Kernel subspace-based anomaly detection for hyperspectral imagery

Abstract

This paper provides a performance comparison of various linear and nonlinear subspace-based anomaly detectors. Three different techniques, Principal Component Analysis (PCA), Fisher Linear Discriminant (FLD) Analysis, and the Eigenspace Separation Transform (EST), are used to generate the linear projection subspaces. Each of these three linear methods is then extended to its corresponding nonlinear kernel version. The well-known Reed-Xiaoli (RX) anomaly detector and its kernel version (kernel RX) are also implemented. Comparisons between all linear and non-linear anomaly detectors are made using receiver operating characteristics (ROC) curves for several hyperspectral imagery.