Assessment and validation of individual water intake of dairy cows from reticular boluses

The whole dairy cattle community, scientists and farmers in particular, are putting efforts to improve environmental sustainability of the sector. Future generations of dairy cows must optimize natural resources use without impairing quantity and quality of their output. Water use in cattle, for example, can be more sustainable and efficient. Sensor data can often be good predictors of complex traits, like health status, fertility and stress. However, a large quantity of reference data is needed to develop robust predictive models and decision-support systems. In this study we explored the possibility to predict cow’s water intake (WI) by mean of longitudinal temperature recorded by specific reticular boluses (smaXtec animal care GmbH, Austria) inserted into 8 cows undergoing a feeding trial at the experimental farm of the University of Bologna (October to December 2020). Cows, 4 primiparous and 4 pluriparous, were farmed in tie-stall with an individual drinker available. The WI was registered every day for 28 d, i.e. 7 consecutive d per each diet (n = 4); thus, data referred to adaptation phases were excluded. To quantify the number of times the reticular temperature dropped and duration of such drops, a daily ‘drop area’ was obtained for each cow by subtracting the area under the reticular temperature curve from the area under the body temperature curve. The ‘drop area’ was subsequently linked to the respective daily WI, body weight (BW), and milk yield (MY). Before boluses validation, the reference WI (140 ± 34 L/d) was adjusted for systematic effects (cow, parity, diet, and days in milk) using the GLM procedure of SAS v. 9.4. Then, the GLMSELECT procedure selected the most important predictors of WI among all imputed traits: mean, SD and minimum of both body and reticular temperature, MY and BW. Different models were tested by combining the predictors offered using a 5-fold cross-validation (75% training, 25% validation). Based on the mean square error in cross-validation, the best model was the one with all predictors available (R2 CV = 0.90). When MY and BW were masked, the accuracy decreased (R2 CV = 0.65). In both models, the ‘drop area’ was selected as one of the most important predictors and, when used alone, the R2 CV was 0.61. Results are promising and suggest that there is potential to explore dairy cattle water efficiency: WI could be in fact estimated from sensor data in free stall barn/commercial contexts where recording individual intakes is not feasible.