Overview
How can geospatial models transfer across regions, scales, and sensor modalities without extensive retraining? This research line tackles the generalization challenge in GeoAI through LLM-driven model discovery, graph-based community detection, and scalable urban analytics.
GeoEvolve uses multi-agent large language models to automate geospatial model development — enabling AI to self-evolve spatial models. Complementary work on overlapping community structure in urban mobility reveals how spatial communities explain movement patterns at scale. These methods have been applied to shared micromobility, logistics optimization, and urban land-use mapping.
Key Contributions
- Label propagation for detecting overlapping communities in shared micromobility (CEUS 2025)
- Graph-based deep spatial regression for uncovering spatial process heterogeneity (ISPRS JPRS 2026)
- GIMI: geographically generalizable image-to-image search engine (ICLR 2024 workshop)
- Scalable urban logistics optimization with hybrid sparrow search algorithm
Software
- GeoEvolve — LLM-based automatic spatial model development
Selected Publications
Modeling shared micromobility as a label propagation process for detecting the overlapping communities
Transferable GeoAI
CEUS, 2025
Uncovering spatial process heterogeneity from graph-based deep spatial regression
Spatial ExplicitTransferable GeoAI
ISPRS JPRS, 2026
GIMI: A Geographical Generalizable Image-To-Image Search Engine
Transferable GeoAI
ICLR 2024 Workshop, 2024
Sensing overlapping geospatial communities from human movements using graph affiliation generation models
Transferable GeoAI
ACM SIGSPATIAL Workshop on AI for Geographic Knowledge Discovery, pp.1-9, 2022