Skip to content
Daniel Wamriew
Research
PublishedScientific Machine Learning

A Deep Learning Approach for the Automated Extraction of Natural Fracture Traces from Two Dimensional Orthoimages Derived from Digital Outcrop Models: A Study Case from Jebel Madar Anticline in Oman

A. E. Sanchez Garcia, D. Wamriew, R. Iakusheva, Z. Tariq, P. Richard, N. Menegoni, V. Vahrenkamp, B. Yan, T. Finkbeiner

ARMA/DGS/SEG International Geomechanics Symposium 2025 (ARMA-IGS-2025-0341) · 2025

Abstract

Outcrop studies help in the understanding of analog reservoir characteristics, such as their natural fracture systems. In recent years, Digital Outcrop Models (DOMs) have increased the availability of geological data. However, the interpretation of natural fracture traces from DOMs continues to be a manual and time intensive process. To change this, we propose a deep learning approach that automates natural fracture trace extraction from high-resolution DOM-derived 2D orthoimages while preserving mapping accuracy compared to manual interpretation. We applied a 2D U-Net model to a geological pavement of the Shuaiba Formation outcropping in the southern part of Jebel Madar, a salt-cored anticline in Oman. The model achieved strong performance, correctly identifying 90% of natural fracture pixels, with the following evaluation metrics: Intersection over Union (IoU) of 0.8229, Precision of 0.9048, Recall of 0.9009, F1 Score of 0.9029, False Positive Rate (FPR) of 0.00049, and a Precision–Recall Area Under the Curve (PR-AUC) of 0.9678. We observe model limitations in terms of fracture network connectivity and minor spatial misplacement of natural fracture traces. This work demonstrates the potential of combining high-resolution DOM-derived 2D orthoimages with deep learning techniques with the aim to accurately extract natural fracture traces.