Combining Aerial Lidar and Deep Learning to Detect Archaeological Features in the Piedmont National Wildlife Refuge, Georgia

Summary

This is an abstract from the "Big Ideas to Match Our Future: Big Data and Macroarchaeology" session, at the 89th annual meeting of the Society for American Archaeology.

A growing number of archaeologists are using lidar-derived high-resolution Digital Terrain Models (DTM) to detect and document archaeological features. Early adopters used visualizations to manually detect archaeological features; however, recent technological advances provide new tools that can considerably increase the efficiency and effectiveness of archaeological feature detection in DTMs. Our team uses Deep Learning Convolutional Neural Network (CNN) models—a subset of the larger Machine Learning (ML) toolset—to quickly detect archaeological features over large areas. CNNs specialize in the analysis of visual imagery, making them particularly useful for the detection of visual features. For this project, we trained U-Net models with VGG and ResNet backbones to detect the presence and location of archaeological cotton terraces in the Piedmont National Wildlife Refuge, Georgia. While not all models produced compelling results, the best ones reached recall (also known as the True Positive Rate) values >85% and detected terraces that had been missed during manual annotation, thus illustrating the robustness of this method. In this presentation, we demonstrate the steps required to train our models, best practices for a successful use of this methodology, and the resulting maps and outputs it can create.

This Resource is Part of the Following Collections

• Society for American Archaeology 89th Annual Meeting, New Orleans, LA (2024)   •
• Big Ideas to Match Our Future: Big Data and Macroarchaeology

Keywords

Spatial Coverage

Record Identifiers

Abstract Id(s): 38618.0