Cassava cooking properties characterization using NIRS on fresh ground cassava
This study concerns 1101 genotypes from several breeding populations harvested and analyzed at CIAT (Colombia) over 4 years between 2019 and 2022. The near-infrared spectra of the mashed fresh roots of all the genotypes were recorded according to the SOP (Belalcazar John, 2020) developed within the RTBfoods project. Water absorption at 30 minutes boiling was measured according to the SOP developed by CIAT (Escobar Salamanca Andrés Felipe, 2022). The distribution of the WA30 values is asymmetric on the left, 1692 samples (58%) presented WA30 values lower than 12%, and 1213 samples (42%) presented WA30 >12%. The dispersion of the data increases with year, the WA30 standard deviations varies from 5,1% (2019) to 10,4% (2022), the variability of WA30 values increases by a factor 2. This may reflect the increasing diversity of the populations screened by the CIAT, with more progeny populations being screened by WA30. The WA30 value is inversely correlated to cooking time, therefore samples with high values of WA30 correspond to candidate genotypes with low cooking time (CT) as well as softer, more mealy texture. A limit between low and long cooking time genotypes can be arbitrarily set at 12% for WA30, corresponding to ~35 minutes cooking. Samples lower or equal to 12% of WA30 can be considered as poor genotype (too long CT) and samples with WA30 values higher than 12% correspond to suitable genotypes for end user with a low CT. Two sets of data were constituted: one for learning, tuning the parameters, and one for testing, evaluating the error of the predictive model. To do this, 70% (n = 2033) of the samples were randomly selected for learning set and the remaining 30% (n = 872) were used for testing. The repartitions per WA30 classes and year were maintained within the two sets. Two modelling strategies were investigated: 1) an indirect classification based on a regression step to predict WA30 using spectral fingerprints and then class attribution according to WA30 predicted values; 2) a direct classification using discriminant procedure based on classes defined by WA30 laboratory values and spectral fingerprints. The two approaches: regression (Ridge Regression) and classification (PLSRDA), based on different methods for regressor selection within the spectral data, lead to similar performances in terms of classification according to WA30 classes. Nevertheless, PLSRDA leads to better classification and is easier to implement and interpret. The classification accuracy is 81,4% when predicting test set with a sensitivity = 79,4% and a specificity of 82,9% and a false positive rate equal to 17,1% while false negative rate is 20,6%. This model is efficient and can be implemented in a selection scheme 1) as is if the next year/generation variability remains similar to the current database 2) or with controlled update if next year/generation variability differs from current database (e.g. if some samples have WA30 values higher than the range already in the database).
| Main Authors: | , , , , |
|---|---|
| Format: | monograph biblioteca |
| Language: | eng |
| Published: |
RTBfoods Project
|
| Online Access: | http://agritrop.cirad.fr/603459/ http://agritrop.cirad.fr/603459/1/RTBfoods_NIRS%20calibration_Cooking%20properties_Fresh%20ground%20cassava.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|