Learning effective geospatial embeddings is crucial for a series of
geospatial applications such as city analytics and earth monitoring. However,
learning comprehensive region representations presents two significant
challenges: first, the deficiency of effective intra-region feature
representation; and second, the difficulty of learning from intricate
inter-region dependencies. In this paper, we present GeoHG, an effective
heterogeneous graph structure for learning comprehensive region embeddings for
various downstream tasks. Specifically, we tailor satellite image
representation learning through geo-entity segmentation and point-of-interest
(POI) integration for expressive intra-regional features. Furthermore, GeoHG
unifies informative spatial interdependencies and socio-environmental
attributes into a powerful heterogeneous graph to encourage explicit modeling
of higher-order inter-regional relationships. The intra-regional features and
inter-regional correlations are seamlessly integrated by a model-agnostic graph
learning framework for diverse downstream tasks. Extensive experiments
demonstrate the effectiveness of GeoHG in geo-prediction tasks compared to
existing methods, even under extreme data scarcity (with just 5% of training
data). With interpretable region representations, GeoHG exhibits strong
generalization capabilities across regions. We will release code and data upon
paper notification.

Geospatial embeddings are crucial for city analytics and earth monitoring, but learning comprehensive region representations is challenging; this paper introduces GeoHG, an effective heterogeneous graph structure, to learn region embeddings by addressing intra-region feature representation and inter-region dependencies, demonstrating superior performance in geo-prediction tasks even with limited training data.