Needles in the Landscape: Semi-Supervised Pseudolabeling for Archaeological Site Discovery under Label Scarcity

Technical Summary: Needles in the Landscape: Semi-Supervised Pseudolabeling for Archaeological Site Discovery under Label Scarcity

Problem Statement
Archaeological Predictive Modeling (APM) faces a fundamental challenge: archaeological sites are extremely rare (<0.004% of survey areas) and sparsely distributed across heterogeneous landscapes. Traditional supervised learning fails because the vast majority of the landscape is unlabeled rather than confirmed negative; treating unlabeled areas as negatives introduces severe bias. This constitutes a Positive-Unlabeled (PU) learning problem. Existing statistical methods (e.g., LAMAP) rely on hand-crafted features and strong assumptions about site absence, while standard deep learning approaches either collapse (predicting sites everywhere when trained only on positives) or fail catastrophically (inverting probability rankings when trained on uncertain negatives). The core difficulty lies in learning from sparse positive labels without assuming the absence of sites in unlabeled regions.

Methodology
The authors propose Asymmetric Dual Pseudolabeling (DPL), an end-to-end deep learning framework for semantic segmentation that addresses label scarcity and PU uncertainty without hand-crafted feature engineering.

• Architecture: The model utilizes a UNet architecture with a shared encoder and dual independent decoders. This dual-branch structure allows for the generation of diverse predictions necessary for pseudolabeling while maintaining computational efficiency.
• Asymmetric Pseudolabeling Strategy: The core innovation is an asymmetric thresholding mechanism applied during training:
  • Positive Pseudolabels: Require high confidence (conservative). A region is labeled as a positive site only if both decoder branches agree with high probability (minimum prediction > $\tau_{pos}$). These labels are soft-scaled using a temperature parameter.
  • Negative Pseudolabels: Require only weak evidence (permissive). A region is labeled as a negative (non-site) if at least one branch is skeptical (minimum prediction < $\tau_{neg}$). These labels are hard-coded to 0.
  • Dynamic Scheduling: Thresholds evolve during training via a sigmoid ramp-up function, starting with permissive negative thresholds and becoming more selective over time.
• Loss Function: The total loss combines four terms:
  1. Supervised Loss ($L_{SL\_Pos}$): Computed only on confirmed positive pixels to avoid learning from uncertain background.
  2. Pseudolabel Loss ($L_{pseudo}$): Computed on the generated pseudolabels using Focal Loss to focus on uncertain pixels.
  3. Consistency Regularization ($L_{cons}$): Enforces agreement between the two decoder branches.
  4. Entropy Regularization ($L_{ent}$): Applied to pixels where branches disagree to minimize uncertainty.
• Training Strategy: The model employs hybrid sampling (10% patches centered on known sites, 90% random) and is trained on multi-band geospatial imagery (DEM-derived features and historical infrastructure maps) without manual feature engineering.

Key Contributions

1. Formulation: Frames APM as a semi-supervised semantic segmentation task under extreme label scarcity and PU uncertainty.
2. Algorithm: Introduces an asymmetric pseudolabeling strategy that mirrors archaeological risk: conservative for false positives (requiring consensus) and permissive for false negatives (requiring only dissent), enabling learning without assuming background is truly negative.
3. Empirical Bounds: Establishes that standard supervised learning (SL) fails under negative uncertainty (inverting rankings), and positive-only training (SL_Pos) collapses to predicting everywhere. DPL operates effectively between these empirical bounds.
4. Validation: Provides rigorous evaluation on two datasets with complementary strategies: Sagalassos (independent field survey validation) and Cyprus (pure PU setting with spatial k-fold validation).
5. Reproducibility: Releases all code, preprocessing scripts, and the Cyprus dataset; Sagalassos data is proprietary but code is shared.

Experimental Results

• Sagalassos Dataset (Turkey): Evaluated against an independent field survey (Survey 2) using Survey 1 for training.
  • Performance: DPL outperformed the state-of-the-art LAMAP baseline by 12.5% in F1 and 29.3% in Recall across seven historical periods.
  • Comparison: While DPL excelled in discovery metrics (F1, Recall), LAMAP maintained superior probability ranking (PR-AUC and ROC-AUC), attributed to its explicit probabilistic modeling.
  • Baselines: Standard SL missed 85% of sites; SL_Pos collapsed to 100% recall with no discrimination.
• Cyprus Dataset (Pure PU Setting): Evaluated via uniform site-level and spatial k-fold cross-validation without confirmed negatives.
  • Performance: DPL operated nominally between the empirical bounds (SL and SL_Pos) but showed high variance and training instability. It failed to distinguish individual sites from broad coastal terrain gradients.
  • LAMAP Comparison: LAMAP substantially outperformed all deep learning methods in this setting (ROC-AUC 0.765 vs. DPL's ~0.51), suggesting that when only coarse terrain gradients are available, feature-similarity methods are more robust than pseudo-label bootstrapping.
  • Analysis: Feature separability analysis revealed that DPL's success on Sagalassos relied on spatially specific historical infrastructure features (e.g., distance to roads), which were absent in the Cyprus dataset.

Significance and Claims
The paper claims that DPL offers a tractable path for learning from incomplete, positively biased labels in spatially structured domains (relevant to ecology, epidemiology, and remote sensing) without requiring confirmed absences or class-prior estimation.

The authors emphasize that DPL is most effective when spatially specific auxiliary features (e.g., historical infrastructure maps) provide fine-grained discrimination. In landscapes where only broad geographic gradients exist (as in the Cyprus case study), the method struggles to bootstrap stable pseudo-labels, and traditional feature-similarity approaches like LAMAP remain more appropriate. The work delineates the boundary of applicability for pseudo-labeling in archaeological discovery: it is a powerful tool for site discovery when specific discriminative signals are present, but it cannot replace the need for targeted field survey or domain-calibrated models in data-scarce, feature-poor environments.