Predicting individual water intake in beef cattle using longitudinal data and long short-term memory models.

Area of Science:

• Animal Science
• Data Science
• Environmental Science

Background:

• Water intake (WI) is crucial for beef cattle health and resource efficiency but is poorly understood.
• Current predictive models, like NASEM, use outdated data and lack individual animal resolution.
• Dynamic environmental factors significantly influence cattle water needs.

Purpose of the Study:

• To develop and evaluate a Long Short-Term Memory (LSTM) model for predicting daily water intake in beef cattle.
• To compare the LSTM model's performance against existing methods using longitudinal and engineered environmental data.
• To identify key environmental drivers of water intake in beef cattle.

Main Methods:

• Collected longitudinal data from 2,268 beef cattle across diverse systems (2019-2024).
• Engineered dynamic environmental features from NOAA and NASA data, including rolling deltas and temporal encodings.
• Trained and evaluated an LSTM model using animal characteristics, dry matter intake (DMI), and engineered environmental features.

Main Results:

• The LSTM model achieved high predictive accuracy (RMSE = 3.85 L/day, R² = 0.74) on training data.
• The model generalized well to unseen regional and non-regional drylot datasets (RMSE ≈ 4.20 L/day, R² ≈ 0.62).
• Models without engineered features or NASEM predictions showed poor performance and systematic underestimation.

Conclusions:

• Sequence-based models with dynamic environmental covariates significantly enhance beef cattle water intake prediction.
• Engineered features, particularly short-term environmental stress indicators like THI, are vital for accurate WI prediction.
• The developed LSTM model offers a scalable framework for water-efficient genetic selection and adaptive management in beef cattle systems.