EVALUATION OF INDONESIAN LOCAL SOYBEAN BASED ON CHEMICAL CHARACTERISTICS AND VISIBLE - NEAR INFRARED SPECTRA WITH CHEMOMETRICS

PCA analysis of the comprehensive reflectance dataset showed distinct patterns based on different preparation methods, as indicated in Figure 4a. Generally, two main data groups were found depending on the treatment applied to the samples. These included Group I, consisting of intact, crumble, and flour, and Group II, comprising paste samples. Specifically, Group II accounted for 83% of the variance in PC-1 and 11% in PC-2. Most samples in Group I tended to have negative PC-2 values, primarily in Quadrants III and IV. A significant distinction appeared when water was added to soybean paste in Group II, resulting in positive PC-2 values in Quadrants I and II. This phenomenon showed a spectrum response significantly different from other treatments.

Various treatments showed distinct positions along the axes, with intact samples primarily located in the negative range of PC-1, spanning from <-25 to >-150. Meanwhile, crumble samples extended in the positive direction, ranging from >-50 to approximately 0. An outlier, representing the intact black soybean (K8) sample, was found at the left end of Quadrant III. The flour samples ranged from >-25 to <50 in terms of PC-1, with significant overlap.

In Figure 6(b), PCA loading plot showed a prominent peak at approximately 0.02 to 0.03 on PC-1 at 350 nm wavelength, which gradually decreased to 998 nm. Although no significant anomalies were observed, the peak showed wavelength with the most substantial effect on PCA grouping.

As shown in Figure 6a, the application of Multiplicative Scatter Correction (MNC) was used for optimal visualization to show the loading performance of PC-1 and PC-2 concerning genotypes. The distinction among the various groups became evident, primarily characterizing the yellow soybeans (K1-6), with significant contributions from greenish (K7) and black soybean (K8). Except for paste form, in the case of intact, crumble, and flour forms, soybean samples were dispersed across PC-2 negative region in Quadrants III and IV. Greenish soybean (K7) showed a more extensive spread, while black (K8) was positioned with the lowest PC-2 negative values.

Figure 6b shows the loading value performance of PCA results. Anomalies in PC-1 were evident in peaks and valleys spanning from 380 nm to 790 nm, while PC-2 ranged from 430 nm to 780 nm. In PC-1, approximately 82% of variances were described, and the line descended to -0.02 at its lowest point around 370 nm, followed by an increase to 0.02 at 460 nm. This value remained relatively constant until approximately 750 nm, which gradually decreased to -0.02 and continued steadily to the end of the range. The pattern observed in PC-2 followed a similar trajectory but with distinct lowest and peak points. The line initially descended from 380 nm to almost -0.05 at the lowest point and ascended to peak value of 0.02, which was maintained until 700 nm. Subsequently, a decline to -0.02 was observed, followed by an increase to approximately -0.01 at about 820 nm. These anomalies affected sample grouping, with pronounced deviations in the spectral regions of 380-470 nm and 700-800 nm due to variations in pigment levels, specifically carotene and chlorophyll, in the samples.

Black soybean is characterized by black outer skin, which contains anthocyanins, as documented by Yang et al. (2022). These anthocyanins significantly impact the interaction between black skin layer and light spectra, either reflecting a specific spectrum of light or being absorbed by particular molecules. Specifically, when examining the distribution of genotypes through PCA scores, some dissimilarities are frequently observed between paste and other forms.

The implementation of physical treatment led to significant patterns in PCA plots. The condition of intact soybean, which was not subjected to any physical treatment, primarily reflected the composition of outer skin, resulting in a marked divergence from other soybean types, particularly the yellow ones. In contrast, physical treatments such as grinding exposed the inner layers of the soybeans, leading to a more uniform composition compared to other varieties.

The results of PCA analysis showed that adding water served as a discriminative factor, segregating the samples into two clearly defined groups. Specifically, other preparation methods did not show different clusters, indicating that the specific crushing or sample size reduction had no significant influence on sample response to Vis-NIR spectrum.

Partial Least Square Regression (PLSR)

Table 3 shows the coefficient of determination (R2) and root mean square error (RMSE) of calibration (C) and prediction (P) of various quality parameters for intact, crumble, flour, and paste of soybean obtained from Partial Least Squares (PLS) analysis. Several parameters, including protein, fat, ash, carbohydrate, and water contents as well as a* color, yielded inadequate models. In all cases, R2 values were consistently below 0.7, showing relatively low performance levels. This showed that the parameters could not be effectively detected in the specific Vis-NIR range of observation.

Despite being a protein source, soybean showed a low R2 prediction across all groups, with the highest of 0.64 obtained in the paste form. Protein identification in spectroscopy is often detected by amino acids referring to the nitrogen content above 300 nm, specifically at 346 nm in the visible range. However, traces of histidine were found in the Dika nut, contributing to the failure to detect amino acid responses at the energy levels observed at 380-400 nm (Okoronkwo et al. 2017). R2 value for fat was also low at 0.50, but peaks related to fat were found at 537-770 nm without contributing to PLSR model. In contrast, Lapčíková et al. (2018) detected wavelength range of 300 to 550 nm for unsaturated fatty acid, with ash showing the highest R2 value of 0.37. Preece et al. (2009) reported significant wavelength for predicting ash content at 350, 360, 390, 410, 430, 480, and 630 nm. Based on the results, similar peaks were found at 460-590 nm, with low concentrations of ash contributing to PLSR model. Although Chandaka et al. (2012) reported that the significant 750 nm could determine the amount of carbohydrate, wavelength at 630-750 nm found in this study did not affect the PLSR mode. Wavelength contributing to PLSR model for determining water was found at 882 nm and 760–970 nm, in line with the second overtones of O-H bands, which indicated water absorption (Devianti et al. 2023). The calibration and prediction PLSR model for water was also low due to reduced intensity at wavelength above 900 nm which was typical Vis-NIR spectroscopy. Based on Figure 1, the model demonstrated low ability in detecting a* value, as indicated by green or yellow with a specific black color shown by soybean samples.

