Forecasting surface movements based on PSI time series using machine learning algorithms

Abstract

Surface movements pose a critical issue requiring prompt detection and monitoring to ensure safety. To address this concern, the detection and continuous monitoring of surface deformations have gained importance. While numerous point-based surveying methods exist, multi-temporal interferometric synthetic aperture radar (InSAR) analysis has emerged as a powerful technique capable of detecting deformations across large areas with freely available SAR images and software. However, the significance of surface movement forecasting must be addressed, as it plays a significant role in preventing accidents and implementing precautionary measures. In light of this, a comprehensive study was conducted at the new Istanbul Airport in Turkiye, employing the Persistent Scatterer InSAR (PSI) method with a collection of 211 Sentinel-1 SAR images. The results obtained were investigated, and six test case regions were determined for the forecasting analysis in terms of time series movement type (subsidence, uplift, and stable) and structure type (runway and building) using various machine learning (ML) algorithms. To accomplish this, regression-based and sequential-based ML algorithms were employed and compared, showing the potential of cutting-edge techniques. The Stacked LSTM algorithm was the most effective in generating accurate forecasts. Furthermore, by incorporating exogenous variables into the forecasting analysis, the study improved the accuracy of the results. To gain deeper insights into the impact of these exogenous variables, feature importance analysis was conducted using the permutation feature importance method. The results obtained were examined, considering both the structure type and the characteristics of the time series data.