Water Body Identification Based On Multispectral Remote Sensing Imagery Based On Deep Learning

Water-body segmentation is a basic operation for many important applications such as water resources allocation,ecological evaluation,and flood control.The current mainstream segmentation algorithms based on deep learning are far from satisfactory in terms of water and land segmentation.On the one hand,due to the complex background of remote sensing images,diverse features,and rich details,traditional semantic segmentation algorithms cannot be applied to remote sensing images,and the segmentation accuracy is low.Especially in the complex natural water boundary,it is easy to cause pixel misclassification and water body details.There are few reservations;on the other hand,since the current water recognition algorithms usually use fully-supervised image segmentation methods,a large number of pixel-level annotations are required to improve the accuracy of the segmentation,which restricts the improvement of model segmentation performance and the generalization ability of the model.It is necessary to study The application of weak supervision method and the feasibility of water body identification can improve the efficiency of water identification.In order to solve the above problems,this paper firstly proposes a strong supervised water recognition method based on feature pyramid enhancement and pixel pair matching.This method constructs a multi-scale feature enhancement subnet,and obtains a shallow feature pyramid of the corresponding scale through the input image pyramid,so as to retain relatively complete image structure information and transmit it to the backbone network,thereby alleviating the common loss of details in the deepening network problem.In addition,a new loss term is used for each pixel,so that the network learns from the classification results of the selected adjacent pixel group to smooth out small local errors.Experimental results prove that this method has achieved better segmentation results on self-made data sets.Secondly,in view of the time and economic cost of pixel-level labeling,the feasibility of weakly supervised semantic segmentation applied to water recognition is studied,and a weakly supervised water recognition algorithm combined seed,expansion and boundary loss(SEC)that based on location information is proposed.This method first obtains the rough distribution of key areas through the classification network,and then filters them to obtain accurate water position information.The position information is used as the SEC weakly supervised semantic segmentation network Seed pixels solve the shortcomings of previous weak-supervised methods that do not make full use of location information,and its segmentation performance is significantly improved compared with other weak-supervised semantic segmentation methods.The main contents of this article are as follows:1)Based on the water distribution characteristics of the Dongting Lake basin,the original remote sensing image data is obtained as the research area,and after a series of operations such as data preprocessing,semantic labeling,data segmentation,and data enhancement,it is constructed to be available for end-to-end training The total number is10,000 pairs of remote sensing images and the corresponding label map of the water body identification data set.2)Aiming at the problem that the current strong-supervised image segmentation method can easily lead to inaccurate classification and few boundary details in the complex natural land and water boundaries,a feature pyramid enhancement model is proposed based on the existing spatial information enhancement method to prevent The details of the water body are omitted in the deep convolutional neural network.At the same time,the pixel pair matching loss function is constructed,and the pixel similarity is introduced to obtain the connection between pixels,which reduces the classification error of the pixel water body boundary.3)In view of the insufficient training data and the high cost of pixel-level labeling,and the existing weak supervision methods have low segmentation accuracy and low practicability,a SEC weakly supervised water identification method based on location information guidance is proposed,and the classification network is used to identify The approximate area of the water body is filtered to obtain the position information of the water body,which is used as part of the supervision information of the SEC segmentation network,that is,the seed pixel.Finally the water segmentation performance is improved through the three loss functions of seed loss,expansion loss,and boundary constraint loss.