Monitoring a portfolio of railway bridges using spaceborne InSAR measurements and transfer learning: methods and applications.

This thesis presents a population-based framework for monitoring railway bridges using spaceborne interferometric synthetic aperture radar (InSAR), environmental variables, and transfer learning within a structural health monitoring perspective. The study investigates seven steel railway bridges along the Po River in Italy using European Ground Motion Service (EGMS) Level 2a line-of-sight displacement time series from ascending and descending satellite tracks. The proposed methodology transforms heterogeneous bridge-track observations into a common feature-based representation for bridge-to-bridge comparison. After spatial filtering and coherence-based screening, bridge-level features are extracted from persistent scatterers. The feature set consists of five statistical descriptors of the displacement field together with air temperature and river water level. These variables are synchronized on a regular temporal grid and processed through principal component analysis and subspace alignment, allowing all bridge tracks to be represented in a common latent domain. Within this domain, a population-relative anomaly index is defined as the distance between each bridge-track state and the instantaneous centroid of the monitored population, enabling the identification of tracks whose behavior deviates from the collective response. The results identify bridge B4 as the most anomalous case, with one track showing a persistent increase in anomaly index. Feature-level analysis and spatial diagnostics reveal that the anomaly is localized to specific portions of the bridge rather than uniformly distributed. The study demonstrates the potential of transfer-learning-based anomaly detection for portfolio-scale bridge monitoring with InSAR data while highlighting limitations related to temporal interpolation, track heterogeneity, and the lack of a unified bridge-level indicator. It provides a basis for future developments toward more robust bridge-monitoring frameworks.