Combined Hyperspectral Imaging with Wavelet Domain Multivariate Feature Fusion Network for Bioactive Compound Prediction of Astragalus membranaceus var. mongholicus

Abstract

The pharmacological quality of Astragalus membranaceus var. mongholicus (AMM) is determined by its bioactive compounds, and developing a rapid prediction method is essential for quality assessment. This study proposes a predictive model for AMM bioactive compounds using hyperspectral imaging (HSI) and wavelet domain multivariate features. The model employs techniques such as the first-order derivative (FD) algorithm and the continuum removal (CR) algorithm for initial feature extraction. Unlike existing models that primarily focus on a single-feature extraction algorithm, the proposed tree-structured feature extraction module based on discrete wavelet transform and one-dimensional convolutional neural network (1D-CNN) integrates FD and CR, enabling robust multivariate feature extraction. Subsequently, the multivariate feature cross-fusion module is introduced to implement multivariate feature interaction, facilitating mutual enhancement between high- and low-frequency features through hierarchical recombination. Additionally, a multi-objective prediction mechanism is proposed to simultaneously predict the contents of flavonoids, saponins, and polysaccharides in AMM, effectively leveraging the enhanced, recombined spectral features. During testing, the model achieved excellent predictive performance with R² values of 0.981 for flavonoids, 0.992 for saponins, and 0.992 for polysaccharides. The corresponding RMSE values were 0.37, 0.04, and 0.86; RPD values reached 7.30, 10.97, and 11.16; while MAE values were 0.14, 0.02, and 0.38, respectively. These results demonstrate that integrating multivariate features extracted through diverse methods with 1D-CNN enables efficient prediction of AMM bioactive compounds using HSI.

1. Introduction

Astragalus membranaceus var. mongholicus (AMM), a leguminous herb widely distributed across Inner Mongolia, Ningxia, Shaanxi, and Shanxi, serves as a cornerstone of traditional Chinese medicine due to its high concentrations of pharmacologically bioactive compounds, including polysaccharides, saponins, and flavonoids [1]. These compounds exhibit immunomodulatory, antioxidative, and bone metabolism-regulating properties, necessitating precise quantification for effective quality control and optimized cultivation practices [2,3].

Conventional analytical techniques, such as ultraviolet spectrophotometry, high-performance liquid chromatography, and gas chromatography–mass spectrometry [4,5,6], provide accurate measurements but involve destructive sampling, extensive laboratory preparation, and require specialized expertise [7]. These limitations hinder large-scale screening and real-time monitoring, highlighting the need for rapid and non-destructive alternatives [8].

Hyperspectral technology has emerged as one of the most prominent cutting-edge tools in precision agriculture, enabling comprehensive analysis of optical reflectance spectra. It allows for precise identification and characterization of plants and their health indicators. Neri et al. developed and implemented a real-time AI-assisted push-broom hyperspectral system for plant recognition, which provides unprecedented detail and accuracy in crop monitoring and achieves high-precision classification of lettuce and arugula leaves [9]. Hyperspectral imaging (HSI), which integrates spatial and spectral data across hundreds of contiguous bands, has emerged as a transformative tool for non-invasive chemical analysis. Its applications span agricultural quality assessment, environmental surveillance, and pharmaceutical research, and it is increasingly adopted in traditional Chinese medicine for the quantification of bioactive compounds [10].

Recent advances in HSI applications for traditional Chinese medicine analysis have demonstrated promising results in compound prediction for various medicinal plants [11,12,13,14]. Yet, these methods face a fundamental challenge in processing high-dimensional spectral data characterized by inherent feature redundancy. This dual characteristic of hyperspectral datasets not only increases computational complexity but also impairs analytical performance because of strong inter-band correlations and overlapping information among adjacent spectral bands [15,16].

To overcome these limitations, recent studies have demonstrated various spectral preprocessing approaches for medicinal plant analysis. Wang et al. employed second derivative processing of attenuated total reflectance–Fourier transform infrared spectroscopy to achieve rapid species identification and polysaccharide prediction in Polygonatum [17]. Hao et al. systematically compared multiple preprocessing methods for analyzing different target components (e.g., total flavonoid content and antioxidant activity) in plum fruits [18]. In another study, Wang et al. applied standard normal variate transformation to preprocess near-infrared (NIR) spectra of Hypericum perforatum [19]. Ma et al. utilized orthogonal signal correction for NIR spectral processing, enabling rapid prediction of rosmarinic acid, tanshinones, and salvianolic acids in Salvia miltiorrhiza [20]. Furthermore, Liu et al. identified Savitzky–Golay smoothing and detrending as optimal preprocessing methods for the analysis of gentiopicroside and loganic acid, respectively, using Fourier transform infrared photoacoustic spectroscopy coupled with chemometric analysis for the quality evaluation of Gentiana macrophylla [21]. However, these methods typically rely on single-feature extraction, failing to fully utilize the multidimensional feature information of hyperspectral data.

Integrating multivariate feature extraction with deep learning presents an opportunity to improve prediction accuracy. Discrete wavelet transform (DWT), for instance, can decompose spectra into high- and low-frequency components, while cross-fusion networks can enhance hierarchical feature representation. Such advancements could enable non-destructive root compound prediction via leaf spectral analysis, supporting precision agriculture and intelligent phenotyping.

