303
EFFICIENT DETECTION OF ANOMALIES IN HYPERSPECTRAL IMAGES
S. R. Soofbaf 3 , M. J. ValadanZoej b ’ *, H. FahimnejacT, H. Ashoori b
a ISA (Iran Space Agency), Sayeh Alley, Africa St., Tehran, Iran - sr.soofbaf@gmail.com
b K.N.Toosi University of Tech. Geomatics Eng., Mirdamad Cross, Valiasr St., Tehran, Iran - (valadanzouj@kntu.ac.ir,
Hamed_Ashoori@yahoo.com)
c Sepehr Geomatics Engineering Co., East Hoveizeh St., North Sohrevardi Ave, Tehran, Iran -
Hamed_Fahimnejad@yahoo.com
Commission VI, WG VI/4
KEY WORDS: Hyperspectral, Target detection, Anomaly detection, RX algorithm, ROC
ABSTRACT:
By reason of recent advances in airborne and ground-base hyperspectral imaging technology, many applications have been
developed.
One of the most important hyperspectral images applications involves automatic detection of hidden objects without any prior
knowledge about them such as man-made targets, rare minerals in geology, vegetation stresses in agriculture, poisonous wastes in
environments, cancerous cells or tumors in medical imaging, etc.
The most robust class of algorithms for detection of this type of targets is arguably the one that searches the pixels of image cube for
rare pixels whose information significantly differs from their surrounding pixels and local background. These targets are known as
"Anomaly" in image processing and remote sensing literature.
Considering mentioned concepts, in this research, a host of different anomaly detectors such as RX-base anomaly detectors (Basic
RX, Modified RX, Normalized RX, Weighted RX, Causal RX, UTD, RX-UTD, and ACAD), Dual window-base Eigen Separation
Transform (DWEST) method, Nested Spatial Window-base Target detector (NSWTD) and Combined F-Test (CFT) algorithm are
investigated and compared.
1. INTRODUCTION
Hyperspectral imaging sensors acquire images in many
contiguous and very narrow spectral bands in visible, near-
infrared, and mid-infrared portions of electromagnetic spectrum.
This type of digital data shows vast potential for use in
automatic target detection since it provides useful information
about the spectral characteristics of the materials and targets in
the image.
This means that many targets that generally cannot be resolved
by multispectral images can be located as a consequence of high
spectral resolution in hyperspectral images according to the
concept of a spectral signature which uniquely characterizes any
given material. In this process, many factors should be
considered such as variations in atmospheric conditions,
location, noise of sensor, material composition, adjacent
materials, etc. But in many cases there is not any prior
information about the target. In this respect the most practical
approach is to search for anything that displays significantly
different spectral characteristics from its surroundings. This
process is known as "anomaly detection" in remote sensing
literature.
The main purpose of these target detection methods is to locate
targets which are commonly unknown, relatively small and only
occur in the image scene with low probabilities. Also these
algorithms can be applied directly to the radiance at the sensor
level. Thus they do not require any training or difficult step of
atmospheric correction and they are usually simple to
implement, even in a real-time or near real-time manner.
Nowadays anomaly detection has found in a broad variety of
applications ranging from defence, agriculture, geology,
environmental monitoring, medical imaging and etc.
In this research, a group of different anomaly detectors are
investigated and compared to select the best algorithm
according to purpose of aimed application, data specifications
and its quality (number of spectral bands, signal-to-noise ratio,
etc) and other effective parameters.
2. ANOMALY DETECTION ALGORITHMS
In this section most important anomaly detection algorithms are
briefly described theoretically which will be used for
comparative analysis.
2.1 Basic RX algorithm
The basic RX algorithm is the benchmark anomaly detection
algorithm, originally developed for multispectral images by
Reed and Yu (1993) that is formulated based on two hypotheses.
The first one models the image background as a Gaussian
distribution with zero mean and an unknown covariance matrix
which is estimated globally or locally from the data (N (0,X)).
The second hypothesis models the target as a linear combination
of a target signature and background noise. So, under a spectral
vector is represented by a Gaussian distribution with a mean
equal to the signature of the target (s) and an additive noise
equal to the background covariance matrix in hypothesis
(N(s,I)).
H Q :r = n
U (1)
H j :r = as + n
In this case detection process is based on exploiting the
difference between the spectral signatures of an input pixel and
its surrounding pixels that is very similar to the well-known
Mahalanobis distance and given by (Chang, 2003):