Multiscale and Multidirectional Multilooking for SAR Image Enhancement

Abstract

With the steadily increasing spatial resolution of synthetic aperture radar images, the need for a consistent, but locally adaptive image enhancement rises considerably. Numerous studies already showed that adaptive multi-looking, able to adjust the degree of smoothing locally to the size of the targets, is superior to uniform multi-looking. This study introduces a novel approach of multi-scale and multi-directional multi-looking based on intensity images exclusively, but applicable to an arbitrary number of image layers. A set of two-dimensional circular and elliptical filter kernels in different scales and orientations (named Schmittlets) is derived from hyperbolic functions. The original intensity image is transformed into the Schmittlet coefficient domain where each coefficient measures the existence of Schmittlet-like structures in the image. By estimating their significance via the perturbation-based noise model, the best-fitting Schmittlets are selected for image reconstruction. On the one hand, the index image indicating the locally best-fitting Schmittlets is utilized to consistently enhance further image layers, e.g. multi-polarized, multi-temporal, or multi-frequency layers, and on the other hand, it provides an optimal description of spatial patterns valuable for further image analysis. The final validation proves the advantages of the Schmittlets over six contemporary speckle reduction techniques in six different categories (preservation of the mean intensity, equivalent number of looks, and preservation of edges and local curvature both in strength and in direction) by the help of four test sites on three resolution levels. The additional value of the Schmittlet index layer for automated image interpretation, though obvious, still is subject to further studies.