~ienne
à la
i. On
niques
ar les
vastes
(e pas
| étre
blème.
nt été
llites
Lreurs
ner la
champs
Ces
on des
1881 à
cision
aphies
océdés
apport
ésager
de la
Evaluation of spectral classifiers for separating Sea Ice
from open water in preparation for Radarsat.
T. Heacock!, T. Hirose!, M. Manore?
1. Noetix Research Inc. 9^ Floor, 280 Albert St. Ottawa, Ontario, Canada, K1P 5G8.
2. Canada Centre for Remote Sensing (CCRS), Ice Applications Group, 588 Booth St, Ottawa, Ontario, Canada,
K1A 0Y7.
ABSTRACT
The launch of Radarsat in 1995 confirms Canada's commitment to the use of spaceborne radar data for the
monitoring of its land and ocean areas. Large volumes of data will be used by the sea ice community, and in
particular the Ice Centre Environment Canada to monitor the shipping lanes in and around Canada. In an effort
to assist in data analysis automated systems are being developed to extract value-added information products for
end users. To date, research has been conducted on methods of extracting ice information from both calibrated and
uncalibrated SAR data. The first step in the development of a fully automated system is the separation of areas of
ice and water. Five algorithms that use only tone and local scene texture extracted from an uncalibrated SAR scene
(so called spectral algorithms) were evaluated. These algorithms, representative of all spectral algorithms, were
selected because of their computational speed and efficiency. The results illustrated that all the algorithms have
the ability to separate ice from water in SAR imagery under ideal conditions. The performance varied on an image
to image basis dependant on the amount of systematic or geophysical variability within each scene used for the
evaluation. Observations about the relative strengths and weaknesses of the algorithms are made.
KEY WORDS
Synthetic Aperture Radar, Sea Ice, Classification, Algorithm, Tone, Texture.
Environment Canada (ICEC), and it's client services
1.0 INTRODUCTION
Remotely sensed data is a practical and cost-effective
method of acquiring detailed, timely information over
Canada's ice-infested waters. With the launch of
RADARSAT in early 1995 (Langham 1992), large
volumes of Synthetic Aperture Radar (SAR) data will
be acquired over Canada's oceans. It is anticipated
that manual analysis of the data on a regional scale
will be overwhelming. To overcome this potential
bottleneck, researchers have proposed the
development of automated systems to generate
products for end users in all applications disciplines,
and in particular for sea ice applications because of
the operational nature of the Ice Centre,
(Ramsay et al. 1993).
An automated system for classifying Radarsat
imagery must be able operate on data acquired from
a single channel, single polarized SAR (C-band hh).
The system must be robust enough to operate on
imagery illustrating a diversity of ice conditions,
collected from all seasons and all geographic regions
of operational interest to ICEC. Furthermore, the
data product which will be used by ICEC operations
will be ScanSAR data, generated through a quick
delivery system. This data will contain image
ambiguities that are a function of the unique nature of
the ScanSAR image creation process.
To date, ice classification research has been carried
out for both calibrated (Kwok et al. 1992) and
427