Ellipsoid-simplex hybrid for hyperspectral anomaly detection

The problem of anomaly detection in hyperspectral imagery is expressed in terms of a minimal volume set in a high-dimensional space that encloses the bulk of the data samples. The venerable RX algorithm employs an ellipsoid for this volume, but endmember methods can be used to create a simplex volume. This paper considers a hybrid ellipsoid-simplex volume and characterizes its performance on hyperspectral imagery by computing a plot of volume versus false alarm rate. This plot provides a generic measure of quality (smaller volumes are better) without requiring the identification of specific anomalies in the data.