Small Sample Classification Of Hyperspectral Image Based On Generative Adversarial Network

Hyperspectral remote sensing can acquire hyperspectral images that contain rich surface coverage information,which can be used in national economy construction and national defense and military informationization.The fields all have broad and far-reaching development potential.Specifically at the application level,the hyperspectral image classification task is one of the core techniques for its application.In recent years,the increasing spectral resolution of hyperspectral remote sensing systems and the increasing quantization depth of imaging equipment make the acquired images become more and more sophisticated,but also cause many challenges to the classification task.To solve the problem of easy overfitting and poor generalization in smallsample conditions with depth learning-based classification methods,the paper focuses on how to improve the accuracy of hyperspectral image classification in small-sample conditions using a generative adversarial network based on band selection.The main contents and innovations of the thesis are as follows:1.Aiming at the feature extraction method will destroy the band correlation or even lose key discriminative information during the transformation of hyperspectral data,a hyperspectral image band selection method combining orthogonal subspace projection and information divergence is proposed based on the maximum ellipsoidal volume principle band selection method.By optimizing the maximum ellipsoidal volume band selection method,the fast version of it is obtained and the generalized band selection objective function is generalized,then the information divergence is introduced as a measure of the band information,and the band selection algorithm combining orthogonal subspace projection and information divergence is obtained.The algorithm effectively reduces the spectral dimension of the original image while preserving the physical significance of the original data features,and the computational complexity is low,which can achieve better classification accuracy while effectively reducing the dimension of the hyperspectral data,and can give the recommended number of the optimal number of bands2.The idea of generative adversarial is introduced to design a deep convolutional generative adversarial network structure for hyperspectral image classification.The network removes the fully-connected hidden and pooling layers,where the generator takes fractional-strided convolution as an up-sampling strategy and obtains fake samples with random noise and category labels of the samples as input.Both the discriminators and classifiers adopt strided convolution as a downsampling strategy,with generated fake samples and true training samples as inputs to discern authenticity and category labels.The parameters are optimized according to the multicategorization loss in the training,and the loss function can be optimized more reasonably than the traditional generating adversarial network.The real probability distribution of the hyperspectral image dataset can be learned by the model and the hyperspectral images can be classified according to the learned representative spatial-spectral features with good classification performance.3.In order to further address the problem of poor accuracy due to the small number of labeled samples,and to improve the generalization capability of the classification model,a method for improving the accuracy of hyperspectral image classification based on depth learning by applying migration learning techniques in the case of small samples of hyperspectral images is proposed.Firstly,the migration of the generalized image classification model and the hyperspectral image classification model was carried out using the deep network fine-tuning method,and the experimental results showed that the applied migration learning was effective when the labeled samples were sparse,but the specific classification effect was limited by the use of the original model and the similarity of the new task to the original task,and the applicable scenario was limited.An adversarial network model is then generated by combining the open set domain adaptation technique and the hyperspectral image classification designed in this paper,by adjusting the loss function during training so that it can identify samples of unknown categories that are not present in the source domain.The experimental results show that the application of open set domain adaptation can apply the trained model to new data sets and classify them into known and unknown classes,accelerating the classification process while improving the classification accuracy of other known classes.