SAR Image Segmentation Based on Unsupervised Classification of Log-Cumulants Estimates

Abstract

This letter proposes an unsupervised algorithm for synthetic aperture radar (SAR) image segmentation based on the statistical properties of log-cumulants estimates of SAR data modeled by the $G^{0}$ distribution. We also investigate its application in SAR images for remote sensing of environment. The proposed method adjusts a multi-class two-Gaussian mixture to the distributional parameters estimated by the log-cumulants method. Our method generates an initial classification using the Gaussian mixture and the probability density of the $G^{0}$ parameters to provide the training samples for the support vector machine (SVM) algorithm. For comparison purpose, we used three classifiers from literature: a Bayesian approach, an enhanced $K$ -nearest neighbors method, and an iterative SVM. The evaluation of the results was performed in terms of the error of segmentation (EoS) and cross region fitting (CRF) measures. The segmentation results, based on EoS and CRF, showed that our method was more efficient than the three classifiers for synthetic and real SAR images.