Integration of drill-core hyperspectral and geochemical data by deep learning to enhance drill-core mineral mapping

Abstract

Drill-core mineral mapping is critical in orebody modeling, but traditional analysis based on manual drill-core geological logging is time-consuming, subjective, and prone to bias. Hyperspectral (HS) imaging provides a quick, high-resolution, and non-invasive alternative; however, traditional analysis of HS data also rely on human interpretation, which limits its efficiency and reliability. Recent advances in deep learning (DL) offer ways to handle these challenges through automatic extraction of complex spatial and spectral features from HS data, eliminating bias in human interpretation. This study integrated drill-core HS image data and geochemical data (major element concentrations) with DL methods to enhance drill-core mineral mapping. The research aimed to: (i) quantify uncertainty in drill-core mineral mapping using DL approaches, (ii) evaluate and compare the performance of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), specifically long-short term memory (LSTM) models, and (iii) develop a hybrid CNN-LSTM model to enhance drill-core mineral mapping. The results demonstrated that all three models (CNN, LSTM, CNN-LSTM) had high predictive performances with low mean absolute errors (MAEs) and uncertainty. The hybrid CNN-LSTM model outperformed the stand-alone models (CNN, LSTM), having the lowest MAE (0.0321), followed by the LSTM model (MAE = 0.0358), CNN model (MAE = 0.0393), as well as artificial neural network (0.0576) which was used as shallow neural network to compare with the deep neural networks. The improved performance of the hybrid CNN-LSTM model demonstrated the benefit of integrating spatial–spectral feature extraction with sequential dependency modeling, whereas the relative strength of LSTM over CNN emphasizes the importance of sequential features.