The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXV//. Part B7. Beijing 2008
306
2.9 Combined F-Test anomaly detector (CFT)
CFT method is based on a nonparametric model that compares
two sets of data (random variable), background data and target
data by using a typical inner/outer window mechanism to
sample the image for local detection, or using full image data to
global detection. In this method after combining two mentioned
samples, statistical parameters of them (mean and variance) are
estimated. Then ratio of these parameters are calculated (Z CF r)
and tested by Fisher distribution of (1,1) degree of freedom.
A decision threshold T is determined via defining a type I error
for F distribution function and compared with the result of ratio.
If Z CFT is greater than T, it means that these two sets are most
likely sampled from different distributions. Therefore they are
anomalous to each other. If not, they are likely sampled from
the same distribution.
The main assumption in this method is asymptotic behaviour of
Fisher’s F distribution for data sets which are examined by a
common statistical test (Rosario, 2005).
and their performances are compared by discrete Receiver
Operating Characteristic (ROC) curves and their area under
curves (AUC) as appropriate criterions for evaluation of
detection algorithms. Therefore seven threshold values are used
to achieve confidence levels about 93% to 99% for calculating
probability of correct detection and false alarm.
For example figure (2) shows the visual result of local RX
method with local window size 15, correct detected targets and
ROC curve of this algorithm
3. EXPERIMENTAL RESULTS
Two sets of hyperspctral data have been used for these
experiments.
The first one is a real hyperspectral data (AVIRIS image of
Cuprite region which is available in ENVI sample data) for
visual inspection of detection results. The second one is a
synthesized hyperspectral data for quantitative evaluation which
was simulated by sampling of contiguous spectral curve with 20
nanometre spectral resolution for some natural and man-made
targets such as dry grass, sandy loam, sagebrush, galvanized
iron and aluminium metal using ENVI spectral libraries and
MATLAB software. This synthetic hypercube’s size was
100x100 pixels which made up linear combination of some
above signatures for background region and anomaly panels.
In background region first pixel is started with 100 percent dry
grass and 0 percent sandy loam. Then by moving to next pixel
percentage of dry grass is decreased and percentage of sandy
loam is increased. This process is repeated until the last pixel
has 0 percent of dry grass and 100 percent sandy loam. Then
anomaly pixels are added to this data in various locations.
Anomaly panels in each column have the same pure signature
with various sizes and they have the same signature in each row.
All of anomaly pixels have 10 percent abundance of anomaly
signatures such as galvanized iron and 90 percent abundance of
background signatures. Also Gaussian noise is added to each
pixel to achieve 30:1 signal-to-noise ratio.
Figure (1) shows the simulated hypercube in band number 20
(800 nanometre) and location of considered anomaly pixels in it.
(a) (b)
Figure 1. Simulated image in band number 20 (a),
location of anomalies (b)
Figure 2. Result of local RX (a), correct detected anomalies in
99% confidence level (b), full ROC curve (c),
comparable part of ROC curve (d)
Local DWEST and NSWTD algorithms are tested with different
combination of local window sizes. Figure (3) shows the best
visual results of these methods.
Figure 3. Global DWEST-RX (a), global DWEST (b),
local DWEST with in/out window size 1/7 (c),
NSWTD with in/mid/out window size 1/11/25 (d)
Also figure (4) shows the visual results of global RX-base
methods.
All mentioned algorithms (in global or local manner with
various window sizes) are implemented in MATLAB software