Enabling scalable spectral clustering for image segmentation
Pattern Recognition
| Hi-index | 0.00 |
The use of synthetic aperture radar (SAR) has become an integral part of sea-ice monitoring and analysis in the polar regions. An important task in sea-ice analysis is to segment SAR sea-ice imagery based on the underlying ice type, which is a challenging task to perform automatically due to various imaging and environmental conditions. A novel stochastic ensemble consensus approach to sea-ice segmentation (SEC) is presented to tackle this challenging task. In SEC, each pixel in the SAR sea-ice image is assigned an initial sub-class based on its tonal characteristics. Ensembles of random samples are generated from a random field representing the SAR sea-ice imagery. The generated ensembles are then used to re-estimate the sub-class of the pixels using a weighted median consensus strategy. Based on the probability distribution of the sub-classes, an expectation maximization (EM) approach is utilized to estimate the final class likelihoods using a Gaussian mixture model (GMM). Finally, maximum likelihood (ML) classification is performed to estimate the final class of each pixel within the SAR sea-ice imagery based on the estimated GMM and the assigned sub-classes. SEC was tested using a variety of operational RADARSAT-1 and RADARSAT-2 SAR sea-ice imagery provided by the Canadian Ice Service (CIS) and was shown to produce successfully segmentation results that were superior to approaches based on K-means clustering, Gamma mixture models, and Markov Random Field (MRF) models for sea-ice segmentation.