Forestry biosecurity relies heavily on surveillance traps. Thousands of them are deployed around ports, nurseries, and high-risk sites across Australia, each collecting insects for identification. The problem? Actually identifying what’s in those traps requires trained entomologists examining specimens under microscopes. It’s time-consuming, expensive, and creates a bottleneck in our early detection systems.

That’s starting to change as image recognition AI learns to do the job.

The Identification Bottleneck

A typical trap servicing operation goes like this: field staff collect traps on a regular schedule (weekly, fortnightly, or monthly depending on risk). The trap contents get sent to a laboratory where entomologists sort through everything, identifying species and checking for anything unusual.

For common species, experienced technicians can identify things pretty quickly. But for potential new detections, you need expert entomologists. And when you’re dealing with thousands of specimens from hundreds of traps, it takes time.

During that time lag, if there’s actually a new pest in the trap, it’s out there spreading. Every week of delay between an insect hitting a trap and someone identifying it as a problem is a week for the incursion to expand.

In 2023, a brown marmorated stink bug detection in Melbourne was confirmed about 10 days after the trap was collected. That might sound fast, but during those 10 days, the local population could have increased substantially. Earlier detection would have made eradication easier.

What AI Can Do

Modern image recognition systems are getting remarkably good at identifying insects from photographs. The technology that can tell a golden retriever from a labrador can also tell a Ips grandicollis from an Ips hauseri (both are bark beetles, and they look very similar).

The process involves training neural networks on large sets of labelled images. You show the system thousands of photos of different insect species, telling it what each one is. The AI learns to recognise the visual patterns that distinguish one species from another.

Once trained, the system can examine photos of trap contents and identify what’s there. Not just at a family or genus level, but down to species—the level you need for biosecurity decisions.

Some systems now achieve over 95% accuracy on species they’ve been trained on. That’s comparable to human expert identification for many groups.

The Practical Implementation

Several Australian research groups and forestry organisations are working on practical implementations of this technology.

The basic workflow is straightforward: instead of sending trap contents to a lab, field staff photograph them in a standardised way. The photos get uploaded to a cloud system where AI models process them, identifying all the insects present. Anything unusual or concerning gets flagged for human expert review.

The University of Melbourne’s Forest BioSystems Lab has been developing exactly this kind of system for pine plantation pest surveillance. They’ve trained models on common Australian forest insects plus key exotic species we’re watching for.

Custom AI development for specialised applications like this involves some unique challenges. You need high-quality training data, which means properly identified insect specimens photographed in consistent conditions. You need models that can handle the enormous variation in how insects can appear depending on angle, lighting, and condition.

And you need the system to be conservative—it’s better to flag something for human review unnecessarily than to miss a genuine new detection.

The Training Data Challenge

Building accurate AI models requires lots of training data. For common species, that’s not too hard—there are plenty of museum specimens and reference collections to photograph.

For rare species or pest species that don’t occur in Australia yet, it’s trickier. You need to get specimens from overseas collections or coordinate with international partners. Some research groups are using specimens intercepted at borders—insects that were caught before they established but can now serve as training data for future detection.

There’s also the question of variation within species. Insects can look different at different life stages, in different seasons, or even depending on their host plant. Male and female of the same species sometimes look quite different. All this variation needs to be represented in the training data, or the AI will struggle with real-world specimens.

The iNaturalist project and similar citizen science databases have been valuable sources of training data. Millions of insect observations, many with expert-verified identifications. While these aren’t from biosecurity traps specifically, they help AI models learn what different species look like in various conditions.

Advantages Over Human Identification

AI identification isn’t necessarily better than expert entomologists in all cases, but it has some significant advantages.

Speed is the big one. An AI can process photos from hundreds of traps in minutes. A human expert might need days. That faster turnaround means quicker detection of problems.

Consistency is another factor. Human identification accuracy varies with fatigue, experience, and individual expertise. An AI model performs the same way whether it’s the first photo or the ten-thousandth.

Scalability matters too. There are only so many trained entomologists, and training new ones takes years. You can deploy an AI system anywhere you have internet connectivity, and it works 24/7.

