Statistical Edge Detection Operators for Linear Feature Extraction in SAR Images

Abstract

Edge detection in synthetic aperture radar (SAR) images is rendered difficult by the presence of speckle. The data are often filtered using adaptive filters independently of the edge detection process when, in fact, the two steps should be coupled (i.e., the local homogeneity criterion employed by an adaptive filter should be consistent with the edge detector criterion). Five different edge detection algorithms for SAR images are evaluated and compared. The detection algorithms are comprised of two operators based on non-parametric statistical tests, the Ratio of Averages test, difference of averages (essentially a gradient method), and a test based on the mean squared to variance ratio. Two edge thinning and thresholding operations are also compared: an algorithm proposed by Nevatia and Babu (1980), and one based on mathematical morphology (Serra, 1980). Initial testing is carried out on simulated imagery for accurate control of the signal being masked by speckle noise. We obtain the best results using the ratio operator in combination with the morphological thinning operations. High real edge recovery rates are required for segmentation (Fmn > 0.95), and these levels of Fmn are only produced by high-contrast, low mean, local grey level boundaries. We show that a complete boundary delineation for segmentation purposes cannot be expected from a typical Seasat SAR agricultural scene due to the large number of low contrast edges. These methods are applicable in situations where there is a greater contrast between the targets to be discriminated. A suitable application of edge extraction from SAR, for example, is lake boundary extraction; a sample image is presented.