242
R.KwOK
rithm works only during the winter and also that it provides a fine age resolution of
only the young end of the age distribution. We refer to the time interval between
sequential images as a time step. During each time step, the cell areas are computed. A
positive change indicates that new ice was formed in the cell. A negative change is
assumed to have ridged the youngest ice in the cell, reducing its area. The assumption
here is that once ridging starts, the deformation tends to be localized in the thinner,
weaker ice that recently formed in the lead systems. The age classes are determined by
the lengths of the time steps. The area of ice in each age class in each cell is updated at
each time step. In this manner, we keep track of the age distribution of the young ice
within a cell.
The area of multiyear ice in each cell is also computed at each time step using the
ice classification algorithm described in Kwok et al. (1992). The ice type algorithm
uses a maximum likelihood classifier and a look-up table of expected backscatter
characteristics to assign each image pixel to one of four classes: multiyear ice,
deformed first-year ice, undeformed first-year ice, and a low-backscatter type characteristic of smooth, younger ice types and calm open water. Two possible sources
of error are from wind-roughened open water and from frost flowers growing on
new ice. Both of these physical phenomena exhibit highly variable radar backscatter, causing the classifier to label these pixels sometimes incorrectly as multiyear ice
because of overlapping signatures. We use the time series of multiyear ice area for
each cell to resolve these ambiguities, resulting in a more accurate classification.
Since the area of multiyear ice within a cell should remain constant (no multiyear
ice is created in the winter), any anomaly which shows up as a transient spike or
hump could be filtered out. The ice classification algorithm is not used to identify
different types of first-year ice because the accuracy of the classifier is lower for these
ice types. Since the areas of young ice and multiyear ice in each cell are accounted
for by the above procedures and the total cell area is known from its geometry, the
residual area is simply labeled as first-year ice. For a series of five images with, say,
3 days between successive images, the age classes would be: 0-3 days, 3-6 days, 6-9
days, 9-12 days, first-year ice, and multiyear ice.
With records of the near-surface air temperature, the young end of the age distribution is converted to a thickness distribution using a simple empirical relation
between accumulated freezing-degree days and ice thickness (Maykut 1986). We
note that this scheme does not provide estimates of the thickness of first-year or
multiyear ice. Our present method provides a two-dimensional, potentially basinwide picture of the thickness of young ice, although it does not give any information about the mean thickness of the ice cover as a whole, since young ice occupies
only a small fraction of the total area. The main output product of our age analysis procedure is the thickness distribution of young ice at a fine spatial resolution
and the areal fraction of first-year and multiyear ice at regular time intervals. Further details of how ice age is related to cell area changes can be found in Kwok et
al. (1995).
We select two examples to illustrate the procedure used for ice age computation within a cell. These cells are extracted from ERS-l image sequences of the central Beaufort
Sea acquired during 1992. The time series spans a period of 12 days from March 18 to
March 30. The sampling interval or time step of the sequence is 3 days. The initial cell
size is 5 X 5 km. We assume, in the two examples shown here, that the initial distribu-
R.KwOK
rithm works only during the winter and also that it provides a fine age resolution of
only the young end of the age distribution. We refer to the time interval between
sequential images as a time step. During each time step, the cell areas are computed. A
positive change indicates that new ice was formed in the cell. A negative change is
assumed to have ridged the youngest ice in the cell, reducing its area. The assumption
here is that once ridging starts, the deformation tends to be localized in the thinner,
weaker ice that recently formed in the lead systems. The age classes are determined by
the lengths of the time steps. The area of ice in each age class in each cell is updated at
each time step. In this manner, we keep track of the age distribution of the young ice
within a cell.
The area of multiyear ice in each cell is also computed at each time step using the
ice classification algorithm described in Kwok et al. (1992). The ice type algorithm
uses a maximum likelihood classifier and a look-up table of expected backscatter
characteristics to assign each image pixel to one of four classes: multiyear ice,
deformed first-year ice, undeformed first-year ice, and a low-backscatter type characteristic of smooth, younger ice types and calm open water. Two possible sources
of error are from wind-roughened open water and from frost flowers growing on
new ice. Both of these physical phenomena exhibit highly variable radar backscatter, causing the classifier to label these pixels sometimes incorrectly as multiyear ice
because of overlapping signatures. We use the time series of multiyear ice area for
each cell to resolve these ambiguities, resulting in a more accurate classification.
Since the area of multiyear ice within a cell should remain constant (no multiyear
ice is created in the winter), any anomaly which shows up as a transient spike or
hump could be filtered out. The ice classification algorithm is not used to identify
different types of first-year ice because the accuracy of the classifier is lower for these
ice types. Since the areas of young ice and multiyear ice in each cell are accounted
for by the above procedures and the total cell area is known from its geometry, the
residual area is simply labeled as first-year ice. For a series of five images with, say,
3 days between successive images, the age classes would be: 0-3 days, 3-6 days, 6-9
days, 9-12 days, first-year ice, and multiyear ice.
With records of the near-surface air temperature, the young end of the age distribution is converted to a thickness distribution using a simple empirical relation
between accumulated freezing-degree days and ice thickness (Maykut 1986). We
note that this scheme does not provide estimates of the thickness of first-year or
multiyear ice. Our present method provides a two-dimensional, potentially basinwide picture of the thickness of young ice, although it does not give any information about the mean thickness of the ice cover as a whole, since young ice occupies
only a small fraction of the total area. The main output product of our age analysis procedure is the thickness distribution of young ice at a fine spatial resolution
and the areal fraction of first-year and multiyear ice at regular time intervals. Further details of how ice age is related to cell area changes can be found in Kwok et
al. (1995).
We select two examples to illustrate the procedure used for ice age computation within a cell. These cells are extracted from ERS-l image sequences of the central Beaufort
Sea acquired during 1992. The time series spans a period of 12 days from March 18 to
March 30. The sampling interval or time step of the sequence is 3 days. The initial cell
size is 5 X 5 km. We assume, in the two examples shown here, that the initial distribu-