Forecasting urban expansion: A dynamic urban growth model using DS-ConvLSTM to simulate multi-land regulation scenarios

This research addresses the computational inefficiency problem in deep learning-based urban growth modeling. This study proposes a novel Depthwise Separable Convolutional Long Short-Term Memory (DS-ConvLSTM) model to predict the urban growth patterns in Hanam City South Korea by 2030. The model incorporates six scenarios that reflect diverse land demands and urbanization patterns. Integrating 40 years of data, DS-ConvLSTM demonstrated superior performance compared to existing models, such as Convolutional Long Short-Term Memory (ConvLSTM), achieving an accuracy, F1-score, and Figure of Merit of 0.9801, 0.9510, and 0.8092, respectively. Notably, its efficient design reduces the network parameters by more than half compared to the ConvLSTM model, thereby decreasing model complexity. The study further explores potential land demand based on population and economic growth projections, ranging from 27.15 km 2 to 29.31 km 2 . The analysis reveals trade-offs between development approaches. Business-as-usual scenarios lead to agricultural and forestland loss, while ecologically-focused development prioritizes forest preservation but increases development pressure on agricultural land. Sustainable compact development reduces land loss due to urban expansion through high-density redevelopment. However, high-density areas can lead to concentrated traffic congestion and environmental pollution. These findings provide valuable insights for urban planners, enabling them to make data-driven decisions regarding future land use policies. • DS-ConvLSTM outperformed existing models in terms of prediction overall accuracy. • DS-ConvLSTM outperforms ConvLSTM in capturing spatiotemporal urban change, while using less than half its parameters. • DS-ConvLSTM is more efficient than ConvLSTM in data-driven decision-making for determining potential future urban land.