Laser scans flown over forests can strip away the trees and reveal the bare ground at half-meter resolution, exposing old canals, mounds and mining pits. Deep segmentation models trained on such scans work well inside the landscape they were trained on. They have no reliable way to move to a new landscape, where both the ordinary background and the ancient structures look different. This note proposes a cheap, label-free correction borrowed from text retrieval. First, describe each small patch of terrain as a "word" from a learned dictionary of landforms. Then, in every new region, count how common each landform word is and down-weight the ubiquitous ones before the segmentation network sees them. This is the inverse-document-frequency half of term frequency–inverse document frequency (TF-IDF), the classic weighting that suppresses words like "the" and boosts rare, informative ones. The bet is that the landforms that are everywhere in a region, such as paddy-field texture or uniform slopes, are the background cues a model latches onto in one landscape and that fail to carry over to the next.
The argument is shaped by an analogy from genetic epidemiology, and it should be read as an analogy. In Mendelian randomization, genetic variants that track a causal variant in European cohorts often stop tracking it in other ancestries, because the background correlation structure differs. The field's response is to re-estimate that background per population and to check whether an effect replicates across ancestries (Zhang et al.; Zhu et al.). We take two things from that practice. One is per-population recalibration, which here means recomputing the landform frequencies on each target region. The other is cross-environment replication as the test of whether a learned cue is real. Architecture enters as the place where we expect the effect to differ. Flat vision transformers, which let every patch attend to every other patch from the first layer, integrate more region-wide context than convolutional networks do. We therefore predict they gain the most from suppressing region-common background.
Core Claim
Recomputing landform rarity on the unlabeled target region and using it to re-weight the input should improve cross-landscape segmentation of archaeological features. The gain should beyond what per-region intensity normalization, per-region re-estimation of normalization-layer statistics, or class-rebalanced losses provide. The gain should also be largest for architectures that pool global context. For a region $r$ with tiles $D_r$, a landform token $t$ from a fixed dictionary, and a tile $d \in D_r$, the weight is
$$w_r(t,d) = \mathrm{tf}(t,d)\,\log\frac{|D_r|}{1+|\{d' \in D_r : t \in d'\}|},$$
where $\mathrm{tf}(t,d)$ is the fraction of pixels in tile $d$ assigned to token $t$. The second factor is computed only from unlabeled scans of region $r$. This is the property that separates the proposal from loss reweighting: loss reweighting needs labels and is fixed at training time.
The claim is falsifiable in one specific way. If re-estimating normalization-layer statistics on the target region matches the rarity weighting, then the landform-word view adds nothing beyond moment matching, and the method should be dropped in favor of that simpler approach.
Research Program
The dictionary is built from multi-scale local relief descriptors computed on bare-earth terrain models. These include depression and elevation measures that do not depend on illumination angle, of the kind used in Saxon placer-mining surveys (Schröder et al., Part 1). Each descriptor vector is clustered into $K$ tokens with k-means, fit once on pooled unlabeled terrain, and $K$ is swept from 64 to 1024. Each pixel gets a token, and each region gets its own rarity table. The weighted map $w_r$ enters the network in one of two ways. It can be stacked as an extra input channel beside elevation, or it can gate early features multiplicatively. Both variants are run on every backbone.
The architectural lens uses the backbone families already benchmarked on Archaeoscape, which evaluates convolutional networks (U-Net, DeepLabv3), flat transformers (ViT and hybrids) and hierarchical transformers (Swin, PVTv2) on 888 km² of Khmer terrain. That paper found that elevation drives detection, that RGB-only models fall under 30% mIoU, and that flat ViTs trail U-Net. It motivates our architecture prediction but tests nothing across regions. We hold the training recipe fixed and ask whether the cross-region gap, and the gain from rarity weighting, ordered by how global each architecture's context is.
The replication diagnostic is the second component taken from Mendelian randomization practice. For each token, we estimate its association with each label class in the source region and again in a held-out region. A token whose association does not replicate plays the role of a decoupled tag variant. We test whether the model's attribution mass on non-replicating tokens predicts its transfer loss, and whether rarity weighting reduces that mass. Invariant Ancestry Search searches for minimally invariant predictor sets across environments and takes their union. It suggests a further analysis: treat the set of tokens that replicate across held-out environments as the stable cues, rather than requiring invariance in every environment at once. This is a secondary analysis, not the main method.
Baselines
The naive reference point is each backbone trained on the source environment with the standard Archaeoscape recipe and applied unchanged to the target. The strongest in-domain numbers on Lacuna are PVTv2-b1 at 52.1 mIoU and 88.7% overall accuracy, and U-Net with 512-pixel inputs at 52.8 mIoU (Archaeoscape). We re-run both configurations exactly, so the in-domain ceiling and the transfer drop are measured on our splits. No reported result was found on Lacuna for any cross-region or cross-culture transfer on terrain data.
