Skip to main content
Top of the Page

Integrating Machine Learning and Data Augmentation for Automated Texture Classification in Borehole Image Logs

Borehole image logs have played an essential role throughout the evolution of hydrocarbon exploration and production. Information extracted from image log data assists in a more accurate understanding of depositional environments and the geological evolution of the area of interest. However, interpreting geological features from image logs presents challenges due to their complexity and the laborious manual work associated with interpretation procedures by experts. An approach to interpretation is to visually correlate Borehole Imaging (BHI) data with the textures of rocks from analogous fields or cores obtained from the same study field to draw the most accurate conclusions. In this regard, Machine Learning (ML) methods present a promising direction to facilitate or automate specific tasks related to image log interpretation, such as identifying textural patterns. Furthermore, data augmentation techniques, widely used in ML-related studies, could help address the limited availability and small size of image log datasets, which could substantially enhance ML model performance. Thus, this study aims to (1) employ computer vision techniques for the classification of textural patterns in resistivity image logs and (2) propose a set of appropriate data augmentation techniques for image log data and assess their impact on the computer vision models’ performance in the classification task. To enable the training of deep learning models, we analyze and define four texture classes for this case study: Laminated, Granular, Massive, and ‘Others.’ We utilized annotated data from five distinct wells, in which the resistivity image log matrix is available as Comma-Separated Values (CSV). To convert the resistivity data matrix to images, we develop and employ a data loading and preprocessing pipeline with multiple steps, such as removing invalid entries, value scaling, histogram equalization, and dividing the image logs into square tiles to properly use the image data in computer vision models. Due to the significant data requirements of the deep learning models, we propose a data augmentation pipeline, i.e., a process with a set of image transformations that we can combine to augment the dataset sizes considerably while retaining the characterization of relevant textures, thus leading to more efficient learning and accurate predictions. Then, we train and evaluate classification models from multiple architectures, namely ResNet, VGG, and Vision Transformers (ViT), to predict the most prominent texture class for any image log input region (tile). To properly validate the classification models, we separate the datasets into subsets for training and validation and evaluate the predictions’ accuracies and Jaccard coefficients. In our experimental evaluation on a small, real-world dataset of five wells, spatial and vertical shift augmentations consistently delivered the highest performance gains across the three evaluated architectures. Notably, combining spatial and vertical shift transformations boosted the overall accuracy by up to 4.7% (ResNet) and 2.9% (VGG). In contrast, a mix of blur, noise, and spatial transformations increased ViT accuracy to 76% (with a 4.5% Jaccard improvement over the baseline). These results highlight the effectiveness of domain-aligned augmentations to mitigate dataset scarcity and improve model robustness in classifying laminated, granular, and massive textures despite the challenges of class imbalance. As practical implications, employing computer vision techniques may facilitate the validation of analyses,allowing specialists to observe the results of the textural classification alongside the investigated wells.
Consequently, this may reduce the misinterpretation and the costs associated with complex interpretations.
Furthermore, data augmentation techniques could be helpful in various learning tasks in similar domains (e.g., acoustic image logs), thus aiding in advancing the adoption of ML as a valuable tool in hydrocarbon exploration and production.
Standard price: 10.00
Discounted price: 1.00
10.00
You could save: 90%
There are no records.
Back to Top
SPWLA Business Office Address
8866 Gulf Freeway, Suite 320
Houston, TX 77017

SPWLA Contact Numbers
Phone: +1 (713) 947-8727
Fax: +1 (713) 947-8747

Copyright 1959 – 2025 SPWLA

Terms of Use  |  Privacy & Cookie Policy

Our vision is to advance the science of formation evaluation through electrical, nuclear, lithological, and other well logging techniques; to promote the proper application of these techniques to exploration for and exploitation of gas, oil and other naturally occurring substances; to maintain high ethical standards and to further the interests of persons professionally engaged in these endeavors. With the support of our Oil and Gas, Service, Software and Educational Institutions we can make a difference in the industry.