Rapid discrimination of Brassica napus varieties using visible and Near-infrared (Vis-NIR) spectroscopy

2.1 Plant materials

Eight varieties of oilseed rape belonging to the B. napus L. variety were selected from the Korean penisula with the following local names: ‘Youngsan’ ‘Hanla’ ‘Joongmo 7002′ Joongmo7001′ ‘Naehan’ ‘Tamla’ ‘Tammi’. A commercial variety, ‘Westar’ (Fig. 1) was also procured from the National Agrobiodiversity Center, Jeonju, Republic of Korea. All the varieties were grown in soil pots at the greenhouse of the National Institute of Agricultural Sciences, Jeonju, Republic of Korea, during May–July 2020. A total of 80 plants were used in this study, 10 samples from each of the eight varieties (8 × 10 = 80).

Close

Fig. 1

The representative figures of eight Brassica napus varieties used in the study. A, Westar; B, Naehan; C, Youngsan; D, Joongmo 7001; E, Joongmo 7002; F, Tamla; G, Tammi; H, Hanla.

Export to PPT

2.2 Vis-NIR spectral data collection

The Vis-NIR diffuse reflectance spectra of intact leaves of eight B. napus varieties were acquired using a handheld integrated portable spectrum analyzer (FieldSpec HandHeld 2, ASD Inc., Longmont, CO, USA) in the range of 325–1075 nm with a stepping of 1.5 nm in reflectance mode (log/R). The spectra were taken on the fully inflated leaves' adaxial surface, which may easily capture light. In each group, 10 spectra were acquired from three distinct sections of the leaf blade. Each group yielded a total of 300 spectra (3 × 10 × 10 = 300) to use for further investigation. To remove unnecessary noise, the Vis-NIR device's optical window was placed directly on the leaf's face during each spectrum capture, ensuring that the sensor window was entirely covered.

2.3 Preprocessing and chemometric analysis

In general, the background signals appeared in the raw spectra of samples due to the system settings and external noise. Hence, to minimize spectral noise and increase the accuracy of modeling approaches, different preprocessing methods, namely, normalization (area), standard normal variate (SNV), and derivatives (Savitzky–Golay (first differentiation)) were applied. The effectiveness of preprocessing methods was compared with raw spectra. Unscrambler X software, version 10.5.1, was used to execute the preprocessing computations (CAMO ASA, Oslo, Norway).

Several machine learning approaches were utilized and compared for effective visualisation and discrimination of spectral data. RapidMiner Studios Version 9.0.002 (RapidMiner, Inc., Boston, MA, USA) was used for the modeling. In this study, deep learning, decision tree, support vector machine (SVM), random forest, generalized linear model, rapid large margin, Naive Bayes, and linear discriminant analysis were employed to discover the optimal modeling strategy with the highest classification accuracy. The Aquap2 package developed by Pollner and Kovacs, (2014) was also used to apply the different preprocessing techniques and to perform linear discriminant analysis in R-studio. The inputs for each method were the spectral data points, and the classes were the identifying labels for eight B. napus varieties. For the reliability of the models in predicting multiple sample types, cross-validation was performed. The data was divided into two sets for this purpose: a training set and a validation set. The training set had two-thirds of the data and the rest was used as the validation set. The data splitting was done three times in order to make sure that each sample was tested at least once in the calibration and validation set.

2.4 Statistical analysis

The influence of (1) the scatter correction method, (2) the eight machine learning methods, and (3) the interaction between preprocessing and machine learning methods was determined using one-way analysis of variance (ANOVA). Tukey's range test was employed as a mean comparison procedure with a significance level of p ≤ 0.05.

3.1 Spectral analysis and preprocessing

Fig. 2 shows the average Vis-NIR spectra obtained from the eight B. napus varieties. That includes raw spectra and preprocessed with three different preprocessing methods. The raw spectra (Fig. 2A) are the spectra with no changes. However, other spectra (Fig. 2B-2D) were preprocessed with three different methods, namely Savitzky–Golay, standard normal variate and normalization, respectively. There are numerous crossovers and overlapping across the eight samples (Fig. 2); in other words, the spectra of each variety are quite similar to those of other varieties. Consequently, discrimination of varieties directly based on absorbance spectra is difficult. Therefore, it is necessary to use machine learning methods for the effective discrimination of eight varieties. From 325 to 550 nm, the spectral curve is flat and between 550 and 600 nm there is a small peak before dropping back to their normal position. This demonstrates that the leaves actively absorb blue (400–500 nm) and red (680 nm) light while reflecting green light (550 nm) in the visible range (Li and He, 2008) which is responsible for chlorophylls and carotenoids (Xu et al. 2009; Smith et al. 2017). From 650 to 750 nm, there was a sharp increase in the peak that remained the highest absorbance value; later, there were no variations in the remaining wavelength until 1200 nm. The spectra were preprocessed to reduce systemic noise and emphasize differences between samples. Using a number of preprocessing methods simultaneously will help us obtain a greater degree of classification accuracy and will allow us to select the best preprocessing approach for each sample (Feng et al. 2017; Sohn et al. 2022). It is difficult to discriminate the plant varieties only with the spectra shown in Fig. 2. For effective discrimination, Vis-NIR spectroscopy was combined with several models and machine learning methods such as discriminant analysis and principal component analysis (PCA) (Sohn et al. 2021a). To investigate the qualitative differences between the eight B. napus varieties, PCA analysis was performed using raw spectra. PCA analysis is a powerful data mining technique. The principle of PCA is to determine the linear combinations of the initial variables that contribute to the differences between samples (Li et al. 2007). These combinations are referred to as principal components (PCs). As shown in Fig. 3A, all of the different PCs showed the same slight pattern of separation for the different samples in the PCA paired plot from PC1 to PC6, but PC1 vs PC2 (Fig. 3B) showed the most visual differences. Therefore, outlier detection was performed using these two PCs before initiating preprocessing for the machine learning methods. Generally, the computerized iterations make PC1 have the maximum amount of information and PC2 have the maximum share of residual information (Li et al. 2007).

