uncalibrated (Barber et al. 1993 , Wackerman 1988), 
SAR systems. A calibrated SAR system allows a 
direct relationship between the pixel value and the 
backscatter of the target. If knowledge of the 
backscatter characteristic each ice type is known, the 
classification process simply assigns each pixel to the 
most similar ice type. However, this approach relies 
on each ice type having a unique and known radar 
signature, which is often not the case, particularly for 
new ice types. A calibrated classification system, 
for use with ERS-1 data (Cvv), is currently in place 
at the Alaska SAR Facility (ASF). It is anticipated 
that a calibrated system will be implemented at ICEC 
in the future for use with RADARSAT data (Chh). 
An uncalibrated SAR system, on the other hand also 
relates the pixel intensities to the backscatter from the 
target, but includes contributions from the radar 
system itself, and the variation of the signal within 
the scene and between scenes. The latter is a 
function of the backscatter response of the target, 
unequal weighting of the antenna pattern, and 
modulation within the radar scene. 
The cornerstone of a fully automated ice classification 
system is the development of an ice/no ice classifier 
from which further value-added products, such as ice 
concentration can be generated. The fundamental 
component of the ice/no ice classifier will be the 
image segmentation algorithms. A human interpreter 
uses a variety of image attributes to segment an 
image, these include tone, texture, structure, shape, 
size and content (relationship between features). The 
tone, or pixel intensities, and texture properties are 
the lowest level elements that can be used to 
discriminate the major ice types within an image. 
Tone represents the backscatter (the amount of 
microwave energy reflected back to the sensor) while 
texture represents the spatial arrangement of the pixel 
intensities, which provides information on structure. 
The philosophy behind starting with elementary 
image properties is that the algorithms will be 
relatively simple and computationally fast, which is 
important to an operational implementation.d More 
complex properties can be added latter as needed. 
In this paper the results of an evaluation of five 
algorithms that use pixel intensities and local texture 
to separate ice from open water in SAR imagery are 
presented. These algorithms, referred to as spectral 
classifiers, were representative of algorithms listed to 
date in the literature which use tone and local texture 
to segment a SAR image. Section 2 describes the 
datasets which were used to evaluate the algorithms. 
The datasets are described in terms of their unique 
geophysical and sensor characteristics. Section 3 
describes the algorithms, and section 4 describes the 
processing results. In section 5 the causes of the 
algorithms successes and failures are explored, as 
well as what the anticipated ambiguities of Radarsat 
ScanSAR data will be and how they will affect the 
algorithms. The conclusions drawn from this 
research will be detailed in section 6. Finally, 
section 7 will comment on the future directions that 
need be considered. 
2.0 DATASETS 
Seven datasets were selected for evaluating the 
algorithms. These images provided a representative 
sample of the images that are analyzed by Ice Centre 
Operations. The characteristics of each of the 
datasets is described in table 1. The Ice Centre 
analyzes Star-2, X-band imagery (Falkingham, 1993), 
images numbered 3 to 7. In anticipation of Radarsat, 
a C-band image (image #1), collected by the CCRS 
Convair 580 SAR system (Livingstone et al. 1987) 
was included for evaluation. Finally, for comparison 
between the C- and X-band data, an X-band image, 
(image #2), collected by the CCRS SAR coincident 
with image number 1 was also included in the 
evaluation. All images were horizontally polarized 
(transmitted and received). 
The images illustrated tonal and textural variations 
within and between scenes,which are a function of 
the geographic location and season (i.e. ice type, ice 
surface wetness), and imaging characteristics (i.e. 
sensor frequency, range fall-off). 
3.0 ALGORITHMS 
A large number of spectral algorithms exist in the 
literature, each of which can be categorised by; 1) the 
discriminant function, 2) use of spatial statistics 
within a scene, and, 3) the strategy used to segment 
the scene. Five algorithms were selected for 
evaluation of their respective ability to separate ice 
from open water within an uncalibrated SAR image. 
428 
Table 1 
  
  
Loc 
& 
  
Gran 
East 
Marc 
  
Granc 
East 
Marc 
  
Beauf 
Augu: 
  
St-La 
Feb 
  
Barro 
High 
Augu: 
  
Jones 
High 
Feb 
  
  
M 
Labrac 
Jar 
1¢