Non-Destructive Viability Discrimination for Individual Scutellaria baicalensis Seeds Based on High-Throughput Phenotyping and Machine Learning

Abstract

It is crucial to identify and select high-quality seeds for improving Scutellaria baicalensis yield. In this study, we present a non-destructive and accurate method for predicting Scutellaria baicalensis seed viability that used seed phenotypic data with machine-learning algorithms to distinguish between vital and dead seeds. Meanwhile, the SMOTE was used to balance the dataset and make the established viability discrimination model more efficient by avoiding problems of overfitting or under-fitting. The results showed that hyperspectral imaging (HSI) combined with detrend (DT) preprocessing and a support vector machine (SVM) model could predict Scutellaria baicalensis seed viability with a 93.3% accuracy, and increased the germination percentage of the seed lot to 99.1%, while machine vision imaging provided the highest 87.9% accuracy and 87.0% germination percentage. This strategy is suitable for large-scale Scutellaria baicalensis seed viability discrimination operations for ensuring seed quality, expanding the cultivation and production scales of Scutellaria baicalensis, and accelerating the present solving of the problem of short supply. It can help to accelerate the breeding of quality Scutellaria baicalensis varieties.

1. Introduction

Scutellaria baicalensis (Scutellaria baicalensis Georgi, or Huangqin in Chinese), as a major traditional Chinese medicine herb, contains a large number of active ingredients that have a variety of pharmacological effects (e.g., antiviral [1], antibacterial [2], anti-injury [3], anti-inflammatory [4], anti-tumor [5,6,7], etc.). Therefore, Scutellaria baicalensis is widely used in clinical treatments [8] and is used to treat nearly 160 diseases [9], causing great market demand. The limited wild resources alone cannot meet the need; the artificial cultivation of Scutellaria baicalensis, therefore, has been a priority [10]. High-quality seeds are the basis and prerequisite of standardized Scutellaria baicalensis production, and also an essential condition of high-quality production for Chinese medicinal materials [11]. However, the confusion about germplasm resources and the lack of improved variety has greatly hampered the development of the Scutellaria baicalensis industry [12]. Hence, it is urgent to study how to improve the germination ability of Scutellaria baicalensis seeds and ensure seed quality, which has an important realistic impact on expanding the cultivation and production scale of Scutellaria baicalensis, as well as accelerating the present solving of the problem of short supply [9].

Compared with crop seeds, Scutellaria baicalensis seeds are simply very small (long axis of about 2 mm). Accordingly, it is challenging to apply some routine seed viability testing methods to them, e.g., germination test, tetrazolium staining, or conductivity testing, which has drawbacks such as complicated operation and time-consuming procedures or are destructive to seeds [13,14]. There is hope that non-destructive testing may be a reliable tool, which can be used for the time-efficient and non-destructive viability determination and prediction of each Scutellaria baicalensis seed in the early stage of planting.

Due to its low-cost and easily-obtained characteristics, machine vision is the most widely used non-destructive seed quality testing method [15]. It can quickly obtain phenotypic features from seed surfaces, typically in the form of RGB images. When using those shape, color, or texture features, extracted from seed images, coupled with machine learning algorithms, seeds with different qualities can be efficiently classified [16,17,18,19,20,21]. With this strategy, single pepper seeds’ viability was conveniently and effectively discriminated, and the germination percentage was increased by about 20% [22].

Another non-destructive and high-throughput phenotyping method—hyperspectral imaging (HSI) can obtain spectral and spatial information simultaneously [23], which can reflect the external features and also the internal physical structure and chemical composition differences among samples [24]. This method has been successfully applied to differentiate and classify a variety of crops and agricultural products including seeds, according to detecting subtle differences in chemical composition and distribution [25,26,27]. Many pieces of research have proved that the application of HSI in the field of seed quality evaluation is successful. The research of Yang et al. [24] showed that hyperspectral image processing technology combined with machine learning algorithms could accurately and non-destructively predict the germination status of sugarbeet seeds, thus realizing the prediction of the sugarbeet seeds’ germination percentage. Their results also found that a support vector machine (SVM) algorithm performed better than k-nearest neighbor (KNN) and random forest (RF), with the best prediction accuracy of 95.5% using the full wavelength. When using the HSI to detect the viability of naturally-aged rice seeds, a SVM and a convolutional neural network (CNN) both performed well, with the best accuracy of over 85% [28]. In addition, HSI coupled with SVM or RF was adopted for rapid seed viability prediction of Sophora japonica [29,30] and Japanese mustard spinach seeds [31].

