Research On Water Extraction Method Of High-Resolution Remote Sensing Images By Using Restricted Labeling

Water extraction from high-resolution remote sensing images is one of the important research topics in the field of remote sensing applications.Although traditional algorithms have made some progress in water body extraction,there are still problems such as cumbersome artificial features and insufficient extraction accuracy.At the same time,thanks to the rapid development of remote sensing technology,high-resolution remote sensing images contain richer texture information and clearer edge structure information.Traditional water extraction methods are difficult to make full use of high-resolution remote sensing images with rich semantic information and meet the growing needs of remote sensing applications.In recent years,with the rise of deep learning technology and the improvement of computer hardware performance,deep learning,especially convolutional neural networks,can be successfully applied to the field of remote sensing image processing.However,high-resolution remote sensing images have a large amount of data,high resolution and wide range,which makes manual pixel level annotation extremely time-consuming and laborious.Therefore,using simpler or fewer labels for water extraction has become a key issue in the field of remote sensing image applications.Combined with the characteristics of remote sensing image,this dissertation studies the water extraction method of high-resolution visible light remote sensing image in the two directions of using point label and few label supervision.It is introduced as follows:(1)Aiming at the problem that point labels lose a lot of semantic information and cannot effectively extract water bodies,this dissertation proposes a method called neighbor feature aggregation network.The method utilizes a neighbor sampler and a neighbor feature aggregation module to learn neighbor information and aggregate neighbor features,enabling convolutional neural networks to learn more representative features by utilizing neighbor features instead of global or local features.On the one hand,neighbor sampling can map adjacent pixels of remote sensing images to the same spatial position,which is beneficial to the neural network to extract adjacent features,thereby ensuring the correct mapping of spatial information.On the other hand,neighbor feature aggregation can complement feature information and extract effective semantic information,thereby improving the generation quality of pseudo-labels.Experimental results show that this method improves MIoU by 9.0 percentage points over other weakly supervised methods when using point labels.(2)Aiming at the problem that few labeled data make it difficult for convolutional neural networks to learn key features,this dissertation proposes a semi-supervised algorithm based on confidence filtering.This method retains high-confidence regions locally,eliminates lowconfidence regions,and avoids introducing too much false information in pseudo-labels,thereby achieving high-precision water extraction with a small number of labels.On the one hand,the consistency criterion can introduce random disturbances and force the model to predict the same,so that the model can learn consistency and enhance the generalization ability.On the other hand,confidence filtering minimizes the entropy of the prediction results of the model,allowing the model to learn key features.The experimental results show that the method reduces the MIoU by only 1.5 percentage points when reducing 95% of the labels.(3)Water extraction software design and application examples.We have implemented the design of water body extraction software in the development environment based on image processing platform Opencv,interactive development platform Pyqt5,artificial intelligence platform Pytorch and other development environments,and verified it with real high-resolution remote sensing images.