EXTENDED SEMI-SUPERVISED LEARNING GENERATIVE ADVERSARIAL NETWORKS FOR FEATURE ANALYSIS AND CLASSIFICATION OF HYPERSPECTRAL IMAGERY

Abstract

Deep learning has been increasingly used for feature extraction and analysis of high-dimensional data. Hyperspectral imagery (HSI) cubes are high-dimensional data sets that can be difficult to analyze. One deep learning approach, generative adversarial networks (GANs), has been shown to be very effective in classification and generation of accurate synthetic data in computer vision problems. In this work we propose extended semi-supervised learning (ESSL) GAN and show that the ESSL GAN scheme produces effective classification and generation networks. We propose novel evaluation methods, novel performance metrics, and statistical correlation for the ESSL GAN model scheme and validate it using HSI data. Additionally, we introduce a novel class weighting scheme to improve the ESSL GAN generator performance when processing imbalanced data sets. Compared to standard classifiers, ESSL GAN is shown to produce classifiers that improve accuracy, precision, recall, and F1 score in all data sets evaluated.