TLDR: A new deep learning method uses semi-supervised pseudolabeling and Conditional Random Fields to find archaeological sites in vast, sparsely labeled landscapes. It learns from few known sites and abundant unlabeled data, outperforming traditional methods in some aspects and offering a more scalable, flexible approach for archaeological site discovery using both terrain data and raw satellite imagery.
Archaeological site discovery has long been compared to searching for “needles in a haystack.” Vast, diverse landscapes hold the secrets of past human activity, but the actual sites are rare and often leave only subtle traces. Traditional methods, such as extensive field surveys, are incredibly labor-intensive and time-consuming. While statistical models like LAMAP have aided in identifying potential sites, they often struggle with the complexity of modern data and require significant manual effort to define features.
A new research paper, “Needles in the Landscape: Semi-Supervised Pseudolabeling for Archaeological Site Discovery under Label Scarcity,” introduces a promising deep learning approach to this challenge. The core problem in archaeological predictive modeling (APM) is “label scarcity” – we know where some sites are (positive labels), but the vast majority of the landscape is unlabeled, not necessarily confirmed as “non-sites.” This makes traditional supervised learning difficult.
A Smarter Way to Learn: Semi-Supervised Pseudolabeling
To overcome the scarcity of labeled data, the researchers adopted a semi-supervised, positive-unlabeled (PU) learning strategy. This means the model learns not just from the few known archaeological sites, but also intelligently uses the abundant unlabeled data. They implemented this as a semantic segmentation model, which essentially assigns a “likelihood of being a site” to every pixel in a given landscape image.
The key innovation lies in “dynamic pseudolabeling,” a technique where the model generates its own “fake” labels for unlabeled areas based on its current understanding. These pseudolabels are then refined using a Conditional Random Field (CRF), which acts like a spatial filter, ensuring that predictions are consistent and smooth across the landscape, respecting natural boundaries and features. This process helps increase the confidence of the labels, even when dealing with a severe imbalance between known sites and the vast unlabeled areas.
Testing the Approach
The model was evaluated on two distinct datasets from the Sagalassos Study Area in southwestern Turkey, a region known for its rich archaeological history. The first dataset used features derived from a Digital Elevation Model (DEM), which represents terrain as continuous elevation values along with other geographical descriptors like slope and aspect. The second dataset utilized raw Landsat 9 satellite imagery, providing multispectral data without prior engineering of features.
When tested against LAMAP, a state-of-the-art traditional method, the new deep learning approach (specifically the CRF-DPL model) performed comparably in terms of overall ranking of positive samples (AUROC) and achieved higher Dice scores, which measure the overlap between predicted and actual site areas. This indicates a better ability to delineate the actual extent of potential sites. Crucially, the deep learning models demonstrated a significant improvement in overall performance across various metrics compared to LAMAP, especially when considering a broader range of evaluation criteria.
Beyond Hand-Crafted Features
One of the significant advantages of this deep learning framework is its ability to move beyond rigid, hand-crafted models. Instead of requiring experts to manually define features that might indicate a site, the neural networks learn these complex representations directly from the data. This makes the approach more flexible and scalable, particularly when dealing with raw satellite imagery where manual feature engineering would be prohibitively complex.
The study also showed that the semi-supervised models produced predictive maps that were more interpretable, especially when trained on raw Landsat 9 imagery. While LAMAP often produced very fine-grained details, the deep learning models, particularly with the CRF refinement, generated cleaner and more stable probability surfaces, making it easier for archaeologists to understand and act upon the predictions.
Also Read:
- AI’s Eye on Bangladesh: Mapping River Erosion and Lost Villages with Satellite Imagery
- Enhancing Uncertain Knowledge Graph Completion with Confidence Distribution Learning
Future Prospects for Archaeology
This research highlights the immense potential of deep learning in advancing archaeological predictive modeling. By effectively addressing the challenge of label scarcity, these semi-supervised methods offer a scalable and adaptive framework for identifying undiscovered sites across large, sparsely annotated landscapes. This could significantly reduce the time and resources required for initial prospecting, allowing archaeologists to focus their efforts more efficiently and uncover more of our shared human history.
