A generative deep learning approach with multi-perspective map representations for simulating land use changes

Da Wang

doi:10.1080/13658816.2026.2663121

A generative deep learning approach with multi-perspective map representations for simulating land use changes

2026-05-11

0

PRE

AI

DOI：10.1080/13658816.2026.2663121

Original

Original request for help

Save

Abstract

En 中文

Land-use simulation is essential for understanding future spatiotemporal changes on the Earth’s surface. Most existing methods simulate land use change by modelling dynamic processes among cell labels, often overlooking the landscape-scale patterns that emerge from these cells. Moreover, reliably constructing evolutionary rules and key parameters to simulate future land use change remains a fundamental challenge. We proposed MapsGT, a novel generative deep learning approach for the simulation of land use change over multiple classes. The framework consists of two core components: MapsVAE, which learns compact landscape encodings to capture spatial information from historical maps of land-use and geographic environment; and EMTrans, which models spatiotemporal dependencies within these encodings to generate future maps of land use. Experiments in the Pearl River Delta and the Changsha-Zhuzhou-Xiangtan urban agglomerations showed that MapsGT excels in short-term simulation, achieving a high Figure of Merit and User’s Accuracy. Additionally, a newly developed regional consistency index (RCI) reveals the model’s ability to capture and reproduce complex landscape patterns with high spatial authenticity. MapsGT provides a valuable framework for geospatial artificial intelligence that complements existing simulation methods of land use change.

Keywords:

Land use simulation

generative deep learning

geospatial artificial intelligence

A generative deep learning approach with multi-perspective map representations for simulating land use changes

Abstract

Journal

Organization