Comparaison de méthodes pour valider l’estimation par scanner à induction magnétique de la composition de jambons et de poitrines

Gérard Daumas et al., 52e Journées de la Recherche Porcine (FRA), 4 et 5 février 2020, Paris, p. 59-60, poster

Poster.

La validation statistique est une étape à ne pas négliger dans le processus de test d’une technologie. Néanmoins, il n’y a pas de consensus sur la méthode à appliquer. Les résultats semblent dépendre de la nature des données. Aussi, il est souvent conseillé de tester plusieurs méthodes. Ayant estimé la composition de jambons et poitrines par un scanner à induction magnétique sur un échantillon de calibrage (Daumas et al., 2019), les auteurs souhaitaient passer à l’étape de validation. Pour cela, les auteurs ont comparé la performance de cinq méthodes de régression linéaire parmi les plus courantes.

ENG

Poster.

Comparison of methods to validate magnetic induction scanner estimation of ham and belly composition

Magnetic induction scanning is a promising technology for carcass grading and sorting of cutting parts. The objective of this study was to compare the performance of five prediction methods of the composition of hams and bellies by magnetic induction, based on observations of a calibration sample. The five prediction methods tested were Ordinary Least Squares (OLS), Lasso, Ridge, Partial Least Squares (PLS), and complete selection of sub-models by minimizing Bayesian information criterion (Subset). For each statistical method, R2 and RMSEP were calculated in a 10-fold cross validation repeated 100 times with random division of the data into 10 segments. Data for two calibration samples were used: one for 100 hams and the other for 80 bellies. Hams and bellies were scanned with a recent commercial device using a low-intensity magnetic field. The four response variables, weights and contents of fat and muscle, were measured by computed tomography. Based on the median values, the PLS gave the best performance for hams. The dispersion of results was lowest with the PLS as well. For bellies, Ridge regression was the most successful, except for fat content, for which Subset was better. Muscle content of hams and fat content of bellies were estimated respectively with a median R2 of 0.64 and 0.66. The ranking of methods based on their prediction performance depended on the cut. Subset, Ridge and Lasso seemed to show the most stable prediction performance results among the cuts and response variables, always being close or equal to the best performance.