A Residual Convolution Neural Network for Sea Ice Classification with Sentinel-1 SAR Imagery

Abstract

Sea ice type classification is critically important for sea ice monitoring, and synthetic aperture radar (SAR) has become the main data source for sea ice classification. With a large number of SAR images produced every day, a more intelligent sea ice classification process is urgently needed. In this paper, we constructed a four-type sea ice classification dataset using Sentinel-1 SAR images with the reference of Canadian Ice Service’s ice charts and designed a residual convolution network for sea ice classification: Sea Ice Residual Convolutional Network (SI-Resnet). We further designed a multi-model average scoring strategy with the idea of ensemble learning to improve the classification accuracy between closely-associated ice types. Based on the experiments, our proposed method outperformed MLP, AlexNet, and traditional SVM methods, reaching the overall accuracy of 94% and Kappa coefficient of 91.9. For the evaluation on regional ice concentration, the values computed from the SI-Resnet’s classification results are more consistent with ice chart’s regional concentration data than those of MLP, AlexNet and SVM. Compared with the manually generated ice chart of CIS, our method can work automatically and provide more detailed ice distribution to a useful reference for ship route planning and sea ice changes monitoring.