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 rooted 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 paper discusses the project's significance in enhancing Indigenous science and technology education and outlines its results in utilising 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 labelled 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 three 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.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to classify rock art patterns using machine learning (ML) techniques to assist archaeologists and rangers in the work of site archiving and protection. Specifically, the paper explores the following key issues: 1. **Insufficient educational resources and educational gap**: - By developing a citizen - science application, the project aims to increase the participation of indigenous students in STEM (science, technology, engineering, and mathematics) fields and bridge their educational gap in these subjects. This not only helps to enhance their academic interests but also provides them with the opportunity to be exposed to modern digital technologies. 2. **Challenges in cultural heritage management**: - Australia has a rich rock art heritage, but recording and classifying these sites face both financial and technological limitations. By combining machine - learning techniques, the project aims to improve the efficiency of cultural heritage management and protection. For example, using ML models to automatically identify and classify rock art images can significantly reduce the workload of manual annotation and improve the accuracy of classification. 3. **Application of machine learning with small data sets**: - The data set of rock art images is relatively small and difficult to obtain, which poses a challenge to training deep neural networks. The paper explores how to overcome this problem through the method of transfer learning, using pre - trained models (such as VGG19, ResNet50, and EfficientNet V2 S). The results show that even with limited training resources, transfer learning can still achieve a relatively high classification accuracy (Top - 1 accuracy of 79.76% and Top - 5 accuracy of 94.56%). ### Core objectives of the paper The core objective of the paper is to show how to apply machine - learning techniques to the classification of rock art patterns through interdisciplinary cooperation, thereby promoting the development of indigenous education and cultural heritage protection. Specific methods include: - **Developing a citizen - science application**: Let students participate in annotating rock art images to generate data for training ML models. - **Application of transfer learning**: Use pre - trained models for fine - tuning to adapt to the rock art image classification tasks in specific fields. - **Evaluating the performance of different models**: Through experiments with different data split ratios, evaluate the performance of various models in the rock art classification task. ### Conclusion By combining machine - learning techniques and indigenous cultural education, this project not only improves the automation level of rock art classification but also provides valuable STEM learning opportunities for indigenous students. Ultimately, this innovative method is expected to further improve the efficiency of cultural heritage protection and management in the future.

Digitising the Deep Past: Machine Learning for Rock Art Motif Classification in an Educational Citizen Science Application