Table 3 
Algorithm Failures Successes 
A e Any range ambiguities accentuated. e Open water lead in c-band image was properly 
water. 
as water. 
e Smooth first year & new ice classified as water. 
(Entropy) e Wet multiyear & first-year ice misclassified as 
e Smooth fast ice with low return signature classified 
classified. 
e Easily interpreted output. 
  
B e Any range ambiguities very evident. 
(Means) water. 
* Noisy output. 
e Classifies new & thin first year ice as water. 
(Migrating) |® Wet multiyear & first-year ice misclassified as 
e Handles variable signal return due to range fall 
off. 
  
C e Noisy output. 
(Polynomial) 
e Smooth first year & new ice classified as water. 
e Better handling of range effects than algorithms 
A or B. 
e Open water generally well delineated. 
  
  
* Negative results. 
(Hierarchical) 
(Network) 
  
  
D * Uncontrolled number of classes. e Best handling of range effects. 
e First year ice classified as water. e Separates wet ice into distinct intermediate 
(Mask) e Lead in c-band image partly mis-classified as ice. class. 
E e Processing intensive. 
  
  
  
This problem is evident in the output for images 2 
and 4. In these scenes, the majority of the pixel 
were high values (i.e. 180-255 range) resulting in a 
discrimination function being drawn in the middle of 
the ’ice mode’, and everything lower than this 
function was classified as open water. 
It should be noted that the overestimation of open 
water was accentuated in scenes where a ’fall-off” in 
the radiometric values was observed in range, and 
where floes had a low to medium pixel intensity. 
Algorithm C, (polynomial) filters the image to find 
pure samples of ice and open water from the original 
image. The filter is used for finding uni-modal 
samples of ice and water. When a uni-modal (one 
class) sample is found, the mean is saved. The 
dynamic range described by the saved "water means" 
and "ice means" are used in a migrating means 
procedure to generate a discriminant function for 
two-class separation. In this case the choice of filter 
size is critical as it will be directly related to the size 
of the features within the scene. In our testing, a 
filter size of 40 by 40 pixel was used. This filter size 
worked well in certain situations where the scene was 
not complex, yet failed more in scenes where few 
single class areas of that size could be found. With 
only a few means saved for each class the 
discriminant function was not accurately defined, and 
the resulting function was not based on true class 
information. 
The typical output from this algorithm was much 
noisier when compared with algorithm B. It is 
anticipated that preprocessing the imagery with a 
noise reduction filter would result in more 
homogenous areas of ice and open water classes. 
However, this would aggravate the problem of 
finding regions of pure class samples. 
Algorithm D (Mask) is identical to algorithm B 
except that it takes the processing one step further. 
Once two clusters have been established, each cluster 
is further subdivided into two, provided the total 
number of samples exceeds a threshold value. 
In complex scenes containing floes with a variety of 
pixel intensities (tones) and textures, this algorithm 
captured  mis-classified ice floes. All other 
algorithms misclassified the intermediate toned ice 
floes as open water. By further subdivision of 
classes the thin ice within leads in image number 3, 
and the wet-surfaced firstyear and multiyear ice in 
image number 5 were not incorrectly classified as 
432 
open wat 
class int 
where ba 
This als 
identifyii 
However 
contains 
an imprc 
should b 
The resul 
two class 
the size o 
the numb 
less favo 
number 3 
new ice ( 
tones). I 
water in 
Several i 
implemer 
processin 
interest, 
characteri 
and water 
The 500 
range ove 
] illustra 
incidence 
during the 
figure illt 
band sign 
that will t 
hand has : 
incidence 
in target 1 
is signifi 
winter tin 
where as 
Compared 
response i 
Spectral « 
computati 
importanc