Open Set Domain Adaptation for Hyperspectral Image Classification Using Generative Adversarial Network

Hyperspectral image (HSI) classification attracted lots of attention due to its complexity in dealing with large dimensions. In recent years, the techniques for dealing with the HSI have been evolved, ensuring the increase in efficiency to some extent in classification and other perspectives. Domain adaptation is a well-established technique for using any trained classification model, when the feature space from target domain is a subset of feature space from source domain. The objective of this paper is to create an efficient and effective model for HSI classification by implementing open set (OS) domain adaptation and generative adversarial network (GAN). This has advantages in quite few ways, such as creating a single training model that deals with various HSI data set with common classes, classifying the features in any data to specific trained classes and unknown (to be labelled) making it easy to annotate. The proposed open set domain adaptation for HSI classification is evaluated using Salinas and Pavia. The proposed method resulted in the classification accuracy for unknown classes as 99.07% for Salinas and 81.65% for Pavia.