Fusion of Spectral and Textural Data of Hyperspectral Imaging for Glycine Content Prediction in Beef Using SFCN Algorithms

Glycine, the simplest free amino acid. It is one of the important factors affecting the flavor of beef. In this study, a fast and non-destructive method combining near-infrared hyperspectral (900–1700 nm) and textural data was first proposed to determine the content and distribution of glycine in beef. On the basis of spectral information pre-processing, spectral features were extracted by the interval variable iterative space shrinkage approach, competitive adaptive reweighting algorithm, and uninformative variable elimination (UVE). The glycine content prediction models were established by partial least squares regression, least squares support vector machine, and the optimized shallow full convolutional neural network (SFCN). Among them, the UVE-SFCN model was found to show better results with prediction set determination coefficient (RP2) of 0.8725. Furthermore, textural features were extracted by the gray-level co-occurrence matrix and fused with the spectral information of the best feature band to obtain an optimized UVE-FSCN-fusion model (RP2 = 0.9005, root mean square error = 0.3075, residual predictive deviation = 0.2688). Compared with the full spectrum and characteristic wavelength spectrum models, RP2 was improved by 6.41% and 3.10%. The best fusion model was visualized to represent the distribution of glycine in beef. The results showed that the prediction and visualization of glycine content in beef were feasible and effective, and provided a theoretical basis for the hyperspectral study of meat quality monitoring or the establishment of an online platform.


Introduction
Beef has gradually become the main edible meat product with abundant proteins and similar amino acids ratio to human's needs (Yda et al. 2020). In the past, amino acids in meat were generally studied in the form of total nitrogen or total amino acids, for example, characterization of meat spoilage by total volatile basic nitrogen (Alaa et al. 2021) and supplementation of total amino acids to affect broiler performance (Pratima et al. 2022). Free amino acid, a single amino acid molecule in a free state, could be directly absorbed and utilized. It is the main nutrient source in beef products and is also an important odor substance. The flavor function of amino acids is related to its hydrophobicity to a certain extent. When the hydrophobicity of amino acids is low, they are mainly sweet; on the contrary, they are bitter. Amino acids with acidic side chains are one of the most significant prerequisites for sourness and freshness (Chong and Ham 2014). The flavor of broth can be replicated through different free amino acids, proving its important role in flavor generation (Golbraikh 2014). Glycine is the simplest flavor amino acid among free amino acids with low hydrophobicity, so it contributes to sweet taste. It also plays a variety of vital roles, including immune regulation (Liem-Moolenaar and Zoethout 2010), antibacterial, and bactericidal action (Li 2015), enhancing Maillard reaction, enhancing meat taste (Chen et al. 2016), and improving the body's antioxidant capacity (Girmatsion and Mahmud 2021). Too much or too little glycine intake may affect the absorption balance of amino acids in the human body, leading to nutritional imbalance and poor health (Morikawa et al. 2019;Kehchiro muramatsu 1971). Therefore, the quantitative detection of glycine content is particularly important.
Previous studies on free amino acids mainly focused on the innovative applications of traditional physical and chemical detection methods, including high-performance liquid chromatography (HPLC), reverse high-performance liquid chromatography (RP-HPLC), and amino acid automatic analyzer. HPLC and RP-HPLC were used to determine the content of free amino acids in poultry and plasma (Kucha and Liu 2018;Kobayashi et al. 2010). Electronic tongues and electronic noses were used to identify the geographical origin of two salmonid species based on free amino acid contents (Duan et al. 2021). Amino acid analyzer was used to evaluate the free amino acid content of beef according to the country of origin and marbling score (Jackman and Sun 2008). Traditional testing methods are accurate and high-throughput, but they are time-consuming, destructive, and tedious and cannot meet the requirements of rapid nondestructive testing.
With the development of science, hyperspectral imaging (HSI) technology, an efficient and non-destructive detection method, has been employed in food detection. HSI integrates with the main features of imaging technology and spectroscopy and is used to obtain continuous spectral and spatial information from objects simultaneously. Ripeness of Camellia oleifera fruit was assessed by near-infrared (NIR) hyperspectral combined with PLS-DA (Jiang et al. 2022). Weng et al. (Weng et al. 2021) identified mutton geographical origin and breed accurately by HIS combined with RF. Using VIS/NIR spectroscopy combined with a variety of methods to quickly detect the adulteration of ground meat, the accuracy was improved, compared with the original method (Weng et al. 2020). The non-destructive detection of beef corruption index was carried out by fluorescence hyperspectral, and a model for the corruption index was established (Liu et al. 2020). In addition, hyperspectral imaging has great potential in the study of free amino acids. Textural information, as important feature information of hyperspectral technology, was fused with spectral information for food detection. For example, Palmitic and oleic acid content in mutton was established by integrating NIR hyperspectral with texture information (Wang and Wang 2020). Christopher et al. (Christopher et al. 2022) predicted pork intramuscular fat content, and the results of the fusion data were better than those predicted using spectral detection in isolation. However, to our knowledge, the prediction of free amino acid content in beef by hyperspectral and textural data fusion has not been reported.
In recent years, convolutional neural network (CNN) plays an important role in the field of machine learning. CNN reduced the complexity of the traditional neural networks and optimized mass data, becoming the ideal choice for analyzing and processing hyperspectral data. The mixed deep learning model based on hyperspectral remote sensing was used to predict strawberry yield with the model accuracy rate of 70% (Nassar et al. 2020). HSI was combined with CNN in soybean price prediction and Conv-LSTM used to improve RMSE by 31%. Fully convolutional network (FCN) was a framework used for image semantic segmentation (Zhang et al. 2020). FCN replaced the full connection layer with the convolution layer, accepting any input size while increasing the network depth. It becomes more efficient with head of the problems of double storage and computing convolution. Remote sensing hyperspectral and FCN were used to predict the depth of mature rice, and its performance was much better than the regression model based on Vis-HSI (Morikawa et al. 2019). However, the up-sampling results of SCN are fuzzy and not sensitive to details. Therefore, a lightweight shallow full convolutional neural network (SFCN) was proposed in this study. The network depth and the up-sampling influence were reduced to fit in smaller datasets during the running process of SFCN (Cui et al. 2022).
In the current paper, spectral data of representative beef varieties in Northwest China were collected by NIR hyperspectral (900-1700 nm) technology. After processing, the spectral data of the optimal characteristic wavelength was obtained. Textural information was got through gray-level co-occurrence matrix (GLCM) method. The deep learning model SFCN of glycine content was established and optimized with fusion of hyperspectral and textural information. The study may provide a reference for the rapid quantitative detection of glycine in beef and a theoretical basis for the development of hyperspectral quality monitoring online platform and help improve beef quality and the sustainable and healthy development of the industry. The complete test process was shown in Fig. 1.

