TY - JOUR

T1 - Characterization and prediction of InSAR-derived ground motion with ICA-assisted LSTM model

AU - Peng, Mimi

AU - Motagh, Mahdi

AU - Lu, Zhong

AU - Xia, Zhuge

AU - Guo, Zelong

AU - Zhao, Chaoying

AU - Liu, Qinghao

N1 - Funding Information: This research is supported by National Key Research and Development Program of China (No.2022YFC3004302), the National Natural Science Foundation of China (Grant No. 41929001), the Fundamental Research Funds for the Central Universities, CHD (No. 300102269722), the China Scholarship Council (No. 202006560072). The authors would like to thank the European Space Agency for freely providing the Sentinel-1 SAR data. SRTM digital surface model is provided by USGS via Earth Explorer (https://earth explorer.usgs.gov). We also thank the anonymous reviewers and editors for their constructive and valuable comments on the manuscript.

PY - 2024/2/1

Y1 - 2024/2/1

N2 - Interferometric Synthetic Aperture Radar (InSAR) is a highly effective and widely used approach for monitoring large-scale ground deformation. The precise and timely prediction of deformation holds significant importance in mitigating and preventing geological hazards, particularly considering the long revisit cycle of satellites and the considerable time required for data processing. In this study, we propose a strategy that predicts spatiotemporal InSAR time series based on Independent Component Analysis (ICA) and the Long Short-Term Memory (LSTM) machine learning model. Unlike traditional methods that rely on physical or statistical models, the proposed strategy leverages the power of ICA and LSTM to achieve accurate predictions without such dependencies. ICA is employed to decompose and capture the InSAR displacement signals of interest caused by various natural or anthropogenic processes and to characterize each individual signal. The spatiotemporal unsupervised K-mean cluster method is then applied to partition large-scale deformation fields into homogeneous subregions, considering the spatial variations and temporal nonlinearities of time series. This process facilitates the refinement of the model, thereby enhancing the accuracy of large-scale predictions. The neural network models are then individually constructed for each cluster, and the optimal parameters are determined through a grid search strategy. Subsequently, the proposed framework is implemented and assessed using two datasets featuring distinct deformation patterns: Case I involves land subsidence in Willcox Basin, USA, while Case II focuses on post-seismic deformation following the 12 November 2017 Mw 7.3 Sarpol-e Zahab earthquake. The results demonstrate that our proposed ICA-assisted LSTM outperforms the original LSTM model on large-scale deformation prediction, with the average prediction accuracy for one-step prediction (12 days in our case) being improved by 34% and 17% for cases I and II, respectively. Furthermore, we perform iterative predictions on the spatiotemporal InSAR measurements with varying temporal characteristics for the subsequent five steps using Sentinel-1 data and evaluate its performance and limitations. The successful prediction of land subsidence and post-seismic deformation provides further evidence that the proposed prediction strategy can be effectively employed in monitoring other large-scale geohazards characterized by prolonged and gradual deformation. This capability enables expedited decision-making and timely implementation of risk mitigation measures.

AB - Interferometric Synthetic Aperture Radar (InSAR) is a highly effective and widely used approach for monitoring large-scale ground deformation. The precise and timely prediction of deformation holds significant importance in mitigating and preventing geological hazards, particularly considering the long revisit cycle of satellites and the considerable time required for data processing. In this study, we propose a strategy that predicts spatiotemporal InSAR time series based on Independent Component Analysis (ICA) and the Long Short-Term Memory (LSTM) machine learning model. Unlike traditional methods that rely on physical or statistical models, the proposed strategy leverages the power of ICA and LSTM to achieve accurate predictions without such dependencies. ICA is employed to decompose and capture the InSAR displacement signals of interest caused by various natural or anthropogenic processes and to characterize each individual signal. The spatiotemporal unsupervised K-mean cluster method is then applied to partition large-scale deformation fields into homogeneous subregions, considering the spatial variations and temporal nonlinearities of time series. This process facilitates the refinement of the model, thereby enhancing the accuracy of large-scale predictions. The neural network models are then individually constructed for each cluster, and the optimal parameters are determined through a grid search strategy. Subsequently, the proposed framework is implemented and assessed using two datasets featuring distinct deformation patterns: Case I involves land subsidence in Willcox Basin, USA, while Case II focuses on post-seismic deformation following the 12 November 2017 Mw 7.3 Sarpol-e Zahab earthquake. The results demonstrate that our proposed ICA-assisted LSTM outperforms the original LSTM model on large-scale deformation prediction, with the average prediction accuracy for one-step prediction (12 days in our case) being improved by 34% and 17% for cases I and II, respectively. Furthermore, we perform iterative predictions on the spatiotemporal InSAR measurements with varying temporal characteristics for the subsequent five steps using Sentinel-1 data and evaluate its performance and limitations. The successful prediction of land subsidence and post-seismic deformation provides further evidence that the proposed prediction strategy can be effectively employed in monitoring other large-scale geohazards characterized by prolonged and gradual deformation. This capability enables expedited decision-making and timely implementation of risk mitigation measures.

KW - Clustering

KW - ICA

KW - InSAR time series prediction

KW - Land subsidence

KW - Large-scale

KW - LSTM

KW - Postseismic

UR - http://www.scopus.com/inward/record.url?scp=85178624394&partnerID=8YFLogxK

U2 - 10.1016/j.rse.2023.113923

DO - 10.1016/j.rse.2023.113923

M3 - Article

AN - SCOPUS:85178624394

VL - 301

JO - Remote sensing of environment

JF - Remote sensing of environment

SN - 0034-4257

M1 - 113923

ER -