Drainage network generation using Deep Convolutional Generative Adversarial Networks

Topological and geomorphological characteristics of drainage networks are essential in order to assess the hydrologic response of catchments. It is also crucial for cost-optimal design and reliable operation of urban drainage networks. However, accurate modeling and prediction of such system are nontrivial because of inherent modeling uncertainty from the lack of information in multiscale river systems and climate conditions. Network theory with stochastic modeling has been applied to investigate the topological and functional aspects of both river networks and engineered urban drainage networks and support the optimal planning and management of catchment systems, but under several statistical assumptions that specific parametric models make.

In this presentation, we propose a deep-learning based urban drainage network modeling approach when a large number of data sets from catchment orthoimagery is available. In specific, Generative Adversarial Networks (GANs) is used to construct a generative river network model without any statistical assumption, only trained from river network data sharing similar network complexity and topological properties. To validate our proposed approach, Gibbs' model is used to generate stochastic network models for training. The results show that once trained, our proposed approach generates river networks efficiently and quickly. Each generated network is statistically close to the ones from Gibbs' model, which show a great potential to reliable urban drainage network stochastic modeling and efficient infrastructure design. We applied the proposed methodology to 10 urban catchments in Southeastern Chicago Areas and evaluated the applicability of the method.

Sandia National Laboratories is a multi-mission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-NA0003525.