A hybrid spatiotemporal convolution-based cellular automata model (ST-CA) for land-use/cover change simulation

Abstract

Accurate land-use/-cover change (LUCC) simulation is of great significance to issues closely related to regional planning and policy-making. Many models have been committed to conducting LUCC simulations for better decision-making. However, LUCC is a nonlinear spatiotemporal process with complex links and feedback as well as latent dependencies in both spatial and temporal neighborhoods. They are challenging to be integrally utilized using existing models that employ conventional statistical or machine learning methods, inevitably leading to inaccurate LUCC simulations. Aiming to handle this problem, this paper innovatively proposed a hybrid spatiotemporal convolution-based cellular automata model (ST-CA) by coupling nonlinear spatiotemporal dependency learning and CA-based spatial allocation. A three-dimensional convolutional neural network (3D-CNN) was introduced in the model to assimilate both the nonlinear driving mechanism and spatiotemporal dependencies. It contributes to generating more elaborate development potentials to increase simulation accuracy. To evaluate the model performance, an LUCC simulation was applied on a national scale in China by ST-CA. Four traditional CA models, namely, logistic regression (LR)-CA, random forest (RF)-CA, full-connected neural network (FCN)-CA, and convolutional neural network (CNN)-CA, were also developed for accuracy comparisons. The results demonstrate that the simulation by ST-CA reached an FoM of 18.42%, which outperformed the other models with accuracy increases of 11.65%, 13.11%, 7.01%, and 2.29%, respectively. The proposed model incorporating 3D-CNN effectively captured the nonlinear spatiotemporal properties in the LUCC process, which is promising for more accurate LUCC simulations.