Juvenile Palm Cockatoo, photographer Celina Cacho

Using AI to help conserve “Ringo Starr” bird hollows

We’re advancing palm cockatoo research with technology trials at Weipa

Last updated: 12 April 2024

 

It’s a steamy summer sunrise in Far North Queensland, Australia.

Our small group steps as quietly as we can through the leafy bush as we look up, carefully scanning each branch, hoping to win the bird-watching lottery.

“Quick – up in that tree!” Each of us peers through binoculars in turn, catching glimpses of sleek black feathers and bright red cheeks before the majestic parrot takes flight.

We’re now among the lucky few who have seen the elusive palm cockatoo in the wild.

Our guide, Celina, was a Threatened Species Researcher at our bauxite operations at Weipa. She’s driven us deep into the bush at the crack of dawn to see firsthand these rare birds she’s so passionate about conserving.

As the leader of our Palm Cockatoo Research Program, Celina worked with colleagues and industry experts to improve palm cockatoo conservation using a tool you might not expect – machine learning.

Site visit with Naziah and Alice
Celina (left) on-site at Weipa with colleagues Naziah and Alice who helped train YOLOv5 for the project.

Finding and protecting rare palm cockatoos

Palm cockatoos are a pretty big deal. Literally. Of the 28 cockatoo species found across Australia and South-East Asia, they’re the largest – fully grown, they’re around 60 centimetres (2 feet) tall, nearly a third bigger than their common sulphur-crested cousins.

They’re also known as the “Ringo Starr” of the bird world, thanks to their unique rhythmic style – they’re the only bird in the world that manufactures drumsticks to drum with the same rudiments as human instrumental music.

But numbers of these avian rockstars are also dramatically declining in Cape York Peninsula, their only stronghold in Australia. They’re fussy nesters, who only breed once every two years on average, and lay a single egg with low offspring success when they do.

Palm cockatoos nest on the lands where we mine bauxite at Weipa. So when we clear land for mining, we survey carefully to make sure we’re not impacting their breeding habitat.

After a lot of walking through the bush, when our teams find a suitable nesting hollow, we use camera “traps” to confirm their breeding sites, and then create a permanent buffer of trees around their hollow until the tree falls of natural causes.

Across a huge mining site, though, it’s a challenge to manually achieve this at such a large scale. So we’re working with technology partners to streamline how we identify certain breeding behaviour on camera traps, which also helps us gather important data on nesting success.

  • Male Palm Cockatoo, photographer Sean Craven
    Male Palm Cockatoo. Image: Sean Craven
  • Breeding hollow at Weipa
    Breeding hollow at Weipa.
  • Male Palm Cockatoo, photographer Adrian May
    Male Palm Cockatoo. Image: Adrian May
  • Top view of a Palm Cockatoo Hollow, photographer Celina Cacho and Ejsha Thorpe
    Top view of a Palm Cockatoo Hollow. Image: Celina Cacho and Ejsha Thorpe

Using big data to find big birds

Previously, our ecologists had to create datasets by manually poring over each image to find and tag cockatoos. But we’re now trialling a machine-learning pipeline that’s helping us significantly speed up the process.

We’ve trained an open-source model base – an object-detection model called YOLOv5, short for ‘you only look once’ – to identify palm cockatoos and their competitors in images. It can process thousands of images rapidly, sort them into classes and make verifying images much easier.

Model Improvement using an object-detection model called YOLOv5
Over four months, the model improved significantly in its ability to spot the elusive palm cockatoo, as shown in the progression of shots here. Image: Celina Cacho

The unique way we deploy the camera traps was developed alongside Ecotone Flora Fauna Consultants specifically for the project back in 2015. The traps capture roughly 20,000 time-lapse images at each nesting hollow every three weeks.

“As a conservation tool, this has the potential to help our team identify key areas of critical nesting habitat for long-term protection, without limiting our research aims by using resources on data review,” Celina says.

The trained model also reduces double-handling and data entry, and each reviewed image also goes straight into the broader dataset for automatic model retraining.

So we can spend less time analysing and entering data, and more on understanding the attendance patterns, breeding behaviour and timelines of nest success or failure.

We’re using this insight to prepare better land management strategies.

"Minimising the manual data review means we can increase our research effort,” Celina says.

“We’re aiming to use those freed up resources to understand more about how these birds breed across the wider landscape, and how many breeding hollows each pair maintains and defends in their network.”

Automating a strong base for future research

The initial trial was successful. We already had the most comprehensive long-term breeding-behaviour dataset for the palm cockatoo to date, which provided the initial training data. Now, we have more than 10,000 images tagged to identify the cockatoos, with more added regularly.

The model’s reliability and accuracy will continue to improve as it develops, meaning we can refocus our research resources for broader work in the field, like targeting entire hollow networks to better understand how pairs select and defend their breeding sites.

We’re also exploring how we can adapt the model for other types of conservation research.

"Palm cockatoos have such unique and complex calls, and each population has slight differences in their primary contact calls,” Celina says.

“So, thanks to our Brisbane team, we’re trialling a proof-of-concept model to process bioacoustic data as well. A call recogniser for palm cockatoos could shed some more light on how closely connected these populations are, and add another feasible survey tool to our kit.”

“The underlying architecture and process could also translate well to other applications, sites and species – as long as someone is happy to do some initial hard yards of manual tagging!”

The second phase of the trial was successfully completed in 2023, and we have now operationalised the technology, using it as part of our annual Palm Cockatoo hollow monitoring program. This innovation and use of AI has significantly reduced data analysis and processing times, freeing resources up to focus on other conservation and land management projects.

Header image: Celina Cacho

Related content

Autonomous vehicles

New tech for renewable transport

Transitioning the Pilbara to zero-emission trucks and trains
Autonomous truck, Gudai-Darri, Pilbara

Look inside a mine of the future

Our most intelligent mine yet is pioneering new mining technologies
Pilbara AutoHaul train

How did one of the world's largest robots end up here?

From big data to building the first autonomous train, innovation is part of our DNA