Research On Classification Method Of Hyperspectral Image Based On Improved 3D Convolutional Neural Network

In recent years,Convolutional Neural Network(CNN)has been widely used in the field of hyperspectral image classification,in which 3D-CNN can be used to extract the spectral-spatial joint features of hyperspectral images according to the threedimensional data characteristics of hyperspectral image.3D-CNN has proved to be an effective classification method.However,there are still the following problems when3D-CNN is used for hyperspectral image classification: 1.The classification method based on 3D-CNN requires a large-scale labeled data set to train the network.The insufficient number of samples will lead to the phenomenon of overfitting of the network and reduce the classification effect.2.The spectral information is redundant and there are interfering pixels.3D-CNN will use context information when extracting features.However,pixels in the context usually contain different categories,and pixels that are different from the categories of central pixels will interfere with feature extraction.In view of the above problems,this paper conducts the following research:Aiming at the problem that insufficient number of samples will cause overfitting in the network,this paper combines data augmentation(mixup)with 3D-CNN,and proposes a hyperspectral image classification method based on M-3DCNN.Mixup is used to generate virtual data.Original data and virtual data are used to train the network together,which expands the number of samples in the training set and alleviates the3D-CNN overfitting phenomenon.3D-CNN is used for feature extraction and classification.Experiments were conducted on three general hyperspectral image classification data sets to verify the effectiveness of the method.The experimental results show that the classification effect of this method is improved compared with3D-CNN on the three data sets.Aiming at the problem of spectral information redundancy and interference pixels,the attention mechanism is introduced into the M-3DCNN network,and a hyperspectral image classification method based on the AM-3DCNN network is proposed.The useful features of the data through the attention mechanism are strengthened and the useless features are suppressed,which improves the feature extraction ability of the network,reduces the influence of spectral information redundancy and interference pixels on classification,and obtains better classification results.The channel attention mechanism compresses the relationship between spatial location information modeling channels on the global spatial location information receptive field,and the spatial location attention mechanism compresses the relationship between channel information modeling spatial locations on the global channel information receptive field.Both of the two attention mechanisms use Sigmoid function to complete the mapping of different weights.Experiments were performed on three general hyperspectral image classification data sets to verify the effectiveness of the method.The experimental results show that compared with other methods,the method has achieved the best classification effect on the three data sets.