224
C. BERTOIA, J. FALKINGHAM, F. FETTERER
ally a measure of how well the look-up table distributions match the actual backscatter distributions for the four ice types. This is valuable information because any
between- or within-class variability (that is, changes in the means or standard deviations of the backscatter classes) increases algorithm error.
It was understood at the outset that the algorithm would fail to identify new ice and
open water because these classes have a backscatter signature which can overlap that
of all other classes. Fortunately, these classes make up only a small percentage of the
area covered in the algorithm domain (winter, far from the MIZ). Other errors revealed
by the assessment included:
Inability to distinguish forms of ice younger than MY. While analysts could usually distinguish between two distinctly different types of FY ice ( one brighter overall, one with
lower apparent ridge density), the backscatter for these classes overlapped to such an
extent that automated classification failed. Backscatter distributions for images of winter ice in the Beaufort Sea tended to be bimodal, with one peak for MY and one peak
for all other ice.
Overestimate of FY ice concentration at high latitudes. The algorithm erroneously
classified MY ice pixels at the low end of the MY backscatter distribution as FY, unless
all MY pixels were above the FY/MY decision threshold of about -12 dB. The error
became more apparent at high latitudes, where the concentration of MY is usually
greater than 95%. When FY ice is present in sufficient quantities (generally at lower latitudes), this error tends to be counteracted by a corresponding misclassification of FY
ice as MY, so that the percentages of the FY and MY for the image as a whole are about
right.
Misclassification of FY ice in the MIZ. Large expanses of FY ice in the Chuckchi Sea
had a backscatter of about -12 dB, leading to misclassification as MY ice. At 69°N, this
area is south of the stated algorithm domain, but the reason why this ice was brighter
than "average" FY ice is not known.
Degradation of performance in the East Siberian Sea. Regional variability in the
backscatter of MY ice is reported by Kwok and Cunningham (1994). In the Beaufort Sea
data set, Fetterer et al. (1994) found that changes in MY backscatter over latitude were
matched by corresponding changes in FY ice backscatter. Therefore, the separation
between types remained consistent and the sliding scale feature of the algorithm preserved classification accuracy. The root mean square error difference between algorithm
and supervised classification for MY /FY separation was only 5%. The backscatter of FY
and MY in the East Siberian Sea sector of the Arctic was less well-behaved. Figure 10
displays backscatter samples for different ice types as a function oflongitude. Note the
increasing variability in MY backscatter from east to west, and the greater overlap
between FY and MY ice west of 175 oW. Statistics from ice backscatter samples show that
the standard deviation of MY backscatter is relatively high in the East Siberian Sea, and
the distribution is skewed, with the median lower than the mean. This leads to root
mean square error rates that are 6-12% higher than the 5% found in the Beaufort Sea.
In the East Siberian Sea, small, rounded MY floes appear embedded in a matrix of mixed
ice types. The Beaufort Sea SAR imagery does not exhibit this matrix.
C. BERTOIA, J. FALKINGHAM, F. FETTERER
ally a measure of how well the look-up table distributions match the actual backscatter distributions for the four ice types. This is valuable information because any
between- or within-class variability (that is, changes in the means or standard deviations of the backscatter classes) increases algorithm error.
It was understood at the outset that the algorithm would fail to identify new ice and
open water because these classes have a backscatter signature which can overlap that
of all other classes. Fortunately, these classes make up only a small percentage of the
area covered in the algorithm domain (winter, far from the MIZ). Other errors revealed
by the assessment included:
Inability to distinguish forms of ice younger than MY. While analysts could usually distinguish between two distinctly different types of FY ice ( one brighter overall, one with
lower apparent ridge density), the backscatter for these classes overlapped to such an
extent that automated classification failed. Backscatter distributions for images of winter ice in the Beaufort Sea tended to be bimodal, with one peak for MY and one peak
for all other ice.
Overestimate of FY ice concentration at high latitudes. The algorithm erroneously
classified MY ice pixels at the low end of the MY backscatter distribution as FY, unless
all MY pixels were above the FY/MY decision threshold of about -12 dB. The error
became more apparent at high latitudes, where the concentration of MY is usually
greater than 95%. When FY ice is present in sufficient quantities (generally at lower latitudes), this error tends to be counteracted by a corresponding misclassification of FY
ice as MY, so that the percentages of the FY and MY for the image as a whole are about
right.
Misclassification of FY ice in the MIZ. Large expanses of FY ice in the Chuckchi Sea
had a backscatter of about -12 dB, leading to misclassification as MY ice. At 69°N, this
area is south of the stated algorithm domain, but the reason why this ice was brighter
than "average" FY ice is not known.
Degradation of performance in the East Siberian Sea. Regional variability in the
backscatter of MY ice is reported by Kwok and Cunningham (1994). In the Beaufort Sea
data set, Fetterer et al. (1994) found that changes in MY backscatter over latitude were
matched by corresponding changes in FY ice backscatter. Therefore, the separation
between types remained consistent and the sliding scale feature of the algorithm preserved classification accuracy. The root mean square error difference between algorithm
and supervised classification for MY /FY separation was only 5%. The backscatter of FY
and MY in the East Siberian Sea sector of the Arctic was less well-behaved. Figure 10
displays backscatter samples for different ice types as a function oflongitude. Note the
increasing variability in MY backscatter from east to west, and the greater overlap
between FY and MY ice west of 175 oW. Statistics from ice backscatter samples show that
the standard deviation of MY backscatter is relatively high in the East Siberian Sea, and
the distribution is skewed, with the median lower than the mean. This leads to root
mean square error rates that are 6-12% higher than the 5% found in the Beaufort Sea.
In the East Siberian Sea, small, rounded MY floes appear embedded in a matrix of mixed
ice types. The Beaufort Sea SAR imagery does not exhibit this matrix.