The Mahalanobis distance is a kind of generalized squared distance, which is based on the theory of multivariate normal distribution and effectively takes into account the three parameters of mean, variance, and covariance, and it is a comprehensive indicator that can comprehensively describe the overall multivariate structure (32). Principal component analysis (PCA) is a method used in mathematics for dimensionality reduction, where a set of multiple variables are recombined into a new set of mutually unrelated composite variables by orthogonal transformation to reduce the number of variables, and the information of the original variables is represented by selecting the principal components (PCs) whose contribution accounts for the larger cumulative contribution (37). The combination of PCA with Mahalanobis distance, whereby principal component scores are used in place of the original data to calculate Mahalanobis distance, not only reflects all the data information but also compresses the number of variables participating in the calculation of Mahalanobis distance and ensures that there is no covariance in the M-matrix (38).

This study used PCA combined with Martensitic Distance for outlier rejection work on all pea spectral data before spectral pre-processing. MATLAB R2021a software was used for the procedure writing of this method.

When calculating the results using NIR spectroscopy, factors such as the particle size of the sample to be measured, the homogeneity of the internal structure of the sample, the stability of the sample itself, and the noise of the instrument itself during the detection of the sample will all have an impact on the results (39, 40). To reduce the influence of the intrinsic factors of the instrument and the sample itself on the accuracy and stability of the quantitative analysis model, improved the quality of acquired spectral data, it is necessary to use a highly selective method to pre-process the NIR spectral data first, and then build the corresponding quantitative analysis model on this basis.

Commonly used spectral preprocessing methods mainly include Smoothing, Normalization, Baseline, standard normal variable (SNV), Detrend, multiple scattering corrections (MSC), 1st Derivative, and 2nd Derivative (41). The derivative algorithm is used to eliminate the effect of the background or drift of the measuring instrument on the signal. The MSC and SNV transform belong to the scattering correction, which is used to eliminate the effect of scattering on the spectrum caused by the uneven distribution of particles and differences in particle size. Smoothing is mainly used to remove random noise from the spectral signal (33, 42, 43). In this study, two, three, or even four of the above seven single preprocessing methods were selected to combine for composite spectral preprocessing, and the best spectral preprocessing method was finally selected.

Near-infrared spectroscopy is an indirect analysis technique, through the establishment of a calibration model to quantitatively or qualitatively analyze unknown samples, the main process includes the establishment of the model and the detection of unknown samples. The specific NIR spectroscopy techniques analyzed are shown in Supplementary Figure S3 (44, 45). This flowchart was drawn by Figdraw¹.

Quantitative analysis aims to establish stable and reliable models for quantitative analysis. The modeling consists mainly of the choice of calibration methods and the determination of chemical values, the choice of stoichiometry, and the prediction of unknown samples. Among them, the selection of calibration methods is the core of chemometrics. Currently, the commonly used calibration methods are principal component regression (PCR), partial least squares (PLS), multiple linear regression (MLR), artificial neural network (ANN), and topology, TP, etc. (46). Partial least squares regression (PLSR) is a regression modeling method that addresses the interdependence between two sets of multi correlated variables. It is used to investigate the use of one set of variables (independent or predictor variables) to predict another set of variables (dependent or response variables) (3). PLSR is an appropriate methodology for modeling when the number of variables in the two groups is considerable and linearly correlated, and the number of observations is limited. Compared with traditional multiple linear regression, PLSR can effectively solve the problem of multicollinearity between variables, has the advantages of simple calculation, high prediction accuracy, and easy qualitative interpretation, and its constructed model can more accurately identify the effective information (3, 47). This technique has been successfully used to predict the functional properties of grains, such as the content of protein subunits, gelatinization, solubility, etc. (32). Therefore, in this study, the preprocessed spectral and chemical values were analyzed by PLSR.

In this experiment, the Kennard-stone algorithm (3:1) is used to divide the correction and validation set samples during the modeling process (13, 48): the Kennard-stone algorithm is one of the commonly used techniques for selecting the samples of the correction set and the samples of the validation set, which means that all the samples are considered to be the candidate samples of the correction set, and the first one to be selected is the one with the furthest Euclidean distance from a pair of samples is chosen first, then the two samples that are farthest as well as closest to the selected samples are found by calculating the Euclidean distance from each of the remaining samples to each of the known samples in the validation set, and these two samples are then selected into the correction set, and the process is repeated until the desired number of samples is reached. The remaining samples are used as the validation set to validate the model, and the procedure is repeated several times to obtain the correction sample correlation coefficient (R_c²), root mean standard error of calibration (RMSEC), residual predictive deviation of calibration (RPDC), prediction sample correlation coefficient (R_p²), root mean standard error of prediction (RMSEP), and residual predictive deviation of prediction (RPDP), which are the metrics for evaluating the performance of the model: R², the larger residual predictive deviation (RPD), the smaller square error of calibration (SEC), the standard error of cross-validation (SECV), the square error of prediction (SEP) indicate better model performance (39, 46, 49).

