Computer vision and weigh scale-based prediction of milk yield and udder traits for individual cows

Abstract

The measurement of the performance of dairy cow milk production contributes to improve the productivity and profitability of dairy farms. Milk yield is used on-farm for decisions around culling, drying off, heat detection and off-farm for estimation of sire breeding values. Milk yield is measured by herd testing at multiple times through a lactation and by individual milk meters, although the rate of herd testing is decreasing and only a minority of milking systems have milk meters. The goal of this study was to develop a non-contact computer vision and/or walk-over weigh scale technology based on cow 3D images and/or weights before and after milking to determine individual cow milk yield and the lactation milk yield for individual cows. This technology could provide some of the benefits of individual milk meters, although only 1–2 sensors would be required per milking apparatus as opposed to a milk meter per bale on the milking apparatus. An algorithm was developed to estimate the udder volume before milking, udder volume after milking and the difference in udder volume before and after milking. A calibration model based on in-line milk meter measured milk yield from 20 cows was developed to predict milk yield per milking. The model based on udder volume before milking and the difference in udder volume before and after milking had a calibration R2 = 0.92 and RMSE = 0.84 L. The algorithm RMSE was inferior to the in-line milk meter (RMSE = 0.5 L for this trial). This imaging technology provides an inexpensive method to determine the milk yield of individual cows, which can be used for better on-farm management of cows in the herd. The feasibility of determining lactation yield from walk-over weigh scale measurements of each cow before and after each milking was also demonstrated, with algorithm prediction (based on a 2.5% weigh scale error) of lactation milk yield estimated to be similar to that based on two herd tests per lactation. Furthermore, the potential of image-based phenotyping of udder traits such as front/rear teat placement, teat length and teat orientation were demonstrated, with an estimated model accuracy of 94% for front/rear teat placement on a 9-point scale. Average rear teat length was 39 ± 1.5 mm and average front teat length was 49 ± 1.3 mm. This highlights the potential for using computer vision for on-farm prediction of milk performance and udder traits from dairy cows.