Considering correlation properties on statistical simulation of clutter

FLESIA, ANA G.; LUCINI, MARÍA MAGDALENA; PEREZ, DARIO JAVIER

Learning and Inferring

College Publications

Lugar: Londres; Año: 2015; p. 72 - 89

egin{abstract} Statistical properties of image data are of paramount importance in the design of pattern recognition technics and the interpretation of their outputs. Image simulation allows quantification of method´s error and accuracy. In the case of SAR images, the contamination they suffer from a particular kind of noise, called speckle, which does not follow the classical hypothesis of entering the signal in an additive manner and obeying the Gaussian law, make them require a more careful treatment. Since the seminal work of~citet{Frery1997} a great variety of studies have been made targeting the specification of statistical properties of SAR data beyond classical assumptions. The ${cal G}$ distribution family proposed by Frery has been proved a flexible tool for the design of pattern recognition algorithms based on statistical modeling. Nevertheless, most of such work does not consider correlation present in the data as significant, which introduces an error in the model of particular regions of the imagery. The autocorrelation function can represent the structure of sea waves and the random variation made by the height and width of trees, along with the variability introduced in forests by the variation of wind intensity. Using the roughness parameter of the ${cal G}$ family for target discrimination alleviates this modeling error, since it was shown by~citet{Frery1997} that it characterizes heterogeneity in data. Classification accuracy is then tied to parameter estimation, which in this case it has been proved difficult,~citet{Lucini2002}, ~citet{Bustos2002}. In this paper we review some of our own simulation techniques to generate SAR clutter with pre-specified correlation properties, ~citet{Flesia1999}, ~citet{Bustos2001}, ~citet{Bustos2009}, and release a new set of routines in R for simulation studies based on such techniques. We give an example of the code versatility studying the change in accuracy of non-parametric techniques when correlated data is classified, compared with classification of uncorrelated data simulated with the same parameters. All code is available for download from AGF´s Reproducible Research website, cite{AGF}.