Skip to main content

Ice Motion Vectors: Data Processing Description

First the ice tracker algorithm must select two input images:

  • Reference Image - the earlier image of a particular region (usually three days before the target image, as the ERS-1 satellite retraces its path in that time)

  • Target Image - the later image of the specified region

These inputs are ERS-1 SAR low-resolution geocoded images of the same region, taken at different times. The input images represent a 100 km by 100 km region with a pixel spacing of 100 m and a resolution of 240 m. "Geolocated" means that each image is mapped onto the SSM/I latitude/longitude reference grid so that it may be more easily and accurately compared. The images are also radiometrically corrected to ensure better ice feature identification. More work is done before the two images are approved as inputs. The tracker algorithm takes the reference estimate of the region's ice motion and, using geostrophic wind data which accounts for 70 percent of ice motion variance (on the days timescale), estimates how the prominent ice features probably moved. The program tests its estimate against selected target images, choosing as an input the image with the best correspondence of expected ice features. For images taken of the marginal ice zone or affected by costal influences, even more processing is done before a target image is selected. For this further processing, the reference and target images undergo low-resolution feature-tracking and area-tracking schemes to determine how well the ice features match. A target image is only selected when at least 50 percent of the reference image's ice features are found in it.

Once the two input images have been selected, the algorithm branches depending upon the region involved. Features in the central ice pack do not deform or rotate as rapidly as those in more dynamic regions, so the less intensive area-matching techniques can be successfully utilized for them. If the region covers the ice margin, however, feature-matching techniques must be used as well. For this more complicated case, the reference image's distinct features (region boundaries) must first be determined. Regions are defined as image areas that exhibit similar attributes in terms of texture (how the pixel intensity varies across a feature - like describing in mathematical terms the pattern on an object's surface) and average pixel intensity. The image is segmented into statistically similar regions, and cluster boundaries (regions of similar attributes) are identified. The boundaries are converted from an array of (x, y) coordinate pairs into (psi, s) curve points so that one-dimensional computation, rather than two, may be done. The output of the features' (psi, s) comparisons/matching attempts is the rotational and translational motion of the region boundaries.

The results of the feature-matching processes are non-uniform in the sizes represented, so an area-based matching technique is next used in order to fine tune the feature-matching's motion estimates and present the data in a more evenly spaced, understandable manner. This refining process is also performed on the images covering the more stable ice pack region; the ice pack images thus go through two area-based matching algorithms, the second one at a finer level. The translational and rotational estimates from the previous processor are used to estimate the ice motion at each grid point. The algorithm then determines, using the motion estimate, which grid points in the next image would be the most likely to correspond to a selected grid point in the reference image. Using Fast Fourier Transforms to rotate and translate the reference image area into possible future representations, the "moved to" grid points are tested to determine which one best fits. When the best match is found, the original grid location is then assigned its motion vector. Note that regions of no visible motion, such as smooth water, will not register results. A number of filtering processes and verification checks are performed throughout the ice motion vector program to ensure quality results. The output product has a sample spacing of 5km on the SSM/I grid, an accuracy to 300 meters, and a reliability of 90-95%.

Due to previous users' requirements, most of the input images were of the Beaufort Sea and the high Arctic region up to 85N latitude. Some came from the Chikchi and East Siberian seas, with only a few images south of 65N latitude utilized. The ice motion products were processed at a rate of about 60 per week from February 1992 until December 31, 1994, during the fall and winter months. A new GPS, the RADARSAT Geophysical Processor System (RGPS), will build upon the experiences of the previous GPS group in providing high quality SAR-derived geophysical products. RGPS will begin production in early 1997.

(Please see the article, "An Ice-Motion Tracking System at the Alaska SAR Facility" starting on page 44 of the January 1990 edition of the IEEE Journal of Oceanic Engineering, upon which this page was based, for more details.)