Bayesian正则化BP神经网络拟合两类似大地水准面  

Application of Bayesian Regulation BP Neural Network to Fit Two-Kind Quasi-geoid

作　　者：宋雷[1] 方剑[2] 周旭华[2] 黄腾[1] 

机构地区：[1]河海大学土木工程学院,南京市210098 [2]中国科学院测量与地球物理研究所,武汉市430077

出　　处：《武汉大学学报（信息科学版）》2009年第5期552-555,560,共5页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金资助项目(40674013);国家863计划资助项目(2006AA09Z153)

摘　　要：为限制重力似大地水准面拟合到GPS/水准似大地水准面上的模型代表性误差,提出了Bayesian正则化BP神经网络拟合两类似大地水准面的新方法。利用某区域的重力似大地水准面模型和GPS/水准数据,将新方法与传统的曲面拟合法进行比较。在较大区域和两类似大地水准面差别不规则的情况下,Bayes-ian正则化BP神经网络有效地减少了拟合模型的代表性误差,而且通过Bayesian正则化算法对网络权值进行限制,抑制了过拟合现象。新方法提高了两面拟合结果的内、外符合精度。In order to restrict the models error of fitting gravimetric quasi-geoid to GPS/levelmg quasi-geo neural netwo certain area, the new method of fitting two kind quasi-geoid using Bayesian regulation BP was proposed. Using the gravimetric quasi-geoid and GPS/leveling data in a e new method was compared with polynomial surface fitting method. In the case with biggish area and anomalous difference between two kind of quasi-geoid, Bayesian regulation BP neural network could reduce the erros of models, and Bayesian regulation arithmetic could improve the structure of network by restricting weights to produce a smoother network response . The experimental result shows that the new method can improve the inner and outer precisions of fitting two kinds of quasi-geoid clearly.

关 键 词：Bayesian正则化 BP神经网络 似大地水准面 高程异常 

分 类 号：P223.0[天文地球—大地测量学与测量工程]