Close

Fig. 2

Average raw and preprocessed spectra of eight B. napus varieties. Average raw (A) and preprocessed with different preprocessing methods, namely Savitzky–Golay (B), standard normal variate (C), and normalization (D).

Export to PPT

Close

Fig. 3

Principal component analyses based on the Vis-NIR spectra of eight B. napus varieties. Raw spectra have been used. (A) Paired blot; (B) axes are first and second principal components.

Export to PPT

3.2 Chemometric analysis for discrimination of eight B. napus varieties

The potential of visible-NIR spectroscopy to discriminate or identify plant varieties is based on leaf spectral properties related to biochemical composition and structure, which are influenced by a variety of factors such as plant species, leaf development or microclimate position of the plant, etc. (Jacquemoud and Ustin, 2001; Xu et al. 2009). To determine the most accurate method for distinguishing eight B. napus varieties, the classification accuracy of various chemometric methods combined with different preprocessing methods was assessed. Table 1 shows a summary of the classification accuracy for the various methods. The classification accuracy of the various methods ranged from 22.0 to 98.2 %. Both raw and preprocessed spectra analyzed using chemometric methods demonstrated effective discrimination with diverse classification accuracies.

In most chemometric analyses, however, preprocessed spectra were found to have a higher classification accuracy than raw spectra. In some cases, the use of raw spectra yielded very less classification accuracy with the use of Decision tree (21 %), Random Forest (22 %) and Naïve Bayes (28.9 %). The maximum classification accuracy (98.2 %) was achieved with the preprocessing method SNV in combination with SVM. As shown in Table 1, among different preprocessing methods, Savitzky-Golay found to have higher classification accuracy with the combination of all other chemometric methods with a range of 55.6 to 97.9 %. Among the different chemometric analyses, SVM, deep learning, linear discriminant analysis and fast large margin were found to have higher levels of classification accuracy (SVM/SNV, 98.2 %; deep learning/SG, 97.9 %; SVM-SG, 96.8 %; LDA-SNV, 95.2 %; FLM/SG, 95.1 %). Even when using the raw spectrum without preprocessing the data, the SVM model had a high accuracy of 78.8 %. The SVM is particularly well suited to high-dimensional data because the value of each attribute is arbitrary (Gaye et al. 2021). Overall, the combination of SVM and SNV was found to be more effective in the discrimination of eight B. napus varieties. The preprocessing method Savitzky-Golay (97.9 %) was the best preprocessing method for use with multiple chemometric methods. Previously, several reports used multiple preprocessing and chemometric methods for discriminating between of plant varieties. For the discrimination of potato tuber varieties, Yee et al. (2006) used NIR spectra in combination with LDA analysis, resulting in a classification accuracy of 93 %. Chen et al. (2007) studied the differentiation of three tea varieties using SVM. Similarly, Vis-NIR spectroscopy combined with artificial neural networks (ANN) discriminated tea plant varieties with an accuracy of 77.3 % (Li and He, 2008). Xu et al. (2009) used PCA, LDA, and discriminant partial least squares (DPLS) regression methods for the on-site discrimination of tomato varieties. The LDA plot for the discrimination of eight B. napus varieties is shown in Fig. 4. The variety “Hanla” was completely separated from the clusters of other varieties, while other clusters of the other seven varieties were closely placed. This suggests that the other seven varieties share a higher level of similarity in their biological composition, whereas “Hanla” shares very little with the other varieties. Previously, the LDA analysis has been used for the discrimination of several plant varieties, such as sprouted mung bean (Tjandra et al. 2021) and melon varieties (Li et al. 2019).

Close

Fig. 4

Linear discriminant analysis for the effective discrimination of eight B. napus varieties without confidence circles (A) and with confidence circles (B).

Export to PPT

3.3 Selection of significant preprocessing and chemometric methods for discrimination

The efficiency of preprocessing and machine learning approaches were statistically analyzed (Table 2). After cross-validation, the mean percentage of classification accuracy of each chemometric method combined with various preprocessing methods revealed the significant modeling for the discrimination of eight B. napus varieties (Table 2). The statistical analysis using ANOVA (Table 3) revealed that the sum of square and mean sum of square values had statistical significance at p ≤ 0.0001. However, when a combination of preprocessing and multiple machine learning approaches was used, there was no significance at p ≤ 0.0001 (p value of 0.0003). The confusion matrix illustrates the degree of error in the identification of the assessed plants, suggesting that SNV combined with SVM was the most accurate classification method (Table 4). Similar results were witnessed by the use of Vis-NIR spectroscopy in the discrimination of Amaranthus sp. (Sohn et al. 2021b).