Measuring Synthetic Aperture Radar target differences with stochastic distances

Synthetic Aperture Radar (SAR) imagery plays a central role as a source of unique data for Geographic Information Systems. These data sets provide complementary information to that provided by optical and infra-red sensors as, for instance, Landsat TM, CBERS-2, IKONOS and SPOT, to name a few. SAR sensors capture information about the target roughness and its dielectric properties, and their imaging capabilities are able to penetrate clouds, fog, rain and even some types of land cover as, for instance, forest canopies. A major issue related to the use of SAR images is their statistical behavior. It is well known that classical Gaussian and additive models do not hold for such data. The multiplicative model (MM) has been extensively tested with success, and it is able to explain phenomenological aspects of the image formation. One of the most important distributions related to the MM is the G° law. The G° distribution, as all other laws related to the MM, greatly departs from the Gaussian model. This paper assesses the SAR image discrimination capabilities of selected parametric methods based on divergences measures, when compared to the nonparametric Kolmogorov-Smirnov testing methodology. The importance of the Triangular and Arithmetic-Geometric distances is quantified with respect to the Kullback-Leibler parametric and Kolmogorov-Smirnov non-parametric classical distances by means of Monte Carlo simulation.