mineral-specific classification is more robust when multispectral or hyperspectral sensors are incorporated. For this reason, the present paper focuses on lithological-domain recognition and identifies spectral sensing as a future extension rather than as a claim of completed mineralogical automation. Table 1 - Initial lithological-domain classes for the anonymized pilot Domain class Geological meaning Initial images Main annotation type Carbonatedominated zone Pale carbonate host rock; relatively uniform light texture. 100 Segmentation mask Skarn-rich zone Dark calc-silicate texture with high visual contrast. 100 Segmentation mask Mixed transition zone Diffuse contact where visual textures overlap. 100 Mask + contact guidance Brecciated zone Fragmented rock with irregular clast or matrix texture. 100 Mask + boxes Altered or wet reflective zone Wet or altered surface with reflective artifacts. 100 Mask + boxes 4. MATERIALS AND METHODS 4.1 Image acquisition and metadata protocol The acquisition protocol is designed for operational repeatability. Images or video frames should be captured after access authorization and, where applicable, from a safe distance using handheld cameras, pole-mounted systems, drones or robotic platforms. The protocol requires overlapping coverage, consistent camera orientation where possible, and multiple exposures when wet surfaces or lamps generate glare. Original files are retained as immutable records, while normalized copies are generated for training and inference. Each image must be linked to metadata, including level, working area, date, approximate camera position, viewing direction, lighting condition and the name or code of the geologist who validated the annotation. These metadata are essential for traceability and for future integration with 3D geological models. Where LiDAR or photogrammetric geometry is available, it should be used as a spatial scaffold for georeferencing, while the lithological classification remains driven by supervised image analysis. Table 2 - Recommended data protocol for a first training cycle Stage Minimum requirement Reason for inclusion Acquisition Overlapping RGB images or video frames with consistent standoff distance where possible. Reduces geometric and illumination variability. Metadata Level, chainage or local reference, date, camera position and lighting condition. Allows traceability between image, location and interpretation. Pre-processing Resize, white-balance normalization, contrast control and retention of original images. Creates a stable training copy without losing the raw record.
RkJQdWJsaXNoZXIy MTM0Mzk2