Coupling cellular automata with area partitioning and spatiotemporal convolution for dynamic land use change...

Area of Science:

• Geographic Information Science
• Urban Planning
• Computational Geography

Background:

• Urbanization drives significant land use change (LUC), necessitating accurate simulation models for effective urban planning.
• Existing cellular automata (CA) models often overlook spatial heterogeneity and spatiotemporal dependencies in neighborhood interactions.
• Traditional CA models apply uniform transition rules across diverse sub-regions and use single time slices for neighborhood analysis.

Purpose of the Study:

• To develop a novel hybrid cellular automata model (PST-CA) that addresses limitations in simulating land use change.
• To incorporate spatial heterogeneity through area partitioning and capture spatiotemporal neighborhood dynamics.
• To enhance the accuracy and reliability of land use change modeling for urban planning.

Main Methods:

• A machine-learning-based partitioning strategy using Self-Organizing Maps (SOM) to divide regions into homogeneous sub-regions.
• Extraction of spatiotemporal neighborhood features using a 3D Convolutional Neural Network (3D CNN).
• Integration of neighborhood features, driving factors, and constraints within an Artificial Neural Network (ANN) to generate sub-region specific conversion probability maps.

Main Results:

• The proposed PST-CA model demonstrated superior performance compared to four traditional models (LR-CA, SVM-CA, RF-CA, ANN-CA), achieving a 4.66%–6.41% increase in overall accuracy.
• SOM partitioning revealed distinct spatial patterns in built-up area growth, confirming its effectiveness in handling spatial heterogeneity.
• Optimal time steps in 3D CNN correlated positively with built-up area growth rates, indicating the importance of capturing long-term temporal dependencies.

Conclusions:

• The PST-CA model offers a significant advancement in land use change simulation by effectively integrating spatial partitioning and spatiotemporal feature learning.
• The findings highlight the necessity of considering sub-regional characteristics and long-term neighborhood dynamics for accurate urban modeling.
• This approach provides a more robust tool for supporting urban planning and decision-making in rapidly changing environments.