Automated trap already reveals soil life in greater detail than expected

How can computer vision, image processing and artificial intelligence help to map soil life without harming it – and in doing so contribute to biodiversity conservation?

That question lies at the heart of an ongoing research project led by, among others, Marcel Polling and Gerwoud Otten. Polling is a molecular ecologist at Wageningen University & Research (WUR), specialising in animal ecology, and serves as project leader of Technological Innovations for Ecosystem Monitoring (TIEM). This knowledge development project is funded by the Ministry of Agriculture, Fisheries, Food Security and Nature (LVVN) and runs from 1 January 2025 to 31 December 2028. One of its key aims is to develop advanced sensor systems capable of continuously monitoring biodiversity in a broad sense.

Within the project, Polling works closely with Gerwoud Otten, a researcher in computer vision and robotics at WUR. As soon as both researchers start talking about the project – and about the first prototype of their automated trap for identifying and counting soil insects – their enthusiasm is unmistakable.

Polling explains: “What makes this so appealing is that nobody specifically asked us to do it this way. It really is our own initiative, and it is incredibly interesting to bring different areas of expertise together.”

Otten adds: “For me, what makes it particularly interesting and enjoyable is that this research is very different from the plant and fish research we usually do.”

Too little time and too little expertise

Soil fauna is traditionally monitored using pitfall traps – cups with a perforated lid, filled with ethanol, that are buried in the study area. Polling explains: “By soil fauna we mean species such as ground beetles, spiders and springtails. Ground beetles in particular are a very good environmental indicator and a useful measure of biodiversity at a specific location. Their presence is influenced by factors such as temperature and moisture, which means the placement of the traps partly determines how many beetles you catch. Weather conditions also play a role.”

Finding suitable locations for the traps and installing them is time-consuming. Researchers sometimes place up to 50 traps per day at different sites. Once the traps are filled with captured insects, they have to be collected and transported to Wageningen, where experts analyse the contents – another extremely labour-intensive task. A single trap usually contains many different animal groups, each of which requires a specialist to identify them.

“Besides being painstaking work, those specialists are simply no longer available in sufficient numbers,” Polling says. “On top of that, some insects become very difficult to recognise with the naked eye after spending a long time in ethanol. To analyse species diversity, we therefore carry out DNA sampling on the contents of the traps, but that is also highly labour-intensive.”

Combining research ambitions with vision and AI

“From a research perspective, the current pitfall-trap method is far from ideal,” Polling continues. “Ideally, we would like to observe and count animals directly in the field – preferably continuously, and in a non-destructive way. It is easy to see that this is no longer feasible if everything has to be done manually.”

In an attempt to automate the counting process, a colleague pointed Polling and Otten to research on spotted wing drosophila in Portugal. “That involved a compact, solar-powered trap that attracted fruit flies with a lure and automatically took photographs of them. We buried a similar trap in the ground. However, it missed about 90 per cent of the soil insects entering the trap, because it only took a photo every 30 seconds. On top of that, the image quality was poor due to a low-quality sensor. ‘I can do better than that,’ Gerwoud said.”

Otten elaborates: “At the top of our wish list were high-resolution images, so we could photograph a relatively small circular surface with a diameter of 3.5 centimetres in high resolution. Those images are essential for training algorithms, at least in the longer term. That turned out to be quite a challenge, because we were starting from scratch. There is nothing comparable available worldwide. We tested various camera types and eventually found a 20-megapixel camera that produces RGB images and is also affordable. The next challenge was the lens. It has to be able to zoom under low-light conditions and produce razor-sharp images of insects of very different sizes – from soil mites of just a few millimetres to insects several centimetres long. We also see harvestmen, earthworms, slugs and occasionally even mice appearing in front of the camera. For lighting, we currently use a small lamp that remains on continuously. Thanks to the design of the trap, the camera view is largely shielded from daylight. This gives us consistent light levels day and night and under different conditions, allowing us to create uniform images. Our original idea was to combine an RGB camera with an infrared camera and infrared lighting, so insects could be detected in the dark without being attracted by visible light. Only a brief flash would then be needed when taking the high-resolution images. That remains an ambition, but so far the attractive effect of the small amount of light visible on the outside of the trap appears to be very limited.”

Technology defines the size of the trap

The height, diameter and material of the automated trap were primarily determined by practical considerations. The height is based on the distance required between camera and lens to capture sharp images of the various animal species.

Polling explains: “The diameter is the same as that of conventional pitfall traps: around 11 to 12 centimetres, because we use a soil auger with that diameter to drill the holes. Roughly half of the trap, which is 70 centimetres high in total, is underground.”

Above ground, the trap houses the camera and lighting system, as well as a Raspberry Pi computer with 128 gigabytes of storage. Below ground is the mechanism that centres the animals beneath the camera and then collects them in ethanol.

Building the trap

Otten built the first trap himself using a PVC tube. “At ground level, there are openings in the tube through which insects can walk into the trap. Via a funnel, they fall onto a circular, white, illuminated flap at the bottom centre of the trap. The camera takes two images per second of this flap. The images are not stored unless one or more insects are detected. Detection is fully automated by mathematically subtracting successive images and then further processing and filtering the result. When an insect is detected, the image is saved. After three consecutive detections, a small servo motor – borrowed from the model-building world – opens the flap, allowing the animal to fall into a container filled with ethanol. That sample is analysed for DNA by Marcel’s team. While we are already able to automatically identify the type of animal, we cannot yet determine the species automatically. For that, DNA analysis is still required.”

Surprising results already

Although the automated trap is still a very early prototype and had only been operational for a few months by the end of November 2025, both researchers are pleasantly surprised by the results so far.

“The most surprising aspect for me is the image quality,” says Gerwoud Otten. “Even the smallest animals are photographed so sharply, under a wide range of conditions, that identification with the human eye is virtually always possible – including soil mites, which are extremely small.”

Marcel Polling adds: “What I find most surprising, and at the same time impressive, is that we are already able to visualise even the tiniest soil organisms in such great detail. In terms of image quality, this is far more than we had hoped for. I also think the current setup is very well suited to assessing the biodiversity profile of a specific area. And what is most rewarding about this research is how many different disciplines and aspects come together. Entomologists, molecular ecologists like myself, and specialists in vision, robotics and artificial intelligence who combine knowledge, software and technology. In my view, that is something you can really only do in Wageningen. I therefore invite companies and other organisations to get in touch if they are interested in the technology and algorithms we are developing. Of course, there are still ambitions and improvements to be made – especially now that we have seen how quickly and effectively the first prototype performs. And those, too, we will solve together.”

Next steps and scaling up

The additional ambitions are both ecological and technical in nature. Polling explains: “We would like to make the automated trap more animal-friendly by releasing the animals back into the environment after photographing them – for example by guiding them through a small tube. That requires technical adjustments and is only possible once the recognition algorithms are sufficiently mature. We also need access to large numbers of images and datasets. At the moment, an entomologist still annotates the images manually, and we want to automate that process. However, there is currently no openly accessible model of sufficient quality to do so.”

Otten adds that he would like to further refine image filtering and replace continuous lighting with flash lighting. “That would make the trap more energy-efficient and could allow us to power it with solar energy. At present, we still rely on a continuous low-voltage power supply. I would also like to apply edge computing, so that animals can be identified in the field using AI. Finally, we want to further develop the trap design into a more compact model that can be produced at scale. We may be able to do that ourselves, but there could also be external partners who are interested.”

Researcher