Research On Dimension Reduction And Ground Object Classification Algorithm Of Hyperspectral Remote Sensing Images Based On Deep Learning

Hyperspectral remote sensing technology,as a frontier remote sensing technology,can provide multispectral remote sensing images containing rich spectral information by combining imaging technology and spectral technology.Aiming at the problems of high data dimension,complex feature types and limited labeled samples in hyperspectral images,this paper analyzes and studies the dimensionality reduction and classification in hyperspectral images.The specific research contents and work results are as follows:(1)Research on dimension reduction and ground object classification of hyperspectral remote sensing images based on grouping principal component analysis.In this study,GF-5 hyperspectral remote sensing image was used as the data source.Firstly,the correlation between bands was used to group the original hyperspectral images,then principal component analysis(PCA)was applied to each group of band data,and the most important principal component bands are extracted as the dimensionality reduction results.Finally,CNN fusion SVM classifier was used to realize the accurate classification of ground object types.The dimensionality reduction method proposed in this paper not only improves the classification accuracy,but also greatly shortens the network training time.Compared with the traditional PCA dimensionality reduction method,the classification accuracy is improved by about 6%.(2)Research on automatic hyperspectral optimal band selection method based on similarity grouping.Aiming at the phenomenon of redundancy and noise between bands of hyperspectral images,this study first grouped the original hyperspectral images,then used the structural similarity between bands(SSIM)to update the grouping and the clustering center of each group,and finally selected the best band in each group as the band selection result.In this paper,experiments were carried out on WHU-Hi hyperspectral dataset,using three-dimensional convolution network(3DCNN)and SVM as classifiers.The effects of different band selection numbers on classification accuracy are compared,and the optimal band selection number is given.The experiments show that the proposed algorithm can automatically give the optimal band selection results.Compared with the dimensionality reduction method based on PCA,the classification accuracy is improved by about 2%.(3)Classification of hyperspectral objects based on convolutional attention mechanism and 3D residual network.In this study,On the basis of dimensionality reduction based on optimal band selection,3D residual network(3D-Res Net)based on convolutional attention mechanism(CBAM)was constructed for the characteristics of hyperspectral data,and the effects of different image block sizes and network depth on classification accuracy were compared.Experiments on WHU-Hi dataset show that the classification results obtained by 3D-Res Net based on CBAM are better than those obtained by traditional SVM classifier,3D-CNN based on image block classification and TSRN model based on residual network.In the dimension reduction data based on optimal band selection,the classification accuracy of the network has been improved by about 2%,and the training speed has been significantly improved.