Practical Evaluation of Max-Type Detectors for Hyperspectral Images

This paper evaluates performance of max-type detectors introduced in earlier work as a special case of generalized fusion. The max-type detectors combine several base detectors, e.g., matched filters (MF), based on information from identified background segments. For identification of background segments, we use a method that we call Flexible Spectral Clustering (FSC) since this turns out to be a flexible algorithm particularly suited for spectral data. For our numerical results, we use an AVIRIS urban scene with a wide range of natural and man-made materials. Our technique of target implantation to each image pixel allows taking into account various realistic combinations of the background materials present in the pixel being tested for the target. We evaluate performance of several max-type detectors in order to understand the impact of the number of background segments being used. We also make a comparison to the base detector, which is the global MF (GMF) in this case. In most cases considered in this paper, the max-type detectors were at least as good as the GMF, and they were often much better than the GMF. The difficulty of a given detection scenario was described by the size of the target and the strength of its spectral signal. In very difficult detection scenarios (small targets that are not very different from the background), the GMF performs much worse than the max-type detectors. In medium-difficulty scenarios, the detection capability of the GMF gets closer to that of the max-type detectors. In the scenarios with easy detection (large targets that are very different from the background), all detectors considered here perform very well and their performance is almost identical to each other. Similar results are also obtained for detection of a real target known to be present in the image. This paper also provides some valuable evaluation tools for re- searchers to investigate performance of other detectors. By using our tools, one can understand how the detectors' performance depends on the target size and the strength of its spectral signal.