For indirect root bioactive compound prediction in AMM based on leaf hyperspectral data, this study proposes a custom-designed Wavelet and Cross-Fused Multivariate Feature Network (WCT-MCFNet), based on a one-dimensional convolutional neural network (1D-CNN). Specifically, the model combines first-order derivative (FD) and continuum removal (CR) algorithms with DWT to extract multivariate spectral features, processes these features through parallel cross-fusion branches for high- and low-frequency components, followed by feature fusion and a multi-objective prediction mechanism that simultaneously estimates flavonoid, saponin, and polysaccharide concentrations.

This study aims to develop an indirect and non-destructive method based on leaf HSI to achieve simultaneous and accurate prediction of multiple bioactive compound contents in the roots of AMM. To this end, this study proposes to integrate traditional feature extraction methods with deep learning techniques to construct an end-to-end predictive model. By combining multiple feature extraction algorithms and enhancing spectral response features using DWT, a parallel cross-fusion strategy is further designed to effectively integrate multi-scale spectral information for the synchronous prediction of the three target components. This approach offers a novel and reliable computational framework for high-throughput and high-precision quality prediction of traditional Chinese medicinal materials. The key contributions are as follows: (1) a non-destructive prediction model based on leaf HSI, thereby eliminating the need for destructive sampling of AMM roots; (2) a DWT-based and 1D-CNN-based tree-structured (WCT) feature extraction module that integrates FD and CR to enhance the robustness of feature extraction; (3) multi-feature-based cross-fusion (MCF), in which the proposed MCF feature fusion module leverages complementary spectral features to improve prediction accuracy.

2. Materials and Methods

2.1. Data Acquisition

2.1.1. Bioactive Compounds Measurement

AMM plants for this study were collected in late October 2023 from the AMM planting demonstration base in Guyang County, Inner Mongolia Autonomous Region (41°16′55.17″ N, 109°58′46.56″ E). Figure 1 displays the geographical locations where AMM samples were collected. Taking into account the spatial distribution characteristics of the planting area, a grid-based sampling method was employed. The study area is characterized by a typical mid-temperate semi-arid continental climate, with significant diurnal temperature variations. The mean annual temperature ranges from 2 to 5 °C, and the annual precipitation averages approximately 290 mm. The dominant soil types are Kastanozems and Calcisols, with an organic matter content of 19.30 g/kg. The total nitrogen, available phosphorus, and available potassium contents are 1.12 g/kg, 10.30 mg/kg, and 156.00 mg/kg, respectively. These climatic and soil conditions collectively facilitate the synthesis and accumulation of bioactive compounds in the roots of AMM, providing a sufficient nutritional foundation and creating favorable conditions for its growth and development [22].

After collection, the AMM plants were sealed, labeled, and transported to the laboratory for subsequent analysis. Representative samples of the collected AMM leaves and roots are shown in Figure 2a,b.

The labeled AMM roots were stored in refrigeration cabinets and sent to Shanghai Dingtai Testing Technology Co., Ltd., for the analysis of their bioactive compounds, including flavonoids, saponins, and polysaccharides.

Table 1. Statistics on the contents of 3 bioactive compounds in AMM samples.

Compound	Maximum (mg/g)	Minimum (mg/g)	Average (mg/g)	Standard Deviation
Flavonoids	11.94	3.0	7.53	2.44
Saponins	4.02	1.81	2.47	0.51
Polysaccharides	59.89	18.75	41.99	9.58

displays the statistical summary of flavonoid, saponin, and polysaccharide concentrations measured in AMM.

2.1.2. Acquisition of Hyperspectral Data from AMM Leaves

A handheld Specim IQ hyperspectral imaging system was used to capture images of AMM leaves. The acquired hyperspectral data had a spatial resolution of $512\times 512$ pixels, a spectral resolution of approximately 3 nm, and covered a spectral range of 400–1000 nm. During the acquisition process, a vertical incidence approach was adopted, utilizing a fixed support to maintain a consistent 90° angle between the sensor and the leaf surface, thereby minimizing the influence of incidence angle variations on reflectance measurements. Meanwhile, to minimize the influence of ambient lighting on the leaf spectra, image acquisition was performed in a controlled, darkened indoor setting. Two halogen lamps were employed to provide consistent illumination of the AMM leaves. To ensure lighting stability, the halogen lamps were preheated for approximately 20 min after their angles were fixed, prior to image acquisition. Additionally, white–black panel calibration was conducted before imaging to mitigate the effects of background noise and uneven illumination.

In this study, all samples were collected uniformly when the plants had reached physiological maturity and were in the vigorous growth stage. To comprehensively represent the overall physiological status of the plants and minimize within-individual variation, a systematic stratified sampling strategy was adopted. According to the vertical structure of the canopy, 16 healthy, fully expanded leaves free from pests and diseases were selected from each of the upper, middle, and lower layers of every AMM plant. A total of 720 leaf samples were collected from 15 plants. This sampling approach effectively enhanced data reliability and ecological validity by intensifying sampling effort per plant while maintaining a limited number of sampled individuals. The SPXY algorithm [23] was used to allocate 80% of the leaf spectral data, along with corresponding bioactive compounds content (576 datasets), to the training set. The remaining 20% was divided equally into the validation and test sets, each comprising 72 datasets.

2.1.3. Preprocessing of Leaf Hyperspectral Data

Due to the small area of AMM leaves, the acquired hyperspectral data includes leaf information and considerable background information. This study selected the mid-lateral regions of the leaf samples as the regions of interest (ROIs), deliberately avoiding areas such as the leaf tip, base, and midrib where compositional heterogeneity is likely to occur. This strategy ensured that the acquired reflectance data were representative of the overall biochemical characteristics of the leaf, reducing measurement bias caused by spatial heterogeneity. In this study, the “LabelMe” annotation tool was used to outline the leaf contours, and ROIs were manually defined. The obtained masks were multiplied by the original hyperspectral data to separate the leaf hyperspectral data from the background. The pixel values within the ROI were averaged across the spatial dimension to calculate the average value for each wavelength, resulting in representative spectral curves. This process resulted in the spectral data for each leaf sample [24]. The corresponding spectral curves of the leaf samples are shown in Figure 3a.

