Sea Ice Identification in Liaodong Bay Based on Semantic Segmentation and U-Net Modeling

Liaodong Bay is one of the low-latitude ice-covered areas in China, where sea ice disasters occur frequently in winter. Accurate identification of sea ice is important for marine resource utilization and disaster prevention and mitigation. Currently, sea ice identification mostly utilizes traditional supervised classification, with cumbersome sampling steps and low identification accuracy. This article used Landsat 8-OLI images for creating the dataset, which includes 818 groups of data, of which 570 groups for the training set, 164 groups for the validation set, and 82 groups for the test set. Based on semantic segmentation with U-Net model, it achieved 92.77%, 93.03%, and 97.98% of accuracy, precision, and recall, respectively. On the testing set, the accuracy of this model is 91.67%, the precision is 92.09%, and the recall is 95.25%. Moreover, when the model was applied to identify the sea ice area in Liaodong Bay on January 19, 2018, the estimation results had small differences with the national published data. This study provides an important research idea for the identification and classification of sea ice, which can help to improve the accuracy and efficiency in the overall identification.