Four hundred seventy-five accessions of cowpea consisting of indigenous and exotic collections were taken from MTS of National Gene Bank at ICAR-NBPGR, New Delhi, India, having high variability in seed morphology. Using Augmented Block Design accessions were grown in Issapur experimental farm, New Delhi, India, following standard agronomic practices (29). Matured and dried seeds were collected, and extraneous material was removed.

Using FOSS NIRS 6500, 475 accessions were scanned, and the reflectance spectrum was recorded from 400 to 2,400 nm. Hierarchical clustering was done by Ward's method using squared Euclidean distance of 5 on the normalized spectral data of 475 samples. Major clusters and sub-clusters were identified and separated in the same manner. A set of highly diverse 121 accessions were selected to cover the entire range of variability in the data set. The wet chemistry values of these accessions were used as reference values. These samples were homogenized, ground, and sieved through a 1 mm sieve in FOSS cyclotec, and the flour thus obtained was used for scanning and wet chemistry analysis.

FOSS NIRS 6500 spectrophotometer equipped with Win ISI Project Manager Software Version 1.50 was calibrated using a reference tile (100% white). Approximately 5 gm homogenized samples were loaded and scanned in a circular ring cup with a quartz window (3.8 cm and 1 mm thickness). The average spectrum was recorded by scanning the sample 32 times at 400–2,500 nm and was registered as log (1/R) at increments of 2 nm, where R is the respective reflectance.

For development of calibration and validation set, the accessions were arranged in ascending order and every second value was taken out to make the calibration set. Therefore, calibration and validation sets were obtained in the ratio of 2:1, which ensured uniform variability in both the sets (32). Thus, the accessions were divided into two sets for modeling, i.e., 81 accessions in the training set and 40 accessions in the validation set for all the traits.

Win ISI Project Manager Software Version 1.50 was used to develop calibration equations using multivariate analysis by regressing spectral data with laboratory values. MPLS regression with cross-validation was used to develop equations on the above software using full spectra. Various mathematical algorithms, such as SNV–DT (SNV with detrend) were used for scatter correction and pre-processing the spectral data for each biochemical parameter. Moreover, different mathematical treatments like “2,4,6,1”, “2,8,8,1”, “2,4,4,1”, “3,4,4,1”, and “2,8,8,1” were used to develop models where the first digit represents the order of derivative; the second digit represents the gap (data points), third and fourth digits represents the data point in first and second smoothening. The developed calibration equations were assessed by different parameters such as coefficient of determination (RSQ), standard error of cross-validation [SEC(V)], standard deviation (SD), one minus variance ratio (1-VR). SEC(V) of RSQ_internal was calculated by Win ISI Project Manager Software V 1.50. The portion of the variation in reference data that may be characterized by the variance in predicted data is displayed using RSQ. High RSQ and lower SEC models are superior to low RSQ and higher SEC values. Only cross-validation was insufficient to assess the accuracy of the models [i.e., with SEC(V) and 1-VR], so RSQ_external (coefficient of determination in external validation), bias (difference between predicted and reference values), SEP (standard error of performance), SEP(C) (corrected standard error of performance) and RPD (ratio of performance to deviation) values are used. RPD values are used for accuracy of MPLS models where if RPD <1.5, the model is not reliable, in between 1.5 and 2.0, it indicates the capacity of a model to distinguish high and low values, in between 2 and 2.5, indicating approximate quantitative prediction, in between 2.5 and 3.0, indicates good quality prediction and if it is >3.0, then the prediction is excellent (13).

All of the calibration and prediction was done using Win ISI III Project Manager Software Version 1.50, which applied various mathematical treatments based on spectral and analytical data. Reference and predicted values were monitored using Win ISI Project Manager Software V 1.50 with the developed equation. Using global statistical values like RSQ, slope, bias, RPD and SEP(C), the accuracy and predictive capacity of the model were evaluated. The coefficient of determination (RSQ_{internal/external}) was externally plotted using Veusz statistical package for graphs of all the biochemical parameters.

A paired sample t-test was performed between reference and predicted values at 95% confidence interval using Jamovi statistical software package v1.6.9 (33). Strict parallel analysis was performed to check the reliability of the developed models and the reliability score was calculated by IBM SPSS v17.3 between the predicted and laboratory validated samples.

