A feature space for scatterer characterization is constructed of the geophysical parameters of scatterers on size, shape, and orientation. Dense divisions are defined to discriminate target classes of possibly subtle distinction. The statistical similarity and uncertainty of overlapping cluster pairs are evaluated with Wishart based likelihood ratio and at a desired level of false classification the dense set of clusters is hierarchically pruned. Wishart based classification is then applied to the whole imagery, accomplishing a physical-based unsupervised classification algorithm. The algorithm is illustrated using an AIRSAR dataset of San Francisco to evaluate its capability in characterizing complex terrains. As an optional step, the K-Means or Expectation-Maximization iteration is performed to further adapt the cluster centers.