An Advancement of Minimum MSE Space Varying Filtering of SAR Interferogram Based on K-SVD Technique

SAR interferometric technique is widely used for analysing the geodynamics of earth surface phenomena. Basically, such imaging technique is accomplished by considering two coherent SAR images from two passes of a single SAR antenna (repeat pass interferometry) or with the single pass of two-antenna system (single pass interferometry) [1], [2]. However, the efficacy of interferometric technique is mostly affected by multiple decorrelation effects, which collectively results as SAR interferogram phase noise. Besides the disturbances such as orbital errors, typical sources of noise is the decorrelation associated with thermal noise, image co-registration error and temporal and geometrical decorrelation. To mitigate decorrelation effects, in this paper we introduce a two steps approach, which combines a minimum mean square error space varying filtering [3] for filtering of the spatial decorrelation and phase denoising based of sparse representation [4] of the interferometric signal. In this context, we start with mitigation of geometrical decorrelation effects as proposed by [3]. Geometrical decorrelation depends mainly on the angular diversity between the two focused images, which causes baseline decorrelation as well as Doppler centeroid decorrelation due to different viewpoint of antenna beam. But presence of nonplanar topography can even increase the limit of geometric decorrelation effect. To overcome the limitation of such critical aspect, the SV-MMSE technique [3] has been successfully used as preliminary step in our procedure. To further reduce the decorrelation effects due to other noise sources, we followed an approach that intend to deal with the residual noise reduction problem by using sparse and redundant representation over " trained dictionaries " [5], strategy that we found to be highly effective and promising. Consequently, we choose K-SVD algorithm [4] as signal representation technique, which efficiently separates signal from noise, through a suitable choose of elementary signals, named " atoms " organized in a matrix form, named " dictionary ". In this framework, we choose a data pair of ERS-1/ERS-2 (Tandem) data acquired over the area of Mt. Vesuvius volcano near Naples, Italy, with a spatial baseline of about 250 m