The Schmittlets for automated SAR image enhancement

Kurzfassung

Multi-looking is the essential step in SAR image preprocessing with respect to distributed targets. The presumably independent single intensity measurements of the same target are averaged in order to reduce the local variability and to retrieve a stable “mean” intensity for the target of interest [1]. In practice, multi-looking commonly is performed with a uniform number of looks all over the image though numerous studies already proofed that this is not suitable because signal of smaller targets possibly get mixed while larger targets are insufficiently smoothed. Thus, it is necessary to identify single targets, i.e. their location and their shape, to aggregate sample coming from the same main unit (statistically speaking). Three different approaches can be found in literature and in practice so far. Firstly, locally adaptive filtering techniques switch the filtering kernel of a fixed extension according to the local environment (multi-directional) [2]. Unfortunately, the extension of the local environment to be considered is uniform for the whole image. Hence, the scale of the targets of interest must be known in advance. Secondly, image segmentation aggregates neighboring pixels to segments of arbitrary shape according to homogeneity criteria (multi-directional and quasi multi-scale) [3]. The introduction of sharp edges between neighboring pixels, though very practical for computation, is not justified from an image processing perspective because a higher image resolution is feigned than existent. Furthermore, it denies the existence of mixed pixels in the single look image which are unavoidable due to the limited resolution of SAR images [4]. And thirdly, alternative image representations deliver an optimal multi-directional and multi-scale image description [5]. Evaluating the difference between neighboring scales of the image their application is restricted to images showing additive characteristics, i.e. a normal distribution as commonly accepted for optical images or logarithmic SAR intensities [6]. Therefore, this contribution introduces the Schmittlets as first alternative image representation (multi-directional, multi-scale, and multi-shape) that is applicable to SAR intensities in linear scale. Accordingly, it can easily be utilized for SAR image enhancement and SAR image analysis as well. The Schmittlet index layer indicating the best-fitting Schmittlet out of a selection of 35 geometric primitives for each location in the image gives valuable texture or structure information which can be used for scene characterization or classification purposes.