Using tuned LS-SVR to derive normal height from GPS height

This paper presents the application of least squares support vector regression (LS-SVR) with radial basis function (RBF) kernel in deriving orthometric height from GPS heights. First, identical survey points in both height systems of a D-order GPS network are picked up as the data set used in analysis. Second, LS-SVR is performed to model the height anomaly and then to convert GPS height to normal height for practical use. Standard grid-search and particle swarm optimization (PSO) was adopted to tune the hyperparameters. Finally, the results are compared with that of genetic algorithm based back-propagation neural network (GA-BPNN) and conicoid fitting. It is found that the tuned LS-SVR has a better fitting and predictive ability and the normal height derived from GPS height can arrive at centimeter level in a plain region.