Graph Network Based Generative Adversarial Network for Hyperspectral Image Classification

Abstract

Generative Adversarial Network is one of the emerging deep learning technologies that is used for many applications. Semi-supervised classification is one of the application of GANs by using the discriminator as a classifier. Semi-supervised GANs are used in Hyperspectral image classification to address the limitation of labeled dataset for the training. Most semi-supervised GAN models utilizes Euclidian structure of Hyperspectral image in the discriminator and other models like GCN is used to utilize non-Euclidian structure of Hyperspectral image in form of graph. But most of the semi-supervised GAN models were not address the utilization of graph like structure which contains abundant information compared to Euclidean form. In the proposed solution, GNGAN model are used as semi-supervised classifier by utilizing both Euclidian structure which has joint spatial and spectral features and non-Euclidian structure in the form of graph which has deep spectral features in order to increase the performance of the classification processes. The proposed model uses Convolutional block and Graph network block networks for utilizing Euclidian and non-Euclidean structures of Hyperspectral image respectively. Concatenation mechanism was used as combination mechanism after compared with Hadamard product and Element wise Addition methods to combine the two blocks. The proposed classifier is carried out with two mostly used Hyperspectral image datasets: Salina Valley and Pavia University. The contribution of the proposed work is two which are: it address the long term dependencies of spectral features in non-Euclidian structure using Graph network block and it learns joint spatial spectral features in Euclidian structure using Convolutional block. The obtained results showed that the proposed model provide competitive results compared to other methods: SSGAN and GCN. The GNGAN outperform SSGAN and GCN by 0.5% and 3.66% for Salina Valley dataset and 3.8% and 6.17% for Pavia University dataset respectively using overall accuracy as evaluation matrix