12 Towards Operational Monitoring of Arctic Sea Ice by SAR
erally distinguishable but independent rotation of floes is also likely. Algorithms that
perform well with one type of ice are generally not best suited for determining motion
in other ice conditions. Following the example of the Alaskan SAR Facility (Kwok et al.
1990; Rothrock et al. 1992), it was therefore decided to adopt a hybrid solution for the
IPAP system: i.e., to develop one algorithm of each type to be selected as best appropriate to the prevailing ice conditions.
The main methods that have been developed which are appropriate for pack ice are
area correlation (Fily and Rothrock 1986, 1987), matched filtering (Ninnis et al. 1986;
Collins and Emery 1988), and optical flow methods (Yan 1992). Of these algorithms,
the area correlation method has been the most widely used, and as one of the algorithms implemented under IPAP is fully described in Sect. 12.3.2 below. Since the main
drawback of the method is in terms of computational efficiency, the matched filtering
technique (in which the correlation is performed in the Fourier domain) would seem
to hold out significant potential. In particular, the cross correlation function of the
power spectra might be used as a first step to rotationally correct the images prior to
applying a conventional area correlation, the power spectra being related by a rotation which is the same as the relative rotation between the two images in the spatial
domain [Fily and Rothrock (1987) estimate that rotations beyond 7" will start to affect
the correlation values obtained]. For the optical flow method, in which the components of ice velocity are calculated at each image pixel in terms of partial derivatives
(temporal and spatial) of the image intensity, the assumptions made regarding stationary image intensity and gradient are likely to be valid only for small velocities of
the order of a few pixels, and so this method may not be appropriate for large-scale
ice displacement.
The main methods that have been developed which are appropriate for marginal
ice are tie-point extrapolation (Zhang et al. 1989), edge-feature matching (Vesecky et
al. 1988; McConnell et al. 1991), postsegmentation region matching (see Sect. 12. 3.3),
invariant moment matching (Zhang 1992), and stochastic approaches (Banfield
1991). For the tie-point extrapolation method, the degree of manual intervention
makes it too time-consuming for routine use within an operational environment. The
other approaches can all be considered as variations on a common theme, i.e. matching common features (generally floes) between images based on some measure of
shape. For each of these techniques, the primary requirement is the input of a good,
repeatable segmentation. Assuming that there is no noise in the boundary determination, then the best representation of the region shape will be obtained from one of
the edge-feature matching techniques which characterize region boundaries at the
pixel level. At the other extreme, the invariant moment matching method is based on
a parameterization of the region shape in terms of its moments and, whilst being less
sensitive to boundary noise, would not be able to cope with tracking regions that split
or merge. The stochastic approach represents an intermediate solution in that floe
boundaries are approximated by principal curves from which edge-pixel probability distributions are derived. The postsegmentation region matching scheme also
offers an intermediate solution in that segmented regions are approximated to by
polygons (of user-defined maximum deviation from the true boundary position), but
also offers the flexibility of incorporating tests based on other attributes such as classification label. This algorithm was that selected for implementation as part of the
IPAP system (see Sect. 12.3.3).
erally distinguishable but independent rotation of floes is also likely. Algorithms that
perform well with one type of ice are generally not best suited for determining motion
in other ice conditions. Following the example of the Alaskan SAR Facility (Kwok et al.
1990; Rothrock et al. 1992), it was therefore decided to adopt a hybrid solution for the
IPAP system: i.e., to develop one algorithm of each type to be selected as best appropriate to the prevailing ice conditions.
The main methods that have been developed which are appropriate for pack ice are
area correlation (Fily and Rothrock 1986, 1987), matched filtering (Ninnis et al. 1986;
Collins and Emery 1988), and optical flow methods (Yan 1992). Of these algorithms,
the area correlation method has been the most widely used, and as one of the algorithms implemented under IPAP is fully described in Sect. 12.3.2 below. Since the main
drawback of the method is in terms of computational efficiency, the matched filtering
technique (in which the correlation is performed in the Fourier domain) would seem
to hold out significant potential. In particular, the cross correlation function of the
power spectra might be used as a first step to rotationally correct the images prior to
applying a conventional area correlation, the power spectra being related by a rotation which is the same as the relative rotation between the two images in the spatial
domain [Fily and Rothrock (1987) estimate that rotations beyond 7" will start to affect
the correlation values obtained]. For the optical flow method, in which the components of ice velocity are calculated at each image pixel in terms of partial derivatives
(temporal and spatial) of the image intensity, the assumptions made regarding stationary image intensity and gradient are likely to be valid only for small velocities of
the order of a few pixels, and so this method may not be appropriate for large-scale
ice displacement.
The main methods that have been developed which are appropriate for marginal
ice are tie-point extrapolation (Zhang et al. 1989), edge-feature matching (Vesecky et
al. 1988; McConnell et al. 1991), postsegmentation region matching (see Sect. 12. 3.3),
invariant moment matching (Zhang 1992), and stochastic approaches (Banfield
1991). For the tie-point extrapolation method, the degree of manual intervention
makes it too time-consuming for routine use within an operational environment. The
other approaches can all be considered as variations on a common theme, i.e. matching common features (generally floes) between images based on some measure of
shape. For each of these techniques, the primary requirement is the input of a good,
repeatable segmentation. Assuming that there is no noise in the boundary determination, then the best representation of the region shape will be obtained from one of
the edge-feature matching techniques which characterize region boundaries at the
pixel level. At the other extreme, the invariant moment matching method is based on
a parameterization of the region shape in terms of its moments and, whilst being less
sensitive to boundary noise, would not be able to cope with tracking regions that split
or merge. The stochastic approach represents an intermediate solution in that floe
boundaries are approximated by principal curves from which edge-pixel probability distributions are derived. The postsegmentation region matching scheme also
offers an intermediate solution in that segmented regions are approximated to by
polygons (of user-defined maximum deviation from the true boundary position), but
also offers the flexibility of incorporating tests based on other attributes such as classification label. This algorithm was that selected for implementation as part of the
IPAP system (see Sect. 12.3.3).