Abstract
Mass earth movements require reliable monitoring as well as defensible interpretation of spatial variability. This study evaluates a reproducibility-aware semivariogram workflow using two geodetic datasets with markedly different provenance: a reproducible two-epoch total-station network at the Kadzielnia landslide and archival UAV/Surfer products from a mining-affected road embankment in Bytom. The central question is not whether the two deformation mechanisms can be directly compared, but how much spatial inference remains defensible when sampling support, reference control, and data provenance differ. For Kadzielnia, 27 rock bolts generated 351 unique point pairs. Empirical semivariograms were recalculated with bootstrap confidence intervals, competing nugget, spherical, exponential, and Gaussian models, lag-bin sensitivity tests, leave-one-out kriging, simple prediction baselines, directional semivariograms, and circular statistics. Under the primary binning scheme, dh is best represented by a pure nugget, whereas dX, dY, and dr show non-unique model fits and parameter sensitivity; prediction gains over simple baselines are small. Horizontal directions are strongly concentrated (R = 0.970), but the archived network adjustment does not permit exclusion of common-mode reference-frame translation or orientation bias. For Bytom, recovered 2018 and 2019 point clouds contain more than 25 million records per epoch, but a coarse translation-only diagnostic leaves a 0.627 m centered residual RMSE, much larger than the deformation scale of interest. In addition, the exact time interval and point-level provenance of the archived semivariograms cannot be reconstructed. Consequently, Bytom is used only as an archival reproducibility example. The results show that semivariograms are useful exploratory diagnostics, but physical interpretation requires stable reference control, stationarity checks, robust binning, transparent model comparison, and reproducible data provenance.