Characterization of the spectral distribution of hyperspectral imagery for improved exploitation.

摘要

Widely used methods of target, anomaly, and change detection when applied to spectral imagery provide less than desirable results due to the complex nature of the data. In the case of hyperspectral data, dimension reduction techniques are employed to reduce the amount of data used in the detection algorithms in order to produce "better" results and/or decreased computation time. This essentially ignores a significant amount of the data collected in k unique spectral bands. Methods presented in this work explore using the distribution of the collected data in the full k dimensions in order to identify regions of interest contained in spatial tiles of the scene. Here, interest is defined as small and large scale manmade activity. The algorithms developed in this research are primarily data driven with a limited number of assumptions. These algorithms will individually be applied to spatial subsets or tiles of the full scene to indicate the amount of interest contained. Each tile is put through a series of tests using the algorithms based on the full distribution of the data in the hyperspace. The scores from each test will be combined in such a way that each tile is labeled as either "interesting" or "not interesting." This provides a cueing mechanism for image analysts to visually inspect locations within a hyperspectral scene with a high likelihood of containing manmade activity.