Fowler interpolated data from AVHRR channels 2 and 4 to a 5-km resolution EASE-Grid. Channel 4 was used throughout the year, but it is not as useful during melt periods when the ice surface and water are nearly the same temperature. Channel 2 is only useful during sunlit periods in the late spring, summer, and early fall.
The following shows a single orbital swath from AVHRR.
One AVHRR orbital swath
The visible channel is useful during summer melt conditions when microwave and infrared data are unreliable. Thermal imagery is used throughout the year except during the summer melt season when surface ice and water have nearly the same temperature. Ice displacement was derived from four daily satellite passes (0000, 0600, 1200, and 1800 GMT) and matched with the corresponding passes of the next day; therefore, ice velocities from AVHRR data represent average velocities over 24 hours. These four orbits provide full coverage of the ice-covered areas at least once per day, and up to four times near the North Pole, as the following figure shows.
Four daily AVHRR orbital swaths
Calibration is not an issue with ice motion calculations; however, image registration must be optimal, otherwise errors will show a bias in the ice motion. Fowler applied roll, pitch, and yaw corrections to the navigation stage to minimize error.
Detection of ice displacement in SMMR, SSM/I, and AVHRR imagery was achieved with maximum cross correlation (MCC) techniques described in Emery et al. 1995. Fowler compared 10 x 10 pixel rectangular subsets of the same spatial locations between two consecutive days, and chose the location with the best correlation coefficient. The change in location is considered the ice displacement, which allows ice motion to be calculated. This method applies to each of the channels for each of the four passes.
While AVHRR provides complete satellite coverage of both polar regions, the presence of clouds obscures the surface in both the the near-infrared and thermal channels. This is the main limitation with using AVHRR data for ice motion.
Fowler used an AVHRR Polar Pathfinder ice extent mask with the MCC algorithms over ice-covered areas. To detect registration problems, the MCC method was also applied over land. If any displacements between image pairs occured over land, Fowler removed this bias from the calculated ice motions. He applied a filter to the calculated vectors to remove "bad" vectors. Assuming the ice moves locally and uniformly, a spatial coherence filter is used to compare each vector with neighboring vectors. If a vector had at least three neighboring vectors that moved in a similar direction to within two pixels, that vector was considered "good."
Cloud contamination is the main reason for eliminating vectors. The MCC method assumes that ice motion of the chosen rectangular area is linear with no distortion or rotation. For most small areas, the non-linearity is small enough to not violate this assumption; however, in areas where ice motion is non-linear, the MCC method cannot effectively track the ice. This is not a significant problem for the 5-km resolution AVHRR data.
All of the vector data for four passes and two channels are averaged together to create a set of daily vectors. Following is an example of daily-averaged vectors from 20 April 2000.
Daily-averaged ice motion vectors from AVHRR
After filtering and merging different passes and channels, the percentage of AVHRR temporal coverage ranges from 0% to about 40%. The following image shows that coverage in the Canada Basin is nearly 40% in the winter, but much lower in other areas because of persistent cloud cover.
Winter coverage from AVHRR
Cloud cover is more prevalent during the summer, resulting in fewer vectors from AVHRR.
Summer coverage from AVHRR
The following figure shows that the temporal coverage of vectors derived from AVHRR in the Antarctic region is less than that in the Arctic. Cloud cover is much more persistent and frequent.
Summer coverage from AVHRR
The MCC method can often detect sub-pixel resolution displacement of ice motion vectors. With satellite imagery, accuracy is between 1/3 and 1/2 pixel displacements. With AVHRR 5-km imagery, ice motion accuracy is about 2 cm/sec. Fowler compared vectors derived from AVHRR imagery with those from buoy data. He found 26,820 pairs of AVHRR and buoy vectors that were less than 50 km apart. The mean difference in the u component was -0.12 cm/sec with a Root Mean Square (RMS) error of 3.31 cm/sec. The mean difference in the v component was 0.07 cm/sec with an RMS error of 3.29 cm/sec.