A multi-scale technique for detecting forest-cover boundary from L-band SAR images

This paper reports the development of a new multi-scale boundary-detection technique suitable for extracting forest-cover boundaries from L-band Synth etic Aperture Radar (SAR) imagery. Speckle characteristics of SAR data requ ire the smoothing of an image at a rather coarse scale (resolution) so that subsequent edge detection produces a level of detail that is easily interp retable and appropriate for the application. At finer scales, detected edge s are as much due to speckle noise as to true boundary features. In order t o detect interpretable forest boundaries from Japanese Earth Resource Satel lite (JERS)-1 L-band SAR images, a suitable scale was empirically determine d by considering the speckle noise and the spatial resolution of the data. This scale is referred to here as the 'critical scale' because of its impor tance. However, edges detected at this critical scale are distorted geometr ically, while at finer scales edges have progressively better localization but are increasingly noisy. This difficulty in single-scale edge detection is well explained by Canny's uncertainty principle. To overcome this diffic ulty, a centroid attraction algorithm was formulated that integrates edges detected at a range of scales (with the critical scale as the coarsest) to produce a forest-cover boundary map. Such boundary results are shown to be more accurate and clean than those detected at any single scale.