Transformer-based three-dimensional crack detection in lava tubes: from a terrestrial analogue toward lunar subsurface exploration

<title>Abstract</title> <p>Crack detection inside lava tubes is critical for structural integrity assessment, yet conventional tunnel-inspection methods are ineffective here: lava tubes have irregular, continuously varying cross-sections and host primary volcanic textures at scales comparable to the cracks themselves. We introduce LavaCrack_PT, a Transformer-based model for 3D crack detection in unstructured point clouds, featuring a five-level multi-scale backbone and an enhanced Dice loss designed for severe class imbalance. Tested on LiDAR data from Xianren Cave (Hainan, China), it achieves a mean IoU of 0.69 and class accuracy of 0.80, surpassing CNN-, graph-, and Transformer-based baselines while maintaining robustness across terrains, crack scales, and point-cloud densities. Used as a measurement tool, the detector yields a field inventory of 893 cracks with position-dependent orientations, axis-parallel on the roof, transverse on the walls, consistent with thermoelastic stress-field predictions. This validated workflow offers a transferable basis for future Moon subsurface exploration.</p>