Experiment 1

The first experiment was conducted between 10 and 19 May 2016 on a group of 18 spring calving dairy cows on the Teagasc research farm in Moorepark, Fermoy, Ireland. The data of six of these cows were analyzed to align times between the visual observations and the automated sensor (MooMonitor+). Due to the technical specification of the internet connection at the research farm, the timestamp on the base station linked with the sensors was not correct. Therefore, to investigate the correct time stamp, data of six cows were used to validate the automated sensor against human observer. Those data were excluded afterwards from the experimental data set.

Twelve cows were used for validation. This group consisted of six Jersey crossbred (JEX) and six Holstein-Friesian (HF) cows. There were four primiparous and eight multiparous cows involved, the range of lactation was from 2 to 6. The mean body condition score (BCS)± SD was 2.8±0.2 (based on a 1 to 5 scoring system with 0.25 increments; Edmonson et al., 1989). Average BW was 477±65 kg. The milk yield was 22.5±4.5 kg/cow per day over the experimental period and average days in milk (DIM) was 91±12 at the beginning of the experiment. All cows followed a similar milking schedule, being milked twice daily at 7:00 and 14:30 h with approximately 1.5 to 2.0 h away from the paddock during each milking. Cows were fed with only grass on the paddocks with no additional supplementation of concentrate. A fresh allocation of pasture was provided after each milking. Pre- and post-grazing grass heights were measured daily using a rising plate meter (diameter 355 mm and 3.2 kg/m²; Jenquip, Fielding, New Zealand). Pre-grazing heights and post-grazing heights were 11.9±2.5 cm and 4.5±0.8 cm, respectively, during the experimental period. These values represented a non-restrictive grazing management strategy in Ireland, where cows received a daily herbage allowance of 16.3±2.6 kg dry matter (DM)/cow per day, measured above 3.5 cm sward height, on average during the experimental period (McCarthy et al., 2013). The chemical composition was analyzed once weekly resulting in an average DM content of 15%, an average CP content of 22% and an average content of NDF and ADF of 41% and 23%, respectively.

Grazing and rumination time data were collected by visual observation according to a 1-min scan sampling protocol, similar to the method used in the study of Büchel and Sundrum (2014) and by the MooMonitor+. Two previously trained observers were monitoring 18 cows in total (12 cows for validation; 6 cows for time alignment). The cows were divided into six subsets with three cows each for the purpose of observation. Each subgroup was observed by each observer on three occasions over 6 days (Table 1). Observations took place over 2-h periods between dawn (05:00) and dusk (21:00) excluding milking times from 07:00 to 09:00 h and 14:00 to 17:00 h. After the first 3 days, the times were changed to cover the full range of daylight hours within the days.

Experimental protocol for cow grazing and rumination data collection by visual observation

Behavioral data of each minute were categorized into grazing and rumination, considering the main activity within each minute. Grazing was defined as cow’s muzzle being located near or above the grass and making a biting motion to ingest grass or chewing ripped grass with the head position down, or cow’s head position up and making a chewing motion to masticate the grazed grass. Alternatively, rumination was defined as regurgitation, chewing, salivation and swallowing of ingested grass (Bikker et al., 2014). The data were recorded on a manual spread sheet. Subsequently, the data were transferred manually to an electronic spread sheet (Microsoft Excel Version 2010; Microsoft Corporation, Redmond, WA, USA).

The MooMonitor+ was used for the automated data collection. It is a collar device on the cow’s neck containing a box with a 3-axis accelerometer was positioned on the right sight of the neck. This accelerometer measured activity in a 10 Hz resolution. On-board data analysis with a generic algorithm, which was identifying specific pattern for different categories such as rumination, grazing, resting, developed by Dairymaster, summarized activities occurring in the raw data into time spent at those activities for 15-min periods. These summarized periods were then transmitted wirelessly to a base station with a range of up to 2000 m. The base station is usually linked with the internet connection in a normal farm environment and corrected itself in time, based on a deviation of ±5 min. This time is also corrected on the sensors once they were in the range. To ensure correct positioning of the accelerometer box on the cow’s neck, a weight was applied at the lowest point of the collar. Cows within the group were identifiable by numbers painted on their sides.

The 1-min visually recorded data in the experimental dataset were summarized in 15-min and 1-h grazing and rumination periods to allow direct comparison with the data recorded by the automated method. The automatically captured data were classified into the categories of grazing and ruminating in 15-min summaries. Then four 15-min summaries were totaled to form 1-h summaries. Consequently, there were 504 15-min periods and 72 1-h periods of valid observations across the full database.

For the statistical analysis, R version 3.3.1 (R Foundation for Statistical Computing, Vienna, Austria) was used (statistical code see Supplementary Material S1). To assess agreement between numeric value data of the MooMonitor+ and visual observation, the Spearman’s Rank correlation (r_s) and a concordance correlation coefficient (CCC) was calculated. The interpretation of r_s-values and CCC were based on definitions by Hinkle (2003) as follows: Negligible = 0.0 to 0.3, low = 0.3 to 0.5, moderate = 0.5 to 0.7, high = 0.7 to 0.9 and very high = 0.9 to 1.00. Furthermore, the Bland–Altman analysis was applied to assess the agreement between visual observation and automated system. This was conducted in Microsoft Excel calculating the mean differences (bias; MooMonitor+ – visual observation) against the means of visual observation and MooMonitor+. The limits of agreement were calculated as ± 1.96×standard deviation from the mean difference. Although the parameters themselves were not normally distributed, the Bland–Altman analysis was used as the differences between the paired values did follow a normal distribution.