Research On Neural Network Elevation Fitting Model Based On Robust Estimation

The refining geoid and elevation transformation of control points are the basis content of modern control survey. It is the crucial link of elevation transformation for the data transferring and the benchmark elaboration in surveying and mapping. To transfer the GNSS height in the practical application, the gravity measurement method and the traditional numerical fitting method are used. It is carried out in-depth analysis, which is mainly aimed at the mathematical model and calculating precision in the process of calculating the height anomaly with the neural network.The model is formed by BP neural network fitting model with robust estimation based on MEA optimization, which is applied in the wide range and varied terrain area.Based on the features of weaker differential resistance and local optimization of BP network, this model improves the differential resistance, which optimizes the input of model and method of weights and threshold of hidden layer. Experimental results prove that the model has shown a certain superiority, which is test by using the engineering data respectively. In order to meet the demand of height anomaly fitting in a small range, a new model with robust estimation is proposed, which is based on the regularization of RBF neural network. With the idea of robust estimation in data processing, the model improves the randomness in the process of value selection of hidden layer. It presents the principle that the low precision and small steady power generate the longer distance between the center and input. The model improves the performance of resistance and shows the superiority, which is verified with engineering data.In view of mathematical modeling, the internal and external precision of the model is analyzed and the corresponding conclusion is drawn, which has a certain practical significance in elevation fitting with the diversity of observed values in the process of automated data collection. Considering the influence of terrain features and the quality of data, the model still need further improvement in practical.