首页|基于时序InSAR与机器学习的大范围地面沉降预测方法

基于时序InSAR与机器学习的大范围地面沉降预测方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
地面沉降是由自然因素和人为因素综合作用下形成的地面标高损失,为预防这种累进性的缓变地质灾害,高效的大范围地面沉降预测显得尤为重要.现有的预测方法忽略了地面沉降的空间特征,且基于单点循环预测存在高耗时现象.针对上述问题,提出了一种基于时序InSAR与机器学习的大范围地面沉降预测方法.首先,利用SBAS-InSAR技术获取大范围的地面沉降时序信息;其次,采用经验正交函数(empirical orthogonal function,EOF)提取时序信息的空间模态及对应的主成分(principal components,PCs);最后,采用基于误差反馈的岭多项式神经网络(ridge polynomial neural network with error-output feedbacks,RPNN-EOF)模型训练与预测PCs,将预测结果重构回地面沉降时序.以延安新区 2018 年 8 月至 2021 年 5 月的 84 景Sentinel-1A数据为例,获取了新区的地面沉降时序.同时,EOF所提取的空间模态能清晰地表达整个新区的空间变化特征.预测结果显示,相较于传统点循环模式以及主流的时间序列预测方法,本文方法的均方根误差至少降低了 22.7%,建模耗时至少降低了 27.5%,因此该方法具有良好的实用性.
The Prediction Method of Large-Scale Land Subsidence Based on Multi-Temporal InSAR and Machine Learning
Land subsidence is the loss of land elevation formed by a combination of natural and human factors.To prevent the delayed progressive geohazards,it is essential to predict large-scale land subsidence with high efficiency.However,the current prediction methods usually neglect spatial characteristics of land subsidence,which are time-consuming due to the issue of single-point cycle.To address the problem,a new prediction method of large-scale land subsidence based on multi-temporal InSAR and machine learning is proposed.Firstly,the large-scale land subsidence time series information is obtained by the SBAS-InSAR technique.Secondly,the spatial modes and the consistent principal components(PCs)are extracted from the time series information with the empirical orthogonal function(EOF).Finally,the PCs are trained and predicted by predictive model based on the ridge polynomial neural network with error-output feedbacks(RPNN-EOF),and the outcomes are reconstructed back to the land subsidence time series.The 84-view Sentinel-1A data from August 2018 to May 2021 of Yan'an New District were adopted in the land subsidence time seriesacquisition.Simultaneously,the spatial modes extracted by EOF can clearly reveal the spatial variation characteristics of the whole new district.The prediction results show that the root mean square error and modeling time of the proposed method is reduced by at least 22.7%and 27.5%respectively,in comparison with that by the single-point cycle pattern and the prevailing time series methods.Thus it has good practicality and applicability.

machine learningtime series predictionempirical orthogonal functionsspatial modalitiesneural networksmulti-temporal InSARland subsidencehazards

罗袆沅、许强、蒋亚楠、孟冉、蒲川豪

展开 >

成都理工大学地质灾害防治与地质环境保护国家重点实验室,四川成都 610059

成都理工大学地球科学学院,四川成都 610059

机器学习 时间序列预测 经验正交函数 空间模态 神经网络 时序InSAR 地面沉降 灾害

国家自然科学基金国家自然科学基金国家重点研发计划地质灾害防治与地质环境保护国家重点实验室自主研究项目

41790445416306402021YFC3000401SKLGP2023Z026

2024

10.3799/dqkx.2023.048

地球科学
中国地质大学

地球科学

CSTPCD北大核心
影响因子：1.447
ISSN：1000-2383
年,卷(期)：2024.49(5)
  • 41