The method of correlation analysis was taken as Pearson’s correlation analysis as represented in Figure 2A: the closer the graph is to blue, the weaker the correlation between the indicators; the closer the graph is to red, the stronger the correlation between the indicators. The results show that pea height had a highly significant positive correlation with pea starch (r = 0.47), but a highly significant negative correlation with crude fiber (r = −0.51); seed shape also showed a significant positive correlation with crude fiber (r = 0.39), which is similar to the results of previous studies (53). Additionally, there was a significant negative correlation between pea starch and moisture (r = −0.53), crude fiber (r = −0.57), crude protein (r = −0.36); there was a significant positive correlation between crude protein and crude fiber (r = 0.64), crude fat (r = 0.66). This may be because the higher content of crude fiber aids in the efflux of proteins and fats, thereby increasing the protein extraction rate. Crude fiber was highly significantly and positively correlated with crude fat (r = 0.60). In summary, the selection of low-fat, high-protein pea varieties should consider height varieties, a trend also observed in sweet potato leaves (54). The heat map was mapped using ChiPlot (see Footnote 2).

According to the chord diagram (Figure 2B), it can be seen that there is a significant relationship between pea sensory quality and physicochemical and nutritional quality and that one indicator may constrain or contribute to changes in several indicators, so that trends in one indicator may indicate trends in several indicators the contents of basic components of different varieties of peas are correlated, with the largest pathways for starch and crude protein. At the same time, it also demonstrates that peas have a higher starch content and protein content.

The PCA was used to explore the relationship between the content of essential components in different pea samples. Figure 2C shows the PCA labeled plot (PCA scoring labeled plot combined with K-means clustering as an unsupervised clustering method). The sum of PC1 and PC2 was 51.2%. At the same time, As can be seen in the Figure 2C the pea starch was negatively correlated with moisture, crude fiber, crude protein, and crude fat because the angles of their arrows were all greater than 90°. This is similar to previous studies (36, 40–42). The study also found that crude protein was also negatively correlated with the thickness of pea seeds. The angles of crude protein, crude fat, and moisture are all less than 90°, indicating a positive correlation between these three indicators, which is consistent with previous findings (52, 55). In summary, pea varieties with higher crude protein content tend to be lower in starch. In addition, the arrows pointing to PC2 have the longest arrow lengths for pea height and starch, indicating that pea height and starch contribute the most to PC2.

K-means as an unsupervised method was used to cluster the peas based on all components analyzed. As shown in Figure 2C, the 30 pea varieties were categorized into three classes. Starch, protein, and fat are categorized into three separate classes. The blue dots represent the first category, which consists of six varieties from Hebei, Ningxia, Yunnan, and Sichuan, respectively. Compared with other experimental pea varieties, it is characterized by lower fat content, which can be classified into three levels according to the fat content in total, ≤5% for the first level, 5–9% for the second level, and ≥9% for the third level, and therefore it is recommended to use “Cui Lu 99” to make a low-fat functional food (56). The yellow triangle represents the second category, including 16 varieties from Hebei, Yunnan, Sichuan, and Shandong, which are characterized by higher starch content and larger seed shape and can be classified into three levels according to the starch content, with the content ≥55% for the first level, 45–55% for the second level, and ≤45% for the third level. It is therefore recommended that “Ba Wan 1” be used for canned goods, dried fruit, and other products, or processed into pea flour for bread, biscuits, and other products (56). Gray squares represent the third category, consisting of eight varieties from Shandong and Ningxia, which are characterized by high crude protein content and can be divided into three levels according to protein content, with ≥25% as the first level, 20–25% as the second level, and ≤20% as the third level. For this reason, it is recommended to use “Qi Zhen 76” as a pea protein powder or pea protein meat substitute to provide a source of protein for dieters or vegetarians (56). Specific pea variety screening can be seen in Table 3.

To extract information related to chemical composition from spectra and eliminate interfering factors, it is crucial to use appropriate spectral preprocessing methods to build a stable and reliable model. Therefore, several spectral preprocessing methods were investigated to improve the signal-to-noise effect, including single and combined preprocessing methods. These methods were evaluated based on correlation coefficients and standard errors to determine the best pre-processing method. The effects of different spectral pre-treatments on the content of essential components of peas are shown in Supplementary Tables S2–S6. The optimal preprocessing method for the crude protein model was Derivative (2nd), for the crude fiber model was Derivative (1st) + Derivative (2nd), and for the crude fat model was Derivative (1st) + Derivative (2nd) + Detrend. The best preprocessing method for the starch model is Derivative (2nd), and the best preprocessing method for the moisture model is Derivative (2nd).

To better and more intuitively evaluate the model’s generalization ability, this study utilized the optimal preprocessing method identified in section 3.6.1 to construct the model for pea basic component content. Specifically, RMSEC, RMSEP, R_c², and R_p² were used as the evaluation indexes of the model, in which the larger the values of R_p² and R_c², and the smaller the values of RMSEP and RMSEC, indicated the higher the model performance (18). The PLSR method was used to quantitatively analyze the data, aiming to identify the optimal regression curve and establish a quantitative analysis model for the basic components of peas.