However, so far, there is barely any research on the germination prediction of Scutellaria baicalensis seeds using these non-destructive methods. This is mainly because Scutellaria baicalensis seeds are small and dark, causing a weak characteristic spectrum, therefore increasing the difficulties in obtaining phenotypic traits and the subsequent RGB or spectral image analysis. Given the existing problems, this study adopted machine vision and HSI separately to non-destructively obtain and extract Scutellaria baicalensis seeds’ phenotypic traits combined with machine learning algorithms to establish, and, finally, select an accurate model to realize the non-destructive viability discrimination (germination prediction) of Scutellaria baicalensis seeds.

2. Materials and Methods

2.1. Test Materials

A total of 600 Scutellaria baicalensis seeds were randomly selected as the materials in this study. The seeds were produced in Anguo City, Hebei province, China (38°42’48″ N, 115°33’30″ E); harvested in August 2020, and stored at room temperature (about 25 °C, RH 60%) during the experiments. After obtaining the seed phenotype information, a standard germination experiment was conducted to obtain the viability of each seed.

2.2. Seed Phenotypic Traits Acquisition

2.2.1. RGB Image Acquisition and Feature Extraction

When acquiring phenotypic traits, all seed samples were imaged directly, without any preconditioning. For the machine vision portion, 600 Scutellaria baicalensis seeds were placed neatly on a scanner (Microtek scanmaker i360, Microtek, Shanghai, China) without touching each other, to obtain scanned images. Then the RGB images were saved as tagged image file format (TIFF) with a resolution of 300 dpi. With that, the shape, color, and texture features (a total of 54 features) of each seed were extracted using AIseed (a software program developed by our lab and Nanjing AgriBrain Big Data Technology Co., Ltd. (Nanjing, China)). Meanwhile, all seeds were numbered.

2.2.2. Hyperspectral Reflectance Data Collection and Preprocessing

In the HSI portion, this study used a proto-type visible/near-infrared (VIS/NIR) hyperspectral imaging system to obtain hyperspectral images for each numbered Scutellaria baicalensis seed. The wavelength ranged from 311nm to 1090 nm, with a spectral resolution of 0.77 nm. The HSI system used was installed at the Beijing Key Laboratory of Crop Genetic Improvement, China Agricultural University [23], and equipped with a linear scanning V10E imaging spectrometer, a CCD camera, a 150 W halogen tungsten lamp, and a mobile platform (Isuzu Optics Corp., Hsinchu, Taiwan, China). All HSI images were acquired in a darkroom. We also collected a standard whiteboard with a white Teflon plate, and a standard blackboard by covering the lens cap of the spectrum camera in the same collection environment with the samples. After HSI image acquisition, all spectral images were corrected according to the following formula:

$$I_{RT}=\frac{I_{raw}-I_{dark}}{I_{white}-I_{dark}}$$

where I_RT represents the corrected hyperspectral image, I_raw represents the raw hyperspectral image, I_dark represents the standard blackboard image, and I_white represents the standard whiteboard image.

We used HSI Analyzer software (Isuzu Optics Corp, Hsinchu, Taiwan, China) to calibrate all the hyperspectral images and extract the reflectance data of each seed from the corrected hyperspectral images. After reflectance data collection, 311–400 nm and 1000–1090 nm were removed from the original data to avoid significant impacts on the reflectance bands at both ends of the hyperspectral reflectance spectrum, which are usually caused by stochastic noise. Consequently, a total of 765 reflectance data between 400 nm and 1000 nm were obtained for each seed, which served as input variables in the following analysis and modeling.

Even though we made a lot of efforts to avoid the spectra being affected by various adverse factors during the spectra collection process, there existed some unavoidable problems, such as background noise, baseline drift, stray light, and electronic noise [24]. Consequently, we chose standard normal variable (SNV) transformation, detrend (DT), and multiplicative scatter correction (MSC) to preprocess the raw spectral data. SNV can eliminate the stray light caused by different sizes of seed samples and uneven distribution [32]. The spectral errors caused by scattered light on the seed surface and the optical path change can be eliminated by MSC [33]. DT can effectively solve the problem of baseline drift of the original spectrum [34].

