For nearly thirty years now, airborne and satellite hyperspectral imaging sensors have been used to collect high spatial resolution (1-30 meter) imagery of the earth's surface in hundreds of co-registered, contiguous spectral channels. These data have been shown to enable the detection of objects smaller than a pixel due to the spectral information present. However, it is not always obvious beforehand if a given object will be detectable in a given scene, as performance has been observed to depend on many factors including illumination conditions, scene spectral complexity, target variability, sensor artifacts as well as algorithm variations. Over the past fifteen years our research has been exploring ways to predict and assess performance of hyperspectral subpixel detection. Our methods have included analytical modeling tools, empirical blind tests, and quality metrics for spectral imagery. Results of this work have confirmed the feasibility of hyperspectral subpixel objection detection and have provided tools for quantification of the performance.
展开▼
