This paper formulates the hyperspectral anomaly detection problem in terms of a local context by modeling the relationship of individual pixels with the annuli of pixels that surround them. In this formulation, a locally anomalous pixel is one that might even be quite typical in
the context of the whole image, but is “out of place” with respect to its local neighborhood. The problem of anomaly detection is cast as a supervised learning problem, in which samples from one class (normal) are provided by pixel/annulus pairs that occur in the scene, and samples
from the second class (anomalous) can be created by making pixel/annulus pairs in which the pixels and annuli are effectively scrambled with respect to each other.
Although the formulation is in terms of machine learning, the experiments performed here use a simplified approach in
which parametric (multivariate Gaussian and fatter-tailed multivariate t) distributions are fit to the data. This leads to a suite of local anomaly detectors that we compare to standard local RX and global RX detectors.