SAR imagery is becoming more and more used. Among its advantages over visible range imagery, one could mention those related to the facts that there is no need to operate the satellite with daylight, that images could be obtained even in the presence of clouds, fog, etc., and that the returned signal carries information about the dielectrical properties of the soil. For a comprehensive summary of the statistical properties of SAR images, the reader is referred to Kelly et al. (1988), Derin et al. (1990), and the references therein.
A simple, though theoretically tractable and physically sensible, model for the marginal distribution of the observed values in every pixel, is the Rayleigh distribution. This assumption departs from the more classical hypothesis of normality, commonly used for observations in the visible range (e. g. SPOT and LANDSAT images).
Given a SAR image, it is often neccessary to estimate the parameters of the Rayleigh distributions whose outcomes we are seeing, for instance, to be used as input information for filters, segmentation algorithms, etc. (see Duda and Hart (1973) for details about these and other related procedures). These estimations are usually carried out by selecting a window (a subset of the image) of homogeneous observations, and then estimating the parameter using the data there contained. A common problem is that the chosen window might not be completely homogeneous, and some observations within it could have come from other population(s): we then say that there is a contamination. It is therefore interesting to evaluate the performance of several estimation procedures in order to compare them, either with and without the presence of contamination.
In Frery and Sant'Anna (1993) an implementation of the estimators that will illustrate this paper is shown. They are used as SAR image filters, and are applied in the reduction of the speckle noise present in a real image. The results presented there show that the overall noise is efectivelly reduced, and that details and sharp features are better preserved with the use of the robust estimators (filters) than with the use of the maximum likelihood and moments estimators (filters).