Unsupervised Time Series-InSAR Deformation Analysis via Joint Dimensionality Reduction and Clustering

Ground deformation monitoring is essential for assessing natural and anthropogenic hazards. Advanced Differential Interferometric SAR (DInSAR) techniques, such as Persistent Scatterer Interferometry (PSI) and Small Baseline Subset (SBAS), enable large-scale spatiotemporal analysis of ground deformation. However, TS-InSAR datasets may contain a large number of measurement points (MPs) with different levels of reliability, making appropriate quality control essential before subsequent data analysis. In this study, PSI and SBAS datasets were processed using SARProz, where interferometric quality indicators were used to identify and retain reliable measurement points while excluding low-quality or unstable observations. The reliable TS-InSAR measurements were subsequently integrated through a multi-geometry spatial and temporal processing framework to derive vertical and horizontal displacement time series. To efficiently characterize the resulting large-scale dataset, an automated unsupervised machine learning framework combining dimensionality reduction and clustering is proposed. Two dimensionality reduction approaches are evaluated and compared: Principal Component Analysis (PCA), a linear statistical method used as a baseline, and an Autoencoder (AE) neural network capable of capturing nonlinear structures within the data.Both approaches are assessed in terms of their ability to preserve meaningful deformation information while reducing data complexity. The resulting low-dimensional representations are subsequently processed using K-means clustering to identify distinct deformation regimes at a regional scale. Furthermore, each identified cluster is decomposed into trend and seasonal components to provide a comprehensive characterization of temporal displacement behavior. The proposed methodology provides a scalable and data-driven framework for the automated interpretation of TS-InSAR observations and supports geohazard mapping and geological risk assessment by facilitating the identification of spatially and temporally distinct ground deformation patterns.