Are seed potato fields ready for robots to detect and remove diseased plants?

In seed potato cultivation – the production of high-quality planting material for potato growers – selecting and removing diseased plants is still mostly done by hand. It is a specialised task requiring training and experience. But now, several developers are working on robots that use artificial intelligence to identify infected plants. If the technology proves reliable, it could mark a breakthrough in reducing manual labour, improving crop health, and boosting efficiency across the seed potato industry.

Next year, the first pre-series robot should enter the field. That is the ambition of several developers working on a potato selection robot. The arguments for continuing development are becoming stronger, and the core technology is already available. While further development is still required, current progress is promising. What can we expect from this technology, and what are developers focusing on? The Dutch agricultural magazine Boerderij investigated and discovered that multiple parties are actively preparing to bring artificial intelligence-based robotic technology for potato selection into practice.

Significant savings

Seed potato selection is a time-consuming task that also requires considerable expertise. Automation could help reduce costs and address labour shortages. The disease pressure is a key factor, but in general, seed potato cultivation requires 20 or more hours of selection work per hectare, with fields being selected 3 to 5 times per season. If that human effort can be significantly reduced through robotisation, it would justify a substantial investment.

Labour is one aspect, and other alternatives are also being explored, such as breeding for resistance, applying straw, or net cultivation, but the need for selection will remain. “If a robot could do this even better than a human, that would be a real step forward,” says Frank van der Werff, head of seed potato production at HZPC. Robotisation—if done well and reliably—could benefit seed potato cultivation, which is why various trading houses are not only following the development of selection robots with interest but are actively supporting it. At least half a dozen companies are working on a potato selection robot, and it seems likely that next year one, or possibly more, suppliers will offer robots that can significantly reduce the workload for potato selectors.

Following in the footsteps of tulip growers

In tulip cultivation, disease detection using an autonomous robot has gained rapid traction since 2020. Within just a few years, dozens of robots have been operating in tulip fields each spring, already saving a great deal of labour. In this case, it is the Selector developed by H2L Robotics in Delft. Although the underlying technology had been available behind the scenes for some time, it took quite a while before a practically usable selection robot reached the market. But once it was there — and proved to deliver quality results — growers were quick to invest.

What started in tulips is now happening in potatoes. Or perhaps: if it works in tulips, it can work in potatoes too? The success of the H2L Selector in tulips has encouraged developers to turn their ideas into practice. Creating a reasonably functional robot can happen relatively quickly, but achieving a high-performing machine — preferably better than a human — takes time, which makes development a substantial investment. The tulip area in the Netherlands covers 14,000 to 15,000 hectares, roughly half of the total flower bulb acreage. For seed potatoes, the Dutch area is around 38,000 hectares.

Whereas tulip cultivation is largely a Dutch affair, seed potato production is on a much larger global scale. This gives the selection robot much greater potential, which explains why now — with time and technology on their side — several parties are seriously working to bring a robot to market. In tulips, the robot has already surpassed human capabilities.

How will it work?

At present, developers are mainly focused on collecting large volumes of images of diseased and healthy potato plants. Behind the scenes — quite literally behind a computer screen — the core process takes place: analysing the photos to develop an algorithm capable of identifying diseased and healthy plants in other images as well. This is time-consuming. The expression of Y-virus infections varies by variety and must also be recognised accurately in different growth stages. So, there is a lot to photograph and analyse.

The expectation is that once the algorithms can reliably identify diseased plants in several varieties, it will become easier to recognise them in others too. In tulips, the leap was made immediately to an all-in-one autonomous robot system. While driving, it detects diseased plants and applies a few drops of a substance that kills the plant. It does not work particularly fast, but it runs 24 hours a day without human intervention. For potatoes, selection — at least in this phase — will be a separate operation. While driving, diseased plants are marked with dye, after which they are manually removed in a separate round.

Additional data

Gathering extra data is a by-product of robotised disease detection. Every potato plant is inspected individually. The robot can count the plants, but may also be able to monitor plant growth or map symptoms of drought, fertilisation, or damage during the same operation. Electronically recording the locations of diseased plants in the field makes it easier and more efficient to revisit these spots after scanning. Even more interesting: the location where a diseased plant is found often turns out to be the place where infected plants reappear in a subsequent selection round. This can signal the robot to scan those areas even more thoroughly next time. The data may also help identify the source of an infection or how it spreads. Speaking of data, information collected by the robot can later be used to analyse a problematic batch. Although it should not happen, there are cases where a batch approved by NAK (the Dutch General Inspection Service for agricultural seeds and seed potatoes) causes unexpected problems in the following crop. One of the puzzle pieces in such cases is whether the batch came from a grower where hardly any disease was detected — or from one where only after intensive selection the batch just passed inspection. Measuring is knowing.

