AN ASSESSMENT OF SOME ANOMALY AND TARGET DETECTION ALGORITHMS FOR HYPERSPECTRAL IMAGES

Abstract

Anomaly and target detection is of particular interest in hyperspectral image analysis as many unknown and subtle signals (spectral response) unresolved by multispectral sensors can now be uncovered in hyperspectral images. The detection of signals in the form of small objects and targets from hyperspectral sensors has a wide range of applications both civilian and military. Anomaly detection is an ability to find spectral outliers within a complex background in an image without a priori information about the image or its specific contents. This is an unsupervised way of target detection where the knowledge of target to be detected from the image is required. Various algorithms have been developed for anomaly and target detection from hyperspectral images. For anomaly detection, there are some algorithms which detect the anomalies based on the local normal model of the pixel neighbourhood. Another approach for anomaly detection is based on Gaussian Mixture Model which can be used to extract global anomaly from the heterogeneous background. In addition, several anomaly detection algorithms are based on the transformation of the data in . a space which highlight the anomalies and some algorithms are nonlinear making the use of kernel functions. Likewise, a large number of algorithms have been developed in the field of target detection from hyperspectral images. A number of target detection algorithms are'based on linear unmixing of endmembers present in the scene which require complete knowledge of image endmembers. Whereas, a few algorithms have been developed which require the knowledge of only target endmembers, there are some nonlinear algorithms that make use 'of kernel functions for target detection. From an extensive literature review on anomaly and target detection algorithms, it has been 'observed that each algorithm has its 'own advantages;- disadvantages and assumptions. The selection of a . particular, algorithm may depend on the amount , of information available as per the . requirement of the algorithm, -application area, :.the computational complexity etc. In the present study, some -. algorithms for anomaly and target detection have been investigated. The algorithms that have been selected for anomaly .detection include Reed Xiaoli (RX) algorithm as it is considered as :a benchmark for all the anomaly detection , ' algorithms. Some variants of RX. have been studied, which include Reed Xiaoli-Uniform iiiPage Target Detector (RX-UTD), and Orthogonal Subspace Projection-Reed Xiaoli (OSP-RX). These algorithms detect anomalies by first suppressing the background and then performing the anomaly detection. All these algorithms have been examined for the cases in which the background has been estimated locally from a neighbourhood around the pixel under test or globally from the whole image. Similarly, for target detection, the Orthogonal Subspace Projection (OSP) which is based on linear unmixing and requires complete knowledge of , the image endmembers, Constrained Energy Minimization (CEM that requires only the knowledge of target, endmember, have been selected. Further, a nonlinear version of OSP called Kernel Orthogonal Subspace Projection (KOSP) has also - been selected. The selected anomaly and target detection algorithms have been implemented` via an inhouse developed software in Visual C++. The efficacy of these algorithms has been examined over three different hyperspectral datasets which include a synthetic image, an image acquired from AVIRIS and HyMap hyperspectral sensors. The quality of anomaly and target detection from these algorithms has been evaluated through visual interpretation and also -through receiver' operating characteristic (ROC) curves. Based on various experiments, it has been found that the quality of the anomaly detection algorithms that make use of an outer window and guard window highly depends- on the size of these windows. RX-UTD performs better when the input data follow uniform distribution. In case of OSP-RX algorithm, the number of endmembers should be chosen cautiously otherwise it may lead to the suppression of anomalies or may increase the false alarm rate. In general, the anomaly detection algorithms which are based on the estimation of background statistics from the local neighbourhood yield higher performance as compared to those which estimate background statistics from the whole image. In case of target detection algorithms, the spectral variability among the targets does matter when the separation between the background and the target class is very less. The performance of OSP algorithm has been found _ to be better than or comparable to CEM algorithm. However, KOSP algorithm performs better than OSP algorithm but the key limitation of this algorithm lies in the selection of appropriate value of the width of the kernel function. ... .._