Individual wheat kernels vigor assessment based on NIR spectroscopy coupled with machine learning methodologies

Abstract

Knowledge of the seed vigor status of individual wheat kernels could provide scientific evidence for the screening of excellent germplasm and the breeding of seedlings. Although many factors collaborate to reduce or render seed vigor, many methods have been employed to detect individual kernel vigor. This study aims to demonstrate the feasibility for using near-infrared (NIR) spectroscopy to detect individual wheat seed vigor and determine suitable machine learning classification models. For this study, 1152 wheat kernel samples were selected, and five-sixths of the portion was treated by artificial aging (AA). All seeds spectra were acquired using a single-seed near-infrared system covering the spectral range of 1200–2400 nm. After NIR spectral collection, all kernels underwent a germination test to confirm their vigor. The spectral data from kernels within 3 germination days, 5 germination days and the non-germination kernels were further used for the development of three-category classification models. After pretreatment by using Savitzky-Golay (SG) second derivative-method and standard normal variate (SNV) correction, the high-dimension spectral data were smoothed, and then were reduced to select most effective wavelengths by two spectral dimensional reduction algorithms: principal component analysis (PCA) and successive projections algorithm (SPA). Four machine learning methodologies, support vector machine (SVM), extreme learning machine (ELM), random forest (RF) and adaptive boosting (AdaBoost) were combined with the two spectral dimensional reduction algorithms to build eight models to discriminate and predict each wheat kernel’s vigor. The results demonstrated that the eight three-category machine learning classification models developed with the two spectral dimensional reduction algorithms provided comparable results for individual wheat kernel vigor. The accuracies of the eight models were higher than 84.0%, and PCA-ELM and SPA-RF models afforded the two highest classification accuracies at 88.9% and 88.5%, respectively. The macro-average F1 of these two models were at the same level of 0.887, which means these two models had almost the same ability to assess kernel’s vigor. This study could serve as a major step towards the development of a fast and non-destructive high-throughput NIR-based sorting system of individual wheat kernel vigor determination for plant breeders, wheat quality inspectors, wheat processors, etc.