An explainable semi-supervised deep learning framework for mineral prospectivity mapping: DEEP-SEAM v1.0
Over recent decades, the rate of mineral deposit discovery has been steadily declining, intensifying the demand for novel exploration methodologies. Deep learning (DL), with its capacity to effectively extract information relevant to target tasks, is emerging as a pivotal tool in mineral prospectivity mapping. However, the application of DL to multi-source, complex exploration datasets presents significant challenges, including nonlinear and highly heterogeneous features, sparse and imbalanced classes, and limited interpretability. In this study, we develop DEEP-SEAM v1.0, a novel interpretable semi-supervised DL framework. This framework introduces a comprehensive data preprocessing pipeline and incorporates a semi-supervised anomaly detection DL model, along with SHapley Additive exPlanations (SHAP) methodology. This approach leverages a limited number of positive samples together with extensive unlabeled data to effectively establish mappings between multi-source exploration data and mineralization probability, while enhancing understanding of the underlying decision-making mechanisms and logic of the DL model. This study targets rare earth element (REE) mineralization in the northern Curnamona Province, South Australia, utilizing a substantial open-source exploration dataset to generate evidence layers reflecting target mineralization. The prospective mineralization zones delineated by our framework demonstrate strong spatial correlation with known REE deposits. By comparing DEEP-SEAM's predictions with our understanding of the mineralization system, the established framework enhances both the reliability and credibility of prospective mineralization zone predictions.
This directory contains two subdirectories and essential configuration files necessary for reproducing the DEEP-SEAM framework: Input_Data_Layers/ - Contains the input datasets required for DEEP-SEAM framework execution Output_Features_Generated/ - Contains the generated datasets produced by the framework after execution requirements.txt - Specifies the exact versions of all required libraries to ensure reproducible execution
This notebook contains the main implementation of the proposed framework, encompassing three primary components: data processing, model training and prediction, and model interpretation.
This R script performs essential preprocessing operations for geochemical data, including Isometric Log-Ratio (ILR) transformation and Robust Principal Component Analysis (RPCA).