Subsurface and Remote Archaeology: GPR, LiDAR, and the Data Sovereignty Debate

Subsurface and Remote Archaeology: GPR, LiDAR, and the Data Sovereignty Debate

Advanced remote sensing, airborne laser scanning (LiDAR), and machine learning are enabling archaeologists to map buried and obscured historical structures across vast landscapes without breaking ground. However, as these technologies scale, they are colliding with intense debates over data sovereignty, indigenous heritage rights, and the ethics of crowdsourced "digital colonialism."¹

The OpenAI to Z Challenge: Crowdsourced Discovery and Its Backlash

In May 2025, artificial intelligence giant OpenAI launched the OpenAI to Z Challenge, a public hackathon offering up to $250,000 in cash and API credits. The competition challenged participants to use OpenAI's generative models (such as GPT-4o and o3/o4 mini) alongside public datasets—including high-resolution satellite imagery, published LiDAR tiles, and historical records—to identify undiscovered archaeological sites in the Amazon rainforest.

On August 28, 2025, OpenAI announced the winning three-person team, "Black Bean" (led by Meta software engineer Yao Zhao). The team trained deep learning models on Google Earth Engine and NASA digital elevation models, using GPT-4o to learn the spatial patterns of known archaeological sites. Their model successfully flagged 67 distinct, square-mile patches across the Amazon basin (primarily in Brazil) that are highly likely to contain ancient settlements, with a strong clustering near accessible water sources.

Despite the technical efficiency, the challenge ignited a fierce international backlash:

• Lack of Informed Consent: Archaeologists and Indigenous leaders pointed out that OpenAI failed to consult any of the 300+ Indigenous groups living in the target areas, violating international treaties like the 1989 Indigenous and Tribal Peoples Convention (ILO 169).
• Vandalism and Security Risks: Daniel Kuikuro, president of the Kuikuro Indigenous Association of the Upper Xingu, warned that publishing site coordinates could expose sacred and historically sensitive sites to looters, land grabbers, and hostile political groups.
• Corporate Model Extraction: Technology ethicists argued that OpenAI used the guise of archaeology to harvest free public labor to test and refine its proprietary geospatial and multi-modal models.
• State Intervention: On July 1, 2025, the Brazilian government's Ministry of Indigenous Peoples demanded that OpenAI halt the publication of any findings, coordinates, or data until the company clarified its legal and methodological framework.

"The OpenAI to Z Challenge has mobilized tech-savvy researchers worldwide, but has also faced criticism from archaeologists, Indigenous communities, and tech ethicists who argue it ignores important research norms, including consultation with the more than 300 Indigenous groups who live in the rainforest. Last week, the Brazilian government demanded OpenAI address the concerns." — Sofia Moutinho, Science (July 2025)

"Daniel Kuikuro, president of the Kuikuro Indigenous Association of the Upper Xingu, says he fears data generated by the contest could lead to vandalism of Indigenous sites from groups opposed to Indigenous land rights. University of Virginia technology ethicist Lori Regattieri adds that under the challenge rules, all data submitted by contestants and their computer models become the property of OpenAI. She suggests the company is selling the challenge as an archaeology research project when its real goal may be to enlist free labor to test AI geospatial models." — Sofia Moutinho, Science (July 2025)

DINO-CV: Self-Supervised Learning and Collaborative Heritage Management

To map obscured archaeological features without the ethical pitfalls of crowdsourced extraction, researchers are turning to collaborative, self-supervised machine learning frameworks.

In November 2025, researchers from the University of Melbourne, in formal collaboration with the Gunditj Mirring Traditional Owners Corporation, published DINO-CV. This self-supervised, cross-view pre-training framework is designed to map low-lying, ancient dry-stone walls in the Budj Bim Cultural Landscape in Victoria, Australia (a UNESCO World Heritage site).

Because low-lying stone walls are heavily obscured by dense vegetation in standard optical satellite imagery, the researchers utilized airborne LiDAR data collected during the dry season. They processed the LiDAR data into Digital Elevation Models (DEMs) and generated two complementary terrain visualizations:

1. Multi-directional Hillshade (MHS): Simulates illumination from multiple sun angles to highlight linear geometry.
2. Visualization for Archaeological Topography (VAT): Combines slope, local relief, openness, and sky-view factors to emphasize micro-topographic variations.

DINO-CV utilizes a self-supervised knowledge distillation framework (based on DINO) to train a student network to learn view-invariant structural features by matching representations across MHS and VAT views, without requiring manual annotations.

Overcoming the Annotation Bottleneck

When fine-tuned on a labeled dataset (BudjBimArea), DINO-CV demonstrated extraordinary label efficiency:

• 10% Labeled Data: With only 10% of annotated training data, the model retained a 63.8% mean Intersection over Union (mIoU), compared to just 21.4% for a randomly initialized model and 49.3% for a standard supervised model.
• Generalization: The model successfully delineated stone wall networks beneath dense forest canopies in unseen regions located 10 km away from the training data.

Crucially, the project was conducted under a formal Memorandum of Understanding (MoU). All raw data and derived archaeological models remain the sole property of the Gunditjmara Traditional Owners, serving as a model for ethical, indigenous-led digital archaeology.

"Applied to the Budj Bim Cultural Landscape (Victoria, Australia), a UNESCO World Heritage site, the approach achieves a mean Intersection over Union (mIoU) of 68.6% on test areas and maintains 63.8% mIoU when fine-tuned with only 10% labeled data... Beyond archaeology, this approach offers a scalable solution for environmental monitoring and heritage preservation across inaccessible or environmentally sensitive regions." — Huang et al., arXiv:2510.17644