With GeoPACHA-AI, archaeologists and machine learning experts have teamed together to design and build a new Vision Transformer (ViT) foundation model (DeepAndes) from WorldView 2 and WorldView 3 high resolution multispectral satellite imagery, with potential applications in many areas of research, from the earth sciences to urban planning and disaster preparedness. We then fine-tuned DeepAndes with a range of human-labeled data, resulting in our archaeology-specific AI model, DeepAndesArch.
Our AI framework includes both semantic segmentation models of agricultural field systems (including active and abandoned terrace complexes, for example) and object detection models for identifying features associated with human settlements, such as abandoned architectural structures (archaeological buildings) and features related to pastoralist settlements in the high puna grasslands. As a federated, collaborative effort, GeoPACHA-AI harnesses the regional and domain expertise of a network of archaeological experts and their teams. It builds on our prior efforts that used "brute force" (manual) imagery survey across eight survey zones and documented about 40,000 archaeological loci. Our teams are expanding on those findings by generating expertly-curated training datasets and vetting the results of the AI detections in an iterative process that will result in an order of magnitude greater coverage than was possible via brute force methods. This "expert-in-the-loop" approach is central to how we envision GeoPACHA-AI will bring high accuracy detections of archaeological loci across the Andes.
The GeoPACHA-AI technology stack is built on two AI models: DeepAndes, a Vision Transformer (ViT) model that we have created from the DINOv2 self-supervised framework, using a large sample (3 million image patches), and DeepAndesArch, a fine-tuned archaeology-specific model based on the training data inputs from our collaborating teams. We are now in the process of generating a revised DeepAndes foundation model (DeepAndesV2) based on DINOv3 and a larger sample of high resolution multispectral satellite imagery and environmental rasters. DeepAndesV2 is enabled by the supercomputing resources of Oak Ridge National Laboratory. Both models and the resulting database will be shared with the archaeological research and conservation communities to support investigation and heritage management.
Deployment of DeepAndesArch then enables autonomous detection of a range of archaeological features over large areas of Andean South America. The GeoPACHA-AI platform enables international teams of archaeologists to systematically and efficiently audit these autonomous detections. We will then use these audited data to improve DeepAndesArch performance to human- or better-than-human sensitivity and specificity, for deployment over virtually all of the Andes—an area of about 2 million square kilometers in an area approximately coincident with the historic footprint of the Inka Empire.
Our goal is to detect the visible relict architectural features, active and abandoned agricultural infrastructure (terrace and field systems) across this vast area for the archaeological research community and heritage management entities.