FU3D: the first global projections of future urban three-dimensional (3D) expansion for the 21st century under shared socioeconomic pathways

Global-scale and long-term projections of future urban 3D expansion are essential for understanding the environmental effects of future urbanization. We develop a multi-task learning-based and end-to-end cellular automaton model for urban 3D projection (named MECA-3D). Within the MECA-3D model, the top-down component determines urban built-up volume demand based on panel data regression model and socioeconomic factors, while the bottom-up component estimates urban land suitability and built-up height by a multi-task residual neural network model. Using the MECA-3D model, for the first time, we present the global projections of future urban 3D expansion dataset (named FU3D) from 2020 to 2100 at a 1km resolution under the five Shared Socioeconomic Pathways (SSPs). Our projections show that by 2100, global urban built-up volume will increase to about 184%–409% under the five SSPs, with the largest 3D expansion (exceeding 4000 km3) projected under the fossil-fuelled development scenario (SSP5). The validation procedures confirm that the FU3D dataset exhibits sufficient accuracy, long-term reliability and reasonable uncertainty. In general, our FU3D dataset overcomes the limitations of 2D projections, provides valuable 3D morphological information for future sustainable cities and can serve as a valuable input in relevant fields.

全球尺度、长期的未来城市三维扩张预测，对于解析未来城市化的环境影响至关重要。本研究构建了一款基于多任务学习的端到端元胞自动机（cellular automaton）模型，用于城市三维扩张预测，该模型命名为MECA-3D。在MECA-3D模型中，自上而下模块依托面板数据回归模型（panel data regression model）与社会经济因素确定城市建成体量需求；自下而上模块则通过多任务残差神经网络（multi-task residual neural network）估算城市土地适宜性与建成高度。依托MECA-3D模型，本研究首次构建了五种共享社会经济路径（Shared Socioeconomic Pathways，SSPs）下、2020至2100年、分辨率为1km的全球未来城市三维扩张预测数据集，命名为FU3D。预测结果显示，至2100年，在五种SSPs情景下，全球城市建成体量将增长至初始水平的184%~409%；其中化石燃料驱动发展情景（SSP5）下的三维扩张规模最大，将超过4000立方千米。验证流程证实，FU3D数据集具备充足的预测精度、长期可靠性与合理的不确定性范围。总体而言，本研究构建的FU3D数据集突破了二维预测的局限性，可为未来可持续城市提供极具价值的三维形态信息，同时可作为相关领域研究的重要输入数据。