Abstract: Amid intensifying global competition for mineral resources, raising domestic security of strategic minerals requires higher-precision, more explainable exploration. Although petrology, spectroscopy, and mineralogy have amassed large volumes of heterogeneous data, limited cross-source fusion, weak semantic linkage, and misaligned taxonomies hinder their use in exploration. This paper proposes a systematic “Rock-Mineral-Spectrum” Knowledge Graph (RMS-KG) to address these gaps. We integrate remote-sensing imagery, reflectance spectra, mineral characteristics, and geological literature using a hybrid ontology approach that combines top-down domain modeling with bottom-up data construction. The schema covers core concepts in rock taxonomy, mineral attributes, and spectral features. Deep learning and semantic parsing extract entities, attributes, and relations from structured databases, semi-structured reports, and unstructured texts; knowledge is then fused in a graph database to enable semantic linkage, visual querying, and dynamic reasoning. RMS-KG contains on the order of tens of thousands of nodes and edges and includes more than 1, 000 rock-mineral types. It unifies the “rock-mineral-spectrum” semantics, supports mapping spectral fingerprints to minerals and rocks, and enables metallogenic-type inference from mineral assemblages. Two application scenarios, “spectrum-guided mineral identification” and “metallogenic-type inference”, demonstrate its effectiveness and interpretability. RMS-KG provides a reusable knowledge substrate and reasoning capability for rock-mineral recognition and prospecting, improving the retrievability, computability, and reusability of geological knowledge and offering a generalizable paradigm for knowledge-centric geological AI.