Spectral image utility prediction

The utility of an image is an attribute that describes the ability of that image to satisfy performance requirements for a particular application task. The robust ability to predict the utility of an image for a given application would facilitate sensor design trade studies, provide a basis for tasking image collection activities, and create the foundation for an image archive indexing scheme. In this paper, we examine methods for predicting the utility of spectral images for detecting sub-pixel targets using the constrained energy minimization matched filter detector. The result of our initial work is a prediction of the likelihood of finding a synthetically implanted target in a target-free image in advance of actually applying the detector. We define image utility for the target detection application as the probability of detection at a specified probability of false alarm. We analytically predict this utility for a given image by first estimating statistical parameters directly from the image, then operating on these parameters with the matched filter detector. Three parametric statistical models are used for characterizing the image background: the global Gaussian, the multiple-class Gaussian, and the elliptically contoured t-distributions. The target models come from a library of target materials and are assumed to be multivariate Gaussian with known mean and covariance. We benchmark prediction performance by comparing predicted detection probabilities to empirical results obtained by applying the detector to the data for two HYDICE images. Our longer term objective is to build on this initial result by developing a more general spectral image quality and utility framework, and specific metrics for the prediction of utility across many potential applications.