Parameter Estimation Assessment by Information Content in Speckled Imagery

LUCINI, MARÍA MAGDALENA; GAMBINI, JULIANA; FRERY, ALEJANDRO CESAR

Valdivia

Conferencia; XIX Conference on Nonequilibrium Statistical Mechanics and nonlinear Physics,; 2016

In this work we analyze a SAR (Synthetic Aperture Radar) image featureextraction method based on the estimation of the parameters of a statisticalmodel largely used for this kind of data. The statistical model used is the $\mathcal{G}_I^0$distribution,which with very few parameters is able to characterize a large numberof targets in monopolarized SAR imagery, deserving the denomination of ¨UniversalModel¨.One of these parameters, denoted $\alpha$, is instrinsecaly related to the roughness or texture of the backscatter, being this fact one of the reasons why its estimation is of paramount importance and receives a great deal of attention in the cientific community.We here analyze the results of extracting this feature using several methods andalgorithms for parameter estimation, including Maximum Likelihood, methodsbased on fractional moments , log-cumulants and a procedure based on asymmetric kernels and stochastic distances, as a previous step in SAR image classification.The quality of these estimates is assessed by the information content they convey.Maps generated by different estimators of $\alpha$ are used as the input for classificationmethods to discriminate land cover types and their performances are addressedin terms of the accuracy of the classification. The methodology here presentedcan also be applied to any image generated by coherent illumination, as is thecase of ultrasound, laser and sonar images which can also be modeled by the $\mathcal{G}_I^0$distribution.