Research On The Segmentation Method Of Sea Area Monitoring Elements Using Deep Learning

Sea area monitoring has an important supporting role in protecting the marine environment,preventing marine disasters,and protecting the offshore economy and other aspects of security.Traditional sea area monitoring is to detect and analyze single,or multiple monitoring elements of the ocean using techniques such as field measurement,image processing or remote sensing observation,but there are disadvantages such as low efficiency,high cost and limited monitoring coverage area.Therefore,how to effectively use advanced remote sensing technology,and computer vision technology to study the theory and method to improve the accuracy and efficiency of sea area monitoring has important scientific significance and application value.The "air,sky,ground,and bottom" three-dimensional remote sensing observation technology provides multimodal remote sensing data for sea area monitoring;the new computer technology provides technical feasibility for the rapid extraction of sea area monitoring elements.In this paper,with the goal of improving the level of intelligent monitoring of sea area elements,we take the sea area monitoring elements such as island waterlines,mangroves and Spartina alterniflora Loisel as examples,and study the semantic segmentation method of sea area monitoring elements using deep learning.Details of the study are as follows.(1)The island waterlines segmentation model(DANet-SMIW)with improved DANet is proposed.There are problems such as low extraction efficiency and poor accuracy when using deep learning methods for island waterlines segmentation in remote sensing images.In this paper,we propose the DANet-SMIW semantic segmentation model.Firstly,adopting NDWI and OTSU to expand the input channels of the segmentation model to improve the model’s capability to analyze the spectral features of remote sensing images;Secondly,improving the DANet backbone network to improve the feature extraction ability and efficiency of the model for island waterlines,especially for small-area islands.Thirdly,Introducing the boundary optimization module and improving the loss function to optimize the segmentation accuracy of the model for island waterlines and reduce the phenomenon of missed segmentation and mis-segmentation.The DANet-SMIW model is compared with the FCN-32 s,Deep Labv3+,PSPNet,Dense-ASPP,PSANet,ICNet,and Du Net semantic segmentation models.The results show that the proposed DANet-SMIW model has higher accuracy and efficiency in segmenting island water edges with 99.08%m Io U,96.36% PA,and 12 frames per second FPS.(2)Segmentation model of mangroves and Spartina alterniflora Loisel(SwinUper Net)with improved Uper Net is proposed.The simultaneous segmentation of mangroves and Spartina alterniflora Loisel from remote sensing images has the problems of unbalanced sample distribution and low segmentation accuracy.In this paper,we propose the Swin-Uper Net semantic segmentation model.Firstly,designing a data fusion module to make full use of the multispectral characteristics of remote sensing images;Secondly,selecting Swintransformer as the backbone network and proposing the boundary optimization module to improve the segmentation accuracy of the model for mangroves and Spartina alterniflora Loisel;Thirdly,designing the image pre-processing method and replacing the loss function with the combination of Cross-entropy loss and Lovasz softmax loss to solve the problem of unbalanced distribution of samples in different categories,which affects the segmentation accuracy.The proposed Swin-Uper Net model is compared with other models,including PSPNet,PSANet,Deep Labv3,DANet,FCN,OCRNet,and Deep Labv3+,to evaluate their segmentation performance.The results show that the Swin-Uper Net model has higher segmentation accuracy and efficiency for mangrove and Spartina alterniflora Loisel,where m Io U is 90.0%,PA is 98.87%,and FPS is 10 frames per second.In summary,this paper proposes a semantic segmentation model of island water edge(DANet-SMIW)and a semantic segmentation model of mangrove and Spartina alterniflora Loisel(Swin-Uper Net)using deep learning methods.And the research results provide the new method to realize quasi-real-time and high-precision segmentation of sea area monitoring elements,and this can provide the technical support and theoretical basis for integrated sea area management and marine ecological environment protection.