A Model for Fat Content Detection in Walnuts Based on Near-Infrared Spectroscopy

Abstract

Near-infrared spectroscopy (NIR) is an efficient and accurate method for fat content detection in walnuts. ‘Wen 185’ walnut is grown in large quantities in southern Xinjiang, and its fat content is an important indicator for evaluating the intrinsic quality. The excessive pursuit of yield efficiency, combined with the neglect of quality, agricultural product safety and other factors, has led to the production of poor-quality walnuts. Moreover, research on predicting walnut kernel fat content based on near-infrared spectroscopy technology is rarely reported. Therefore, a technical framework for walnut kernel detection based on near-infrared spectroscopy and the technical standards for ‘Wen 185’ are urgently needed. After first optimizing the initial spectrum data using five preprocessing methods, we established separate prediction models for walnut kernel fat content based on either a back propagation neural network or a support vector regression (SVR) algorithm. The results show that the correction set and validation set coefficients of determination of the walnut kernel fat content prediction model using the back propagation neural network algorithm were 0.86 and 0.89, respectively, with root mean square errors of 1.56 and 1.58, and an RPD value of 2.57; the coefficients of determination for the calibration and validation sets of SVR were 0.90 and 0.83, respectively, with root mean square errors of 1.76 and 1.70, respectively, and an RPD value of 1.70. Thus, near-infrared spectroscopy can be used to establish a foundation for realizing the rapid detection of walnut kernel fat content.

1. Introduction

China is the largest producer of walnuts in the world, with a history of cultivation spanning more than 2000 years, and it possesses very rich walnut resources. Walnut (Juglans regia L.) is one of the most important woody oil crops in China, and walnut kernels are rich in fat (with linoleic acid being the main component), protein, carbohydrates and other components, representing high economic and nutritional value [1]. Xinjiang is one of the birthplaces of walnut cultivation in China. Moreover, the development of the walnut industry in other provinces and cities not only has unique advantages but also serves as a pillar for farmers to increase their income and prosperity, particularly in the southern border areas [2]. At present, there is significant focus on the quality and nutritional value of walnuts [3]. The fat content of walnuts is a critical indicator for evaluating their quality and is central to their nutritional value. Thus, testing the fat content of walnut kernels is essential for ensuring the healthy development of the walnut industry [4]. However, traditional detection methods have problems such as long analysis times, excessive use of organic reagents, high costs and low efficiency, and are thus unsuitable for high-volume testing and may not be compatible with modern production requirements. As a detection technology, near-infrared spectroscopy has the characteristics of fast speed, high efficiency, low analysis cost and simple operation [5,6]. Therefore, these features can effectively solve the problems of traditional methods and provide a feasible method to realize the rapid detection of walnut kernel fat content [7,8,9].

In recent years, near-infrared spectroscopy has been increasingly applied in nut quality testing. It was used by Yi J et al. [10] to detect and analyze the water, protein and fat content in 116 walnut powder samples. Their analysis demonstrated that near-infrared reflectance spectroscopy had good predictive ability for moisture, protein and fat content, indicating its reliability as a technique for predicting walnut composition. In 2019, Ma Wenqiang et al. [11,12] established an IPLS quantitative correction prediction model based on NIR for the simultaneous detection of protein components in walnut samples in different physical states. With the root mean square error and correlation coefficients of walnut whole kernel sample validation set to 0.021 and 0.913, they demonstrated that their method can be used to optimize model quality and achieve the rapid, nondestructive detection of walnut kernel fat content. Chengji Li et al. [13] investigated the application of near-infrared diffuse reflectance spectroscopy to fresh walnuts during storage, establishing a two-dimensional correlation visible–NIR kinetic model for the rapid detection of their moisture content at room temperature. Qingping Wang et al. [14] established a calibration model for crude fat and protein content in hickory based on NIR, and their results showed a correlation coefficient as high as 0.99 for crude fat and a coefficient of variation of 0.83. This demonstrated the good predictive performance of NIR, suggesting it can effectively detect crude fat and protein content in pecans.

Overall, walnut research based on NIR is similar to research using other techniques, although there are differences in the methods for data preprocessing, selection of spectral characteristic bands and inversion. Therefore, our study aims to explore the influence of different spectral pretreatment methods on the inversion results of ‘Wen 185’ walnut in southern Xinjiang, comparing the influence of different feature screening and inversion models on the inversion results and establishing both back propagation neural network and SVR prediction models for the fat content of ‘Wen 185’ walnut. Subsequent rapid characterization provides a relevant theoretical basis for rapidly determining the quality of ‘Wen 185’ walnut in southern Xinjiang.

