SAR Sea Ice Image Segmentation And Evaluation Using EPR Based Markov Random Fields

In this paper, we propose a new approach to sea ice segmentation in SAR intensity imagery by combining an edge-preserving region (EPR)-based representation with region-level MRF models. The EPR-based representation gives an initial segmentation of SAR images in the form of primitive regions, with the goal of efficiently suppressing oversegmentation within objects while accurately locating region boundaries at object edges. To achieve this goal, object edges and within-object homogeneity in SAR images are first characterized by instantaneous coefficient of variation (ICOV) derived using the speckle reduction anisotropic diffusion (SRAD) algorithm , which has the distinctive advantages of detecting edges in the presence of speckle noise. The watershed algorithm is further applied to partition the image into disjoint regions. In contrast to the previous ICOV-based segmentation algorithm , our scheme aims at a segmentation result more appropriate for a combination with region-level MRF models towards segmentation purpose. Specifically, two new metrics for quantitative assessment of region characteristics (region accuracy and region redundancy) are defined and used for parameter estimation in the ICOV extraction process. Combined with the EPR-based representation, the regionlevel MRF approach to SAR image segmentation is facilitated in the following two aspects: First, the segmentation process is largely accelerated. In contrast to existing regionbased representations, the EPR-based representation has much fewer regions which correspondingly narrows down the search space associated with the optimization process Second, the segmentation accuracy is improved over existing region-based representations. By alleviating speckle impacts on the segmentation, the EPR-based representation has much larger regions with boundaries positioning at object edges. In this manner, feature statistics of regions are less sensitive to speckle noise which leads to more accurate estimate of feature model parameters, and consequently reduces the probability of false segmentation.