Vigor Detection for Naturally Aged Soybean Seeds Based on Polarized Hyperspectral Imaging Combined with Ensemble Learning Algorithm

Abstract

To satisfy the increasing demand for soybeans, identifying and sorting high-vigor seeds before sowing is an effective way to improve the yield. Polarized hyperspectral imaging (PHI) technology is here proposed as a rapid, non-destructive method for detecting the vigor of naturally aged soybean seeds. First, the spectrum of 396.1–1044.1 nm was collected to automatically extract the region of interest (ROI). Then, first derivative (FD), Savitzky–Golay (SG), multiplicative scatter correction (MSC), and standard normal variate (SNV) preprocessed hyperspectral and polarized hyperspectral data (0°, 45°, 90°, and 135°) for the soybean seeds was obtained. Finally, the seed vigor prediction model based on polarized hyperspectral components such as I, Q, and U was constructed, and partial least squares regression (PLSR), back-propagation neural network (BPNN), generalized regression neural network (GRNN), support vector regression (SVR), random forest (RF), and blending ensemble learning were applied for modeling analysis. The results showed that the prediction accuracy when using PHI was improved to 93.36%, higher than that for the hyperspectral technique, with a prediction accuracy up to 97.17%, 98.25%, and 97.55% when using the polarization component of I, Q, and U, respectively.

1. Introduction

Soybean, known as the ‘meat of the field’, is one of the world’s most economically and nutritionally important crops, containing 40% protein, 20% oil [1], and isoflavones; it dominates global oilseed production, and it is one of the world’s most important protein sources and also the primary protein supply in animal feed [2]. According to the growth rate, the world’s population is expected to reach 9.8 billion in 2050 and 11.2 billion in 2100. The increasing world population means that the world needs more soybean production every year [3]. However, crop yield largely depends on seed vigor, which is an important indicator of seed characteristics, including the seed germination rate, field emergence rate, and lodging resistance of sprouted plants [4]. In other words, seed vigor directly affects the yield production.

In general, most performance factors of the seeds, such as germination, vigor, emergence speed, and lodging resistance, will decline with the increase in storage time, especially for short-lived seeds like soybean, resulting in crop loss [5]. Therefore, accurate detection of soybean seed vigor before sowing can avoid the problem of non-germination caused by low seed vigor, ensure seed batch vigor, improve seed yield, increase storage time, and benefit agricultural producers. In addition, because the level of seed vigor is mainly affected by its own genetic characteristics, it is possible to select the seeds with high vigor to breed new varieties with higher vigor and improve the vigor of seeds from a genetic perspective. Hence, soybean seed vigor detection is very important for both farmers and seed breeding researchers.

Traditional seed viability detection methods, such as the tetrazolium chloride staining method, pH, and the potassium content and conductivity (EC) determination method [6], can meet the basic needs of seed germination detection, but require a long operating time, can damage seeds, may require additional samples, and cannot achieve rapid and contact-free on-line non-destructive detection. Therefore, in view of the above problems, the emergence of a variety of spectroscopic techniques provides a direction for scholars to carry out non-contact, rapid, and non-destructive pipeline detection research on seed quality. Machine vision was first used to replace human observation for seed quality detection to achieve a rapid assessment of the quality of seed, but it is limited to the detection of the external physical characteristics of seeds, and the material information inside the seeds cannot be known [7]. Near-infrared spectroscopy is also an efficient contactless detection technology which can recognize the quality of seeds [8]. Because the internal content of seeds will change correspondingly during the aging process, the subtle differences in the near-infrared (NIR) spectra of seeds with different vigor can be amplified by a pretreatment method, and the analysis and determination of the substance composition or content in the seeds can be realized after the extraction and modeling of characteristic bands. NIR technology, unlike machine vision, can identify the internal components and content of seeds and obtain more complete seed information, allowing for a qualitative and quantitative analysis of seed vigor. Because the amount of chlorophyll in the seed coat decreases as the storage time increases, fluorescence was used to measure seed vigor based on the content of chlorophyll in the seed coat; however, there are seeds that are high-vigor despite being stored for a long time. In recent years, with the development of detection technology, hyperspectral technology has been used for seed quality detection because it can simultaneously detect the appearance and epidermal substance components of seeds through image and spectrum fusion [9].

Polarized hyperspectral imaging (PHI) incorporates polarization detection based on hyperspectral analysis and has evolved into a new detection technology that not only retains the advantages of hyperspectral imaging technology but also can simultaneously obtain multiple pieces of information such as spatial, intensity, and polarization information about experimental objects, so as to achieve more refined detection and recognition of experimental objects even in a more complex background [10]. Currently, PHI, as a rapid development in recent years and a new detection technology, has played an important role in many fields, such as in the field of medicine, where it can be used to evaluate collected skin images, detect and inspect the presence of diabetes and its complications [11], and provides a new direction for detecting infectious skin diseases. In other words, doctors and patients can quickly detect a condition without contact, detect the oxygen content of the skin [12], and check the damage degree of caries [13]. In the field of environmental detection, the thickness of oil slicks on the sea [14] and water pollution can also be detected [15] to quickly judge water quality. In the field of military defense, PHI technology can be used to suppress the background, so as to highlight the intended objects in order to conduct rapid exploration of camouflaged military targets [16]. In the field of agriculture, it can not only detect defects in crops [17] and diseases in crops [18], but also detect the content of sugar and other substances in crops [19]. These research results all confirm the advantages, feasibility, and universality of PHI technology and provide a theoretical basis for the method of this study.

Although some scholars have made contributions to soybean vigor detection methods, the research on soybean seed vigor detection is relatively scarce compared with that for other kinds of seeds. Meanwhile, the research objective of seed vigor mainly includes studying naturally aging seeds and artificially aging seeds. Natural aging means that seeds are stored under certain environmental conditions and aging occurs naturally without human intervention, while artificial aging uses extreme methods such as high temperature and high humidity to accelerate seed aging in a short timeframe [20]. Although the artificial aging method is simple and controllable, it differs significantly from the natural aging conditions, and the results may also have deviations [21]. In addition, natural aging is a more direct way to measure the storage capacity of soybean, but there are relatively few studies on naturally aging seeds, especially for soybean seeds.

