Random forest regression for estimating dry matter intake of grazing dairy cows

The understanding of individual feed intake of grazing cows is essential for monitoring the nutrient intake of the animal, to calculate feed efficiency and increase animal and pasture productivity as well as tailored herd and pasture management. However, the measurement of individual herbage dry matter intake (hDMI) of individual cows is laborious and may be challenging in commercial grass-based dairy systems. The aim of this study was to assess the application potential of machine learning algorithms (random forest) to estimate the individual hDMI of grazing dairy cows in an intensive grazing management system using animal behaviour characteristics. This study was performed at a research farm on n = 41 animals with the established reference value for measuring feed intake based on the n-alkane technique. The database for model development included individual cow information, on-field grass measurements, grass quality as well as detailed individual cow behavioural characteristics based on the RumiWatchSystem. Random forest regression was used to predict hDMI. Recursive feature elimination (RFE) was used to select the best subset of predictors to be included in the model. To overcome issues associated with the relatively small sample size of n = 68 weekly values in total a nested cross-validation procedure was implemented. Results showed that RF has good potential for the prediction of hDMI in grazing dairy cattle. However, further studies are required to fully assess performance of this method and identify new potential predictors for hDMI.