Deep GMDH Neural Networks for Predictive Mapping of Mineral Prospectivity in Terrains Hosting Few but Large Mineral Deposits

There has been in recent years a trend towards adopting deep neural networks for addressing earth science problems. Of the various deep neural networks applied to different problems in earth sciences, this study aimed to demonstrate how to apply the group method of data handling (GMDH) neural networks for mineral prospectivity mapping (MPM). GMDH neural networks are sophisticated, multilayered, robust tools for addressing complex regression problems. However, labeled data for MPM, which constitute multivariate attributes of known deposit and non-deposit sites, are often (if not always) insufficient to train neural networks adequately. Such issue triggers networks' poor generalization; that is, the networks developed fit the labeled data perfectly (i.e., low bias) but cannot predict unseen data accurately (i.e., high variance). Given this, GMDH neural networks were, in this study, coupled with a window-based data augmentation technique, an approach to generate additional geologically constrained labeled samples for MPM, and applied to a district hosting few but giant porphyry copper deposits. It was recognized that coupling the data augmentation technique with GMDH neural networks yielded robust predictive models that can handle the bias-variance tradeoff, making this combined methodology a viable option for MPM in terrains that host few but large mineral deposits.

Automated summary

In recent years, there has been a trend towards utilizing deep neural networks in earth science research, with a focus on solving complex regression problems. However, the lack of labeled data for mineral prospectivity mapping (MPM) often leads to poor generalization of neural networks. This study demonstrates that coupling the group method of data handling (GMDH) neural networks with a window-based data augmentation technique can generate robust predictive models for MPM in terrains with few but large mineral deposits, effectively addressing the bias-variance tradeoff.