Digitising the Deep Past: Machine Learning for Rock Art Motif Classification in an Educational Citizen Science Application
Published
10-12-2024
Type of Work
Article - Journal
Advisors and Contributors
Firstly, for their generous and ongoing collaboration, we are indebted to the Laura community, who are committed to preserving their cultural heritage, often in the face of limited resourcing. The Laura Rangers, in particular, have been instrumental in the design and implementation of this project and hold the collection of rock art images used for this project. The Laura State School, particularly principal Melissa Shepherd, are owed a debt of thanks for their enthusiastic support and participation. The students of Laura State School have contributed materially to this research and the expanded understanding of the researchers. The pedagogy team, particularly Dr Troy Meston and Dr Tasha Riley, were not directly involved in the app’s design and analysis of ML data but were extremely helpful in informing and advising the programme’s rollout in the school environment; Dr Tristen Jones also assisted in lending her experience in cultural heritage programs geared towards Indigenous primary school students, Dr Andrea Jalandoni reviewed a draft version of the manuscript and Fiona McKeague is also thanked for her enthusiastic engagement in many aspects of the project, including input on designing the app, testing the app’s functionality and traveling to Laura to facilitate the classroom delivery.
Project Affiliation
Griffith University, Griffith Centre for Social and Cultural Research
Abstract
Digitising the Deep Past (DDP) is an interdisciplinary project based at Griffith University, Australia, that innovates in three areas: Indigenous cultural heritage, Indigenous education, and Machine Learning (ML) and Artificial Intelligence (AI). The project investigates the use of a purpose-built citizen science application that engages Indigenous youth in educational exercises grounded in local cultural heritage, specifically rock art, making learning more engaging and exposing them to digital technologies. Furthermore, ML models trained with the data gathered through these educational activities can then assist with classifying new rock art images and assisting rangers and archaeologists with site archiving and conservation efforts. This article discusses the project’s significance in enhancing Indigenous science and technology education and outlines its results in using ML for rock art classification. Adopting deep learning in rock art classification offers a compelling avenue for the automated analysis and interpretation of heritage objects and places. However, training deep neural networks from scratch often requires enormous datasets and computational resources, posing challenges for domain-specific applications with smaller datasets. With a dataset comprising approximately 3,100 labeled rock art images, we evaluated various tools within the transfer learning toolbox using three prominent pre-trained architectures: VGG19, ResNet50 and EfficientNet V2 S. Through the collaborative efforts of Indigenous students and ML, we demonstrate that even with limited training resources, using transfer learning to re-purpose an existing model can achieve motif classification Top 1 accuracy of 79.76% and Top 5 of 94.56%. The project ran from 2021 to 2023, including 3 week-long sessions with students of Laura State School to trial the citizen science app and the evaluation, development and refinement of the ML models. The DDP project not only serves as a beacon for community-centric research but also forges a new frontier in integrating Indigenous cultural heritage with modern technology. The impact reaches beyond academia, directly enriching the educational experience for Indigenous students in Laura and equipping local rangers and archaeologists with advanced tools for rock art conservation.
Citation
Turner-Jones, R. N., Tuxworth, G., Haubt, R. A., & Wallis, L. (2024). Digitising the Deep Past: Machine Learning for Rock Art Motif Classification in an Educational Citizen Science Application. ACM Journal on Computing and Cultural Heritage, 17(4), 1-19. https://doi.org/10.1145/3665796
Link to Published Work
Income (grants or other funding)
Other
Spotlight grant, Griffith University