Cutting-Edge ML for Pasture Management

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Aug 4 2024

In an article published in the journal Nature, researchers evaluated the minimum data required to train a machine-learning decision-support tool (DST) using satellite images for accurate pasture biomass prediction. They compared the DST's estimates against a calibrated rising plate meter (RPM) across various training datasets.

The authors demonstrated that using satellite-based DST with minimal training data could achieve accuracy comparable to the RPM, the gold standard for pasture biomass monitoring.

Background

Pasture-based dairy systems are appealing due to low initial infrastructure investment and minimal reliance on market commodity prices. Despite their benefits, including positive environmental impacts and high-quality food production, effective grazing management remains crucial for profitability. Traditional methods for monitoring pasture biomass are labor-intensive and often perceived as inaccurate, leading to hidden losses in pasture management.

Recent advancements in machine learning and satellite remote sensing offer new DST for pasture monitoring. However, their on-farm application is limited due to concerns about accuracy and the need for substantial training data. Previous studies have developed methods for estimating pasture biomass using satellite imagery but lacked large-scale validation, particularly in regions with diverse grass species.

This paper addressed these gaps by evaluating the necessity and amount of additional training data required for accurate pasture biomass estimation using a satellite- and machine learning-based DST. The study quantified the minimal data needed to improve DST accuracy, focusing on precision grazing management and considering different pasture compositions.

Methods and Data Collection

The research was conducted on 14 commercial dairy farms across the coastal regions of New South Wales (NSW), Australia, from May 2022 to April 2023. The farms, covering a total of 2080 hectares, were located in the mid-coast (Taree and Tocal), north coast (Casino), and south coast (Bega). These regions typically had a mild temperate climate with moderate to high rainfall. Selected farms operated on annual ryegrass–kikuyu grazing rotation, with ryegrass sown in early autumn and kikuyu dominating during the summer.

Pasture biomass was monitored weekly using a calibrated RPM, with measurements taken from fixed and pre-grazing paddocks. Biomass was assessed through linear regression equations derived from monthly pasture cuts. Additionally, pasture botanical composition was visually assessed and categorized based on grass species.

The researchers aimed to evaluate the training requirements of a DST provided by Pasture.io, which used satellite imagery and machine learning to estimate pasture biomass. Remote sensing data was sourced from the PlanetScope CubeSats and Copernicus Sentinel-2 missions, providing multispectral data integrated with weather, soil, and farm management information.

Results and Evaluation

The field data collection for RPM calibration involved monthly pasture biomass cuts, revealing a mean absolute error (MAE) of 356 kg dry matter (DM)/hectare (ha) and normally distributed residuals. Weekly pasture biomass was calculated using a calibrated RPM, with the south coast showing the highest variation. The model evaluation showed that the initial training phase did not improve modeling performance significantly, but accuracy increased consistently with the inclusion of more weekly data.

The north-coast region had the lowest MAE, with mid and south-coast regions showing higher errors. Evaluations demonstrated that using pre-selected paddocks for model training improved prediction accuracy across the grazing platform. Despite the model's difficulty predicting higher canopy biomass, systematic training data inclusion reduced errors across all terciles. DST had lower accuracy for rainfed paddocks compared to irrigated ones but showed error reduction with more training.

Kikuyu-based pastures had approximately 20% lower MAE initially compared to annual ryegrass-based pastures, but the error reduced significantly from the second week onwards. The study confirmed that systematic incorporation of RPM data into the DST enhanced prediction accuracy for pasture biomass across various regions and conditions.

For model training and evaluation, pasture biomass data was split into four training sets: 25% (week 1), 50% (weeks 1 and 3), 75% (weeks 1, 3, and 4), and 100% (weeks 1 to 4). The DST's estimates were evaluated against RPM data using root-mean-square error (RMSE), MAE, and RMSE to standard deviation ratio (RSR). The authors also validated the DST's accuracy across different training datasets to enhance precision grazing management.

Conclusion

In conclusion, the researchers demonstrated that a machine learning-based DST for pasture biomass prediction could achieve accuracy comparable to a calibrated RPM with minimal training data. By training the DST with data from just 10% of the farm area at fortnightly intervals, precision in biomass estimates was significantly improved.

This approach highlighted the feasibility of using satellite imagery and minimal in-field data to enhance pasture management. The findings paved the way for the broader adoption of remote sensing technologies in dairy systems, promising efficiency and accuracy in grazing management. Future research should address potential overfitting and refine practical implementation strategies.