2.3. Germination Test

After completing seed phenotypic traits collection, a single kernel germination test was conducted, in order of number. The whole germination test took place over 11 days, to mark whether the seeds germinated or not, which was a surrogate measure of seed viability in this study. Statistics showed that 480 Scutellaria baicalensis seeds germinated (marked as “1”), and 120 Scutellaria baicalensis seeds did not germinate (marked as “0”).

2.4. Dataset Balancing

The imbalance in germinated and non-germinated seeds could cause underfitting problems. Theoretically, under-sampling the majority category or oversampling the minority category can be used for data balancing, but they make the information learned by the model lack generalization, leading to the problem of overfitting [24]. The synthetic minority over-sampling technique (SMOTE) can analyze the minority samples, generate artificial samples, and then interpolate them to the minority class. The specific details are consistent with those described in the article by Yang et al. [24]. After the SMOTE, the number of germinated and non-germinated samples were equal, both at 480.

2.5. Viability Discrimination Modle Establishment

In this study, we chose and compared two classic machine algorithms, RFand SVM, to achieve accurate discrimination and prediction of individual Scutellaria baicalensis seeds’ viability.

The decision tree is the basis classifier of RF, which can resample the same data set and then generate multiple similar basic classifiers. Finally, the RF can output the overall classification results using averaging or voting methods to process the classification and prediction results output by the basic classifiers [23,24]. For both machine vision and HSI datasets, the number of estimators was searched from 1 to 100 with a step of 10, and the max features were searched from 1 to 30 with a step of 1, with both through a ten-fold cross-validation operation and grid search.

The SVM used in this study was based on an RBF kernel to provide more efficient processing of the nonlinear phenotypic data. In addition, a ten-fold cross-validation operation and grid search program were carried out to calculate the optimal penalty coefficient c and the kernel parameter g. For the machine vision dataset, the optimal c was searched from −0.2 to 4 with a step of 0.2, and the g was searched from −0.002 to 0.05 with a step of 0.002. The searching ranges of c and g for the HSI data were set from 10 to 200 (step: 10) and from 1 to 50 (step: 5), respectively.

According to the introduction in Section 2.4, we used the SMOTE to balance the two categories of seeds. After the equalization treatment, there were 480 vital and 480 dead seeds, respectively. The total amount of 960 seed samples was randomly divided into the training and test sets, with a ratio of 3:1. In this study, the accuracy of the test set for predicting seed viability and the finally germination percentage (the percentage of truly germinated seeds among the predicted germinated seeds in the test set) were used as indicators to determine the optimal spectral preprocessing and modeling method. The entire seed phenotypic traits acquisition process involved in the experiment and the modeling part is shown in Figure 1.

2.6. Characteristic Wavelength Extraction

For the collected hyperspectral data containing 765 bands, the high-dimensional data included much redundant information, especially for adjacent wavelengths [24]. Therefore, it could be an effective method to extract the effective characteristic wavelength and reduce the dimensionality, for hyperspectral data processing. As a frequently chosen method, the successive projections algorithm (SPA) selects variables forward. SPA selects characteristic variables with the least collinearity and redundancy after several iterations, based on the principle of root mean square error (RMSE) minimization [34]. In this study, we used the SPA method to select the effective wavelength from the raw spectra of Scutellaria baicalensis seeds.

3. Results

3.1. Viability Discrimination Model Based on Machine Vision

To non-destructively predict, whether a Scutellaria baicalensis seed could germinate, we first tested machine vision technology to obtain the phenotypic information of each seed, and combined those features with different modeling algorithms to evaluate their abilities to distinguish vital and dead Scutellaria baicalensis seeds. As shown in Figure 2, we plotted the probability density distributions of 54 shape, color, and texture features extracted from vital and dead seeds’ RGB images. However, these features were mainly overlapping, indicating that relying on only some of these features could not provide enough information for distinguishing vital seeds from dead Scutellaria baicalensis seeds. More elaborate analytical methods may be needed to accurately classify them.

Accordingly, these features served as inputs for both the RF and SVM models to establish viability discrimination models for predicting vital and dead Scutellaria baicalensis seeds. The optimal parameters for the different models are indicated in

Table 1. The optimal parameters for different models based on machine vision and HSI data.

Dataset	Preprocessing	RF		SVM
		Number of Estimators	Max Features	c	g
Machine vision	/	91	19	1.6	0.004
HSI	Raw	71	23	150	16
	SNV	91	9	20	1
	DT	91	17	70	46
	SMC	71	17	160	26

. After using the SMOTE for dataset balancing, the original minority non-germinated samples were amplified to 480, which was consistent with the number of germinated samples. From Figure 3, it can be noted that the SVM model’s accuracy was 7.1% higher than that of RF, and its improvement in germination percentage was 2.7% higher than that of RF. However, on the whole, these results indicated that the models established by machine vision phenotypic data and machine learning algorithms were insufficient to efficiently predict the Scutellaria baicalensis seeds’ viability.

