10 Polar SAR Data for Operational Sea Ice Mapping
229
but fast analysis. At the operationally important ice margins, a more sophisticated classifier, such as an expert system, or an experienced human analyst, will take over.
10.8
Automated Ice Motion Tracking
10.8.1
The Development of Automated Ice Motion Algorithms
Operationally, ice motion vectors are useful because they indicate areas of ice convergence and divergence. Motion vectors are also useful in forecasting ice drift. Compared
with ice classification, tracking ice motion using SAR is a relatively straightforward
problem because floes or recognizable patterns tend to maintain their appearance in
successive images. An algorithm need only identify the same feature and calculate its
displacement in two georeferenced images. Fily and Rothrock (1987) first implemented a method for accomplishing this automatically by finding an area in two images separated in time in which the pattern of pixels was highly correlated. Subsequent work
(e.g. Collins and Emery 1988; Vesecky et al.1988) refined the method and documented
its limitations. This area correlation method works well for the central Arctic where ice
undergoes displacement with little rotation. In the MIZ, however, floes may rotate, and
it is necessary to use an algorithm that first extracts individual floes or other distinct
features and identifies them with rotation invariant parameters, such as floe outline
shape. Daida et aI., (1990), Banfield (1991), and McConnell et aI. (1991) propose methods for this kind of feature tracking.
The ice tracking algorithm developed for ASF (Kwok et aI. 1990) was designed to use
area correlation for the central Arctic and feature tracking for ice in the MIZ. If area
correlation failed (that is, the number of erroneous vectors became high), the feature
tracker estimated rotation and translation. Evaluation of approximately 80 ERS-1
image pairs (R. Kwok pers. comm. May 1996) the feature tracker worked well when a
few large floes were present in the scenes. It failed, however, in dynamic areas such as
the Bering Sea, where small floe sizes and the changeable nature of ice and water signatures made it impossible for the segmentation scheme to extract consistent feature
boundaries for the tracker.
In the central Arctic in winter, the algorithm worked quite well. Currently products
from the algorithm are being used to study the deformation characteristics of ice (Li
et al., this volume), and the algorithm forms the basis of the RGPS (Kwok, this volume).
The performance of both area correlation and feature tracking degrades in summer
because ice appears nearly featureless in SAR imagery and the signature of open water
is variable. A neural-network-based feature tracker is being developed at the University of Kansas to improve tracking capability in the MIZ and in summer (Silveira et aI.
1994).
10.8.2
SAR Ice Motion Algorithms at US and Canadian Operational Centers
The NIC has adopted the ASF ice motion algorithm but it was seldom used with ERS-1
data because of the difficulty in rapidly acquiring SAR imagery in which the same ice
229
but fast analysis. At the operationally important ice margins, a more sophisticated classifier, such as an expert system, or an experienced human analyst, will take over.
10.8
Automated Ice Motion Tracking
10.8.1
The Development of Automated Ice Motion Algorithms
Operationally, ice motion vectors are useful because they indicate areas of ice convergence and divergence. Motion vectors are also useful in forecasting ice drift. Compared
with ice classification, tracking ice motion using SAR is a relatively straightforward
problem because floes or recognizable patterns tend to maintain their appearance in
successive images. An algorithm need only identify the same feature and calculate its
displacement in two georeferenced images. Fily and Rothrock (1987) first implemented a method for accomplishing this automatically by finding an area in two images separated in time in which the pattern of pixels was highly correlated. Subsequent work
(e.g. Collins and Emery 1988; Vesecky et al.1988) refined the method and documented
its limitations. This area correlation method works well for the central Arctic where ice
undergoes displacement with little rotation. In the MIZ, however, floes may rotate, and
it is necessary to use an algorithm that first extracts individual floes or other distinct
features and identifies them with rotation invariant parameters, such as floe outline
shape. Daida et aI., (1990), Banfield (1991), and McConnell et aI. (1991) propose methods for this kind of feature tracking.
The ice tracking algorithm developed for ASF (Kwok et aI. 1990) was designed to use
area correlation for the central Arctic and feature tracking for ice in the MIZ. If area
correlation failed (that is, the number of erroneous vectors became high), the feature
tracker estimated rotation and translation. Evaluation of approximately 80 ERS-1
image pairs (R. Kwok pers. comm. May 1996) the feature tracker worked well when a
few large floes were present in the scenes. It failed, however, in dynamic areas such as
the Bering Sea, where small floe sizes and the changeable nature of ice and water signatures made it impossible for the segmentation scheme to extract consistent feature
boundaries for the tracker.
In the central Arctic in winter, the algorithm worked quite well. Currently products
from the algorithm are being used to study the deformation characteristics of ice (Li
et al., this volume), and the algorithm forms the basis of the RGPS (Kwok, this volume).
The performance of both area correlation and feature tracking degrades in summer
because ice appears nearly featureless in SAR imagery and the signature of open water
is variable. A neural-network-based feature tracker is being developed at the University of Kansas to improve tracking capability in the MIZ and in summer (Silveira et aI.
1994).
10.8.2
SAR Ice Motion Algorithms at US and Canadian Operational Centers
The NIC has adopted the ASF ice motion algorithm but it was seldom used with ERS-1
data because of the difficulty in rapidly acquiring SAR imagery in which the same ice