III: Medium: Collaborative Research: Deep Generative Modeling for Urban and Archaeological Recovery

Modeling and understanding the evolution of urbanization over the course of human history elucidates a key aspect of human civilization, and can significantly help stakeholders today make better informed decisions for future urban development. However, the modeling of current and past urban spaces remains extremely challenging and a rigorous comparison between ancient and modern urban form is lacking. In this project, the team will provide an artificial intelligence based framework for discovering a relatively complex urban model (walls, corners, rooms, orientation, and built area clusters) from a sparse number of remote sensing and field observations. As opposed to cities present today, modeling a historical urban site is fundamentally limited to sparse (and few) data observations because most of the structures have been eroded or destroyed. The research team will provide a preliminary cyberinfrastructure, pursue 3D re-creations of historical sites, create a feature- and time-based urban taxonomy of ancient sites from the late Prehispanic and Colonial period Andes and the Bronze/Iron Age South Caucasus periods, while leveraging the NEH and American Council of Learned Societies funded GeoPACHA web platform for result dissemination. Moreover, the project spans three major US universities and five departments, led by five experienced senior researchers and a team of at least six multidisciplinary graduate students, as well as additional undergraduates, who will produce publications in top tier venues, conference workshops, as well as theses and PhD dissertations. To assist with modeling and understanding the evolution of urbanization over the course of human history, this project seeks a computational methodology for discovering a relatively complex urban model from a sparse number of observations. While performing a dense acquisition of a current city implies focusing on sensor deployment and on big data issues, modeling a historical urban site is fundamentally limited to sparse (and few) data observations because most of the structures have been eroded or destroyed. Inferencing approaches show significant promise, but they struggle in a situation of relatively sparse data and obscured structure. As a first domain application, the team will assist computational archaeologists having relatively sparse data but of an underlying structured site. First, they will solve a set cover problem to determine a discrete set of atomic elements and rules that are minimal yet sufficient to span the sparse data. Second, they will use these atomic elements and rules to produce sufficient data samples for training deep networks in a self-supervised manner in order to learn how to perform segmentation, classification, and completion. Finally, they will use the learned representations to model archaeological sites resulting in reconstructions, semantic understandings, and site taxonomies, for instance. Further, the team anticipates that the developed models can be re-tooled to assist with other domains also limited to sparse observations of an underlying structured region. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.