Research On Classification Method Of Typical Object Of Hyperspectral Image Based On Deep Learning

With the launch of domestic hyperspectral satellite and the rapid development of UAV hyperspectral remote sensing,people can more easily obtain a large number of hyperspectral remote sensing images,which are also widely used in the field of national economy and military.Hyperspectral image classification technology is the basis for further application of these hyperspectral images.However,manual acquisition of labeled samples in hyperspectral images is usually time-consuming and laborious,so the number of labeled samples that can be used for supervised training is limited.Therefore,this paper mainly studies how to make better use of deep learning method to improve the classification and recognition accuracy of typical objects in hyperspectral images in the absence of labeled samples.The main research work of this paper are as follows:(1)Aiming at the high dimension of hyperspectral image data,this paper explores an automatic band selection method for hyperspectral image based on attention mechanism.Specifically,attention mechanism is introduced to describe the importance of different bands in fully connected network,and then the whole model is trained to learn the importance of different bands based on reconstruction error,and several bands are selected according to the importance.The designed method trains the deep network model unsupervised for band selection,so no manual marking information is needed.Compared with ISSC,Spa BS,MVPCA and MOBS,the overall classification accuracy of band selection method based on deep learning(FC-BSNet)is improved by 9.26%,7.36%,9.45% and 2.59% on Pavia university data,and by7.47%,1.26%,7.06% and 0.81% on Salinas data.(2)On the basis of band selection,based on the idea of transfer learning,the deep convolution neural network model pre-trained by the Image Net is used to extract the spatial features of hyperspectral images,and the support vector machine and other classifiers are used to complete the classification task.Experimental results show that the proposed method can quickly extract the features of hyperspectral images,and the extracted features can effectively improve the classification accuracy of hyperspectral images.Classification experiments on two sets of data show that the overall classification accuracy of the feature extraction method based on Res Net18 is more than 1% higher than that of SVM,3D-CAE,3D-CNN,CNN-PPF and other methods.(3)In order to make better use of the spatial spectral information of hyperspectral image to improve the classification accuracy,the spatial spectral feature fusion method based on multi granularity scanning is studied,and the deep random forest is used to complete the classification.The proposed method only needs a small amount of label information to obtain good classification results,which effectively alleviates the problem of lack of label samples in hyperspectral image classification.The experimental results show that the overall classification accuracy of the multi-granularity scanning and deep random forest classification method is 0.27%and 0.40% higher than that of the Resnet18+SVM method on the two groups of hyperspectral images.