Clutter complexity analysis of hyper-spectral bands

This work investigates the use of a clutter complexity measure for the selection of bands for a detector operating over forward looking hyper-spectral image cubes. By clutter complexity, we mean an aggregation of statistical image features that predict the "degree of difficulty" to detect and/or identify a target object in an image. We show that clutter complexity correlates well with single-band automatic target recognition (ATR) performance. We also show the performance of the multi-band ATR when the bands are selected based on clutter complexity. To baseline this ATR performance, we consider a uniform band ordering strategy and an "optimal" ordering strategy determined by an exhaustive search. Our results show that the ordering by clutter complexity results in an improvement of the ATR performance when compared to the uniform ordering strategy. This improved performance is however not as good as the performance obtained for the optimal ordering.