局部区域高程异常拟合的二次曲面／RBF神经网络组合方法

Quadratic Surface/RBF Neural Network Combination in Regional Height Anomaly Fitting

在地形复杂区域，用二项式曲面进行GNSS高程拟合时存在模型误差大和转换精度不高的问题。RBF（Radio Basis Function）径向神经网络具有全局最优、局部逼近以及良好的泛化能力等特性，研究构建了”二次曲N／RBF神经网络”组合的GNSS高程转换模型。通过实例数据对单一的二次曲面拟合、RBF神经网络拟合及组合模型转换精度进行了比较，结果表明：该组合模型转换精度高于单一的转换模型，能达到较好的拟合效果，可适用于地形复杂区域。

In complex terrain, the problem of great model error and low precision of conversion accuracy remain when using binomial surface for GNSS height fitting. RBF （ Radio Basis Function） radial neural network has the characteristics of global optimization, local approximation and good generalization ability. A combination model of quadric surface and RBF neural network for GNSS height is studied and established. Through example data,the single quadric surface fitting, RBF neural network model and the combined model of GNSS height conversion accuracy are compared and analyzed. The results show that the conversion accuracy of the combined model is higher than that of a single conversion model, which can achieve better fitting effect and is suitable for the terrain complex region.