Therefore, this research used naturally aging soybean seeds stored for varying numbers of years as experimental subjects, and it was the first to apply PHI technology in the field of seed vigor detection, which is a no-contact method for quick and non-destructive vigor detection in large quantities of soybean seeds. We were able to verify that our method reached an ideal prediction accuracy for soybean seed vigor. At the same time, it provided a new solution for on-line mass assembly-line detection for seeds, which has great potential for seed vigor detection in the agricultural industry and research.

2. Materials and Methods

2.1. The Experimental Materials

The experimental subjects were three naturally aged soybeans with different harvest years (2017, 2018, and 2019), which were all produced in the experimental field of Nanjing Agricultural University (118.637503 E, 32.040436 N) and stored in the seed storage bank of Nanjing Agricultural University after harvest, which is one of the largest seed banks in China. To increase the robustness of experimental results, six soybean strains with three different colors were selected in this study, among which three were yellow (Gutian, 85 growing days, 100-grain weight of 17 g, hereinafter referred to as BY-1; Diandou 1, hereinafter referred to as BY-2, required 132 growing days and had a 100-grain weight of about 22 g; and Zhonghuang 1, which had 96 growing days and a 100-grain weight of about 19 g, was bred by crossbreeding Precocious 6 and Jindou 4 and is referred to as BY-3 below). The Huai Yangqing variety was selected for green and is referred to as BG below; the Wo Yanghei and Yan Yaohei bean varieties were selected for black and are referred to below as BB-1 and BB-2, respectively.

Due to the limited number of samples harvested and stored, the number of soybean experiments for each strain was not completely consistent. Among them, BY-1 had 40 granules per aging gradient for a total of 120 granules. BY-2 had 25 granules per aging gradient for a total of 75 granules. BY-3 had 25 grains in 2017 and 2018 and 21 grains in 2019, for a total of 71 grains. BG, BB-1, and BB-2 varieties all had 40 grains per aging gradient, with a total of 120 grains per variety. The soybean seeds were randomly selected according to the requirements of uniform size, full grain, no shrinkage, and no mildew. After the selection, the seeds were packed in kraft paper ziplocks, and the year and variety were marked on the bags. Then, the seeds were stored in the fresh-keeping area of the refrigerator in the laboratory (4 °C) for subsequent experimental use. The ambient humidity was controlled to be below 70%.

During the experiment, soybean samples were first taken from the refrigerator, and then one side of the soybean seeds was pasted using double-sided adhesive. The size of the soybean particles in accordance with the order of the year. Then, the soybean seeds were uniformly pasted on the black cardboard with the germ side up, from top to bottom, in the order of 2017, 2018, and 2019. The seed line name and natural aging year were marked on the back of the cardboard and stored. The setup of the test samples can be seen in Figure 1.

2.2. PHI Principle

Hyperspectral imaging technology is the integration of imaging and spectrum for seed target detection. The collected dataset is a 3D matrix, usually called hypercube, which not only has spectral information, but also spatial information of the seed to be tested [22]. Due to the dispersion effect, dozens or even hundreds of narrow band spectra can be formed for each spatial pixel, and the spectral information of the same pixel can draw a complete and continuous spectral curve; usually through the difference between the spectral curves, the seed viability can be effectively detected.

Polarization spectroscopy is a method to achieve non-contact measurements based on the difference in polarization characteristics of different substances [23]. Because light carries information about spatial intensity, radiation length, and change of polarization direction, when linearly polarized light penetrates a given seed, the polarization state of the beam transmitted or reflected by the detected target changes, so it can be used to analyze the seed to be tested [24]. The Jones vector, Stokes vector, and Muller matrix can usually characterize the polarization characteristics. However, since Stokes vector can represent both completely polarized light and non-polarized light, Stokes vector is often used to describe polarization imaging detection. The polarization characteristics of light waves are characterized by four real parameters representing the time average of light intensity. The BRDF parameter equation of polarized light is defined as follows [25].

$$\mathit{S}=\left[\begin{array}{c}{\mathit{S}}_0=\langle \left|{\mathit{E}}_{\mathrm{x}}^2(\mathit{t})\right|\rangle +〈\left|{\mathit{E}}_{\mathit{y}}^2(\mathit{t})\right|〉\\ {\mathit{S}}_1=\langle \left|{\mathit{E}}_{\mathrm{x}}^2(\mathit{t})\right|\rangle -〈\left|{\mathit{E}}_{\mathit{y}}^2(\mathit{t})\right|〉\\ {\mathit{S}}_2=2\langle \left|{\mathit{E}}_{\mathrm{x}}^2(\mathit{t})\right|\rangle 〈\left|{\mathit{E}}_{\mathit{y}}^2(\mathit{t})\right|〉\cos ({\mathit{\zeta }}_{\mathit{y}}(\mathit{t})-{\mathit{\zeta }}_{\mathit{x}}(\mathit{t}))\\ {\mathit{S}}_3=2\langle \left|{\mathit{E}}_{\mathrm{x}}^2(\mathit{t})\right|\rangle 〈\left|{\mathit{E}}_{\mathit{y}}^2(\mathit{t})\right|〉\sin ({\mathit{\zeta }}_{\mathit{y}}(\mathit{t})-{\mathit{\zeta }}_{\mathit{x}}(\mathit{t}))\end{array}\right]=\left[\begin{array}{c}{\mathit{S}}_0=\\ {\mathit{S}}_1=\\ {\mathit{S}}_2=\\ {\mathit{S}}_3=\end{array}\begin{array}{c}{\mathit{I}}_0={\mathit{I}}_{0^{°}}+{\mathit{I}}_{{90}^{°}}={\mathit{I}}_{{45}^{°}}-{\mathit{I}}_{{135}^{°}}\\ {\mathit{I}}_{0^{°}}-{\mathit{I}}_{{90}^{°}}\\ {\mathit{I}}_{{45}^{°}}-{\mathit{I}}_{{135}^{°}}\\ {\mathit{I}}_{\mathit{l}}-{\mathit{I}}_{\mathit{r}}\end{array}\right],$$

