Organizations: Applied Artificial INtelligence and Intelligent Systems (AAIINS) Laboratory, Dhaka 1217, Bangladesh · Department of Computer Science and Engineering, United International University, Dhaka 1212, Bangladesh · Department of Data Science and Artificial Intelligence, Monash University, Melbourne 3800, Australia · Department of Software Systems & Cybersecurity, Monash University, Melbourne 3800, Australia · Energy and Resources Institute, Faculty of Science and Technology, Charles Darwin University, NSW 2000, Australia · Faculty of Science and Technology, Charles Darwin University, Casuarina, NT, 0810 Australia · School of Engineering and Energy, Murdoch University, Murdoch, WA 6150, Australia · Energy and Resources Institute, Faculty of Science and Technology, Charles Darwin University, Darwin, NT, 0810, Australia
Precision livestock farming requires accurate and timely heat stress prediction to ensure animal welfare and optimize farm management. This study presents a physics-informed digital twin (DT) framework combined with an uncertainty-aware, expert-weighted stacked ensemble for multimodal forecasting of Core Body Temperature (CBT) in dairy cattle. Using the high-frequency, heterogeneous MmCows dataset, the DT integrates an ordinary differential equation (ODE)-based thermoregulation model that simulates metabolic heat production and dissipation, a Gaussian process for capturing cow-specific deviations, a Kalman filter for aligning predictions with real-time sensor data, and a behavioral Markov chain that models activity-state transitions under varying environmental conditions. The DT outputs key physiological indicators, such as predicted CBT, heat stress probability, and behavioral state distributions are fused with raw sensor data and enriched through multi-scale temporal analysis and cross-modal feature engineering to form a comprehensive feature set. The predictive methodology is designed in a three-stage stacked ensemble, where stage 1 trains modality-specific LightGBM 'expert' models on distinct feature groups, stage 2 collects their predictions as meta-features, and at stage 3 Optuna-tuned LightGBM meta-model yields the final CBT forecast. Predictive uncertainty is quantified via bootstrapping and validated using Prediction Interval Coverage Probability (PICP). Ablation analysis confirms that incorporating DT-derived features and multimodal fusion substantially enhances performance. The proposed framework achieves a cross-validated R2 of 0.783, F1 score of 84.25% and PICP of 92.38% for 2-hour ahead forecasting, providing a robust, uncertainty-aware, and physically principled system for early heat stress detection and precision livestock management.
Figures & tables
Study
Multimodal Fusion
Physics- Informed
Continuous CBT Forecasting
Uncertainty Quantification
Personalization
Becker et al. [ 25 ]
✓
✗
✗
✗
✗
Georgiades et al. [ 26 ]
✗
✗
✗
✗
✗
Woodward et al. [ 27 ]
✗
✗
✗
✗
✗
Daniels et al. [ 28 ]
✗
✗
✗
✗
✗
Li et al. [ 22 ]
✓
✗
✓
✗
✗
Chapman et al. [ 24 ]
✓
✗
✗
✗
✗
Table 1: Comparison of the Proposed Framework with State-of-the-Art Methods.
Figure 1: High-level architecture of the proposed multimodal forecasting framework. Raw data from cow-worn and environmental sensors undergo temporal alignment and preprocessing before being fed into a Physics-Informed Digital Twin. The digital twin dynamically couples a behavioral Markov model, a thermodynamic ODE, a Kalman filter, and a Gaussian process to generate uncertainty-aware physiological and behavioral state features. These are fused with multi-scale engineered features to train modality-specific experts and an optimized meta-model, ultimately yielding continuous 2-hour ahead core body temperature (CBT) forecasts and heat-stress classification. Copyright Free Cow Image source: Freepik 1 1 1 https://www.freepik.com/free-photo/cow-grazing-green-meadow_11244779.htm .
Figure 2: Overview of the proposed pipeline. Sensor data from eight modalities are aligned and preprocessed, then passed through a physics-informed DT that models thermal and behavioral dynamics. Multi-scale features are engineered and fed into modality-specific base models. An expert-weighted ensemble fuses predictions, with bootstrap-based uncertainty quantification yielding final CBT forecasts and heat-stress risk estimates.
Figure 3: Internal Computational Loop of the physics-informed DT. The diagram illustrates the recursive state estimation cycle performed at each time step t . It unifies the mechanistic prediction (ODE), behavioral context (Markov), sensor correction (Kalman Filter), and uncertainty quantification (Gaussian Process) into a closed-loop adaptive system.
