Arctic Sea Ice and Open Water Classification From Dual-Polarization Synthetic Aperture Radar Imagery and Deep Learning Models

Abstract

In this study, we present a deep learning (DL)-based method for classifying Arctic sea ice and open water using spaceborne C-band RADARSAT-2 dual-polarization (HH, HV) synthetic aperture radar (SAR) images. The HH- and HV-polarization radar. backscatter and cross-polarization ratios (HH/HV) are used as input for the DL model. First, we combine an unsupervised clustering method with an adaptive thresholding technique to generate accurately labeled samples, thereby minimizing subjective errors and reducing the time required for visual interpretation. Second, we employ an enhanced U-Net architecture to develop the proposed classification method. The modified Atrous spatial pyramid pooling module is integrated into a customized dilated U-Net to create a multiscale feature extraction model (MS-DUNet). MS-DUNet is then trained and validated using 11 485 labeled patches extracted from 3565 RADARSAT-2 SAR images. In addition, 5004 SAR images are used to test the model. Compared to representative DL models, MS-DUNet demonstrates better performance, achieving an overall classification accuracy of 99.3% and an intersection over union value of 98.6%. Furthermore, we compare MS-DUNet's predictions with the daily sea ice extent data from the Interactive Multisensor Snow and ice mapping system, achieving an average accuracy of 91.9%. The results indicate that the proposed approach effectively distinguishes between sea ice and open water, using wide-swath SAR imagery. The method also demonstrates strong performance in the complex marginal ice zone during the melting season, and in distinguishing sea ice leads from the surrounding areas.