STUDY ON GPS HEIGHT TRANSFORMATION USING BP NEURAL NETWORK

摘要

Positions determined by GPS receivers are expressed in geocentric coordinate orrngeodetic coordinate defined by WGS-84 ellipsoid, but in engineering application, thoserncoordinates need to be transformed to local coordinate system (such as Northing and Easting), andrnellipsoidal (geodetic) heights (h) need to be transferred to physical heights, such as orthometricrnheights (H). The height difference between the geodetic height h and the orthometric height H isrncalled undulation N. If the undulation N for a point determined by a GPS receiver is available,rnthen the 3-D geocentric coordinate in WGS-84 can be transformed to local coordinate system andrnorthometric height.rnIn this paper a transformation method from GPS geodetic height (h) to orthometric height (H)rnusing back-propagation (BP) artificial neural network is proposed. A series of 283 bench marks,rnlocated at the central of Taiwan, are selected as the test data. The preliminary test results indicaternthat: (1) the 'trainbr' training algorithm should be used to transform GPS height if both ANNrntraining speed and accuracy are considered, (2) the transformation accuracy decreases fromrn0.0510 m to 0.0331 m when the number of neurons increases from 5 to 40, if the 'trainbr' trainingrnalgorithm was adopted, (3) the performance of GPS height transformation using BP artificialrnneural network is better than the other two estimation methods. The transformation accuracy forrn'BP ANN', 'Curve Fitting', and 'MOI Model' are 0.044 m, 0.199 m, and 0.201 m respectively.