The algorithms evaluated provided a good 
representative sample of spectral  classifier 
algorithms. Table 2 summaries the properties of 
each of the algorithms evaluated. Detailed 
descriptions of each algorithm follow. 
Algorithm A (Entropy) separates ice and water, based 
on local measurements of information randomness. 
A discriminate functions is used to describe the 
boundary between pixel representing ice and those 
representing water within scatterplots extracted from 
local windows passing over the entire image. The 
function which is selected for each window will have 
a negative slope and an intercept which is less than 
128. This algorithm was proposed by Shokr et al. in 
1991 and modified by Noetix Research. The 
modification was simply to use the mean pixel value 
and the standard deviation within the scatterplots, 
rather than using the pixel intensity and variance. 
An assumption with this algorithm is that the 
backscatter from open water will always be lower 
than that of ice, and will have intensity values less 
than 128. When the boundary is set all pixels below 
it are classified open water, and all values above it 
are classified as ice. 
Algorithm B (Migrating Means) separates classes 
based on a clustering technique which uses a 
histogram of the pixel intensities. In the first step, 
two randomly selected mean points are defined from 
the histogram. Two clusters are defined based on 
their proximity to the two mean points. Iterating 
redefines the mean points of each cluster and the data 
values are regrouped. This carries on until the 
clustering stabilizes. 
Algorithm C (Polynomial) relies on determining the 
modality within a local window as it passes over the 
image. The modality is determined through the use 
of a fourth order polynomial fit which best describes 
the histogram for each local window. If the function 
is bimodal, the window is positioned over an area 
representing more than one class so it is ignored. If 
the function is uni-modal the mean of the local 
window is saved, and used as either an ice or water 
value. The separation of these means into ice and 
water is accomplished through the use of the 
migrating means theory. 
  
  
  
  
  
  
Table 2 Algorithms 
Algorithm Name Discriminant Properties Reference 
Function 
Entropy Linear Samples from entire image Shokr et al. (1991) 
B Migrating Non-linear Samples from entire image Everitt (1974) 
Means 
C Polynomial 2™ Order, Linear | Samples from partitions Wackerman et al. 
(1991) 
D Mask Non-Linear Initially, samples from entire image, then Noetix Research 
focuses on ice and open water areas (1993) 
E Hierarchical Non-Linear Samples are local on the high resolution Pietkainen (1981) 
Network image and progressively increases to be 
global in successive coarser images. 
  
  
  
  
  
  
  
Algorithm D (Mask) is based on an extension of the 
migrating means theory. The migrating means 
algorithm was modified to recursively continue to 
separate classes, only stopping after a pre-determined 
maximum number of clusters are separated, or a 
predetermined number of pixel values per cluster is 
achieved. 
Algorithm E (Hierarchical Network) segments an 
image using a parent-child linked pyramid structure. 
The approach is a layered arrangement of arrays in 
which each array is half the size of the array one 
level up, and the bottom level contains the image 
which is be classified. A parent-child relationship 
between each pixel value and it’s position within the 
image is not fixed and may be redefined at each 
iteration. For each node in level / there is a 4x4 
sub-array of ’candidate child’ nodes at level / -1. 
The node itself is a member of four such sub-arrays 
430 
for level © 
linked to 
between | 
node is 
’legitimat 
define wi 
image seg 
All the al, 
from ov 
boundarie 
scenes th 
radiometri 
signal at t 
change in 
incidence 
significan! 
imaging ir 
2, and 5( 
expected 
Radarsat. 
will be us 
mosaic of 
covering a 
concerns t 
in the text 
Table 3 p 
weaknesse 
by evalu: 
summary, 
complex s 
tones. F« 
Strait (ima 
to the pres 
signal. In: 
ice will ha 
of volume 
algorithms 
water. T] 
algorithms 
the inherer 
characteris 
1) filtering 
these chat 
algorithm 
Algorithm 
output. À 
homogeno 
other class