Applications of data augmentation in mineral prospectivity prediction based on convolutional neural networks

The supervised deep learning methods applied in mineral prospectivity mapping usually need sufficient samples for training models. However, mineralization is a rare event. Insufficient known mineral deposits cannot meet the sample requirement of supervised learning methods, resulting in lower predictive accuracies and poor generalization abilities. For the purpose of solving this issue, this paper adopted a data augmentation method to make mineral prospectivity prediction of gold deposit in the Fengxian Region, China. This data augmentation method adopted cropping operations to generate sufficient training samples without changing spatial directions of geological data. Meanwhile, this paper utilized the continuous buffer distance method to quantify faults and anticline axes, overcoming the loss of geological information caused by using the discrete buffer distance mode. To prove the effectiveness of the data augmentation method, this paper utilized three different convolutional neural networks (LeNet, AlexNet, and VggNet) to extract relationships between multisource ore-indicating factors and mineral deposits. In addition, this paper discussed effects of different parameters on predictive performances. According to series of comparisons, the LeNet model outperformed other models, achieving superior values of accuracy (91.38%), Kappa coefficient (0.8119), and AUC (0.958). Moreover, the LeNet model successfully caught 81.8% of known gold deposits within 18.6% of the study area. The delineated high potential areas offer intuitive guides for exploring more gold deposits in the Fengxian region. The proposed data augmentation method is available for mineral prospectivity modeling by supervised deep learning methods for the areas of lower exploration degrees. For mineral prospectivity modeling based on convolutional neural networks, utilizing the continuous buffer distance to transform faults and anticline axes into predictor variables of the image form is conducive to improve the predictive performance than utilizing the discrete buffer distance.