Unsupervised detection of InSAR time series patterns based on PCA and K-means clustering

Abstract

The need for implementing efficient value-adding tools able to optimise Earth Observation data usage, compels the scientific community to find innovative solutions for the downstream of Earth Observation information. In this paper we present an unsupervised and automated approach based on Principal Component Analysis (PCA) and K-means clustering to detect patterns of natural or anthropogenic ground deformation from Interferometric Synthetic Aperture Radar (InSAR) Time Series. For our proof-of-concept, we focus on the Valle d’Aosta region (Northwest Italy) where mass wasting processes frequently occurs, interacting with human activities and infrastructures. The large volumes of Sentinel-1 data produced allows for retrieving horizontal and vertical Time Series from multi-geometry data fusion of Line-of-Sight (LOS) InSAR measurements. The added benefit of combining ascending/descending InSAR data and interpolating displacements in time at different time steps is here explored prior to data dimensionality reduction and feature extraction through PCA. The retrieved principal components serve as a continuous solution for cluster membership indicators in the K-means clustering method, allowing to define spatially and temporally coherent displacement phenomena. The signal of the ground deformation clusters is then deconstructed into the underlying trend and seasonality components to enhance the interpretability of the classified satellite InSAR features. Using InSAR Time series data spanning 2014–2020, the proposed approach detects several slope movements and anthropogenic deformations with both linear and seasonal displacement behaviours. The results demonstrate the potential applicability of our transferable approach to the development of automated ground motion analysis systems.

Keywords

• InSAR time series
• PCA
• Unsupervised learning
• Data mining
• Automated approach
• EO data downstream