Sample preparation
A representative breed of cattle from Northwest China was selected, with average fresh carcass weight of 285 ± 45 kg. The cattle were slaughtered in accordance with the principles and guidelines established by the Animal Care and Use Committee of the Institute of Food Science and Technology, Chinese Academy of Agricultural Sciences. Acid was released in 24 h after death, before dissecting. The three muscles, foreleg (FL), hindquarter (HQ), and longissimus dorsi (LD), were labeled, vacuumed into polyethylene bags, and transported to the Agricultural Products Non-destructive Testing Laboratory, Food College, Ningxia University, China, by an ice box. Visible fat and connective tissue were removed and cut into slices (approximately 4 cm × 3 cm × 1 cm). They were then transported to the Spectral Analysis Laboratory and Food Testing and Analysis Center, Ningxia University, respectively. The glycine content was measured by an amino acid automatic analyzer, and the sample sections of the same site were measured by NIR hyperspectral. The samples were placed at room temperature for spectral measurement, and a total of 360 beef samples (120 samples in each part) were obtained. The key to establishing an effective multivariate prediction model is to ensure that the chemical and spectral values of the same cross-section sample are measured at the same time.

Chemical Analysis
A mixed beef sample (2.00 g) was put into a 10-mL centrifuge tube, then 0.02 mol/L hydrochloric acid was added to dissolve in a constant volume. Then, 5.00 mL methanol and 5.00 mL water were added, respectively, to activate the C 18 pre-treated column. A 2.50-mL sample was added, followed by 1.50 mL 0.02 mol/L hydrochloric acid. After passing the column, 0.02 mol/L hydrochloric acid was added to a final volume of 5.00 mL, then filtered through a 0.45-μm membrane. Parameter settings for the amino acid automatic analyzer were: wavelength 570 nm and 440 nm; injection volume 20.00 μL; reaction temperature 135 ± 5 ℃; separation column 57 ℃; and chromatographic column sulfonate cation resin separation.
During spectral scanning, light source intensity, exposure time, meat surface texture shape, and color may affect the size and spatial resolution of hyperspectral images collected, so reasonable scanning parameters should be set before image collection. Based on the preliminary experiment, the micro-hyperspectral scanning parameters were determined as follows: scanning speed 1.4 mm/s, back off speed 2 mm/s, exposure time 10 m/s, and gain 5. Finally, the resolution of the cubic spectral image was 320 Pix 360 Pix 256 Pix.
Black and white correction was performed on the original image before the experiment began to eliminate the influence of instrument light source. Briefly, the white image reflectivity of a standard total reflection whiteboard (100% of the Teflon board), and an all-black image reflectivity covering the lens (0% of the reflectivity) were obtained, then the black-and-white correction image of the sample was calculated using the following formula 1: where I is the obtained calibration image, R is the original obtained image, R d represents dark current image, and R w represents white reference image.