2. Materials and Methods

2.1. Material

Samples were selected from the ‘Wen 185’ walnut forest farm in Wensu County, Aksu Prefecture, Xinjiang, and were obtained from 180 10-year-old trees in 9 different walnut orchards (3 each in high-yield management orchards, middle-class management orchards and low-yield management orchards), each with a row spacing of 4 m × 6 m. During the walnut ripening period (1 September 2021), 20 walnut fruits were collected from each tree for a total of 3600 walnuts. After harvesting, the green peels were removed and dried in a ventilated area. After drying, the walnuts were broken and shelled for collecting the kernels, and the walnuts were crushed for 3 min using a FW-80 high-speed universal crusher with even mixing. They were then packed in plastic bags, sealed and stored in a refrigerator at 4 °C until spectral scanning and the determination of the contents of each component were carried out.

2.2. Main Instruments

The instruments used in these tests were the Antaris II FT-NIR Fourier Near Infrared Spectrometer from Thermo Fisher, Waltham, MA, USA. SAE-D Active Automatic Fat Analyzer from Jinan Alva Instrument Ltd., FW-80 High-speed Universal Crusher from Shanghai Jinpeng Analytical Instrument Ltd., Shanghai, China. Model DHG-9240A Electrothermal Constant Temperature Blast Drying Oven from Shanghai Qixin Scientific Instrument Co., Ltd., Shanghai, China. Model BCD-216TX Refrigerator from Qingdao Haier Co., and BCD-216TX Refrigerator from Qingdao Haier Co., Qingdao, China.

2.2.1. Method for Acquisition of the Initial Spectrum of Walnut Kernels

Before the spectra were collected, the test samples were stored in the room where the instrument was located to ensure the consistency of the spectral acquisition conditions for all samples [15,16]. When collecting spectra, the instrument must be turned on and allowed to warm up for 60 min, and the walnut kernel spectrum was collected in the range of 10,000–4000 cm⁻¹. The crushed walnut kernels were placed in a quartz sample cup (diameter 30 mm, height 5 mm, wall thickness 1 mm) and filled until parallel to the cup mouth, and the sample surface was then pressed to ensure it was level after filling. After scanning each sample three times, 540 spectra were obtained, and the spectra of replicates were averaged to obtain the final spectra of 180 samples. After each measurement, the sample cup was rinsed with tap water, then washed with distilled water, and finally wiped clean with alcohol for the spectral acquisition of the next sample.

2.2.2. Method for Determining the Fat Content of Walnut Kernels

The fat content of the walnut kernel samples was determined using the SAE-D6 active automatic fat analyzer. First, six 1 g samples of walnuts (accurate to 0.001 g) were weighed, wrapped in filter paper and placed in filter paper cartridges (with each sample measured three times, allowing two samples to be processed in one day), and 70 mL of petroleum ether (boiling point: 30–60 °C) was then poured into the aluminum cup for extraction. The leaching time (480 min) and drenching time (40 min) were set according to the fat meter’s operating procedure, and the program set to run automatically. After running the program, the aluminum cup was transferred into a 105 °C oven, dried for 2 h and then removed for cooling and weighing. The fat content was calculated as follows:

$${\mathrm{X}}_1=\frac{{\mathrm{M}}_1-{\mathrm{M}}_2}{{\mathrm{M}}_0}\times 100$$

(1)

X₁ indicates the fat content in the specimen (g); M₁ indicates the quality of aluminum cup and walnut grease (g); M₂ denotes the constant weight (g) of the empty aluminum cup; M0 represents the mass of the weighed walnut sample (g); and 100 indicates the conversion factor. The fat content in walnut kernels can be calculated using this formula, and it can be seen from

Table 1. Statistical analysis of physicochemical values of walnut kernel fat.

Component	Maximum Value (%)	Minimum Value (%)	Mean Value (%)	Standard Deviation
Fat	74.76	47.40	63.22	3.38

that there are stark differences in the maximum and minimum values of fat content across samples, demonstrating strong differentiation and representativeness.

2.2.3. Spectral Preprocessing and Modeling Methods

The spectrometer used for sample measurement acquires information related not only to the components being analyzed but also various noise signals [17]. The sources of most of these noise signals are uncertain and may arise from environmental conditions (light, humidity, etc.), instrument operation, and other factors. These noise signals hinder the spectral analysis process and can introduce significant errors into the calibration model during construction, thereby affecting the prediction of the composition of the sample being measured [18], which in turn impacts the prediction results of subsequent models. Therefore, the main spectral pretreatment methods used to address noise in this study include multivariate scattering correction (MSC), standard normal variate (SNV), first derivative (FD), second derivative (SD), multivariate scattering correction (MSC) and first derivative (FD) combination algorithms [17]. Partial least squares (PLS), back propagation neural network, support vector regression (SVR) and other methods are mainly used in the quantitative analysis of spectra. BP neural network and SVR are currently the most widely used modeling methods in spectral quantitative analysis [19,20], and back propagation neural network and SVR are used to establish the mathematical models in our study. MATLAB R2017a software was utilized for code writing as well as for prediction, and the BP neural network was structured as a single hidden layer neural network of 129-7-1. Among them, the activation function of the hidden layer selects the trainlm function, the activation function of the output layer selects the tansig function, and the activation function of the output layer selects the logsig function. The parameters were set to a maximum iteration number of 100, an error threshold of 0.001, and a learning rate of 0.01. The kernel function of the SVR algorithm was chosen to be the RBF function, with a penalization factor c of 0.2 and a radial basis function g of 0.15. The key parameters of the CARS were as follows: fold (the number of cross-validated pairwise folds) was set to 5, num (the number of MC sampling times) was set to 100, and selectLV (the criterion for selecting the best Criteria for Variables) was set to 0.

