Estimation Model for Maize Multi-components Based on Hyperspectral Data

1. Introduction

As the third largest grain crop in the world, maize not only plays a significant role in human diet, but also serves as an important raw material for industrial production. The quality of its seeds is directly related to the yield of corn and the quality of its derivative products. The accurate and efficient detection of component content in corn seeds is of great significance in elevating the overall standards of the corn planting industry and promoting the growth of related industries. Among the various components of corn seeds, the content of water, fat, protein, and starch are key indicators for evaluating seed quality. Water content is an important factor affecting seed storage stability, transportation safety, and germination rate after sowing. High levels of moisture content can accelerate seed respiration, consume nutrients, and even lead to mold and decay; however, too low a water content may cause the seeds to lose vitality and affect germination. Fat, protein, and starch are the main components that constitute the nutritional and economic value of corn seeds, and their content directly determines the nutritional quality and market value of corn. Conventional methods for detecting the composition of corn seeds, such as drying and titration, can meet detection needs to some extent, but they have the disadvantages of being time-consuming, cumbersome to operate, and causing damage to samples, making it difficult to fall short of meeting the needs of rapid, non-destructive, and batch detection in modern agricultural production. Hence, investigating a novel and effective detection technology holds immense importance in enhancing the efficiency and precision of determining the compositional content of corn seeds.

Hyperspectral imaging technology (HSI), an advanced detection method that amalgamates traditional near-infrared spectroscopy with machine vision technology, has gained significant traction in the realm of agricultural product quality assessment in recent years. This technology facilitates swift, non-invasive, and precise analysis of sample composition by capturing spectral data from every pixel on the sample surface and integrating it with spatial information. Over the past few years, it has been subject to extensive research and application in the quality evaluation of agricultural products and food items. [1,2,3,4,5,6]. Liu et al. established a SSA-SVM model based on hyperspectral data to predict the protein content in milk. The prediction results met the accuracy requirements for dairy testing, providing a feasible new method for the rapid detection of milk protein content [7]. Xuan et al. established a multivariate linear regression model for protein content in rice grains based on HSI, with an $R_c^2$ of 0.9393 in the validation set, providing the possibility for non-destructive protein content detection [8]. Van Haute established a quantitative analysis model between carotene and spectrum based on LSSVM and PLSR, accurately predicting the content of carotene in carrots [9]. Khajehyar et al. used hyperspectral signatures to predict foliar nitrogen and calcium content of tissue cultured little-leaf mockorange (Philadelphus microphyllus A. Gray) shoots, and the $R^2$ values of the established RF regression model reached 0.72 and 0.99 [10]. Zhang et al. combined hyperspectral technology with deep learning algorithms to establish a CNN-LSTM corn seed moisture content detection model, achieving a prediction set R² of 0.937 [11]. In summary, hyperspectral imaging has shown great potential in predicting the composition content of crops. However, previous studies have primarily focused on detecting individual components, with limited research on simultaneous detection of multiple components, and a lack of in-depth exploration into model optimization. Based on this, this research aims to utilize HSI to quickly and non-destructively detect the moisture, fat, protein, and starch content of corn seeds. By collecting hyperspectral images of corn seeds and combining advanced data processing and analysis methods, a detection model for corn seed composition content based on HSI is established to provide a scientific foundation and technical support for quality control and production management in the corn seed industry.

2. Materials and Methods

2.1. Samples

The samples used in the experiment comprised 80 varieties of corn seeds cultivated in 2023, all from Jilin Province, China. The hulls of these seeds are of similar color, uncoated, and they have plump kernels. Figure 1 shows the photos of some varieties of seeds.

To achieve a stable condition of the seeds and ensure consistency in experimental conditions, prior to the experiment, all seeds were left to stand in the same experimental environment for more than 72 hours, followed by the collection of hyperspectral images. Subsequently, each variety was divided into four aliquots, and their moisture, fat, protein, and starch contents were determined using chemical quantitative methods. Finally, a dataset correlating spectral data with the contents of these components was established.

2.2. Experimental Equipment

The hyperspectral image acquisition equipment used in the experiment mainly including: a 1000–2500 nm imaging spectrometer (ImSpertor N25E); a 1600×1200 pixel CCD camera (Bobcat 1410); a bilateral 150W halogen light source (IT3900); and a one-dimensional precision mobile platform (IRCP-0076-400). During image acquisition, the entire system is placed in a closed dark box to prevent interference from ambient light. The structure of the acquisition system is shown in Figure 2. The spectral acquisition range of the system is 935.5-2539nm, with a resolution of 6.3 nm. In the experiment, through repeated debugging, the lens height was set to 44 cm, the exposure time was set to 10 ms, and the speed of the platform movement was set to 8 mm/s, allowing for the collection of clear images of corn kernels.

