Key points are not available for this paper at this time.
Significant progress has been achieved in the applications of multisource datasets, comprising satellite observations, geophysical measurements, geochemical analysis, and geological information, for mineral prospectivity research. However, conventional mineral exploration techniques are increasingly challenged in complex geological settings, especially for deep-seated or concealed deposits, due to their limited capability to integrate and interpret highly heterogeneous multi-source datasets. Deep learning, as a state of the art methodology, can learn joint representations from such heterogeneous Earth observation and geoscientific data, integrate subtle spectral, spatial and structural patterns, and thereby support more efficient and targeted field campaigns by prioritising prospective areas and reducing unnecessary ground surveys. In this systematic review, 349 studies on deep learning for mineral exploration published between 2018 and 2025 are examined. A stage-dependent data framework is proposed to detail how data requirements evolve from regional reconnaissance to deposit-scale targeting. Subsequently, we critically evaluate major deep learning architectures and relate their specific strengths to distinct data characteristics, such as spatial grids, sequences, and graph topologies. Special attention is given to multisource fusion strategies, emphasizing the necessity of rigorous spatial consistency and physical complementarity. Furthermore, persistent challenges limiting operational uptake are identified, including discovery bias in label representativeness, class imbalance, interpretability, and cross-region generalization. Finally, emerging frontiers are discussed, including the role of geospatial foundation models and Large Language Models in automating knowledge extraction.
Yuan et al. (Fri,) studied this question.