Hyperspectral Image Classification Based On Generative Adversarial Networks With Small Samples

Hyperspectral images classification is an important branch of remote sensing hyperspectral images processing.Its purpose is to determine the true labels of each pixel in the remote sensing hyperspectral images through computer analysis and other means.Labeling the training samples required for hyperspectral images classification is time consuming and laborious,so there is a problem of small samples with few labeled training samples in hyperspectral images classification.In addition,there are phenomena of different objects have the same spectrum and different spectrums have the same object in the hyperspectral images,and the hyperspectral images have the characteristic of spectrum unification.Therefore,how to extract the joint feature of spatial-spectral is another key problem in the task of hyperspectral classification.To solve these problems,this paper proposes a hyperspectral images classification algorithm based on generative adversarial networks(GAN)with small samples.Firstly,aiming at the small sample problem of hyperspectral images classification,a hyperspectral images classification algorithm based on data augmentation of GAN is proposed.The algorithm uses the characteristics of GAN to generate high-quality images,and designs a GAN model with gradient penalty based on auxiliary classifiers to generate labeled spectral samples to extend the real training set,as a data augmentation method to improve the classification results of the classifier.In order to ensure the diversity of the generated samples,an online sample generation mechanism is used.In order to remove some generated samples that are too far from the true sample distribution,an algorithm based on K nearest neighbor classifier is used to select the generated samples.Smoothing the labels of the generated samples by label smoothing regularization reduces the noise in the labels.Through experimental analysis,it is shown that the proposed algorithm achieves a consistent improvement effect on three hyperspectral data sets,and compared with other advanced algorithms,the superiority of the proposed algorithm is verified.Secondly,aiming at the problem of lack of labeled samples in the classification of hyperspectral images,but a large number of unlabeled samples are available,a hyperspectral images classification algorithm based on adversarial representation learning is proposed.The proposed algorithm will improve the GAN by adding an encoder and modifying the discriminator,the adversarial network will be extended to the field of representation learning.The features extracted by the encoder will be used for classification.The discriminator is modified into a multi-class discriminator.The discriminator is used to guide the generator to reconstruct the original images patch based on the high-level semantic information of the category,and the unlabeled samples are introduced through conditional entropy for semisupervised adversarial training to make the features extracted by the encoder more suitable for classification tasks.The experimental analysis verifies the robustness of the proposed method to the number of training samples on three hyperspectral data sets and its advancedness compared to other algorithms.Finally,aiming at the small sample problem and the extraction and fusion of spatial-spectral joint features,a hyperspectral images classification algorithm based on spatial-spectral dualchannel adversarial representation learning is proposed.The algorithm aims to solve the small sample problem and the extraction and fusion of spatial-spectral joint features through the dual-channel adversarial representation learning network and class consistency loss.The proposed algorithm takes into account the characteristics of spectrum unification of hyperspectral datasets,and introduces the consistency principle of spatial-spectral classification results to open up a new idea for spatial-spectral feature fusion,increases intraclass region consistency and inter-class edge preservation effect of classification results.The experimental results on three hyperspectral images datasets illustrate the advantages of the proposed algorithm over other advanced classification algorithms,especially when the number of samples is small.