Explainable convolutional neural network for iron ore prospectivity mapping: a case study of the Yemaquan district, Qinghai, China

Abstract

Convolutional neural networks (CNNs) have been widely applied to mineral prospectivity mapping; however, their “black-box” nature limits geological interpretability and practical acceptance. To address this issue, this study proposes an explainable convolutional neural network (Ex-CNN) framework that integrates deep learning with SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP). The Yemaquan iron ore concentration area in Qinghai Province, northwestern China, was selected as a case study. Seven predictive factors of metallogenic prediction factors were constructed, including stratigraphy, intrusive bodies, magnetic anomalies, gravity anomalies, apparent polarizability, and apparent resistivity. These multi-source datasets were standardized into spatial grids and input into the CNN model to capture nonlinear relationships between geological features and mineralization. The results demonstrate that the proposed Ex-CNN model outperforms conventional machine learning and non-explainable deep learning models. Interpretability analysis based on SHAP, permutation importance, and PDP consistently indicates that apparent resistivity is the most influential factor controlling mineralization, reflecting the key role of hydrothermal alteration and structural disruption. Stratigraphy and intrusive bodies provide important geological constraints, while magnetic and gravity anomalies contribute auxiliary information. Several high-prospectivity targets were delineated in underexplored areas, confirming the effectiveness of the proposed framework. This study demonstrates that integrating explainable AI with deep learning significantly enhances both predictive performance and geological interpretability, providing reliable support for intelligent mineral exploration.