PSX-27 Machine learning prediction of individual beef cattle grazing dry matter intake.

Abstract Accurate estimation of individual dry matter intake (DMI) in pastured beef cattle is crucial for grassland management, genetic improvement, and sustainable animal agriculture. Current methods for measuring DMI are labor-intensive, costly, and spatially demanding, with limited options for cattle producers. To address this, we developed a machine learning approach to predict DMI using only body weight, water intake, and open-source climate variables. Pasture data were collected from 2021 to 2023 in Wardensville, WV at the West Virginia University Research, Outreach, and Education Center. Each year, 36 crossbred steers were phenotyped for Residual Feed Intake (RFI) in a drylot, and after a grazing acclimation period, were rotationally grazed through seven 0.05 ha plots. Daily weights and water intake data were collected throughout the experiment using an RFID-equipped front-end scale and metered waterers. Fecal samples were collected to estimate individual DMI using an inert marker. This process was repeated as often as forage growth allowed from May–November, with four grazing tests in 2021, five in 2022, and three in 2023. A pasture training dataset was created from daily bodyweight, water intake, marker-derived intakes, and local climate data. This dataset was combined with a corresponding dataset containing the same features, but with ground truth DMI measured at a Growsafe 800 feed bunk, to train a Gaussian Process Boosting (GPBoost) model—a machine learning approach—for predicting individual DMI. The model was initialized, trained, and optimized in MLflow and Docker. Missing data were handled through a robust imputation strategy, including linear interpolation, Piecewise Cubic Hermite Interpolating Polynomial (PCHIP), K-Nearest Neighbors (KNN), and forward/backward filling techniques. Data were standardized and outliers were detected using interquartile range (IQR) and Z-score methods. An 80:20 train-test split was used for model validation. Model complexity was controlled by optimizing tree depth, number of leaves, learning rate, and regularization parameters, ensuring balance between generalization and predictive accuracy. The Gaussian process component enabled structured residual modeling, capturing latent dependencies across repeated measures for individual animals. Our GPBoost model trained on both drylot and pasture data was able to predict individual grazing DMI with an RMSE of 1.65 kg and an MAE of 1.2. Our model outperformed the current NASEM equations for pasture intake by 70%, demonstrating that this machine learning approach provides a more efficient and accurate method for predicting individual DMI in pastured beef cattle. This work represents a novel, scalable solution for researchers and producers seeking to estimate individual DMI with greater accuracy, improving both efficiency and sustainability in animal agriculture. Our model can be applied to season-long grazing datasets, allowing access into questions regarding differences in barn and pasture-derived RFIs, grazing efficiency and methane emissions, and lifetime ecological impacts of grazing cattle.