3.2. High-Accuracy Viability Discrimination Model Based on HSI Data

To improve the accuracy of viability discrimination for Scutellaria baicalensis seeds, VIS/NIR hyperspectral imaging system was then adopted to explore whether it could classify between vital and dead seeds using subtle differences, through spectral reflectance. After noise removal, the full wavelengths retained 756 variables between 400 and 1000 nm, which were used for the subsequent analysis. The raw and preprocessed spectra of the Scutellaria baicalensis seeds are shown in Figure 4. The raw spectral reflectance for all the Scutellaria baicalensis seeds was less than 0.35, the vital and dead seeds showed similar trends, and most vital and dead seeds were intertwined without obvious boundaries, especially at 400~7000 nm (Figure 4a,e). In addition, it seemed that most vital seeds had higher raw reflectance than dead seeds in the band of 750~1000 nm.

Preprocessing was conducted to remove high-frequency noise existing in the raw spectrum. There appeared an obvious absorption peak at 650 nm~750 nm, which increased significantly after DT treatment (Figure 4c,g). It can be seen from Figure 4 that the spectral curves of both vital and dead seeds were distributed closer and showed substantial overlap, after SNV, DT, and MSC preprocessing. Moreover, the relative differences between the two categories grew smaller no matter the spectrum, for every single seed (Figure 4b–d), or the average spectral curve (Figure 4f–h). This result thus indicated that distinguishing between the dead and live Scutellaria baicalensis seeds with spectral data only was still difficult in accuracy and efficiency.

We subsequently used the RF and SVM algorithms to analyze and establish models for the viability discrimination of Scutellaria baicalensis seeds. The best parameters of the different models are shown in Table 1. The full-band raw HSI data and the different spectral datasets preprocessed by SNV, DT, and MSC were used as the inputs in modeling, and the viability state (germinated, “1”, or non-germinated, “0”) of each seed sample was used as the output. The accuracies of the test sets for several different models are shown in Figure 5, and the germination percentages for each model are presented in Figure 6. As can be seen from Figure 5, the results of the test sets showed apparent differences in accuracy between the models, but all of them performed better than the model based on the machine vision dataset. In addition, compared comprehensively, the overall performance of the SVM model was better than that of RF, with the highest discrimination accuracy of 93.3% using the DT-preprocessed HSI data. However, none of the RF models’ accuracies was over 90%.

When considering the actual improvement in germination percentage, except for the raw-based RF model (83.2%), the remaining HSI-based models were better in improving the germination percentage (Figure 6), compared to models using machine vision data (Figure 3). At this time, according to the horizontal comparison, the germination percentage of the SVM model fluctuated little under different preprocessing methods, and better improvement in germination percentage results could be obtained than in all the RF models. Notably, SNV pretreatment seemed to perform better. The DT-based SVM model increased the germination percentage from the original 80.0% to 99.1%. When a comprehensive analysis is drawn, in this study, DT preprocessing was determined as the optimal pretreatment method, and SVM was selected as the best modeling method. Therefore, this study determined that the HSI-DT-SVM model was the final accurate and efficient Scutellaria baicalensis seed viability discrimination model.

3.3. Characteristic Wavelength Extraction for HSI Data

To reduce the possible adverse effects caused by high-dimensional full-wavelength data, this study selected the SPA to extract the most informative characteristic wavelength. During the SPA analysis, with increasing characteristic wavelengths, the RMSE value gradually decreased. In this study, when analyzing the raw full wavelengths, the RMSE changed slowly, which happened in numbers of characteristic wavelengths greater than six. Therefore, six characteristic wavelengths were obtained as follows: 401.1, 408.5, 716.4, 784.2, 876.4, and 968.7 nm, of which the distribution is shown in Figure 7. Then, we used the spectral data of these effective wavelengths as training data to establish the SVM model. However, the model accuracy for Scutellaria baicalensis seeds’ viability discrimination in the test set was as low as 73.8%, and the final germination percentage was 71.8%. This result indicated that the six characteristic wavelengths selected by SPA did not contain adequate information about the viability of Scutellaria baicalensis seeds, and thus could not replace the full wavelengths.

