Multitemporal Analysis of Multipolarization Synthetic Aperture Radar Images for Robust Surface Change Detection

Abstract

This thesis addresses two approaches for change detection from multipolarization, multilooked SAR images: a post-classification comparison and a direct change detection. We consider the complete workflow associated with performing post-classification change detection from time series of multipolarization SAR (PolSAR) images acquired with different imaging geometries and polarimetric configurations. The application is connected to monitoring of changes in Arctic glaciers. The images are corrected for terrain effects by thoroughly reducing topographic effects on both geolocation, radiometry and polarization signature. The matrix-variate U distribution is found to enable proper statistical representation of the variable texture in our data. An unsupervised contextual non-Gaussian clustering algorithm is employed for segmentation of the terrain corrected images. This algorithm has built in contextual smoothing by MRF modeling, and yields homogeneous segmentation, leading to robust change results. The clustered data is subsequently labeled into glacier zones with the aid of ground truth data. The consistency of the segmentation algorithm is also demonstrated by characterizing the expected random error level for SAR images under different imaging conditions. Finally, the classified images of succeeding years are compared, and temporal changes are identified in the location of boundaries between glacier zones. The thesis also proposes a novel method for direct unsupervised change detection from PolSAR data. We assume that the matrix variates follow the complex Wishart distribution, and the complex Hotelling-Lawley trace statistic is applied as a new test statistic for change detection. The sampling distribution of the test statistic is then approximated by a Fisher-Snedecon (FS) distribution. The proposed method is to match the population moments of the FS distribution with those of the HL statistic. The no change hypothesis of equal covariance matrices may then be rejected at a predefined significance level. The performance of the method is demonstrated with good results on simulated and real PolSAR data.