Due to the optical properties of the hyperspectral imager’s lens, it generates a large amount of dark current, and the distribution of light intensity also exhibits non-uniformity in certain specific wavelength bands. These factors can result in images obtained from the hyperspectral imager being riddled with a lot of noise. Therefore, to extract useful information from hyperspectral images, it is necessary to perform correction processing on the original image to improve its quality. The image correction formula can be expressed as:

$$R={\displaystyle \frac{I_o-I_{dark}}{I_{white}-I_{dark}}}$$

(1)

where, R represents the corrected image, I_o denotes the original sample image, I_white signifies the image captured by scanning the standard whiteboard, and I_dark represents a dark field image obtained by covering the camera lens during the experiment

2.3. Method and Principle

2.3.1. Preprocessing Methods

The spectral signal of complex samples is frequently impacted by various factors, including noise, stray light, and baseline shift, with the particle size of the sample being a pivotal factor in spectral measurement [12,13]. Specifically, with the increase in sample particle size, the reproducibility of the spectrum decreases, while the variability of the spectrum increases exponentially. Due to the variations in shape and diameter of individual corn kernels, the resulting spectral data show significant differences, which further complicates the elimination of measurement errors and ultimately impacts the accuracy of quantitative analysis [14]. Therefore, to effectively mitigate spectral noise interference due to factors like seed morphology, it is crucial to employ suitable spectral preprocessing techniques to enhance the quality of spectral data.

To enhance the quality of spectral data and minimize noise and interference, seven preprocessing methods are applied to the spectral data: moving average (MA), Savitzky-Golay derivative convolution (SG), normalization (NOR), baseline correction (BC), multiplicative scatter correction (MSC), standard normal variate transformation (SNV), and detrending transformation (DT). Among them, the window size for MA is 7, and for SG is 7, the polynomial order for SG is 2, and for DT is 2.

2.3.2. Feature Wavelength Extraction Algorithm 

1) Successive Projections Algorithms (SPA) Method

SPA is a forward-selection technique that has been traditionally utilized for selecting feature wavelength variables, with the purpose of identifying the combination that reduces redundant information to the greatest extent [15]. The algorithm works by projecting each wavelength onto other bands and comparing the magnitudes of the projection vectors. It selects characteristic wavelengths with minimal redundant information and collinearity, ultimately reducing computation time and enhancing model classification accuracy. SPA selects only a small subset of wavelength variables from the original spectral data, yet these variables capture most of the useful information from the original data. This approach avoids information redundancy, significantly reduces model complexity, and improves the accuracy, speed, and robustness of predictions [16].

2) Uniformative Variable Elimination (UVE) Method

UVE is an algorithm for wavelength selection that utilizes PLSR coefficients. Its primary objective is to refine the modeling dataset by eliminating wavelength variables whose stability is poorer than noise, ultimately improving the predictive power of the model [17]. The method involves combining artificial random noise information with PLS to build a regression cross-validation model. By computing the ratio of the mean to the standard deviation of the regression coefficients, an evaluation metric is established to assess the importance of the characteristic wavelength variables. Furthermore, the maximum value of the noise matrix serves as the upper and lower limits for the algorithm’s threshold, with feature variables exceeding or falling below these thresholds being identified as the final optimized feature vectors.

3) Competitive Adaptive Reweighted Sampling (CARS) Method

The CARS feature extraction algorithm combines Monte Carlo sampling with PLS model regression coefficients [18]. By mimicking Darwin’s “survival of the fittest” principle, it uses adaptive re-weighted sampling (ARS) to progressively discard points with regression coefficients of lower absolute weights, thereby forming a new subset and re-establishing the PLS model. After several iterations, it selects the wavelength subset with the minimum root mean square error of cross-validation (RMSECV) from the PLS model as the characteristic wavelengths. This approach, widely used in spectral data analysis and image processing, effectively reduces data dimensionality and enhances model prediction accuracy.

2.3.3. Modeling Algorithm

1) Partial Least Squares Regression (PLSR)

PLSR, a multivariate data analysis technique in statistics and machine learning, is particularly adept at addressing scenarios involving multicollinearity between dependent and independent variables [19,20]. It merges the strengths of principal component analysis (PCA) and multiple linear regression (MLR), with its central concept revolving around identifying novel orthogonal projection directions, or principal components, that optimize the covariance between the projected dependent and independent variables. This approach facilitates the establishment of a predictive model. Specifically, PLSR introduces latent variables by concurrently analyzing the relationship between the independent variable X and the dependent variable Y, and employs this to construct a linear model between them. Thanks to its unique principles and processes, PLSR is especially proficient in tackling high-dimensional data and small sample regression challenges [21,22].

