328 Late-Breaking: Transfer learning for improving dairy cattle body weight prediction in small farms using depth images and point cloud data.

Abstract Accurate body weight (BW) prediction is essential for dairy cattle health monitoring and feed management. Traditional approaches are not practical for routine use on farms as they typically require intensive labor. New technologies, such as computer vision, provide a more efficient and scalable approach for predicting BW. Depth images have been widely used for BW prediction, and three-dimensional point cloud data have also been investigated in previous studies. However, direct comparisons between the performance of depth images and point cloud data in dairy cattle BW prediction remain limited. In addition, training a computer vision model for BW prediction usually requires a large dataset, which presents challenges for small farms that have fewer animals. The objective of this study was to compare the performance of depth images and point cloud data for BW prediction and to investigate the use of transfer learning from a large farm to improve BW prediction in a small farm. Top-view depth images and point cloud data were collected from 1,201, 215, and 58 cows at three dairy farms of large, medium, and small sizes, resulting in 13,458, 2,122, and 211 samples, respectively. Within each farm, we randomly split the data from 60% of the animals into a training set for model training and the data from the remaining 40% of the animals into a testing set for model evaluation. Two deep learning models, ResNet50 and PointNet++, were trained on depth images and point cloud data for BW prediction, respectively. In both the large and medium farms, PointNet++ outperformed ResNet50, achieving lower Mean Absolute Percentage Error (MAPE) values of 4.04% and 6.89%, compared to 5.75% and 8.97% from ResNet50, respectively. However, ResNet50 performed better in the small farm, achieving 8.50% MAPE compared to 12.10% from PointNet++. To improve performance in the small farm, we applied transfer learning by fine-tuning the model pretrained on the large farm using the training set from the small farm. This approach reduced the MAPE in the small farm testing set from 8.50% to 5.71% for ResNet50 and from 12.10% to 6.76% for PointNet++. These results suggest that point cloud data can lead to better BW prediction performance than depth images, especially when the training data are large. The findings also indicate that transfer learning using large farm data can substantially improve BW prediction in the small farm. Further studies are needed to investigate the broader application of transfer learning for improving BW prediction in small-scale dairy operations.