Research On Automatic Extraction Method Of Instantaneous Coastline Based On Improved CycleGAN Model

The coastline is an essential part of China’s national resources,and under the influence of natural conditions and artificial development activities,the coastline has been in a state of flux and has a sinuous and undulating form.The coastal zone is rich in natural resources,including biological,mineral,energy,and land.Coastal cities have long been using the seabed to develop aquaculture,and maritime shipping and fishing are also developing,driving the economic development of coastal cities.China is a large marine country.In building a solid marine country with all efforts,the precise extraction of the coastline is very significant for developing China’s marine resources,ecological,environmental protection,and national defense construction.The pixel-level sea-land segmentation of remotely sensed images is a fundamental task for transient coastline extraction.Traditional shoreline measurement uses field survey,which is inefficient and accompanied by certain dangers.With the development of remote sensing technology and the abundance of image data,the study of shoreline extraction has developed from the original visual interpretation method to the automatic computer interpretation method.However,these methods still consume high labor costs and are susceptible to subjectivity,noise,and background complexity.Moreover,if the traditional single feature is used for sea-land segmentation,the segmentation effect does not perform well and will lead to unclear extracted sea-land boundaries,which is not conducive to the work of high-resolution remote sensing image resolution.With the rise of big data and the improvement of computing resources,deep learning is widely used in computer vision because of its robust feature extraction ability and the ability to fit complex problems.Research on using deep learning to achieve automatic coastline extraction is also being carried out.At present,the extraction task of coastline from remote sensing images is mainly implemented in deep learning using image segmentation techniques.Existing deep learning-based image segmentation algorithms need to rely on many pairs of remotely sensed images and manually accurately labeled images for training,but such publicly available datasets are currently scarce and more difficult to obtain.In this paper,remote sensing images and sea-land segmentation images are regarded as two forms of image representation.The image translation technique is used to realize the conversion from remote sensing images to sea-land segmentation images for coastline extraction.Due to the dynamic changes of coastlines,it is not easy to obtain accurate coastline marker datasets.In this paper,we use Google Aerial photo-Maps paired samples and construct a new paired dataset after the sea-land binarization process of Google Maps.In this paper,we propose DAM-Cycle GAN based on the Dual Attention Mechanism based on the Cycle Generative Adversarial Network(Cycle GAN)model to address the problem of undersized samples in the new dataset.In this paper,we focus on the extraction of transient shorelines,so all the shorelines mentioned in this paper are transient.The specific research contents are as follows:(1)By studying the role of Cycle GAN loss function in model training and analyzing the advantages and disadvantages of commonly used loss functions for generating adversarial models,WGAN-GP and WGAN losses are selected to replace the original LSGAN and GAN losses as the adversarial losses of the new model,and for the problem of detail distortion in the images generated by the Cycle GAN model,an improved cycle consistency loss is proposed that improves the quality of sample generation and enhances the stability of the model by fully considering the structural relationship between the source and target images.(2)Since publicly available remote sensing images and manually accurately labeled data are challenging to obtain,this paper selects the data eligible for experimental research based on the Cycle GAN public dataset Aerial photo-Maps dataset.It constructs a new database by sea-land binarization processing.(3)By studying the role of the dual attention mechanism and discussing the location of the introduction of the dual attention module,the best experimental results are obtained by introducing the dual attention module between the deconvolutional and convolutional layers of the generative network decoder through the experimental comparison of several schemes.The average cross-merge ratio MIo U reaches 91.63%,which is about 3.5%higher than that of the original Cycle GAN model,thus proving that the Cycle GAN algorithm is based on the effectiveness of the Cycle GAN algorithm based on the dualattention mechanism is thus demonstrated.(4)Finally,ablation experiments and comparison experiments were designed to verify the effectiveness of the DAM-Cycle GAN model.In the three evaluation indexes of mean square error,average pixel accuracy,and average cross-merge ratio,the experimental results compared with those of the fully convolutional neural network model and Deep Lab model under the training of multiple scale data sets,the sea-land binarized images transformed by the improved model are more consistent with the truevalue images.The MIo U values are improved by at least 7% and 6%,respectively,thus verifying the effectiveness and feasibility of the proposed method.