About the competition

The CSIRO – Image2Biomass Prediction competition challenges participants to develop machine learning models capable of accurately predicting pasture biomass. The project addresses the existing difficulty farmers face in estimating the amount of feed available in paddocks, a task that is currently performed using slow or unreliable manual methods such, as plate meters or “clip and weigh”. Participants are provided with a professionally annotated dataset of Australian pastures, ground-truth measurements, and publicly available datasets such as Normalized Difference Vegetation Index (NDVI) values to enhance their models. The success of this project is intended to demonstrate a critical transition for the agriculture industry toward reliable, data-driven decision-making, moving past reliance on manual, subjective estimation methods.

“Precision agriculture is key to securing a sustainable future for food production.”

CSIRO

The competition is conducted by CSIRO, Australia’s national science agency, which is dedicated to solving scientific challenges to improve environmental resilience and food security. This research is undertaken in partnership with Meat & Livestock Australia (MLA), whose involvement ensures the project directly supports the red meat industry’s profitability and long-term sustainability.

Relevance

Accurate pasture biomass prediction is crucial for farmers to determine appropriate grazing times for livestock, addressing the existing difficulty of using slow manual methods like plate meters or “clip and weigh”. Achieving this accuracy ensures sustainable land productivity while also supporting animal welfare. By partnering with MLA, CSIRO aims to support the red meat industry’s profitability and sustainability. The effort aligns with maximizing accessibility and real-world impact and supports UN Sustainable Development Goal 2: Zero Hunger.

Technical details

The competition requires models to predict five specific biomass components: total dry biomass, dry green, dry dead, dry clover, and GDM. Total Dry Biomass is the sum of Dry Green, Dry Clover, and Dry Dead, while GDM is the sum of Dry Green and Dry Clover. The evaluation of model performance is based on a single, globally weighted coefficient of determination (R²) calculated across all target pairs, not by averaging R² for each target. The training data consists of images—all resized to 2000×1000, representing 70cm by 30cm tiles—and includes associated metadata like Australian State, dominant species, Sampling date, NDVI, and average pasture height. Inputs are images, and the desired output is a predicted biomass value (regression). To visually explain the relationships between the biomass components (Total Dry Biomass = Dry Green + Dry Clover + Dry Dead; GDM = Dry Green + Dry Clover), a diagram such as a drawing of sets should be included. Furthermore, an initial analysis shows strong correlations between the target components, suggesting that a multi-output model is appropriate. We plan to utilize a two-stream Deep Learning architecture: one stream processing the visual data from the images and a second stream processing the metadata to allow for a combined prediction.

The CSIRO Image2Biomass Prediction project aligns with UN Sustainable Development Goal 2: Zero Hunger by promoting sustainable agriculture and food security. By improving pasture management through data-driven insights, this initiative helps optimize resource use and enhance long-term environmental resilience.