Agricultural grazing systems cover around a quarter of the Earth's land surface and accurately estimating how much feed is available is critical for livestock productivity, land condition and long‑term sustainability.
However, pasture measurement has historically relied on manual sampling and field-based assessments, which can be time-consuming, costly and difficult to scale.
While satellite and other remote sensing approaches have helped broaden monitoring across large areas, high resolution on-ground digital photographs are one approach which could be used to calibrate existing systems, while also revealing new fine-scale features such as species mix and quality.
To this end, Australia's national science agency, CSIRO, in partnership with Google Australia and Meat & Livestock Australia Limited (MLA), launched a global 'Kaggle' challenge to advance the use of AI in agriculture.
The winners of the Image2Biomass Prediction Competition have now been announced, with Team 卷不动了 from China securing first place for an approach that improved accuracy by adapting to changing conditions.
Participants were tasked with training machine learning models to estimate pasture biomass directly from images, using data collected across different Australian regions, seasons and pasture types.
The winning teams demonstrated that advanced models can learn to extract meaningful information from images – such as the amount of plant material, including grass and other vegetation available for livestock to graze – and do so reliably across changing conditions.
This approach supports a shift from broad monitoring to targeted, site-specific management that pinpoints exactly where fertiliser or other interventions are needed.
With a US$75,000 prize pool, the competition attracted nearly 100,000 model submissions from approximately 14,000 registrations across 109 countries, highlighting strong global interest in applying specialised data science to real-world agricultural challenges.
CSIRO Senior Principal Research Scientist, Dr Dadong Wang, said the results were an important step forward for agricultural research, environmental monitoring and sustainable land management.
"Within a short period, competitors tested a wide range of approaches and refined their models in different ways, leading to major improvements in how accurately feed levels could be predicted across different regions, seasons and pasture conditions," said Dr Wang.
"The winning solutions showed that reliable results can be achieved using relatively small amounts of data, making these tools practical for real-world farming environments where conditions are constantly changing."
Rather than building solutions tailored to individual sites or seasons, top‑performing teams focused on enabling systems to perform reliably across different environments by recognising patterns in pasture and capturing fine botanical details in the images, such as dying grass or small clover leaves. This approach helped ensure predictions remained reliable even as landscapes, weather conditions and pasture composition changed.
MLA Group Manager – Science and Innovation, Michael Lee, said the outcomes highlighted growing opportunities to support producers with better information.
"Accurately understanding how much feed is available and what the feed is comprised of is fundamental to grazing management," Mr Lee said.
"The approaches demonstrated through this competition point to future tools that could reduce reliance on manual measurement and provide producers with faster, richer insights to support day‑to‑day decisions."
Google Australia's Partnerships Principal, Mr Scott Riddle, said the competition showed the value of connecting research, industry and the global technology community.
"By bringing together CSIRO's scientific expertise, MLA's industry knowledge and the global Kaggle community, this challenge demonstrates how partnerships can help bridge the gap between research and practical solutions," Mr Riddle said. 
CSIRO will now analyse the winning approaches in detail to inform future research and development, and will continue working with industry partners to explore how the most promising methods could be translated into practical, scalable pasture measurement tools.
This work has been supported by FrontierSI (previously known as the Cooperative Research Centre for Spatial Information).
The winning teams of the Image2Biomass Prediction Competition include:
- Team 卷不动了 from China took a novel approach by treating available feed as a counting problem rather than a simple estimate, enabling models to adapt to new conditions and improve accuracy on unseen data.
- Team dino series from Vietnam focused on understanding where feed appears within an image, estimating spatial distribution and using simulated environmental variation to strengthen performance.
- Team embee from the United States prioritised robustness by combining multiple models into a single system, reducing overfitting and delivering more consistent results across a highly variable dataset.
