Conditional Generative Adversarial Networks for Hyperspectral Image Classification

Though Hyperspectral Image (HSI) classification has been extensively investigated over recent decades, it is still a challenging task especially when the labeled samples are extremely limited. In this paper, we overcome the obstacle by using Conditional Generative Adversarial Networks (CGAN) to generate trainable data set with complete spectral and spatial information. Through comparing generated images of different shape and classification map for Indian pines, the most suitable data are selected and used to train the common model of neural network. Second, three common and latest neural network methods including two-dimensional Convolution (Conv2D), three-dimensional Convolution (Conv3D), Hybrid spectral CNN (Hybrid SN) used for HSI classification, are proposed. After repeating experiments and cross-validation, we have found that the proposed method, enhancing original data, can make model achieve better and robust performance for HSI classification compared to complete original data set, especially when the labeled data is limited.