Journal of Digital Landscape Architecture

Seite drucken

RiverGAN: Fluvial Landform Generation Based on Physical Simulations and Generative Adversarial Network

Physical river models and simulations are used in the field of landscape architecture to study the complex and nonlinear process of river morphology, which is difficult to be studied by numerical simulations. However, the setup of physical simulations is usually sophisticated, costly and time-consuming. Thus, the physical models can only produce a limited number of simulation results for each study, which significantly limits the use of the physical simulation and the designers’ imagination as well. In this research, we take the simulation of fluvial landscape on a hydromophological sand table as an example, to present a novel and alternative framework for faster, more accessible and iterative method for physical simulations for landscape designers. By using the methodological framework of Procedural Terrain Generation (PTG) with Generative Adversarial Network (GANs), we collected 500 pairs of mesh and texture data from a hydromorphological sand table for training a LightweightGAN model. The resulting 3D latent walk is visualized through a customized user interface and can be easily integrated into the creative workflow by landscape designers.

Autor / Author: Liu, Xun; Tian, Runjia
Institution / Institution: University of Virginia, Charlottesville/USA; Harvard University, Cambridge/USA
Seitenzahl / Pages: 7
Sprache / Language: Englisch
Veröffentlichung / Publication: JoDLA – Journal of Digital Landscape Architecture, 7-2022
Tagung / Conference: Digital Landscape Architecture 2022 – Hybrid Landscapes
Veranstaltungsort, -datum / Venue, Date: Harvard University, Cambridge Mass, USA 09-06-22 - 10-06-22
Schlüsselwörter (de):
Keywords (en): Generative Adversarial Network, landscape architecture, Landform Generation, Procedural Terrain Generation, hydromorphology
Paper review type: Full Paper Review
DOI: doi:10.14627/537724011
7387 -