Abstract

Characterizing fractures and bedding from borehole image logs is crucial for reservoir analysis. The conventional method for this is a time-consuming manual process where interpreters pick minimum of 3 points on each feature’s location on the borehole image, which are then used for curve fitting to determine the dip and azimuth. While recent machine learning methods have used segmentation approaces, including YOLO, Mask-RCNN, and Faster RCNN to assist this process, they remain semiautomated and do not eliminate the manual picking step. This paper introduces a novel, fullyautomated end-to-end workflow that directly mimics and replaces the manual interpretation process. We adapt previously developed GeoBFDT model, shifting its objective from direct regression of physical parameters (depth, dip and azimuth value of each feature) to predicting the keypoints associated with each sinusoid of different geological feature. These predicted keypoints are then used to mathematically reconstruct the dip and azimuth values via curve fitting. This keypoint-guided approach forces the model to learn the feature's sinusoidal pattern, leading to improvement in classification accuracy over the direct regression method. Our results show that the classification accuracy for bedding increased from 63% to 76%, while fracture accuracy showed slight improvement. The model achieves an F1-score of 76% for beddings and 65% for fractures at a 3 cm depth threshold. By using the keypoint learning process, our method provides an end-to-end solution for geological feature characterization, overcoming the limitations of both manual and semi-automated techniques