Redefining Regions in Space and Time: A Deep Learning Method for Spatio-Temporal Clustering

Identifying regions that are both spatially contiguous and internally homogeneous remains a core challenge in spatial analysis and regional economics, especially with the increasing complexity of modern datasets. These limitations are particularly problematic when working with socioeconomic data that evolve over time. This paper presents a novel methodology for spatio-temporal regionalization—Spatial Deep Embedded Clustering (SDEC)—which integrates deep learning with spatially constrained clustering to effectively process time series data. The approach uses autoencoders to capture hidden temporal patterns and reduce dimensionality before clustering, ensuring that both spatial contiguity and temporal coherence are maintained. Through Monte Carlo simulations, we show that SDEC significantly outperforms traditional methods in capturing complex temporal patterns while preserving spatial structure. Using empirical examples, we demonstrate that the proposed framework provides a robust, scalable, and data-driven tool for researchers and policymakers working in public health, urban planning, and regional economic analysis.