Identifying classes in SAR sea ice imagery using correlated texture

This paper presents a new technique in identifying classes in Synthetic Aperture Radar (SAR) sea ice imagery using correlated texture. First, we employ dynamic local thresholding to generate a histogram of thresholds. Then, we use a multi-resolution peak-detection method, a strategy used in digital image quantization field, to extract significant intensity thresholds from the histogram and provide an initial segmentation. Next, we compute correlated texture of the result and create a matrix of spatial, probabilistic relationships among the classes. Given the texture, we cluster the classes into different groups. The clustering concept is based on an innovative 'solidification' model that strives to obtain similar auto-correlated textural values for all groups. This process produces a second segmentation with the correct number of classes. We have tested out technique in more than 200 SAR sea ice imagery successfully. The entire process is fully automated and fast.