Research On Optical Image Coastline Detection Technology Based On Deep Learning

Coastline detection using optical remote sensing images is a fundamental work in the field of remote sensing applications,which plays an irreplaceable part in environmental protection,coastal construction,automatic navigation and mapping.However,fast and accurate coastline extraction is still a challenging problem due to the complex and diverse coast types and the lack of obvious spectral features of sea-land boundaries.Traditional shoreline detection methods are based on manually designed features,which are difficult to achieve the expected results when facing remote sensing images with complex textures and intensity distributions.In recent years,the development of deep convolutional neural networks has broken the bottleneck of coastline detection.Therefore,this thesis takes the offshore coastal area of China as the study area and Landsat8 OLI images as the data source,and conducts a study on the optical image coastline detection based on deep learning,with the following details:1)A HED-UNet network based on deformable convolution is designed to address the problem of ambiguous shoreline localization caused by the simplicity of a single ordinary convolution function.First,the network fuses the semantic segmentation framework and the edge detection framework to form a multi-task model,and refines the segmentation results using additional edge detection branches to effectively improve the edge accuracy.Second,the standard convolutional layer is redesigned using the residual fusion Inception module based on deformable convolution,which enhances the feature extraction capability of the network.Finally,five upsampling and downsampling blocks are used on the encoder-decoder structure,which can obtain a larger perceptual field and richer semantic information.The experimental results show that the HED-UNet network based on deformable convolution can effectively solve the problem of weak boundaries in coastline extraction and obtain clearer boundary results.2)A WHED-UNet network combining an articulated attention mechanism and a deformable convolution is designed to solve the problem of unsatisfactory accuracy in the coastline detection task.Two improvements are made on the basis of HED-UNet in Chapter 3,and two HED-UNet networks with the same structure of deformable convolution are cascaded to effectively enhance the contour and location information of the images.Secondly,in order to focus on the sea-land boundary features in a targeted manner,an attention mechanism bridging the networks with fused space and channels is introduced,which can fully consider the effective information of different levels and achieve more accurate detection results.The experimental results show that the proposed WHED-UNet network exhibits better extraction accuracy and fault tolerance compared to other methods.