Detection of Anomalies in Hyperspectral Images

Summary:

Hyperspectral images are digital images, often taken either from an airplane or satellite, in which each pixel has not just the usual three visible bands of light (red at 650nm, green at 550nm, and blue at 450nm), but on the order of hundreds of wavelengths so that spectroscopy can be conducted on the materials on the ground. The user is then able to identify, for instance, the species of trees and other vegetation, crop health, mineral and soil composition, moisture content of soils and vegetation, and pollution quantities. In images over water, it is possible to identify and quantify water particulates, chlorophyll levels, erosion patterns, and in some cases water depth. The technology also has clear military and intelligence applications, as it enables the identification of man-made materials such as buildings and vehicles, even with attempts to camouflage. It is also possible to detect and identify gas plumes such as those arising from leaks even when the gasses are invisible to the human eye. In fact, hyperspectral imaging was used following the attack on the twin towers and the hurricane Katrina disaster to identify dangerous gas leaks, providing guidance and protection to rescuers. Hyperspectral sensors are flown on aerial platforms by several private companies, and NASA runs a hyperspectral sensor called Hyperion on their EOS-1 satellite.

Although this technology has incredible potential, its utility is currently limited because of the enormous quantity and complexity of the data it gathers. A single image can consist of millions of pixels, each with hundreds of bands, and can require tens of gigabytes of memory to store. The standard data analysis algorithms for spectral and multispectral imaging, which use linear methods and multivariate statistics, break down on complex scenes and are unable to extract information from the nonlinear structures embedded in hyperspectral data sets. It has recently been demonstrated, by researchers at RIT and others, that topological tools along with ideas from the theory of networks are well suited to successfully handle hyperspectral data sets.

The standard algorithm for detecting anomalous pixels in an image, the well-known RX algorithm, simply ranks pixels according to their standard deviations from the mean in a multivariate sense. Specifically, the Mahalanobis distance from the mean, is computed for every pixel in the image and the output is displayed as a grayscale image. As before, this algorithm performs well in simple images such as a large forest, but fails in complex urban scenes and in the presence of even mild sensor noise. Clearly, this algorithm assumes that the data has a Gaussian distribution which is often not true. The Topological Detection Algorithm (TAD) constructs a graph on the data: each pixel is represented by a node and two nodes are connected via an edge if they are closer than some threshold. In TAD, the threshold is low so that only the most dense background pixels are collected in large components of the graph. Every pixel in the image is then ranked by its distance to its closest neighbors in the background and the result is output as a grayscale image.

Publications:

1. Anomaly Clustering in Hyperspectral Images, T.J. Doster, D.S. Ross, D.W. Messinger, W. F. Basener, Proceedings, SPIE Conference on Defense, Security, and Sensing, Orlando, (2009).


2. Anomaly detection using topology, W.F. Basener, E.J. Ientilucci and D. Messinger, Proc. SPIE 6565, 65650J, (2007).