Estimation of precipitable water vapor using least squares support vector regression and comparison with other models

The LS-SVR model uses simple linear equations in the training phase. As a result, the complexity of the computational algorithm is reduced; the speed of convergence and the accuracy of the results are increased. Seven parameters of longitude and latitude of GPS station, day of year (DOY), time to universal time (UT), relative humidity (RH), temperature (T) and pressure (P) are considered as inputs of LS-SVR model. And the PWV corresponding to these seven parameters is the output of the model. After the training step, the PWV value was estimated with the trained model and compared with the PWV values obtained from the radiosonde station, the empirical model of Saastamoinen and GPT3, the support vector regression model (SVR) and the radial basis neural network model (RBNN) in the control stations. Statistical indices of relative error, correlation coefficient and root mean square error (RMSE) have been used to evaluate the accuracy of the models. The conducted analyzes show that the average RMSE of RBNN, SVR, LS-SVR, GPT3 and Saastamoinen models in 3 control stations is to 4.92, 4.13, 2.87, 4.22 and 4.29 mm, respectively. Also, the average relative error of the models in 3 control stations is calculated as 38.06, 30.77, 22.37, 34.63 and 32.80% respectively. Analysis of the PPP method shows an improvement of 33 mm in the coordinate components using the LS-SVR model. The results of this thesis show that the LS-SVR model can be considered as an alternative to the empirical troposphere models in the studied area. The LS-SVR model is a local troposphere model with high accuracy.