| The change of plateau lake area is affected by climate and human activities,and it is an important carrier to record the natural and man-made changes including temperature,precipitation,environment and human activities.Therefore,obtaining lake information efficiently and accurately plays an important role in studying climate change,regional water cycle and ecological protection measures.The improvement of extraction methods has always been the research focus of remote sensing image information extraction of plateau lakes,and its trend is from visual interpretation to remote sensing index,machine learning and now the development of deep learning methods.The traditional methods of plateau lake extraction have many problems,such as heavy workload,low degree of automation,dependence on a large number of human participation,and higher cost than deep learning methods.Traditional machine learning cannot make full use of spectral information and spatial information of remote sensing image,and generally only uses shallow model with simple structure to learn artificial features,so it has certain limitations in the accuracy of ground object extraction.At present,the direction to improve the extraction accuracy of plateau lakes is to use deep learning method to fully mine the rich information of remote sensing images.Because the image width,space and time resolution and data availability and other factors,the plateau lakes monitoring using satellite remote sensing image is primarily a series of Landsat satellites,but based on the depth study of remote sensing image feature extraction is mainly using high resolution remote sensing image,thus to break through the deep learning technology in moderate resolution remote sensing image extract the limitations in the plateau lakes,It is necessary to find a deep learning method to extract plateau lakes from medium resolution remote sensing images.Deep network architecture U~2-Net has high overall accuracy and finer boundary extraction after migration,but the details in the extraction results are not good.Therefore,in order to further improve on the plateau lakes in moderate resolution remote sensing image extraction accuracy,based on the plateau lakes U~2-Net network internal fine part of discrimination and extract the effect not beautiful missions,combined with the U-Net in detail such as the island of qinghai lake in partial extraction performance,to qinghai lake boundary extraction and U~2-Net performance,Double U-Net was selected to extract plateau lakes from medium resolution remote sensing images.The improved Double U-Net based on U-Net fusion image depth features has strong robustness and generalization ability.The SE module added in double U-Net can use weighting to strengthen the most critical features to achieve the purpose of automatic discrimination and learning of important features,so as to improve the extraction accuracy of plateau lake.In this paper,medium resolution remote sensing image plateau lake data set was made for experiment,and the experimental results show that:(1)U~2-Net has a deep network architecture,and its structure is a two-level nested U-shaped structure.The pooling layer of RSU block in its bottom structure increases the depth of the whole network,so it can obtain more characteristic information of plateau lakes.The overall accuracy of prediction results of Namco Lake reaches 97.53%.Higher than U-Net network,the deep network architecture can obtain more abundant image information,so it can improve the precision of semantic segmentation.(2)Although the prediction results after U~2-Net network migration to Qinghai Lake have high accuracy,the details in the extraction results are not good,and the details such as islands inside Qinghai Lake are not fully extracted.U~2-Net cannot effectively improve the accuracy of lake extraction only by acquiring more and more detailed image features of plateau lakes,and is not suitable for the extraction of fine parts of plateau lakes in medium resolution remote sensing images.The inner fine part of Qinghai Lake is more complicated and there is some difficulty in extracting it.Therefore,only improving the model's ability to obtain image features is not suitable for the discrimination and extraction of inner fine parts.(3)Double U-Net combines the advantages of U-Net network structure and SE module,with U-Net network image features after encoding and decoding and network deep and shallow semantic features fusion characteristics;The addition of SE module weighted the original feature map at the channel level,so that the network can automatically learn the output features of different channels according to their importance.The overall accuracy of double U-Net prediction results for Qinghai Lake is 98.55%,and the other precision indexes are better.Therefore,Double U-Net has the ability to independently distinguish and learn important features of plateau lakes in medium resolution remote sensing images,and can complete the task of plateau lake extraction.(4)Take Qinghai Lake as an example to verify the generalization of double U-Net model in the long time series extraction task,predict the remote sensing image of Qinghai Lake from 2000 to 2021,and analyze the spatio-temporal changes of qinghai Lake area according to the prediction results.The temporal variation trend of Qinghai Lake area was that the lake area decreased from 2000 to 2003,and increased from 2003 to 2021.In the spatial variation,the greatest changes in the boundary of Qinghai Lake are mainly Shadao Lake and Haiyan Bay on the east bank of Qinghai Lake,the entrance of Shaliu River on the north bank and Bird Island and Tibuka Bay on the west bank.In addition,the boundary of Qinghai Lake showed a trend of outward expansion from 2000 to 2021,and the boundary of Qinghai Lake changed more dramatically after 2009.The results show that the Double U-Net model can provide data support and basis for relevant policies and management,and has the value of popularization and application. |
PDF Full Text Request