DataSheet4_Supervised learning techniques for dairy cattle body weight prediction from 3D digital images.PDF

Introduction: The use of automation and sensor-based systems in livestock production allows monitoring of individual cows in real-time and provides the possibility of early warning systems to take necessary management actions against possible anomalies. Among the different RT monitoring parameters, body weight (BW) plays an important role in tracking the productivity and health status.
Methods: In this study, various supervised learning techniques representing different families of methods in the machine learning space were implemented and compared for performance in the prediction of body weight from 3D image data in dairy cows. A total of 83,011 records of contour data from 3D images and body weight measurements taken from a total of 914 Danish Holstein and Jersey cows from 3 different herds were used for the predictions. Various metrics including Pearson’s correlation coefficient (r), the root mean squared error (RMSE), and the mean absolute percentage error (MAPE) were used for robust evaluation of the various supervised techniques and to facilitate comparison with other studies. Prediction was undertaken separately within each breed and subsequently in a combined multi-breed dataset.
Results and discussion: Despite differences in predictive performance across the different supervised learning techniques and datasets (breeds), our results indicate reasonable prediction accuracies with mean correlation coefficient (r) as high as 0.94 and MAPE and RMSE as low as 4.0 % and 33.0 (kg), respectively. In comparison to the within-breed analyses (Jersey, Holstein), prediction using the combined multi-breed data set resulted in higher predictive performance in terms of high correlation coefficient and low MAPE. Additional tests showed that the improvement in predictive performance is mainly due to increase in data size from combining data rather than the multi-breed nature of the combined data. Of the different supervised learning techniques implemented, the tree-based group of supervised learning techniques (Catboost, AdaBoost, random forest) resulted in the highest prediction performance in all the metrics used to evaluate technique performance. Reported prediction errors in our study (RMSE and MAPE) are one of the lowest in the literature for prediction of BW using image data in dairy cattle, highlighting the promising predictive value of contour data from 3D images for BW in dairy cows under commercial farm conditions.

引言：在畜牧生产中应用自动化与基于传感器的系统，可实现对单头奶牛的实时（real-time, RT）监测，并为针对潜在异常状况采取必要管理措施的预警系统提供支撑。在各类实时监测参数中，体重（body weight, BW）在追踪奶牛生产性能与健康状态方面发挥着关键作用。
方法：本研究针对奶牛三维（3D）图像数据的体重预测任务，选取了机器学习领域中不同算法家族的多款监督学习算法进行实现与性能对比。本研究共使用来自3个不同牧场的914头丹麦荷斯坦奶牛与娟姗牛的相关数据，包含83011条三维图像轮廓数据记录与体重测量记录。本研究采用皮尔逊相关系数（Pearson’s correlation coefficient, r）、均方根误差（root mean squared error, RMSE）以及平均绝对百分比误差（mean absolute percentage error, MAPE）等多项指标，对各监督学习算法开展严谨评估，以便与其他研究进行对比。预测任务分别针对单个品种开展，随后又在多品种联合数据集上进行。
结果与讨论：尽管不同监督学习算法与数据集（品种）间的预测性能存在差异，但本研究结果显示模型预测精度较为理想，平均相关系数最高可达0.94，平均绝对百分比误差与均方根误差最低分别为4.0%与33.0kg。相较于品种内分析（娟姗牛、荷斯坦奶牛），采用多品种联合数据集开展预测时，在高相关系数与低平均绝对百分比误差层面均展现出更优的预测性能。额外测试表明，预测性能的提升主要源于联合数据集带来的数据量增长，而非联合数据的多品种属性。在所实现的多款监督学习算法中，基于树的监督学习算法家族（CatBoost、AdaBoost、随机森林）在所有评估指标下均取得了最优的预测性能。本研究报告的预测误差（均方根误差与平均绝对百分比误差）在现有利用图像数据预测奶牛体重的相关文献中处于较低水平，这表明在商业化牧场条件下，三维图像轮廓数据用于奶牛体重预测具备良好的应用价值。