GRSG Conference 2022: Orbit to Outcrop

Title: Deep learning for mineral exploration: New neural network approach for classification of hydrothermal alteration using PRISMA spaceborne hyperspectral data

Author: Adrian Ctvrtnicek


Deep learning has become increasingly powerful due to the rise in data volume and computational power available. However, the applications of neural networks are still very much unexplored in mineral exploration and remote sensing. Moreover, the advent of hyperspectral satellites in recent years has created an urgent need to find a new approach to learning high-dimensional patterns from increasingly complex data. Thus, this project aims to investigate the potential use of artificial intelligence for satellite mapping and image spectroscopy that might transform the future of geosciences.

The machine and deep learning algorithms will be trained using PRISMA satellite hyperspectral data with supreme spectral resolution (239 bands) and a ground-truth alteration map from the Quellaveco porphyry deposit (S Peru, provided by Anglo American). Subsequently, the trained models could be applied to other porphyries around the world to classify similar alteration signatures in a tiny fraction of the time required for traditional field-based mapping.

One of the most important research outcomes is comparing the accuracy and efficiency of different machine and deep learning models to examine their benefits or disadvantages. These methods include simple machine learning models (e.g., support vector machines, random forests) and deep learning (from artificial neural networks to computer vision). Equally, testing the trained models on unseen datasets from similar localities will be crucial to prove if the concept works in various geological environments while avoiding common problems such as overfitting or underfitting. Additionally, the project also assesses if dimensionality reduction methods could alleviate potential problems including overfitting and computational cost.

Lastly, many industries are already using AI as a vital tool for their decision-making or product recommendations. So, this project is one of the first to capitalise on the ever-growing complexity of remote sensing data by applying state-of-the-art deep learning algorithms to extract geologically significant information.