Spatial sensitivity analysis for urban land use prediction with physics-constrained conditional generative adversarial networks

• DOI: 10.48550/arxiv.1907.09543
• Publication Date: 2019-01-01
• Authors: Adrian Albert; Jasleen Kaur; Emanuele Strano; Marta C. González
• Keywords: Generative adversarial network

Accurately forecasting urban development and its environmental climate impacts critically depends on realistic models of the spatial structure built environment, dependence key factors such as population economic development. Scenario simulation sensitivity analysis, i.e., predicting how changes in underlying at a given location affect urbanization outcomes other locations, is currently not achievable large scale with traditional growth models, which are either too simplistic, or depend detailed locally-collected socioeconomic data that available most places. Here we develop framework to estimate, purely from globally-available remote-sensing without parametric assumptions, (\textit{static}) rate change sprawl macroeconomic indicators. We formulate this regression problem an image-to-image translation task using conditional generative adversarial networks (GANs), where gradients necessary for comparative static analysis provided by backpropagation algorithm used train model. This allows naturally incorporate physical constraints, e.g., inability build over water bodies. To validate model-generated environment distributions, use statistics commonly form analysis. apply our method novel dataset comprising layers nightlighs measurements (a proxy energy use), density world's populous 15,000 cities.