10 Polar SAR Data for Operational Sea Ice Mapping
223
tie points for SSM/I ice type estimates was proposed by Fetterer et al. (1993). Further
investigation showed that high variability in SSM/I multiyear ice signatures meant that
the calibration could not be extended much beyond the specific area of the SAR pass.
With the exception of Kalman filtering for blending buoy data with passive microwave
data (Thomas and Rothrock 1993), data blending has yet to be demonstrated on an operationally viable scale.
Much of the ground-laying research on SAR ice signatures was performed using aircraft SAR and SLAR data (e.g., Ketchum 1977; Onstott et al. 1982). Ice features such as
ridges are apparent in aircraft SAR imagery but are not always distinguishable in satellite SAR imagery, although if present these features contribute to overall pixel backscatter. With satellite imagery, more than one ice type often contributes to an image pixel.
These factors hamper the ability to distinguish surface types. Based on previous aircraft
SAR and ground-based scatterometer measurements, it was reasonable to expect to be
able to separate different types of younger ice from multiyear ice in ERS-1 imagery
(Kwok et al. 1992). However, it was discovered that it is not generally possible to distinguish between new, young, and first-year ice using backscatter alone. New and young ice
display significant signature variation, and small scale inhomogeneities serve to smear
the backscatter distributions for these classes across one another (Fetterer et al.1994).
10.6.1
Automated Ice Classification with ERS-1 at NIC
An ice type classification algorithm was developed to produce ice type maps at the Alaska SAR Facility (ASF). The US Navy funded the adaptation of this ice algorithm for NIC,
a major step toward establishing automated ice analysis on a larger scale with
RADARSAT data. Some problems encountered in applying the algorithm are described
below.
The algorithm [described by Kwok et al. (1992)] is based on a look-up table of
backscatter distributions for four ice types: multiyear (MY), first-year rough (FYR),
first-year smooth (FYS) and new ice/open water (NI/OW). The algorithm clusters a sample of pixels from an image. The centroid of the brightest cluster, if within 2 dB of -10
dB, is assumed to give the mean MY backscatter for the image. The look-up table distributions are then fixed using the mean for MY ice. The mean for FYR ice falls 4 dB
below that of MY, and the means of FYS and NI/OW are lower still. This "sliding scale"
feature makes the algorithm robust to errors in absolute calibration. Bayesian maximum likelihood classification is then used to assign each pixel to a class based on the
look-up table distributions.
The performance of the algorithm in the Beaufort Sea is assessed by Fetterer et al.
(1994). This assessment was extended to the East Siberian Sea by Gineris and Fetterer
(1994}.Algorithm classification results were compared to those of a supervised maximum likelihood classification using manually selected training sets on an
image-by-image basis. This method has two drawbacks: manually classified images are
not validated with surface observations, and manually supervised classification with
training sets is not as accurate as strictly manual interpretation in which an analyst outlines areas of different ice types using contextual as well as backscatter clues.
The number of images (68 in the Beaufort, 86 in the East Siberian Sea) in the assessment made it necessary to use supervised classification. Therefore, the analysis is actu-
223
tie points for SSM/I ice type estimates was proposed by Fetterer et al. (1993). Further
investigation showed that high variability in SSM/I multiyear ice signatures meant that
the calibration could not be extended much beyond the specific area of the SAR pass.
With the exception of Kalman filtering for blending buoy data with passive microwave
data (Thomas and Rothrock 1993), data blending has yet to be demonstrated on an operationally viable scale.
Much of the ground-laying research on SAR ice signatures was performed using aircraft SAR and SLAR data (e.g., Ketchum 1977; Onstott et al. 1982). Ice features such as
ridges are apparent in aircraft SAR imagery but are not always distinguishable in satellite SAR imagery, although if present these features contribute to overall pixel backscatter. With satellite imagery, more than one ice type often contributes to an image pixel.
These factors hamper the ability to distinguish surface types. Based on previous aircraft
SAR and ground-based scatterometer measurements, it was reasonable to expect to be
able to separate different types of younger ice from multiyear ice in ERS-1 imagery
(Kwok et al. 1992). However, it was discovered that it is not generally possible to distinguish between new, young, and first-year ice using backscatter alone. New and young ice
display significant signature variation, and small scale inhomogeneities serve to smear
the backscatter distributions for these classes across one another (Fetterer et al.1994).
10.6.1
Automated Ice Classification with ERS-1 at NIC
An ice type classification algorithm was developed to produce ice type maps at the Alaska SAR Facility (ASF). The US Navy funded the adaptation of this ice algorithm for NIC,
a major step toward establishing automated ice analysis on a larger scale with
RADARSAT data. Some problems encountered in applying the algorithm are described
below.
The algorithm [described by Kwok et al. (1992)] is based on a look-up table of
backscatter distributions for four ice types: multiyear (MY), first-year rough (FYR),
first-year smooth (FYS) and new ice/open water (NI/OW). The algorithm clusters a sample of pixels from an image. The centroid of the brightest cluster, if within 2 dB of -10
dB, is assumed to give the mean MY backscatter for the image. The look-up table distributions are then fixed using the mean for MY ice. The mean for FYR ice falls 4 dB
below that of MY, and the means of FYS and NI/OW are lower still. This "sliding scale"
feature makes the algorithm robust to errors in absolute calibration. Bayesian maximum likelihood classification is then used to assign each pixel to a class based on the
look-up table distributions.
The performance of the algorithm in the Beaufort Sea is assessed by Fetterer et al.
(1994). This assessment was extended to the East Siberian Sea by Gineris and Fetterer
(1994}.Algorithm classification results were compared to those of a supervised maximum likelihood classification using manually selected training sets on an
image-by-image basis. This method has two drawbacks: manually classified images are
not validated with surface observations, and manually supervised classification with
training sets is not as accurate as strictly manual interpretation in which an analyst outlines areas of different ice types using contextual as well as backscatter clues.
The number of images (68 in the Beaufort, 86 in the East Siberian Sea) in the assessment made it necessary to use supervised classification. Therefore, the analysis is actu-