Estimation of tropospheric delay in GNSS observations using support vector machines algorithm

The tropospheric delay occurs not moving linear of GNSS signals as a result of interaction with the gas masses in the troposphere and so, the signals do not reach the receiver in the expected time. In this respect, the accessibility and precise modeling of tropospheric delay play an important role in meteorological studies and weather forecasts, as well as in GNSS positioning applications. Today, while determining the zenith tropospheric delay effect with the GNSS technique, troposphere models are used, which are insensitive to sudden air changes, include parameters based on standardized and experimental methods. These models determine the delay effect at other points (station-wise) with interpolating methods or experimental equations by referencing GNSS stations containing the meteorological sensors. Therefore, with these models, it is not possible to precisely determine or reach the delay effect of every time during the day. Besides, if a GNSS station cannot be accessed due to technical or hardware problems, the quality of the delay forecast deteriorates due to gaps in the data archive. In line with this purpose, new alternative approaches are required to determining the tropospheric delay effect due to sudden weather changes instantly, continuously and accurately, making the reference stations independent from each other in predictions or preventing technical/hardware problems. With the inclusion of machine learning algorithms, which are widely used in science and engineering, into GNSS technology, new approaches to current problems, explanation and interpretation of relationships, and inferences for new cases have been provided. In this study, it is aimed to estimate the zenith tropospheric delay effect from machine learning models created with actual meteorological and GNSS observation data. Support Vector Regression (SVR), which is a sub-form of Support Vector Machines (SVM) in learning models, was used as a method. In application, when creating SVR models, a data set using actual ZTD values and meteorological parameters is needed. Meteorological parameters were obtained for the years 2019-2020 at the "GOPE" station selected as the study area in the IGS/EPN network. Actual ZTD data were also taken from the VMF1 model, which was determined by observations in field conditions with 6-hour intervals (for the years 2019-2020). Afterward, all the collected data were arranged and the model forecasts were realized by using Linear (Linear-SVR), Polynomial (Polynomial-SVR) and Radial-Based (RBF-SVR) kernel functions belong to different mathematics in the structure of SVR. According to the estimation results, the highest success was calculated from the RBF-SVR model with an R2 value of 0,84. The RBF-SVR model was chosen as the most appropriate prediction model in the study because it performs better than other SVR models. At the last stage of the study, the ZTD values forecasted from the RBF-SVR model have compared with the tropospheric products published from the IGS/EPN network and the ZTD values forecasted from the CSRS-PPP within the online-PPP services. In addition, the LOF (Local Outlier Factor) technique, one of the multivariate outlier observation analyzes, was used to improve the RBF-SVR model created using all data and to see how the removal of outlier data between samples groups affected the model result. According to the LOF, the data detected as outlier were removed from the data set and the RTF-DVR (LOF) model was established. Performance analyzes of the newly created models were examined. According to the analysis results, root mean square error differences between the ZTD values obtained from the RBF-SVR (LOF) and RBF-SVR models with the actual ZTD values (VMF1) were found to be ± 1,69 cm and ± 2,08 cm, respectively. When compared with other GNSS services, IGS/EPN and CSRS-PPP, the forecast results were determined to be very close to each other (~0,6 cm). It has been seen that alternative approaches to GNSS and troposphere topics can be presented with the realized application. It shows that machine learning can play an important role in increasing the efficiency of new solutions or existing solutions for problems in GNSS applications.