Fake it till you make it: Deep learning detection of archaeological features using simulated training data

Summary

This is an abstract from the "Practice, Theory, and Ethics of Machine Learning in Archaeology" session, at the 90th annual meeting of the Society for American Archaeology.

High resolution digital surface datasets have become increasingly accessible over the last two decades. Archaeologists have responded by developing methods to streamline locating archaeological sites in these data at a landscape scale. As high-powered computing hardware and cloud computing solutions improve, deep learning models are now a feasible approach to mapping archaeological sites using geospatial big data. However, training a deep learning model requires a comprehensive dataset of annotated features. While data augmentation methods can expand small datasets, there are some archaeological site or feature types for which a training dataset might not exist. This issue becomes a bottleneck for archaeologists who wish to incorporate deep learning into their research and management toolkits. In this paper, we propose several procedural generation workflows to create simulated training datasets for different historic archaeological features (tar kilns, railroad grades, foundations). We then use these training datasets to train U-Net and Mask R-CNN convolutional neural network models and evaluate each model’s metrics when detecting real features in the Kisatchie National Forest, LA. Model performance varies, but successful models generally have high recall and low precision, suggesting that while this approach may have utility, it requires additional fine-tuning and robust post-processing to be reliable.

This Resource is Part of the Following Collections

• Society for American Archaeology 90th Annual Meeting, Denver, CO (2025)   •
• Practice, Theory, and Ethics of Machine Learning in Archaeology

Keywords

Record Identifiers

Abstract Id(s): 51932