Urban environments are challenging to model because they are both very large and very diverse. The size of the environment makes capturing the details of every structure prohibitive, leaving us with only sparse information. Photogrammetric reconstruction attempts to build 3D structures but is often not robust and requires manual input, thus making it difficult to scale to large areas. Procedural modeling of urban environments provides a means for quickly creating architecture, but it does not address populating the database of urban features and it does not use real-world information to mimic actual urban areas. In this paper, we use sparse aerial-based information about an urban environment that is readily available in many towns and cities. From this information, we isolate a small number of canonical urban structures that can be assembled algorithmically using only a few simple operations. The result is an approximation of the entire urban environment to within a prespecified error threshold of the original data. Unlike synthetic approaches that prescribe a grammar, our task is to discover a small dictionary of words and associated application rules for an actual urban environment. Once the dictionary and rules are known, an urban environment in the style of the original can be instantiated. The canonical set of words can be used to prioritize acquisition efforts, to reduce image data, to compress geometry of the model, to find urban patterns, and to visualize properties of the urban environment. CR Categories: I.3.5 [Computer Graphics]: Computational Geometry and Object Modeling, I.3.7 [Computer Graphics]: Three-dimensional Graphics and Realism, I.6.3 [Simulation and Modeling]: Applications.