Mineral prospectivity mapping for multi-source geoscience data: A novel unsupervised deep learning method

Mineral prospectivity mapping provides important information on the distribution of potential mineral resources, which is helpful for formulating reasonable resource development strategies and is an important step in mineral exploration. Recent methods mainly focus on deep learning methods, which can directly learn and extract information from relevant data through neural networks and output map identifying potential areas of minerals. Therefore, this study has developed an unsupervised deep learning method for mineral prospectivity mapping based on vision Transformer. This method is trained in an unsupervised way, without the need for labels and additional manpower. It takes multi-source geoscience data as input data. Multi-source data fusion convolution layer fuses the feature information among the input data. The image characteristics are mined through vision Transformer to provide geochemical anomaly and geological constraint information for samples. This study predicts chromite deposits in the Heishantou area of Balikun, Xinjiang, China for demonstration. Seven comparative case studies were conducted from both visual and quantitative perspectives, all of which demonstrated the superiority of this method. The reliability of this method was further verified through multiple experiments from different perspectives.