Statistical models for physically derived target sub-spaces

Abstract

Traditional approaches to hyperspectral target detection involve the application of detection algorithms to atmo- spherically compensated imagery. Rather than compensate the imagery, a more recent approach uses physical models to generate target sub-spaces. These radiance sub-spaces can then be used in an appropriate detection scheme to identify potential targets. The generation of these sub-spaces involves some a priori knowledge of data acquisition parameters, scene and atmospheric conditions, and possible calibration errors. Variation is allowed in the model since some parameters are di±cult to know accurately. Each vector in the subspace is the result of a MODTRAN simulation coupled with a physical model. Generation of large target spaces can be computationally burdensome. This paper explores the use of statistical methods to describe such target spaces. The statistically modeled spaces can then be used to generate arbitrary radiance vectors to form a sub-space. Statistically modeled target sub-spaces, using limited training samples, were found to accurately resemble MODTRAN derived radiance vectors.