Abstract
Even though urban growth in metropolitan areas is driven by the interaction between city- and neighborhood-level factors, most conventional deep-learning-based urban growth models rely only on neighborhood-level grid variables, distorting the heterogeneity within and between cities. In addition, their single convolutional kernel learns spatial patterns at a single fixed scale. We propose a multi-modal multi-kernel convolutional long short-term memory (MM-MK-ConvLSTM) model to address the single-scale learning issue in existing deep-learning-based urban growth models. The multi-modal design encodes city-level factors as a city-level context that interacts with neighborhood-level features, enabling the model to learn both levels of factors across the metropolitan area. Furthermore, the attention-based multi-kernels and an encoder–decoder structure learn spatial patterns across multiple scales. Trained on the Seoul Metropolitan Area and applied to Yongin City, the proposed model outperformed ConvLSTM and DS-ConvLSTM across all metrics in late 2010s predictions, reproducing fine-grained fragmented patterns that the other models missed. Moreover, despite having more parameters, the proposed model achieved lower computational operations and faster convergence than both baselines. This demonstrates that multi-scale learning enhances the reliability and efficiency of metropolitan urban growth predictions. Urban expansion to 2040 was simulated using 12 scenarios integrating land demand, slope regulations, and development-scale constraints. Results revealed trade-offs: enhanced slope regulations suppress slope development but increase fragmented development on flat farmlands, whereas combining high-density and development-scale constraints most effectively restricts both slope development and fragmentation. These findings indicate that sustainable land-use planning requires a multi-dimensional policy approach rather than a single regulation.