Modeling Noise in a Framework to Optimize the Detection of Anomalies in Hyperspectral Imaging

Abstract

Hyperspectral imagery (HSI) has emerged as a valuable tool supporting numerous military and commercial missions. Environmental and other effects diminish HSI classification accuracy. Thus there is a desire to create robust classifiers that perform well in all possible environments. Robust parameter design (RPD) techniques have been applied to determine optimal operating settings. Previous RPD efforts considered an HSI image as categorical noise. This paper presents a novel method utilizing discrete and continuous image characteristics as representations of the noise present. Specifically, the number of clusters, fisher ratio and percent of target pixels were used to generate image training and test sets. Replacing categorical noise with the new image characteristics improves RPD results by correctly accounting for significant terms in the regression model that were otherwise considered categorical factors. Further, traditional RPD assumptions of independent noise variables are invalid for the selected HSI images. Introduction: Hyperspectral imagery (HSI) has emerged as a valuable tool supporting numerous military and commercial missions including counter concealment, camouflage and deception, combat search and rescue, counter narcotics, cartography and meteorology to name a few (Manolakis (2002); Landgrebe (2003)). A hyperspectral image, also called an image cube, consists of k spectral bands of an m by n spatial pixel representation of a sensed area. Each pixel in the spectral dimension represents an intensity of energy reflected back to the sensor. All spectral dimensions for a given pixel represent a potential target signature. HSI, by its very nature, can provide a method for identifying at most (n 1) unique spectral signals, where n is the number of independent bands in an HSI image cube. This is (n 1) rather than n because one band is used to define the background or noise present in an image. Since HSI contains typically hundreds of bands, this number of signals or targets for classification can be large although bands affected by atmospheric absorption contain little useful information and must be removed and bands that are close to each other are typically correlated. Davis (2009) describes some pitfalls when performing target classification on hyperspectral images. For instance, the spectral library will most often not contain every possible object, manmade or other to be classified. Some objects may be concealed or disguised to make the spectral signature different from what is contained in the library. In addition, environmental effects such as time of day, relative humidity and imaging angle greatly impact the data reflectance values observed by a sensor. Finally, target prior probabilities can be very small in comparison to the number of pixels being considered.