Sea Ice Segmentation Of SAR Imagery With Fusion Of Semantic Features And Edge Information

Accurately distinguishing between sea ice and water is an important guarantee for safe navigation and marine activities in ice zones.Synthetic Aperture Radar(SAR)images have become an effective means of sea ice monitoring.However,the sea ice segmentation based on SAR images still faces significant challenges.Firstly,due to the physical properties of sea ice and meteorological conditions,the ice-water boundary presented in SAR images is complex and dynamic.As a result,many deep learning-based image segmentation algorithms struggle to achieve accurate segmentation of the ice-water boundary in SAR images.Secondly,supervised-learning segmentation algorithms are highly dependent on human interpretations of sea ice.How to effectively utilize a large number of unlabeled sea ice SAR images to improve the accuracy and real-time performance of sea ice segmentation is a worth area of research.The main focus of this study is to fully exploit the deep features of SAR sea ice images by utilizing the edge information and semantic features of sea ice,in order to improve the effectiveness of ice-water segmentation and the segmentation of highly dense sea ice.The specific work is carried out in two aspects:(1)In response to the high-precision automatic segmentation requirements for SAR image ice-water scenes,a SAR ice-water segmentation network that combines edge information and semantic features,E-MPSPNet,is proposed.This network better integrates ice-water semantic features by introducing a multi-scale attention mechanism,and an Edge Supervision Module(ESM)is designed to learn ice-water edge features.ESM not only provides ice-water edge predictions,but also imposes constraints on multiscale semantic feature extraction to make it include richer edge information expression.In addition,a loss function is designed that simultaneously focuses on ice-water edge loss and ice-water semantic loss,to achieve overall segmentation network optimization.Experimental results on the AI4Arctic/ASIP sea ice dataset show that compared to other commonly used segmentation models,E-MPSPNet performs the best,with an accuracy of 94.2%,an F-Score of 93.0%,and an MIoU of 89.2%.Furthermore,E-MPSPNet has a relatively small model size and faster processing speed.The application of E-MPSPNet in a single SAR scene demonstrates its potential business applications in drawing nearreal-time ice charts.(2)To reduce the dependence on expert interpretation of ice charts and improve the real-time performance of SAR sea ice image segmentation,a semi-supervised learningbased E-MPSPNet+ network is proposed to achieve multi-concentration sea ice segmentation(sea ice concentration is divided into five categories: <1/10,1-3/10,4-6/10,7-8/10,9/10+).The network consists of a segmentation network and a discriminator network.The segmentation network improves on the basis of E-MPSPNet by using its edge supervision module to extract rich boundary information of sea ice.The discriminator network infers pseudo labels from unlabeled images to help achieve semisupervised learning and further improve the performance of the segmentation network.Without increasing the annotation workload,this network can obtain more accurate segmentation results,especially for sea ice with similar concentrations.Experimental results on the AI4Arctic/ASIP sea ice dataset show that this model has the best segmentation performance compared to other commonly used segmentation models.In the fully supervised training mode,E-MPSPNet+ reaches accuracy of 0.633,F-Score of0.722,and MIoU of 0.730;while in the semi-supervised training mode only using 1/2 of the labeled data,the accuracy,F-Score,and MIoU are 0.640,0.730,and 0.722,respectively.Moreover,the semantic segmentation model under the semi-supervised learning framework has better segmentation ability and robustness to adversarial interference.The main contributions of this study are as follows: Firstly,the proposed ice-water scene segmentation network,E-MPSPNet,effectively captures the edge features of ice and water.Based on the idea of deep supervision,the edge supervision module constructed by the network can directly predict the ice-water edge feature map and provide additional edge constraints for feature extraction.Meanwhile,E-MPSPNet takes into account class imbalance problem between edge pixels and non-edge pixels and combines edge loss with semantic loss for network optimization.Secondly,a multiconcentration sea ice segmentation network,E-MPSPNet+,is proposed.With the assistance of semi-supervised learning methods,the segmentation model relying on a small amount of labeled data has better segmentation ability for different sea ice concentration recognition than the model trained by fully supervised learning methods and achieves optimization in edge refinement processing.