Modeling High Resolution Sar Images of Urban Areas Using Mixture Models and Hidden Markov Model

A distribution of Synthetic Aperture Radar (SAR) amplitude data in urban areas has a heavy tail. Most of the energy, i.e. composed of low radar return from natural areas and shadow regions, concentrates in a tiny portion of the distribution. Some strong backscattering from man-made structures lie in the long tail. The contrast between the clutter and building layover is extremely high. Furthermore, the buildings spans a much bigger extent along the distribution, which results in wide diversity and complexity for building analysis. Various distributions have been proposed for SAR amplitude data, including K-, Weibull, Log-normal, Nakagami-Gamma, generalized Rayleigh, Fisher distribution [1], and so on. Fig. 1 shows four examples of SAR amplitude data fitted by Fisher, Gamma and Log-normal distribution. Fisher distribution is capable of capturing the variations in amplitude distribution. In [1] each object class in an urban area is modeled by a Fisher distribution. However, a single distribution is not enough to describe the variations inside an object class. A more refined model, e.g. a mixture model, is required.