Application of Local Supervised Feature Selection Approach to Target Detection in Hyperspectral Imagery

Abstract

Feature selection (FS) for target detection (TD) attempts to select features that enhance the discrimination between the target and the image background. Moreover, TD usually suffers from background interference. Therefore, features that help detectors suppress the background signals and magnify the target signal effectively are considered more useful. Accordingly, in this paper, a supervised FS method, called autocorrelation-based feature selection (AFS), is proposed based on the TD concept. This method uses the image autocorrelation matrix and the target signature in the detection space (DS) for FS. Features that increase the first-norm distance between the target energy and the mean energy of the background in DS are selected as the optimal features. To evaluate the proposed method and to explore the impact of FS on the TD performance, the target detection accuracy (TDA) measure is employed. The experiment shows that the proposed FS method outperforms the two existing FS methods used for comparison. In fact, AFS achieves the maximum TDA value of 19.02% using 58 features while, compared to FS, the other methods achieve much lower values. Furthermore, the effect of image partitioning on the TD performance in both full-band and reduced-dimensionality feature spaces is investigated. The experiment results show that partitioning, as a way of adding local spatial information to TD, dramatically improves the TD performance. For experiments, the HyMap dataset is employed.