Classifying Sea Ice Types with a U-Net Model from Dual-polarized Sentinel-1 Images and GLCM Texture Feature

Abstract

Classifying different types of sea ice plays a significant role in marine navigation, especially on the Northwest Passage in the Arctic. In this paper, a U-Net model is trained to identify multi-year ice (MYI), first-year ice (FYI), and open water (OW) from synthetic aperture radar (SAR) images. 14 Sentinel-1 A/B Extended Wide (EW) mode images, acquired in the Beaufort Sea and Svalbard between January and March 2020, and their gray level co-occurrence matrix (GLCM) texture feature are split into chips to be fed into the model for training and testing. The GLCM is a set of images reflecting the comprehensive information of direction, adjacent interval, and variation amplitude of the gray level of original image. It can also be understood as texture features of original image. It is the basis of analyzing local patterns in an image and their arrangement rules. Dissimilarity of HH (HH-Dis) chosen from 20 kinds of available GLCM texture features is combined with horizontal-vertical (HV) polarization, horizontal-horizontal (HH) polarization, and incident angle as inputs. The LabelMe software creates labels with visual interpretation. The proportions of FYI and MYI data in training data were adjusted to approximately equal, to avoid bias of model prediction results towards dominant categories. The result shows that the combination of HH, HV, incident angle and HH-Dis has the highest precision of 92.10%, although the precision is 89.35% when only HH and HV information combined. This result reflects the potential of adding GLCM for sea ice classification.