Mapping prospectivity for regolith-hosted REE deposits via convolutional neural network with generative adversarial network augmented data

Abstract

The regolith-hosted rare earth elements (REE) deposits are the dominant source of the global heavy REE resources. This study proposed a convolutional neural network (CNN) architecture to integrate the multi-source data (e.g., geological, geochemical and geomorphological data) to map mineral prospectivity of regolith-hosted REE deposits. To solve the lack of labelled data for the training of the supervised CNN, a generative adversarial network (GAN) was applied for data augmentation. The proposed GAN is trained in an unsupervised way, which regards the random downscaled real samples as the inputs. The GAN-based augmented data is validated by comparison with real data and the peak signal-to-noise ratio value, respectively. A case study of mapping prospectivity for regolith-hosted rare earth elements deposits in Southern Jiangxi Province of China further illustrates and validates the procedure. The final mineral prospectivity map is obtained by the CNN with all geological, geochemical and geomorphological data augmented by GAN. The CNN reaches 99.7% training accuracy and 98.9% validation accuracy. All the known mineral deposits are located in the prospective areas delineated by the CNN, which only occupy 2.36% of the study area. The obtained results indicate that the proposed framework, integrating CNN and the GAN-based augmentation method is an effective way for the application of the supervised deep learning algorithms in mineral prospectivity mapping of the regolith-hosted REE deposits and the prospective areas could be used for guiding further exploration in the study area.