Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Simon Jaxy, Anton Theys, Patrick Willett, W. Chris Carleton, Ralf Vandam, Pieter Libin

TL;DR

This study tackles archaeological site discovery under extreme label scarcity by framing predictive modeling as end-to-end semantic segmentation informed by semi-supervised dynamic pseudolabeling and CRF refinement. By integrating a dual-branch DPL mechanism with a CRF-RNN refinement layer, the approach propagates confident site signals across large landscapes using both DEM-derived features and raw Landsat9 imagery. Across two modalities and seven historical periods, the method achieves substantial Dice gains over naive supervision and performs on par with or above the LAMAP baseline in terms of predictive surface quality and calibration, demonstrating robust, scalable performance under sparse annotations. The work highlights the potential of semi-supervised learning for large-scale archaeological prospecting, offering a flexible framework that can incorporate historical context and multi-modal data while enabling end-to-end data-driven discovery with future field validation.

Abstract

Archaeological predictive modelling estimates where undiscovered sites are likely to occur by combining known locations with environmental, cultural, and geospatial variables. We address this challenge using a deep learning approach but must contend with structural label scarcity inherent to archaeology: positives are rare, and most locations are unlabeled. To address this, we adopt a semi-supervised, positive-unlabeled (PU) learning strategy, implemented as a semantic segmentation model and evaluated on two datasets covering a representative range of archaeological periods. Our approach employs dynamic pseudolabeling, refined with a Conditional Random Field (CRF) implemented via an RNN, increasing label confidence under severe class imbalance. On a geospatial dataset derived from a digital elevation model (DEM), our model performs on par with the state-of-the-art, LAMAP, while achieving higher Dice scores. On raw satellite imagery, assessed end-to-end with stratified k-fold cross-validation, it maintains performance and yields predictive surfaces with improved interpretability. Overall, our results indicate that semi-supervised learning offers a promising approach to identifying undiscovered sites across large, sparsely annotated landscapes.

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

TL;DR

This study tackles archaeological site discovery under extreme label scarcity by framing predictive modeling as end-to-end semantic segmentation informed by semi-supervised dynamic pseudolabeling and CRF refinement. By integrating a dual-branch DPL mechanism with a CRF-RNN refinement layer, the approach propagates confident site signals across large landscapes using both DEM-derived features and raw Landsat9 imagery. Across two modalities and seven historical periods, the method achieves substantial Dice gains over naive supervision and performs on par with or above the LAMAP baseline in terms of predictive surface quality and calibration, demonstrating robust, scalable performance under sparse annotations. The work highlights the potential of semi-supervised learning for large-scale archaeological prospecting, offering a flexible framework that can incorporate historical context and multi-modal data while enabling end-to-end data-driven discovery with future field validation.

Abstract

Archaeological predictive modelling estimates where undiscovered sites are likely to occur by combining known locations with environmental, cultural, and geospatial variables. We address this challenge using a deep learning approach but must contend with structural label scarcity inherent to archaeology: positives are rare, and most locations are unlabeled. To address this, we adopt a semi-supervised, positive-unlabeled (PU) learning strategy, implemented as a semantic segmentation model and evaluated on two datasets covering a representative range of archaeological periods. Our approach employs dynamic pseudolabeling, refined with a Conditional Random Field (CRF) implemented via an RNN, increasing label confidence under severe class imbalance. On a geospatial dataset derived from a digital elevation model (DEM), our model performs on par with the state-of-the-art, LAMAP, while achieving higher Dice scores. On raw satellite imagery, assessed end-to-end with stratified k-fold cross-validation, it maintains performance and yields predictive surfaces with improved interpretability. Overall, our results indicate that semi-supervised learning offers a promising approach to identifying undiscovered sites across large, sparsely annotated landscapes.
Paper Structure (46 sections, 6 equations, 29 figures, 4 tables)

This paper contains 46 sections, 6 equations, 29 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Pseudolabels
  4. Positive-Unlabeled Learning
  5. Semantic Segmentation with Sparse Labels
  6. Related Work
  7. Deep Learning in Archaeological Predictive Modeling
  8. Methodology
  9. Data and Feature Modalities
  10. Deep learning methods
  11. Supervised learning
  12. Dynamic Pseudolabels
  13. Conditional Random Fields
  14. LAMAP
  15. Experimental Design
  16. ...and 31 more sections

Figures (29)

  • Figure 1: Computational graphs for (a) Supervised Learning (SL), (b) Dynamic Pseudolabeling (DPL), and (c) Conditional Random Field Dynamic Pseudolabeling (CRF-DPL).
  • Figure 2: Surface plots comparing (a) CRF-DPL and (b) LAMAP. Heatmap (c) shows differences: red indicates higher probability for CRF-DPL, blue for LAMAP. LAMAP produces more fine-grained predictive surfaces.
  • Figure 3: DEM-model multi-objective results averaged over all time periods. We show the volume gain compared to the Baseline LAMAP.
  • Figure 4: Positive ratios of each DEM-model per discretized probability bin.
  • Figure 5: Probability densities per DEM-model (left) and Landsat9 model (right).
  • ...and 24 more figures