Australia’s national science agency has turned one of grazing’s most tedious jobs into a global machine-learning contest, and the results suggest a smartphone photo may soon do work that used to require a farmer, a quadrat and a pair of shears.
CSIRO, working with Google Australia, Meat & Livestock Australia (MLA) and spatial-data body FrontierSI, has announced the winners of a worldwide Kaggle competition that trained artificial intelligence to estimate pasture biomass directly from images. The challenge drew nearly 100,000 model submissions from roughly 14,000 registrations across 109 countries, competing for a US$75,000 prize pool.
Teams from China, Vietnam and the United States took the podium in what CSIRO called the Image2Biomass Prediction Competition. Team 卷不动了 from China placed first, followed by Team dino series from Vietnam and Team embee from the United States, according to results reported by Mirage News.
From quadrats to computer vision
Measuring how much feed is standing in a paddock is deceptively hard. The usual methods involve manual sampling, cutting and drying grass, or walking transects with a rising-plate meter, all of which are slow, labour-intensive and easy to get wrong across a large or varied property.
The competition asked entrants to do it from pictures instead. As CSIRO explained when it launched the challenge in October 2025, participants trained models on pasture photographs paired with detailed field measurements, including plant height and vegetation indices that reflect pasture health and light reflectivity. The images spanned seasons, regions and species so models could not simply memorise one type of paddock.
“By combining images with field data, we’ve collated a dataset that allows AI models to learn in more than one way,” CSIRO Senior Principal Research Scientist Dr Dadong Wang said at launch.
The winning approaches show how varied the solutions were. China’s Team 卷不动了 reframed feed estimation as a counting problem rather than a single guess, which helped its model adapt to conditions it had not seen. Vietnam’s Team dino series leaned on spatial distribution and simulated environmental variation, while the United States’ Team embee combined several models into one system to reduce overfitting, Mirage News reported.
Small data, real farms
The most consequential finding for Australian producers may be about data, not accuracy. According to CSIRO, Dr Wang said the winning solutions showed reliable results were achievable using relatively small amounts of data, making the tools practical for real-world farming environments where conditions constantly change.
That matters because a model that only works after ingesting thousands of carefully labelled images from every region would never reach an ordinary farm. One that performs on modest data is far easier to fold into an app or an on-farm sensor.
The models do more than weigh grass. CSIRO says they can identify species composition and the amount of plant material available for grazing, distinguishing details such as dying grass and clover leaves. That opens the door to a shift the agency frames as moving from broad monitoring to targeted, site-specific decisions, applying fertiliser where it is actually needed and timing stock movements on data rather than a walk of the paddock.
“Accurately understanding feed availability and composition is fundamental to grazing management,” MLA Group Manager for Science and Innovation Michael Lee said, adding that the competition pointed to tools that could reduce reliance on manual measurement.
Google Australia framed the exercise as a bridge between research and the paddock. “By connecting CSIRO’s deep scientific expertise and MLA’s industry knowledge with the 26 million innovators on Kaggle, we’re putting a global AI community to work,” the company’s Partnerships Principal Scott Riddle said at launch.
Why it matters for Australia
Grazing systems cover roughly half of Australia’s landmass, and the red meat and livestock industry recorded turnover of $81.7 billion in 2022-23, according to MLA figures. Small percentage gains in how well producers match stock to available feed translate into large national numbers, both in profit and in land condition.
The approach also fits a wider pattern in Australian agtech. MLA is already funding data-driven grazing projects that blend in-field weigh stations, biomass data and climate datasets to benchmark paddock productivity. Cheap, photo-based biomass estimates would slot neatly into that stack, giving smaller producers a low-cost entry point to precision grazing that until now has favoured well-resourced operations.
There is a strategic dimension too. By owning the dataset and running the challenge, CSIRO keeps the intellectual groundwork for a nationally relevant tool onshore, even as the winning code came from overseas. Crowdsourcing the modelling through Kaggle let three partners tap a global talent pool at a fraction of the cost of building the models in-house.
The obvious caveats remain. Competition performance on a curated dataset is not the same as reliability on a windswept station in a dry season, and CSIRO has not committed to a product or a timeline. The winning models are a proof of concept, not a paddock-ready app.
Still, the direction is clear. The next test will be whether these approaches survive contact with real conditions, poor light, patchy connectivity and the enormous variety of Australian pastures, and whether a producer can one day point a phone at a paddock and trust the number that comes back. On the evidence of this challenge, that future is closer than the shears-and-quadrat present suggests.
Sources: CSIRO, New AI tools find smarter ways to measure pasture; CSIRO, Global AI challenge to transform pasture management; Mirage News, New AI tools find smarter ways to measure pasture; Meat & Livestock Australia, The red meat industry; MLA, Data-driven system optimising the forage base.
