76
S. LI, Z. CHENG, AND W.E WEEKS
the ASF ice motion products, .1X and L1 Yare 5 km. Once the indices for the row and column positions of a tie point are found, the location of the tie point within our problem-specific 2-D matrix is set. All the other ice motion information, including the positions of both the starting and end points, the displacements in the X and Y directions,
and the rotation angle of the neighborhood surrounding the tie point, are copied onto
the relevant 2-D matrices using i, j as position indexes (Fig. 3).
4.3.2
The Edge Point Trimmer
Defective points (or flyers) in the ice motion products can be a major source of errors
in our procedure. Through inspection, we have found that the majority of points that
exhibit large ice motion errors occur near the edges of the SAR images. This error is
caused by the sharp boundary at the image edge. Before the ice motion product is generated, the images in the matching pair are geocoded and mapped onto a larger frame
based on a standard polar stereographic (SSM/I) map projection. As a result, the area
outside the original image is filled with zeroes as a background, thereby forming a sharp,
but unreal, boundary. When such sharp boundaries exist in the matching cells on both
images, the matching algorithm produces an erroneous best correlation based on the
relative position of the sharp, imaginary boundaries, as the algorithm is unable to distinguish true image data from background. Thus, an erroneous ice motion vector is
formed, resulting in a belt of unreasonably large deformation values near the edge of
the matched area. Obviously, this type of error should be avoided.
Therefore, a special "edge point trimming" algorithm has been developed to eliminate these defective points. The procedure is conservative in that all of the edge points
are eliminated to insure the reliable extraction of the ice deformation parameters. This
procedure is performed after the row and column positions of all the gridded ice
motion vectors are determined. Even the simplest approach, that of eliminating the end
data points in each row and column, has been found to significantly improve the results
by removing the majority of the bad vectors. However, this method commonly eliminates too many edge points, in that the end points for each row or each column are not
necessarily the edge points of the matched area. Quite often, tie points occurring outside the end points were eliminated during the process of ice motion generation
because of poor matching (Kwok and Cunningham 1993).
To eliminate all of the edge points while keeping all the data points inside the overlapped
area, a more sophisticated, several-step procedure has been devised. It first searches
through the first and the last rows and columns of the grid to find the vertices of a polygon that defines the overlapped area. At maxinmm, there can be a total of eight vertices,
with two for each side of the rectangular grid. Then, a side is created by interpolating
between the adjacent vertices of the resulting polygon. Next, the points on that particular
side are marked by "2'S" with all the other points initialized as "o's:' Once all the points on
the sides of the polygon are marked down to the grid by assigning them to nearest cells, a
filling process is used to change the values at all the interior data cells into "1'S:' The resulting pattern forms a mask which is further trimmed removing all the edge points marked
by"2's;' and then complemented with the results of the previously mentioned simple elimination procedure to create a compound mask through a logic "OR" operation. Finally the
compound mask is applied to the data panel to filter out all the true edge points (Fig. 4).
S. LI, Z. CHENG, AND W.E WEEKS
the ASF ice motion products, .1X and L1 Yare 5 km. Once the indices for the row and column positions of a tie point are found, the location of the tie point within our problem-specific 2-D matrix is set. All the other ice motion information, including the positions of both the starting and end points, the displacements in the X and Y directions,
and the rotation angle of the neighborhood surrounding the tie point, are copied onto
the relevant 2-D matrices using i, j as position indexes (Fig. 3).
4.3.2
The Edge Point Trimmer
Defective points (or flyers) in the ice motion products can be a major source of errors
in our procedure. Through inspection, we have found that the majority of points that
exhibit large ice motion errors occur near the edges of the SAR images. This error is
caused by the sharp boundary at the image edge. Before the ice motion product is generated, the images in the matching pair are geocoded and mapped onto a larger frame
based on a standard polar stereographic (SSM/I) map projection. As a result, the area
outside the original image is filled with zeroes as a background, thereby forming a sharp,
but unreal, boundary. When such sharp boundaries exist in the matching cells on both
images, the matching algorithm produces an erroneous best correlation based on the
relative position of the sharp, imaginary boundaries, as the algorithm is unable to distinguish true image data from background. Thus, an erroneous ice motion vector is
formed, resulting in a belt of unreasonably large deformation values near the edge of
the matched area. Obviously, this type of error should be avoided.
Therefore, a special "edge point trimming" algorithm has been developed to eliminate these defective points. The procedure is conservative in that all of the edge points
are eliminated to insure the reliable extraction of the ice deformation parameters. This
procedure is performed after the row and column positions of all the gridded ice
motion vectors are determined. Even the simplest approach, that of eliminating the end
data points in each row and column, has been found to significantly improve the results
by removing the majority of the bad vectors. However, this method commonly eliminates too many edge points, in that the end points for each row or each column are not
necessarily the edge points of the matched area. Quite often, tie points occurring outside the end points were eliminated during the process of ice motion generation
because of poor matching (Kwok and Cunningham 1993).
To eliminate all of the edge points while keeping all the data points inside the overlapped
area, a more sophisticated, several-step procedure has been devised. It first searches
through the first and the last rows and columns of the grid to find the vertices of a polygon that defines the overlapped area. At maxinmm, there can be a total of eight vertices,
with two for each side of the rectangular grid. Then, a side is created by interpolating
between the adjacent vertices of the resulting polygon. Next, the points on that particular
side are marked by "2'S" with all the other points initialized as "o's:' Once all the points on
the sides of the polygon are marked down to the grid by assigning them to nearest cells, a
filling process is used to change the values at all the interior data cells into "1'S:' The resulting pattern forms a mask which is further trimmed removing all the edge points marked
by"2's;' and then complemented with the results of the previously mentioned simple elimination procedure to create a compound mask through a logic "OR" operation. Finally the
compound mask is applied to the data panel to filter out all the true edge points (Fig. 4).