Analysing the back of dairy cows in 3D imaging to better assess body condition
Résumé
Body condition is an important trait in dairy cow management, usually measured with the body condition
score (BCS), which is subjective, and not very sensitive. The aim of this work was to develop and to validate
a method, NEC3D, estimating BCS with 3D pictures of dairy cows back, from the pins to the hooks, which
is commonly used as reference area in the BCS scoring systems. A 57 cows 3D-shapes dataset with large
BCS variability (0.5 to 4.75 on a 0-5 scale), transformed with a principal component analysis, was built
for calibration. The principal components were performed on BCS with multiple linear regressions. Four
anatomical points had to be identi¿ed manually to normalise the pictures. Influence of two different ways
of points’ identification and of the picture resolution on method quality was analysed. External validation
was evaluated on two additional datasets: one with cows used for calibration, but with a different stage in
milking (valididem) and one with cows not used for calibration (validdiff). To fully qualify the method,
the reproducibility was estimated with 6 cows using 8 3D-shapes of each cow obtained the same day. Both
ways of points’ identification had quite good results in terms of calibration (R2=1) and differed slightly on
validation quality (RMSE=0.34 vs 0.32 for validdiff). Nec3d was 2.8 times more reproducible than usual
BCS (σ=0.1 vs 0.28). The lowest resolution implied a loss of reproducibility, but did not increase the error
of prediction. A simplified acquisition system, implying low resolution, could therefore be developed. The
error of prediction was similar for valididem and validdiff, indicating that the NEC3D is not less efficient
for cows not used in the calibration set. Assessing body condition thanks to 3D shapes appears to be a
promising tool which can improve phenotyping of this trait.