Crude fiber parameters, including chlorophyll, total carotene, and vitamin C showed effective model performance, achieving relatively high R2 values across all groups, ranging from 0.80 to 0.98. In this study, wavelength contributing to crude fiber appeared at 510-630 nm, while OIV (2021) reported the carboxymethyl cellulose absorbance at 540 nm.

Chlorophyll in soybean seeds is mainly chlorophyll-a (C55H70MgN4O5) and chlorophyll-b (C55H70MgN4O6). In the visible range, the absorption peaks of chlorophyll-a occur at 430 nm and 662 nm, while chlorophyll-b occur at 453 nm and 642 nm (Lopes et al. 2017; Shi et al. 2022; Yang et al. 2022; Zhu et al. 2018). Longoni et al. (2020) and Gebregziabher et al. (2022) reported chlorophyll and carotenes substances in the visible range, where R2 values in calibration and prediction played a significant role in specific sample classification. At 680 nm, soybean samples showed chlorophyll content that distinguished green color (Pahlawan et al. (2022). In the visible range, the absorption peaks of chlorophyll-a occur at 430 nm and 662 nm, while chlorophyll-b was found at 453 nm and 642 nm ( Lopes et al. 2017; Shi et al. 2022; Yang et al. 2022; Zhu et al. 2018). Peak in absorbance of around 670 nm found in the olive oil spectrum confirmed the presence of chlorophyll pigments (Lapčíková et al. 2018). At 500 nm, soybean samples indicated carotenoid content that distinguished yellow or darker (Pahlawan et al. 2022).

R2 values of vitamin C ranged from 0.87-0.98 for calibration and prediction, with the highest observed for intact samples, particularly K8, as shown in Table 2. However, there is uncertainty regarding the impact of soybean seeds coat as the primary factor contributing to the elevated vitamin C levels. According to (Riscahyani et al. 2019), the analysis of vitamin C can be obtained at the wavelength of 494 nm. The validation of the well-performing model ranged from 0.44 for L* parameter in the crumble sample to 0.96 for vitamin C in paste form. The only parameter with less affirmative validation was L* parameter, registering the lowest R2 value. However, the prediction accuracy at 0.71 indicated a more reliable model outcome, suggesting that physical treatments, such as milling, induced variations in the color characteristics of soybeans. The distinct darkening effect observed with water addition compared to the flour sample could be attributed to the interaction or bonding of water molecules with soybean particles.

PLSR model for determining total carotene yielded high R2 values for calibration and prediction at wavelength 390, 536, 870, and 990 nm. Moreover, the most relevant spectral variables for carotenoids were detected at the wavelength of 449, 448, and 450 nm (Afonso et al. 2017). The results also showed that PLSR model for predicting L* and b* values had high accuracy. The contrast between green (K7) and black (K8) soybeans was particularly pronounced, as indicated in the L* and b* parameters in the intact sample group, with R2 values of 0.80 and 0.92. This significant difference in the characteristics of whole soybeans enhanced the model performance, allowing for the effective differentiation of color (green and black) from other yellowish samples. Subsequently, the results were adjusted by considering R2 values for validation and prediction. L* indicated lightness value in the range 0 (pure black) to 100 (pure white), representing light intensity that increased the absorption at the 400-500 nm spectrum band (Afonso et al. (2017). The a* value was identified as reddish colored sample that presents three peaks of great absorption in the 400-500 nm region of the spectrum. Meanwhile, yellow colored sample represented in b* profile showed significant peaks at 400-500 nm (Afonso et al. 2017).

Table 3 also shows the validation results of the calibration and prediction models, specifically in terms of RMSE, including RMSEC and RMSEP. Most proximate parameters, such as protein, fat, ash, and water, did not correlate significantly. Regarding RMSE values, crude fiber parameters showed higher values in intact preparation mode, while paste form had the lowest RMSE. This phenomenon suggested that adding water to paste form significantly affected samples grouping.

Based on chlorophyll, the lowest RMSE was observed in the flour form, potentially indicating more uniform sample conditions between the outer and inner layers of soybean containing chlorophyll. A similar trend was observed for total carotene, with the lowest RMSE values in the flour and paste forms. This pattern was also found in vitamin C parameter, where the paste form showed the lowest RMSE value.

Regarding color parameters, physical treatment did not yield significant differences in RMSE. The comprehensive Mean Absolute Percentage Error (MAPE) assessment shown in Figure 8 further explored this aspect. The calibration model parameters suggested that the model accuracy was good for proximate parameters, while water showed less accurate measurements. In contrast, chlorophyll, total carotene, and L* were assessed as showing good accuracy, vitamin C and b* as reasonably accurate, and a* as not accurate.