Region of Interest Selection and Spectral Information Extraction
The surface of beef sample was not smooth, and there was a shadowed area at the junction of sample and background in hyperspectral images. Meanwhile, the moisture on the surface of meat samples, fascia, and light source irradiation may have interference to spectral information. To weaken these effects, the size of the image was adjusted. On the premise of retaining the complete information of the tested sample, the unnecessary pixel regions in the original hyperspectral image were clipped by ENVI 4.8 software, and part of the background image was removed to reduce spatial dimensionality. The region of interest (ROI) of the sample was extracted by threshold method (threshold value 0.14). The corresponding spectral information in ROI was extracted, and its average value was calculated as the spectral value of the sample to be measured.

Spectral Data Analysis
A prediction model was established with spectral information. Before modeling, the data were analyzed as follows: (1) Division of correction set and prediction set. Based on sample set partitioning and joint x-y distance (SPXY) algorithm, 75% of the samples were selected as the calibration set, the remaining were modeled for the prediction set, and the prediction error of each sample was obtained.
(2) Pretreatment. Pretreatment improves the spectral resolution of the overlapping data and minimizes the adverse effects of the instrument. The following pretreatment methods were used: normalize, standard normal variable (SNV), detrend, baseline, and multiplication scattering correction (MSC). (3) Characteristic wavelength extraction. The whole band spectral data have high dimension and complicated information, and so may need lots of calculations, with low model accuracy. Therefore, dimensionality reduction of the original spectra is required to extract wavelength information closely related to glycine content. In this experiment, the interval variable iterative space shrinkage approach (iVISSA), competitive adaptive reweighting algorithm (CARS), and uninformative variable elimination (UVE) were used to extract eigenvalues of the spectra.

Establishment of Prediction Model and Performance Evaluation
Traditional linear regression model PLSR, non-linear regression analysis model LSSVM, and deep learning model SFCN were used to establish models for glycine content prediction, and different modeling methods were compared. PLSR, a multi-linear regression modeling method, integrates the ascendancy of multiple linear regression, principal component analysis, and canonical correlation analysis, and is fast in modeling and sensitive to outliers. By standardizing X and Y, the correlation coefficient matrix is established, the contribution rate is calculated, and then the regression equation is made. LSSVM is used for non-linear mapping linearization to increase accuracy and reliability, on the basis of SVM optimization of the problem constraints. Both linear and non-linear models are traditional and classical, and have been widely used in modeling and analysis. 1D-CNN model has been used for spectroscopic and chemometric analysis, but it is limited by the effective features (2) Y = X T * X of data sets and one-dimensional convolution extraction models. In this study, a two-dimensional spectral information matrix was obtained by one-dimensional spectral selfmultiplication (Formula 2). The depth, height, and width were arranged and expressed in SFCN. The network was mainly composed of an input layer, convolution layer, activation function, pooling layer, and other parts (Fig. 2). The neurons in each layer were connected only to a small area of the previous layer, forming a locally connected layer. Rectified linear unit (ReLU) was used as the activation function of the convolution layer. Input 2d data were convolved and pooled, and output parameters of feature extraction module were set (Table 1). The fill value was 0, and the step size was 1.
In this study, R C 2 , RMSECV, prediction set determination coefficient (R P 2 ), root mean square error of prediction (RMSEP), and residual prediction deviation (RPD) were used to evaluate the performance of the quantitative analysis models.  ENVI 4.8 was used to extract the ROI, Unscrambler X 10.4 was used for data preprocessing and PLSR modeling, and Matlab R2021a was used for characteristic wavelength extraction, SFCN, and LSSVM modeling. All data analyses were performed on a computer with Intel Core I7-10,700, 12-GB CPU, 2.9-gHz frequency, NVIDIA 3060 GPU, and 16-GB memory.

