Matched filter stochastic background characterization for hyperspectral target detection

Abstract

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. CAPT WEST JASON E FUNDING NUMBERS PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES)ROCHESTER INSTITUTE OF TECHNOLOGY PERFORMING ORGANIZATION REPORT NUMBERCI04-1012 SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) ABSTRACTAlgorithms exploiting hyperspectral imagery for target detection have continually evolved to provide improved detection results. Adaptive matched filters can be used to locate spectral targets by modeling scene background as either structured (geometric) with a set of endmembers (basis vectors) or as unstructured (stochastic) with a covariance or correlation matrix. These matrices are often calculated using all available pixels in a data set. In unstructured background research, various techniques for improving over the scene-wide method have been developed, each involving either the removal of target signatures from the background model or the segmenting of image data into spatial or spectral subsets. Each of these methods increase the detection signal to background ratio (SBR) and the multivariäte normality (MVN) of the data from which background statistics are calculated, thus increasing separation between target and non-target species in the matched filter detection statistic and ultimately improving thresholded target detection results. Such techniques for improved background characterization are widely practiced but not well documented or compared. This paper provides a review and comparison of methods in target exclusion, spatial subsetting, and spectral pre-clustering using preliminary matched filter detection results from a larger study. The analysis provides insight into the merit of employing unstructured background characterization techniques, as well as the limitations for their practical application.

Comparative evaluation of background characterization techniques for hyperspectral unstructured matched filter target detection

“…The target approach results ( Figure 3 & Results not shown here [6] indicate that this level of performance is not achievable when the man-made targets in the scene are not excluded from computation of the background statistics. Also, the "mixed" background performed particularly well for the high contrast targets indicating that the broader image is well suited to serve as a background, but there was a detrimental impact of including all pixels in the statistics.…”

“…If the target exists in the scene (as is the assumption in the target detection problem), then its signature is incorporated in these estimates and the target one is looking for has been explicitly included in the background. This is obviously less significant if the target occupies a very small portion of the image, but it has been shown that including as few as five target pixels in a test set of 18,000 background pixels can alter the eigenvectors of the covariance matrix computed for that test set of pixels [7]. A change in the eigenvectors indicates a change in the shape of the covariance matrix which can dramatically affect the resulting test statistic for detectors based on a Mahalanobis distance (such as the GLRT in equation 3).…”

“…Target exclusion seeks to minimize the effects of target signatures in the characterization of the background and thus, increase the separation between the target and background spaces. West (2005) and showed that including target pixels in the estimation of the covariance of a single-class background influenced the shape of the covariance matrix. This change in shape corresponds to higher false alarm rates.…”

“…This change in shape corresponds to higher false alarm rates. Improvements on scene-wide background characterization methods have also been achieved with segmentation methods [6,[8][9][10][11]. Subsetting the image for background training seeks to improve the multivariate normality (MVN) of the background by selectively including or excluding pixels from the background statistics pool.…”

A morphological adaptation to spectral target detection searches

“…For instance, anomaly detection attempts to characterize the background and then determine which pixels vary significantly from it (Soofbaf et al 2007, Molero et al 2011. Other detectors use probability distributions and hypothesis testing to identify pixels similar to a target signature that are not the background (Howari 2003, Manolakis et al 2003, West 2005. Commonly used algorithms include spectral angle mapper (SAM), matched filter (in numerous forms), and spectral unmixing (Schaum 2006, Campbell 2007, Schott 2007.…”