Application Of Semantic Segmentation Of Remote Sensing Image Based On Improved Unet In Surface Water Change

The land surface water body is the general name of river and lake.It is a natural complex of water-covered areas,and it is the cradle of nature and human civilization.However,since the human civilization entered the stage of industrialization,some overloads and even destructive developments have caused surface water bodies to be devastated.The change of surface water bodies becomes more and more frequent,and the effective monitoring of water body changes is a prerequisite for the efficient development and utilization of water resources.Along with the rapid development of remote sensing technology,the high-resolution image information provided by the space remote sensing platform is increasingly rich,and the spatial and spectral information of surface water bodies occupy an important part.The abundant surface water data carried by remote sensing images lays a solid foundation for the monitoring of surface water changes,but the rich data volume cannot be directly used,and it is necessary to extract effective information for efficient analysis and quantification.In this paper,the change of surface water body is taken as the research goal,and the purpose is to calibrate the change state of the remote sensing image at different times in the same area.The traditional segmentation change detection uses features such as the distribution,shape,structure,texture,and hue of the feature,and performs well when dealing with a single scene.However,the performance of the algorithm is affected when the local information is complicated or the image resolution is high,and the traditional The threshold is set by human experiment detection,and the algorithm is less robust.In order to solve this problem,this paper uses superpixel as the basic analysis unit,and uses machine learning algorithm to construct a deep learning model with multiple nonlinear transform combinations to model the advanced abstract features of superpixels to improve the accuracy of image segmentation.Great.The main contents of the paper are as follows:(1)Obtain multiple remote sensing images of different time segments in the same region and perform geometric correction,and then use the object-oriented method to mark the image with the eCognition software to obtain the ground truth.(2)In view of the fact that traditional threshold segmentation method can not effectively segment remote sensing images with high background complexity,this paper proposes a water-based transition segmentation method based on deep convolutional neural network.This method is optimized and improved on the basis of Unet network architecture.D-Unet(Deep-Unet)and DS-Unet(Deep Separable-Unet)can accurately extract water body information from remote sensing images,laying a solid foundation for subsequent research on surface water body changes and water area measurement.(3)For the deep convolutional neural network,the rough segmentation has the problems of unsmooth boundary and inaccurate pixel location.The full-join condition is used to refine the coarse segmentation result to achieve more precise and accurate results.(4)Aiming at the characteristics of remote sensing images,a method for visualizing surface water area and water migration is proposed.(5)This paper tests the effectiveness of the method for the benchmark data collected by the dataset.DS-Unet obtains an average segmentation accuracy of 88.74% and achieves 15 s for one 6000*6000px under Intel Core i7(2.2 GHz).The segmentation speed of remote sensing images;D-Unet can achieve an average segmentation accuracy of 91.59% and a segmentation speed of a 6000*6000px remote sensing image processed by Intel Core i7(2.2 GHz)for 45s;after using the fully connected condition random field optimization The result of the segmentation is more detailed.The average processing time under the Intel Core i7(2.2 GHz)is 166 s,the robustness of the environmental noise is improved,and the smoothness of the segmentation edge is higher,which is an effective method for water body extraction.At the end of the paper,the pixel-based surface water body area measurement and surface water body change visualization method proved to be a good calculation method in the experiment,and achieved good display effect.