PNNL

Abstract

We present two fully unsupervised deep learning approaches for hyperspectral anomaly detection. In one approach we formulate the anomaly detection problem as an adversarial game where a generator network learns the distribution of the hyperspectral background pixels comprising a single hyperspectral image and the output of the correspond- ing discriminator network yields a detection statistic. The other approach formulates the detection statistic as the error between an input hyperspectral pixel and the reconstruction of that pixel by an autoencoder network trained on the image. Both methods leverage a sub-sampling scheme that allows for unsupervised training and testing on the same data set. Our approaches are validated on a four-class synthetic hyper- spectral data set and compared to a statistical approach (RX) and a geometric approach (skeleton kernel principal component analysis). The proposed Generative Anomaly Detector algorithm achieves top performance on the data set while the autoencoder detection scheme also demonstrates performance gains relative to the comparison algorithms. Benefits and drawbacks of the approaches are discussed and highlight the many potential directions for future work.