Research On Hyperspectral Image Classification Based On Stereo Convolutional Neural Network

In recent years,the rapid development of hyperspectral remote sensing technology has provided convenience for the production and life of the society.Because hyperspectral images have the characteristics of "map integration" and high spectral resolution,they are widely used in important fields such as military target detection,agricultural development,urban planning,and environmental governance.Hyperspectral image classification has gradually become a hot issue in the field of hyperspectral image processing.Due to the problems of low spatial resolution,spectral information redundancy,and complex distribution of ground features in hyperspectral images,hyperspectral image classification still faces greater challenges.Based on the existing convolutional neural network,this paper makes full use of the rich information and spatial correlation of hyperspectral images.Under the condition of low sampling rate of training samples,the classification accuracy of hyperspectral images is further improved and the classification time is reduced.the complexity.The research content of this article mainly includes:(1)Aiming at the problem of insufficient spatial and spectral feature extraction in current hyperspectral image classification methods based on convolutional neural networks,this paper proposes a hyperspectral image classification model based on cubic convolutional neural networks,through principal component analysis and one-dimensional The convolution joint dimensionality reduction strategy effectively reduces the dimensionality and feature extraction of the hyperspectral image,and then performs convolution operations on the three planes of the hyperspectral data cube to fully mine the spatial spectral information of the hyperspectral image.Compared with the three-dimensional convolution,The proposed cubic convolution is more flexible in the parameter training process,and the convolution kernel is also smaller,which obviously speeds up the training speed.Through experimental comparison,the accuracy of classification is improved,which proves the advantages of the proposed algorithm in classification.(2)In view of the problem that the current hyperspectral image classification methods based on convolutional neural networks are difficult to capture the spatial pose characteristics of objects,and the principal component analysis ignores some important information when retaining few components,this paper proposes an extension based on Multi-morphological attribute contour feature(EMAP)cubic capsule network model to complete the hyperspectral image classification task.EMAP features can effectively extract the contour features of various attributes of entities in hyperspectral images,and the capsule neural network can effectively capture complex spatial features such as the poses of ground objects.The proposed algorithm firstly introduces the EMAP algorithm to extract the morphological attributes of the principal components extracted by the principal component analysis.The processed feature map is used as the input of the network,and the spectral and spatial lower layers of the hyperspectral image are extracted by cubic convolution.Information,using the initial capsule layer and the digital capsule layer to extract the high-level information of the hyperspectral image.Through experimental comparison,the superiority of the proposed algorithm in classification is proved.