Application Of Generalized Regression Neural Network Optimized Based On Genetic Algorithm In GPS Height Conversion

Presently, the application of GPS satellite positioning system in the field of surveyingand mapping is mainly manifested in the horizontal control survey, while the precisionin the area of height measurement has always been seen as unreliable and the markingand claiming precision of the instrument is also relatively lower than horizontalpositioning precision in the effect of the inconsistent of coordinate system, observationerror and so on, which largely restricted the application of GPS technology in heightmeasurement. So far, a large number of scientific research and practice has been doneand a variety of GPS height conversion methods which can be roughly classified intothree categories: geometric analytic method, physical geodesy method and neuralnetwork appeared to realize the application of GPS in height measurement. The GPSheight conversion based on physical geodesy method is not reliable for the scarcity ofgravity resources in our country. Therefore, it is widely pervasive in the research ongeometric analytic method and neural network in GPS height transformation, but thesemethods more or less have some defects.In this paper, a new method of GPS height conversion called GA-GRNN method hasbeen put forward on the basis of summarizing the predecessors’ research, which usesglobal optimization of genetic simulated algorithm to optimize the smooth factor of thegeneralized regression neural network, thereby improving the precision of generalregression neural height conversion network. To achieve this method, feasibility of itshould be illustrated by theoretical arguments of genetic algorithm and generalizedregression neural network, and then explore the superiority of this new method bycompare the transform results between specific experiments and quadratic surfacefitting, BP neural network, RBF neural network.Experiments show that the combination of genetic algorithm and generalizedregression neural network which has obvious effect in optimization of generalizedregression neural network can give full play to the advantages of genetic algorithm inglobal search, and avoid the limitations of smooth factor selection artificially, and thecalculation speed is faster with good integral approximation performance, and theelevation conversion result is ideal which can fully meet the needs of large scaletopographical mapping. Besides, the method of GA–GRNN can avoid the model errorcompared with quadric surface fitting, avoid trapping in local optimum compared with BP neural network, avoid the Cumbersome multiple parameters optimizationcompared with the RBF neural network, avoid keeping trying of SPREAD parametercompared with the GRNN method. So GA-GRNN method has good accuracy andhigher prediction efficiency. Finally GA-GRNN method has high generalization abilitycompared with other traditional GPS height conversion method, and the degree ofdependence upon the point number and distribution that participate in fitting has beenweakened. So give priority to GA-GRNN method when lack of data points or underthe condition of uneven distribution of data points.To sum up, the method is feasible and has certain accuracy that based on geneticalgorithm to optimize the generalized regression neural network for GPS heightconversion.