Subsequently, the Z-Score normalization method was employed to standardize the leaf spectral data [25]. For spectral bands at both the lower and upper ends of the wavelength range, where the signal-to-noise ratio (SNR) is relatively low, this standardization yielded a more uniform and smoother data distribution. As shown in Figure 3b, it effectively enhanced the SNR and mitigated the influence of high-frequency noise. This preprocessing step facilitates faster convergence of gradient descent algorithms and improves the ability of deep learning to fit the data accurately [26]. The calculation formula of Z-Score normalization is presented in Equation (1):

$$Z_{ij}={\displaystyle \frac{X_{ij}-\frac{1}{n}{\sum }_{i=1}^n{X}_{ij}}{\sqrt{\frac{1}{n}{\sum }_{i=1}^n{(X_{ij}-\frac{1}{n}{\sum }_{i=1}^n{X}_{ij})}^2}}},$$

(1)

where $X_{ij}$ represents the value of the ith sample at the jth band in the original data, $Z_{ij}$ denotes the standardized data, and n is the total number of samples.

2.2. Feature Extraction from Leaf Hyperspectral Data

This study uses FD and CR algorithms with DWT to extract informative features from the hyperspectral data of AMM leaves. DWT decomposes spectral features into low-frequency and high-frequency components. The low-frequency component typically captures the main constituents and overall trends of the spectral data, while the high-frequency component preserves detailed and localized spectral features [27]. The corresponding calculations are presented in Equations (2) and (3) [28]:

$$A\left[k\right]=\sum _{n}x\left[n\right]g[2k-n],$$

(2)

$$D\left[k\right]=\sum _{n}x\left[n\right]h[2k-n],$$

(3)

where $A\left[k\right]$ and $D\left[k\right]$ represent the obtained low-frequency and high-frequency components, respectively; $x\left[n\right]$ denotes the original signal; $g\left[n\right]$ and $h\left[n\right]$ represent the high-pass and low-pass filter coefficients, respectively; and $2k$ indicates downsampling by a factor of 2.

The FD algorithm is primarily used to compute the rate of spectral change at each wavelength point, facilitating clearer identification and analysis of spectral absorption and reflection peaks. By enhancing subtle variations in the spectrum, FD improves the accuracy of compound identification. The corresponding calculation is expressed in Equation (4) [29]:

$$R^{\prime }\left({\lambda }_i\right)={\displaystyle \frac{R\left({\lambda }_{i+1}\right)-R\left({\lambda }_{i-1}\right)}{{\lambda }_{i+1}-{\lambda }_{i-1}}},$$

(4)

where ${\lambda }_{i-1}$, ${\lambda }_i$, and ${\lambda }_{i+1}$ represent adjacent wavelengths, and $R^{\prime }\left({\lambda }_i\right)$ denotes the first-order derivative reflectance at wavelength ${\lambda }_i$.

The CR algorithm effectively eliminates background interference, thereby clarifying the overall absorption and reflection characteristics of the spectrum. By emphasizing the dominant features within the spectral data, CR significantly enhances the contrast of the spectral curve. The calculation described by Equation (5) can be expressed as follows [30]:

$$C{R}_j={\displaystyle \frac{R_j}{R{C}_j}},$$

(5)

where $C{R}_j$ represents the continuum-removed spectral reflectance, $R_j$ denotes the original spectral reflectance, and $R{C}_j$ is the continuum line reflectance.

The application of multiple feature extraction methods to spectral data enables the capture of diverse and complementary characteristics. The FD algorithm enhances spectral detail and highlights abrupt changes, thereby extracting more discriminative features. In contrast, the CR algorithm emphasizes dominant spectral components, facilitating the extraction of key features. DWT further decomposes the spectral data into high- and low-frequency components, where high-frequency features capture localized detail, and low-frequency features reflect the global spectral trend. As illustrated in Figure 4 and Figure 5, the combination of FD and CR with DWT effectively separates refined high- and low-frequency features. This decomposition strategy lays a solid foundation for subsequent multi-feature fusion and significantly improves the prediction accuracy of key bioactive compounds—flavonoids, saponins, and polysaccharides—in AMM.

2.3. Overall Structure of WCT-MCFNet

The proposed WCT-MCFNet model is made up of three parts, including the WCT feature extraction module, the MCF feature fusion module, and the SF prediction module integrating 1D-CNN with a fully connected (FC) layer, as illustrated in Figure 6. First, the WCT feature extraction module extracts features from the hyperspectral data of AMM leaves, resulting in two distinct sets of high- and low-frequency features. Next, the MCF feature fusion module performs cross-fusion of the extracted high- and low-frequency features. Finally, the SF prediction module is applied to estimate the contents of the bioactive compounds flavonoids, saponins, and polysaccharides in AMM.

2.3.1. WCT Feature Extraction Module

Using a single-feature extraction method alone may not capture all the valuable information in the spectral data. Features extracted by different methods can exhibit complementarity, and fusing these complementary features allows for a more comprehensive identification of key characteristics, significantly enhancing the model’s accuracy. In this study, the proposed WCT feature extraction module combines the FD and CR algorithms with DWT and the spectral feature extraction network (SpectraNet) for feature extraction, resulting in a four-branch set of high- and low-frequency multivariate features.

