GeoKG for GeoFM: Neuro-Symbolic Integration of Semantically-Enriched Geospatial-Temporal Data

Solving complex real‑world problems requires integrating heterogeneous data. Although recent Geospatial Foundation Models (GeoFMs) excel at pixel‑level analysis, they make only limited use of the geographic information contained in natural language texts and therefore often lack the semantic context needed for complex geospatial reasoning. This vision paper argues for reconceiving unstructured texts as a core component of a neuro‑symbolic framework that populates a geospatial‑temporal knowledge graph (GeoKG). We outline an architecture in which modality‑specific extractors produce multi‑granular footprints that are fused into a GeoKG, synchronizing symbolic representations with embeddings, geometries, and grid cells. This GeoKG supports geospatial reasoning and provides enriched training data for truly multimodal GeoFMs, for which we pose a set of research challenges.