Coherence-Based Prediction of Multi-Temporal InSAR Measurement Availability for Infrastructure Monitoring

Abstract

Predicting the availability of measurement points provided by Multi-Temporal Interferometric Synthetic Aperture Radar (MT-InSAR) poses a challenge due to a nonuniform distribution of Persistent Scatterers (PSs). This article introduces a novel method to estimate the availability of MT-InSAR results on buildings and infrastructure networks, eliminating the need for labor-intensive and time-consuming analyses of the entire SAR data stack. The method is based on an analysis of the interferometric coherence decay characteristics and data regarding buildings and transport infrastructure location as inputs to a convolutional neural network. Specifically, a U-Net architecture model was implemented and trained to predict the PS density of Sentinel-1 data. The methodology was applied to a regional-scale analysis of the Dutch infrastructure, resulting in a low 1.06$\pm$0.10 mean absolute error in the pixel-based PS count estimation on the test data split, with over 80% of predictions within $\pm$1 from the actual value. The model achieved high accuracy when applied to a previously unseen dataset, demonstrating strong generalization performance. The proposed workflow, with its notable ability to accurately predict areas lacking measurement points, offers stakeholders a tool to assess the feasibility of applying MT-InSAR for specific structures. Thereby, it enhances infrastructure reliability by addressing a critical need in decision-making processes and improving the applicability of MT-InSAR for structural health monitoring of infrastructure assets.

Keywords

• Neural networks
• rail transportation
• remote sensing
• road transportation