¹Data Science in Biomedicine, TU Braunschweig, Braunschweig, Germany
²Technische Universität Braunschweig, Braunschweig, Germany
³Thünen Institute of Agricultural Technology, Braunschweig, Germany

Text

Introduction: Cow-calf contact (CCC) systems, allowing dams to nurse calves for varying periods, are gaining attention in dairy farming as an alternative to immediate separation, driven by consumer and welfare concerns. Research indicates that CCC does not significantly increase disease transmission risks, but it consistently reduces saleable milk yield, with varying effects on milk composition depending on contact duration and breed [1]. In pasture-based systems, CCC lowers milk yield without harming calf weight gain, highlighting trade-offs between welfare and productivity [2]. This study investigates the impact of different CCC systems on dairy cows' milk flow patterns using advanced machine learning techniques, aiming to optimize dairy production while supporting more sustainable practices.

Methods: We analyzed daily milk flow data from 82 German Holstein cows at the Thünen Institute of Organic Farming subjected to three CCC treatments: whole-day contact (WDC), daylight contact (DLC), and no contact (NOC). The experiment spanned from August 2020 to June 2022, with milk flow recorded twice daily. Data preprocessing included standardization, normalization, and dimensionality reduction via Principal Component Analysis (PCA) to prepare the time-series data for clustering. We employed baseline clustering algorithms such as K-Means and K-Medoids, alongside Deep Temporal Clustering (DTC) [3], which uses an autoencoder to capture hierarchical and temporal features in the data. Clustering performance was evaluated using accuracy, purity, normalized mutual information, and adjusted Rand index.

Results: DTC outperformed traditional clustering methods across all performance metrics, demonstrating superior ability to capture the underlying patterns in milk flow data related to CCC. The DTC model is better aligned with the distinct CCC groups compared to K-Means and K-Medoids. Additionally, clustering based solely on phenotypic data (e.g., cow weight, lactation number) showed weaker alignment with CCC groups, reinforcing that CCC has a dominant influence on milk flow patterns. These results highlight the effectiveness of advanced clustering techniques like DTC in uncovering complex relationships within time-series agricultural data.

Conclusion: Our findings suggest that CCC significantly impacts milk flow patterns, with DTC emerging as a powerful tool for elucidating these complex relationships. The superior performance of DTC underscores the importance of considering temporal dynamics in analyzing milk production data, offering a scalable approach for large-scale dairy operations. These insights suggest that optimizing CCC systems could enhance both milk production and animal welfare, potentially reducing stress and improving long-term productivity. This research employs unsupervised learning to reveal patterns within milk flow data, without predefined labels, reducing bias and allowing for exploratory analysis. Further research should focus on refining DTC models and incorporating additional phenotypic data to enhance the accuracy and interpretability. Moreover, comparing phenotypic clusters to CCC groups could further clarify the interplay between cow traits and contact systems. Future studies should enhance DTC models with richer phenotypic information and evaluate the long-term effects of CCC on health and behavior.

Deep learning-based clustering can reveal how CCC affects milk production patterns, guiding welfare-conscious, data-driven dairy management. Further refinement of such models and exploration of CCC's broader impacts will support sustainable farming innovation.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.

References

[1] Barth K. Effects of suckling on milk yield and milk composition of dairy cows in cow–calf contact systems. J Dairy Res. 2020;87(S1):133-137. DOI: 10.1017/S0022029920000515
[2] Johanssen JR, Adler S, Johnsen JF, Sørheim K, Bøe KE. Performance in dairy cows and calves with or without cow-calf contact on pasture. Livest Sci. 2024;285:105502. DOI: 10.1016/j.livsci.2024.105502
[3] Madiraju NS, Sadat SM, Fisher D, Karimabadi H. Deep temporal clustering: Fully unsupervised learning of time-domain features [Preprint]. arXiv. 2018;cs.LG/1802.01059. DOI: 10.48550/arXiv.1802.01059