Five nutritional traits were worked out for 121 diverse cowpea accessions, and descriptive statistics are given in Table 1, whereas the box and whisker plots to showcase the variability of each trait are given in Figure 1. Protein was found to be range from 20 to 27.8/100 g, which is the most important trait for any legume. It is within the interval previously reported from 17.4 to 31.7/100 g (34) but lower than the values of 28.1–31.8% for five different Ethiopian cultivars (35). The protein content of cowpea has essential amino acids like lysine, histidine, and aromatic amino acids (36, 37). Starch content varied from 26.7 to 38.7/100 g, which is in agreement with the range of 28.3–36.2/100 g reported (34) but is lower than compared to other legumes like black gram 45/100 g and red bean 46 /100 g (38). Starch content will be useful for making processed products like moin-moin and akara. TDF in our study significantly varied from 12.2 to 22.4/100 g with a mean of 17.3 /100 g, close to the reported values worked by Akissoe et al. (15.6/100 g) but lower than worked values of 27.4/100 g (39, 40). Higher TDF in legumes corresponds to the lowering of cardiovascular diseases, diabetes, obesity, etc. Anti-nutritional factors such as phenolic compounds like phenolic acids, tannins, flavonoids, and phytates were analyzed and found in the range of 0.08–0.545 GAE/100 and 0.583–1.62/100 g, respectively. These compounds, if present in higher quantities, can limit the bioavailability of divalent cations (41).

Combined NIRS spectra of 121 cowpea accessions in the range of 400–2,490 nm is given in Figure 2A. The bands result from overlapping absorption that corresponds to the combination and overtones of vibrational modes N–H, O–H, and C–H, found in proteins, fatty acids, and carbohydrates, respectively. The main absorption bands were observed at 1,196, 1,468, 1,736, 1,934, 2,100, 2,310, and 2,482 nm, as shown in Figure 2B. C–O and N–H stretch, found in the spectral region between 2,000 and 2,222 nm, denoting protein content (42). O–H group can also be found in 1,560–1,640 nm, which can be allocated to the O–H group associated with phytates (18). O–H stretch first overtone of hydroxyl phenol groups present in 1,430–1,470 nm, whereas O–H bending/stretching of polysaccharides was found near the peak of 1,920 nm. The third polysaccharide overtone stretched by asymmetric C–O–O stretch was found near 2,083 nm (43). Unclear peaks were observed near 2,304–2,352 nm; this wavelength characterizes fatty acids and oils; since cowpea has little number of fatty acids, the bands were unclear. Symmetric stretching (–CH) in methyl groups (–CH₃) found in a wavelength of 1,200 nm causing weak absorption bands. Similar bands were found in the study of rice flour and its quality properties using NIRS (44).

A calibration set is generally referred to as training set providing learning and training to build the model. Regression algorithms like MPLS, PLS, and PCR can be used for model development, but as compared to the PLS algorithm, MPLS is supposed to be more stable and accurate (24) and thus was employed in the present study. Both spectra and reference composition were used in the MPLS technique for generating equations, decreasing the effect of irrelevant large spectroscopic variations. Absorption levels are generally altered due to the variation in the scattering of light and path length variation caused by intervention in sample particles and light. Linear calibration and spectral interpretation of NIR spectra becomes very complex and difficult due to the alterations (13). Spectral pre-processing was employed to diminish the multiplicative effect of particles size and scattering including scatter correction and derivatization techniques (22). SNV works by removing the mean from each spectrum, followed by dividing the value of each signal by the SD of the entire spectrum to center it around zero. Along with SNV, detrend is another approach to correct behavior shift. SNV with detrend (DT) in the present study was employed to avoid any noise in the NIRS signal baseline. Table 2 summarizes the calibration models by MPLS for protein, starch, TDF, phenols, and phytates in homogenized cowpea flour. For the development of calibration equations for various parameters, several mathematical treatments like “2,4,4,1”, “2,4,6,1”, “2,8,8,1”, and “3,4,4,1” were finalized. Our calibration equation was based on the highest 1-VR and RSQ_internal, lowest SEC(V) values. Resolution of spectra can be improved by using derivatives 2 and 3, which eliminates baseline shifts and superimposed peaks. The signal-to-noise ratio in the spectral region caused due to erratic high-frequency perturbations can be improved by using gaps 4 and 8 and smoothening (S1, S2). Calibration equations were generated by removing a few outliers (<10), which occurred due to scanning or analytical errors and were removed. As given in the Table 1, RSQ_internal for different traits was obtained for protein (0.800), starch (0.997), TDF (0.934), phenols (0.719) and phytates (0.985) for the given mathematical treatments “2,4,6,1”, “2,8,8,1”, “2,4,4,1”, “3,4,4,1”, and “2,8,8,1” respectively.

