Transfer Learning for Core Image-Based Classification in Diverse Geological Settings with Various Deep Learning Models

Core samples provide valuable geological information for the characterization of subsurface deposits and classification of samples into categories (i.e., lithology, mineralogy, alteration, etc.). Modern core image classification often utilizes machine learning methods, including convolutional neural networks (CNNs) and transformers, to extract information from core samples; however, learning across different geological settings is rarely considered. The present study leverages training on core samples from different deposits to enhance lithology classification with transfer learning (TL) and model fine-tuning. Heterogeneous TL is implemented by transferring information from multiple source datasets to a target dataset of interest. The pretraining process considers the use of four deep learning models: (1) residual network with 50 layers (ResNet50), (2) vision transformer (ViT), (3) attention fusion (AF), which relies on self-attention between feature datasets from ResNet50 and ViT, and (4) simple concatenation (SC), which concatenates ResNet50 and ViT features. Two TL strategies are considered to improve model performance: (1) transferring pretrained weight matrices, and (2) transferring feature matrices. The performance of these four models with different TL strategies is analyzed with five geographically and geologically diverse core image datasets: (1) a South Australian copper deposit, (2) a Canadian gold deposit, (3) a Canadian oil-sands deposit, (4) a US limestone reservoir, and (5) a North-East Australian and Maldivian carbonate deposit. The concatenation hybrid model with weight matrix-based TL considering a minimum sample count per category achieves the highest F1 score; accuracy improvement ranges between 10% and 50% for the evaluated datasets compared with using the base model alone. The results underscore that TL can improve core image classification even when the source datasets are geologically dissimilar from the target dataset. Overall, the findings demonstrate the value of TL across varied geological settings and under both data-limited and data-rich conditions.