DigiAgro3 from 5Groningen testing ground

Scale model 1:3 of the ‘Pieperkieker’ used by Hanze University students to test their navigation software. Photo: Hanze University of Applied Sciences

The ‘Pieperkieker’ first saw the light of day in 2020 as part of a pilot project from the 5Groningen testing ground.

Within the Digital Society Hub, students from different programmes collaborated with Groningen-based Advanced Climate Systems and a lecturer-researcher to develop an autonomous vehicle capable of detecting diseased potatoes using camera technology. A YouTube video still available online shows Advanced Climate Systems director Sietse Damstra saying the results were promising, though considerable development time would still be required to make the system fully field-read In 2024, DigiAgro3 was launched under the leadership of TNO — a consortium of companies and research institutes in the northern Netherlands focused on developing robots for agriculture and horticulture. Potato selection is one of its key themes, and the Pieperkieker has been incorporated into this.

Lecturer-researcher Jos Bredek of Hanze University: “The Pieperkieker is fully electrically powered. At the moment, the chassis is being modified to make the vehicle more robust. The Pieperkieker is being developed to drive completely autonomously and be remotely controlled. Next, we will continue training multiple algorithms. This is an incredibly educational and motivating project for students. Croptimal is also involved, and we are supporting each other in development. That prevents us from reinventing the wheel in isolation.”

2 varieties already well identified

Left: Arnout van der Ploeg, right: Han Hilbrands from Croptimal. Centre: Jacco Tuma (Tuma Mechanisation). The current focus is on data collection, processing and prototype testing. Photo: Mark Pasveer

Croptimal is a start-up founded last year. It originates from Doorgrond, a company specialising in precision farming technology and experienced in working with camera systems that recognise plants.

Jeroen Wolters is one of the initiators and partners in the company. “Last year, we scanned our first potato plants and for the varieties Spunta and Fontane, our algorithm was already able to distinguish diseased from healthy plants with 80% accuracy. Of course, we are working to improve this further — and whether that figure holds up in practice, we will only know for sure later this season. That is when the prototype of the selection robot will be put to work for the first time,” Wolters explains.

So far, most testing has been theoretical. The current prototype handles 4 rows, but we expect to launch a pre-series version in 2026, which will be an 8-row model. Each row is equipped with a camera, and initial results suggest that together with the software, the system will be fast enough to operate at 8 kilometres per hour. That speed may not always be used, but it is important that the detection and marking technology does not limit the working speed. The scan produces a digital map showing exactly how many diseased plants are present and where they are located. A follow-up crew can then locate and remove them with minimal time loss. After all, there is no need to walk or drive along a row with no diseased plants.

Croptimal has partnered with Jacco Tuma in Ulrum, who handles the mechanical construction. This currently involves several selection carts equipped with cameras to collect as much data as possible at various locations in the Netherlands this year. In principle, a scanner could also be mounted on a tractor frame, but practice clearly shows a strong preference for as lightweight a vehicle as possible. At present, image recognition is the foundation for detecting virus infections in potatoes, but other sensor technologies could potentially offer valuable additional data. For that reason, Croptimal has joined the PhytoPhotonic network — an international collaboration of research institutions, universities, and companies focused on applying optical sensor technology, including for plant disease monitoring.

Based on Wageningen expertise

During a lightning visit on Monday 12 May, Minister Wiersma observed precision applications, including robotic disease detection in tulips and potatoes. Photo: Anne van der Woude

Wageningen University & Research has been involved since 2007 in a project aimed at developing a ‘tulip robot.

At the time, the project failed to gain traction due to insufficient support from the business community. In 2010, former project leader Jan Kamp proposed applying the technology to seed potato cultivation. The initial results were promising, but again, there was too little industry interest to seriously invest in further development. It was not until 2015 that a partnership was formed with Kverneland, LTO / BO Akkerbouw, NAK, Agrico and HZPC, and a PPP (public-private partnership) was launched.

