Résumé
The current protocol's main objective is to determine the feasibility to calibrate a qualitative classification model allowing the distinction between good and bad genotypes using NIRS spectra collected on different product states (i.e. raw intact organ, chopped, puree). This protocol focuses on pounded yam and boiled cassava, but if the proof of concept is achieved, it can be extended to RTB product profiles. The main principle is based on the existing traditional knowledge of varieties of good and bad qualities. This knowledge allows us to choose genotypes from both categories and train a binary classification model to predict their belonging. To test feasibility as soon as possible, this preliminary protocol focuses on the already available database. The idea is to train and test a classification model predicting good or bad genotypes using a convolutional neural network (CNN) with a binary cross-entropy loss function. Indeed, this type of algorithms showed excellent results for similar studies. In order to avoid a confusion effect of storage length on tuber quality, the only tuber with similar physiological age (i.e. meaning same storage length) should be kept in each database.