A software tool was developed to allow a dairy producer and/or agricultural advisor to monitor the genetic selection differentials (GSD) that a dairy farm is making. The objectives of this study were (i) to monitor GSD in individual farms, over years, so that producers can be advised as to whether or not they are achieving their selection objectives (and hence optimizing productivity and profitability); (ii) the development of a prototype software tool and visualization model to assist producers in interpreting the results for their individual farm, and to compare their farm results with suitable benchmarks. Data used for this study were the EBVs routinely calculated by the Canadian Dairy Network (CDN), for Milk Yield, Fat Yield, Protein Yield and Lifetime Profit Index. The Canadian Ayrshire breed has been used as a model, test breed; records on animals born between January 1980 and April 2016 were used. For each day between this period, the average sire EBV of all the sires available for use on that day was calculated; sires availability determined from the dates of sires’ first and last usage. Sire GSD was then calculated as the EBV of the Sire of a cow minus the average sire EBV on the date the cow was conceived. The average GSD for the entire population, top and bottom 10% sires, and top and bottom 10% of herds per year of conception were computed and stored in a database. This then allows an individual producer to compare and visualize the individual animal selection he/she is making, and also to compare his/her herd against the average herd and the top 10% of herds (as a reference goal to potentially aim to also achieve). The developed software tool is updatable every time CDN releases new genetic evaluation list. The concept can be equally well applied to the other dairy breeds and livestock species for which genetic evaluations are routinely computed. This methodology is not limited to only the four named traits but can also be used for all traits genetically evaluated (currently approximately 30 traits in Canada), allowing a producer to monitor multiple traits and hence decide upon his/her selection objectives.

Accelerometers support the farmer with collecting information about animal behavior and thus allow a reduction in visual observation time. The milk intake of calves fed by teat-buckets has not been monitored automatically on commercial farms so far, although it is crucial for the calves’ development. This pilot study was based on bucket-fed dairy calves and intended (1) to evaluate the technical feasibility of using an ear-attached accelerometer (SMARTBOW, Smartbow GmbH, Weibern, Austria) to identify drinking events, (2) to develop an algorithm to detect milk intake, and (3) to validate the SMARTBOW sensor incorporating the algorithm developed under (2) for identifying drinking events against observations from video. The acceleration data used in this study were generated from three sensors attached to the ears of three preweaned calves. Sensor data were recorded for 5  d for 24 h/d and calf behavior was video camera-recorded during the same time period . Based on a training data set, an algorithm was developed to identify drinking events. In addition, a mathematical data simulation was performed which generated further 15  d of data. The complete data set was compared with video recordings to analyze whether drinking events (n = 174) were detected correctly. Sensitivity (82.9 %), specificity (96.9 %), and accuracy (96.2 %) were good, but precision (60.4 %) was not yet satisfactory. Cohen’s kappa (0.68) indicated a substantial agreement between sensor and video analysis. Additional work with a larger number of animals is planned to further improve the algorithm.

Machine-learning methods may play an increasing role in the development of precision agriculture tools to provide predictive insights in dairy farming operations and to routinely monitor the status of dairy cows. In the present study, we explored the use of a machine-learning approach to detect and monitor the welfare status of dairy herds in terms of lameness and lesions based on pre-recorded farm-based records. Animal-based measurements such as lameness and lesions are time-consuming, expensive and, thus, typically not collected on a routine basis. A predictive model that is suitable for routine field applications can be thus an efficient strategy to improve dairy cattle welfare. A decision tree approach was therefore used to classify the welfare status of 229 herds. Single measurements were aggregated to a composite index for lameness and lesions, scaled to percentile ranks, and expressed as low, intermediate and high risk that a herd be deficient in lameness and lesions. Routinely collected dairy herd improvement data related to milk production, milk quality, herd size, housing and reproduction were used as potential predictors of the risk level. Model accuracy based on the average of repeated 10-fold cross validation suggests that a simple decision tree algorithm was able to predict welfare level with a mean accuracy of 44%. Ensemble methods such as random forests and boosting methods slightly improved the prediction performance to some extent (up to 51% accuracy). Model specificity for herds at high welfare risk was 91% with a boosting approach, suggesting that only a small proportion of lower risk herds were misclassified as high risk herds. These results suggest that a model based on a machine-learning approach is able to detect herds with potential welfare deficiencies using routine herd data. Additional data are required to improve model performance and validate the approach. Nonetheless, a machine-learning approach may be an appropriate and powerful tool to estimate and monitor the dairy welfare status at herd level, and can be a useful decision support tool for dairy farmers.

