Texture correlation length: a deconvolution method to account for speckle in SAR images

Land-cover classification techniques have long used the image mean and some measure(s) of the spatial variation of that mean in order to better classify and/or quantify classes of interest. The current research is directed at using the textural information in a single-look SAR image to better classify an area. In particular the authors are interested in using the texture from ERS and JERS data from the Amazon to better distinguish different classes with the rain forest. Unfortunately, the speckle noise inherent in SAR imaging causes any estimate of the texture to be corrupted. This research details a technique borrowed from image blurring in order to remove the effects of speckle from their texture estimate. In their case, they are interested in estimating the correlation length, a parameter borrowed form rough surface modeling. They show accurate retrievals of correlation length using modeled SAR images