geological textures and contacts. These factors increase the variability of interpretation, reduce traceability between field observations and digital models, and may delay the incorporation of new geological information into planning workflows. GeoLithAI is used in this paper as a research name for an AI-assisted workflow, not as a commercial product. The proposed contribution is a supervised computer-vision approach designed to support geologists in recognizing lithological domains from underground face images. No company names, mine names, project names or confidential production data are disclosed. The study setting is described only as an anonymized underground polymetallic mine in the central Peruvian Andes. 2. STATE OF THE ART AND INNOVATION GAP Convolutional neural networks have become a standard method for image-based classification because they learn hierarchical representations of edges, textures and spatial patterns from labelled data (Krizhevsky et al., 2012; Goodfellow et al., 2016). For geological applications, this capability is relevant because lithological classes often differ through combinations of colour, grain size, texture, mineral fabric, veining, brecciation and alteration patterns. However, mine-face mapping is more complex than controlled image classification because the image contains geological objects, non-geological occlusions and variable illumination in the same scene. Semantic segmentation models such as U-Net are suitable for lithological domain delineation because they produce pixel-level class maps and can be trained with expert polygons (Ronneberger et al., 2015). Object detection and instance segmentation methods such as Mask RCNN and YOLO-type models are useful when the task requires locating specific objects or artifacts within a cluttered image (He et al., 2017; Bochkovskiy et al., 2020). In the proposed workflow, these methods are combined: the segmentation branch maps lithological domains, while the detection branch identifies mesh, water reflections, paint, support elements and fragmented material that can bias the domain prediction. Digital outcrop modelling, photogrammetry and LiDAR have also influenced geological mapping by providing safer and more reproducible geometric context for inaccessible rock exposures (Bellian et al., 2005; Webster et al., 2006). In underground environments, spatial data can support georeferencing and 3D integration, but it does not eliminate the need for geological interpretation. The innovation gap addressed here is therefore not only image classification, but a complete, reviewer-in-the-loop workflow that converts underground imagery into traceable lithological-domain information suitable for short-cycle model updates. 3. ANONYMIZED GEOLOGICAL SETTING AND SCOPE The anonymized case is framed as an underground polymetallic mine in the central Peruvian Andes. This regional wording is intentionally broad and is used only to define a plausible geological environment. The central Peruvian Andes include extensive polymetallic mineral systems with carbonate host rocks, skarn and replacement associations, hydrothermal alteration and structurally controlled mineralization (Soler and Bonhomme, 1988; Noble and McKee, 1999; Bissig et al., 2008). The proposed lithological classes are therefore expressed as general visual domains rather than deposit-specific units. The scope of the pilot is limited to AI-assisted recognition of visually separable lithological domains in underground face photographs. Mineralogical interpretation from RGB imagery is treated cautiously: RGB images may support domain recognition and textural discrimination, but
RkJQdWJsaXNoZXIy MTM0Mzk2