The structural diagram of the WCT feature extraction module is shown in Figure 5a. Initially, the FD algorithm is applied to preprocess the original spectral information of AMM leaves, enhancing the detailed features in the spectrum. DWT is then used to decompose the spectral data into its high- and low-frequency components. These extracted features are fed into SpectraNet for further feature extraction. Similarly, the CR algorithm is applied to preprocess the original spectral data, emphasizing the primary features. After DWT’s decomposition into high- and low-frequency components, these features are also passed into SpectraNet for extraction.

SpectraNet is inspired by the core architecture of the 1D-CNN hyperspectral image classification model (1DCNN-HSI) proposed by Hu et al. [31]. This model consists of an input layer, a convolutional layer (C1), a pooling layer (M2), a fully connected layer (F3), and an output layer. In this framework, the C1 and M2 layers serve as trainable feature extractors for the input data, while the F3 layer functions as the classifier for the extracted features. The kernel size and number of kernels in the C1 layer are set to $k_1\times 1$ (where $k_1=\lceil n_1/9\rceil$, and $n_1$ is the number of input features) and 20, respectively. The M2 layer uses max pooling with a kernel size of $k_2\times 1$ (where $k_2=\lceil k_1/5\rceil$). After pooling, the resulting multidimensional data is flattened, ensuring a smooth transition from the C1 layer to the F3 layer. The hyperbolic tangent function (Tanh) is used as the activation function for both the C1 and F3 layers [31]. SpectraNet retains the C1, M2, and F3 layers from the 1DCNN-HSI model, further processing the extracted high- and low-frequency features. The structural diagram of SpectraNet is illustrated in Figure 7.

2.3.2. MCF Feature Fusion Module

Fusing the high- and low-frequency features obtained through DWT enhances the model’s sensitivity to detailed changes while preserving the primary characteristics of the data. By cross-fusing the high- and low-frequency features extracted using two different algorithms, the strengths of each algorithm can be fully leveraged. This fusion improves the expressive power of the features, enhancing the overall effectiveness of data analysis and model construction.

The proposed MCF feature fusion module effectively integrates the multivariate features extracted from the WCT feature extraction module. The detailed analysis is provided in Figure 6b; the process begins with the cross-fusion of low-frequency features derived from the FD-DWT combination and high-frequency features obtained from the CR-DWT combination through multi-feature concatenation. Similarly, the high-frequency features from the FD-DWT combination and the low-frequency features from the CR-DWT combination are also cross-fused. Next, a multivariate feature integrator (MFI), implemented through a fully connected layer, is used to further extract features from the fused multivariate features. Finally, the features are concatenated once again to enhance their representation before being passed to the next stages of the model.

2.3.3. SF Prediction Module

In this study, an enhanced version of SpectraNet is employed as the prediction module for estimating the bioactive compounds of AMM. To obtain the final prediction results, an additional FC is appended to the original SpectraNet architecture. Specifically, the fused features generated by the MCF feature fusion module are first input into SpectraNet for further feature extraction. Subsequently, the extracted features are fed into the FC layer to perform the regression prediction. The output of the FC layer comprises three real-valued predictions, corresponding to the concentrations of the three bioactive compounds—flavonoids, saponins, and polysaccharides—in AMM. The architecture of the SF prediction module, incorporating SpectraNet and the FC layer, is illustrated in Figure 8.

2.4. Evaluation Metrics

To evaluate the prediction performance of the proposed model in predicting the concentrations of bioactive compounds in AMM, this study adopts four quantitative metrics, including the coefficient of determination (R²), root mean square error (RMSE), relative percentage deviation (RPD), and mean absolute error (MAE). These metrics collectively assess the model’s predictive accuracy and robustness. The formulas for R², RMSE, and RPD are provided in Equation (6), Equation (7), and Equation (8), respectively [32,33,34], while the calculation for MAE is presented in Equation (9) [35].

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{({\widehat{y}}_i-y_i)}^2}{{\sum }_{i=1}^n{({\widehat{y}}_i-\overline{y})}^2}},$$

(6)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{({\widehat{y}}_i-y_i)}^2}{n}}},$$

(7)

$$\mathrm{RPD}={\displaystyle \frac{\sqrt{\frac{1}{n-1}{\sum }_{i=1}^n(y_i-\overline{y})}}{\sqrt{\frac{1}{n}{\sum }_{i=1}^n(y_i-{\widehat{y}}_i)}}},$$

(8)

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_i-{\widehat{y}}_i\right|,$$

(9)

where $y_i$ is the true value of the ith sample, ${\widehat{y}}_i$ is the predicted value of the ith sample, $\overline{y}$ is the average of all predicted values, and n is the total number of samples.

A higher value of R² (closer to 1) and a lower value of RMSE (closer to 0) indicate the superior predictive accuracy of the model. Conversely, a larger MAE reflects poorer model performance. The RPD metric evaluates the consistency between the predicted and true values, with higher RPD values signifying stronger predictive capability [36,37].

