NIRS & biophysical analyses: Tentative prediction of cassava cooking properties - Year 2. High-Throughput Phenotyping Protocols (HTPP), WP3

The analyses concern 56 cassava genotypes harvested in 2019, 2020 and 2021: 1 genotype was analyzed 1 time, 6 analyzed 2 times, 20 analyzed 3 times, 15 analyzed 5 times, 6 six times and 9 analyzed 9 times. The total number of analyses is 250. The samples were analysed for their cooking properties (cooking time in boiling water), texture parameters (gradient, max force, distance at max force, area, linear distance and end force/ max force), dry matter content and water absorption capacity during cooking. The same genotypes were analysed in near infrared spectroscopy. The absorption spectra were performed on ground fresh roots using a FOSS 2500 spectrometer. The average Dry matter is 39,5 %, this value is constant over months (age of the root). The cooking time average value is 33,7 min, the values range from 10 to 60 minutes. The distribution of the values, allows defining 2 classes: C1 for OCT lower than 33,7 min and C2 for OCT higher or equal to 33,7 min. There is a non-linear relation between Water_Absorption at 20 min or 30 min and optimal cooking time: High time cooking genotypes absorb less water at 20 min than “good cooking” genotypes. The values of gradient range between 170 and 2489 kg/mm with an average value of 1205 kg/mm. The distribution of the values follows a normal law. Gradient is highly correlated to physical values related to Max force, Area and Linear distance. Gradient is also correlated to OCT (r = 0,719). The highest correlation between gradient and Water absorption is for WA 30 minutes (r = - 0,693). The relation between gradient and WA_30 is nonlinear (second order), genotypes with high gradient values absorb less water at 30 mn than genotypes with low gradient values which correspond to genotypes with low optimum cooking time. Different multivariate approaches were investigated to associate spectral data and physico-chemical parameters. The direct calibrations of physico-chemical parameters were not performant. Classification according to 2 cooking time classes was tested using different algorithms. Whatever were the pre-treatments used (SNV, SNVD, first or second derivative…) and whatever the classification approach (K Nearest Neighbors, Support Vector Machine, Naive Bayesian Classifier, Random Forest, Classification Regression Trees…), the predictions of a validation set for the 2 cooking time classes failed. The Lasso approach is encouraging and clearly improved the predictive model for OCT. The classification of the samples using predicted OCT values was 82% correct for learning set and range between 66% and 72% for validation samples depending on the validation set. The model lacks robustness, because of a relatively few number of samples and because of the variability of the samples due to harvest year, as shown by the PCA of the spectra. These results confirms that the spectral signature contains information about textural properties and that nonlinear models or deep learning approaches good help extracting this information.

Bibliographic Details

Main Authors:	Davrieux, Fabrice, Belalcazar, John, Zhang, Xiaofei, Tran, Thierry
Format:	monograph biblioteca
Language:	eng
Published:	RTBfoods Project
Online Access:	http://agritrop.cirad.fr/603474/ http://agritrop.cirad.fr/603474/1/RTBfoods_NIRS%20tentative%20calibration_Biophysical%20analyses%20%26%20cooking%20properties_Boiled%20cassava%20%282%29.pdf
Tags:	Add Tag No Tags, Be the first to tag this record!