You are here

IceSynth: An image synthesis system for sea-ice segmentation evaluation

TitleIceSynth: An image synthesis system for sea-ice segmentation evaluation
Publication TypeConference Paper
Year of Publication2009
AuthorsWong, A., W. Zhang, and D. A. Clausi
Conference Name6th Canadian Conference on Computer and Robot Vision
Date Published05/2009
Conference LocationKelowna, British Columbia, Canada
Keywordsautomatic sea-ice monitoring, ice classification, ice types, IceSynth, image classification, image segmentation, image synthesis system, image texture, oceanographic techniques, polar regions, probability, probability distribution estimation, pseudo-ground truth data, remote sensing by radar, SAR imagery, sea ice, sea-ice segmentation evaluation, sea-ice structures, sea-ice textures, stochastic processes, stochastic sampling, synthetic aperture radar
Abstract

An ongoing challenge in automatic sea-ice monitoring using synthetic aperture radar (SAR) is the automatic segmentation of SAR sea-ice images based on the underlying ice type. Given the intractability of obtaining ground-truth segmentation data from polar regions, the evaluation of automatic SAR sea-ice image segmentation algorithms is generally limited to tests using real SAR imagery based on pseudo-ground truth data (e.g., manual segmentations) and simple synthetic tests using basic shape primitives. As such, it is difficult to evaluate automatic segmentation algorithms in a systematic and reliable manner using realistic scenarios. To tackle this issue, a novel image synthesis system named IceSynth is presented, which is capable of generating a variety of synthetic sea-ice images that are representative of real SAR sea-ice imagery. In IceSynth, SAR sea-ice textures for each ice type are synthesized via stochastic sampling based on non-parametric local conditional texture probability distribution estimates. A stochastic sampling approach based on non-parametric local class probability distribution estimates is used to generate large-scale sea-ice structures of various ice types based on ice classification priors extracted from real SAR sea-ice imagery. Experimental results show that IceSynth is capable of generating realistic-looking SAR sea-ice images that are well-suited for performing objective evaluation of SAR sea-ice image segmentation algorithms.

DOI10.1109/CRV.2009.27