Artificial intelligence weed detection makes mine recultivation easier

By Richard Pallardy

Proofminder, a Hungarian startup, is zeroing in on weeds using computer vision technology. The company specialises in leaf-level artificial intelligence across a wide range of crop species, from corn and cotton to oranges and pineapple, delivering hyper-precise, actionable insights for food producers and leading seed companies. In an Innowwide project, it turned its digital eyes towards a problem that bedevils the rehabilitation of mine sites in Australia: an invasive weed called Chilean needle grass (Nassella neesiana).

A pernicious grass

“Chilean needle grass is a quarantine weed. It is almost impossible to get rid of.”

— Foulcher

Chilean needle grass is a botanical plague in Australia. Native to South America, this seemingly benign weed was first discovered near Melbourne in 1934. The delicate-looking grass spread more widely beginning in the 1970s and became a problem for Australian ecosystems and farmers alike. It is now one of the country’s Weeds of National Significance.

The species is unpalatable to livestock and its namesake seeds can injure cattle and sheep. It reproduces rapidly, with a single plant releasing as many as 20,000 seeds in a season, allowing it to outcompete native vegetation. The seeds are persistent, too. They can remain dormant for more than a decade and germinate when conditions are right.

“It is just replacing native species — they cannot grow,” says Stuart Foulcher, Proofminder’s head of Australia operations.

Its prolific habits allow it to dominate the disturbed soil at former mine sites. From there, it easily moves to neighbouring agricultural land. Removal is now a pressing business concern. In addition to risks posed by the contamination of livestock feed and crops, property owners can be subject to steep fines and other civil liabilities if they fail to take steps to suppress the grass.

Seek and destroy

“It is very important to make sure that this grass is not present when mines are being re-cultivated—and for farmers to eradicate it too.”

— Foulcher

The problem of Chilean needle grass in Australia was a perfect use case for Proofminder’s technology. With funding through Eureka’s Innowwide programme, they refined their artificial intelligence detection software specifically to scan for needle grass at mine remediation sites there.

The grass can appear frustratingly similar to native grass species, especially in its early growth stages. Using highly calibrated drone flights over sample plots infested with the grass, Proofminder’s data scientists zeroed in on the correct parameters for capturing photographs that could then be analysed by the software. The photographs are so detailed that each pixel represents only a few millimetres.

The software can thus identify the unique characteristics of the plant with high accuracy and generate a map of where the grass is most prevalent. It can then be removed by hand or through pesticide application — also using drones.

“The drones can spray areas where the grass is dense and avoid areas where it is not, preserving biodiversity and using smaller amounts of pesticide,” Proofminder co-founder and Chief Technology Officer, Norbert Havas explains.

The project was the result of a collaboration between Proofminder and local Australian partners. “Precision Drones provided the flight and field expertise and relationships with the farms,” explains Foulcher. Foulcher coordinated the deployment of the technology in the field.

The origins of the technology

“There was a market response to our early work. We have scaled to multiple continents and work with some of the world’s top seed companies.”

— Havas

Proofminder first demonstrated its proof of concept in 2021. In a collaboration with Corteva Agriscience, the company showed that the technology could effectively discern between ragweed plants and sunflowers, allowing for targeted elimination of the pest plant.

Photographic information gathered by drone cameras is analysed by the algorithm, which generates maps that indicate relevant points in a field — weeds, disease outbreaks and other problems.

Some problems are relatively easy to spot. “Green on brown” detection can identify weeds in fallow fields, distinguishing them from surrounding soil before crops have grown up around them. “Green on green” detection, as in the case of the ragweed and sunflowers, is more challenging. Harmful plants must be distinguished from desirable plants based on sometimes subtle differences in leaf shape and colour.

“We are using drones and cameras to gather information from the field on a very detailed level,” Havas says. “We then feed this to our algorithm, which is where the magic happens.”

The company was named the most innovative startup by the Hungarian National Chamber of Agriculture for the proof-of-concept project. It has since refined its artificial intelligence technology for use in a variety of other agricultural applications. Proofminder has also gained international recognition, including selection as a semi-finalist in the SVG Ventures THRIVE Global Impact Challenge, winning the EBRD AgVenture Competition in 2024 and featuring in multiple short listings by EU-Startups and other reputable organisations.

Eye in the sky

“We can generate a report on a given field within hours.”

— Havas

Depending on the crop, Proofminder may use either drones, on-the-ground cameras or a combination of both to capture the images needed by its software. Drones are usually sufficient for row crops like corn and soybeans. Signs of disease or weeds are typically visible from above.

The company partners with drone service providers worldwide. “We give them the parameters and flight plans so they can do the job,” Havas enthuses. These providers then upload the images to Proofminder for analysis.

However, crops grown in orchards present an additional challenge due to the canopies of the plants, which can conceal problems with the fruit when viewed from the sky. In these cases, images can be captured on the ground, often by simply driving a tractor or truck through the rows with a GoPro or similar camera mounted on it. In some cases, smaller drones with forward-facing cameras can be piloted through the rows.

A closer look

“The core differentiator of our product is that we are gathering images that are a half centimetre per pixel or less.”

— Havas

Proofminder has distinguished itself by concentrating on accuracy. The ability of their software to read fine details in images of plants is what makes it such a perfect solution to the needle grass problem. Discerning the differences between the invasive weed and similar native grasses makes all the difference. Were it not for this capability, native grasses might be killed by pesticides as well.

A case study for the Innowwide project found that an Australian family-run farm spent four months of the year patrolling for Chilean needle grass and attempting to control it by manual application of pesticides — a dangerous procedure that affected their health. The productivity of their land was also reduced, because livestock could not be rotated between their properties due to fear of spreading the weed. The artificial intelligence technology helped the farmers to eliminate the weed from their properties in a safer, more comprehensive way.

“These models allowed for accurate spot spraying, reducing the amount of chemicals normally sprayed on the fields. The weeds within the fields sprayed were previously being reduced but Chilean needle grass needs multiple applications before it is fully eradicated,” Foulcher says.

This same technology has been more broadly applied in the service of corn detasseling, where the male flowers of corn plants are removed so that they do not self-pollinate the female flowers below. By removing the male flowers, the female flowers can be pollinated by desirable male flowers from neighbouring plants instead, leading to more productive hybrid varieties. Proofminder’s algorithm helps to identify missed tassels, allowing for their early removal and increasing the quality of the seed crop.

Proofminder continuously listens to growers and responds quickly to emerging field challenges, developing new AI models within days to address specific crop needs.

By combining rapid model development with scalable deployment, Proofminder helps make field operations more efficient. Its outputs are fully compatible with existing agricultural machinery and robotics, closing the loop of agricultural automation.

Proofminder’s software is touted to have yet more positive implications that will ripple through Australian agricultural properties and beyond.

---

Innowwide is funded by the European Union as part of the European Partnership on Innovative SMEs.

---