Domain Adversarial Neural Network for Mapping Mineral Prospectivity

Various deep learning algorithms have been employed for mineral prospectivity mapping (MPM) owing to their powerful capacity for automatic extraction of high-level representations from multisource data. However, extending the application of deep neural networks to areas with different ore-forming characteristics remains challenging because it requires integrating the geological knowledge learned in one area with that of other areas. Unsupervised domain adaptation is a major strategy in transfer learning that can address the problem of poor network generalization caused by domain shifts. In this study, an improved unsupervised domain adversarial adaptation network driven by synthetic data was constructed for MPM. Ample synthetic datasets were used to build the training set, which provided a solid foundation for the network training. An improved loss function embedding the maximum mean discrepancy was designed to achieve interdomain feature alignment in an unsupervised scenario. With relatively sufficient data, northwestern Hubei Province, China, was chosen as the source area, whereas southeastern Hubei Province was selected as the target area for gold (Au) polymetallic mineralization. Comparative experimental results demonstrated that the constructed neural network exhibited better performance and domain adaptation capability in the target domain than baseline methods. The delineated high-potential zones displayed strong spatial correlations with known Au polymetallic mineralization in the target area, thus providing critical clues for prospective mineral exploration.