Table 2: Parameters of the physics-based thermal ODE (Eq. 8 ). Each parameter’s status (fixed, calibrated, or learned per cow) and source are specified in the text.
Figure 4: The overview of the hierarchical fusion strategy. Modality-specific “Specialist” models ( M1…Mn ) are trained independently. Their validation performance ( R2 ) determines their contribution weight ( Wi ). These weighted predictions are concatenated with global context features ( Fglobal ) to train the Meta-Regressor ( fmeta ), which outputs both the final forecast and uncertainty bounds.
Model
MAE ↓
RMSE ↓
R2↑
Linear Regression
0.1737 ± 0.0293
0.2112 ± 0.0323
0.4571 ± 0.1688
Random Forest
0.1712 ± 0.0125
0.2206 ± 0.0181
0.4184 ± 0.0669
Decision Tree
0.1576 ± 0.0171
0.2021 ± 0.0262
0.5025 ± 0.1393
Gradient Boosting
0.1552 ± 0.0122
0.1987 ± 0.0210
0.5242 ± 0.0984
XGBoost
0.1538 ± 0.0120
0.1972 ± 0.0203
0.5909 ± 0.0954
SVR
0.1793 ± 0.0227
0.2259 ± 0.0290
0.3908 ± 0.1130
Table 3: Performance Comparison of Baseline Models and the Proposed Multimodal Model.
Figure 5: Modality Prediction Correlation Matrix. This heatmap illustrates the Pearson correlation coefficients between the predictions of the modality-specific base learners. The high correlation between dt_features and phys_cbt (0.67) validates the DT’s physiological fidelity, while lower correlations in environmental features confirm the model diversity required for the stacked ensemble.
Figure 6: CBT Prediction Over Time. The model’s predictions (red dashed line) accurately track the true CBT values (blue solid line) and effectively capture the critical crossing of the heat stress threshold (black dotted line at 38.8 ∘ C). This demonstrates the model’s capability for timely early warnings.
Feature Group
MAE ↓
RMSE ↓
R2↑
phys_cbt
0.1830 ± 0.0145
0.2313 ± 0.0190
0.3786 ± 0.0644
behav_ankle
0.2340 ± 0.0159
0.2925 ± 0.0155
0.3459 ± 0.0330
behav_immu
0.2349 ± 0.0177
0.2923 ± 0.0176
0.2578 ± 0.0422
behav_uwb
0.2351 ± 0.0195
0.2943 ± 0.0217
0.1429 ± 0.0179
env_weather
0.2133 ± 0.0112
0.2600 ± 0.0136
0.2147 ± 0.0321
prod_milk
0.2372 ± 0.0144
0.2978 ± 0.0159
0.0513 ± 0.0369
Table 4: Ablation Study on Feature Group Importance.
Metric
With DT
Without DT
Improvement
MAE
0.1147
0.1385
+17.22%
RMSE
0.1435
0.1756
+18.29%
R2
0.7835
0.6927
+13.06%
PICP
92.38%
86.52%
+6.77%
Table 5: Ablation Study on the Digital Twin Component.
Figure 7: Comprehensive Regression Analysis. (Top-left) Prediction vs True Values scatter plot with color gradient indicating prediction density and red dashed line representing perfect prediction. ( Top-middle ) Error Distribution histogram with KDE overlay showing the distribution of prediction errors. ( Top-right ) Residual Distribution plot with trend line indicating no systematic bias. ( Bottom ) True Values and Predictions over time with uncertainty intervals (shaded area), demonstrating the model’s ability to quantify prediction uncertainty. The model achieves MAE: 0.114°C, RMSE: 0.143°C, and R²: 0.783.
Figure 8: Extended Performance Analysis. (Left) Autocorrelation of prediction errors shows no significant patterns beyond lag 0, indicating that the model has captured the temporal dynamics of the data without leaving systematic patterns in the residuals. ( Right ) Rolling performance metrics ( MAE , RMSE, R²) remain stable over the evaluation period, showing model robustness across different time segments.
Figure 9: Classification Performance for Heat Stress Prediction. (Left) Normalized Confusion Matrix showing high accuracy for both Normal (0.94) and Stress (0.82) classes. (Middle) ROC Curve with AUC of 0.967, demonstrating good discriminative ability. The optimal threshold is identified at 38.630°C. (Right) Precision-Recall Curve with AUC of 0.946, indicating strong performance across different classification thresholds. The model achieves an accuracy of 0.8938, precision of 0.8411, recall of 0.8453, and F1-score of 0.8425.