(1)

where ${\mathit{S}}_0$ is the total light intensity; ${\mathit{S}}_1$ and ${\mathit{S}}_2$ represent the linearly polarized light intensity in two different directions, where ${\mathit{S}}_1$ is the difference of the intensity of horizontal and vertical linear polarization, ${\mathit{S}}_2$ is the difference of the intensity of 45° and 135° linear polarization, and ${\mathit{S}}_3$ represents the circularly polarized light intensity; ${\mathit{E}}_{\mathit{x}}(\mathit{t})$ and ${\mathit{E}}_{\mathit{y}}(\mathit{t})$ represent the amplitude of the electric field in the X and Y directions; ${\mathit{\zeta }}_{\mathit{x}}(\mathit{t})$ and ${\mathit{\zeta }}_{\mathit{y}}(\mathit{t})$ represent the phase in the X and Y directions, respectively; 0°, 45°, 90°, and 135° represent the angles of the polarizer; ${\mathit{I}}_{\mathit{l}}$ and ${\mathit{I}}_{\mathit{r}}$ represent the intensity of left-handed (L) and right-handed (R) circularly polarized light respectively. To measure the polarization information of seeds, the Stokes vector parameters can be calculated first. The Stokes-parameters-defined ${\left[\begin{array}{cccc}{\mathit{S}}_0& {\mathit{S}}_1& {\mathit{S}}_2& {\mathit{S}}_3\end{array}\right]}^{\mathit{T}}$ is generally written as ${\left[\begin{array}{cccc}\mathit{I}& \mathit{Q}& \mathit{U}& \mathit{V}\end{array}\right]}^{\mathit{T}}$.

In practical measurement, because the circular polarization component ${\mathit{S}}_3$ of most ground objects can be ignored, the ${\mathit{S}}_3$ component is usually set to 0, and at least three polarization measurement data in different directions are needed to infer the complete polarization information of the target. According to the Stokes vector, the polarization degree and polarization phase of polarized light reflected from the target can be deduced:

$$\mathit{D}\mathit{o}\mathit{l}\mathit{p}=\frac{\sqrt{{\mathit{S}}_1^2+{\mathit{S}}_2^2}}{{\mathit{S}}_0},$$

(2)

$$\mathit{\psi }=\frac{1}{2}\arctan (\frac{{\mathit{S}}_1}{{\mathit{S}}_2}),$$

(3)

When a beam of polarized light interacts with the seed to be tested, the Stokes vector of the outgoing light and the incident light generally show a linear function relationship, namely

$${\mathit{S}}_{\mathit{o}}=\mathit{M}\cdot {\mathit{S}}_{\mathit{i}},$$

(4)

where the Stokes vector of the input beam ${\mathit{S}}_{\mathit{i}}$ can be used to construct, and the matrix ${\mathit{S}}_{\mathit{o}}$ is represented by the Stokes vector of the output beam after interacting with the seed [26].

$$\left\{\begin{array}{c}{\mathit{S}}_{\mathit{i}}=\left[\begin{array}{cccc}{\mathit{S}}_0^{\mathit{i},0}& {\mathit{S}}_0^{\mathit{i},1}& {\mathit{S}}_0^{\mathit{i},2}& {\mathit{S}}_0^{\mathit{i},3}\\ {\mathit{S}}_1^{\mathit{i},0}& {\mathit{S}}_1^{\mathit{i},1}& {\mathit{S}}_1^{\mathit{i},2}& {\mathit{S}}_1^{\mathit{i},3}\\ {\mathit{S}}_2^{\mathit{i},0}& {\mathit{S}}_2^{\mathit{i},1}& {\mathit{S}}_2^{\mathit{i},2}& {\mathit{S}}_3^{\mathit{i},3}\\ {\mathit{S}}_3^{\mathit{i},0}& {\mathit{S}}_3^{\mathit{i},1}& {\mathit{S}}_3^{\mathit{i},2}& {\mathit{S}}_3^{\mathit{i},3}\end{array}\right]\\ {\mathit{S}}_{\mathit{o}}=\left[\begin{array}{cccc}{\mathit{S}}_0^{\mathit{o},0}& {\mathit{S}}_0^{\mathit{o},1}& {\mathit{S}}_0^{\mathit{o},2}& {\mathit{S}}_0^{\mathit{o},3}\\ {\mathit{S}}_1^{\mathit{o},0}& {\mathit{S}}_1^{\mathit{o},1}& {\mathit{S}}_1^{\mathit{o},2}& {\mathit{S}}_1^{\mathit{o},3}\\ {\mathit{S}}_2^{\mathit{o},0}& {\mathit{S}}_2^{\mathit{o},1}& {\mathit{S}}_2^{\mathit{o},2}& {\mathit{S}}_3^{\mathit{o},3}\\ {\mathit{S}}_3^{\mathit{o},0}& {\mathit{S}}_3^{\mathit{o},1}& {\mathit{S}}_3^{\mathit{o},2}& {\mathit{S}}_3^{\mathit{o},3}\end{array}\right]\end{array}\right.,$$

(5)

At the same time, $\mathit{M}$ is called the Muller matrix, which is a 4 × 4 matrix. When ${\mathit{S}}_{\mathit{i}}$ can be inverted, the Muller matrix of the seed can be expressed as

$$\mathit{M}={\mathit{S}}_{\mathit{o}}{({\mathit{S}}_{\mathit{i}})}^{-1},$$

(6)

Muller matrix representation is also a common method to describe the change in the polarization state of light. When the characteristic information of the sample is detected, the light is passed through two polarizers, one of which is a polarization generator and the other is a polarization analyzer. When the biological material (seed) to be tested is placed between them, the polarization characteristics of the object can be expressed due to the existence of special substances with optical rotation properties [24]. PHI technology is the fusion of hyperspectral and polarization technologies, which can obtain the polarization information of seeds along with spectral information and three-dimensional spatial information.

