C11A-0348:
Effective SAR Image Segmentation in Analysis of Sea Ice Floe Distribution (FSD) Using Graph-cut Based Feature Extraction and Fusion

Monday, 15 December 2014
Soumitra Sakhalkar, University of Strathclyde, Glasgow, G4, United Kingdom
Abstract:
Soumitra Sakhalkar1, Jinchang Ren1 and Byong Jun Hwang2

1 Centre for excellence in Signal & Image Processing (CeSIP), Department of Electronic and Electrical Engineering, University of Strathclyde, Glasgow, G1 1XQ, UK.

2 Scottish Association for Marine Science (SAMS), Oban, PA37 1QA, UK.

Sea ices that grow in the open seas are characteristically different in forms and shapes from the largely smooth sea ice that grows in calm inlets. For example, strong force from winds and waves fractures the thick sea ice into pieces or floes, which then collide with each other. In studies of the Polar Regions with satellite SAR (Synthetic Aperture Radar) imagery, identification of ice floes and their distribution is particularly important for examining for both large and small scale applications.In this paper, a Graph-Cut (GC) based feature extraction and fusion technique has been proposed for effective segmentation of SAR images and following on FSD analysis. Though GC based approach has been used in the segmentation of natural images, the application of it on SAR image in this context is rare. Based on an energy minimization process, the GC technique has utilized a graph based representation in grouping pixels for segmentation. To deal with sparkle noise, effective pre-processing and image filter is also applied.To validate the efficacy of the proposed approach, real SAR images with a high resolution of 16k by 16k are used for both visual assessment and quantitative analysis. In comparison to several state-of-the-art algorithms such as watershed and K-means it is found kernel based GC approach yields the most accurate results as shown in Fig. 1.

Fig. 1: One example image (t-l) and its ground truth (t-m) along with results of segmentation using graph cut (t-r) and (b-l), watershed (b-m) and K-means (b-r).