Mixture Regression for Sea Ice Segmentation

Abstract

The classification of sea ice in SAR imagery is complicated by statistical nonstationarity. Incidence angle effects, heterogeneous ice conditions and other confounding variables contribute to spatial and temporal variability in the appearance of sea ice. I explore a family of models called mixture regressions which address this issue by endowing mixture distributions with class-dependent trends. I introduce mixture regression as a general technique for unsupervised clustering on nonstationary datasets and propose techniques to improve its robustness in the presence of noise and outliers. I then develop region-based mixture regression models for sea ice segmentation, focusing on the modeling of SAR backscatter intensities under the influence of incidence angle effects. Experiments are conducted on various extensions to the approach including the use of robust estimation to improve model convergence, the incorporation of Markov random fields for contextual smoothing, and the combination of mixture regression with supervised classifiers. Performance is evaluated for ice-water classification on a set of dual-polarized RADARSAT-2 images taken over the Beaufort Sea. Results show that mixture regression achieves accuracy of 92.8% in the unsupervised setting and 97.5% when integrated with a supervised convolutional neural network. This work improves on existing techniques for sea ice segmentation which enable operational ice mapping and environmental monitoring applications. The presented techniques may also be useful for the segmentation of nonstationary images obtained from other remote sensing techniques or in other domains such as medical imaging.