An Integrated PS-InSAR and Machine Learning Framework for Bridge Displacement Monitoring

Abstract

Ensuring the safety and long-term performance of bridge infrastructure requires reliable methods for monitoring structural behavior over time. Traditional structural health monitoring (SHM) systems often rely on in-situ sensors that provide detailed measurements, but are typically limited in spatial coverage and access to the bridge site. In contrast, satellite-based interferometric synthetic aperture radar (InSAR) enables millimeter-level displacement monitoring over large areas without the need of installing and maintaining sensors. This thesis develops a new approach for integrating the time-series modeling, SAR satellite interferometry and machine learning techniques, to better support data-driven structural monitoring. As employing ML techniques for structural investigations must be correctly applied, initially, a FEM model of a cantilever beam and dynamic loading was investigated and its response time-series modeling approaches, similar to the ones used in SHM, were examined, with particular attention to deep learning methods such as Long Short-Term Memory (LSTM) networks, multivariate time-series transformers and LSTM-based autoencoders. These models are reviewed and analyzed for their ability to capture complex temporal patterns in structural response data and their potential for supporting data-driven damage detection. Once these techniques were confirmed as accurately performing for structural applications, a real bridge was selected for monitoring with remote sensing methods (satellites) and used the ML techniques for the assessment of the measured responses. Thus, Persistent Scatterer InSAR (PS-InSAR) method was applied to derive displacement time series for persistent scatterer points, over five years, located on the Victoria Bridge in Montreal. Statistical and temporal features were extracted from the displacement signals and an unsupervised anomaly detection framework based on the Isolation Forest algorithm was developed to identify unusual displacement patterns. Spatial filtering, to ensure the neighborhood consistency of the identified anomalous points on the deck, and bridge deck-axis segmentation, to create regular bins of measured data that are easier to interpret statistically but structurally as well, were developed and implemented to improve the robustness of anomaly identification. These techniques are newly developed in the current study. The proposed framework was extended to a multi-satellite analysis of PS-InSAR observations from Sentinel-1, RADARSAT Constellation Mission (RCM), and RADARSAT-2 to identify the challenges and compatibilities between the measurements of the same monitored structure. Having access to data from three different satellites tasked to monitor the same structure, Victoria Bridge in this case, is very rare and very valuable. Using an enhanced anomaly detection workflow that integrates statistical filtering and spatial constraints, displacement trends and anomaly patterns were compared across different satellites datasets to assess the consistency and reliability of satellite-based bridge monitoring. The PS-InSAR investigation of the Victoria Bridge, enhanced with anomaly detection using spatial filtering techniques, demonstrated that the proposed approach can effectively highlight spatial zones with concentrated anomalies, while improving the interpretability of satellite-derived displacement measurements. Together, these contributions illustrate the potential of combining satellite remote sensing and machine learning techniques to support scalable and data-driven approaches for long-term bridge monitoring.