2.5. Technical Roadmap

The technical roadmap of this study is illustrated in Figure 9. Using AMM as the subject of research, we developed a prediction model for bioactive compound content based on HSI. First, white–black panel calibration was performed on the hyperspectral camera to acquire hyperspectral data of the AMM leaves, from which the average spectral curves of ROIs were extracted and standardized. Simultaneously, the contents of bioactive compounds in the roots were determined through chemical analytical methods. Subsequently, the obtained leaf spectral data and bioactive compound data were divided into training, validation, and test sets. The spectral data were then subjected to feature extraction using a combination of FD, CR, and DWT algorithms, and the resulting multivariate features were fully utilized through cross-fusion to exploit their complementary information. Finally, a prediction module was employed to predict the contents of three bioactive compounds (flavonoids, saponins, and polysaccharides) in AMM. The model performance was evaluated using R², RMSE, RPD, and MAE metrics, and the optimized model was applied to predict the bioactive compound contents in the test set.

3. Results

3.1. Experimental Environment and Parameter Configuration

The experiments for the WCT-MCFNet model were conducted on a workstation with an AMD Ryzen Threadripper PRO 3945WX CPU, an NVIDIA RTX A4000 GPU (16 GB memory), and 128 GB of RAM. The operating system was Windows Server 2022 Datacenter. The model was trained using the Adam optimizer, with a batch size of 8, a learning rate of 0.001, and a total of 300 epochs.

Table 2. Prediction results of flavonoid, saponin, and polysaccharide contents under different parameter conditions.

Learning Rate	Batch Size	Epoch	Flavonoids				Saponins				Polysaccharides
			R2	RMSE	RPD	MAE	R2	RMSE	RPD	MAE	R2	RMSE	RPD	MAE
0.0001	8	300	0.884	0.91	2.94	0.62	0.860	0.18	2.67	0.13	0.820	4.06	2.37	2.64
0.001	8	300	0.981	0.37	7.30	0.14	0.992	0.04	10.97	0.02	0.992	0.86	11.16	0.38
0.01	8	300	0.604	1.69	1.59	1.24	0.514	0.33	1.43	0.24	0.642	5.77	1.67	4.52
0.001	4	300	0.943	0.64	4.21	0.23	0.97	0.08	5.78	0.04	0.96	1.90	5.08	0.77
0.001	8	300	0.981	0.37	7.30	0.14	0.992	0.04	10.97	0.02	0.992	0.86	11.16	0.38
0.001	16	300	0.913	0.79	3.39	0.40	0.906	0.15	3.27	0.09	0.916	2.78	3.46	1.64
0.001	8	100	0.954	0.57	4.67	0.29	0.941	0.11	4.13	0.07	0.946	2.29	4.32	1.16
0.001	8	200	0.932	0.90	3.86	0.25	0.971	0.08	2.92	0.04	0.962	1.86	5.18	0.85
0.001	8	300	0.981	0.37	7.30	0.14	0.992	0.04	10.97	0.02	0.992	0.86	11.16	0.38

lists a comparison of the prediction results for the three bioactive compounds in the test set, evaluated under different parameter configurations.

3.2. Training Results

The prediction results for the contents of flavonoids, saponins, and polysaccharides in AMM, obtained from the test set, are shown in Figure 10. The scatter points for all three bioactive compounds are closely aligned with the 1:1 line, with R² values of 0.981, 0.992, and 0.992, respectively. These results indicate that the WCT-MCFNet model can accurately predict the bioactive compounds of AMM.

Figure 11 displays the distribution of prediction residuals for the contents of three compounds—flavonoids, saponins, and polysaccharides—in the test set. As can be seen from the figure, the prediction error for flavonoid content is within ±1.0, the prediction error for saponins does not exceed ±0.1, and the prediction error for polysaccharide content falls within ±0.6. These results indicate that the predicted values of the three bioactive compounds are in strong agreement with the measured values, demonstrating the excellent predictive performance of the model.

3.3. Comparative Experiments

The predictive performance of the proposed WCT-MCFNet model for bioactive compound contents in AMM was rigorously evaluated through a series of carefully designed comparative experiments. The hyperspectral data of AMM leaves were used as input, and the predicted contents of flavonoids, saponins, and polysaccharides served as the output.

To validate the effectiveness of using leaf HSI for indirectly predicting root bioactive compounds, this study compared it with the direct prediction approach based on root HSI.

Table 3. Comparison of prediction results using HSI of different organs.

Compound	Organ	Evaluation Metrics
		R2	RMSE	RPD	MAE
Flavonoids	Root	0.985	0.26	8.87	0.11
	Leaf	0.981	0.37	7.30	0.14
Saponins	Root	0.993	0.03	12.37	0.02
	Leaf	0.992	0.04	10.97	0.02
Polysaccharides	Root	0.992	0.81	11.89	0.33
	Leaf	0.992	0.86	11.16	0.38

indicates that the prediction accuracy achieved by the two methods was comparable. Although root HSI provided a slightly higher accuracy, the difference was not statistically significant. However, the root-based method exhibited clear limitations in data acquisition efficiency, whereas the use of leaf HSI significantly improved efficiency while maintaining comparable precision. As the primary photosynthetic organs, leaves can systematically reflect biochemical changes at the whole-plant level. Leaf spectra are closely related to the overall physiological status of the plant, and their non-destructive nature allows for continuous dynamic monitoring of the same individual. In contrast, acquiring root HSI requires destructive sampling and is further constrained by soil adhesion interference and excavation difficulties, which considerably reduce the efficiency of spectral collection and analysis. Therefore, this study demonstrates that leaf HSI serves as an efficient and reliable alternative technique for non-destructive monitoring and assessment of internal biochemical constituents in plants.

The predictive effectiveness of WCT-MCFNet was systematically validated through benchmarking against established traditional prediction models, including logistic regression (LR), partial least squares regression (PLSR), and support vector regression (SVR). According to the experimental comparison results summarized in

Table 4. Comparison of prediction results using traditional models.

