2 Identifying Ice Floes and Computing Ice Floe Distributions in SAR Images
11
sacrificing its actual size. Second, small ice floes have to be eliminated in order to
achieve necessary separation. For example, for a specified number of iterations, i, ice
floes with size less than (2i +1) x (2i +1) would be discarded. For a low-resolution (100
m per pixel) ERS-l SAR sea ice imagery, if i = 7, then ice floes with size less than 15 x
15 pixels, or 2.25 km>, would be lost. Third, the number of iterations needed for each
image is different, and its determination requires human visual judgment. This
means experiments would have to be performed in order to manually select the optimal number of iterations.
In another study, Korsnes (1993) proposed a mathematical morphology based technique that estimated floe size automatically. A disc-like mask was used to perform morphological operations such as dosing and opening on the floes. The area of the opening operation of the union of disjoint discs of different sizes was used to give the integrated area of the discs with certain radius. The distribution of the integrated areas, by
their different radii, defined size distributions of objects in the image. However, there
are two disadvantages in this scheme. First, the distribution is simply an estimation since
it is based on approximating discs. Second, connected floes are misidentified as a piece
of large floe.
In the next sections, we describe the restricted growing concept that enables extraction of ice floes individually by creating separation among floes while (1) preserving
the original size and shape of the floes, (2) retaining small ice floes as small as a few
pixels, and (3) requiring minimal human intervention.
2.3
The Restricted Growing Concept
The main idea of the restricted growing concept is that growing a shrunken version of
an object within the boundary of its original version facilitates separation and preserves
size and shape. The shrinking process provides separation among neighboring objects.
Once this separation is established, we grow the objects back to their original sizes and
shapes restricted by the boundary of the original objects. Conventional region growing or morphological dilation schemes do not restrict the process of growing objects
such that the acts of achieving both separation and size preservation become contradictory.
According to our definition, an object is a feature in the image and its identification
and localization, as well as its size and shape, are desired. Should only the identification and localization or global interpretation of the image be of interest, a basic image
segmentation technique will suffice to accomplish the task, ignoring the requirements
of accurate size and shape. The object must have definitive shape and size in order to
be useful for further processing. This definition serves as the basic objective of the
restricted growing concept. By this requirement, objects that are fuzzy (composed of
different segment pixels where the object segment dominates), disintegrated (composed of adjacent, disconnected parts due to lost details during the image segmentation or capture process), and connected (linked by pixels misidentified as part of
objects during the preprocessing of the image) are not desirable. Hence, an object identification scheme should be able to produce definitive shape and size of the objects
encountering these three problems. Figure 1 shows examples of fuzzy, disintegrated, and
connected objects.
11
sacrificing its actual size. Second, small ice floes have to be eliminated in order to
achieve necessary separation. For example, for a specified number of iterations, i, ice
floes with size less than (2i +1) x (2i +1) would be discarded. For a low-resolution (100
m per pixel) ERS-l SAR sea ice imagery, if i = 7, then ice floes with size less than 15 x
15 pixels, or 2.25 km>, would be lost. Third, the number of iterations needed for each
image is different, and its determination requires human visual judgment. This
means experiments would have to be performed in order to manually select the optimal number of iterations.
In another study, Korsnes (1993) proposed a mathematical morphology based technique that estimated floe size automatically. A disc-like mask was used to perform morphological operations such as dosing and opening on the floes. The area of the opening operation of the union of disjoint discs of different sizes was used to give the integrated area of the discs with certain radius. The distribution of the integrated areas, by
their different radii, defined size distributions of objects in the image. However, there
are two disadvantages in this scheme. First, the distribution is simply an estimation since
it is based on approximating discs. Second, connected floes are misidentified as a piece
of large floe.
In the next sections, we describe the restricted growing concept that enables extraction of ice floes individually by creating separation among floes while (1) preserving
the original size and shape of the floes, (2) retaining small ice floes as small as a few
pixels, and (3) requiring minimal human intervention.
2.3
The Restricted Growing Concept
The main idea of the restricted growing concept is that growing a shrunken version of
an object within the boundary of its original version facilitates separation and preserves
size and shape. The shrinking process provides separation among neighboring objects.
Once this separation is established, we grow the objects back to their original sizes and
shapes restricted by the boundary of the original objects. Conventional region growing or morphological dilation schemes do not restrict the process of growing objects
such that the acts of achieving both separation and size preservation become contradictory.
According to our definition, an object is a feature in the image and its identification
and localization, as well as its size and shape, are desired. Should only the identification and localization or global interpretation of the image be of interest, a basic image
segmentation technique will suffice to accomplish the task, ignoring the requirements
of accurate size and shape. The object must have definitive shape and size in order to
be useful for further processing. This definition serves as the basic objective of the
restricted growing concept. By this requirement, objects that are fuzzy (composed of
different segment pixels where the object segment dominates), disintegrated (composed of adjacent, disconnected parts due to lost details during the image segmentation or capture process), and connected (linked by pixels misidentified as part of
objects during the preprocessing of the image) are not desirable. Hence, an object identification scheme should be able to produce definitive shape and size of the objects
encountering these three problems. Figure 1 shows examples of fuzzy, disintegrated, and
connected objects.