10
L.-K. SOH, C. TSATSOULlS, AND B. HOLT
1993), and lead analysis (Fily and Rothrock 1990). For these floe-dependent approaches, geometric descriptors of an ice floe such as size, perimeter, orientation, and circularity are used as inputs to label, monitor, register, or classify the ice floe. Thus, such
issues necessitate accurate identification and separation of ice floes.
However, adjacent ice floes touch each other forming a network of connected ice floes
which conceals separation among ice floes and hinders ice floe-dependent analysis. In
the following, we present an innovative concept called the restricted growing concept.
The general model of this concept is aimed at achieving object separation and is applicable to many kinds of imagery. In our particular application to SAR sea ice imagery, we
combined probabilistic labeling and morphological operations as means of implementing the concept. In this paper, we emphasize the application of the restricted growing concept to floe size distribution. We also present the ice floe separation module within an integrated system designed to identify and measure floe size, distinguish open
water areas, and separate out remaining forms of ice not identifiable as floes. Preliminary results using these procedures are provided for a time series analysis in the Arctic.
2.2
Background
In image processing, objects are separated from the background using image segmentation. Conventional image segmentation mechanisms include thresholding (Gonzalez
and Woods 1992), region splitting and merging (Horowitz and Pavlidis 1976), and relaxation (Hummel and Zucker 1983; Rosenfeld and Kak 1982). These techniques have been
able to successfully extract objects from the background, and where segmentation was
the primary objective, these mechanisms have been sufficient. However, in cases where
(1) separation among objects is desirable and (2) objects are known to be contiguous
- for example, detection of individual objects in radiographic imagery (Hall et al. 1971),
recognition of defects in electronic patterns (Ejiri et al. 1973), and identification of ice
floes in SAR imagery (Hall and Rothrock 1981, Daida and Vesecky 1989) - we need an
object separation module designed to accomplish the task.
For ice floe identification in satellite images, Banfield and Raftery (1992) devised
a technique that integrated mathematical morphology and principal curve clustering.An erosion-propagation algorithm (EP) was used to select the potential edge pixels and group them into floe outlines. First, the image was thresholded to obtain its
object-background segmentation. Second, the EP algorithm was performed iteratively. At the first iteration, if a pixel was ice and any of its neighbors was water, the
pixel was eroded and became water. At the second iteration, the same operation was
performed on the image resulting from the first iteration, and so on. After certain
iterations of the eroding process, separation among objects would be observed. To
prevent subdivision within single floes, a method was developed to determine which
of the floes identified by the EP algorithm should be merged, based on an algorithm
for clustering about closed principal curves (Hastie and Stuetzle 1989). The authors
showed that the approach produced reasonably good results - floe identification and
separation. However, the EP algorithm suffers from three disadvantages. First, preservation of floe size and shape is not satisfactory. The morphological effects of the EP
algorithm allow a piece of ice floe to shrink with its overall shape roughly intact while
L.-K. SOH, C. TSATSOULlS, AND B. HOLT
1993), and lead analysis (Fily and Rothrock 1990). For these floe-dependent approaches, geometric descriptors of an ice floe such as size, perimeter, orientation, and circularity are used as inputs to label, monitor, register, or classify the ice floe. Thus, such
issues necessitate accurate identification and separation of ice floes.
However, adjacent ice floes touch each other forming a network of connected ice floes
which conceals separation among ice floes and hinders ice floe-dependent analysis. In
the following, we present an innovative concept called the restricted growing concept.
The general model of this concept is aimed at achieving object separation and is applicable to many kinds of imagery. In our particular application to SAR sea ice imagery, we
combined probabilistic labeling and morphological operations as means of implementing the concept. In this paper, we emphasize the application of the restricted growing concept to floe size distribution. We also present the ice floe separation module within an integrated system designed to identify and measure floe size, distinguish open
water areas, and separate out remaining forms of ice not identifiable as floes. Preliminary results using these procedures are provided for a time series analysis in the Arctic.
2.2
Background
In image processing, objects are separated from the background using image segmentation. Conventional image segmentation mechanisms include thresholding (Gonzalez
and Woods 1992), region splitting and merging (Horowitz and Pavlidis 1976), and relaxation (Hummel and Zucker 1983; Rosenfeld and Kak 1982). These techniques have been
able to successfully extract objects from the background, and where segmentation was
the primary objective, these mechanisms have been sufficient. However, in cases where
(1) separation among objects is desirable and (2) objects are known to be contiguous
- for example, detection of individual objects in radiographic imagery (Hall et al. 1971),
recognition of defects in electronic patterns (Ejiri et al. 1973), and identification of ice
floes in SAR imagery (Hall and Rothrock 1981, Daida and Vesecky 1989) - we need an
object separation module designed to accomplish the task.
For ice floe identification in satellite images, Banfield and Raftery (1992) devised
a technique that integrated mathematical morphology and principal curve clustering.An erosion-propagation algorithm (EP) was used to select the potential edge pixels and group them into floe outlines. First, the image was thresholded to obtain its
object-background segmentation. Second, the EP algorithm was performed iteratively. At the first iteration, if a pixel was ice and any of its neighbors was water, the
pixel was eroded and became water. At the second iteration, the same operation was
performed on the image resulting from the first iteration, and so on. After certain
iterations of the eroding process, separation among objects would be observed. To
prevent subdivision within single floes, a method was developed to determine which
of the floes identified by the EP algorithm should be merged, based on an algorithm
for clustering about closed principal curves (Hastie and Stuetzle 1989). The authors
showed that the approach produced reasonably good results - floe identification and
separation. However, the EP algorithm suffers from three disadvantages. First, preservation of floe size and shape is not satisfactory. The morphological effects of the EP
algorithm allow a piece of ice floe to shrink with its overall shape roughly intact while