The strongest simple competitor is per-target re-estimation of normalization-layer statistics on unlabeled target tiles, together with per-region z-scoring of elevation. Like our method, it is label-free and per-region, so it is the comparison the claim depends on. A second baseline adds the relief-visualization descriptors directly as input channels without tokenizing or weighting them. This isolates whether the gain comes from better input features or from rarity. A third trains with inverse-class-frequency loss weights, which tests whether "rarity" just means class imbalance.
The strongest existing adaptation method is Adaptive Teacher. It combines adversarial feature alignment through gradient reversal with teacher–student pseudo-labeling, and it reports 50.9% mAP on Foggy Cityscapes, 49.3% on Clipart1k and 59.9% on Watercolor2k, all on convolutional detectors. We port its alignment and mutual-learning losses to segmentation on each backbone with the same target data and the same tuning budget. On Archaeoscape, we also compare against the prototype-consistency method of Xu et al., which reports 70.06% mIoU on PASCAL VOC with 1/16 of the labels and is tested only on convolutional networks.
The key ablations separate the ingredients of the method. Source-computed rarity applied to the target removes per-region recalibration. Term frequency alone removes rarity. A random dictionary of the same size removes landform semantics. Rarity weighting stacked on top of Adaptive Teacher tests whether the two are complementary.
Every method gets the same backbone weights, the same training epochs, and a 20-run hyperparameter search on a source-only validation split. Target labels are never used for tuning. Beating a baseline means a mean target mIoU gain of at least 2 points over three seeds, with non-overlapping bootstrap intervals over target tiles. The gain must hold on at least two of the three backbone families.
Why not just…?
Why not just re-estimate normalization-layer statistics on the target region? This is the most serious threat, because it is also label-free and per-region. It differs in that it shifts and rescales whole feature channels, whereas rarity weighting suppresses specific common landforms and leaves rare ones alone. Whether that difference matters is an empirical question. If the two tie on every backbone, we will report that and abandon the dictionary.
Why not just reweight the loss toward rare classes? Class reweighting needs labels and is fixed at training time. It cannot respond to a target region where, for example, ponds are everywhere and so stop being informative. The inverse-class-frequency baseline and the "source-computed rarity" ablation together show whether region-specific recomputation is what helps.
Why not just run off-the-shelf domain adaptation? Adaptive Teacher is the strongest such method on Lacuna. It was developed for weather and style shifts on convolutional backbones, where the objects themselves stay the same. Its stability on transformer backbones and under shifts in what the structures look like is untested. We run it directly; it may well win on convolutional networks. The proposal survives if rarity weighting adds to it or helps where it destabilizes.
Why not just standardize on PVTv2? Its lead on Archaeoscape is small, and U-Net with more context overtakes it (52.8 against 52.1 mIoU). Neither number says anything about transfer. Our architecture comparison is what would justify, or undercut, picking one backbone.
Proposed Experiments
The first experiment uses within-Khmer environments. Coordinates are stripped from Archaeoscape, so we cluster tiles by unlabeled terrain statistics into landscape types, such as low floodplain versus upland, and hold out whole clusters as target "populations." Within one culture, this measures shifts in background landscape.
The second is a harder test across cultures. We will annotate a small set of placer pits, sunken roads and foundations on the published Saxon terrain models from Sauschwemme (Schubert et al.) and Hohenwalde (Faule Pfütze). We will map classes to the closest Archaeoscape categories (mound and hydrology) and report per-class IoU. Annotator disagreement will be recorded, because Archaeoscape itself notes ambiguous boundaries on earthworks.
Three outcomes decide the question. First, does target-computed rarity beat both normalization re-estimation and source-computed rarity on held-out clusters and on Saxony? Second, is the gain ordered flat ViT > hierarchical transformer > CNN, as the context argument predicts? Third, does the model's attribution mass on non-replicating tokens correlate with transfer loss across backbones and seeds, and does weighting reduce it? A positive answer to the first with no architectural ordering would still yield a useful tool. A negative answer to the first ends the method.
References
- Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- Early Placer Mining in Saxony. Part 1: Methodological Principles of Archaeological Exploration, Interpretation and Documentation
- Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era
- Invariant Ancestry Search
- Cross-Domain Adaptive Teacher for Object Detection
- Semi-supervised Semantic Segmentation with Prototype-based Consistency Regularization
- Early Placer Mining in Saxony. Part 2: Archaeological Investigations at the Sauschwemme – New Evidence for Tin Placer Mining during the Bronze Age, Hallstatt Period, and High Middle Ages
- Fortification, mining, and charcoal production: landscape history at the abandoned medieval settlement of Hohenwalde at the Faule Pfütze (Saxony, Eastern Ore Mountains)