139 6. DISCUSSION The main value of the workflow is that it adapts computer vision to geological decision-making rather than treating mapping as a generic image-recognition task. In underground faces, the model must handle geological variability and operational clutter simultaneously. The YOLO branch is therefore not only a detector of objects; it is a quality-control mechanism that helps prevent support mesh, water glare, lamps or paint marks from being misinterpreted as lithological texture. A second advantage is the reviewer-in-the-loop design. Fully automatic mapping would be difficult to justify during a first deployment because geological labels are interpretive and may change as new drilling, sampling or structural information becomes available. A supervised workflow is more defensible: predictions are used to accelerate interpretation, while geologists validate boundaries, correct ambiguous areas and enrich the dataset for subsequent training cycles. The workflow also supports safer and more responsible mineral delivery. Images can be acquired from safer positions or by robotic systems, reducing unnecessary exposure to unstable faces. Faster interpretation can improve the timeliness of grade-control updates, while better traceability helps build confidence in the link between field observations and digital models. These contributions align with the congress theme by addressing speed, intelligence and responsibility in a single geological workflow. Several limitations must be acknowledged. First, RGB imagery alone cannot reliably solve all mineralogical discrimination problems, especially where visually similar minerals or alteration products occur. Second, model performance will depend strongly on lighting standardization and annotation quality. Third, the first pilot dataset must be expanded across more faces, lighting conditions and geological variability before the system can be considered robust at production scale. Future work should integrate multispectral or hyperspectral data, 3D point clouds and geometallurgical validation to strengthen the connection between visual domains and processing behaviour. 7. CONCLUSIONS GeoLithAI is presented as an anonymized AI-assisted workflow for lithological recognition and grade-control support in an underground Peruvian mine. The proposed method combines CNN-based semantic segmentation, YOLO-type object detection, uncertainty flagging and geologist review to generate traceable lithological-domain information from underground face images. The most defensible contribution is not the replacement of geologists, but the acceleration and standardization of short-cycle geological mapping. By separating lithological domains from non-geological occlusions and by returning ambiguous predictions to expert review, the workflow can improve repeatability while maintaining geological accountability. The pilot framework defines five initial domain classes, a 500-image first training cycle, and measurable indicators such as IoU, Dice coefficient, mAP@0.5, review acceptance and mapping-cycle time. The approach is scalable to additional sensors and mining contexts, but its current strength is a confidentially safe, technically grounded framework for underground lithological-domain recognition. ACKNOWLEDGEMENTS The author acknowledges academic mentors and mining professionals whose technical discussions on operational geology, image-based interpretation and responsible digital transformation contributed to the conceptual development of this anonymized workflow.
RkJQdWJsaXNoZXIy MTM0Mzk2