An ICA-assisted LSTM model for spatiotemporal characterization and prediction of ground motion related to geohazards

Abstract

Global earth's surface is subject to adverse effects of a variety of slow and sustained geological hazards, such as land subsidence, earthquake, tectonic motion, mining, landslides, coastal erosion, volcano and permafrost, caused by natural forces and anthropogenic activities disturbance. Earth observation data enable scientists to efficiently assess ground deformation and damages posed by these hazards and result in significant resilience planning. Therefore, preserving a complete record of past, present, and future surface movements is essential for disaster risk mitigation and property protection. It is widely acknowledged that interferometric synthetic aperture radar (InSAR) is a highly effective and widely used geodetic technique for understanding the spatiotemporal evolution of historical ground surface deformation. Meanwhile, deformation prediction is crucial for preventing and mitigating geological hazards, considering the long revisit cycle of satellites and the time it takes to process data.In this study, we propose a strategy that predicts spatiotemporal InSAR time series combining independent component analysis (ICA) and Long Short-Term Memory (LSTM) machine learning model, Here ICA, as a blind signal separation method, is deployed to identify and capture the InSAR displacement signals of interest and characterize each independent time series signal, caused by various natural or anthropogenic processes. In addition, considering that ignoring heterogeneity would reduce the model's accuracy, K-means clustering approach is jointly used to divide the study area into several spatiotemporal homogeneity subregions over a large-scale region, where we assume that the points in the same cluster have similar spatiotemporal behavior. Finally, neural network models for each cluster are constructed and optimal parameters are determined. The suggested study framework is used into two real datasets with diverse deformation characteristics: land subsidence post-seismic time series. The results reveal that our suggested ICA-assisted LSTM model outperforms the original LSTM, with the average prediction accuracy for one-step prediction improved by 34% and 17%, respectively. Furthermore, we mapped the spatiotemporal predicted results of subsidence and post-seismic events in 60 days and examine their performance and limitations, the results of which show high consistency using the enhanced processing technique.The successful prediction on subsidence and post-seismic deformation further indicates that the proposed prediction strategy can be applied to monitor other large-scale geo-hazards with sustained and slow deformation for rapid decision-making and timely risk mitigation. Furthermore, the proposed prediction methodology is applicable to different scenarios of derivative applications. It has application potential in deformation time series fusion across a specified region over the last 20 years, as well as automatic cataloging for abnormal time identification of post-events over broad areas, such as landslides and volcano eruption events, by deviating from the original time series.