What makes a good tree? We used AI to ask birds
For IT teams, the notable pattern here is not the ecology story itself, but the multi-stage AI pipeline used to translate complex real-world geometry into design choices. The workflow combines laser scanning, point-cloud classification, branch-attribute extraction, and behaviour prediction. That is a familiar architecture for any team turning unstructured spatial data into actionable outputs: each stage depends on the quality of the previous one, so capture fidelity and model assumptions matter as much as the final design.
The article also highlights a practical trade-off in applied ML: the system is not optimising for what looks most tree-like to humans, but for what a downstream decision model predicts birds will actually use. In enterprise terms, that is a shift from aesthetic similarity to task-specific utility. It is a useful reminder that model objectives should be aligned to the real operational goal, otherwise AI can produce convincing but ineffective outputs.
Another important angle is deployment and maintainability. The proposed structures are intentionally lightweight, reconfigurable and removable, which reduces installation friction and supports iterative refinement. That matters because the design is not a one-off deliverable; it is a feedback loop that depends on field validation, observational data, and repeated updates. In other words, the system behaves more like a living service than a static asset.
For technical leaders, the broader lesson is about human-in-the-loop design extended to non-human stakeholders through model-mediated inference. The approach will only be as resilient as its data governance, sensor calibration, and validation methods. It also shows why AI projects in complex environments should be evaluated on fit to context, not just predictive accuracy.
A large old tree with a complex canopy near Canberra. We used AI to extract and classify its 4,122 branches.Stanislav Roudavski / Alex Holland
Trees, birds and power poles
Molonglo hosts a once-thriving ecosystem that is now fragmented and damaged. Large old trees are increasingly rare. These trees, some more than 500 years old, provide complex canopy structures that are essential for bird nesting, foraging and roosting. As urban development expands and old trees die, the challenge is to fill the gap left by these giants.
Existing artificial habitat structures, including utility poles (left) and snags (right), cannot replicate canopy structures provided by a large old tree.Stanislav Roudavski / Alex Holland
Modified utility poles and relocated dead trees (or snags) have previously been introduced into the region as substitute habitat. These structures Original Postii/S0006320719300461" shape="rect">can provide important habitat features such as elevated perches, nesting boxes and bark that do not occur in planted tree saplings. However, it is very difficult to understand exactly which features of a large old tree are important to birds โ which limits the value of artificial structures.
Carefully analysing imagery and other data can help us discern these features. For example, we and our collaborators found that birds prefer small horizontal branches for perching and nesting.
From studying birds, we can learn their preferences for certain characteristics that have already been designed by trees. Our next challenge was to use this information to design better habitat structures.
Learning from trees
Our process used laser scanning and AI to recognise tree branches, assess how birds use them, then generates potential artificial structures and assesses how birds might use them.Stanislav Roudavski / Alex Holland
We used a process that involved data capture, predictive modelling and iterative design. AI and machine learning were indispensable in interpreting complex spatial data.
First, we mapped each tree by reflecting many millions of laser beams from each square centimetre of its surface to capture tree canopies as clouds of points. Then we used algorithms to identify and measure significant attributes such as orientation, size and linking of branches. A better understanding of bird preferences for these attributes can inform designs for artificial replacements.
Next, we developed statistical models to predict bird behaviour. These models were based on long-term observations of bird interactions led by Philip Gibbons at the Australian National University. By simulating how birds might use artificial branches, we could refine our designs to better meet their needs.
Reimagining artificial habitats
One version of an artificial tree that uses a lightweight structure of cables and rods to attach to an existing utility pole (right). The suitability visualisation shows inconveniently inclined branches as blue and near-horizontal ones in red. The thickness indicates exposure and therefore the ease of access. Brightness refers to the distance from the ground.Stanislav Roudavski / Alex Holland
To generate a variety of artificial tree crowns we developed further algorithms. Instead of judging the resulting designs by how much they resembled a tree to human eyes, we used our bird behaviour model to figure out how these structures might serve avian inhabitants.
Our additional goal was to create lightweight structures that are easy to install, reconfigure and remove. Our simulations showed that, compared to utility poles and snags, these structures can provide a significant increase in habitat suitability.
Returning to the field
We are currently building prototypes based on our designs, but the final step in this process will be field testing to find out what the birds think. Birds can provide feedback on the characteristics of artificial structures through their interactions with them. This testing will help make the designs even better. Design processes, even for non-human stakeholders like birds and trees, are currently dominated by human perspectives and expertise. Our findings show how broadening the scope of creative contributions and judgements can improve the design process. The outcomes of this design process can take the form of โcontinuous servicesโ, sustainably providing shelter or other resources. While we hope to build better artificial structures, it is important to remember that there is no true substitute for large old trees. We must also preserve the trees we have and plant more for the future.Broader implications for design
The principles of more-than-human design we used in Canberra also have broader applications. Many environments around the world face similar challenges. By rethinking current approaches to design and planning, we can create more inclusive and resilient environments for many different lifeforms. The essential change is to treat other species as innovators and expert participants in design. Extending existing efforts to communicate with whales, bats and honeybees, this approach uses AI to incorporate input from nonhuman lifeforms to produce new and better designs. Our case study shows how participatory approaches that include nonhuman beings can work around human biases. As a result, we unlock a far greater range of possible designs.Fair design
The world faces many urgent environmental crises. We need innovative, inclusive design approaches to meet this challenge. Trees are already excellent designers, just as birds are excellent judges of their work โ and if we include their input we can create better โmore-than-humanโ designs. We believe that using AI to give a voice to non-human stakeholders can lead to better solutions in which many species can live together. Our work in Canberra is an example of how participatory design can create more equitable and sustainable futures for all beings.We acknowledge the initiative of Darren Le Roux in researching and installing artificial habitat-structures to support biodiversity.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.
