Delving Into Classifying Hyperspectral Images via Graphical Adversarial Learning

Recent remote sensing literature has seen generative adversarial network (GAN)-based models developed for hyperspectral image classification, especially in a spatiospectral manner. The intuition is that training classifiers with additional generated hyperspectral data improves model generalization and hence increases classification accuracy. Existing GAN-based hyperspectral image classification methods tend to straightforwardly characterize spatiospectral characteristics of hyperspectral data subject to basic distributions (e.g., Gaussian or uniform distributions). However, hyperspectral imageries have high dimensions in both spectral and spatial representations, which are possibly derived from a latent space following more sophisticated distributions. In this scenario, we believe that comprehensively modeling the latent space would favor accurate hyperspectral classification. To this end, we develop a graphical adversarial learning (GAL) framework that explores the latent variable structure for generating diversified hyperspectral samples. The comprehensive modeling strategy enables GAL to be capable of accurately characterizing full-depth hyperspectral data such that it establishes an end-to-end framework that does not require reduction on spectral bands. We conduct extensive experiments on three public hyperspectral datasets in terms of processing full-depth hyperspectral images without dimension reduction. The experimental results validate that, first, our GAL excels at full-depth hyperspectral image classification, and second, our GAL is competitive with state-of-the-art methods which use global data (i.e., both training and testing data) for learning spectral reduction.