Ancient-terrace point clouds fail across regions: MELD corrects each feature level with unlabeled target data, lifting F1 from 0.397 to 0.459
Synopsis
Between the Teyuna and Miraflores lidar surveys in the Sierra Nevada de Santa Marta, Colombia, terrace detectors trained in one region degrade in the other because terrain, vegetation, acquisition settings and terrace morphology differ. MELD trains a separate diffusion denoiser on unlabeled target features at each of four PointNet++ levels, then uses these priors to refine the multi-scale features fed to the segmentation decoder during training and inference. F1 rises from 0.397 to 0.459 for Teyuna-to-Miraflores transfer and from 0.557 to 0.594 in reverse, the highest among the compared adaptation methods in both directions; adding KNN local-geometry descriptors raises these to 0.542 and 0.606.
Interpretation
A terrace detector can adapt to a new survey using only that survey's unlabeled point cloud, without first annotating terraces in the new region. The prior practice was to train in one survey and apply the model directly to another, where differences in terrain, vegetation, acquisition configuration, point density and terrace morphology degrade a model that performs well in the source region. In bidirectional transfer between Teyuna and Miraflores, the source-only PointNet++ model reaches F1 of 0.397 (Teyuna to Miraflores) and 0.557 (reverse), and MELD raises these to 0.459 and 0.594.
Adaptation acts on the multi-scale features the segmentation network actually consumes, rather than only shrinking an aggregate gap between source and target distributions. Statistical alignment, adversarial alignment and prediction-based regularization each handle one aggregate discrepancy or output confidence, and when background points dominate and terrace points are sparse and varied, that global objective is easily dominated by background. Against the source-only model and nine established domain-adaptation configurations, MELD attains the highest F1 in both transfer directions; for Teyuna to Miraflores the strongest competitor, single-level MMD/DAN, remains below MELD, and in reverse the strongest is multi-level DANN.
Each PointNet++ level gets its own diffusion denoiser, which learns that level's target-domain distribution prior and then corrects source and target features along short reverse-diffusion trajectories. Earlier diffusion-based domain adaptation operated at the input, on representations from pretrained generative models, or at a single learned feature stage, without modeling target distributions separately at the multiple levels a segmentation backbone produces. A level ablation shows all four levels active gives the strongest result in both directions; refining only the deepest level yields a small gain, most clearly for Miraflores to Teyuna.
Adding surface normals, slope and aspect estimated from K-nearest-neighbour neighborhoods further improves cross-region terrace detection beyond diffusion adaptation. With learned hierarchical features alone, the local surface orientation and slope breaks that define terraces are not explicitly represented. With these geometric descriptors, F1 rises from 0.459 to 0.542 for Teyuna to Miraflores and from 0.594 to 0.606 in reverse, the strongest overall results in both directions.
Perspective
Evaluation is confined to the Teyuna and Miraflores airborne and UAV lidar surveys in the Sierra Nevada de Santa Marta, Colombia, on a binary terrace versus non-terrace task, under a transductive setting in which the unlabeled target tiles used for adaptation are the same tiles later evaluated point by point.
The transductive setting does not show whether the adapted model generalizes to target samples unseen during adaptation; a gap to fully supervised in-domain training remains when target labels are plentiful; and the study considers terraces as the only archaeological feature class.