Reliable individual monitoring is essential for dairy cattle welfare and early heat stress detection in Precision Livestock Farming (PLF). However, translating raw multi-camera videos into automated diagnostics remains challenging due to disconnected tracking and recognition pipelines and complex physical environments. Dense mutual occlusions and cross-camera viewpoint variations fragment identities, while environmental noise, including camera vibrations, and limited agricultural data obscure subtle heat stress cues such as rapid breathing and flank undulations. To bridge low-level visual features and high-level behavioral diagnostics, we propose CowCare, a unified framework combining viewpoint-robust tracking with spatiotemporal heat stress recognition. Its tracking module, CowTrack, maintains cross-camera identity consistency through a View-Decoupled Aggregation Network (VDA-Net). Instead of relying on static global representations, VDA-Net aggregates localized textures under heavy occlusion and uses a Viewpoint Decoupling Transformer (VDT) to suppress transient viewpoint variations. CowTrack further incorporates a causal online self-supervision mechanism for continuous adaptation to dynamic farm scenes. The resulting tracklets are subsequently processed by TSDF-Net, a fine-grained video understanding network. Its Temporal Stabilization and Aggregation Module (TSAM) suppresses high-frequency jitter, while a Diffusion Semantic Prior (DSP) uses frozen pre-trained diffusion semantics to enhance subtle behavioral signals and mitigate overfitting. Extensive evaluations on two real-world farm benchmarks demonstrate the framework's effectiveness: CowTrack achieves 83.48% HOTA and 87.72% IDF1 for global multi-object tracking on MVCow, and TSDF-Net achieves 87.72% Top-1 accuracy on HSCow. These results establish CowCare as a robust end-to-end framework for automated livestock health monitoring.
Xin Liu, Weiran Li, Minghao Zhu +2
College of Information and Electrical Engineering, China Agricultural University, Beijing, 100083, China · National Innovation Center for Digital Fishery, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China · Key Laboratory of Smart Farming Technologies for Aquatic Animal and Livestock, Ministry of Agriculture and Rural Affairs, Beijing, 100083, China +1
Commercial grazing systems yield irregular livestock observations, which challenge cattle growth forecasting. This study developed a hybrid machine learning framework for herd level cattle weight forecasting using automated sensing observations collected between 2022 and 2024 in southeastern Australia. Weekly live weight observations, demographic variables, and lagged environmental predictors were integrated into structured forecasting datasets. Herd level forecasting trajectories were generated through temporal aggregation of animal level predictions. Four hybrid architecture families were evaluated, including residual, stacked, cascade, and ensemble assisted frameworks. ARIMA, LSTM, and GRU models were used as comparative baselines. Independent testing demonstrated strong predictive agreement across multiple forecasting horizons. The cascade GB to RF to NN architecture achieved the best performance, with a test R^2 of 0.889, RMSE of 21.319 kg, and MAE of 15.462 kg. Hybrid architectures maintained greater robustness than recurrent sequential models under sparse observation conditions. Forecasting error increased progressively across extended prediction horizons. Feature importance analysis identified animal age, rainfall, and temperature as dominant predictors influencing herd level growth forecasting. The proposed framework may support feed allocation, grazing management, and livestock marketing decisions under heterogeneous sensing environments.
Muhammad Riaz Hasib Hossain, Rafiqul Islam, Shawn R. McGrath +2
a School of Computing, Mathematics and Engineering, Charles Sturt University, Wagga Wagga, NSW 2650, Australia · b School of Computing, Mathematics and Engineering, Charles Sturt University, Albury, NSW 2640, Australia · c Gulbali Institute for Agriculture, Water and Environment, Charles Sturt University, Wagga Wagga, NSW 2678, Australia +4
Accurate prediction of crop states (e.g., phenology stages and cold hardiness) is essential for timely farm management decisions such as irrigation, fertilization, and canopy management to optimize crop yield and quality. While traditional biophysical models can be used for season-long predictions, they lack the precision required for site-specific management. Deep learning methods are a compelling alternative, but can produce biologically unrealistic predictions and require large-scale data. We propose a hybrid modeling approach that uses a neural network to parameterize a differentiable biophysical model and leverages multi-task learning for efficient data sharing across crop cultivars in data limited settings. By predicting the parameters of the biophysical model, our approach improves the prediction accuracy while preserving biological realism. Empirical evaluation using real-world and synthetic datasets demonstrates that our method improves prediction accuracy by 60% for phenology and 40% for cold hardiness compared to deployed biophysical models.
William Solow, Paola Pesantez-Cabrera, Markus Keller +3
Oregon State University · Washington State University