A.J. SEPHTON AND K.C. PARTINGTON
12.3.2
Pack Ice Motion
The area correlation techniques which have been used for assessing pack ice motion in
remote sensing data (see Sect. 12.p) are based on the cross-correlation of blocks of pixels in two images, the displacement vectors between matching points being taken to be
between centroid positions of the blocks with the highest correlation. In order to
reduce the required degree of processing, a hierarchical scheme has been adopted in
the implementation at MRC, similar in principle to that developed by Fily and Rothrock
(1986,1987). In this method, the correlation computation is facilitated by applying the
correlation to an image-pyramid data structure (one pyramid for each image), where
cross-correlation peaks corresponding to motion vectors are first found in highly averaged images, and repeated stepwise (up to three iterations in the algorithm implemented at MRC) in images of finer and finer resolution. Providing that the measured
correlation peaks are statistically significant (at the 1% level for a correlation window
size of 16 pixels), the "best match" positions derived from the coarser resolution images
are taken as the starting point for the correlation at the next level in the pyramid.
At the end of each smoothing iteration, a filter may optionally be applied (in the MRC
implementation) in a user-defined number of iterations in order to correct "spurious"
displacement vectors at odds with their neighbors. This helps to prevent errors in the
top levels of the pyramid propagating down to the lower levels. In the first filter iteration, if the displacements in both the x and y directions for a particular pixel corresponding to the second best correlation differ less than the x and y displacements corresponding to the best correlation from the mean x and y displacements of the pixel's
neighbors, then the displacements for that pixel are updated by the displacements associated with the second best correlation. In subsequent filter iterations, if the displacements in either of the x and y directions for a particular pixel differ from the mean x
and y displacements of the pixel's neighbors by more than a user-defined threshold (in
standard deviations), then the displacements for that pixel are updated by the mean x
and y displacements.
The ice motion displacements are displayed corrected for any geographic offset
between the images, either in the form of a grid representation for all points in the data
or as arrows showing those displacements with the highest correlation. Up to three classes of vector are displayed dependent on the strength of the associated correlation. The
mean vector displacements of ice between the two images can also be generated, averaged over a grid of user-specified dimensions.
Evaluation of the area correlation algorithm based on three pairs of ERS-1 images
each recorded 3 days apart off the east Greenland coast in March 1992 shows that for
pack-ice there is a high degree of consistency in the measured vector displacements
both within and between images, as well as with manual interpretation and with the
known patterns of ice motion in this area. Figure 5 shows by way of example the results
of the correlation obtained using a pair of images recorded on March 24 and 27, 1992.
The image of March 24 is shown overlaid by the grid representation of the total correlation space in red and by arrows showing the vector displacements of ice in the later
image with respect to the earlier image, yellow arrows corresponding to displacement
vectors which have an associated correlation coefficient between 0 and 1 standard deviations above the user-defined threshold of 2 standard deviations above the mean, and
12.3.2
Pack Ice Motion
The area correlation techniques which have been used for assessing pack ice motion in
remote sensing data (see Sect. 12.p) are based on the cross-correlation of blocks of pixels in two images, the displacement vectors between matching points being taken to be
between centroid positions of the blocks with the highest correlation. In order to
reduce the required degree of processing, a hierarchical scheme has been adopted in
the implementation at MRC, similar in principle to that developed by Fily and Rothrock
(1986,1987). In this method, the correlation computation is facilitated by applying the
correlation to an image-pyramid data structure (one pyramid for each image), where
cross-correlation peaks corresponding to motion vectors are first found in highly averaged images, and repeated stepwise (up to three iterations in the algorithm implemented at MRC) in images of finer and finer resolution. Providing that the measured
correlation peaks are statistically significant (at the 1% level for a correlation window
size of 16 pixels), the "best match" positions derived from the coarser resolution images
are taken as the starting point for the correlation at the next level in the pyramid.
At the end of each smoothing iteration, a filter may optionally be applied (in the MRC
implementation) in a user-defined number of iterations in order to correct "spurious"
displacement vectors at odds with their neighbors. This helps to prevent errors in the
top levels of the pyramid propagating down to the lower levels. In the first filter iteration, if the displacements in both the x and y directions for a particular pixel corresponding to the second best correlation differ less than the x and y displacements corresponding to the best correlation from the mean x and y displacements of the pixel's
neighbors, then the displacements for that pixel are updated by the displacements associated with the second best correlation. In subsequent filter iterations, if the displacements in either of the x and y directions for a particular pixel differ from the mean x
and y displacements of the pixel's neighbors by more than a user-defined threshold (in
standard deviations), then the displacements for that pixel are updated by the mean x
and y displacements.
The ice motion displacements are displayed corrected for any geographic offset
between the images, either in the form of a grid representation for all points in the data
or as arrows showing those displacements with the highest correlation. Up to three classes of vector are displayed dependent on the strength of the associated correlation. The
mean vector displacements of ice between the two images can also be generated, averaged over a grid of user-specified dimensions.
Evaluation of the area correlation algorithm based on three pairs of ERS-1 images
each recorded 3 days apart off the east Greenland coast in March 1992 shows that for
pack-ice there is a high degree of consistency in the measured vector displacements
both within and between images, as well as with manual interpretation and with the
known patterns of ice motion in this area. Figure 5 shows by way of example the results
of the correlation obtained using a pair of images recorded on March 24 and 27, 1992.
The image of March 24 is shown overlaid by the grid representation of the total correlation space in red and by arrows showing the vector displacements of ice in the later
image with respect to the earlier image, yellow arrows corresponding to displacement
vectors which have an associated correlation coefficient between 0 and 1 standard deviations above the user-defined threshold of 2 standard deviations above the mean, and