Since its domestication by humans 9,000 years ago, maize has played an increasingly diverse role in global agricultural systems and is an important food and feed crop in the world (Kennett et al., 2020). The purity and authenticity of maize hybrids is not only indication of seed identity, but also at least one of the most important indicators of seed quality. Related studies have shown that for every 1% reduction in seed purity, the yield of maize is reduced by 180 kg/ha (Zhao and Wang, 2013). Authenticity of maize inbred lines, on the other hand, is the critical factor in determining the authenticity of progeny hybrids (Frascaroli et al., 2007).

Currently popular detection methods, seed or seedling morphology testing is simple to operate, but the scope of application is narrow; field planting identification results are intuitive but the cycle time is long; protein fingerprinting technology experimental results are unstable; DNA fingerprinting technology is authoritative and accurate because of its detection of DNA molecular level differences; however, like protein fingerprinting technology, the method must be damaged samples, and requires a certain sample size, the required equipment, personnel costs and technology costs are higher, in addition to the testing process of the waste will also pollute the environment (Ertiro et al., 2015). Therefore, it is necessary to explore a more rapid, convenient, reliable and accurate technique for seed sample authenticity detection (ElMasry et al., 2019).

Near-infrared (NIR) spectroscopy is the qualitative and quantitative detection of hydrogen-containing groups (e.g. C-H, N-H, O-H, etc.) by analyzing the information of octave and combined frequency between them, which is widely used in the identification of the purity of crop seed varieties due to the advantages of non-destructive, rapid and high throughput (Osborne and Fearn, 1986). The PLS-DA classification model was able to achieve 80-100% accuracy in studies applying NIR spectroscopy to differentiate wheat seeds of different ploidy (Ziegler et al., 2016). In addition, previous studies have used chemometric algorithms such as multiple scattering correction (MSC), principal component analysis (PCA), and K-nearest neighbor algorithm (KNN) to test the authenticity of 520 rice samples of different quality grades and origins, and the classification results exceeded 90% (Teye et al., 2019). However, the former researchers mainly focused on factors including basic statistics, selection of classification methods, and data processing methods, while the effect of genetic differences between different samples, i.e. genetic distances, on the identification performance of the model has rarely been published, furthermore, there is as yet no systematic study on the correlation between genetic distances and relative spectral distances, and between the genetic clustering results and the spectral clustering results.

Therefore, in this experiment, different maize inbred lines and hybrids were selected to study the relationship between genetic distance and spectral distance among samples and the performance of the model discrimination. Specifically, the experiments included: 1) using spectral data of samples to build a SVM discrimination model combined with preprocessing; 2) using near-infrared spectral data to calculate the relative spectral distance among samples and carry out spectral clustering; 3) extracting DNA from samples, calculating genetic distance, and performing genetic clustering; 4) comparing and analyzing the results of the two clustering methods, as well as the relationships between relative spectral distance, genetic distance, and model discrimination performance among samples.

Firstly, in this work, the spectral data of single kernel of maize samples were acquired by portable near-infrared spectrometer (900~1700nm), combined with SGD and SNV data preprocessing algorithms to establish SVM classification model, and the average accuracy of inbred and hybrid samples were 93.6% and 93.7%, respectively, which proved the feasibility of this technology in the field of seed authenticity detection (McVey et al., 2021). NIR spectroscopy has medium energy and strong transmission ability, and mainly responds to the vibrational information of atomic groups such as C-H, N-H, O-H, and so on (Bec et al., 2021). While the nutritional qualities such as crude protein, crude fat, crude starch and lysine content ranged from 8.18% to 12.64%, 3.41% to 4.77%, 71.44% to 77.67% and 0.238% to 0.42% with coefficients of variation of 0.10%, 0.08%, 0.02% and 0.12% among different samples, respectively. This variation is reflected in the NIR spectral information, which in turn leads to excellent model performance.

Secondly, the clustering analysis of the hybrid samples indicates that there are differences in the content of biological macromolecules among different groups. These differences may be caused by the genetic characteristics of the samples, because the influence of genetic characteristics on seed chemical components accounts for about 18% of the variation (Ferreira et al., 2014). We therefore analyzed the biomolecule content of the hybrids and found differences in crude protein, crude starch, crude fat and lysine content between the four groups of samples from the spectral clustering results, as shown in Figure 8 . Previous studies on the identification of 25 maize materials through chemical composition and SSR markers showed that Mantel test results indicated a significant positive correlation between seed chemical composition and SSR molecular marker genetic distance (Sofy et al., 2020).

Thirdly, this study found that as the genetic distance increased, the relative spectral distance also increased, along with the accuracy of the model. The researchers found some kind of possible connection between the genetic and relative spectral distances between the samples and the performance of the model, and suggested that the discriminatory performance of the model improved as the genetic distance increased (Liu et al., 2015). However, the above survey involved only five maize inbred materials, while 35 inbreds and 33 hybrids were selected for comparative validation in this work; moreover, portable spectrometers as well as single grain detection are more suitable for online application scenarios of massive quantities and meet the actual needs of production (Pu et al., 2021).

Finally, 5.72% and 6.06% of the two clustering outcomes in the inbreds and hybrids, respectively, differed in this study, which may be mainly due to the effect of the performance of the spectroscopic instruments themselves (Agelet and Hurburgh, 2014). Some of the characteristic bands reflecting the inclusions of the samples may be missing in the wavelength range of 900~1700 nm, i.e., in the long-wave NIR spectral region of 1700~2500 nm, 1923 nm and 2009 nm carry information on the combined frequency of the O-H group in water and starch, whereas 2125 nm and 2173 nm carry information on the diploid frequency of the N-H group in proteins (Ozaki et al., 2007).

This work assessed the applicability of portable near-infrared spectroscopy to identify the authenticity of single maize seeds. The seed samples selected covered both inbred lines and hybrids, which ensures that the technique can be used as a rapid and non-destructive tool for multiple scenarios of maize seed detection in the field, in warehouses, and online by a diverse group of people such as breeders, producers, and inspectors.

The application of NIR spectroscopy to detect the authenticity of crop seeds is not only affected by instrument parameters, preprocessing algorithms, modelling approaches, etc., but the seed genetic distance, as an indicator reflecting the genetic background among the samples, also affects the discrimination performance of the identification model. In this work, a portable near-infrared spectrometer was applied to systematically study for the first time the relationship between genetic distance, relative spectral distance and model performance. In addition, the genetic clustering results performed consistently with the spectral clustering results. NIR spectroscopy can well reflect the genetic relationship between corn seeds. The calculation of relative spectral distance is faster than molecular marker detection methods, and it has lower requirements for operators than establishing classification models. Therefore, in the future, it can be achieved to quickly identify the genetic background of single seeds using spectral information of the samples, effectively shortening the time limit of crop breeding. This discovery brings good news to breeders and seed producers. In summary, NIR spectroscopy is of great value to modern crop breeding, variety identity, and purity sorting as an assistant means of genetic breeding.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

YQY: Data curation, Formal analysis, Software, Writing – original draft, Writing – review & editing. RH: Writing – original draft. DZ: Methodology, Writing – review & editing. BS: Formal analysis, Methodology, Writing – original draft. YLY: Funding acquisition, Writing – original draft. DK: Conceptualization, Funding acquisition, Project administration, Software, Supervision, Writing – review & editing.