Spatial interrelation matters: advancing 3D mineral prospectivity modeling with fully-connected CRFs—insights from Sanshandao Gold Belt, Eastern China

• A fully connected Conditional Random Field (CRF) model that relaxes the independent and identically distributed (IID) assumption in data-driven three-dimensional mineral prospectivity modeling. • Unary and pairwise potential neural networks for modeling mineralization associations and spatial interrelations within the CRF framework. • Integration of the spatial covariance of mineralization into the CRF’s pairwise potential to capture spatial continuity, heterogeneity, and anisotropy. The data-driven Three-Dimensional Mineral Prospectivity Modeling (3D MPM) has become an essential tool for localizing and quantifying concealed mineral resources. Machine learning techniques have become a cornerstone of 3D MPM, enabling the mapping of spatial associations between ore-controlling features and mineralization patterns. However, existing machine learning methods typically rely on the independent and identically distributed (IID) assumption, overlooking the inherent spatial interrelation in mineralization, which limits their predictive effectiveness and accuracy. This paper introduces a novel 3D MPM approach that addresses these limitations by incorporating spatial and contextual cues through a fully-connected Conditional Random Field (CRF) framework. To tailor the CRF for 3D MPM, a unary potential network is designed to capture mineralization associations at the 3D cell level, and a pairwise potential network is developed to model intercell interactions. Specifically, the spatial covariance of mineralization is incorporated into the CRF model to capture spatial continuity, heterogeneity, and anisotropy. This approach allows simultaneous association of mineralization prospectivity across all cells, leveraging their spatial interrelation to improve predictive performance. A case study conducted in the Sanshandao gold belt, Eastern China, compares the proposed CRF with mainstream machine learning-based methods and includes an ablation study. Results demonstrate the superiority of the CRF in prediction accuracy and targeting efficiency, highlighting its effectiveness in utilizing spatial dependencies to enhance 3D MPM performance.