Urban expansion simulation with an explainable ensemble deep learning framework.

Area of Science:

• Urban and Regional Planning
• Geographic Information Science
• Artificial Intelligence

Background:

• Urban expansion simulation is critical for effective land management and policy development.
• Deep learning advancements offer improved accuracy in predicting urban land dynamics.
• Existing methods face challenges in interpretability and intra-urban granular simulation.

Purpose of the Study:

• To propose a novel deep learning-based ensemble framework for intra-urban granular urban expansion simulation.
• To enhance the interpretability of deep learning models in spatial analysis.
• To improve the accuracy and temporal consistency of urban expansion predictions.

Main Methods:

• Developed an ensemble framework utilizing transformers for multi-temporal spatial features and convolutional layers for single-temporal features.
• Integrated a channel-wise attention module to assess feature importance and model interpretability.
• Applied the framework to simulate urban expansion in Shenzhen, China.

Main Results:

• The proposed deep learning framework accurately simulated urban expansion at an intra-urban level.
• The channel-wise attention module improved the interpretability and confidence in simulation results.
• The method outperformed baseline approaches in both spatial accuracy and temporal consistency.

Conclusions:

• The novel deep learning ensemble framework provides a robust and interpretable solution for urban expansion simulation.
• This approach offers significant potential for enhancing land management and urban policymaking.
• Accurate intra-urban scale predictions are achievable with advanced deep learning techniques.