Multithresholding techniques in SAR image classification

Abstract

Beyond bilevel segmentation, multithresholding (MT) is a powerful technique that is seldom considered in image analysis. In active imaging techniques, however, MT appears as a significant alternative due to its obliviousness to the inherent nonlinear noise present in these kind of images. This is especially the case in satellite synthetic aperture radar (SAR) images, which are becoming an increasingly popular information source in remote sensing, given that the acquisition is independent of weather and daylight conditions. In SAR images, data-dependent multiplicative noise (speckle) hampers most of the image filtering techniques available from linear theory, and thus nonlinear techniques like MT appear to be promising. There are, however, a large amount of proposals in this direction, each with different theoretical or empirical justifications, and thus a careful analysis of their advantages in results quality and computational cost have to be assessed. In this paper we survey the most representative MT techniques and methods applied to SAR imagery (both synthetic and actual satellite images), and evaluate them in terms of region segmentation quality and computational cost. Results show that the maximum likelihood method provides the best quality segmentation results at the expense of higher computation times, while a state transition based method provides the fastest results with acceptable quality, and that all methods can be assessed with a simple tradeoff representation.