The Petroglyph Analyzer: CDA

The Arizona State University Artificial Intelligence Cloud Innovation Center (AI CIC), powered by Amazon Web Services (AWS), collaborated with the Center for Digital Antiquity (CDA) at Arizona State University to develop the Petroglyph Analyzer. The CDA is dedicated to the long-term preservation of and access to irreplaceable archaeological data, primarily through tDAR (the Digital Archaeological Record); this is an international repository of digital archaeological datasets, images, reports, and other digital information. 

The project focused on building a research and experimentation platform combining custom instance segmentation (YOLOv8 trained on SageMaker) with LLM-based classification (Claude on Amazon Bedrock) to accelerate and standardize petroglyph documentation. The platform enables researchers to upload batches of field photos, automatically segment individual rock art instances, classify each instance against the team's established coding schema, and return structured results for expert review and correction. This allows the team to develop and validate their own AI-assisted recording methodology.

The CDA's research team documents thousands of petroglyph instances across hundreds of field sites in the American Southwest. This manual process requires physical examination of each panel and hand-coding of motif types, techniques, conditions, varnish classes, superimposition relationships, and other attributes into standardized recording forms. 

This manual process is slow (approximately 30 minutes per panel with multiple glyphs) and subjective, introducing reliability issues when different field teams apply the coding schema inconsistently. The coding schema itself is complex, with hundreds of possible motif codes and multiple technique and condition categories, making consistent human application difficult even with training. This creates a growing bottleneck as sites are surveyed and photographed faster than the resulting data can be processed and entered into the database. 

Existing tools are limited to purely manual data entry via spreadsheets and database forms with no AI assistance. The research team needed a platform to explore whether instance segmentation combined with LLM-based classification can accelerate and standardize the documentation process by providing AI-generated first-pass classifications that domain experts can rapidly review, correct, and refine.

The ASU AI CIC team built this on a fully serverless AWS architecture using AWS CDK (TypeScript) for infrastructure and Python for AWS Lambda functions. The core pipeline operates through the following stages:

• Researchers upload 75–300 petroglyph panel images per batch via presigned Amazon S3 URLs.
• An AWS Step Functions pipeline invokes a YOLOv8 model on Amazon SageMaker to detect individual petroglyph instances and generate bbox-cropped images.
• Cropped instances are sent to Amazon Bedrock (Claude) via batch inference for structured classification (motif code, technique, condition, etc.).
• Researchers review results, correct misclassifications, and flag images with poor segmentation for retraining.
• A continuous training loop runs alongside the classification pipeline, allowing admins to iteratively improve the segmentation model using Amazon SageMaker Ground Truth and Amazon SageMaker Training Jobs.
• The system deploys via a single shell script that orchestrates AWS CodeBuild, requiring no local Node.js, CDK, or Python installation. A GitHub account connects to AWS Amplify to pull frontend code, and the deployment automatically creates and manages IAM roles and policies, CodeBuild projects, AWS Secrets Manager secrets, Amazon CloudWatch Logs, AWS CloudFormation stacks, and Amazon SES email identities.

The platform transforms what was a purely manual, inconsistent process into an AI system that gets better with use. The platform reduces per-panel documentation time from approximately 30 minutes to approximately 5 minutes of review and correction. Financially, configuring Application Auto Scaling with scale-to-zero capabilities eliminates approximately $530/month in idle GPU costs when no batches are being processed. Additionally, utilizing Amazon Bedrock batch inference provides a ~50% cost reduction compared to on-demand per-image API calls.

For the past 10 years I’ve been thinking hard about how to leverage AI in my research, and especially how to use machine learning to streamline and standardize the inventory, coding, and analysis of the more than 100,000 Indigenous petroglyphs I and my colleagues have documented across southern Arizona. I didn’t know where to start and felt I was in over my head. The team at AI CIC pulled it together in a remarkably collaborative, efficient, and transparent manner. I really look forward to putting this tool to use and scaling outward.

- Aaron Wright, Preservation Anthropologist, Archaeology Southwest

A longer-term goal for the CDA is to build a searchable database of petroglyph classifications that enables comparison of glyphs across different sites around the world. Future scope additions include exposing configurable training hyperparameters (epoch count, learning rate, image size, and augmentation settings) via the admin UI, establishing a single-class segmentation mode for sites where individual glyph boundaries are ambiguous, and multi-site batch management tools. 

Additional enhancements include real-time single-image classification previews for prompt testing, automated prompt evaluation scoring based on researcher feedback patterns, and direct export integration with GIS tools (Shapefile/GeoJSON with georeferenced glyph locations).