To validate the accuracy of the pea basic ingredient content model, pea samples (n = 80) were validated using the Kennard-stone algorithm in which all pea samples (n = 80) were evenly divided into the calibration set samples and validation set samples in a ratio of 3:1, where 54 samples were used to construct the calibration model and 26 samples were used as validation set samples to verify the calibration model (48). The results are shown in Table 4. The pea starch R_C² was 0.99, RMSEC was 0.47 (Figure 4A); the pea crude protein R_C² was 0.98, RMSEC was 0.21 (Figure 4B); the pea moisture R_C² was 0.85, RMSEC was 0.90 (Figure 4C); the pea crude fiber R_C² was 0.93 and RMSEC was 0.25 (Figure 4D); the pea crude fat R_C² was 0.89 and RMSEC was 0.09 (Figure 4E); It was verified that the starch R_P² reached 0.77, with RMSEP was 0.86 (Figure 4A); the crude protein R_P² reached 0.77, with RMSEP was 0.24 (Figure 4B); the moisture R_P² was 0.71, with RMSEP was 1.00 (Figure 4C); the crude fiber R_P² was 0.72 and RMSEP was 0.59 (Figure 4D); the crude fat R_P² reached 0.88 and RMSEP was 0.25 (Figure 4E).

As shown in Figure 6, the pea varieties (n = 80) used for modeling were classified into three categories. This demonstrates that the high-throughput NIR technique can quickly hierarchically classify different pea varieties compared to traditional methods. Circle heatmaps are plotted using ChiPlot (see Footnote 2).

A new method using a combination of portable rapid quality testing and near-infrared spectroscopy has been developed for the fast and non-destructive detection of the basic components in peas, to understand the correlation between the quality of pea varieties quickly and thus to carry out the work of pea variety quality grading. The technique starts with a routine chemical analysis of the sensory quality and basic components of peas, followed by correlation and cluster analysis of the data, and finally modeling and validation using near-infrared spectroscopy data.

Correlation analysis results showed that pea starch exhibited a substantial negative correlation with moisture, crude fiber, and crude protein while demonstrating a highly significant positive correlation with the thickness of the pea seed thickness; additionally, pea protein showed a significant positive correlation with crude fiber and crude fat. The combined contribution of PC1 and PC2 in the PCA was 51.2%. Cluster analysis showed that the different pea varieties used in the experiment could be classified into three groups. Starch, crude protein, and fat were classified into three classes, and specialized processing pea varieties were selected. Subsequent modeling using PLSR identified optimal preprocessing methods for various constituents: for pea starch, the Derivative (2nd) method resulted in a model R_C² was 0.99 and RMSEP was 0.86; for crude protein, the Derivative (2nd) method yielded R_C² was 0.98 and RMSEP was 0.24; for moisture, the Derivative (2nd) method R_C² was 0.85 and RMSEP was 1.00; for crude fiber, a combination of first and second derivatives provided R_C² was 0.93 and RMSEP was 0.59; and for crude fat, a combination of first and second derivatives plus detrending resulted in R_C² was 0.89 and RMSEP was 0.25.

This study has successfully demonstrated the use of near-infrared spectroscopy for the simultaneous, rapid, and non-destructive detection of essential ingredient content in whole peas. The technique significantly reduces the reliance on extensive laboratory equipment and lowers costs compared to traditional methods. It eliminates the need for sample preparation, and it takes only 5 s to obtain comprehensive data on the basic ingredient content of peas, which saves a great deal of experimental time. This will allow a quicker understanding of the relevance of the quality and grading of the different varieties of peas, which will lead to the selection of specialized varieties for pea processing. The high efficiency, multifunctionality, and broad applicability of NIR detection technology offer substantial advantages, facilitating the work of breeding experts in developing specialized varieties and assisting processing enterprises in establishing dedicated raw material bases. At the same time, near-infrared spectroscopy technology can not only quickly detect the basic composition content of peas, but also realize the rapid, non-destructive detection of soybeans, lentils, chickpeas, chickpeas and other leguminous crops, and analyze the proteins, lipids, moisture, fibers and other nutrients in them, to provide a scientific basis for the quality assessment of leguminous crops. In addition, near-infrared spectroscopy can be used in breeding, harvesting and sorting to realize the monitoring of the dynamic trend of the internal composition of the fruit, as well as the determination of the optimal harvesting period, thus improving the superiority rate of the fruit and market competitiveness. In the future, NIR spectroscopy technology is expected to be applied in more agricultural product testing fields, such as agricultural product quality and safety testing, origin traceability and crop breeding. Meanwhile, combined with advanced technologies such as IoT and AI, NIR spectroscopy technology will further expand its application scenarios and advantages, injecting new vitality into the development of modern agriculture.