Image-based phenotyping of cassava roots for diversity studies and carotenoids prediction

Abstract

Phenotyping to quantify the total carotenoids content (TCC) is sensitive, time-consuming, tedious, and costly. The development of high-throughput phenotyping tools is essential for screening hundreds of cassava genotypes in a short period of time in the biofortification program. This study aimed to (i) use digital images to extract information on the pulp color of cassava roots and estimate correlations with TCC, and (ii) select predictive models for TCC using colorimetric indices. Red, green and blue images were captured in root samples from 228 biofortified genotypes and the difference in color was analyzed using L*, a*, b*, hue and chroma indices from the International Commission on Illumination (CIELAB) color system and lightness. Colorimetric data were used for principal component analysis (PCA), correlation and for developing prediction models for TCC based on regression and machine learning. A high positive correlation between TCC and the variables b* (r = 0.90) and chroma (r = 0.89) was identified, while the other correlations were median and negative, and the L* parameter did not present a significant correlation with TCC. In general, the accuracy of most prediction models (with all variables and only the most important ones) was high (R2 ranging from 0.81 to 0.94). However, the artificial neural network prediction model presented the best predictive ability (R2 = 0.94), associated with the smallest error in the TCC estimates (root-mean-square error of 0.24). The structure of the studied population revealed five groups and high genetic variability based on PCA regarding colorimetric indices and TCC. Our results demonstrated that the use of data obtained from digital image analysis is an economical, fast, and effective alternative for the development of TCC phenotyping tools in cassava roots with high predictive ability.

Fig 1. Scatter plots of the total carotenoid content and colorimetric indices.

Numerical values represent Pearson’s correlation coefficients between colorimetric indices and total carotenoid content. The blue line represents the 1:1 isoline.

Fig 2. Within-group sum of squares of 228 biofortified cassava genotypes based on phenotypic data of total carotenoid content and colorimetric indices (CIELAB).

The grouping criterion was based on the K-means clustering algorithm.

Fig 3. Principal Component Analysis (PCA) based on phenotypic data of total carotenoid content and colorimetric indices (CIELAB) evaluated in 228 biofortified cassava genotypes.

Fig 4. Boxplot of total carotenoid content and colorimetric indices for each cluster of 228 biofortified cassava genotypes based on principal component analysis.

Different letters represent significant differences between accession groups with p < 0.05 by the Tukey Honest Significant Difference test.

Fig 5. Relative importance of colorimetric indices for predicting total carotenoid content using 12 prediction models: Linear Regression with Forward Selection (LRFS), Linear Regression with Backwards Selection (LRBS), Ridge Regression (RR), Linear Regression with Stepwise Selection (LRSS), Generalized Linear Model with Stepwise Feature Selection (GLMSS), Random Forest (RF), Partial Least Squares (PLS), the Bayesian Lasso (BL), the Bayesian Blasso (BBL), Artificial Neural Network (ANN), Support vector machine (SVM), and Classification and regression trees (CART).

Fig 6. Relationship between observed and predicted values for total carotenoid content of cassava roots.

The prediction was performed based on 12 different models based on random cross-validation without test set (V-Random) (80/20% in training and validation, respectively). Artificial neural network (ANN), Bayesian Blasso (BBL), Bayesian Lasso (BL), classification and regression trees (CART), generalized linear model with stepwise feature selection (GLMSS), linear regression with backward selection (LRBS), linear regression with forward selection (LRFS), linear regression with stepwise selection (LRSS), partial least squares (PLS), random forest (RF), ridge regression (RR), and support vector machine (SVM) were calculated. Each cross-validation fold is represented in different colors. The numerical data included in the graphs represent: linear equations (y) and coefficient of determination (R²).

Fig 7. Relationship between observed and predicted values for total carotenoid content of cassava roots.

The prediction was performed based on 12 different models based on random cross-validation with test set (IV-Random) (60/20/20% in training, validation, and test set respectively). Artificial neural network (ANN), Bayesian Blasso (BBL), Bayesian Lasso (BL), classification and regression trees (CART), generalized linear model with stepwise feature selection (GLMSS), linear regression with backward selection (LRBS), linear regression with forward selection (LRFS), linear regression with stepwise selection (LRSS), partial least squares (PLS), random forest (RF), ridge regression (RR), and Support vector machine (SVM) were calculated. Each cross-validation fold is represented in different colors. The numerical data included in the graphs represent: linear equations (y) and coefficient of determination (R²).

Fig 8. Relationship between observed and predicted values for total carotenoid content of cassava roots.

The prediction was performed based on 12 different models based on PCA clustering-based with k = 5 (IV-Cluster). Artificial Neural Network (ANN), Bayesian Blasso (BBL), Bayesian Lasso (BL), Classification and Regression Trees (CART), Generalized Linear Model with Stepwise Feature Selection (GLMSS), and Linear Regression with Backward Selection (LRBS). The numerical data included in the graphs represent: linear equations (y) and coefficient of determination (R²).

Fig 9. Relationship between observed and predicted values for total carotenoid content of cassava roots.

The prediction was performed based on 12 different models based on PCA clustering-based with k = 5 (IV-Cluster). Linear Regression with Forward Selection (LRFS), Linear Regression with Stepwise Selection (LRSS), Partial Least Squares (PLS), Random Forest (RF), Ridge Regression (RR), and Support Vector Machine (SVM). The numerical data included in the graphs represent: linear equations (y) and coefficient of determination (R²).