Research On Water Body Extraction From Remote Sensing Image Of Lakes In Cold And Dry Regions Based On Fully Convolutional Neural Network

With the development of remote sensing technology,the accuracy and resolution of remote sensing images have been greatly improved,which has become an important data source for monitoring lake ecological environment in cold and arid regions.Lakes in cold and dry regions are frozen for a long time every year.Their surface features vary significantly with the seasons,showing more complex characteristics on remote sensing images.Moreover,there are a lot of mountains and shadows in the background.Therefore,how to extract the water body in the cold and arid areas efficiently and properly becomes an urgent problem to be solved.The Fully convolutional neural network model in deep learning can learn the deep features of the image,and it can classify the pixels in the image one by one to achieve end-to-end semantic segmentation.Therefore,based on the fully convolutional network model,this thesis proposes two new water body extraction methods.Two typical lakes in cold and arid regions,Wuliangsuhai Lake and Hulunhu Lake are selected as the research objects,and their remote sensing image features are utilized to verify the proposed methods.The main work of this thesis are as follows:(1)Build a Semantic Segmentation Dataset.The existing public data sets are all water bodies with relatively uniform colors,which are not suitable for extracting water bodies in Wuliangsuhai and Hulun Lake.Therefore,this thesis first obtains the Landsat8 OLI remote sensing images of the Wuliangsuhai Lake and the Hulunhu Lake,and then performs a series of preprocessing.After that,use a part of the images to make a semantic segmentation dataset suitable for Wuliangsuhai and Hulun Lake,and enlarge the dataset.(2)Propose a new method based on Fully Convolutional Network(FCN)and the GrabCut algorithm.First,the trained FCN model is initially used to extract the water body.Then,treat the water body as the foreground and the area outside the water body as the background.Using the GrabCut algorithm to segment the image again to obtain the complete water body extraction result.(3)Propose an improved U-net network,named as AU-net.This method adds the Convolutional Block Attention Module(CBAM)to the backbone feature extraction network of the U-net model.The obtained new network has the adaptive ability and pays more attention to the water body in the image.Then AU-net uses the Fully Connected Conditional Random Field for back-end processing to optimize the extraction results.(4)The extensive experiments has been carried out using Wuliangsuhai Lake and Hulunhu Lake images that did not participate in the model training in order to verify the efficiency of the proposed methods.In comparison with the traditional method NDWI,the semantic segmentation model Deeplab v3+ and the original Unet model,the proposed methods improve accuracy significantly.The two methods can eliminate the interference of mountains and shadows in the background,and reduce the misdivision and leakage of the water body.Moreover,these two methods have achieved good extraction results for the remote sensing images of Wuliangsuhai Lake and Hulunhu Lake in different seasons.