Comparison of parameter estimators for K-distribution

Abstract

Parameter estimation forms an essential part of many signal- and image-processing tasks. In particular, in the analysis of coherent imagery, such as that provided by synthetic aperture radar (SAR), parameter estimation is required to characterise the statistical properties of homogeneous regions for use in segmentation and target detection algorithms. The statistics of SAR imagery can be modelled by the K-distribution, and so it is of interest to study methods for estimating the parameters of this distribution. The estimation errors of three moment based estimation schemes are compared with the maximum likelihood estimation errors calculated via the Cramer-Rao lower bound. On the basis of this comparison, recommendations are made regarding the number of looks and the parameter estimation scheme that should be used to obtain near optimum estimation performance, without resorting to cumbersome numerical evaluations of the maximum likelihood solution. In particular, it is found that an estimator based on the mean and the variance of the data yields large errors, but an estimator based on the mean of the data and the mean of the log of the data is close to optimum.