The original Wageningen research began with multispectral cameras — a technology based on light measurements — but with the rise of image recognition and AI, the project has entered a new phase. In 2024, Kverneland announced its commitment to developing a disease detection robot as part of the AGROS-II PPP project. This initiative includes specialists from Wageningen University & Research (WUR), potato companies HZPC, Agrico, Meijer Potato and inspection service NAK. The project builds on knowledge and experience gained from previous Wageningen-led projects and now includes H2L Robotics, which is working toward a functional robot.

Data are currently being collected primarily at the WUR experimental farm in Lelystad. Project leader Corné Lugtenburg explains: “We scan plants that NAK inspectors have labelled as diseased. This way, the robot can potentially match the inspectors’ accuracy. But we are also developing techniques to detect disease even before the inspector can. If successful, it will allow fields to be cleaned up earlier, helping to prevent further spread.” H2L Robotics, known for its tulip selection robot ‘Selector’, is also scanning plants this year at eight seed potato farms — four in the Wieringermeer and four in Flevoland — to ‘feed’ the algorithms and develop a working robot. A first pre-series version is expected to be market-ready next year.

In tulips, it took 15 years

The autonomously operating Selector robot is now widely used in tulip cultivation. Its success in tulips is a driving force behind expanding the technology to other crops. Photo: Andrea van Schaik

Since 2020, tulips have been successfully selected by robot on a commercial scale. The technology gained momentum rapidly over just a few years. Still, this breakthrough was a long time coming.

The ability to detect diseased tulips using camera technology already existed over 15 years ago. “Several tonnes needed for ‘Disease Detection Robot’,” headlined the Westland edition of Rodi.nl in September 2015. After a successful lab trial, the PPO launched a tulip disease detection project in 2008. “The particularly challenging yellow and white cultivars are being examined in this project using sensitive cameras for signs of mosaic virus (TBV),” the press release at the time stated. What the sensitive cameras captured was left vague, but in layman’s terms, these were images of tulip leaves analysed by smart software based on visual appearance. The result: a conclusion on whether a plant was diseased or healthy.

In September 2011, Boerderij reported that the tulip disease detection robot was nearly as effective as experienced disease scouts. That was the conclusion of PPO Bulbs, Nursery Stock and Fruit in Lisse after several years of developing the technology. Ton Baltissen, now project leader at Cropey, was at the time the WUR project lead for tulip disease detection. “Even 20 years ago, there were fewer people with the skills and knowledge to scout tulips for disease. That was the reason for starting the project — to replace humans with alternative technologies. We even tried using sniffer dogs,” Baltissen recalls, adding that the entire process took considerable time. You can tell from a tulip’s appearance whether it is diseased — but how certain can you be? Validation came from digging up the bulb and performing DNA testing. Around 2008, the technology had advanced to the point that the camera actually outperformed the two best inspectors in recognising diseased tulips.

Baltissen: “At the time, we estimated we would need €400,000 to develop a working robot. A financier was found — a company in North Holland that was later acquired by a Japanese firm. But ultimately, the project stalled, probably because there was not enough pressure among growers to push forward. Only 15 years later was the time finally right.”

Selecting 7 hectares per hour

If a drone can accurately detect diseased plants, it could result in enormous labour savings and a relatively modest investment. Photo: Ruud Ploeg

Croptic is a Belgian start-up with considerable experience in using drones to map crops.

It is mainly used for herbicide spraying, where drone imagery guides the field sprayer to apply product only where needed (spot-spraying). In potatoes, this technology is ideal for tackling problem weeds that survive blanket treatments — such as nutsedge, thorn apple, chickweed, thistle — and volunteer potatoes in the next crop.

“Last year, Croptic began its first trials using drones to detect diseased potato plants,” explains Lander Pauwels, one of the company’s partners and co-founders. According to Pauwels, their drone technology was able to identify 85% of diseased plants last year and assign them a precise GPS label. That percentage still needs to improve, but Pauwels is confident they will succeed. This year, as many drone flights as possible are being conducted with several seed potato growers in Belgium. They fly once before the manual selection round and again afterwards.

The difference between the two sets of data reveals which plants were identified as diseased — and this is used to train the algorithm. Once the algorithm is sufficiently capable of recognising diseased plants, the aim is to feed the data into a sprayer that uses spot-spray technology to mark the plants with dye. The drone can scan about 7 hectares per hour. Croptic has no ambition to develop equipment themselves. If the potato scanning system proves effective, Croptic will bring it to market in the same way they currently provide weed detection — as a service. They aim to have the system ready for practical use by early 2026, starting with the lower seed potato classes.