ارزيابي كارائي مدل‌هاي يادگيري ماشين در مدل‌سازي ژئوئيد منطقه اي با استفاده از اندازه‌گيري‌هاي GPS/Leveling

The geoid is one of the equipotential surfaces of the Earthʹs gravity field, which approximates the mean sea level (MSL) in the best possible way (according to the theory of the least squares method). This level is known as the height base (elevation datum - physical form) of the earth an‎d is used in civil an‎d construction projects to determine heights. Therefore, accurate modeling of this level is very important in surveying an‎d geodesy. There are different methods for geoid modeling. These methods include: geometrical method, geoid determination by satellite method, gravimetric methods an‎d geoid determination using GPS/leveling. Each of these methods has advantages an‎d disadvantages. However, nowadays, due to the expansion of GNSS networks an‎d also accurate leveling in the stations of these networks, the use of GPS/Leveling method has gained more acceptances. In this method, having the orthometric an‎d normal heights, the height of the geoid can be calculated. In recent years, machine learning methods have been used for local geoid height modeling. The main advantage of machine learning methods is simple an‎d fast calculations, as well as high accuracy an‎d precision in the results. Therefore, in this thesis, the generalized regression neural network (GRNN), adaptive neuro-fuzzy inference system (ANFIS) an‎d support vector regression (SVR) models have been used to estimate the local geoid height in the Iranian region. In order to eva‎luate the proposed models, 31 an‎d 26 GPS/Leveling stations have been used in the Iranian plateau in central Alborz network an‎d northwest network in three different scenarios. In these three scenarios, considering 2, 4 an‎d 7 test stations, the accuracy of the models have been investigated. Also, the results of the machine learning models in the test stations have been compared with the IRG2016 an‎d EGM2008 models. In the first scenario, where the number of training stations is more than the other two scenarios, the RMSE of ANFIS, SVR, GRNN, IRG2016 an‎d EGM2008 models in the central Alborz network is 27/35, 63/69, 62/35, 25/26, an‎d 63.20 cm, respectively. For the northwest network, the RMSE is 3.94, 83.27, 93.11, 23.05 an‎d 45.65 cm, respectively. In this scenario an‎d in both studied networks, the error of the ANFIS model is lower than the other two models. However, in scenarios two an‎d three, the error of the ANFIS model increased, but the error of the SVR an‎d GRNN models decreased. This shows that the SVR an‎d GRNN models provide higher accuracy with less input data. The averaged RMSE of IRG2016 an‎d EGM2008 models in two studied networks is 26.15 an‎d 54.42 cm, respectively. Therefore, the local model of IRG2016 has been more accurate than the global model. But the comparison of the IRG2016 model with the ANFIS model presented in this thesis shows that the new model can provide higher accuracy in estimating the spatial varations of the geoid height if there are more GPS/Leveling station observations with a more uniform distribution.

6

149067