Mineral prospectivity mapping using geological map semantic knowledge graph embedding: a case study of gold prospecting in Ankang, Shaanxi Province, China

Data-driven MPM often overlooks expert knowledge, leading to poor interpretability and overly broad predictions. We convert the semantic information of geological maps into a semantic knowledge graph(Geo-mapSKG). By embedding the Geo-mapSKG using the TransG model and integrating with geochemical data to enhance the knowledge constraints. Given the spatial variability of geological features, we use a window sampling method for data collection to ensure the completeness of geospatial structural features. To improve the model’s ability to learn the complex variations in geospatial features, we employ the Conformer deep learning model for gold prospectivity prediction. This approach combines the local geological feature extraction capability of CNN with the Transformer’s overall geological dependencies. To validate the method effectiveness, a gold prospective exercise was conducted at Ankang in Shaanxi Province (North China). Results show Geo-mapSKG embedding effectively constrains predicted area distribution, yielding a smaller predicted area, and that the geological semantic features of the predicted areas show strong consistency with the ore geological features of known deposits. Compared with the prediction results of the CNN and Transformer models, the accuracy of the Conformer model is 1.38% higher than the CNN model and 2.92% higher than the Transformer model.