Statistics of target spectra in HSI scenes

The majority of spectral imagery classifiers make a decision based on information from a particular spectrum, often the mean, that best represents the spectral signature of a particular target. It is known, however, that the spectral signature of a target can vary significantly due to differences in illumination conditions, shape, and material composition. Furthermore, many targets of interest are inherently mixed, as is the case with camouflaged military vehicles, leading to even greater variability. In this paper, a detailed statistical analysis is performed on HYDICE imagery of Davis Monthan AFB. Several hundred pixels are identified as belonging to the same target class and the distribution of spectral radiance within this group is studied. It is found that simple normal statistics do not adequately model either the total radiance or the single band spectral radiance distributions, both of which can have highly skewed histograms even when the spectral radiance is high. Goodness of fit tests are performed for maximum likelihood normal, lognormal, Γ, and Weibull distributions. It is found that lognormal statistics can model the total radiance and many single-band distributions reasonably well, possibly indicative of multiplicative noise features in remotely sensed spectral imagery.