Compound	Model	Evaluation Metrics
		R2	RMSE	RPD	MAE
Flavonoids	LR	0.712	1.33	1.86	1.05
	SVR	0.609	1.55	1.60	1.24
	PLSR	0.723	1.31	1.90	1.03
	WCT-MCFNet	0.981	0.37	7.30	0.14
Saponins	LR	0.749	0.29	2.00	0.24
	SVR	0.707	0.32	1.85	0.26
	PLSR	0.778	0.28	2.12	0.22
	WCT-MCFNet	0.992	0.04	10.97	0.02
Polysaccharides	LR	0.637	5.48	1.66	4.36
	SVR	0.535	6.20	1.46	4.46
	PLSR	0.651	5.37	1.69	4.31
	WCT-MCFNet	0.992	0.86	11.16	0.38

, the R² values obtained using these traditional regression models for predicting the content of bioactive compounds in AMM did not exceed 0.8. This is mainly because traditional machine learning algorithms are generally limited to predicting the content of a single substance without considering the intrinsic correlations among multiple compounds. In contrast, the WCT-MCFNet model can simultaneously predict the contents of flavonoids, saponins, and polysaccharides, leveraging the interrelationships between different compounds. This enhances both prediction accuracy and the model’s generalization ability.

To demonstrate the superiority of the WCT-MCFNet model, it was compared with two recently proposed models for bioactive compound prediction, TCNA [32] and CSAM-CNN [38]. Full-band original spectral data were used as input variables for establishing comparison models. According to the data provided in

Table 5. Comparison of prediction results using latest models.

Compound	Model	Evaluation Metric
		R2	RMSE	RPD	MAE
Flavonoids	TCNA	0.657	1.57	1.71	1.14
	CSAM-CNN	0.803	1.19	2.25	0.80
	WCT-MCFNet	0.981	0.37	7.30	0.14
Saponins	TCNA	0.448	0.35	1.35	0.25
	CSAM-CNN	0.860	0.18	2.67	0.12
	WCT-MCFNet	0.992	0.04	10.97	0.02
Polysaccharides	TCNA	0.596	6.13	1.57	4.71
	CSAM-CNN	0.745	4.87	1.98	3.32
	WCT-MCFNet	0.992	0.86	11.16	0.38

, the WCT-MCFNet model achieved the best prediction performance, with an average R² value close to 0.99, demonstrating the effectiveness of cross-fusing complementary features obtained from different feature extraction methods. In contrast, the TCNA model showed the lowest R² values for the three compounds, none exceeding 0.7, and RPD values less than 2, indicating that this model provides only rough predictions for the bioactive compounds of AMM. The CSAM-CNN model, while superior to TCNA, still fell short of the WCT-MCFNet model in terms of prediction accuracy.

The advantages of the lightweight network were demonstrated by employing spatial–spectral joint information derived from AMM leaf hyperspectral data as the control group. Prediction modeling was performed using the ResMorCNN (3D) model [39] and the 3D-2DCNN-CA model [40]. The comparison results are summarized in

Table 6. Comparison of prediction results using different dimensional models.

Compound	Model	Time/s	Evaluation Metric
			R2	RMSE	RPD	MAE
Flavonoids	ResMorCNN	856.85	0.776	1.27	2.11	0.80
	3D-2DCNN-CA	1320.78	0.932	0.70	3.85	0.42
	WCT-MCFNet	486.52	0.981	0.37	7.30	0.14
Saponins	ResMorCNN	856.85	0.897	0.15	3.11	0.10
	3D-2DCNN-CA	1320.78	0.951	0.10	4.54	0.07
	WCT-MCFNet	486.52	0.992	0.04	10.97	0.02
Polysaccharides	ResMorCNN	856.85	0.775	4.58	2.10	2.83
	3D-2DCNN-CA	1320.78	0.884	3.28	2.94	1.96
	WCT-MCFNet	486.52	0.992	0.86	11.16	0.38

; the WCT-MCFNet model requires significantly less training time, making it more suitable for deployment on mobile devices or embedded systems. This feature is particularly advantageous for real-time applications and large-scale data analysis.

This study employed a comparative framework incorporating both early and late fusion strategies to systematically evaluate the superiority of intermediate fusion for multivariate feature integration. Figure A1 illustrates the structural diagram of early fusion applied to the hyperspectral data of AMM leaves. Compared to intermediate fusion, early fusion advances the fusion point, directly concatenating the multivariate features obtained after wavelet decomposition into four branches. It then uses SpectraNet for feature extraction. Figure A2 shows the structural diagram of late fusion. Compared to intermediate fusion, late fusion delays the fusion point, allowing each branch to undergo further feature extraction through MFI before concatenating the multivariate features.

According to the data provided in

Table 7. Comparison of prediction results using different fusion strategies.

Compound	Fusion Strategy	Evaluation Metric
		R2	RMSE	RPD	MAE
Flavonoids	Early	0.858	1.01	2.65	0.39
	Intermediate	0.981	0.37	7.30	0.14
	Late	0.952	0.59	4.58	0.21
Saponins	Early	0.922	0.13	3.57	0.07
	Intermediate	0.992	0.04	10.97	0.02
	Late	0.974	0.08	6.25	0.03
Polysaccharides	Early	0.818	4.12	2.34	1.73
	Intermediate	0.992	0.86	11.16	0.38
	Late	0.964	1.82	5.31	0.66

, the R² value obtained using intermediate fusion is significantly higher than that of early fusion, indicating that early fusion may combine features before key features are fully extracted. This premature fusion leads to combined features that still contain a large amount of redundant or noisy information. The R² value for late fusion is also lower than that of the WCT-MCFNet model, suggesting that performing multiple feature extractions on each branch may result in the loss of effective information. This makes it challenging to capture the correlations between branches and fully utilize complementary information among multivariate features. In contrast to early and late fusion, intermediate fusion more effectively leverages the correlations and complementarity between features while extracting useful information and minimizing redundancy.

