Research On Container Cloud Oriented Water Body Extraction Method Using Convolutional Neural Network

In recent years,cloud computing and big data technology frameworks have matured,and artificial intelligence models and algorithms represented by deep learning have been widely used in various fields.With the rapid development of remote sensing sensor technology,the time resolution and spatial resolution of remote sensing data are also getting higher and higher,providing rich data source support for business monitoring applications in water conservancy,agriculture and environment.In order to effectively protect and rationally use water resources,daily monitoring of water changes such as rivers,lakes and lakes has become a key to ecological environmental protection.This paper focuses on the problem that the water extraction algorithm in the complex environment of inland cities is easily affected by the shadow of urban buildings,and the extraction accuracy is not high.A model of convolutional neural network water extraction based on spectral feature matrix is proposed.At the same time,in order to support real-time water monitoring applications,for example,to analyze the change of water area caused by heavy rain,this paper seeks to solve the problem of inefficiency of most remote sensing application algorithms from cloud computing and big data processing underlying technologies.For the first time,the convolutional neural network water extraction model was run in a cloud environment and accelerated by Spark memory parallel technology.The results of this paper show the applicability of the method for real-time monitoring of inland urban waters in the cloud environment to meet the high-precision and high-efficiency operational requirements.The work of this paper as following.(1)Research on Convolutional Neural Network Water Method Based on Spectral Feature MatrixAccording to the difference between the spectral reflection characteristics of the spectral reflectance characteristics of water bodies and other features,water extraction method based on spectral feature matrix of convolutional neural networks was designed for Sentinel-2A/B satellite data.Firstly,by analyzing the spectral information of the ground object and the spectral reflection characteristic of the water body,and considering with the specific band information of the Sentinel-2A/B satellite,the bands feasible for water extraction are determined.Then,the selected one-dimensional spectral vector is mapped to a two-dimensional spectral feature matrix by a mathematical expression.To effectively distinguish the water and other features,water bodies and urban architectural shadows,this paper adds NDWI(normalized water body index)and USI(city shadow index)as two independent band data into the spectral feature matrix.Next,the classification model is trained using manually labeled samples.Finally,the applicability of the water extraction method proposed in this paper was tested experimentally.The experimental results show that the method can effectively overcome the influence of urban architectural shadows,and the model training accuracy can reach 96.37%.However,the extraction results of the algorithm have scattered water body pixels and water body contour fracture phenomenon,and the efficiency is relatively low.(2)Optimization of water body extraction results based on object-oriented thinkingTosolve the problem that the scattered pixels and the boundary deviate from the actual contour,this paper initmatethe object-oriented remote sensing classification method.Specifically,the multi-scale segmentation algorithm is used to divide the research area into multiple scales,and the image is segmented into object maps.Then,the second step consists of taking the object map as the basic unit,combined with the convolutional neural network water body to extract the coarse result,and using the voting method to determine the category of the object map.Finally,the experimental results show that the object-oriented classification method can effectively solve the problem of scattered water body pixels,and the extraction result is more in line with the actual contour.(3)Optimization of water body extraction method based on cloud environmentTo solve the problem of low efficiency of pixel-level convolutional neural network water extraction algorithm,this paper seeks to solve the problem of low efficiency of remote sensing application algorithms from cloud computing and big data processing technology,and proposes a memory parallel processing model which useswindow aggregation functions to promote the efficiency of the Water extraction model.First,the advanced container virtualization technology is used to provide flexible container resources for the memory parallel computing framework.Then,the trained water extraction model and the memory parallel framework are packaged into Docker images and uploaded to a private image storage.Then,Kubernetes is used to manage and coordinate the allocation of dynamic computing and storage resources to provide flexible container resources for applications that invoke the water extraction model.Finally,the experiment analyzes in detail the number of image partitions and the effect of the actuator configuration of the memory parallel computing framework on the execution efficiency of the algorithm.The results show that the memory parallel city water extraction model for window aggregate functions is highly scalable and highly available.