6.4 Automatic Mapping of Sea Ice in SAR Imagery
Sea ice backscatter in SAR imagery depends primarily on the surface roughness and
the dielectric constant of sea ice or open seawater (Carsey 2013). Generally
speaking, rougher surface enables more radiation to be backscattered to the sensor,
causing brighter appearance in SAR imagery. The dielectric constant of sea ice
decreases as the degree of salinity decreases. Lower dielectric constant generally
causes stronger backscatter. This implies that thicker sea ice tends to assume
brighter color due to its lower salinity, and new or fresh sea ice appears darker in
the image due to the relatively higher dielectric constant. Therefore, different
sea-ice types are generally distinguishable on SAR images because they admit
different physical characteristics, i.e. salinity and surface roughness, resulting in
the varying magnitudes of SAR backscattering.
Unfortunately, the complex sea ice physics and ever-changing ocean environment, as well as the numerous sensor parameters cause large inner-class gray tone
variation in the observed SAR sea-ice images. First of all, SAR sensor inherently
produces speckle noise, which renders the observed pixel values either brighter or
darker than the true pixels values, leading to great variability in areas that should be
homogeneous. This variation would reduce the separability of different sea-ice
types. Moreover, the existence of ridges, rubble, rims and deformations produced
by compression forces can also produce inhomogeneity in gray tone values (Shokr
1991). In addition, the same sea-ice type could appear different tone values as the
changing of SAR incidence angle.
The significant inner-class variation imposes a fundamental challenge on the
automatic techniques for sea ice image segmentation. Current algorithms for SAR
sea ice image segmentation can be categorized as pixel-based and texture-based.
The former clusters pixels based on gray tone values, e.g., local thresholding
(Havercamp et al. 1993), Gamma (Samadani 1995) and Gaussian (Karvonen
2004) mixture models, and K-Means clustering (Redmund et al. 1998). Due to
the sensitivity of single pixels to speckle noise, these methods always produce
many artifacts. The suppression of speckle noise by some denoising methods (Lee
1980; Frost et al. 1982; Kuan et al. 1985) however will introduce new problems,
such as the blur of the ice boundaries which serve as important information to
delineate sea ice.
As opposed to using single pixels, the texture-based approaches use for segmentation the texture features, which are linear or nonlinear functions of neighboring
pixels, e.g., variation (Burns et al. 1982; Heolbaek-Hansen et al. 1989), gray-level
co-occurrence matrix (GLCM) (Haralick et al. 1973; Shuchman et al. 1989; Baraldi
and Parmiggiani 1995; Barber and LeDrew 1991; Soh and Tsatsoulis 1999; Clausi
2002), Markov random fields (MRF) (Deng and Clausi 2005) and Gabor filter
(Clausi 2001). Comparing with pixel-based approach, these methods are more
robust to inner-class variation. However, due to the vast amount of texture features
and their varying specializations, it’s difficult to select the best group of features for
the current task at hand. Moreover, most texture-based approaches are
118
L. Xu and J. Li
Sea ice backscatter in SAR imagery depends primarily on the surface roughness and
the dielectric constant of sea ice or open seawater (Carsey 2013). Generally
speaking, rougher surface enables more radiation to be backscattered to the sensor,
causing brighter appearance in SAR imagery. The dielectric constant of sea ice
decreases as the degree of salinity decreases. Lower dielectric constant generally
causes stronger backscatter. This implies that thicker sea ice tends to assume
brighter color due to its lower salinity, and new or fresh sea ice appears darker in
the image due to the relatively higher dielectric constant. Therefore, different
sea-ice types are generally distinguishable on SAR images because they admit
different physical characteristics, i.e. salinity and surface roughness, resulting in
the varying magnitudes of SAR backscattering.
Unfortunately, the complex sea ice physics and ever-changing ocean environment, as well as the numerous sensor parameters cause large inner-class gray tone
variation in the observed SAR sea-ice images. First of all, SAR sensor inherently
produces speckle noise, which renders the observed pixel values either brighter or
darker than the true pixels values, leading to great variability in areas that should be
homogeneous. This variation would reduce the separability of different sea-ice
types. Moreover, the existence of ridges, rubble, rims and deformations produced
by compression forces can also produce inhomogeneity in gray tone values (Shokr
1991). In addition, the same sea-ice type could appear different tone values as the
changing of SAR incidence angle.
The significant inner-class variation imposes a fundamental challenge on the
automatic techniques for sea ice image segmentation. Current algorithms for SAR
sea ice image segmentation can be categorized as pixel-based and texture-based.
The former clusters pixels based on gray tone values, e.g., local thresholding
(Havercamp et al. 1993), Gamma (Samadani 1995) and Gaussian (Karvonen
2004) mixture models, and K-Means clustering (Redmund et al. 1998). Due to
the sensitivity of single pixels to speckle noise, these methods always produce
many artifacts. The suppression of speckle noise by some denoising methods (Lee
1980; Frost et al. 1982; Kuan et al. 1985) however will introduce new problems,
such as the blur of the ice boundaries which serve as important information to
delineate sea ice.
As opposed to using single pixels, the texture-based approaches use for segmentation the texture features, which are linear or nonlinear functions of neighboring
pixels, e.g., variation (Burns et al. 1982; Heolbaek-Hansen et al. 1989), gray-level
co-occurrence matrix (GLCM) (Haralick et al. 1973; Shuchman et al. 1989; Baraldi
and Parmiggiani 1995; Barber and LeDrew 1991; Soh and Tsatsoulis 1999; Clausi
2002), Markov random fields (MRF) (Deng and Clausi 2005) and Gabor filter
(Clausi 2001). Comparing with pixel-based approach, these methods are more
robust to inner-class variation. However, due to the vast amount of texture features
and their varying specializations, it’s difficult to select the best group of features for
the current task at hand. Moreover, most texture-based approaches are
118
L. Xu and J. Li