The external validation statistics for the given traits, protein, starch, TDF, phenols, and phytates of 40 samples are shown in Table 3. No outliers were removed in external validation to show higher prediction power and ensure the robustness of developed models. The best fit models were chosen based on higher RSQ_external, RPD, and low SEP, SD, slope, and bias values. To authenticate the model's validity, RPD value was used, which considers both SEP and variation in values and is more precise than SEP(C) (45). Towett et al. (14) found an RSQ_external of 0.93 for crude protein in cowpea leaves, whereas Pande and Mishra (18) found an RSQ_external of 0.97 for phytates in green gram seeds using FT-NIRS. The regression plot of predicted values versus reference values for protein, starch, TDF, phenol and phytic acid are described in Figure 3 respectively. These plots of predicted values versus reference values were developed through Win ISI III project manager V 1.50. Models for proteins, starch, TDF, phenols, and phytates displayed RPD values of 2.80, 5.32, 3.28, 1.78, and 4.45, respectively, denoting the model's excellent prediction power. Our RPD values for protein are in agreement with Ishikawa et al. (2.88) in cowpea (12). RPD value for phenolics (1.78) which shows that the model can distinguish between higher and lower values. Slope denotes the change in predicted values with a unit change in reference values. The ideal value of the slope is 1, and any value close to 1 would indicate an accurate model. The values of slope in our study for different traits were protein (1.12), starch (1.03), TDF (0.954), phenols (1.18), and phytates (0.929). In determining the model accuracy, bias is an important indicator of similarity between reference and predicted values of the model (23). When the reference and predicted values are the same, the bias would be equal to zero, which is the ideal value for bias. An underestimating model will be signified by negative bias, and overestimating model will be signified by positive bias (46). The values of bias for different traits were 0.197 (protein), 0.029 (starch), −0.026 (TDF), phenols 0.003 (phenols), and 0.009 (phytates), where all the developed models were found to be overestimating except TDF which is found to be underestimating.

To determine whether the mean of a dependent variable is the same as the analytical and predicted values for the examined biochemical parameters, a paired t-test with a 95% confidence interval was performed. In our study, the p-value came out more than 0.05, indicating the accuracy and reliability of the models (Table 4). The p-values of protein, starch, TDF, phenols, and phytates are 0.158, 0.512, 0.637, 0.639, and 0.114 respectively. Hence no statistically significant differences came out between the means in the NIRS method and standard methods used to analyze the traits.

The validated models were used to predict a sample set of 202 cowpea accessions for use in screening germplasm resources. To ensure the reliability of developed equations, 10% of samples were analyzed by standard methods. The predicted data were arranged in ascending order, and every 11th sample was selected for wet chemistry analysis. The result of predicted and laboratory values are given in Table 5. Correlation studies between the datasets show high correlation for protein (r = 0.97, p < 0.001), starch (r = 0.93, p < 0.001), TDF (r = 0.95, p < 0.001) while slightly low correlation was observed for phenols (r = 0.55, p < 0.01) and phytic acid (r = 0.77, p < 0.01) (Table 6). Strict parallel analysis is used to determine the difference in the means and standard deviations of two datasets where in our study there was agreement in the consistency of results, showing high reliability scores (unbiased) for protein (0.91), 0.92 (starch), 0.97 (TDF), 0.87 (phytic acid) and comparatively low reliability for phenols (0.64) (Table 6).