2 Identifying Ice Floes and Computing Ice Floe Distributions in SAR Images
19
an intensity histogram of an ERS-1 SAR sea ice image taken on August 19, 1991, at
74.43°N, 144.29°W. The three smaller modes of histograms are generated by collecting
pixels of floes, water/ice mixture, and water regions after the image segmentation and
water/ice mixture-water distinction processes. It can be seen that the intensities of all
three classes overlap.
The local dynamic thresholding technique in (Haverkamp et al. 1995) has been modified to accommodate SAR sea ice imagery composed of floes, water/ice mixture, and
water. An image is first divided into smaller, overlapping regions such that each region
can be analyzed locally. A histogram is computed for each region, and only regions with
high variance histograms will be further processed. The basic concept is to obtain the
bisector of each region by examining its histogram. We accomplish this by approximating the histogram with a binormal Gaussian distribution. Given the parameters of
the fitting curve, an optimal threshold that dissects the curve into two Gaussian distributions is computed using maximum likelihood, and verified if its histogram passes a
bimodality test. Then, a standard deviation minimization technique is used to divide
all computed thresholds into two groups. Finally, region and point interpolations are
executed to ensure all regions and pixels each have two thresholds. Of the two sets of
thresholds, we choose one to segment the image into floes and nonfloes. This decision
requires a user input to specify which set of thresholds to be used. If floes are brighter
in the image, the set of higher thresholds is used; if floes are darker in the image, the
other is used. Figure 7 shows the result of our local dynamic thresholding technique.
3. Floe Extraction. For floe extraction, we use our implementation of the restricted
growing concept (Sect. 2.3 above). Figure 7C shows the results of our floe extraction phase.
Fig. 7. Part of an ERS-1 SAR sea ice image in process for ice coverage. From left to right: a The original image (copyright ESA), b the segmented image, c the floe-extracted image, d the resultant image
from the water/ice mixture extraction process. Bright regions denote water; gray regions denote
water/ice mixture; dark regions denote floes
19
an intensity histogram of an ERS-1 SAR sea ice image taken on August 19, 1991, at
74.43°N, 144.29°W. The three smaller modes of histograms are generated by collecting
pixels of floes, water/ice mixture, and water regions after the image segmentation and
water/ice mixture-water distinction processes. It can be seen that the intensities of all
three classes overlap.
The local dynamic thresholding technique in (Haverkamp et al. 1995) has been modified to accommodate SAR sea ice imagery composed of floes, water/ice mixture, and
water. An image is first divided into smaller, overlapping regions such that each region
can be analyzed locally. A histogram is computed for each region, and only regions with
high variance histograms will be further processed. The basic concept is to obtain the
bisector of each region by examining its histogram. We accomplish this by approximating the histogram with a binormal Gaussian distribution. Given the parameters of
the fitting curve, an optimal threshold that dissects the curve into two Gaussian distributions is computed using maximum likelihood, and verified if its histogram passes a
bimodality test. Then, a standard deviation minimization technique is used to divide
all computed thresholds into two groups. Finally, region and point interpolations are
executed to ensure all regions and pixels each have two thresholds. Of the two sets of
thresholds, we choose one to segment the image into floes and nonfloes. This decision
requires a user input to specify which set of thresholds to be used. If floes are brighter
in the image, the set of higher thresholds is used; if floes are darker in the image, the
other is used. Figure 7 shows the result of our local dynamic thresholding technique.
3. Floe Extraction. For floe extraction, we use our implementation of the restricted
growing concept (Sect. 2.3 above). Figure 7C shows the results of our floe extraction phase.
Fig. 7. Part of an ERS-1 SAR sea ice image in process for ice coverage. From left to right: a The original image (copyright ESA), b the segmented image, c the floe-extracted image, d the resultant image
from the water/ice mixture extraction process. Bright regions denote water; gray regions denote
water/ice mixture; dark regions denote floes