<p>Spatial relation extraction (SRE) from geological texts is critical for constructing structured knowledge bases in geographic information systems (GIS). Current methods predominantly rely on a closed-set hypothesis to identify predefined relations and operate primarily at the word level. However, these methods struggle with the nuanced semantics, pervasive polysemy, and complex long-range dependencies characteristic of geological texts, and remain vulnerable to Chinese word segmentation errors. To overcome these limitations, we introduce OSRE, an open SRE framework specifically designed for spatial relations in geoscience, which avoids predefined relational schemas and thereby enables the discovery of unconstrained spatial relations. This approach initially generates semantically rich character embeddings using the domain-adapted Geo-BERT model. Subsequently, a structured dependency graph is constructed by connecting words through syntactic dependency edges, while also linking each word to its constituent characters. Following the graph construction, a graph attention network is applied to aggregate multi-hop contextual information over the graph, thereby capturing complex geological semantics. Finally, a novel relation-first two-stage span extraction strategy is employed to achieve geoscientific triplet extraction. Evaluated on specialized geoscience corpora, OSRE achieves state-of-the-art performance with an 8.3% F1-score improvement over baselines, demonstrating enhanced capability for structured spatial knowledge extraction critical to GIS applications including 3D geological modeling and geographic question answering.</p>