2.3. The Experimental Device

The experimental platform was mainly composed of a dual hyperspectral camera system in 400–2500 nm (GaiaSorter-Dual, Shuanglihepu, Wuxi, China) and two linear film polarizers (GSP-50, Hengyangguangxue, Guangzhou, China). The main components of the hyperspectral instrument consisted of a uniform light source, a dual spectrum camera, an electronic control transfer module, a computer with control software, etc. The dual spectrum camera included two hyperspectral cameras, Camera 1 (Image-λ-V10E, wavelength range: 391.6–1044.1 nm, resolution: 2.5 nm) and Camera 2 (Image-λ-N25E, wavelength range: 1044.1–2528.1 nm, resolution: 5.6 nm). A schematic of the experimental device is shown in Figure 2.

As the range of the polarizer was 420–720 nm, only the range of 391.6–1044.1 nm was collected at the same time, and because a set of switches in the hyperspectral instrument was used to control the two lamps at the same time, the light source that was not used was shielded by opaque aluminum profiles.

2.4. Spectral Acquisition

First, the switch of the hyperspectral separator was turned on in accordance with the correct experimental specifications in the correct sequence. Because only one group of halogen lamps was used in this experiment, only one control switch of the halogen lamp was turned on, and 30 min of preheating was carried out before the experiment. Then, the SPECVIEW software (GaiaSorter-Dual, Shuanglihepu, Wuxi, China, https://www.dualix.com.cn/Goods/desc/id/211/aid/1084.html (accessed on 28 June 2023)) was used to focus the camera. The scanning distance was set to 25 cm, and the forward speed of the conveyor belt was set to 0.36 cm/s. After all parameters were set, the seed atlas was collected. To compare the improvement of seed viability detection accuracy with the PHI technique, common hyperspectral data were collected from six soybean strains without polarizing film. Then, a polarizer was installed in front of the lens and the light source, and the PHI information of each sample was collected when the polarizer was at 0°, 45°, 90°, and 135° in front of the rotating lens. At the same time, for the subsequent black-and-white correction processing, black-and-white frames were collected before each experiment and saved for further correction.

2.5. Image Analysis

2.5.1. Black-and-White Correction

Black-and-white correction was performed on the collected hyperspectral and PHI images to reduce the effect of dark current and uneven illumination on the experimental results. The correction formula was [27]

$$\mathit{R}=\frac{{\mathit{R}}_{\mathit{I}}-{\mathit{R}}_{\mathit{D}}}{{\mathit{R}}_{\mathit{W}}-{\mathit{R}}_{\mathit{D}}},$$

(7)

where ${\mathit{R}}_{\mathit{I}}$ is the original image, ${\mathit{R}}_{\mathit{D}}$ is the black frame, ${\mathit{R}}_{\mathit{W}}$ is the white frame, and $\mathit{R}$ is the final spectral image obtained after black-and-white correction.

2.5.2. Region-of-Interest Extraction

Region-of-interest (ROI) extraction is a key step in information analysis using image processing. The image contains a large amount of information, some of which is useless background information. If the analysis and processing are carried out with this background information included, not only is modeling time wasted, but the modeling will not be targeted and focused. Therefore, in most image processing studies, ROI positioning is normally used for extraction. The ROI extraction shape adopted in this study was oval, based on the shape of the soybean. After extracting the most important region of the obtained seed image, the processing analysis was carried out, which not only reduced the processing time but also improved the accuracy of seed vigor detection [28].

Through the hyperspectral imaging system, the spectral image cube of 1024 × 443 × 527 could be obtained. There were 1024 × 443 pixels, and each pixel had 256 spectral information. In this study, the steps of automatic extraction of seed spectral information by using ROI were as follows, and the specific process is shown in Figure 3.

Step 1: According to the obtained hyperspectral stereogram, six random seeds and background regions were extracted manually, and the corresponding mean value data were extracted. Then, the wavelength with the maximum reflected light difference was found through modeling as the characteristic wavelength of the ROI was automatically extracted.

Step 2: The grayscale image at the corresponding characteristic wavelength from the collected stereo image data was found.

Step 3: Final ROI mask image was obtained through threshold segmentation, median calculation, open and close operation, cleaning operation, and expansion corrosion operation.

Step 4: Seed location corresponding to the mask was found by a cell counting method and spectral cube.

Step 5: After extracting the position of each seed in the mask image, all spectral information was automatically extracted and saved after naming each seed.

2.6. Spectral Data Analysis

2.6.1. Spectral Pretreatment

To improve detection speed and accuracy, the vector data of hyperspectral and polarized hyperspectral data of each sample collected was smooth processed by moving average filter. The default window width of the moving average filter was 5. The first derivative (FD) [29], polynomial smoothing algorithm (Savitzky–Golay, SG) [30], standard normal variate (SNV) [31], and multiplicative scatter correction (MSC) [32] were used to process the PHI data of the soybean seeds.

SG performed polynomial least squares fitting on data within the moving window. The number of window points was set to 5 and the polynomial order was set to 2. The SNV algorithm processed each spectrum, and its calculation was essentially a standard normalization of the original spectral data; SNV was calculated by Equation (8):

$${\mathit{R}}_{\mathit{S}\mathit{N}\mathit{V}}=\frac{\mathit{R}-\overline{\mathit{R}}}{\sqrt{\frac{{\displaystyle \sum _{\mathit{i}=1}^{\mathit{p}}\left({\mathit{R}}_{\mathit{i}}-\mathit{R}\right)}}{\mathit{p}-1}}},$$

(8)

where $\mathit{R}$ was the original spectrum of a sample, $\overline{\mathit{R}}$ was the spectral average of all the wavelength points, and $\mathit{i}=1,2,\dots ,\mathit{p}$, $\mathit{p}$ was the number of wavelength points.

