Multi-model decision system: An ensemble deep learning model to enhance predictive power in mineral prospectivity mapping

Deep learning (DL) models have emerged as cutting-edge technologies in the recent decade and have shown remarkable capabilities for metal exploration and mineral prospectivity mapping (MPM). The relevance of DL architectures for MPM applications is attributed to their robust competencies in auto-identifying non-linear features and handling big exploration data in complex Earth systems. However, depending on the use of different models, DL-based MPM procedures may result in diverse mineralization-related spatial patterns. This instability can introduce uncertainty into DL-derived MPM predictions, making it challenging to select the appropriate DL architecture. Here, we conceptualize and discuss an innovative ensemble system designed to create synergies between multiple DL-based predictions, thereby mitigating instabilities from mineralization-related spatial patterns derived from different DL models. The proposed methodology, a multi-model decision system (MMDS), entails a decision-making protocol to fuse MPM predictions from deep neural network, deep belief network, deep forest, and one-dimensional convolutional neural network-type DL models. A decision-making engine inspired by the MARCOS model was also implemented, whereby high-performance DL models are allowed to play a more significant role in generating final MPM predictions based on their corresponding generalizability (i.e., delivered F1-Score values). The relevance of MMDS in MPM was demonstrated through its application to the exploration targeting of IOCG-type mineralization within a brownfield terrain in NE Iran. Success-rate curves and corresponding areas under the curves indicated that the resulting MMDS-derived prospectivity map performed better with regards to vectoring toward mineral exploration targets than stand-alone prospectivity models. The findings of this study suggest that the newly developed ensemble-based decision system can give weight to high-performance DL models more efficiently, thereby enhancing the definition of lower-risk target areas for further exploration.