Deep-learning classification of cave-floor surface types from LiDAR data for detailed cave mapping

High-resolution 3D mapping of subterranean environments remains challenging due to their complex geometry, low-light conditions, and restricted accessibility. Among these environments, caves represent particularly demanding settings where detailed spatial documentation is essential for monitoring processes, supporting exploration and conservation efforts. Laser scanning has become a key technique for capturing accurate and detailed 3D representations of caves that form the basis for this heritage documentation and multidisciplinary research. Despite these advances, the creation of cave maps still commonly relies on traverse-line measurements and field sketches, later digitized using specialized cave-surveying software. In recent years, LiDAR data have been used for deriving the cave extent. While this method effectively captures the general geometry of cave passages, the delineation of cave-floor units, sediments, speleothems, rock blocks, and other features remains largely manual and relies heavily on the surveyor’s interpretation. As a result, feature boundaries vary between authors, and detailed cave-surface representation lacks reproducibility that is problematic for long-term documentation. In this study, we explore the use of deep-learning semantic segmentation for classifying selected cave-floor surface types based on geometric features derived from LiDAR data. Building on previous work focused on semi-automatic cave-map generation from LiDAR point clouds, we extend the workflow from deriving cave extent and floor morphology toward the automated interpretation of surface materials and forms. The method was tested on several common cave-floor surface types, including clastic sediments, flowstone, and bedrock, as well as artificial surfaces and objects typical in showcaves. The resulting classifications show that deep-learning models can distinguish surfaces with subtle geometric differences and produce consistent, reproducible delineations of units that are traditionally mapped by hand. Compared with manual digitization, the approach reduces subjectivity and provides a scalable way to generate polygonal layers used in speleocartographic workflows.