Semantic image segmentation for sea ice parameters recognition using deep convolutional neural networks

Abstract

• A novel CNN-based model was developed for the recognition of sea-ice images. • A modified decoder by combining shallow and deep layer features. • The designed network achieves good results for the segmentation of various scale objects. • A low-cost, and automatic method for the computation of sea-ice parameters. An accurate algorithm for sea ice segmentation is critical for monitoring sea ice parameters of ship navigation in ice-covered seas, as it can automatically extract ice objects and corresponding information to compute essential parameters such as surface ice concentration and ice floe size. In this paper, based on digital images captured by onboard cameras, a novel network called Ice-Deeplab for pixel-wise ice image segmentation is proposed. The Ice-Deeplab network is constructed using the deep convolutional neural network Deeplab and is modified with an attention module and an improved decoding structure. To investigate its reliability, the Ice-Deeplab network is applied to a 320-image dataset, with 80% for training and 20% for validation. The experiments demonstrated that the proposed Ice-Deeplab yields better segmentation results than the original Deeplab model under different validation scenarios, achieving an overall accuracy of 90.5% among the classes sea-ice, ocean, and sky. Moreover, the proposed model was applied to un-labelled test data to demonstrate its generalisation ability for real-time ice segmentation.