2) Back Propagation Neural Network (BPNN)

The BPNN, a widely utilized multilayer feedforward neural network in supervised learning tasks, is trained through the error backpropagation algorithm. This model possesses robust nonlinear fitting capabilities, enabling it to automatically learn intricate relationships between input features [23]. A notable advantage of the BPNN model is its capacity to establish a nonlinear model using a limited set of refined wavelength variables. Structurally, it comprises an input layer, a hidden layer, and an output layer [24]. During the network’s learning process, the signal progresses through both forward and backward propagation. Specifically, the input sample is subjected to forward propagation, where it is sequentially processed by the hidden layer, starting from the input layer, and ultimately transmitted to the output layer. In cases where the actual output of the output layer diverges from the desired output, the error undergoes backpropagation. Subsequently, the weights and thresholds of each layer are adjusted along the gradient direction, progressively diminishing the error and ultimately achieving the target accuracy [25].

The number of hidden layer nodes can be determined by an empirical formula, which is calculated as follows:

$$N=\sqrt{L+M}+\alpha$$

(2)

Where, L and M denote the node counts for the input and output layers, respectively, and α is an integer within the range of 0 to 10.

3) Radial Basis Function Neural Network (RBFNN)

RBFNN is an artificial neural network with a specific architecture, named for the radial basis function (RBF) used as the activation function in the hidden layer [26]. It has found widespread applications in fields such as function approximation, pattern recognition, and control systems. The core idea is to map low-dimensional linearly inseparable data into high-dimensional space, rendering them linearly separable. By carefully selecting the radial basis function and its parameters (such as the center point and width), RBFNN can effectively approximate complex nonlinear functions. As a neural network model, RBFNN possesses strong nonlinear fitting and generalization capabilities. Its unique structure and training algorithm give it significant advantages in dealing with complex nonlinear problems, which is why it has been widely applied in many fields [27,28].

4) Least Squares Support Vector Machine (LSSVM)

LSSVM is an extension of SVM proposed by Suykens et al. It employs the least squares method to find a hyperplane, mapping sample points to a high-dimensional space based on their feature vectors, thus enabling classification or regression predictions for the samples [29]. Compared to traditional SVM, LSSVM replaces the inequality constraints of the slack variable with equality constraints, transforming the problem into solving a system of linear equations, thereby simplifying the calculation process. LSSVM features fast convergence speed, minimal parameter adjustment, strong exploration ability, and high prediction accuracy, making it widely applied in finance, medicine, industry, meteorology, and other fields [30,31].

2.3.4. Model Evaluation Method

The performance of the models was evaluated mainly by the coefficient of determination ($R^2$) and root mean square error (RMSE) [32,33]. The calculation formulas are as follows:

$$R^2={\displaystyle \frac{{{\displaystyle \sum }}_{i=1}^N\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^N{\left(x_i-\overline{x}\right)}^2+{{\displaystyle \sum }}_{i=1}^N{\left(y_i-\overline{y}\right)}^2}}}$$

(3)

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{{\displaystyle \sum }}_{i=1}^N{\left(x_i-y_i\right)}^2}$$

(4)

Where, $x_i$ is the actual measured value, $y_i$ is the predicted value, $\overline{x}$ is the average measured value, and $\overline{y}$ represents the average predicted value, $N$ is the number of samples. $R^2$ is a value range of [0,1]. The closer $R^2$ is to 1, and the smaller the RMSE is, the better the prediction performance of the regression model.

During the modeling process, if the values of $R_C^2$ and $R_P^2$ (determination coefficients of calibration set and validation set) are large with minimal difference, and the values of $RMSEC$ and $RMSEP$ (RMSE of calibration set and validation set) are small and the difference is minimal, the model exhibits high reliability and credibility.

3. Results and Discussion

3.1. Sample Division

The samples were partitioned into a calibration set and a prediction set at a 3:1 ratio using the Sample Set Partition Based on Joint X-Y Distance (SPXY) algorithm. Table 1 displays the content statistics of each component in the dataset. The content range of each component within the calibration set encompasses the range present in the prediction set, suggesting that the division of the sample set is appropriate.

3.2. Spectral Extraction and Analysis

The surface structure of a corn kernel comprises the embryo and endosperm, as illustrated in Figure 3. In order to obtain the spectrum of the complete embryo structure region, the seed image is processed using threshold segmentation. Figure 4 shows the spectra of the embryo region, endosperm region, and background region of the corn kernel.

