Complementarity of data sources for the prediction of duck liver quality

In the duck "foie gras" (fatty liver) industry, the evaluation of the technological quality of raw material is essential to optimize its processing. The main parameter is the cooking loss (%LOSS) which can be up to 50% and is very variable. Reference measurement of %LOSS is unrealistic in practical conditions. It is therefore interesting to predict %LOSS from liver weight, chemical composition (DM, Fat), or to use rapid methods such as color (L*a*b*), or NIR spectroscopy. In the present study we tried to evaluate the complementarity between these sources of information. Liver weight, color and %LOSS were measured on 740 foie gras samples while Fat and DM were measured on a subset of 140 samples. Two procedures were used for NIR spectra acquisition: 1/ on whole liver directly after slaughter, with ASD Labspec Pro (350-2500nm); and 2/ on homogenized sample, with a FOSS NIRSystem 6500 (400-2500nm). Calibration equations were built with PLS regression, using either ASD spectrum, FOSS spectrum, measured variables (weight, color, composition), or any combination of these parameters. Parameters as color and chemical composition led to predictions with SECV values between 8 and 9%. Predictions with NIR had SECV around 5.2% either with ASD or FOSS, and 4.5% with the two spectra used together (two arrays). The use of measured parameters together with NIR spectra did not improve the quality of predictions. It is concluded that the information contained in spectra from FOSS and ASD shows some differences, since the joint use of these two spectra leads to better results than each one individually. On the contrary, all the information brought by measured parameters is already contained in spectra - it is therefore useless to integrate this information in prediction models. Although the multi-array approach is not adapted to industrial conditions in this case, this study demonstrates how it can help identify the most accurate information sources. (Texte intégral