NFNet Based Sea Ice Classification From RADARSAT Constellation Mission Data

As a new Canadian SAR mission, the RADARSAT Constellation Mission (RCM) provides advanced beam modes and data with broader spatial coverage and a shorter revisit time than RADARSAT-1/2. In this study, a deep residual CNN called Normalizer-Free ResNet (NFNet), which achieved advanced classification performance according to the ranking list of the Imagenet in 2021, was employed for sea ice classification from the dual-polarized RCM data. Canadian weekly regional ice charts were used as references for the training and testing of the classification model. A two-level random forest (RF) classifier was trained for comparison with NFNet. In addition, estimations of each region's total concentration and ice-class proportions were verified with the corresponding sea ice chart polygon. The high accuracy of NFNet (98.18 %) proves the sea ice sensing ability of RCM and illustrates NFNet's superior performance over RF in sea ice classification.