The derivative of waveband $\mathit{\lambda }$ was calculated by using the Equation (9):

$${\mathit{FD}}_{\left(\mathit{\lambda }\right)}=\frac{{\mathit{R}}_{\left(\mathit{\lambda }+1\right)}-{\mathit{R}}_{\left(\mathit{\lambda }-1\right)}}{\left(\mathit{\lambda }+1\right)-\left(\mathit{\lambda }-1\right)},$$

(9)

where ${\mathit{R}}_{\left(\mathit{\lambda }+1\right)}$ is the reflectance at the next waveband of $\mathit{\lambda }$, ${\mathit{R}}_{\left(\mathit{\lambda }-1\right)}$ is the reflectance at the last waveband of $\mathit{\lambda }$, $\mathit{\lambda }+1$ is the wavelength of the next waveband of $\mathit{\lambda }$, and $\mathit{\lambda }-1$ is the wavelength of the last waveband of $\mathit{\lambda }$.

2.6.2. CARS Feature Spectrum Extraction

The competitive adaptive reweighted sampling method (CARS) combines the exponential decay function and adaptive reweighting sampling technology to establish a partial least squares (PLS) model and remove wavelength points with small weights. Cross-validation was used to select the subset with the lowest root mean square error of cross-validation (RMSECV) as the optimal wavelength combination [33].

2.7. Germination Experiment

After collecting the polarized hyperspectral imaging data, a total of 360 soybean seeds of six varieties harvested in the past three years were selected randomly and subjected to a germination test in accordance with ISTA. The selected soybeans were soaked in 20 °C clean water for 12 h and then arranged in order according to the species and harvest year. To avoid the possible influence of seed placement on germination, the seeds were uniformly placed at a distance of 3 cm from each other and with the germ side facing up. At the same time, the germinating bed was marked to indicate the variety, year, and the beginning time of the germination experiment, and then the seeds were placed in the incubator with an artificial climate for culturing (20 °C). Finally, after seven days, the germination rate of the seeds was calculated statistically according to Equation (10), as shown in

Table 1. Germination rate results.

Variety	2017	2018	2019
BY-1	0.10	0.50	0.60
BY-2	0.65	0.80	0.90
BY-3	0.70	0.90	0.95
BG	0.25	0.45	0.60
BB-1	0.35	0.60	0.85
BB-2	0.40	0.55	0.75

.

$$\mathit{FYL}=\frac{\mathit{F}}{\mathit{S}}\times 100\%,$$

(10)

where $\mathit{FYL}$ is the germination rate, $\mathit{F}$ is the number of successful germination seeds, and $\mathit{S}$ is the total number of seeds used in seed germination rate detection.

2.8. Modeling Methods of Soybean Seed Vigor

One hyperspectral test and four polarized hyperspectral tests (0°, 45°, 90°, and 135°) were performed on 626 experimental samples. The seeds were mixed and divided into the training and test sets in a ratio of 3:1. After hyperspectral data and PHI data were preprocessed, five machine learning methods were used to predict soybean seed vigor: partial least squares regression (PLSR) [34], back-propagation neural network (BPNN) [35], generalized regression neural network (GRNN) [36], support vector regression (SVR) [37], and random forest (RF) [38]. All model methods were carried out on MATLAB 2022a, and the corresponding parameters are shown in

Table 2. Parameter settings of classifiers.

Classifiers	Parameters	Values
PLSR	principal component	nonlinear iterative partial least squares (NIPALS)
	robustness	root mean square error (RMSECV)
BPNN		MISO structure
		fully connected
GRNN	K-fold cross-validation	K = 10
	kernel function	radial basis kernel (RBF)
	learning rate	0.4
	expected error	4 × 10−5
SVR	kernel function	radial basis kernel (RBF)
	optimal penalty coefficient	c = 5.2
RF	number of trees (ntree)	predictive variables
	predictive variables (mtry)	predictive variables

.

(1)

Partial Least Squares Regression (PLSR)

In the present study, the nonlinear iterative partial least squares (NIPALS) algorithm was used to determine the principal component, and a single cross-validation (CV) was reserved. The root mean square error (RMSECV) was used to verify the robustness of the model.

(2)

Back-Propagation Neural Network (BPNN)

With this method, data normalization is first performed due to the MISO structure of the model. All the layers were fully connected, and there was no connection between them. The activation function was a sigmoid function. The expected error was 4 × 10⁻⁵, and learning rate was 0.4. The gradient descent method was used to calculate the weight coefficients.

(3)

Generalized Regression Neural Network (GRNN)

In the present study, the K-fold cross-validation method was used to train the neural network (K = 10). Moreover, according to the minimum mean square error combined with cyclic discriminant method, the optimal smooth factor spread value of the GRNN was determined (this was determined to be 0.3).

(4)

Support Vector Regression (SVR)

The parameters of the SVR model were set as follows: the radial basis kernel (RBF) was selected as the kernel function type; the optimal penalty coefficient was c = 5.2; and the optimal RBF kernel parameter g = 1.9 was determined by the network search method.

(5)

Random forest (RF)

Random forest has basically two tuning parameters: ntree and mtry. The parameter ntree is the number of trees to grow, and mtry is the number of variables randomly sampled as candidates at each split. The default value of mtry in model is $\sqrt{n}$, where n is the totals of variables. In the present study, the default values for ntree and mtry were used.

To further improve the use of PHI component parameters in mixture model prediction precision, the method of integrating blended learning was adopted; integration in this study involves merging multiple models to complete the study task. This method can be used in the seed vigor classification problem to predict regression problems, and the specific flow chart of the algorithm is shown in Figure 4.

The basic principle of the algorithm was to first detect seed vigor by using a variety of models and selecting the two models with the best modeling effect, namely RF and SVR, through the accuracy of their respective prediction sets. The seeds were mixed and divided into the training and test sets in a ratio of 3:1. The seed vigor prediction models were constructed using the PHI spectral data and germination rate of the training set as inputs, and the obtained prediction values were reset as the input of the new training set. Then, PHI spectral data information of the test set was substituted, respectively, according to the two constructed vitality prediction models, and then predicted values were redefined as the new test set. Finally, the germination rate of the new training set and the original training set were again taken as the input, and the SVR, with the most stable modeling, was selected for modeling, and the prediction result of the new test set was taken as the final prediction result of seed vigor.

