Applying generative adversarial networks for anomaly detection in hyperspectral remote sensing imagery
Abstract
Automatic anomaly detection has previously been implemented on hyperspectral images by use of different statistical methods with good results. We apply a machine learning method using generative adversarial networks (GAN) to a hyperspectral remote sensing image provided by the Norwegian Defence Research Establishment (FFI), using the Wasserstein formulation with gradient penalty. We train a generator network which outputs generated 32x32 pixel image windows with good visual quality, but are only able to make anomaly detection work for 1x1 pixel windows. We show results for both a discriminator based anomaly detection method, which only works properly with very carefully tuned hyperparameters, and for a generator based anomaly detection method making use of a recreation cost, which is robust and achieves useful results that are not much worse than more highly developed methods. Our work demonstrates the possibilities of applying GANs for unsupervised anomaly detection and explores the challenges that this method presents.