Statistical Disorder Parameters Computing For Hyperspectral Image Anomaly Detection

Abstract

Two statistical disorder parameters are defined for hyperspectral anomaly detection in this paper. While the background information is usually located in principal components of the hyperspectral data containing the most energy, the low variance components contain anomaly or noise signals. Two introduced parameters are computed based on the principal components. The first parameter called as entropy contains the randomness value of the spectral measurements while the second parameter called as anisotropy contains the relative importance of the consecutive components of the hyperspectral image. The extracted features can be given to any arbitrary anomaly detector. The experimental results show that feeding entropy and anisotropy features to the RX detector provides a significant improvement in hyperspectral anomaly detection.