Near infrared spectroscopy models to predict sensory and texture traits of sweetpotato roots

High-throughput phenotyping technologies successfully employed in plant breeding and precision agriculture could facilitate the screening process for developing consumer-preferred traits. The current study evaluated the potential of near infrared (NIR) spectroscopy to predict visual, aromatic, flavor, taste and texture traits of sweetpotatoes. The focus was to develop predicting models that would be cost-effective, efficient and high throughput. The roots of 207 sweetpotato genotypes from six agroecological zones of Uganda were collected from breeding trials. The spectra were collected in the wavelengths of 400 – 2500 nm at 2 nm intervals. Using the plsR package, the calibrations were carried out using external validation models. The best calibration equation between the sensory and texture reference values (10-point scales) and spectral data was identified based on the highest coefficient of determination (R 2 ) and smallest RMSE in calibration and validation. Of the visual traits, orange color intensity was well calibrated using NIR spectroscopy (r 2 val = 0.92, SEP = 0.92), and the model is sufficient for field application. Pumpkin aroma (r 2 val = 0.67, SEP = 0.33) was the highest predicted among the aromas. The pumpkin flavour model exhibited the highest coefficient of determination in the calibration (r 2 val = 0.52, SEP = 0.45) for the traits considered under flavor and taste. Different models for textural traits exhibited moderate calibration coefficients: mealiness (chalky/floury) by hand (r 2 val = 0.75; SEP = 1.31), crumbliness (r 2 val = 0.73, SEP = 1.21), moisture in mass (r 2 val = 0.73, SEP = 1.26), fracturability (r 2 val = 0.60, SEP = 1.52), hardness by hand (r 2 val = 0.61, SEP = 1.27) and dry matter (r 2 val = 0.70, SEP = 3.10). The range error ratio (RER) values were mostly >6.0. These models could be used for preliminary screening. The predictability of the traits varied among different modes of samples. Models could be improved with an increased range of reference values and/or exploiting the correlations between chemical compounds and sensory traits.