Non-destructive, AI-based Rock Core Characterization for Automated RQD Assessment in Mineral Exploration

Accurate and efficient rock mass characterization is crucial for achieving sustainable mineral exploration and resource evaluation, especially in the context of increasing global resource scarcity and the urgent need to reduce environmental and operational costs. The Rock Quality Designation (RQD) is a widely used indicator for assessing rock mass integrity in geological and geotechnical engineering. However, conventional RQD determination relies heavily on manual measurements of drill cores, which suffer from low efficiency, poor scalability, and limited integration into data-driven exploration workflows.

To address these limitations, this study proposes an automated approach for RQD computation of drill cores based on computer vision and deep learning. The method integrates image-based sensing with advanced object detection and image segmentation algorithms to achieve non-destructive and automated characterization of drill cores.

First, perspective correction is applied to field-acquired core images to ensure geometric consistency. The principle of perspective correction is to project the two-dimensional original image into a three-dimensional viewing space and then transform the three-dimensional space to the image processing plane. The formulas are as follows:

The 3D viewing space is then mapped to the image processing plane using:

Subsequently, the Segment Anything Model (SAM) is employed to automatically detect and extract core regions based on the similarity of color and texture features. In SAM, the prompt encoder partitions and encodes the image based on object color, texture, and other features using:

On this basis, a YOLOv8-based image segmentation model is constructed to identify gap features between core pieces, enabling precise segmentation of individual core segments. YOLOv8 selects positive samples using the TaskAlignedAssigner strategy, formulated as:

Furthermore, by establishing a mapping between image pixels and physical dimensions, the lengths of core pieces are automatically quantified, enabling RQD computation as follows:

Studies on practical cases indicate that this approach maintains high computational accuracy while significantly improving processing efficiency, highlighting its potential as an AI-driven tool for automated core characterization. This method provides a scalable, non-destructive, and efficient technique for digital and data-driven mineral exploration workflows, supporting more sustainable and scientifically informed decision-making in mineral exploration and resource evaluation.