A novel multi-scale deep learning framework for adaptive urban expansion simulation

Abstract

Unplanned urban sprawl can cause environmental degradation, infrastructural overload, and a diminished quality of life. Consequently, accurate urban expansion prediction models are vital for guiding sustainable city development and aligning policy decisions with long-term community needs. Current urban expansion simulation model often struggles with rigid neighbourhood definitions and static factor weighting in conventional Cellular Automata (CA) models, causing inaccuracies in capturing non-linear, multi-scale dynamics. To address these limitations, we propose an enhanced U-Net++–based framework that adaptively integrates attention mechanisms and autoregressive CA simulation. Tested on Changsha city’s multi-source spatiotemporal data from 2010 to 2022, our model achieved an Overall Accuracy of 0.87 and a Figure of Merit of 0.90, surpassing traditional methods in both quantitative metrics and visual coherence. By leveraging multi-scale convolutional features and attention-driven factor weighting, this approach effectively addresses the challenges of multi-scale and spatiotemporal heterogeneity. Furthermore, its end-to-end design allows for adaptive optimization of the entire prediction process, yielding more coherent and robust urban expansion forecasts. Moving forward, integrating wider socio-political drivers and policy constraints can further enhance the model’s practicality for sustainable urban development, ensuring that city growth management aligns with environmental goals and community well-being. This modelling framework is applicable to predict any urban setting globally.