A Hyperspectral Image Classification Method-Based Auxiliary Generative Adversarial Networks with Probabilistic Graph Model

Abstract

Hyperspectral image classification takes up a significant role in domains like mineral industry, agriculture and military purposes. Nevertheless, due to the availability of few labeled samples of hyperspectral images, it is fairly challenging while performing high-dimensional hyperspectral image classification. Presently, generative adversarial networks (GANs) have been employed widely for creating more synthetic yet realistic image samples. Still, getting high-quality image samples devoid of uncontrolled divergences and unwanted noise is a major challenge. To cope up with such challenges and to generate high-quality hyperspectral image samples, an auxiliary generative adversarial network with probabilistic graph (AGAN-PG) is proposed in this work. Specifically, the model has (1) a spatial-spectral generator unit that exploits the distinguishing spectral-spatial characteristics of the hyperspectral image data and (2) a discriminator unit that identifies the area categories of hyperspectral image cubes. Further, to make advantage of the relatively large amount of unlabeled data available, a conditional random field that refines the preliminary classification results generated by GANs is proposed. Eventually, the model increases the number and quality of training image samples generated, wherein the impact of overfitting is also avoided. Experimental results obtained using two well-known hyperspectral datasets—Pavia University and Indian Pines—demonstrate that the proposed framework AGAN-PG achieved promising classification accuracy even with a small number of initial labeled hyperspectral image samples for training.