Microwave radiation has the ability to penetrate the clouds and offer insights into the structure of the storm and into the precipitation microphysics. For example, in a tropical cyclone microwave imagery can indicate the presence and location of low level circulation centers, banding features, and detect whether or not an eye is developing or already exists. These particular structural features of the tropical cyclones are not always discernible in conventional satellite imagery. Numerous physically based passive microwave precipitation retrieval algorithms have been developed and have seen great advances after the launch of TRMM in 1997. They make use of cloud-radiation databases built from the microphysical output of explicit cloud resolving models coupled with radiative transfer models. The major weakness of this kind of algorithms is the misrepresentation of the cloud model generated 3-D distribution of liquid and frozen hydrometeors, and of their microphysical properties. In order for this kind of algorithm to give reliable results it is essential not only to have a good match between the simulated brightness temperatures and the observations, but also to have cloud models able to reproduce the precipitation structures of the storm (with specific microphysics characteristics) that are actually observed. An unsupervised classification algorithm has been developed in order to identify and characterize hurricane structure using TRMM Microwave Imager (TMI) observations and to assess the ability of cloud resolving models to reproduce the structure of the hurricane. In this paper we will present the results of the classification for the TMI overpasses of Hurricane Bonnie and Hurricane Georges. We will show how the classification algorithm is able to discern between very specific regions in the storm, and how the TRMM Precipitation Radar (PR) can be used to interpret the results of the classification. We will compare these results with those obtained from the classification of cloud-radiation databases produced from explicit cloud resolving simulations of the two hurricanes. We will show how the classification algorithm can be used to: 1) investigate how much passive microwave observations can tell us about the structure of the storm; 2) test the ability of the cloud model to reproduce specific observed regions in tropical clyclones; 3) relate the misrepresentation of specific regions in the hurricane to specific assumptions and parametrizations in the cloud model microphysics scheme.