Adaptive multiresolution quantization for contextual information gain in SAR sea ice images

In this paper we describe an adaptive multiresolution technique that quantizes SAR sea ice images to improve contextual information such as the spatial, relational make up of ice types in a region. First, we use dynamic local thresholding to extract regional intensity threshold values from which a histogram is constructed. Next, we blur the threshold histogram with varying window sizes to build a multiresolution contour map. We identify peaks based on a cumulative distribution function and track each peak on the contour map to assess its significance. Then, for each significant peak identified, we cluster the threshold values extracted during dynamic local thresholding using nearest-neighbors to establish sets of threshold values. Finally, we assign a pixel its quantization value by comparing its original intensity to the set that it belongs to. The technique handles noise and preserves contexts, ensuring a consistent and smooth quantization of the image. We have applied the technique to a large number of ERS-1, ERS-2, and RADARSAT images to obtain quantized representation of the images for contextual information gain. We have also embedded the technique in an unsupervised sea ice segmentation tool that has been installed at the National Ice Center and the Canadian Ice Service.