Prediction of milk yield using visual images of cows through deep learning

Abstract

The objective of the study was to determine, through deep learning, the predictability of milk yield from a cow's side-view, rear-view image, or a combination of the two. The data size of 1 238 image pairs (the side-view and rear-view images) from 743 Holstein cows in their first or second parity, and their corresponding first-lactation, 305-day milk yield values were used to train and test Deep Learning models. The data were first split into the training and testing data at a ratio of 80:20, respectively. The training data were augmented four times, then again divided into training and validation data at the ratio of 4:1, respectively. Three principal analyses were done, the prediction of milk yield using rear-view images, using side-view images, and using a merge of the side-view and rear-view images. In all three analyses, poor predictions were observed, i.e., R² values ranging from 0.30 to 0.38, the mean absolute error ranging from 1112.9 kgs to 1148.3 kgs, the root mean square error values ranging from 1401.2 kg to 1480.5 kg, and the mean absolute error percentages ranging from 17.0 to 17.6%. Hypothesis tests were done to check whether there was any difference between these three prediction models. There was no significant difference in performance between the prediction models. It was concluded that predicting the 305-day milk yield of Holstein cows using either view had the same level of accuracy, and no additional benefits were derived from using both the rear and side views.