Hyperspectral Image Classification Based On Collaboratively Generative Adversarial Networks

Classification of hyperspectral images(HSIs)is a pivotal problem in hyperspectral remote sensing image processing.It is a challenging task to classify hyperspectral images with limited training samples.The generative adversarial network(GAN)can generate samples by the competition between the generator and the discriminator to force the generate samples distribution to approach the real samples distribution,so as to improve the classification performance of the discriminator under small samples.Therefore,GAN has become a promising technique to mitigate small sample size problem in HSI classification.Hyperspectral images usually have multiple classes of samples,the number of labeled samples in each class varies greatly,and the data distribution patterns of different classes of samples are different.GAN can improve the performance of generator and discriminator by the way of confrontation.It uses the judgment of the discriminator to learn the distribution of real samples,which acts as a loss function to provide a learning signal to the generator,but the quality of the samples generated by the generator cannot be evaluated directly.Thus,it is difficult to ensure that the generator is always updated to the real sample distribution.When the HSI data is involved,the generated sample distribution is more difficult to approximate the real sample with complex spatial-spectral distribution,and it is easy to lead to the problems of sample insufficient diversity and model collapse.which may further degrade the classification performance of the discriminator.Therefore,for the various class information,complex spatial distribution and high-dimensional spectral bands of HSIs,how to generate samples with high diversity,strong discrimination,and high-quality is a pivotal issue for GAN-based HSI classification.In addition,in view of the union of imagery and spectrum characteristic of HSIs,how to effectively utilize the spatial and spectral information of HSIs to extract distinguishing features,and achieve efficient and accurate image classification,is still a key problem in the task of HSI classification.To address these problems,a generative adversarial networks based on collaborative learning for HSI classification is proposed.The main contents of this paper are summarized as follows:(1)A GAN algorithm based on joint spatial-spectral hard attention module and convolutional LSTM is proposed.On the one hand,a joint spatial-spectral hard attention module is devised in the generator based on multi-branch convolutional network by using a dynamic activation function.It impels the generated samples approximate to the real samples with spatial-spectral distribution by retaining the discriminative features in the generated sample features and eliminating the confused and misleading ones.On the other hand,a convolutional LSTM layer is merged in the discriminator to capture the long-range dependencies among spectral bands and spatial information of HSIs,to achieve joint spatialspectral features classification.Through experimental verification,it is proved that the proposed method is robust to three challenging HSI datasets,and it obtains satisfactory classification results compared with other advanced methods.(2)A GAN method based on collaborative learning mechanism is proposed.In this method,the shallow to deep features of real multiclass samples in discriminator are utilized to assist the sample generation in the generator.It impels the generator and the discriminator not only compete but also collaborate for optimization and promotion.Finally,the quality of the generated samples and the classification performance of the discriminator are improved by enforcing competitive and collaborative learning between the discriminator and generator.Experiments on three HSI datasets show the proposed method can obtain satisfactory and robust classification results compared with advanced methods.(3)A GAN algorithm based on the multi-class generator and perceptual loss is proposed.The algorithm constructs multiple multi-class generators in the generator,and performs a class-specific loss on each multi-class generator,so that different multi-class generators can capture the data distribution patterns of specific samples of different classes.In addition,a perceptual loss function is constructed in the discriminator,it is used to guide the feature extraction process of the discriminator and the sample generation process of the generator.The experimental results show that,compared with other advanced methods,the proposed method can achieve the best classification results on three different hyperspectral image datasets.