Automated dairy cattle body condition score using side-view images and deep learning

Body condition score is key to assessing the health, productivity, and energy balance of dairy cows. However, traditional manual methods are time-intensive, dependent on evaluator experience, and prone to subjective biases, limiting large-scale, frequent monitoring. In smallholder settings, where regular BCS monitoring and nutrition advisory support are unavailable, rationing often focuses on intake targets, overlooking body condition feedback, resulting in overconditioned cows, particularly during the periparturient period. This increases the risk of metabolic disorders such as ketosis from excessive lipid mobilization. Here, we developed and validated a fully automated regression-based BCS system from side-view images. Given a single on-farm photograph, the system automatically detects and crops the bovine region of interest, preserves the aspect ratio, and applies resizing and padding to the target resolution. The cropped image is then fed into a deep image regression model, estimating the continuous BCS and creating an automated evaluation. We validated the system using 3,208 side-view images of 211 Holstein cows from a commercial herd, capturing a range of real-world scenes and lighting conditions. To ensure reliable ground truth, 3 trained assessors independently scored the cows, with an intraclass correlation coefficient of 0.92 at the cow level before consensus fusion. The method employs a 2-stage design: first, a You Only Look Once version 11 model pretrained on the COCO dataset detects and crops the image with a mean average precision of 99.5%, minimizing background interference and standardizing inputs; second, a side-view image regression model predicts the BCS. To ensure unbiased evaluation, we performed a stratified 5-fold cow-level cross-validation, ensuring images of the same cow appeared exclusively in either the training or validation sets. Comparisons of multiple backbones identified ConvNeXt-Tiny as the optimal model, achieving a mean absolute error of 0.41 and a Pearson correlation coefficient of 0.62. These results indicate that the model's prediction error falls within the typical range of interassessor variability. Explainable artificial intelligence techniques revealed the model focuses on key anatomical regions used by human raters, such as tailhead, hooks, pins, and ribs. This suggests the model learned biologically relevant features, not background artifacts or irrelevant signals. This system is noninvasive, cost-effective, and infrastructure-light. A single side-view image suffices for initial screening, enabling both herd-level monitoring and individual follow-up, facilitating early detection of negative energy balance and timely nutritional intervention. To ensure reproducibility and support edge deployment, all inference scripts, model weights, and logs are archived. Extreme BCS cases are rare in large-scale operations, but future work will expand the dataset, perform external validation across farms, and assess uncertainty thresholds to enhance robustness. In conclusion, this study introduces a proof-of-concept automated side-view BCS regression system with multiexpert consensus labeling, a deployable pipeline, and explainability-based validation, offering potential technical support for precision dairy farming and improving animal welfare.