Unsupervised Multiresolution Segmentation of SAR Imagery Based on Region-Based Hierarchical Model

This paper presents a novel method of unsupervised segmentation for synthetic aperture radar (SAR) images. Firstly, we define a generalized multiresolution likelihood ratio (GMLR), which classifies different kinds of signals more accurately than classical likelihood ratio by fusing more and different signal features. For our SAR image segmentation application, multiresolution stochastic structure inherent in SAR imagery is well captured by a set of multiscale autoregressive (MAR) models. Secondly, good parameter estimates of GMLR can be obtained by estimating several MMARP models using EM algorithm. Thirdly, considering the independence assumption of maximum likelihood estimation of parameter by EM algorithm and reduction of the segmentation time, we present the bootstrap sampling techniques applied above algorithm. Experimental results demonstrate that our algorithm performs fairly well.