roperly 
  
ge fall 
  
jorithms 
  
iate 
  
  
lass the 
ined, and 
rue class 
'as much 
3. Tti 
y with a 
in more 
classes. 
blem of 
yithm B 
) further. 
:h cluster 
the total 
ue. 
;ariety of 
algorithm 
Il other 
toned ice 
vision of 
umber 3, 
ar ice in 
sified as 
open water. However, the division of the open water 
class into two classes was problematic in scenes 
where both subclasses were open water. 
This algorithm provided a unique method for 
identifying ice features that had a range of tones. 
However, a method to determine when a class 
contains a mixture of ice and open water, as well as 
an improved mechanism to decide when a cluster 
should be subdivided would improve the algorithm. 
The results of algorithm E (Hierarchical Network) for 
two class separation indicated promising results over 
the size of areas processed (256 by 256 pixel). When 
the number of classes was increased the results were 
less favourable. The subarea extracted from image 
number 3 included multi-year ice floes (bright tones), 
new ice (medium tones), and open water leads (dark 
tones). In this case the distinction between the open 
water in the leads and the new ice was not possible. 
5.0 RADARSAT ISSUES 
Several issues need to be addressed with respect to 
implementing spectrally based algorithms for the 
processing of Radarsat Scansar data. Of particular 
interest, in the context of this paper, are the 
characteristics of the backscatter coefficients for ice 
and water in c-band horizontally polarized SAR data. 
The 500 km swath of Radarsat ScanSAR data will 
range over incidence angles from 20?-50?. Figure 
1 illustrates the radar backscatter as a function of 
incidence angle for both C- and X-band SAR data 
during the winter and summer (Onstott, 1992). The 
figure illustrates a large amount of variability in C- 
band signatures over the range of incidence angles 
that will be collected. The X-band data on the other 
hand has a relatively uniform return over those same 
incidence angles. For C-band imagery, the variation 
in target response with range for both ice and water 
is significant. For example, for calm water the 
winter time response at 20° is approximately -15dB, 
where as at 50° is -40dB; a variability 25dB. 
Compared with X-band where the variability in 
response is only 10dB (from -25dB to -35dB). 
6.0 CONCLUSIONS 
Spectral classifiers were selected because of their 
computational speed and efficiency, which is of great 
importance for operational implementation. The five 
algorithms that were selected for evaluation were 
representative of all classes of spectral algorithms 
described to date in the literature. Seven image 
scenes from different geographic regions, seasons and 
acquired by different sensors provided a broad dataset 
with which to evaluate the properties of each class of 
algorithm and define their strengths and limitations. 
These data allowed for the evaluation of systematic 
effects observed in uncalibrated SAR data, and 
geophysical features that confuse the signatures of ice 
and water. 
The results indicate that all algorithms have the 
ability to separate ice from water under ideal 
conditions. The performance of each individual 
algorithm on an image to image basis was variable 
depending on the degree of systematic or geophysical 
variability within the image. In particular, variations 
in backscatter as functions of incidence angle, or ice 
type, combined with the characteristics of parameters 
within the algorithm, (such as filter size), all 
combined to generated the results that were achieved. 
This work was an important first step in the 
development of an operational algorithm for the 
separation of ice from water within uncalibrated SAR 
data. The evaluations afforded a greater 
understanding of the issues that need be addressed. 
Several recommendations for further work need be 
examined, they include; 
An approach to process subareas of the image and 
produce a seamless output is required to overcome 
the systematic variation in the signal with range. 
All the tested algorithms extract samples from an 
image without regard to their location and generate 
one discriminant function which is applied to all data 
within that scene. However, when the imagery 
contains systematic variation in pixel intensities in 
range, a single discriminate function is inadequate. 
Characterisation of the expected stability of the 
image backscatter (intensities) from scene to scene 
should be conducted for all SAR systems to be 
used by a classifier. 
A priori knowledge on image characteristics which 
will be stable from scene to scene can be used by an 
algorithm to improve its classification accuracy. This 
applies to the range of intensities for each ice type as 
well as systematic radiometric anomalies introduced 
by the sensor and image formation processes. 
433