As the blended integrated learning method has the ability to combine the advantages of various basic models, it can improve the accuracy manyfold compared with a single modeling algorithm, so it is increasingly favored by relevant scholars and has been proven to be effective in multiple fields [39]. Therefore, the algorithm was used in this study to improve the accuracy of seed mixture modeling. To evaluate the accuracy of the model, RC and RP were used to represent the training samples’ R-square and the test samples’ R-square. REC and REP were used to represent the training samples’ root mean square error and the test samples’ root mean square error.

3. Results and Discussion

3.1. Spectral Pretreatment Results and Analysis

The spectral range was 391.6–1044.1 nm, with 256 wavelengths in total. To compare the influence of the four pretreatment methods on subsequent modeling effects, the collected soybean seed hyperspectral data and PHI data were smoothed, followed by FD, SG, MSC, and SNV pretreatment, respectively. The pretreatment results for soybean PHI of all strains are shown in Figure 5 (taking the polarized angle of 135° as an example). The results showed that the seeds of each strain between different samples of spectra showed the same trend, at the same time, and the same color of the different categories of soybean seed strains also had a similar trend of the polarization spectrum curve. However, with different colors, there were certain differences between soybean seeds because seeds of different colors had different reflection characteristics [40], and differences in reflectance may also be related to the content of chemical components inside soybean seeds of different varieties.

Taking BY-1 as an example, seed samples of the same aging year were averaged at each angle to reduce the randomness of sample selection. Meanwhile, seeds with different aging gradients were compared, and the results are shown in Figure 6. It can be clearly seen from the mean figure that, at the four polarization angles, the reflectivity increased with the increase in aging years, and this rule also existed in the other five soybean strains.

The reason may be due to the increase in storage time of the soybean seeds. When the lipids inside naturally aging seeds experience continual biochemical reactions, such as oxidation of these reactions, a change in the chemical composition content of the seeds occurs when the light source into the seed or the PHI hyperspectral reflectance changes. With an increase in storage time of soybean seeds, the content of soluble sugar and other substances in the seeds continues to decline, resulting in a certain difference in spectral reflectance near 420 nm. Meanwhile, the difference in spectral reflectance between 700 and 1000 nm was due to the difference in the surface structure of soybean seeds in different aging years. The damage degree of cell structure also makes the reflection characteristics different, among which, the spectral difference near 950 nm may be caused by the secondary extensional expansion of O-H bond and the tertiary expansion of C-H bond caused by the decomposition reaction of organic components in the seed [21]. Because seeds of different aging degrees had certain differences in spectral curves at specific wavelengths, these characteristic spectra could be extracted and amplified by algorithm processing, thus becoming an effective basis for detecting soybean seed vigor.

3.2. Extraction Results of Characteristic Spectra

In the subsequent modeling processing results, the FD pretreatment method had the optimal modeling effect after processing the spectral data of seeds. Therefore, the feature spectra extraction process in this section used FD pretreatment data as an example to carry out the measurement of CARS. Meanwhile, as the scope of action of the polarizer was considered, the spectra within the range of 420–720 nm was extracted. The spectral data of the yellow soybean line BY-1, the green soybean line BG, and the black soybean line BB-1 at a 90° polarization angle were taken as examples to reduce the dimension of the spectral data. The optimization results for BY-1 CARS are shown in Figure 7.

When the model was established to the minimum RMSECV, the wavelength values under the sampling times were taken as the number of feature wavelengths. The specific effective wavelengths selected by CARS are listed in

Table 3. Feature wavelengths selected by CARS.

Variety	No.	Selected Wavelengths (nm)
BY-1	21	494.6, 504.5, 512.0, 519.4, 524.4, 541.8, 559.3, 564.3, 566.8, 571.8, 589.3, 599.4, 617.0, 619.6, 627.1, 655.0, 665.2, 690.6, 698.3, 711.1, 718.8
BG	27	524.4, 526.9, 531.8, 541.8, 544.3, 546.8, 549.3, 551.8, 554.3, 556.8, 559.3, 561.8, 564.3, 601.9, 604.4, 607.0, 619.6, 629.7, 632.2, 634.7, 639.8, 655.0, 657.5, 660.1, 662.6, 708.5, 718.8
BB-1	37	430.6, 438.0, 450.3, 455.2, 472.4, 479.8, 482.3, 487.2, 492.1, 497.1, 499.6, 521.9, 531.8, 539.3, 541.8, 544.3, 549.3, 554.3, 571.8, 581.8, 586.8, 596.9, 629.7, 632.2, 637.3, 647.4, 652.5, 655.0, 657.5, 660.1, 667.7, 683.0, 690.6, 698.3, 706.0, 713.7, 721.3

.

3.3. Seed Vigor Modeling Results and Discussion

Five machine learning algorithms, including SVR, PLSR, BPNN, GRNN, and RF, were used to model and analyze the hyperspectral and PHI data collected from soybean seeds in order to find the optimal pretreatment methods of FD, SG, SNV, and MSC in the study of soybean seed vigor detection. Taking BY-1 as an example, the modeling results for the four pretreatment methods are shown in

Table 4. Modeling results for polarized hyperspectral imaging at 45°.

Test	Model	Pretreatment	RC	REC	RP	REP
Polarized Hyperspectral Imaging +45°	SVR	FD	0.9335	3.1318	0.9262	3.1956
		SG	0.9292	3.2207	0.8113	4.7923
		SNV	0.9306	3.2245	0.9088	3.4756
		MSC	0.9110	3.6746	0.9173	3.5903
	PLSR	FD	0.8705	4.2784	0.8572	4.3048
		SG	0.7918	5.3101	0.7000	5.8661
		SNV	0.7589	5.6613	0.7833	5.1442
		MSC	0.7997	5.2198	0.8069	4.7885
	BP	FD	0.5998	6.9945	0.5987	6.4640
		SG	0.9930	1.0340	0.5404	8.9853
		SNV	0.9945	0.9187	0.4945	9.5940
		MSC	0.9778	1.8318	0.7784	5.2053
	GRNN	FD	0.9424	2.9091	0.8951	4.3828
		SG	0.8113	5.3132	0.7691	5.3441
		SNV	0.9600	2.2742	0.9018	3.4882
		MSC	0.9317	2.9019	0.9192	3.4415
	RF	FD	0.9338	3.4211	0.9215	4.0108
		SG	0.8924	4.3877	0.6392	6.0900
		SNV	0.9316	3.3441	0.8421	4.4359
		MSC	0.9290	3.3787	0.8028	4.8811

and

Table 5. Results of hyperspectral modeling.

