Spectral signatures derived from a multispectral or hyperspectral imager can be used in matched filter algorithms to help distinguish targets from background. In this paper we demonstrate the use of these matched filters for different target implantation models. We show that even though a specific matched filter is designated for a particular implantation model, we can use other matched filters and obtain higher detection values for low false alarm rates. We evaluate the efficiency of the algorithms by systematically implanting the target's signature into every pixel in the image and obtaining its score; the lowest scores are those pixels in which the target may be missed. For every algorithm, we generate histograms for the no-target and target cases and then analyze using the classical ROC curve.
We present a method to evaluate point target detection algorithms. For any particular algorithm, a datacube without a target is evaluated with each pixel being assigned a score; the highest scores belong to those points which are potential false alarms. We then systematically implant a selected target signature into every pixel in the image and evaluate the resulting scores; the lowest scores are those pixels in which the target may be missed. ROC curve analysis can then be made. In this paper, we evaluate a new algorithm which we have developed; we use the evaluation of this algorithm as a paradigm for the efficacy of our algorithm evaluation tool.