Mapping Australia's history with Mantel and the Murujuga Aboriginal Corporation
Looking north across Enderby Island.
Recently inscribed as a UNESCO World Heritage Site, the Murujuga landscape holds one of the world’s largest collections of rock art. Experts estimate the oldest petroglyphs to be approximately 50,000 years old, making them among the earliest known records of human existence and environmental change. The collection spans depictions of flora and fauna through to what’s believed to be the earliest-known representation of a human face, estimated at 30,000 years old.
For Traditional Custodians, Murujuga is not just a gallery; it is the physical starting point for many of Australia’s major songlines (Dreaming tracks). The narratives and songs that begin at Murujuga stretch thousands of kilometers across the continent, reaching as far as Alice Springs, Cape York and even Tasmania. As Custodians travel the songlines or pass down Lore, these specific petroglyphs act as vital mnemonic markers, confirming the stories embedded in the songs.
Traditional Owners refer to the site as a place of worship and describe the rock art as their “Bible”. The carvings serve as a permanent historical and ecological record of the songlines. They depict extinct megafauna, changes in sea levels since the last Ice Age and the creation beings who shaped the land.
The Murujuga Aboriginal Corporation manages a collection of global significance which is too large to document by hand. Rangers can walk the land, but Elders are ageing, sites are remote and the existing approach covers only a fraction of the terrain.
Looking south across the expanse of rolling rock-pile hills on Dolphin Island.
The Murujuga Aboriginal Corporation came to Mantel with a clear problem: mapping up to two million petroglyphs across massive expanses of rugged, remote terrain isn’t feasible using current methods. Existing processes rely on people walking the landscape on foot, which is prohibitively slow at this scale and increasingly inaccessible for senior Elders who hold deep knowledge of the sites.
While multi-modal Large Language Models (LLMs) excel at broad visual reasoning and text generation, they are fundamentally mismatched for high-volume spatial detection. Processing massive volumes of high-resolution drone imagery through heavy vision-language models is computationally slow and cost-prohibitive at scale. More critically, generalist LLMs lack the fine-grained domain specificity required to distinguish subtle, 50,000-year-old rock carvings from natural textures, shadows and erosion. The task demands high-speed spatial feature extraction, not conversation.
To solve this, Mantel developed a bespoke computer vision model using an open-source transformer, fine-tuned on custom-labelled data specific to Murujuga’s unique motifs. This specialised approach delivers rapid inference and precise detection at a fraction of the compute cost and environmental footprint. The pipeline ingests footage from drone flights, isolates individual rocks through advanced image segmentation, performs feature extraction and uses the vision model to detect and assign a confidence score to likely petroglyphs. Mantel designed the system to keep Traditional Owners in control and aligned to their cultural protocols. AI detections feed directly into a custom, human-in-the-loop web application where Murujuga Aboriginal Corporation experts can validate or reject the model’s findings, continuously enhancing model performance over time.
To facilitate visualisation and interaction, Mantel built a mapping pipeline and a 3D navigable digital twin enabling interactive flyovers and centimetre-accurate, contextual display of identified petroglyphs. These visualisation tools provide rangers and Elders with a navigable record of the landscape and the art together, without requiring physical access to the land.
Time-lapse of an experimental night-time scan using thermal imaging on Enderby Island.
Mantel delivered a pioneering spatial AI architecture, integrating custom computer vision, a spatial mapping engine, and a sovereign human-in-the-loop web application, to solve a problem previously considered insurmountable. Utilising AI effectively throughout the delivery enabled the team to achieve outcomes on timelines that were previously unfeasible.
The human-in-the-loop design means the Murujuga Aboriginal Corporation retains full authority over what’s documented, how it’s classified and how that knowledge is used. The solution respects Indigenous Data Sovereignty principles, ensuring data ownership and hosting rights remain 100% with the Murujuga Aboriginal Corporation. The ultimate vision is documentation of the entire area, creating a sovereign digital record that allows Traditional Owners to “walk the land” virtually while preserving Elders’ cultural knowledge on their own terms.
Field visit to Enderby Island for drone image capture.
Complex, high-stakes environments cannot be solved with generic AI. Off-the-shelf models and broad LLMs are built for average use cases; they are often not the appropriate tool when dealing with extreme data volumes, specialised domain nuances or strict governance needs. Solving challenges at this scale requires a purpose-built architecture: one engineered specifically to maximise detection accuracy, minimise compute costs and respect data sovereignty.
Mantel designs and builds custom machine learning systems for environments where off-the-shelf tools fall short, off-the-shelf costs explode and getting it right is non-negotiable. If your organisation is exploring computer vision, automated detection, or human-in-the-loop AI, we welcome the conversation.
