76. Predicting Body Weight and Dry Matter Intake in Lactating Dairy Cows Using Depth Cameras.

Abstract

Dry matter intake (DMI) is typically estimated at the group level or predicted using equations based on milk yield, body weight (BW), and days in milk (DIM). Computer vision systems (CVS) can generate large-scale phenotypes such as body volume, area, and length that may better predict intake than BW alone, because cows with similar BW can differ in body frame and intake capacity. Therefore, this study aimed to (1) predict BW using CVS to demonstrate the association between image-based features and BW and (2) predict DMI using image-based features rather than BW. Data were collected from September 2024 to June 2025 at CERZOO Ltd. farm (Piacenza, Italy) from 93 lactating Holstein cows (151 ± 82 DIM, 41 ± 8 kg milk yield, 27 ± 5 kg DMI, and 3 ± 1 lactations), totaling 8,587 BW measurements obtained using a walk-over scale and 8,168 DMI records. For individual identification, a depth camera was installed at the sorting gate where animals were identified. For feature extraction, a segmentation model using a U-Net architecture was trained on 4,328 top-down depth images and validated on 50 images. After segmentation, depth frames were processed to extract body biometric (BB) features from the dorsal body surface, including area, volume, circularity, extent, eccentricity, perimeter, and major axis length, after removing noise and correcting invalid depth pixels. In addition, a YOLOv8 pose estimation model trained on 2,098 annotated images and validated on 25 images detected seven anatomical keypoints (hips, pin bones, tail head, sacral vertebra, and cervical vertebra). Model performance was evaluated using a threshold equal to 5% of the square root of the image area, and Euclidean distances between keypoints were used as features. Extracted features were summarized by animal and week. For BW prediction, a random forest (RF) model using BB features was trained. For DMI prediction, RF models were trained using (1) milk yield, DIM, and BW or (2) milk yield, DIM, and BB features. Both models were validated using leave-one-animal-out cross validation. All keypoints were predicted with accuracy greater than 88%. The segmentation model achieved an average IoU of 0.86 (95% CI: 0.85, 0.87). The BW prediction model achieved an R2 of 0.53, CCC of 0.70, and RMSE of 39.2 kg, representing 5.9% of the observed BW. The DMI model using milk yield, DIM, and BW achieved an R2 of 0.29, CCC of 0.50, and RMSE of 2.9 kg (11.0% of the average observed DMI). When BB features replaced BW, model performance improved to an R2 of 0.33, CCC of 0.52, and RMSE of 2.8 kg (10.7% of the observed values). These results suggest that BB features extracted from depth cameras are associated with BW and better predict DMI than BW from walk-over scales.