确定区域大地水准面的K-means-RBF神经网络方法  

The K-means-RBF Neural Network Method for Determining Regional Geoid

作　　者：黄子炎 王庆宾[1] 冯进凯 谭勖立 范雕 黄炎 HUANG Ziyan;WANG Qingbin;FENG Jinkai;TAN Xuli;FAN Diao;HUANG Yan(Information Engineering University, Zhengzhou 450001, China;61243 Troops, Urumqi 830002, China)

机构地区：[1]信息工程大学,河南郑州450001 [2]61243部队,新疆乌鲁木齐830002

出　　处：《测绘科学技术学报》2021年第3期246-252,共7页Journal of Geomatics Science and Technology

基　　金：国家自然科学基金项目(41774018);信息工程大学科研团队发展基金(f3212,f5206)。

摘　　要：确定区域大地水准面的几何水准方法在拟合大地水面时未顾及重力场信息,仅是一种单纯的数学拟合,忽略了重力场数据自身的物理性和不同数据间的相关性。近年来,深度学习方法得到广泛重视与研究。本文提出了一种有监督学习的RBF神经网络精化大地水准面的方法,使用包含重力异常和大地水准面高的重力场数据进行神经网络训练,并采用K-means聚类算法为RBF神经网络的径向基函数进行初始化,提高神经网络的收敛速度和精度。实验结果表明,该方法确定的平原、丘陵和山地复杂实验区域大地水准面高标准差分别为0.044、0.159和1.075 cm,优于使用几何水准直接拟合大地水准面高的精度,且在重力异常中加入蒙特卡罗随机噪声模拟的观测误差后,3类实验区域标准差总体仍在cm级,误差增幅不显著,表明该方法在确定大地水准面时,能够抑制观测误差的影响。The information of gravity field is not took into account when fitting the regional geoid in the geometric leveling method which is only a mathematical fitting ignoring the physical properties of gravity data and the correlation between different data.In recent years,deep learning method has been widely valued and studied.A kind of RBF neural network with supervised learning to refine geoid is proposed in the paper,where the data of gravity anomaly and geoid height are brought to train neural network,and K-means clustering algorithm is exploited to initialize radial basis function of neural network and to improve the convergence speed and accuracy of neural network.The experimental results show that the standard deviations of geoid height in plain,hilly and mountainous areas calculated using this method are 0.044,0.159 and 1.075 cm respectively,which are better than those of using geometric leveling.Then,some certain observation errors are added to the gravity anomalies,which are simulated by Monte Carlo random noise.It turns out that the standard deviations of three experimental areas are still in cm level,which shows that the method can restrain the influence of observation error when determining geoid.

关 键 词：大地水准面 重力异常 深度学习 K-MEANS聚类 RBF神经网络 

分 类 号：P223[天文地球—大地测量学与测量工程]