Extraction of Image Texture Information
Hyperspectral technology can not only provide important image information, but also provide rich texture information. Texture information is the projection of high-dimensional spectral information to low-dimensional space. The objects around can be effectively distinguished by measuring the difference between pixels and spatial neighborhood. GLCM is a comment method for texture extraction analysis. Texture features based on GLCM were reported to be very helpful for distinguishing accuracy of classification of rice infected by Chilo suppressalis and the accuracy of the model was above 97% (Wan et al. 2022).
ENVI 4.8 software was used to perform PCA transformation on hyperspectral images, and GLCM was applied for effective PC images of ROI. Four mutually independent GLCM-based texture features, including contrast (CONT), homogeneity (HOM), energy (ENG), and correlation (CORR), were extracted from gray co-occurrence matrices with different angles (0°, 45°, 90°, and 135°), and the distance between pixel pairs was set to 1 pixel. The values in four different directions were averaged to record the texture features of each sample. Four texture factors were obtained by the calculation formula (Table 2).

Glycine Visualization in Beef
Hyperspectral images of characteristic wavelengths were selected, each beef sample image was segmented (ENVI 4.8 software), and the information of each pixel point was extracted (Matlab 2021A). A model was established based on the data, and the glycine content of each pixel of the sample was calculated to obtain a grayscale image. Pseudo-color processing was performed on grayscale images to predict the content of glycine in beef samples through different color bands, resulting in color visualization.

Chemical Characteristics
The contents of glycine in different parts of selected beef were significantly different (Fig. 3). The content of glycine was the highest in LD, and was significantly lower in HQ (p < 0.05) and FL (p < 0.01). This may be related to the content of free amino acids. When the energy supply of free amino acids exceeds the energy required by the body, it participates in the tricarboxylic acid cycle and carbohydrate metabolism, and oxaloacetate (4C), citric acid (6C), cis aconitic acid (6C), α-ketoglutaric acid (5C), fumaric acid (4C), and other intermediates were produced. Free amino acids in HQ and FL of beef are decomposed for energy during exercise, while LD has less movement, so it needs more free amino acids to participate in the synthesis. This is consistent with the results of fatty acid content in Tan mutton (Wang and Wang 2020).

Table 2 Calculation formula of GLCM parameters
Where X is the column number of GLCM, Y is the row number of GLCM, and G (i, j) is the gray co-occurrence matrix Feature parameters Equation Fig. 3 The contents of glycine in in different parts of beef sample. Note: ** represents extremely significant, p < 0.01, and * represents significant, p < 0.05 The test samples (n = 360) were divided into calibration sets and prediction sets (Table 3). The best results were obtained with SPXY method, and the calibration set and the prediction set were not very different with similar data distribution. The results indicated that the glycine distribution was uniform between different muscles.

Spectral Features
The average spectral curves of three different parts of the beef sample in the spectral range of 900-1700 nm can be seen that LD tended to have the higher spectral reflectance than HQ and FL and the overall trend of the three parts were roughly the same (Fig. 4). This is similar to the results obtained in the near-infrared spectroscopy (Nolasco Perez et al. 2018). In the range of 1050-1400 nm, the spectral reflectance values of the three parts were different. After 1400 nm, the spectral reflectance values were lower than 0.1 due to the strong water absorption in the muscle (Bergmann et al. 2021). Peaks (Fig. 4a) and valleys (Fig. 4b) occurred in the entire spectral reflectance curve, mainly because chemical bonds absorb energy at specific wavelengths, and the spectral intensity is determined by changes in the molecular dipole moment (Zou et al. 2010). Therefore, the band in the range of 1400-1700 nm was the water absorption band of the second and first overtones related to O-H bond frequency doubling, indicating that water accounted for the largest proportion and absorbed most of the near-infrared radiation in the beef muscle. At 1050 nm, 1200 nm, and 1400 nm, O-H, C-H and N-H overtones appeared, respectively.

Prediction Models Using Full Spectral Range
Normalize, standard normal variables (SNV), detrend, baseline, and multiplicative scattering correction (MSC) were used to process the data, and a PLSR model of pretreated spectral reflectance was established to predict the content of glycine. Comparison of the R C 2 /R P 2 and RMSEC/RMSEP values of the PLSR model and the prediction set showed that a good result was obtained by the data preprocessing (Fig. 5).
Baseline was found to be the best preprocessing method for glycine prediction models in beef samples. The data obtained by other preprocessing methods also showed an improvement trend, indicating that most preprocessing methods provided favorable trends for the spectral data. Thus, selecting the right pretreatment method may increase the reliability of the prediction model. The spectral curves of data preprocessed by different methods were similar to baseline method (Fig. 6).

