A comprehensive TLS-based framework for cave ice monitoring under adverse surface conditions: application at Scărișoara Ice Cave

Abstract. Rapid degradation of cryospheric components requires reliable quantification of volumetric change to understand the leading processes and mechanisms behind the loss of ice. However, such monitoring in cave systems, a particularly complex morphological and climatic environment, remains methodologically challenging due to irregular geometry, limited accessibility, and increasingly dynamic ice-surface conditions. This study presents a comprehensive terrestrial laser scanning (TLS) data-processing framework from noise assessment, surface reconstruction, 3D volumetric change computation and error propagation as a transferable methodology for reliable monitoring of cave ice change even under suboptimal data acquisition conditions. The framework is demonstrated at Scărișoara Ice Cave (Romania) using three annual campaigns acquired with pulse-based and phase-shift scanners under conditions ranging from optimal dry and frozen surfaces to widespread meltwater presence. Scanner-specific noise assessment on different ice surface types reveals pronounced range, intensity and incidence angle dependent errors, demanding an efficient filtering strategy that reduces point cloud dispersion. A hybrid Poisson–MeshFix reconstruction strategy effectively fills meltwater-induced data gaps, reducing modelling error by up to five times. Multiresolution modelling tests show that 5 cm resolution provides an optimal balance between volumetric precision and computational efficiency with errors under 1 % compared to finer resolution models. Comparison of conventionally applied 2.5D change detection and 3D approach reveals ~20 % volumetric discrepancy, confirming that full 3D analysis is essential in such geometrically complex settings. The results reveal a cumulative ice loss of 1521 ± 65 m³ over three years, with heterogeneous spatial patterns controlled by cave morphology and drip-water distribution. Our data clearly shows that omitting any of the evaluated processing steps would greatly bias the volume estimates.