Seminar Details

Global spatially-explicit long-term simulations of urban land expansion under alternative scenarios are essential for studying interactions between human societies and global environmental change. This talk presents a data-science approach using newly available time series of fine-spatial-resolution remote sensing observations of urban land spanning the past 30 years, along with spatial population, economic growth, and environmental variables, to developing a spatiotemporal simulation model accommodating the needs of global environmental change impact studies. The model makes future projections by allowing spatial entities of different scales (nations, regions, and grid cells) to organically evolve along and switch between different urbanization trajectories identified via mining past urbanization patterns. It was used in combination with Monte Carlo experiments to generate urbanization pathways reflecting narrative-based global socioeconomic scenarios. Our approach is unique and improves existing large-scale spatial urban land modeling efforts in at least four ways: close calibration to historical time series data, explicit portrait of subnational, local spatial variations, robust long-term extrapolation, and integration of both quantitative and qualitative information. This work demonstrates challenges and rewards of applying machine-learning models in studying long-term large-scale human-environment interactions. The results show that the amount of urban land on Earth by 2100 strongly depends on societal choices and actions in years to come. For example, the total amount of urban land in 2100 in the U.S. can be 1.3 – 4.0 times its present-day amount, and the range for China is 1.9 – 5.9.