Effective Wavelengths
In this study, the iVISSA, CARS, and UVE were used to extract effective feature bands in the whole band range, and then the feature wavelength extraction method with the best effect was selected through modeling. The iVISSA algorithm extracted characteristic wavelengths through two iterative cycles of global and local search and finally obtained a set of optimized effective wavelengths. In the iterative operation, the sampling weight values of wavelengths varied with the number of iterations, and the weights ranged from 0 to 1 (Fig. 7a).
After 10 iterations, the sampling weight value was basically stable, and a total of 46 characteristic wavelengths were extracted. The 46 characteristic wavelengths selected by the iVISSA method were widely and uniformly distributed in the whole waveband, indicating that the iVISSA algorithm may well preserve the synergistic effect between wavelengths (Fig. 7b).
When the CARS method was used to extract characteristic wavelengths, the variable subset with the smallest RMSECV value was selected as the optimal subset. During the running process, the number of variables and the number of runs showed an exponentially decreasing relationship (Fig. 8a). RMSECV first decreased slowly, with a minimum value at the 275th sample run, and then increased, the number of optimal variable subsets decreased, and some valid information was removed, resulting in a decrease in model accuracy (Fig. 8b). The regression coefficient paths were different for different runs, reflecting the changes in the regression coefficients. The vertical line represented the minimum RMSECV value, which was 180 runs (Fig. 8c).The extracted 19 characteristic wavelengths were mainly distributed between 1100-1165 nm and 1200-1400 nm. The selected characteristic wavelengths were little but uniform, thus preserving some of the synergistic effects between wavelengths (Fig. 8d).
The UVE method preserves the information variables, reduces the frequency band dimension, and improves the prediction accuracy of the model. The left side of the vertical line showed 233 wavelength variables under full-spectrum conditions, while the right side showed 117 random variables (Fig. 9a). Two horizontal dashed lines separate useful variables (outside the dashed line) and useless variables (inside the dashed . The 84 characteristic wavelengths extracted by UVE method were evenly distributed, mostly in the wavelength range, and the combined effect of wavelengths was observed (Fig. 9b).

Prediction Models Using Effective Wavelengths
The three models, PLSR, LSSVM, and SFCN, based on different algorithms to extract feature wavelengths could effectively predict glycine contents in beef samples, and the models of all characteristic wavelengths showed good results (Table 4). In the full-band modeling of beef glycine content, the prediction set performance of PLSR method was 0.0481 higher than that of LSSVM method, indicating that the modeling effect of linear model PLSR was better than that of non-linear model LSSVM; the linear relationship between glycine content and spectral absorption in beef samples was greater than the non-linear relationship. The prediction set performance of SFCN method was 0.0423 higher than that of PLSR method, suggesting that SFCN had better potential in predicting the content of glycine than traditional models.  The number of bands retained by all feature wavelength processing methods ranged from 67.18% to 92.57%, which could avoid data redundancy and eliminate useless information. The modeling effect of extracting characteristic wavelengths based on different methods was better than that of full-band data modeling. The model after extracting characteristic wavelengths by the UVE method achieved the best results, with R p 2 being 0.0298 higher than that of the fullband model, indicating that extracting characteristic wavelengths retained useful information and improved model accuracy. Li et al. (Li et al. 2019) studied the hardness of different pear varieties based on Vis-NIR spectroscopy, and obtained the optimal UVE-SPA-LSSVM hardness prediction multi-variety model, with its R P 2 , RMSEP, and RPD P values of 0.94, 0.91, and 2.93, respectively. Our results suggested that the characteristic wavelengths selected by UVE were more accurate and effective. On the contrary, among the three modeling methods, CARS method was slightly worse and related to the less feature band extraction and the elimination of part of the effective information.

