above it 
: classes 
uses a 
irst step, 
ned from 
based on 
Iterating 
| the data 
until the 
ining the 
over the 
h the use 
describes 
function 
- an area 
iored. If 
he local 
or water 
) ice and 
| of the 
1991) 
74) 
ot al. 
arch 
981) 
  
arrays in 
rray one 
1e image 
ationship 
"thin the 
at each 
is a 4x4 
vel / -1. 
1b-arrays 
for level | +1 nodes. On each iteration the node is 
linked to a single one of these four higher nodes 
between level candidate parent nodes. After each 
node is linked, there will be between 0-16 
'legitimate' children. The non-parent links then 
define windows in the image, and ultimately the 
image segments. 
4.0 RESULTS 
All the algorithms evaluated used samples extracted 
from over the entire image to derive class 
boundaries. As a result, they all had difficulties with 
scenes that exhibited any systematic variation in 
radiometric values with range. The fall-off in the 
signal at the far range is due to a power loss and the 
change in backscatter from a target with imaging 
incidence angle. Both of these factors can be 
significant in airborne SAR systems, which have low 
imaging incidence angles (between 60°-85° for Star- 
2, and 50°-70° for the CV 580 scenes) but are 
expected to be less with satellite platforms such as 
Radarsat. However, Radarsat ScanSAR data which 
will be used by ICEC is unique in that it will be a 
mosaic of several imaging modes of the satellite 
covering a range of incidence angles (20°-49°). The 
concerns that this brings forth will be discussed latter 
in the text. 
Table 3 provides an overview of the strengths and 
weaknesses of each of the algorithms as determined 
by evaluation with the datasets selected. In 
summary, all algorithms performed marginally in 
complex scenes that contained floes with a range of 
tones. For example, multiyear ice in the Barrow 
Strait (image #5) had moderate to dark signatures due 
to the presence of surface melt water attenuating the 
signal. In most cases (standard mid-winter) multiyear 
ice will have a bright signature due to the dominance 
of volume scattering. In these instances, the spectral 
algorithms misclassified the flooded ice floes as open 
water. The problems which were observed in the 
algorithms’ handling of the datasets can be related to 
the inherent characteristics of the algorithms. These 
characteristics fell into primarily into two categories; 
1) filtering, and, 2) population bias. The effects of 
these characteristics are described fugther on an 
algorithm by algorithm basis. 
Algorithm A (entropy) generated easily interpreted 
output. Areas classified as ice or open water were 
homogenous regions unlike the output of many of the 
other classifiers. This characteristic is attributed to 
the filtering of the images prior to input into the 
Entropy classifier. In this evaluation filter sizes of 9 
by 9 and 11 by 11 were selected for examination. 
These filter sizes were considered optimum for speed 
of processing and acceptable level of image noise. 
The reduction of noise by filtering, however, that 
leads to more homogenous classifications, is traded- 
off with a loss of spatial detail. This is particularly 
evident for small features such as fractures, ridges, 
waves in ice and ice edges. Linear features are 
expanded and irregularly shaped features are reduced 
in extent. Furthermore, any illumination or range 
fall-off problems with the data are accentuated. On 
the other hand, in some instances, this attribute 
favours on the side of the algorithm. For example, 
areas of wave broken and flooded multiyear and 
firstyear ice floes with a similar signature as the open 
water in image #3 were correctly classified as sea 
ice. This is a result of fine detailed features (rubble 
and ridges) which had a strong return being blurred 
and identified as ice. Similarly, in the C-band scene 
(image #1) the filtering was instrumental in properly 
classifying the open water by subduing some of the 
strong texture observed in the lead in image # 1. 
This smoothing characteristic which result from 
filtering will have repercussions on any further value- 
added products, suck as ice concentration estimates, 
which may be generated from this output. 
Algorithm B (migrating means) achieved similar 
results to those observed from algorithm C 
(polynomial). The similarity is expected as the only 
difference between the two algorithms is the method 
by which they each extract samples from the image. 
Algorithm B uses all samples extracted from the 
imagery, while algorithm C uses only pure one-class 
samples. The input image and the classifier are 
identical for the two approaches. 
The output from this algorithm was noisy due to the 
input of the raw image and not a mean filtered image 
file, however, the delineation of fine detailed features 
was captured. Moreover, the migrating means and 
its derivatives (algorithm C and D) overestimate the 
amount of open water in all the test scenes. This 
tendency is attributed to the bias of the algorithm 
toward the class with the smallest population, which 
in our cases was open water. The discrimination 
function is drawn towards the largest mode of the 
histogram, resulting in a large number of that mode’s 
pixel being mis-classified . 
431