4. Discussion

The traditional Chinese medicine Scutellaria baicalensis has high medicinal value, leading to huge market demand. Owing to perennial digging, the wild resources of Scutellaria baicalensis are decreasing. A artificial planting also has the confusion of lack of high-quality Scutellaria baicalensis seeds [10], which limits the development of the Scutellaria baicalensis industry and the high-quality production of Chinese medicinal materials [11,12]. Hence, it is urgent to study how to select high-quality Scutellaria baicalensis seeds, to help accelerate the cultivation and production scales of Scutellaria baicalensis, and solve the problem of short supply [9].

Due to small particle size and dark appearance, it is difficult to assess Scutellaria baicalensis seed quality. In this study, we innovatively used non-destructive machine vision and HSI technology to obtain Scutellaria baicalensis seeds’ phenotypic traits (e.g., shape, color, texture features, and spectral data) combined with machine-learning algorithms, to establish an accurate and efficient viability discrimination model to predict vital seeds and thus improve the seed quality of Scutellaria baicalensis. The results showed that, compared with machine vision, hyperspectral imaging performed better, probably because HSI could not only collect information from the seed surface but also reflected subtle differences in the internal chemical composition of seeds through reflectance, which is consistent with the research of Tu et al. [23]. This proved that whether Scutellaria baicalensis seeds could germinate or not was closely related to their internal components, which ias confirmed in the study of Yang et al. [24]. Especially after DT pretreatment, an obvious absorption peak appeared at 650~750 nm, which reflected the vibration information of the N-H in the amino acid contents of Scutellaria baicalensis seeds [24].

Although the six characteristic wavelengths selected by SPA could not completely replace all the wavelengths in modeling, they could reflect the substances with the greatest difference between vital and dead Scutellaria baicalensis seeds. Among them, spectral reflectance in the 401 nm range is related to sucrose [35], which can not only act as a signal molecule coordinating cell division in the tip meristem but also promote the growth of the root. It was reported that reflectance in 683, 709, 714, and 740 nm had a strong relationship with vitamin C, which can promote metabolism and the synthesis of nucleic acid and protein, as well as accelerate the degradation of stored lipids [24]. The characteristic wavelength at 716.4 nm might indicate that the vitamin C content of dead and live Scutellaria baicalensis seeds were different. The characteristic band at 968.7 nm was close to strong glucose-associated bands (892 and 1000 nm), which is the energy source of seed life activities and could be used as a signal molecule to regulate seed growth and development [36,37].

In these HSI models, SVM models always performed better than RF. It can also be found that the model’s accuracy could be improved by using a preprocessing method when classifying the vital and dead Scutellaria baicalensis seeds [38]. The DT and SNV preprocessing methods obtained the highest accuracy and the highest germination improvement, respectively. Through comparing synthetically, SNV was determined as the optimal pretreatment method for the Scutellaria baicalensis seeds’ viability discrimination SVM model. Because SNV preprocessing obtained a model accuracy of 92.5% and improved the germination percentage to 100%, this meant that the model correctly discriminated all the vital Scutellaria baicalensis seeds and did not wrongly predict any dead seed as a vital one. We firmly believe that this non-destructive and efficient method for determining the viability of Scutellaria baicalensis seeds is of great significance for providing high-quality seeds and promoting the industrialization of Scutellaria baicalensis planting, which can also accelerate the breeding of high-quality varieties of Scutellaria baicalensis.

5. Conclusions

This study presents, to the best of our knowledge, the first description of a non-destructive prediction method for Scutellaria baicalensis seed viability based on hyperspectral-imaging data and an SVM model with SNV preprocessing. The SMOTE was used to balance the dataset, and to make the established viability discrimination model more efficient by avoiding the problem of overfitting or underfitting. The best full wavelength HSI-DT-based SVM model achieved a 93.3% accuracy in seed viability prediction, and increased the germination percentage of the seed lot to 99.1%, superior to the best model based on machine vision data (the best accuracy and germination percentage were 87.9% and 87.0%, respectively). These results showed that it was practicable to use hyperspectral imaging technology coupled with an SVM algorithm to non-destructively predict the germinability of Scutellaria baicalensis seeds, which could be significant for Scutellaria baicalensis breeding and yield increase. However, only one Scutellaria baicalensis seed lot was used in this study, the model, therefore, needs to be applied to other Scutellaria baicalensis seed lots for verifying the performance and updating in the future.