Application of Back Propagation Artificial Neural Network for Modelling Local GPS/Levelling Geoid Undulations: A Comparative Study

SUMMARY The fast development of Global Positioning System (GPS) technology provides more precise and rapid surveying in geodetic applications than the traditional terrestrial positioning techniques. Therefore, considerable savings on time, labour and cost are achieved by GPS measurements. The geometric height supplied by GPS is ellipsoidal height and it needs to be transformed to orthometric height for geodetic applications. For the transformation between ellipsoidal heights and orthometric heights, local and global geoid models generated. In the present study, a local geoid model was first generated according to interpolation methods such as polynomial, KRIG, INDW, MSHP, RBAF and LPOL from the geoid undulations obtained by using GPS/Levelling data for a study area. Subsequently, a back propagation artificial neural network (BPANN) that has been more widely applied in engineering among all other neural network models was used to generate the local geoid model of the study area with the same data. The selected interpolation methods and BPANN are evaluated, in terms of root mean square error (RMSE). In the BPANN method, RMSE was calculated as ±0.0185 m for the reference points and as ±0.0202 m for the test points. These values are smaller than the values obtained by the classical interpolation methods. Large Scale Map and Map Information Production Regulation (LSMMIPR) requires ±5 cm, accuracy level for local geoid determination, in Turkey. Therefore, it was concluded that BPANN can be used for local geoid undulation modelling as an alternative to the interpolation methods.