To evaluate the correlation between the multivariate features obtained by combining different feature extraction methods with wavelet decomposition, an ablation experiment was conducted. Two control groups were established, a single-branch group and a multi-branch group. The single-branch group consists of scenarios without feature extraction and those using only a single-feature extraction method, while the multi-branch group includes combinations of original data with a single-feature extraction method, excluding the SpectraNet model. The structural diagrams for using a single-feature extraction method and for excluding SpectraNet are illustrated in Figure A3 and Figure A4, respectively.

As shown in

Table 8. Comparison of results obtained using single-branch model.

Compound	Feature Extraction	Single Branch	Evaluation Metric
			R2	RMSE	RPD	MAE
Flavonoids	Not employed	Original	0.935	0.68	3.93	0.22
	Single	FD	0.974	0.43	6.17	0.27
		CR	0.938	0.67	4.02	0.19
Saponins	Not employed	Original	0.926	0.13	3.67	0.05
	Single	FD	0.961	0.09	5.05	0.07
		CR	0.974	0.08	6.17	0.03
Polysaccharides	Not employed	Original	0.919	2.75	3.50	1.04
	Single	FD	0.964	1.82	5.28	0.95
		CR	0.952	2.11	4.56	0.76

, in the single-branch group, the use of a single-feature extraction method improved the model’s accuracy compared to the scenario without feature extraction. However, due to the limitations of a single model, it could only capture a limited set of spectral features, missing some that are strongly correlated with the bioactive compounds of AMM, thus failing to achieve optimal results.

The experimental comparison results are summarized in

Table 9. Comparison of results obtained using multi-branch model.

Compound	Feature Extraction	Multiple Branch	Evaluation Metric
			R2	RMSE	RPD	MAE
Flavonoids	Original + Single	Original + FD	0.909	0.81	3.32	0.29
		Original + CR	0.914	0.79	3.41	0.20
	Multiple	FD + CR	0.981	0.37	7.30	0.14
	Removing SpectraNet	FD + CR	0.958	0.55	4.88	0.32
Saponins	Original + Single	Original + FD	0.969	0.08	5.68	0.05
		Original + CR	0.970	0.08	5.75	0.02
	Multiple	FD + CR	0.992	0.04	10.97	0.02
	Removing SpectraNet	FD + CR	0.962	0.09	5.10	0.06
Polysaccharides	Original + Single	Original + FD	0.937	2.42	3.98	0.97
		Original + CR	0.959	1.95	4.94	0.53
	Multiple	FD + CR	0.992	0.86	11.16	0.38
	Removing SpectraNet	FD + CR	0.915	2.81	3.42	1.36

. In the multi-branch group, combining original data with a single-feature extraction method did not significantly improve model performance. In fact, some models showed a decline in effectiveness, suggesting that noise in the original data interfered with the model’s training process. When compared to the WCT-MCFNet model, the prediction performance deteriorated after removing SpectraNet, with R² values for all three bioactive compounds decreasing, particularly for polysaccharides, which dropped by 0.08. This indicates that SpectraNet, as a feature extraction method, plays a crucial role in enhancing the prediction accuracy of bioactive compounds in AMM.

4. Discussion

AMM holds significant medicinal value in both traditional Chinese medicine and modern medicine. Traditional methods for detecting its bioactive compounds typically require destructive sampling and chemical analysis of the roots, which are complex, time-consuming, and costly [41,42]. However, research on indirectly predicting the content of bioactive compounds in the roots using hyperspectral data from AMM leaves remains limited. The WCT-MCFNet model proposed in this study, based on hyperspectral data from AMM leaves, successfully achieved high-precision prediction of flavonoid, saponin, and polysaccharide contents in the roots. This model innovatively integrates detailed and global spectral features from various feature extraction algorithms, with R² values exceeding 0.98 and RPD metrics greater than 7, significantly outperforming single-feature extraction methods. These results highlight the correlation between leaf hyperspectral features and root bioactive compound content, offering a new, non-destructive approach to monitoring medicinal plant compounds, thus breaking away from traditional methods that rely on direct root analysis.

Table 8 and Table 9 compare the performance of different feature extraction strategies. As shown, the WCT-MCFNet model, which combines the FD and CR algorithms, outperforms other methods, achieving the highest R² and RPD values, as well as the lowest RMSE and MAE for the prediction of the three bioactive compounds. Due to the high dimensionality and redundancy of hyperspectral data, directly using original spectral data for modeling often leads to interference from noise and irrelevant information. In contrast, feature extraction algorithms can effectively enhance model performance [43,44,45]. The results in Table 8 demonstrate that, compared to models without feature extraction, introducing a single-feature extraction algorithm significantly improves the R² values for flavonoids, saponins, and polysaccharides. This indicates that feature extraction algorithms can effectively highlight the key spectral features sensitive to the bioactive compounds of AMM, underscoring the importance of feature extraction in hyperspectral data analysis [46,47,48].

