Study On Hyperspectral Image Classification Based On Generative Adversarial Network

Hyperspectral images are obtained by continuous scanning or shooting of hyperspectral imaging equipment in a certain band range,which contains rich spectral and spatial information of ground objects.Hyperspectral image classification is a basic and important research direction in the field of remote sensing.It uses modern scientific methods to classify each pixel in the obtained hyperspectral image.However,the insufficient number of hyperspectral image samples is still the main problem that limits the further improvement of classification accuracy.In addition,in hyperspectral images,the spectral information of some ground objects is very close,and different ground objects may reflect similar spectral curves,The spectral curves reflected by the same ground objects may also be different.For the characteristic that hyperspectral images contain both spectral information and spatial information,how to make full use of these information data and deeply mine their potential information features is an important research direction to improve the classification accuracy of hyperspectral images.Aiming at how to deal with these problems,this paper aims at the lack of prior data samples,based on the method of deep learning,This paper proposes a method to fully mine the spatial spectral information of hyperspectral image data by using the generation countermeasure network.（1）Aiming at the problem of insufficient a priori samples in hyperspectral images,a hyperspectral image classification method based on Ga N window generation is proposed.The generator and discriminator are designed to generate the window data with the target pixel as the reference point through confrontation learning,and use the generated large number of samples to make up for the problem of insufficient samples.The generated samples are put into the classification network for training and verified and analyzed by experiments,Compared with the traditional method,the classification accuracy of this method is improved.（2）Aiming at the problems of complex hyperspectral image data,more noise and high redundancy,a hyperspectral image classification method based on Ga N band is proposed.The generation countermeasure network is used to learn each band of the whole hyperspectral image,and the whole hyperspectral image with relatively clear features and relatively high quality is generated.Finally,the generated band samples are accumulated with the original dimensionality reduction data samples,and the 3dcnn network is used to complete the category output,Through experimental verification,the classification accuracy of hyperspectral images after being put into the generated band is improved.（3）Aiming at the problem of how to make full use of the information contained in hyperspectral images,a method of feature extraction and generation confrontation based on spectral information and spatial information of hyperspectral images is proposed to realize classification.One channel extracts spatial information through 2DCNN,the other channel extracts spatial information through 1DCNN,and the last two channels output results at the same time through class consistency loss,This method makes full use of the limited spectral information and spatial information to output high-precision classification results in the case of insufficient samples.（4）Aiming at the problem of how to use spectral and spatial information concisely and efficiently,it is different from the previous methods of using two channels to extract respectively.Instead,3dcnn is used to extract the features of hyperspectral image data,which can obtain spectral information and spatial information at the same time.In addition,through the introduction of multiple attention mechanism,the spectral space is strengthened and the weight of feature information is highlighted.The temporal attention mechanism and spatial attention mechanism perfectly fit the spectral information and spatial information of hyperspectral images.Compared with the hyperspectral image classification methods in previous chapters,the algorithm in this chapter obtains the best classification effect in solving the problem of insufficient samples.