Comparison of RX-based anomaly detectors on synthetic and real hyperspectral data

Abstract

Anomaly detection refers to detecting the deviations from the normal background behavior without any prior information about the target or the background. For hyperspectral image analysis, Reed-Xiaoli (RX) algorithm is arguably the most popular anomaly detector. It models the background as a multidimensional Gaussian distribution and computes how much a test vector is deviating from the background model. Over the years, many versions of RX have been developed and compared on VNIR or SWIR data, but longwave-infrared (LWIR) data comparisons are very few. In this paper, a comprehensive comparison of six different anomaly detectors, namely the global RX, local RX, dual window RX, subspace RX, kernel RX and the global RX combined with a uniform target detector, have been presented. The comparisons have been made on real LWIR hyperspectral data and synthetic data with varying noise levels and target sizes. Several factors to consider such as parameter selection, resilience to noise, effect of window size, computational complexity have been discussed and the detection performance have been presented on receiver operating characteristic curves.