Graph-Based Anomaly Detection and Localization in InSAR Data

The generation of ground-movement maps using interferometric SAR is subject to faults during the acquisition and prepossessing of SAR data. This paper addresses the task of identifying and localizing such faults specifically within the context of data provided by the European Ground Motion Service (EGMS). The lack of a regular spatial domain in geospatial data poses a challenge for signal processing and machine learning. Here, we investigate graph-based methods, that overcome the spatial irregularity of the data, to perform detection and localization of data anomalies. We demonstrate on synthetic data that graph-based frequency analysis and filtering yield superior performance at anomaly localization when compared to state-of-the-art machine-learning methods such as graph U-nets. The proposed methods are employed for anomaly detection and localization on real EGMS data.