Hyperspectral analysis paper, 2007

Home > Publications > By date

Title

Hyperspectral clutter, phenomenology and detection algorithms

Abstract:

In this paper we report on improved hyperspectral detection algorithms by appealing to the natural processes that govern the competition between the different vegetation species that provide the clutter against which the detection problem is posed.

Vegetation growth is restrained by inter- and intra-species competition for finite resources, such as light and water. Insight into these processes can be given by simple models using cellular automata. Two specific models have been chosen based on this physical motivation and developed further. Cellular automata are grid-based models that have simple interaction rules that act in the local neighbourhood of each member of the species. Despite the local and conceptually simple nature of the interaction rules, complex, long-range spatial correlations of the vegetation population can be generated.

Abundance maps for each species are generated from these simulations. These maps measure the relative spectral contribution of each species to the simulated observed spectra. By invoking the linear mixing model, these models are then validated against real clutter distributions and used to develop improved hyperspectral detection processing.

The validation method uses abundance maps together with a radiometrically accurate Monte-Carlo ray-tracing scene simulator (Cameosim) to produce hyperspectral cubes. Discussion is directed towards a comparison of the statistical properties of the simulated and observed background distributions.

The paper also describes the statistical properties that have been used to characterise the phenomenology of real hyperspectral clutter. Examples include the nature of the tails of the clutter distribution, whether the distributions are elliptically contoured, and the structure of the spatial correlations present in vegetation backgrounds. A number of cellular automata models are then developed whose rules are motivated by top-level biological competition principles. These cellular automata are used to generate abundance maps for various vegetation species.

The statistical properties derived have a direct importance in automatic anomaly detection methods, as many of these methods assume an underlying background distribution. For instance, the RX anomaly detection assumes an elliptical contoured background distribution. Insight gained from the vegetation clutter models provides evidence of the local probability distribution around each pixel. Having gained some understanding of this link between hyperspectral vegetation clutter and the natural processes that give rise to the spatial distributions of vegetation, candidate hyperspectral detection algorithms were developed to attempt to exploit this phenomenology. The performance of these algorithms was compared on real hyperspectral image data and benchmarked against existing techniques, and example results are reported in this paper.

Download:

[Adobe Acrobat icon] Adobe Acrobat document, xxxK

If you have problems viewing the PDF document, you may need to update Acrobat Reader to a more recent version. Click on the image below to visit the Adobe download site.

[Download Adobe Acrobat]

Authors:

C. A. Steer, Waterfall Solutions, UK;
M. Bernhardt, Waterfall Solutions, UK;

Conference:

SPIE Europe Security & Defence, Florence, Italy. 17 - 20 September 2007