Bias Correction and Modified Profile Likelihood under the Wishart Complex Distribution
This work provides incremental improvements to parameter estimation methods for PolSAR image analysis, specifically benefiting researchers and practitioners in remote sensing.
This paper tackles the problem of estimating the equivalent number of looks parameter L in polarimetric synthetic aperture radar (PolSAR) images, which controls signal-noise ratio to reduce speckle noise. The authors propose two new maximum likelihood estimators based on bias correction and modified profile likelihood, which outperform existing methods in homogeneous scenarios with lower mean square error, bias, and coefficient of variation.
This paper proposes improved methods for the maximum likelihood (ML) estimation of the equivalent number of looks $L$. This parameter has a meaningful interpretation in the context of polarimetric synthetic aperture radar (PolSAR) images. Due to the presence of coherent illumination in their processing, PolSAR systems generate images which present a granular noise called speckle. As a potential solution for reducing such interference, the parameter $L$ controls the signal-noise ratio. Thus, the proposal of efficient estimation methodologies for $L$ has been sought. To that end, we consider firstly that a PolSAR image is well described by the scaled complex Wishart distribution. In recent years, Anfinsen et al. derived and analyzed estimation methods based on the ML and on trace statistical moments for obtaining the parameter $L$ of the unscaled version of such probability law. This paper generalizes that approach. We present the second-order bias expression proposed by Cox and Snell for the ML estimator of this parameter. Moreover, the formula of the profile likelihood modified by Barndorff-Nielsen in terms of $L$ is discussed. Such derivations yield two new ML estimators for the parameter $L$, which are compared to the estimators proposed by Anfinsen et al. The performance of these estimators is assessed by means of Monte Carlo experiments, adopting three statistical measures as comparison criterion: the mean square error, the bias, and the coefficient of variation. Equivalently to the simulation study, an application to actual PolSAR data concludes that the proposed estimators outperform all the others in homogeneous scenarios.