Prediction Models Using Combinations of Spectral and Textural Features
The optimal grayscale image had a critical influence on the optimal texture features. Principal component analysis was performed on beef sample images, and the cumulative contribution rate of the first three principal component images was 92.12% (Fig. 10). PC1, PC2, and PC3 explained 89.67%, 3.69%, and 0.76% of the image information, respectively, suggesting that the texture information of the first principal component could be used as the representative information of the sample. Therefore, the GLCM method  . 10 Images of the first three principal components of principal component analysis in beef samples Fig. 11 Correlation of glycine content with texture parameters extracted using GLCM was used to extract the texture information of CONT, HOM, ENG, and CORR in four directions (0°, 45°, 90°, and 135°) in pc1 images. Pearson correlation plot showed informative parameters and glycine content, positive in red and negative in blue. The darker the color, the greater the correlation. HOM and ENG were positively and CONT and CORR were negatively correlated with glycine (GLY) content (Fig. 11).
The different correlation coefficients were numerically similar, with the highest correlation between HOM and glycine content (0.47), indicating that there was a certain correlation between texture information and glycine content, which was worthy of further studies. PLSR, LSSVM, and SFCN models were established, based on the fusion of extracted texture information by PC1 image and optimal spectral values ( Table 5). The R P 2 values of the PLSR and LSSVM prediction models were increased by 0.04 and 0.05 respectively, compared with the optimal wavelength model, indicating that the linear and nonlinear relationship between glycine content and fusion information was enhanced to different degrees. The SFCN model established by the fusion of spectral information and textural information had the best performance (R P 2 = 0.9005, RMSEP = 0.3075, RPD P = 0.2688). Compared with the fused data of the other two modeling methods, the R P 2 was 0.0303 and 0.0666 higher, respectively, indicating that the SFCN method still had better predictive ability in the prediction model established after fusing the textural and spectral information of beef glycine. In summary, the fusion of textural and spectral information can describe beef quality information more effectively. This may be because the texture image of beef muscle reflected the arrangement characteristics or degree of density of muscle fibers, which was related to the amino acid content of beef.
At present, it was common for deep learning models to outperform traditional modeling methods. A combination of E-nose and HSI was used for the evaluation of volatile salt nitrogen in mutton, and a combination of image features and IMCNN was superior to spectral data modeling, with the root mean square error, correlation coefficient, and ratio of performance deviation of the prediction set of 3.039 mg/100 g, 0.920 and 3.59, respectively . When HSI was used to predict the hardness and pH value of Geofeng grape, the performance of CNN model was better than that of PLS-DA and SVM models, and the SAENN-LSSVM model had the best performance (R P 2 = 0.9232). The results showed that CNN was more effective than the traditional method and more suitable for fruit internal quality detection (Xu et al. 2021).

Visualization of Glycine Content Prediction
On the basis of optimal spectral information and textural information fusion modeling, specific model coefficients were applied to obtain the dot product of coefficients and spectra to obtain a single pixel in the image. The predicted value of pixels was obtained by displaying distribution, and visualization showed the distribution of glycine content in beef samples (Fig. 12). The color bar showed different concentration values, the concentrations of glycine were between 1.30 g/100 g and 4.97 g/100 g, and the color was not uniform, which may be caused by the uneven distribution of glycine. Some of the mutated blue and red areas may be caused by the failure to remove clean fine fascia and adipose tissue from the sample or by water loss during the experiment. It can clearly and intuitively express the glycine content in different areas of samples after pseudo-color treatment. Compared with traditional prediction methods, this can help consumers quickly and directly judge the quality of beef without damage. It can be said that these studies were effective and necessary.

Conclusions
In this study, on the basis of pre-processing, the traditional models and the SFCN model were built and their performances were compared. The results showed that the models based on the full spectrum band showed good results, indicating that the prediction model of glycine content in beef samples was feasible and effective. In addition, the UVE algorithm was used to extract 84 effective characteristic wavelengths, and the established UVE-SFCN model was the best in the characteristic band range (R P 2 = 0.8427), indicating that removing redundant data and retaining effective bands are helpful to the prediction performance of the model. Finally, the spectral data of the optimal characteristic wavelength was fused with the textural data extracted by GLCM to establish a prediction model. The prediction results of the UVE-SFCN-fusion model were ideal, the model R P 2 was 0.9005, RMSEP was 0.3075, and RPD P was 0.2688, which were 6.41% and 3.10% higher than the full-band model and the extracted characteristic wavelength, respectively. The fusion of textural and spectral information improved the performance of the model while obtaining better prediction results. The best fusion model was used to generate visual color images of glycine in beef samples, which provided a non-destructive and rapid detection method for the prediction of glycine content in beef.
Although the performance gains of building predictive models using hyperspectral techniques were limited, some progress has been made by combining different data fusion and deep learning algorithms. In the future, in order to develop more online platforms for meat quality monitoring by hyperspectral research, we will find the correlation of more meat quality indicators and begin to resolve the quantum mechanism of spectral characteristic peaks by density functional theory gradually.