循环神经网络的地面沉降预测方法  

Study on prediction method of land subsidence based on recurrent neural network

作　　者：岳振华 沈涛[1] 毛曦[1] 马维军[1,2] YUE Zhenhua;SHEN Tao;MAO Xi;MA Weijun(Chinese Academy of Surveying and Mapping,Beijing 100036,China;Capital Normal University,Beijing 100048,China)

机构地区：[1]中国测绘科学研究院,北京100036 [2]首都师范大学,北京100048

出　　处：《测绘科学》2020年第12期145-152,共8页Science of Surveying and Mapping

摘　　要：针对现有的地面沉降预测方法中确定性模型应用复杂,对数据参数要求高,而基于历史观测数据的时间序列模型局限于单观测点的预测,对单观测点历史数据量要求较高等问题,该文提出了一种利用互相关对地面沉降时间序列进行聚类,然后用循环神经网络(RNN)对聚类后各个数据类分别建立预测模型的方法。实现了对多观测点建立统一的沉降预测模型,且应用简单,降低了对单个观测点历史数据量的要求。基于抚顺市合成孔径雷达干涉测量(InSAR)沉降观测数据的实验结果表明,基于互相关的聚类方法能够有效地区分不同沉降趋势,而对沉降趋势相似的时间序列通过循环神经网络建立沉降预测模型有较高的精度。In view of the complexity of the application and the high requirements of data parameters of deterministic models,while the existing time series prediction models based on historical observation data are limited to the prediction of single observation point,which require a high amount of historical data of single observation point in the existing land subsidence prediction methods,this paper proposed a method of clustering the time series of land subsidence by using cross-correlation,and then using the recurrent neural network(RNN)to establish the prediction model for each data class.A unified land subsidence prediction model for multiple observation points was realized,which was simple in application and reduced the requirement of historical observation data.The experimental results based on interferometric synthetic aperture radar(InSAR)land subsidence observation data in Fushun showed that the clustering method based on cross-correlation could effectively distinguish different land subsidence trends,while the prediction models based on RNN for land subsidence time series with the same subsidence trend had a good performance.

关 键 词：地面沉降 互相关 聚类 循环神经网络 沉降预测 

分 类 号：P208[天文地球—地图制图学与地理信息工程]