179 Predicting individual beef cattle water intake.

Abstract As global climate patterns shift and population growth accelerates, optimizing freshwater resource use is critical. Animal agriculture is frequently criticized for its perceived role in excessive water consumption; however, prevailing estimates of beef cattle water intake warrant reassessment. Existing beef cattle water intake standards are based on data collected in 1956, failing to account for decades of genetic progress and management improvements. Recent advances in precision livestock monitoring have provided unprecedented access to high-resolution animal data. However, such installations primarily record feed intake and lack the equipment to measure individual water intake. Machine learning models are well suited for leveraging large feed intake datasets to predict individual water intake in systems where feed intake is measured but water intake is not. We deployed RFID-enabled front-end weighing stations integrated with custom metered waterers at the Reymann Memorial Research, Outreach, and Education Center (REOC) in Wardensville, WV (2019) and the WVU Animal Sciences REOC in Morgantown, WV (2022) alongside Growsafe 8000 feed bunks. These systems facilitated the collection of daily body weight, feed intake, and water intake data at the individual animal level. From 2019-07-08 to 2024-02-16, we evaluated 2,020 animals for individual daily weight, feed intake, and water intake across both research locations. Data collection spanned 1,055 days, covering all months of the year. The dataset encompasses 11 breeds: Angus, Limousin, Angus cross-breed, Charolais, SimAngus, Hereford, Simmental, Charolais cross-breed, Black Hereford, Akaushi Wagyu, and Red Angus. Animal ages ranged from 162 to 753 days. The dataset includes water intake records from 1,255 bulls, 434 steers, and 331 heifers. Hourly climate data were collected from open-source climate stations and transformed into daily summaries of mean, maximum, minimum, and range. The resulting high-resolution dataset was used to train a Long Short-Term Memory (LSTM) recurrent neural network (RNN) model, well-suited for modeling longitudinal data. Features were engineered to capture temporal dynamics, cohort effects, and intake distribution patterns. Temporal features included cyclical encoding of month and day-of-year, polynomial transformations of test progression, and lag-based statistics to model intake trends. Pairwise distance metrics (Euclidean, Mahalanobis) and clustering methods characterized intake similarity within cohorts. These transformations enhanced the model’s ability to learn seasonality, group effects, and temporal dependencies in intake behavior. A 70:30 train-test split was used to evaluate model performance. After hyperparameter tuning, the model predicted individual water intake with an RMSE of 5.16 L, an MAE of 4.08 L, and an adjusted R² of 0.72. A generalization gap of 1.08 indicated no major overfitting. Our model can predict individual water intake in systems that only measure individual feed intake, enabling the identification and selection of water-efficient animals and improvement of water use in beef production systems.