Discrimination measures for target classification

Target detection does not necessarily yield target classification since the detected targets may belong to different classes and cannot be differentiated one from another by target detection. This often occurs when no prior target knowledge is available. In order to resolve this dilemma, four measures are proposed for target discrimination in this paper, two of which are designed based on the Bhattacharyya distance and the other two are derived from the concept of the matched filter. They will be used to cluster the detected anomalies into different types of targets in an unsupervised manner. These four target discrimination measures take advantage of the second-order statistics of the image data to account for sample spectral correlation. Consequently, they all outperform the commonly used single-pixel based similarity measures, such as spectral angle mapper (SAM), Euclidean distance. An experiment-based quantitative study is conducted for their performance evaluation. Interestingly, all these four measures perform very similarly