A Machine Learning Detection of Outliers in InSAR Displacement Time Series

Multi-temporal SAR interferometry (InSAR) estimates the displacement time series of coherent radar scatterers. Current InSAR processing approaches often assume the same deformation model for all scatterers within the area of interest. However, this assumption is often wrong, and time series need to be approached individually. Individual, point-wise approach for large InSAR datasets is limited by high computational demands. The additional problem is imposed by the presence of outliers and phase unwrapping errors, which directly affect the estimation quality. This work describes the algorithm for (i) estimating and selecting the best displacement model for individual point time series and (ii) detecting outlying measurements in the time series. The InSAR measurement quality of individual scatterers varies, which affects the estimation methods. Therefore, our approach uses a priori variances obtained by the variance components estimation within geodetic InSAR processing. We present two different approaches for outlier detection and correction in InSAR displacement time series. The first approach uses the conventional statistical methods for individual point-wise outlier detection, such as median absolute deviation (MAD) confidence intervals around the displacement model. The second approach uses machine learning principles to cluster points based on their displacement behaviour as well as the temporal occurrence of outliers. Using clusters instead of individual points allows for more efficient analysis of average time series per cluster and consequent cluster-wise outlier detection, correction, and time-series filtering. The two approaches have been applied on the Sentinel-1 InSAR time series of a case study from monitoring landslides in Slovakia. The area of interest is affected by characteristic non-linear progression of the movement. Our post-processing procedure parameterized the displacement time series despite the presence of a non-linear motion, thus enabling reliable outlier detection and unwrapping error correction. The validation of the proposed approaches was performed on an existing network of corner reflectors located within the area of interest.