AI-Assisted 3D Rock Fracture Characterization from Multi-View Smartphone Image–LiDAR Data in Underground Environments

Three-dimensional characterization of rock fractures in underground environments is challenging because restricted access, uneven illumination and partial visibility hinder mapping and structural measurements. This study presents an AI-assisted workflow that combines text-prompted Segment Anything Model 3 segmentation with smartphone image–LiDAR reconstruction to derive three-dimensional fracture supports. Candidate masks are combined into graded image evidence, transferred to the reconstructed surface and consolidated across views. Assessment combines manual image annotations, terrestrial laser scanning (TLS) surface and fracture references, field-compass measurements and conditional bootstrap analysis. On 25 annotated images, pixel-level precision and recall were 90.4% and 73.5%. Following alignment to TLS, block-level median absolute M3C2 surface-normal distances were 2.2–5.0 mm, while independent three-dimensional checkpoint RMSEs were 1.10–2.55 cm. For four fractures in a calcarenite chamber, full-image no-threshold supports covered 83.7–100% of the independent TLS traces within 5 cm. TLS plane discrepancies ranged from 0.3–32.7∘ without thresholding to 0.6–8.0∘ after thresholding, with fracture-specific coverage trade-offs. Controlled aggregation comparisons showed no general advantage of weighting, while reduced-image characterization decreased coverage for all four fractures. The workflow links image evidence to independently assessed fracture geometry and orientation. Its structural reliability depends on the spatial distribution of the recovered support, motivating joint evaluation of recovery, plane agreement and resampling stability.