Feature extraction methods can effectively enhance and extract key information from the spectrum that is sensitive to target parameters, reducing interference from background noise and redundant data [49,50,51,52]. However, most existing prediction models based on hyperspectral data typically rely on a single-feature extraction algorithm, overlooking the limitations of this approach and the complementary information that multiple feature extraction methods can provide [53,54,55]. By comparing the performance of single-feature extraction methods with the combined use of multiple methods, this study found that the fusion of multiple feature extraction techniques significantly outperforms individual methods in the prediction of flavonoids, saponins, and polysaccharides. The joint application of multiple feature extraction methods enables the capture of diverse spectral information, fully exploiting the complementarity between features to enhance model prediction accuracy [56,57,58,59]. According to the data provided in Table 9, multi-branch models using original data or data processed by a single-feature extraction algorithm did not show significant improvement in prediction performance compared to single-branch models using only one extraction method, with some even experiencing a decline in accuracy. This confirms that directly using original hyperspectral data for modeling may introduce background noise and redundant information, which interferes with the model’s ability to identify effective features, thereby obscuring the key information extracted by feature extraction algorithms and reducing model performance [60,61,62].

The high-precision prediction performance of the WCT-MCFNet model is primarily attributed to the feature enhancement provided by different feature extraction methods and the scientifically designed multi-level feature fusion strategy. At the data input stage, the FD and CR algorithms enhance the detailed and global features of the spectrum, respectively. Subsequently, DWT further decomposes these features into high- and low-frequency sub-components, which are then combined with SpectraNet for feature extraction. This approach effectively addresses the challenges of high dimensionality and redundancy in hyperspectral data. Meanwhile, the MFI leverages the complementarity of multi-branch features through a cross-fusion network. Compared to traditional serial or parallel fusion methods, such as PCA or simple CNN stacking, this design better aligns with the complex, multi-factor synergistic effects in plant physiological metabolism, significantly improving the prediction accuracy of bioactive compounds like flavonoids, saponins, and polysaccharides in AMM.

Although the WCT-MCFNet model proposed in this study makes significant advancements in detecting the content of bioactive compounds in AMM, it still has some limitations. First, the hyperspectral data used for model input was collected from AMM leaves grown in a specific environment and at a single growth stage. As a result, the model’s generalization ability to plants from different growing regions or stages has not been validated. Second, the model relies solely on spectral information from hyperspectral data and does not fully utilize the accompanying image data. This limitation may result in the loss of some key features, which could hinder further optimization. Additionally, the model has not yet explored the correlation between specific spectral bands and particular bioactive compounds, which may require validation using dynamic chemical composition monitoring data. Furthermore, the relatively small amount of experimental data poses a risk of overfitting, particularly in deep learning, which require large datasets to support the many parameters involved in achieving adequate model fitting.

This study aimed to indirectly estimate the content of bioactive compounds in the roots of AMM using hyperspectral data from its leaves combined with deep learning. The proposed WCT-MCFNet model achieved high-precision prediction of three key bioactive compounds in AMM roots—flavonoids, saponins, and polysaccharides—providing a rapid and non-destructive solution for the quality detection of traditional Chinese medicinal materials. Future research could enhance the model’s generalizability by constructing a large-scale hyperspectral-component database that spans different regions and multiple growth stages. Additionally, incorporating image information from hyperspectral data (such as texture and structural features) and combining it with three-dimensional convolutional neural networks or spatial–spectral joint analysis methods could fully leverage the synergistic effects of spectral and spatial features, improving model accuracy. Further research could also explore the development of lightweight embedded models based on the WCT-MCFNet model, integrating hyperspectral sensors with mobile computing devices to enable real-time field detection, thus facilitating the transition of this technology from the laboratory to industrial applications.

5. Conclusions

To effectively and non-destructively predict the bioactive compound content of AMM, a cross-fusion network integrating HSI and deep learning was developed. To address the limitation of conventional feature extraction methods in fully utilizing spectral information, both the FD algorithm and CR algorithm were introduced for preprocessing and feature extraction. The results demonstrated that the combined use of different feature extraction methods leveraged their respective advantages, significantly enhancing the capability to extract effective spectral features. Furthermore, considering the complementary information among different features, a multivariate feature cross-fusion network based on DWT was proposed. Comparative results showed that the cross-fusion network comprehensively captured the information related to bioactive compound content in the hyperspectral data of AMM leaves, achieving outstanding prediction performance with an average R² of 0.988 for the three bioactive compounds. These findings strongly demonstrate the high performance and reliability of the proposed method in predicting AMM bioactive compound content using hyperspectral data.

This study also confirmed the feasibility of using leaf-based hyperspectral data to predict root-derived bioactive compounds (flavonoids, saponins, and polysaccharides) in AMM. It advances the application of hyperspectral technology combined with deep learning for bioactive compound prediction and provides a new reference for the rapid and effective estimation of other bioactive compounds using leaf hyperspectral data.

Looking ahead, HSI holds significant potential for applications in precision agriculture and research on traditional Chinese medicine resources. On one hand, future efforts could focus on integrating HSI with unmanned aerial vehicles or near-ground mobile platforms to achieve synchronous and integrated monitoring of physiological and biochemical parameters across multiple plant organs at field scale. On the other hand, developing multi-source data fusion strategies—incorporating environmental factors and growth models—may enable the construction of more robust prediction models for medicinal constituent accumulation across multiple growing years. These innovative directions are expected to overcome the current limitations of single-organ and static analyses, thereby advancing HSI technology from mere component retrieval toward holistic physiological diagnostics of plants and end-to-end quality tracking of medicinal materials.