And there’s an archiving benefit—every trap photo creates a permanent record. If identification standards change or new research suggests we should be looking for something we previously didn’t think was important, you can go back and reanalyse old photos. You can’t do that once physical specimens are disposed of.

Current Limitations

AI insect identification isn’t perfect, and there are real limitations to understand.

The systems struggle with rare species they haven’t been trained on. That’s a problem for biosecurity because new arrivals are by definition rare initially. If the AI hasn’t seen it before, it might misidentify it or just flag it as “unknown.”

Photo quality matters enormously. Images need to be clear, properly lit, and show diagnostic features. Insects photographed at odd angles or with poor focus might stump the AI even if a human could identify them.

Some insect groups are just hard. Tiny insects where important features are at the limit of camera resolution. Cryptic species that look identical but are genetically distinct. Damaged specimens where key identifying features are broken off.

And there’s still the issue that for formal regulatory purposes—like declaring an area infested with a quarantine pest—you generally need physical specimens examined by qualified entomologists. An AI identification alone might not be sufficient for official decisions.

Integration With Existing Workflows

The most practical approach seems to be using AI as a first-pass screening tool rather than a complete replacement for human experts.

AI reviews all trap photos and sorts them into: definitely nothing concerning, definitely needs expert review, and uncertain. The “definitely nothing concerning” category can be quickly approved by technicians. The other categories go to experts.

This dramatically reduces the expert workload while still catching everything important. Instead of examining thousands of routine trap samples, entomologists focus their time on the interesting or ambiguous ones.

Several commercial pest monitoring companies are already implementing this approach. Traps get serviced and photographed by field staff with basic training. AI does the initial triage. Experts handle exceptions.

Real-World Results

How well does this work in practice? We’re starting to get real-world data.

A trial in New Zealand pine plantations used AI to monitor for bark beetles. Over a six-month period, the system correctly flagged 87 out of 89 trap samples that contained species of concern. The two it missed were both severely damaged specimens that even human experts initially struggled with.

It also correctly cleared about 2,400 routine trap samples that contained only common, non-threatening species. That saved approximately 200 hours of entomologist time.

In Queensland, biosecurity surveillance for fruit flies has been testing AI identification with similar results. High detection rates, significant time savings, and better turnaround times for getting results to decision-makers.

The Next Generation

Current systems mostly identify one insect at a time from close-up photos. Next-generation systems are getting more sophisticated.

Some research groups are working on analysing photos of entire trap contents at once—identifying multiple species in a single image. This is harder (lots of overlapping insects, varying orientations, different scales) but would further streamline the workflow.

Others are developing systems that can work with smartphone photos taken in the field, allowing preliminary identification before samples even get back to a laboratory. This could enable rapid response to obvious new detections.

There’s also work on combining image recognition with other data. A system that considers what species are expected in a particular location and season, cross-references that with what it sees in the photo, and flags things that are unexpected rather than just things it can’t identify confidently.

Economics and Access

One challenge is that developing AI identification systems is expensive. It requires computational resources, programming expertise, entomological knowledge, and extensive training data collection.

Some research groups and universities are working on this, but their systems don’t always translate easily to operational use. Commercial companies are developing products, but these can be expensive and proprietary.

There’s a growing push for open-source approaches where training data and models are shared. If Australian organisations collaborate on building these tools rather than each developing their own, we’d probably get better results faster and cheaper.

The Global Biodiversity Information Facility and similar international initiatives are working on exactly this—shared infrastructure and open data for biodiversity monitoring, including pest surveillance.

What It Means for Biosecurity

AI-assisted insect identification won’t replace entomologists, but it will change how they work. Instead of routine identification work, experts will focus on confirming unusual detections, training and refining AI systems, and dealing with complex identification challenges.

The potential biosecurity benefits are significant. Faster detection of new arrivals. More comprehensive surveillance coverage because AI makes processing more traps economically viable. Better documentation and archiving of what’s in our environment.

We’re still in the early stages, but the trajectory is clear. In five years, AI-assisted identification will probably be standard practice for routine surveillance traps. In ten years, we might look back and wonder how we ever managed with purely manual systems.

For an industry that depends on early detection of threats, that’s a very welcome development.