222
C. BERTOIA, J. FALKINGHAM, F. FETTERER
10.6
Automated Ice Classification
Producing ice type and concentration estimates automatically from SAR data has been
a longstanding goal of both the operational and scientific communities. Scientists
desire a stable, long-term data record of ice type and concentration estimates with
known accuracy in order to address questions concerning ice mass balance and heat
flux. The most feasible way to analyze the volume of SAR imagery required for a climatology of ice characteristics is to use an automated algorithm.
Operational centers require automated algorithms for the same reason: the volume
of SAR imagery received precludes manual analysis. At the same time, the high resolution of SAR and the distinctiveness of forms of ice (floe size, shape, and deformation)
in SAR imagery give the analyst more operationally useful information than any other satellite sensor. In the US, the science community has benefited from the operational
requirements for and Department of Defense (DoD) funding of SAR ice signature and
automated image analysis research. In Canada, the Canadian Centre for Remote Sensing (CCRS) and CIS have directly funded ice type algorithm development work.
The operational community has requirements that the science community does not
share. For instance, algorithms must run quickly (NIC anticipates receiving about 3 Gb
RADARSAT imagery per day), and data cannot be analyzed retrospectively (although
past images may aid in classifying a current image). While many studies have shown
that multifrequency or multipolarization SAR data can improve classification (e.g.,
Drinkwater et al.1992; Rignot and Drinkwater 1994; Collins and Livingstone 1996), the
ice centers are not likely to have access to operationally useful multichannel data for
decades. Another difference is that the ice centers are required to map ice features. Even
if it could be produced in near-real time, the RGPS Arctic Snapshot product (Kwok, this
volume) would be oflimited use because it does not show the location of ice type features, providing only a distribution of ice type concentrations within 5-km-grid cells.
Operational algorithm development is focused on rapid segmentation of an image
so that each pixel is accurately assigned to an ice type or to water. A much more difficult task is automatically extracting and labeling features such as floes, leads, and
ridges. Extensive work in this area has been done by Banfield and Raftery (1989) and
Vesecky et al. (1990). Data blending, or combining SAR with other satellite data and
meteorological analyses to assess ice conditions, is another area which promises operationally useful results. As pointed out Sect. 10.5, analysts currently blend data subjectively and manually. This blending can potentially be performed automatically, using
a formal framework such as an expert system or probabilistic inference (in which evidence given by data from different sources, or degree of belief in data, dictates the decision concerning the surface type present). Collins (1992) gives a summary of these and
other methods of data blending for sea ice analysis. Currently, all satellite data and meteorological analyses are available as digital data on workstations at both CIS and NIC -
the first step in operational data blending.
Beyond the problem of simply blending coincident ice information from different
sources is the problem of optimally assimilating and interpolating information which
is received at the ice centers irregularly in space and time and at resolutions ranging
from 25 km for passive microwave data to 100 m for SAR to a point measurement for
buoy data. A form of data blending in which ERS-1 estimates are used for "calibration"
C. BERTOIA, J. FALKINGHAM, F. FETTERER
10.6
Automated Ice Classification
Producing ice type and concentration estimates automatically from SAR data has been
a longstanding goal of both the operational and scientific communities. Scientists
desire a stable, long-term data record of ice type and concentration estimates with
known accuracy in order to address questions concerning ice mass balance and heat
flux. The most feasible way to analyze the volume of SAR imagery required for a climatology of ice characteristics is to use an automated algorithm.
Operational centers require automated algorithms for the same reason: the volume
of SAR imagery received precludes manual analysis. At the same time, the high resolution of SAR and the distinctiveness of forms of ice (floe size, shape, and deformation)
in SAR imagery give the analyst more operationally useful information than any other satellite sensor. In the US, the science community has benefited from the operational
requirements for and Department of Defense (DoD) funding of SAR ice signature and
automated image analysis research. In Canada, the Canadian Centre for Remote Sensing (CCRS) and CIS have directly funded ice type algorithm development work.
The operational community has requirements that the science community does not
share. For instance, algorithms must run quickly (NIC anticipates receiving about 3 Gb
RADARSAT imagery per day), and data cannot be analyzed retrospectively (although
past images may aid in classifying a current image). While many studies have shown
that multifrequency or multipolarization SAR data can improve classification (e.g.,
Drinkwater et al.1992; Rignot and Drinkwater 1994; Collins and Livingstone 1996), the
ice centers are not likely to have access to operationally useful multichannel data for
decades. Another difference is that the ice centers are required to map ice features. Even
if it could be produced in near-real time, the RGPS Arctic Snapshot product (Kwok, this
volume) would be oflimited use because it does not show the location of ice type features, providing only a distribution of ice type concentrations within 5-km-grid cells.
Operational algorithm development is focused on rapid segmentation of an image
so that each pixel is accurately assigned to an ice type or to water. A much more difficult task is automatically extracting and labeling features such as floes, leads, and
ridges. Extensive work in this area has been done by Banfield and Raftery (1989) and
Vesecky et al. (1990). Data blending, or combining SAR with other satellite data and
meteorological analyses to assess ice conditions, is another area which promises operationally useful results. As pointed out Sect. 10.5, analysts currently blend data subjectively and manually. This blending can potentially be performed automatically, using
a formal framework such as an expert system or probabilistic inference (in which evidence given by data from different sources, or degree of belief in data, dictates the decision concerning the surface type present). Collins (1992) gives a summary of these and
other methods of data blending for sea ice analysis. Currently, all satellite data and meteorological analyses are available as digital data on workstations at both CIS and NIC -
the first step in operational data blending.
Beyond the problem of simply blending coincident ice information from different
sources is the problem of optimally assimilating and interpolating information which
is received at the ice centers irregularly in space and time and at resolutions ranging
from 25 km for passive microwave data to 100 m for SAR to a point measurement for
buoy data. A form of data blending in which ERS-1 estimates are used for "calibration"