CNN-Transformers for mineral prospectivity mapping in the Maodeng-Baiyinchagan area, Southern Great Xing'an Range

There have been recent breakthroughs in exploration for Sn and Ag mineral resources in the Southern Great Xing'an Range, making this region a world-class Sn and Ag metallogenic area. However, the thick surface cover and concealed deposits have made exploration challenging, severely limiting the effectiveness of traditional exploration methods. The utilization of deep learning techniques in conjunction with multi-source geoscience datasets for comprehensive metallogenic prognosis has emerged as a novel means of geological prospecting. In this study, a comprehensive geological data collection for the Maodeng-Baiyinchagan area was systematically conducted to form a multi-disciplinary geoscience information database. The convolutional neural network (CNN) method was applied for extracting and predicting ore-forming anomalous information. Addressing the issue that CNNs in practical applications focus predominantly on computing short-distance local dependencies and struggle with long-distance dependencies, thereby neglecting the interrelationships of different parts globally and reducing the capability to capture global features, we innovatively integrated the Transformer method. This method effectively captures global feature representations, ensuring that both local and global key information are captured for ore-forming prediction and delineating prospective exploration areas. The results show that the CNN-Transformer model, capturing both local and global features, outperforms the CNN model with an accuracy of 0.92. The mineral prospectivity mapping by the model effectively correlate multi-disciplinary geoscientific data with known locations of deposits, significantly enhancing the precision of potential exploration areas for mineral deposits.