Accuracy of Fisher Distributions and Log-moment Estimation to describe Amplitude Distributions of High Resolution SAR Images over Urban Areas

The framework of this study is classification of high resolution SAR images over urban areas. Statistics of these images reflect the presence of strong reflectors scattered all over; therefore histograms have a heavy tail. We propose a new distribution model (Fisher distribution) to fit such probability density functions. As its moments are not defined for all parameter values, we use a second kind statistics based estimation (log-moment estimation). The purpose of this article is the validation of both estimation method and distribution model. We first prove that, in this context, log-moment method is more accurate than moment method. We also demonstrate that Fisher functions are the most accurate for man-made structures. Finally these distributions are used in a Markovian classification.