Test	Model	Pretreatment	RC	REC	RP	REP
Hyperspectral	SVR	FD	0.9429	2.9048	0.9161	3.1756
		SG	0.9296	3.2089	0.7604	5.6392
		SNV	0.9375	3.0273	0.9094	3.4853
		MSC	0.9197	3.4192	0.8959	3.7495
	PLSR	FD	0.9063	3.6743	0.8957	3.6198
		SG	0.7979	5.2406	0.7632	5.3728
		SNV	0.7309	5.9332	0.6554	6.1595
		MSC	0.7584	5.6665	0.7124	5.7946
	BPNN	FD	0.8265	4.9211	0.7812	5.0884
		SG	0.9947	0.8967	0.6437	8.5866
		SNV	0.9892	1.2799	0.8175	4.6081
		MSC	0.8531	4.5612	0.5749	7.2405
	GRNN	FD	0.9667	2.2278	0.9076	3.3251
		SG	0.8608	4.6082	0.7119	5.5028
		SNV	0.8830	4.3908	0.8743	3.9880
		MSC	0.8733	4.4110	0.8666	4.2317
	RF	FD	0.9447	2.8982	0.9140	3.3840
		SG	0.9122	4.0256	0.5959	6.5527
		SNV	0.9068	3.8296	0.7695	5.2354
		MSC	0.9093	3.8349	0.7782	5.1752

, for PHI technology with a detection angle of 45° and hyperspectral technology, respectively. For the PHI method of soybean seed vigor prediction, the accuracy with FD, SG, SNV, and MSC preprocessing for the four kinds of models was greater than 90%, with 3, 0, 1, and 1, respectively; for the hyperspectral technology method of soybean seed vigor prediction, the prediction accuracy was greater than 90%, with 3, 0, 1, and 0, respectively. The modeling results show that FD was the best among the four pretreatment methods, whether using the hyperspectral method or the PHI method. The reason may be that the main purpose of FD was to differentiate the original spectrum, and then magnify the sample difference.

The best modeling method, FD + SVR, was used, and the modeling results for all seeds at different polarization angles are shown in

Table 6. Optimal modeling results for all seeds at four polarization angles.

Variety	Testing	RC	REC	RP	REP
BY-1	hyperspectral	0.9429	2.9048	0.9161	3.1756
	0°	0.9732	2.0120	0.9109	3.2971
	45°	0.9335	3.1318	0.9265	0.0663
	90°	0.9741	1.9695	0.9309	3.1479
	135°	0.9971	0.6673	0.8761	4.1594
BY-2	hyperspectral	0.9847	0.0306	0.9160	0.0638
	0°	0.9660	0.0469	0.9418	0.0640
	45°	0.9673	0.0449	0.9505	0.0663
	90°	0.9959	0.0161	0.9199	0.0667
	135°	0.8771	0.0852	0.8943	0.0705
BY-3	hyperspectral	0.9861	0.0258	0.9186	0.0620
	0°	0.9962	0.0135	0.9532	0.0506
	45°	0.9714	0.0368	0.8741	0.0828
	90°	0.9862	0.0273	0.9384	0.0604
	135°	0.9936	0.0177	0.9665	0.0481
BG	hyperspectral	0.9999	0.0025	0.8913	0.1018
	0°	0.9670	0.0509	0.9293	0.0842
	45°	0.9412	0.0676	0.9499	0.0705
	90°	0.9301	0.0737	0.9564	0.0693
	135°	0.9638	0.0546	0.9428	0.0759
BB-1	hyperspectral	0.9168	0.1149	0.7664	0.1831
	0°	0.9608	0.0817	0.8320	0.1615
	45°	0.9541	0.0861	0.6779	0.2204
	90°	0.9797	0.0579	0.8026	0.1692
	135°	0.9693	0.0722	0.9131	0.1310
BB-2	hyperspectral	0.8287	0.0711	0.8131	0.0871
	0°	0.8940	0.0575	0.8843	0.0674
	45°	0.9933	0.0147	0.8535	0.0745
	90°	0.9462	0.0434	0.8414	0.0897
	135°	0.9195	0.0518	0.8446	0.0799

. The final results showed that the accuracy of soybean viability detection results for PHI was higher than that for the ordinary hyperspectral method at almost all four polarization angles. When using hyperspectral technology for detection, the prediction accuracies for BY-1, BY-2, BY-3, BG, BB-1, and BB-2 were 91.61%, 91.60%, 91.86%, 89.13%, 76.64%, and 81.31%, respectively. The prediction accuracy could be improved to 93.09%, 95.05%, 96.65%, 95.64%, 91.31%, and 88.43%, respectively, by using PHI. It can be seen that the use of PHI technology to predict soybean seed vigor was effective and feasible, and that the detection accuracy was improved compared with the ordinary hyperspectral detection technology.

In the results, the reason for the accuracy being improved at 0° was that the polarizer was added in front of the lens to eliminate the reflection effect and increase the light intensity reflected from the seed surface.

Considering that the optimal polarized angles vary with different colors and varieties of soybean seeds, the polarized components I, Q, and U were used for soybean seed vigor prediction model construction to build a unified and universal prediction model of soybean seed vigor for different colors and varieties.

The I, Q, and U components and polarization state Dol parameters were calculated by using the spectrum of the four polarization angles under the optimal FD preprocessing mode. After the preprocessing, the CARS algorithm was also used to screen the characteristic bands of PHI data. Finally, these extracted polarization components were used as input in the modeling of seed vigor detection for each strain category, and the modeling results are shown in

Table 7. Optimal modeling results for polarization components of all seeds.

