Target detection in hyperspectral images based on multi-component statistical models for representation of background clutter

Hyperspectral imaging has potential for detection of low-contrast targets in the presence of significant background clutter. We consider here the important case of detecting small targets as anomalies in a spatially cluttered natural background. In order to achieve a low false alarm rate, the properties of the background must be captured by the analysis procedure in sufficient detail to represent the full range of natural variation. Here we examine a statistical background model where background variations are represented by a sum of several multivariate normal probability distributions. The parameters of the statistical model are estimated using the stochastic expectation maximization (SEM) method. The quality of the resulting model's representation of natural backgrounds is discussed in terms of detection performance as a function of model complexity. Results are given for various illumination conditions and targets with different contrast to the background. We show that detection performance can be drastically improved by using multicomponent background models, and that a low number of components is sufficient for detection of quite low contrast targets. The study is based on data with high spectral and spatial resolution from the Airborne Spectral Imager (ASI) hyperspectral sensor.