The objective of this study was to apply principal component analysis (PCA) and multiple correspondence analysis (MCA) on Dairy Herd Improvement (DHI) data of animals on their first lactation to discover the most meaningful set of variables that describe the outcome on the first test day. Data collected over 4 years were obtained from 13 dairy herds located in Québec – Canada. The data set was filtered to contain only information from first test day of animals on their first lactation, resulting in 1637 observations and 35 variables. Eight additional variables were created from the existing DHI metrics. PCA was performed on numeric variables (n = 14) after they were standardized to mean = 0 and standard deviation = 1. MCA was performed on categorical variables (n = 20). Seven numerical variables and eight categorical variables were selected as meaningful to describe the variation on the first test day using PCA and MCA. These variables could be used to evaluate the outcome on the first test day of animals on their first lactation and assist in the evaluation of their transition period. Future work could focus on modeling the relationship between those variables.

Pulse oximetry is a well-established technique in nowadays human and veterinarian medicine. Also in the farm animal sector, it could be a useful tool to detect critical conditions of the oxygen supply and the cardiovascular system of the patient. However, its use in ruminant medicine is still limited to experimental application. The objective of this study was to evaluate the accuracy of a Radius-7 Wearable Pulse Oximeter (Masimo Corporation, Irvine, CA) for monitoring the vital parameters of Holstein Frisian calves. The sensor of the pulse oximeter was placed in the interdigital space of the calf’s front leg. The arterial oxygen saturation (SO2) of 40 newborn calves was measured and compared with the corresponding results from a portable blood gas analyzer which served as reference method. The arterial blood sample was taken from the medial intermediate branch of the caudal auricular artery. The pulse rate was measured on 10 calves aged 0 to 7 days with the pulse oximeter and a heart rate belt simultaneously and their level of agreement was evaluated. Spearman correlation coefficient was 93.8 % for the SO2 parameter between the pulse oximeter and the blood gas analyzer and 97.7 % for the pulse rate between the pulse oximeter and the heart rate belt. The pulse oximeter overestimated the SO2 by 2.95 ± 6.39 % and underestimated the pulse rate by -0.41 ± 3.18 bpm compared with the corresponding reference methods. This pulse oximeter seems to be suitable for continuous monitoring of SO2 and pulse of Holstein-Friesian calves.

Body condition score (BCS) is considered as one of the most important indices for managing dairy cows, which is used to evaluate fat cover and changes in body condition. Dairy farmers should be aware of their cows BCS to be able to identify the patient cows on time and manage diets when needed. In this study, we have introduced a new index which uses Radial Descriptor Lines (RDL) for BC scoring. Based on the fact that the fatter the cow the smoother the back surface, we hypothesised that the changes on the cow’s back at different BCSs could be tracked through the changes on the radial lines centred on the hook bone emitting outward on the surface of the cow’s back.

Images were captured using a Kinect sensor installed in the milking parlour of a dairy farm. To provide the required data for model development and assessment, 165 images were captured from 55 cows with different BCSs. Algorithms were developed in MATLAB environment. The consecutive steps designed in the algorithm were firstly distinguishing the hook bone based on the local maxima on the depth data taken from the Kinect sensor from the cow’s back. Secondly, radial descriptor lines were taken out of the back surface with interval angles of 1 degree outward the hook bone toward the edges of the cow’s back. Overall variations of the descriptor lines respect to a polynomial modelled datum line were measured and used as the extracted features. To include the most related and exclude non-related variations from the BCS estimation model, four orders (from 2 to 5) of polynomial datum curves were tested.

Effective features were selected using correlation-based feature selection (CFS) and fed to artificial neural networks to provide the corresponding BCSs. Results showed a correlation between the estimated BCSs from the model and the scores determined by the experts with a coefficient of determination (R2) of 0.87 and a root mean square error (MSE) of 0.036.