Water Extraction Research In Urban Areas Combined Deep Learning With Vector Constraint

With the development of social economy,the problem of water pollution is becoming more and more serious.At present,water pollution incidents in China mainly occur in cities or urban agglomerations,and the clear distribution of water in cities has become an important prerequisite for local departments to carry out relevant work.Aiming at the application demand of high spatial resolution remote sensing water extraction,this paper studied the automatic extraction technology of urban water information on high spatial resolution remote sensing image.The main work of this paper is as follows:(1)Water extraction based on object-oriented convolutional neural network and vector constraints.Combining object-based image analysis technology with convolutional neural networks in deep learning,using image segmentation methods to obtain object primitives,and building the correspondence between segmented object primitives and model block primitives;collecting samples and training deep learning models,The model is used to predict the classification block primitives,and the prediction results of the classification block primitives are mapped to the segmentation primitives to obtain the water body prediction probability of the segmentation primitives.Then,the overlapped analysis is performed on the segmented object primitives and the existing vector data to obtain the overlap relationship between them,and then the overlap relationship between the vector data and the object primitives and the model prediction results,area and spectral characteristics of the object primitives,constructing extraction rules,combining model prediction results with extraction rules to achieve refinement of water body extraction results.(2)Water extraction with full convolutional neural network and vector constraints.The full convolutional network is used to extract water body.Firstly,the water body information of the sample area is annotated,the sample area and its mask layer are converted into a labeled sample,the model is trained to obtain a semantic segmentation model,and the image is predicted by the model.Then,the vector data and the prediction probability map are superimposed to establish a water body extraction knowledge rule to achieve refinement of water body extraction results.The high-resolution remote sensing images are selected and conducted experiments on the extraction of water information in urban areas.The results show that both deep learning methods can effectively remove the effects of shadows and buildings,and have higher extraction accuracy.The former uses the relationship between constraint vectors and segmented patches to construct extraction rules,and combines the model prediction results of segmented patches to achieve refinement of water extraction results.The method combines object-oriented technology,which makes the sample selection more convenient,but also causes the model to rely heavily on the segmentation accuracy.The latter uses the spatial relationship between the constraint vector and the pixels,and uses high and low probability thresholds to achieve the refinement of water extraction results.The method sample selection requires manual sketching,the process is more complicated,but the model has certain advantages in extracting small water bodies.