Advanced Quality Control Workflow for Full-Size Core Images: Enhancing Machine Learning Training for Geological Applications

Machine learning (ML) has shown substantial promise in automating geological core description, supporting tasks such as lithology classification, porosity estimation, and texture prediction. However, the performance and reliability of ML models are heavily dependent on the quality and consistency of the input data. Core photographs, often collected under varied field conditions, frequently suffer from distortions such as color tint shifts, crushed or missing intervals, and the presence of non-geological visual artifacts like plastic wraps or labels. These issues reduce model accuracy and generalizability, particularly when applied across different fields or data collection settings. This paper presents an advanced quality control (QC) workflow specifically designed for preprocessing full-size geological core images before their use in ML-based interpretation workflows. The system comprises three modular components: a Random Forest classifier for detecting tinted images based on RGB histogram statistics; a U-Net-based convolutional neural network (CNN) for segmenting crushed core intervals; and a MobileNetV2 CNN for identifying non-core elements. The workflow was trained and validated on available industrial datasets and demonstrated improvements in lithology prediction performance. The QC system is fully modular, customizable, and suitable for integration into broader digital subsurface platforms. It enables scalable, reproducible filtering and standardization of geological imagery at field-wide scale, reducing manual review and improving downstream ML performance. This work introduces the novel fully documented and operational end-to-end ML-based QC solution tailored for full-size core images. This approach supports robust AI deployment in digital reservoir characterization, particularly in multi-well, multi-field contexts where data variability is high.