Edge-aware segmentation in satellite imagery: A case study of shoreline detection

Shoreline extraction algorithms from multispectral imagery depend on threshold selection over spectral values and segmentation in general. Although this method gives high performance values for water delineation, error is accumulated on pixels near shoreline and complicates detection of nearby ships, docks etc. Water-shadow spectral mixing and spectral difference in water regions are two of the reasons for such untrustworthy shoreline results. With only four bands available, improvement in water detection depending only on pixel values is not very promising. Therefore, segmentation gains importance. By an edge-aware segmentation method, we aim to improve overall water and shoreline detection performances. In this study, a robust three-stage shoreline extraction algorithm is proposed. In the first stage, segmentation is applied over spectral values and then, some segments are combined according to edge information. In the second stage of the algorithm, pixel-based water information is combined with segmentation. The last step consists of enhancement of water regions based on local optimization by merging regions near shore boundary. Additionally, two new boundary-sensitive performance metrics are introduced for measuring the accuracy of the detected boundaries.