results indicate that the GPT3-1 model performs better than the GPT3-5 model,and its accuracy in the southern region is higher than in the northern and western areas.2.Based on 150 radiosonde station data from 2007 to 2016(covering China and some surrounding countries and regions)and the ability of three machine learning methods(backpropagation neural network,random forest,generalized regression neural network)to deal with nonlinear fitting problems,constructing three new Tm combined models with higher accuracy and fewer parameters in China,namely,combined model B(HM-B),combined model R(HM-R)and combined model G(HM-G).The results show that:(1)taking the Tm data of radiosonde stations as the reference value,the RMSE of HM-B,HM-R,and HM-G is 2.95 K,2.70 K,and 2.76 K,which are much lower than the RMSE of Bevis,GPT3,and HGPT.(2)compared with Bevis,GPT3,and HGPT models,HM-B,HM-R,and HM-G models can maintain better accuracy in different latitudes,heights,and seasons and achieve more even accuracy in space and time.In addition,the memory occupied by the three combined models is significantly lower than that of the GPT3 model,and the number of parameters is reduced considerably.3.A Tm combined model(E-HM-R model)that does not require measured meteorological parameters is constructed by using ERA5 reanalysis data meteorological data to replace measured data to solve the problem of calculating T_m in some areas where it is challenging to obtain measured meteorological data.The results show that:(1)taking the air temperature data of radiosonde stations as the reference value,the average bias of ERA5 reanalysis data is 0.25K,and the average RMSE is 2.30 K.Overall,the temperature data of ERA5 reanalysis data have high accuracy.(2)taking the T_m data of radiosonde stations as the reference value,the RMSE of the E-HM-R model is 2.72 K,which is 38.7%,38.8%,and 37.6%lower than Bevis,GPT3,and HGPT.(3)the E-HM-R model can maintain better and more uniform accuracy than the three comparison models in space and time.4.The HM-R model,with the best precision among the newly built T_m combined model,is used for water vapor retrieval,and the retrieval accuracy is compared with that of Bevis,GPT3,and HGPT models.The results show that:(1)the PWV retrieval of the HM-R model on the four radiosonde stations with uniform distribution in the study area is closer to the reference PWV,and the retrieval accuracy is higher than that of Bevis,GPT3,and HGPT models,and can meet the accuracy requirements of GNSS meteorology.(2)the RMSE of PWV retrieval by the HM-R model on almost all sites in the Chinese mainland region is lower than that of the comparative model.At the same time,the Z-test hypothesis test results further verify that the improvement of the HM-R model is significant.