Variety	Testing	RC	REC	RP	REP
BY-1	hyperspectral	0.9429	2.9048	0.9161	3.1756
	I	0.9451	2.8859	0.9419	3.4286
	Q	0.9723	2.0685	0.7533	5.7877
	U	0.9840	1.5952	0.6244	0.6244
	Dol	0.9694	2.2345	0.2936	7.6085
BY-2	hyperspectral	0.9847	0.0306	0.9160	0.0638
	I	0.9994	0.0063	0.9248	0.0690
	Q	0.9109	0.0732	0.8397	0.0884
	U	0.8773	0.0858	0.9267	0.0579
	Dol	0.5465	0.1477	0.4164	0.1511
BY-3	hyperspectral	0.9861	0.0258	0.9186	0.0620
	I	0.9663	0.0439	0.9292	0.0734
	Q	0.9796	0.0317	0.9000	0.0682
	U	0.9978	0.0103	0.5682	0.1546
	Dol	0.9982	0.0095	0.6203	0.1314
BG	hyperspectral	0.9999	0.0025	0.8913	0.1018
	I	0.9749	0.0444	0.9342	0.0959
	Q	0.5798	0.1630	0.5791	0.1844
	U	0.8825	0.0945	0.7278	0.1556
	Dol	0.9786	0.0429	0.5636	0.1877
BB-1	hyperspectral	0.9168	0.1149	0.7664	0.1831
	I	0.9736	0.0657	0.9090	0.1270
	Q	0.8391	0.1584	0.4424	0.2617
	U	0.9482	0.0914	0.8046	0.1699
	Dol	0.9195	0.1181	0.3279	0.2816
BB-2	hyperspectral	0.8287	0.0711	0.8131	0.0871
	I	0.9026	0.0547	0.8942	0.0644
	Q	0.8064	0.0765	0.4668	0.1345
	U	0.9930	0.0154	0.7222	0.0985
	Dol	0.9766	0.0276	0.7804	0.0945

.

From the experimental results, it can be seen that in the four I, Q, U, and Dol polarization parameters, the use of I, Q, and U polarization component modeling could achieve a certain modeling effect, but it was not higher than that of hyperspectral imaging where modeling accuracy was concerned; relative to the hyperspectral data modeling, the Dol component had the most unstable modeling effect, while the I component had the best modeling effect and was relatively stable. The prediction accuracy could be improved in the prediction of all varieties. The prediction accuracy of BY-1, BY-2, BY-3, BG, BB-1, and BB-2 could be improved to 94.19%, 92.48%, 92.92%, 93.42%, 90.90%, and 89.42%, respectively. The experimental results verified the feasibility of using the I component in the study of soybean seed vigor using PHI.

To unify the model, the soybean seeds of all color lines were mixed, and to improve the accuracy of mixing model, the polarization component was used. Because the modeling effect of Dol parameters was not ideal, three polarization component parameters were used to model the soybean seed mixing. To further improve the accuracy of blended modeling when polarization components were used, the blending theory was used to integrate learning. The ensemble learning was composed of weak classifier and strong classifier, which can improve the stability of the model. It is suitable for a small amount of data. RF and SVR were used as modeling methods for the first layer, and SVR with the best and most stable modeling effect was used for the second layer. The final blended modeling results are shown in

Table 8. Hybrid modeling results.

Testing	RC	REC	RP	REP
0°	0.9938	3.5472	0.9030	15.6814
45°	0.9824	5.8694	0.9624	9.9169
90°	0.9816	6.0160	0.9808	7.1554
135°	0.9876	4.9346	0.9422	12.2301
mean	0.9864	5.0918	0.9471	11.2460
I	0.9767	6.7598	0.9717	8.7390
Q	0.9851	5.4121	0.9825	6.8643
U	0.9948	3.2454	0.9755	8.0817
I + Blending	0.9994	1.1119	0.9760	8.1498
I + Blending	0.9987	1.5992	0.9848	6.5136
Q + Blending	0.9987	1.6198	0.9886	5.4898

. The modeling results showed that the accuracy of PHI hybrid modeling was 94.71% after taking mean values at four angles, and the accuracy improved to 97.17%, 98.25%, and 97.55% when using I, Q, and U polarization components, respectively. After the blending algorithm, the modeling accuracy of I, Q, and U components improved to 97.60%, 98.48%, and 98.86%, respectively. Therefore, the ensemble learning was more effective.

He [41] used near-infrared hyperspectral imaging technology to detect the vigor of rice seeds of different ages, and the classification accuracy of the LS-SVM model based on the characteristic wavelength could reach 94.38%. Zhang [42] used visible and near-infrared hyperspectral imaging techniques to identify wheat seed vigor, and the classification accuracy of the SNV-SPA-PLS-DA model was about 89.5%. After using the blending integrated learning algorithm, the modeling precision of the U components was improved to 0.9886. Overall, polarized hyperspectral imaging combined with an ensemble learning algorithm can improve the detection accuracy.

4. Conclusions

Seed aging has been a popular issue in seed research, as it affects the seedling emergence rate in the field and has a serious impact on agricultural production and food safety. However, most of researchers have only carried out research on seed vigor detection under artificial aging. Therefore, in this paper, we proposed a vigor detection method for naturally aged soybean seeds based on polarized hyperspectral imaging combined with an ensemble learning algorithm.

Research showed that the prediction accuracy of seed vigor by the PHI technique was higher than that by the hyperspectral technique. The average prediction accuracy for BY-1, BY-2, BY-3, BG, BB-1, and BB-2 was 87.03% using the hyperspectral technique, which was improved up to 93.36% by using the PHI technique. The modeling effect for different colors of soybean seeds using the polarization component was better than using each polarization angle. The blended ensemble learning algorithm may further improve model accuracy compared with traditional machine learning. The prediction accuracy of soybean seed vigor was improved to 97.17%, 98.25%, and 97.55% by using the polarization component of I, Q, and U.

In the future, more soybean varieties and a larger number of samples are needed to improve the accuracy of detection. In conclusion, this research method can be used to identify soybean seed vigor and it provides a potential tool for seed companies and market regulators to quickly detect soybean seeds’ vigor in the actual production process.