Deep panel flow inference paper published on IJGIS

GeoDI lab's recent research article, introducing a gravity-informed Deep Spatial Evolution Network (DSEN) to infer panel spatial flows, has now been published in the International Journal of Geographical Information Science.

Zhu, D., & Ma, Z. (2025). Gravity-informed deep flow inference for spatial evolution modeling in panel data. International Journal of Geographical Information Science, 1-29. https://doi.org/10.1080/13658816.2025.2536512

Urban space is dynamic, but our data often comes in discrete snapshots of spatial distributions. What’s missing is a way to directly observe or recover the evolutionary process in between.

In this study, we present DSEN (Deep Spatial Evolution Network) — a gravity-informed, graph-based deep learning model that infers panel flows: the unobserved population movement that drives spatial changes between two time points. DSEN moves beyond the limitations of traditional cross-sectional models, including classical gravity and deep gravity models, by modeling flows as the mechanism of spatial evolution.

This paper builds upon two of our prior works:
🔹 https://www.tandfonline.com/doi/full/10.1080/13658816.2017.1413192
🔹 https://www-nature-com.ezp1.lib.umn.edu/articles/s41598-021-87902-8

With real-world mobile positioning data in the Twin Cities, DSEN shows strong performance in both event-driven and routine urban mobility — including holiday return migration and COVID-induced shifts — and demonstrates generalizability under sparse data.

📂 Code & data: https://github.com/GeoDI-Lab/DSEN-IJGIS

We hope this work contributes to advancing GeoAI, spatial interaction modeling, and practical flow inference in urban and environmental systems.