Detection of Linear Objects in ERS-1 SAR Images Using Neural Network Technology

A classification method for the automatic detection of linear objects in synthetic aperture radar (SAR) images is proposed. It is based on feature extraction using a line model, some basic cues from human vision and a neural network classification considering local and global parameters. The method is applied to ERS-1 SAR images to derive the locations of lake and forest boundaries. didate. Either geometric or statistic parameters are derived from the window area. Methods evaluating geometric parameters like gradients of the intensity function or zero crossings of the second derivative have been developed by Burns et al. (1981), and Kwok (1989). Several investigations conclude that a local gradient analysis is not suitable for edge detection in SAR images because of the speckle influence (e.g. Giess, 1984; Adair & Guindon, 1990). Statistic methods based on local contrast ratio have for instance been developed by Touzi et al. (1988), and Adair & Guindon (1990). A more global assessment is conducted by Samadani & Vesecky (1990) using a maximum-a-priori estimation and Arduini et al. (1992) deriving Markov random fields. Bovik's (1988) and Touzi et al.'s (1988) work is based on statistic as well as geometric parameters. Methods based on a global estimation are those using the Hough transform (e.g. Skingley & Rye, 1987), or dynamic programming (Wood, 1985). INTRODUCTION