Prediction performance of NIR calibration models developed with different chemometric techniques to predict oil content in a single kernel of maize

Abstract

Determining the biochemical content of intact seeds without damaging them provides significant advantages in plant breeding programs. Determination of oil content is one of the most tedious analyses at single kernel level among biochemical analyses. Near infrared reflectance (NIR) spectroscopy is one of the methods that can be an alternative to biochemical analyses in order to determine the oil content at the single seed level without damaging the sample. The aim of this study was to develop calibration models that will enable the determination of oil content in a single maize kernel by means of NIR spectroscopy and to compare the predictive power of the models developed using different chemometric techniques. A total of 500 seeds from 10 different genotypes that differ from each other in terms of oil content (from 1.11% to 10.9%) were used as experimental material. Spectral data were collected between 8333 and 4166 cm−1 on a desktop NIR device. Prediction models were constructed using partial least squares regression (PLSR) and support vector machines (SVM) methods. The model development process was carried out in the SelectWave (https://bafr.shinyapps.io/SelectWave/) application and models (n = 360) were created to determine oil content at single seed level by using 5 different pretreatments, 4 different derivative options, and 9 different wavelength selection methods. Model robustness was evaluated for the calibration samples (n = 341), external validation samples (n = 98), and test samples (n = 50). The most successful prediction result was obtained from the SVM model with the pretreatment combination of None+SVM+None (RMSECal=0.46, R2Cal=95.11, RPDCal=4.53, RMSEVal=0.78, R2Val=84.50, RPDVal = 2.55, RMSETest=0.83, R2Test=82.59, RPDTest = 2.42). Results showed that oil content in single kernel of maize could be correctly predicted by NIR calibration models based on SVM method coupling with the pretreatment of None+SVM+None combination.