PATSA-BIL: Pipeline for automated texture and structure analysis of borehole image logs

Abstract

High-resolution image logs are essential for understanding rock compositions and geological features, but their analysis requires expertise and can be tedious and error-prone. This manuscript presents PATSA-BIL, a novel data-driven pipeline that uses Artificial Intelligence (AI) and computer vision techniques to enhance the analysis of resistivity image log data in the oil industry. PATSA-BIL enables deep-learning-based classification of resistivity textures and automatic segmentation of geological structures in borehole images. In addition to texture classification, PATSA-BIL provides of a series of chained Machine Learning (ML) tasks that comprise Simple Linear Iterative Clustering (SLIC) superpixels, radiomics embeddings, and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for geological structure segmentation. Experiments with the proposed pipeline, including comparisons against five fine-tuned, deep-learning models (Mask R-CNN, Faster R-CNN, RetinaNet, DETR, and DINOv2), demonstrate that PATSA-BIL achieves up to 90% accuracy in detecting resistivity textures. Furthermore, PATSA-BIL’s segmentation chain outperforms the closest competitor (DINOv2) by achieving a 16% higher Mean Average Precision (mAP), significantly improving the clustering of borehole structures by up to 7% while reducing computational demands compared to other approaches. Statistical validation further confirms the robustness of the proposed approach, performed using a Friedman test ($p$-value = $6.7\cdot 10^{-23}$, vs. DINOv2) and a Wilcoxon–Holm post-hoc comparison between the top models ($p$-value = $1.4\cdot 10^{-3}$, vs. DINOv2) with very large effect size.