A Semi-Supervised Approach for Ice-Water Classification Using Dual-Polarization SAR Satellite Imagery

Abstract

The daily interpretation of SAR sea ice imagery is very important for ship navigation and climate monitoring. Currently, the interpretation is still performed manually by ice analysts due to the complexity of data and the difficulty of creating fine-level ground truth. To overcome these problems, a semi-supervised approach for ice-water classification based on self-training is presented. The proposed algorithm integrates the spatial context model, region merging, and the self-training technique into a single framework. The backscatter intensity, texture, and edge strength features are incorporated in a CRF model using multi-modality Gaussian model as its unary classifier. Region merging is used to build a hierarchical data-adaptive structure to make the inference more efficient. Self-training is concatenated with region merging, so that the spatial location information of the original training samples can be used. Our algorithm has been tested on a large-scale RADARSAT-2 dual-polarization dataset over the Beaufort and Chukchi sea, and the classification results are significantly better than the supervised methods without self-training.