Machine learning applications in geoarchaeology: Using digital elevation data to model locations of natural and cultural features

Authors: Leila Character*, University of Texas at Austin, Tim Beach, University of Texas at Austin, Cody Schank, University of Texas at Austin, Takeshi Inomata, University of Arizona, Agustin Ortiz JR, Underwater Archaeology Branch, Naval History and Heritage Command
Topics: Spatial Analysis & Modeling, Historical Geography
Keywords: machine learning, deep learning, artificial intelligence, geoarchaeology, remote sensing, Maya, caves, shipwrecks
Session Type: Virtual Paper
Day: 4/11/2021
Start / End Time: 11:10 AM / 12:25 PM
Room: Virtual 36
Presentation File: No File Uploaded


This project entails creating a series of supervised machine learning models to predict the locations of both natural and cultural features using digital elevation data. The goal of this work is to bridge the gap between the field of machine learning pursued by computer scientists and the types of on-the-ground projects of interest to geoarchaeologists. This project began in 2018 with the goal of creating a targeted method of finding cave entrances in the dense tropical forest of Guatemala and Belize. In 2019, we used a random forest classifier, airborne laser scanning (ALS) data, and a training dataset of known caves to successfully identify several previously undocumented caves in northwestern Belize. Building on this work, modeling has been expanded to include other types of hidden and obscured features that colleagues are interested in studying. These include ancient Maya archaeological features in Guatemala and Mexico and shipwrecks off the coast of the United States. The models for the archaeological features take lidar and sonar imagery as input, are based on existing convolutional neural network architectures, and make use of transfer learning. These models can be used to create more accurate maps of natural and archaeological features to aid management objectives, study patterns across the landscape, and find new features. Such models can easily be adjusted to identify other types of features, even using multispectral or RGB imagery as input. This work seeks to make machine learning methods accessible to non-computer scientists interested in study, management, and/or conservation of the landscape.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login