In a study published in Australian Archaeology this week, researchers from Flinders University say machine learning was effective in identifying the stylistic differences in 98 Aboriginal rock art paintings, with a computer system able to categorise the artefacts with 72 per cent accuracy despite the AI having no prior exposure to the paintings.

The computer software used in the study was pre-exposed to a database of 14 million generic photos curated by Stamford University. From this information, the AI is able to recognise and sort new generic objects presented to it.

Flinders University Archaeologist Dr Daryl Wesley said the idea for the project came a few years ago when a team of European researchers discovered machine learning was capable of sorting German renaissance paintings by historical period.

“That gave us an idea of well, what if we can we apply that to aspects of rock art [and] look at grouping it by similarity and difference,” Wesley said.

“Rock art is certainly an area in archaeology that has a lot of different opinions about how we can study difference and similarity and what constitutes a style, a manner, or a way of painting.”

The team of researchers travelled to the Arnhem Land wilderness and the nearby Wilton River region at the top of Northern Territory to document 98 pieces of Aboriginal rock art over the course of a week.

Wesely said the researchers’ surveying work, which required extensive photography, 3D modelling and photogrammetry, was informed by the land’s traditional owners – the Mimal and Marrku people.

The 98 paintings all depicted human figures in six well-known stylistic classes: Dynamic Figures, Post-Dynamic Figures, Northern Running Figures, Simple Figures with Boomerangs, Simple Figures Round Headdresses, and Wilton River Region Simple Figures.

“Because Arnhem Land rock art is very well known, and there’s a good understanding of the different styles, we wanted to see whether the machine would generate the same observations that generations of archaeologists have made,” Wesley said.

Black and white traces of the images were sent back to the research team in Adelaide for the computer to analyse.

The AI placed the images along a spectrum of stylistic differences, providing researchers with not only an accurate categorisation of style, but also insights into the art’s history.

“We analysed a group of figures, and it then reproduced the results and ordered the different styles in chronological order, which was quite interesting, and [it] showed where the styles fit and where they overlapped,” Wesley said.

“Why this is important is because normally when we look at a particular rock art motif, we might try and look at a dozen different aspects or characteristics to compare with.

“What the machine does is it analyses each image on thousands of points of difference, which we don’t exactly have.”

The AI also placed Arnhem Land and Wilton River region art on different parts of the spectrum, which researchers say shows the potential for the technology to give archaeologists geographic information about rock art that was not previously available.

Daryl Wesley, Jarrad Kowlessar, Desmond
Lindsay and Peter Cooke on the Wilton River (Photo: MLMAC)

The land’s traditional owners were excited by the findings and hoped further research could provide insights on the movement of their ancestors, according to Wesley.

“What they really would like to know more about is what is happening on their country, what their ancestors were doing and how they might have been interacting with their neighbours, because that’s very important in terms of kinship and ceremony and other things,” he said.

“In provinces like Arnhem Land where there’s millions of paintings throughout the whole area, we might actually get to look at movement of ideas, concepts and designs and styles as they emerge and evolve.”

PhD student Jarrad Kowlessar – who set up the machine learning system at Flinders University alongside fellow researcher James Keal – said it can be very difficult for humans to identify geographic differences in rock art style.

“What excites me most is that this method removes kind of the bias that a researcher might have when they’re interpreting something about how similar two things might be to one another stylistically,” Kowlessar said.

“This is obviously a huge advantage to rock art because it can be so subjective and you could have so many biases about what you believe that can kind of influence your interpretation.”

Kowlessar said the next step was to analyse rock art paintings of non-human figures.

“The next thing we’d really like to do is to see if an approach like this can separate out animal species from one another,” he said.

“For example, different kinds of macropods can be really hard to differentiate for us subjectively.

“The silhouette of a wallaby looks similar to the silhouette of a kangaroo, and knowing which one was being painted is very hard for an interpreter.

“But an approach like this might help with things like species differentiation.”