Figure 4 demonstrates that the spectral curve of the embryo exhibits a similarity to the trend observed in the endosperm, yet it shows notable differences when compared to the background spectrum. At 1272.7 nm, 1455.4 nm, and 1871.2 nm, the reflectivity of the embryo and endosperm varies greatly, and the difference in reflectivity between 1272.7 nm and 1871.2 nm is smaller in the embryo structure area than in the endosperm structure area. It is difficult to distinguish the embryo and endosperm regions of some uneven brightness samples by only using the reflection intensity of one band. The band ratio ${\displaystyle \frac{{\mathsf{\lambda }}_{1272.7}+{\mathsf{\lambda }}_{1455.4}+{\mathsf{\lambda }}_{1871.2}}{{\mathsf{\lambda }}_{1272.7}-{\mathsf{\lambda }}_{1871.2}}}$ increases the band ratio of the embryo region and decreases the band ratio of the endosperm region. By threshold segmentation, the embryo structure can be effectively segmented from the image. Figure 5 shows the extraction process of the maize kernel embryo.

The wavelength range of hyperspectral data spans from 935.5 nm to 2539 nm, encompassing a total of 256 bands. Due to significant noise interference at the beginning and end of the spectral data during the acquisition process, these segments of data should be excluded from further analysis. Consequently, we utilized 218 bands, ranging from band 22 to band 239 (i.e., the spectrum spanning from 1065 nm to 2432 nm), for our analysis. The average spectral curve of the embryo structure is shown in Figure 6.

Based on the spectral absorption curve, it is observed that the absorption peak located near 1450 nm corresponds to the first-order overtone (or fundamental) vibration of the NH bond present in proteins. Proteins are important components of corn seeds, which are composed of amino acids, and amino acid molecules contain NH bonds. Therefore, when these NH bonds vibrate at the fundamental frequency, an absorption peak near 1450 nm will be generated in the spectrum. The absorption peak near 1950 nm may be related to the carbohydrates in corn seeds, especially the second-order overtone vibration of C-O bonds. Carbohydrates are one of the main components of corn seeds, and they contain a large number of C-O bonds. Therefore, when these C-O bonds vibrate in the second-order overtone mode, corresponding absorption peaks will appear in the spectrum. The absorption peak near 2200 nm involves the combined vibration of OH and NH bonds [34]. In corn seeds, these two bonds are found in molecules such as water and proteins. When they vibrate in specific combinations, they may produce an absorption peak near 2200 nm on the spectrum. These absorption peaks provide us with useful information about the chemical composition and molecular structure of corn seeds. Therefore, it is feasible to quantitatively determine the main component content through the spectral curve of the seeds.

3.3. Spectral Preprocessing

Despite the presence of high multicollinearity among the independent variables, the PLSR model remains applicable for regression modeling. Hence, this study opts for PLSR to assess the effectiveness of different preprocessing methods. The regression model was evaluated using the fit index $R^2$ and the error index RMSE. With the original and preprocessed spectral data as inputs, four component content detection models were established, with the results shown in Table 2. In the table, $R_C^2$ and $R_P^2$ represent the determination coefficients for the training and test sets, respectively. The closer the value of $R^2$ is to 1, the better the model’s fit. Generally, a model with an $R^2$ above 0.4 is considered well-fitting. RMSEC and RMSEP respectively denote the root mean square errors of the training and testing sets, and the smaller thet are, the superior the model’s fit.

According to the Table 2, on the whole, the prediction accuracy of the moisture content detection model is higher than that of the other three components. The prediction effect of fat content is relatively poor, with the minimum $R_P^2$ only having 0.6588. This is primarily because of the low fat content in corn kernels, which involves certain errors in both physical and chemical measurements as well as hyperspectral measurements. The $R_P^2$ of the four component prediction sets for the original spectra without preprocessing are 0.9748, 0.7483, 0.8995, and 0.8249, respectively, which can better predict the content of each component. For different preprocessing methods, the detection results of different components vary inconsistently. After preprocessing with the MA and SG algorithms, the prediction performance of the model is inferior to that established using the original spectra; NOR improved the prediction effect of protein and starch content, but decreased the prediction effect of moisture and fat content; The BC, MSC, and SNV algorithms have all improved the prediction accuracy for fat, protein, and starch content, but have deteriorated the prediction accuracy for water content; Only when the DT algorithm is applied to preprocess the original spectra, the prediction performance of all components improves, with $R_P^2$ values being 0.9906, 0.8706, 0.9535, and 0.9145, respectively, and RMSEP values being 0.0393, 0.0606, 0.1057, and 0.2170, respectively. Therefore, in subsequent analysis, DT is selected for spectral preprocessing. The spectral curve after DT preprocessing is shown in Figure 7.

