Application Of Improved U-Net Network In The Extraction Of Inland Lake Water Body Information From Hyperspectral Images

Lakes and reservoirs water bodies are an important components of the earth’s water resources.Monitoring them quickly and effectively will help maintain the ecological balance and stability of the earth.In recent years,some conventional water extraction algorithms have achieved good results in multispectral images.However,due to the strong band correlation and redundant new information in hyperspectral remote sensing images,the conventional algorithms have poor results in extracting water bodies.Deep learning network has the characteristics of ”self-learning” of data features,and can perform effective feature extraction and fitting on high-dimensional image data,making it robust in remote sensing image information extraction.So,in this paper,the improved U-Net deep learning network is used to extract large water bodies such as lakes from hyperspectral images.One of the main tasks can be summarized in the following three aspects.Firstly,the downscaled features obtained by principal component analysis(PCA)are fused with the Normalized difference water index(NDWI)features to obtain a new multifeature image,and the randomly selected training samples are automatically labeled from the new image by using a multi-sub-region growing algorithm,thus reducing the time taken to manually creating sample labels.Secondly,in the case of training based on small samples,this study increases the convolution number of encoder and decoder in the U-Net model to obtain deeper feature information of the image,and adds batch normalization(BN)and dropout layers to normalize the input data,accelerate the training and convergence speed of the network,and inhibit the over-fitting of the training model.Finally,the mathematical morphology algorithms is used to optimize the extraction results of the improved U-Net network,reducing the over-segmentation and under-segmentation.In this study,experiments are carried out based on the Zhuhai-1 hyperspectral image,and the improved method is compared with NDWI,Markov random field(MRF),random forest(RF)and classical U-Net network.The experimental results show that the improved method can obtain the best results of lake water bodies extraction in both quantitative and visual aspects,and effectively restrain the influence of the shadow in the process of lake water bodies extraction,which provides a promising method for breaking through the limitations of hyperspectral images in lake water bodies extraction.