Adaptive feature extraction techniques for subpixel target detections in hyperspectral remote sensing

Most target detection algorithms employed in hyperspectral remote sensing rely on a measurable difference between the spectral signature of the target and background. Matched filter techniques which utilise a set of library spectra as filter for target detection are often found to be unsatisfactory because of material variability and atmospheric effects in the field data. The aim of this paper is to report an algorithm which extracts features directly from the scene to act as matched filters for target detection. Methods based upon spectral unmixing using geometric simplex volume maximisation(1.2) (SVM) and independent component analysis(3) (ICA) were employed to generate features of the scene. Target and background like features are then differentiated, and automatically selected, from the endmember set of the unmixed result according to their statistics. Anomalies are then detected from the selected endmember set and their corresponding spectral characteristics are subsequently extracted from the scene, serving as a bank of matched filters for detection. This method, given the acronym SAFED, has a number of advantages for target detection, compared to previous techniques which use orthogonal subspace of the background feature 4. This paper reports the detection capability of this new technique by using an example simulated hyperspectral scene. Similar results using hyperspectral military data show high detection accuracy with negligible false alarms. Further potential applications of this technique for false alarm rate (FAR) reduction via multiple approach fusion (MA-F), and, as a means for thresholding the anomaly detection technique, are outlined.