3.4. Feature Wavelength Extraction

Three feature wavelength selection algorithms, SPA, UVE, and CARS, were used to extract the characteristic wavelengths from the preprocessed spectra for various components. The characteristic wavelength positions corresponding to different components are displayed on the average spectral curve, and the extraction results are shown in Figure 8, Figure 9 and Figure 10.

In order to make quantitative predictions for four components simultaneously, the four sets of characteristic wavelengths extracted by each algorithm were combined into a single set. The SPA algorithm extracted a total of 51 characteristic wavelengths, the UVE algorithm extracted a total of 105 characteristic wavelengths, and the CARS algorithm extracted a total of 89 characteristic wavelengths. Table 3 presents the coefficients of determination for the prediction sets of the four components.

3.5. Establishment of a Model for Detecting the Content of Multiple Components in Maize Seeds

The original spectra were preprocessed using the DT algorithm, followed by the extraction of characteristic wavelengths from the preprocessed spectral data. Subsequently, multi-output content detection models were established for PLSR, RBFNN, BPNN, and LSSVM, each based on both full-band spectra and characteristic wavelengths. The prediction performance of these models was evaluated using the coefficient of determination for the prediction set of each component. The results of the model predictions are presented in Table 4, while the prediction effects are illustrated in Figure 11, Figure 12, Figure 13 and Figure 14.

According to the model prediction results, it can be seen that the detection effect of water content in several components is the best, followed by starch content, and the prediction effect of fat content is relatively the worst. Using full-band spectra for modeling, the model prediction performance is better because the spectral data contains information about all seed components. Among the models established based on the full-band spectra, the PLSR model has the best prediction performance, with $R^2$ values of 0.9874, 0.9472, 0.9833, and 0.9532 for moisture, fat, protein, and starch content, respectively. Although the detection of fat content is not as good as that of the full-band RBFNN model, the detection of the other three components is higher than the RBFNN model.

Using feature selection modeling can significantly improve prediction accuracy and modeling efficiency, enhance model stability, and facilitate online real-time monitoring. However, it also has some disadvantages such as a complex algorithm selection process, potential loss of useful information, sensitivity to initial conditions, and potentially high computational demands. Based on the prediction models of PLSR and RBFNN, from the perspective of prediction accuracy, the modeling performance of feature selection is not as good as that of full-spectrum modeling, indicating that during the feature selection process, some wavelengths with weak correlation to the target substance but containing useful information may be removed, resulting in a greater loss of spectral information related to the substance being tested. On the contrary, the BP neural network and LSSVM models have better prediction performance than the full-spectrum model when using feature selection for modeling. The best performance is achieved by the CARS-LSSVM model, with the $R^2$ values of the four components all improving compared to full-spectrum modeling, indicating that these two models can achieve more precise results with less sensitive information. It is not difficult to find that the prediction performance of the LSSVM model is relatively stable across several different input wavelengths, and the detection performance of the four components is relatively good. Based on the comprehensive comparison results, the optimal model for detecting the composition content of corn seeds was determined to be CARS-LSSVM.

4. Conclusions

This study employs HSI to achieve rapid and non-destructive precise detection of multiple key components in corn seeds, including water, fat, protein, and starch. Initially, the embryo structure of the corn seeds was precisely extracted as the region of interest using hyperspectral image segmentation technology. Subsequently, seven spectral data preprocessing algorithms were applied, and DT was identified as the optimal preprocessing method through a comprehensive comparison with the PLSR model. To further our research, we employed three characteristic wavelength extraction strategies: SPA, CARS, and UVE, with the aim of precisely capturing the spectral characteristics of each component. Finally, multi-component content detection models based on raw spectra and characteristic wavelengths were constructed using PLSR, BPNN, RBFNN, and LSSVM, respectively, to achieve comprehensive and in-depth quality assessment of corn seeds. The experimental results demonstrated that the DT-CARS-LSSVM model exhibited outstanding accuracy and stability in predicting the four key components, with coefficients of determination reaching 0.9877, 0.9344, 0.9827, and 0.9592, respectively, highlighting the significant advantages and application potential of this method in the field of corn seed main component content and quality detection. This study not only provides a scientific basis for the quality assessment of corn seeds but also opens up a new path for the development of non-destructive testing technology in related fields.