Research Of Hyperspectral Image Classification Algorithm Based On Attention Mechanism And Sample Expansion

At present,convolutional neural network(CNN)is widely used in the field of hyperspectral image classification.Among them,3D convolutional neural network(3DCNN)has the advantage of simultaneously extracting the joint information of spectral dimension and spatial dimension of hyperspectral images.However,the following problems still exist when classifying networks based on 3DCNN: 1.As3 DCNN is a deep network structure,a large number of training parameters will be generated in the training process,while the number of mark samples of hyperspectral images is small,it is easy to produce the phenomenon of training overfitting caused by insufficient training in the training process,which reduces the classification accuracy;2.There is a high redundancy among hyperspectral image data.When extracting features,3DCNN will not only consider the favorable discriminant information for the current feature category,but also consider the non-discriminant information that interferes with the current feature category classification,thus affecting the classification effect.3.The classification methods based on 3DCNN usually take a long time to train,and the classification speed is too slow.To solve the above problems,this paper conducted the following research:First of all,due to the small sample size of hyperspectral image labeling,it is easy to produce the phenomenon of training overfitting,this paper uses Mixup algorithm to construct hyperspectral image virtual dataset to expand the original data.The expanded data amount is twice that of the original data amount,which to a large extent alleviated the training overfitting phenomenon caused by the small sample characteristics of hyperspectral images.Secondly,in order to solve the problem of information redundancy and insufficient feature extraction capability of 3DCNN in hyperspectral images,the structure of 3DCNN was improved,and a convolutional block attention module(CBAM)is added between each 3D convolutional layer and Re LU layer,and a total of three CBAMs are used.Through network training,discriminant features in spectral and spatial dimensions of hyperspectral images can be highlighted and non-discriminant features can be suppressed,thus improving the role of discriminant features in recognition.Combined with Mixup algorithm,a hyperspectral image classification method based on M-3DCNN-Attention was proposed.The comparative experiments are conducted on three hyperspectral data sets(Indian Pines,University of Pavia,and Salinas),and the overall accuracy of M-3DCNN-Attention is 99.90%,99.93%,and 99.36%,respectively,which is better than the comparative methods.Finally,aiming at the problem that the classification speed of hyperspectral image classification method based on M-3DCNN-Attention is slow,this paper takes Hybrid SN network as the basis,reduces the number of layers of 3DCNN network and replaces the traditional 2DCNN in Hybrid SN with a two-dimensional depth separable convolutional network,which can effectively reduce a large number of training parameters generated by 3DCNN.To some extent,it can further alleviate the overfitting phenomenon.Then,combining Mixup algorithm and convolution block attention module,a hyperspectral image classification method based on M-Hybrid SN(DS)-Attention is proposed.The comparison experiment of classification accuracy on Indian Pines and Salinas data sets show that the overall accuracy of the proposed method is slightly higher than that of M-3DCNN-Attention,which is99.95% and 99.57%,respectively.In the classification speed comparison experiment,the training speed of M-Hybrid SN(DS)-Attention is significantly improved compared with that of M-3DCNN-Attention.