6.6 Results and Discussion
In this section, the results achieved by the proposed method are reported and
discussed, in comparison with several other popular methods. Before reporting
the results on real SAR images, we first examine the results on simulated SAR
sea ice images, where clean images with ice-like gray tones were degraded by
speckle noise. In simulated study, the clean images were used as ground truth to
produce numerical measures for performance evaluation. For real SAR images, the
evaluation of segmentation result was by visual interpretation based on prior
knowledge and ice-chart concerning sea ice types and their spatial distributions.
6.6.1 Results on Simulated Imagery
One clean image (Fig. 6.6a) was degraded by speckle noise that satisfies a squaredroot Gamma distribution (Xie et al. 2002). Simulated images with different noise
levels measured by equivalent number of looks (ENL) were used to feed segmentation methods, in order to examine their robustness to varying noise levels.
On the simulated images, we compared the proposed method with four popular
methods, such as K-Means clustering (Redmund et al. 1998), Gamma mixture
(Samadani 1995), GLCM (Clausi and Yue 2004) and VMLL technique (Deng
and Clausi 2005). Moreover, we also used for comparison the combination of
KPCA features and K-Means clustering model, where the KPCA features are
used to feed K-Means method. For the MRF-based method, we adopted the gray
tone values as features, used 150 EM iterations and 10,000 simulated-annealing
iterations. For GLCM, we used 12 features (i.e. entropy, dissimilarity and correlation in four directions), 64 quantization levels and the patch-size of 15 Â 15.
Figure 6.3 presents the segmentation results obtained by different algorithms
when ENL ¼ 4. The results suggest that the proposed method outperformed the
other three techniques. As we can see, Gamma mixture and K-Means approaches
produced intense artifacts, due to the sensitivity of single pixel to speckle noise. The
GLCM and MRF methods, although were better at resisting the influence of speckle
noise, produced certain misclassifications. For example, GLCM failed to delineate
the boundaries accurately; MRF also failed to provide smooth boundaries. KPCA
also produced artifacts due to the fact that the label correlation effect is not
addressed. The proposed Bayesian method nevertheless produced segmentation
results that are very similar to the true image.
We adopt the overall accuracy for quantitative evaluation. The overall accuracy
is calculated as the ratio between the number of pixels that are correctly classified
and the total number of pixels. Figure 6.4 shows the overall accuracy of different
algorithms as a function of noise level measured by ENL. As we can see, the values
of overall accuracy of the proposed method are above all the other methods. With
the increasing of noise level, Gamma mixture deteriorated sharply, while MRF
6 Mapping Sea Ice from Satellite SAR Imagery
127
In this section, the results achieved by the proposed method are reported and
discussed, in comparison with several other popular methods. Before reporting
the results on real SAR images, we first examine the results on simulated SAR
sea ice images, where clean images with ice-like gray tones were degraded by
speckle noise. In simulated study, the clean images were used as ground truth to
produce numerical measures for performance evaluation. For real SAR images, the
evaluation of segmentation result was by visual interpretation based on prior
knowledge and ice-chart concerning sea ice types and their spatial distributions.
6.6.1 Results on Simulated Imagery
One clean image (Fig. 6.6a) was degraded by speckle noise that satisfies a squaredroot Gamma distribution (Xie et al. 2002). Simulated images with different noise
levels measured by equivalent number of looks (ENL) were used to feed segmentation methods, in order to examine their robustness to varying noise levels.
On the simulated images, we compared the proposed method with four popular
methods, such as K-Means clustering (Redmund et al. 1998), Gamma mixture
(Samadani 1995), GLCM (Clausi and Yue 2004) and VMLL technique (Deng
and Clausi 2005). Moreover, we also used for comparison the combination of
KPCA features and K-Means clustering model, where the KPCA features are
used to feed K-Means method. For the MRF-based method, we adopted the gray
tone values as features, used 150 EM iterations and 10,000 simulated-annealing
iterations. For GLCM, we used 12 features (i.e. entropy, dissimilarity and correlation in four directions), 64 quantization levels and the patch-size of 15 Â 15.
Figure 6.3 presents the segmentation results obtained by different algorithms
when ENL ¼ 4. The results suggest that the proposed method outperformed the
other three techniques. As we can see, Gamma mixture and K-Means approaches
produced intense artifacts, due to the sensitivity of single pixel to speckle noise. The
GLCM and MRF methods, although were better at resisting the influence of speckle
noise, produced certain misclassifications. For example, GLCM failed to delineate
the boundaries accurately; MRF also failed to provide smooth boundaries. KPCA
also produced artifacts due to the fact that the label correlation effect is not
addressed. The proposed Bayesian method nevertheless produced segmentation
results that are very similar to the true image.
We adopt the overall accuracy for quantitative evaluation. The overall accuracy
is calculated as the ratio between the number of pixels that are correctly classified
and the total number of pixels. Figure 6.4 shows the overall accuracy of different
algorithms as a function of noise level measured by ENL. As we can see, the values
of overall accuracy of the proposed method are above all the other methods. With
the increasing of noise level, Gamma mixture deteriorated sharply, while MRF
6 Mapping Sea Ice from Satellite SAR Imagery
127