2.2.4. Evaluation of Model Performance

The accuracy, precision, stability and ability to predict truth values of a model are the main aspects of assessing how well it works. Common criteria for model evaluation include determination coefficient (R²), root mean square error of calibration and residual prediction deviation (RPD). the RPD refers to the ratio of the standard deviation to the root mean square error, proving the predictive power of the model. In general, as a model’s prediction is considered to improve as its R² value becomes larger and its RMSE value smaller [21]. Under ideal conditions, the coefficient of determination has a value of 1, and the root mean square error of prediction is 0. In this paper, the walnut kernel analytical model was evaluated based on coefficient of determination (R²), residual prediction deviation (RPD), root mean square error of calibration (RMSEC) and root mean square error of prediction (RMSEP). When R² ≤ 0.5 and RPD ≤ 1.4, the model has poor predictive performance but can be used for qualitative prediction, and quantitative prediction is also possible; when 0.5 < R² < 0.8 and 1.4 < RPD < 2.0, the model has good predictive performance, allowing for rough quantitative prediction; when R² ≥ 0.8 and RPD ≥ 2.0, the model is effective for accurate quantitative prediction [22].The formulas for calculating R², RMSE and RPD are as follows [23,24,25]:

$${\mathrm{R}}^2=1-\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{p}}_{\mathrm{i}}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}$$

(2)

$$\mathrm{RMSE}=\sqrt{\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{p}}_{\mathrm{i}}\right)}^2}{\mathrm{n}}}$$

(3)

$$\mathrm{SD}=\sqrt{\frac{1}{\mathrm{n}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{p}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}$$

(4)

$$\mathrm{RPD}=\frac{\mathrm{SD}}{\mathrm{RMSE}}$$

(5)

2.2.5. Reject Abnormal Samples

During this trial, the Mahalanobis distance method was used to eliminate 5 existing outliers. Moreover, the presence of natural variability (extreme values) in the same data type impacted the effectiveness of modeling, so data on the remaining 174 samples were included. First, each sample measurement was calculated to the center of the sample, and then the count, mean μ and standard deviation σ of the sample measurements were determined. Next, we judged the threshold M with μ + 3σ. Then, we calculated the mean value μ and standard deviation σ of the sample measurements, using μ + 3σ as the threshold value M. Finally, if the Mahalanobis distance of the measured sample was greater than the threshold M, the sample was considered to be abnormal and was rejected. If the measured value of the sample was greater than the threshold M, the sample was considered to be abnormal and was rejected. Figure 1 illustrates the process of removing outliers, with the red line indicating the threshold. Values greater than or equal to the threshold are classified as outliers, while those below the threshold are considered acceptable.

2.2.6. Sample Division

The methods for sample division include the random division (RS), Kennard–Stone (K-S) and SPXY methods. When using SPXY to divide samples, both the spectral information variables and the physical and chemical values of walnut samples are taken into account [26], resulting in a more uniform sample range distribution, which can improve the calculation accuracy of the model. In this study, the SPXY method was used to divide the samples into the correction set and the prediction set according to a ratio of 3:1, resulting in 42 samples in the prediction set.

3. Analysis of the Results

3.1. Near-Infrared Spectroscopy Analysis

The image below shows the initial near-infrared spectrum of 180 groups of walnut samples. Each line in the figure represents the shape of the original absorption spectra of different samples. The absorption peaks corresponding to fat are mainly found in the 1070–1440 nm (9345–6944 cm⁻¹) and 1520–1680 nm (6578–5952 cm⁻¹) ranges [27]. It can be seen from Figure 2 that the spectra basically show the same overall trend; however, due to differences in reflectivity, it is difficult to distinguish the relationship between the spectrum and the content of walnut components with the naked eye. The are three distinct absorption peaks observed at 5360 cm⁻¹, 6155 cm⁻¹ and 7721 cm⁻¹. Although the initial absorption spectral shapes of different samples are basically the same for the different management levels, there are certain differences in the reflection intensity corresponding to each spectral line, which facilitates modeling.

3.2. Effects of Different Spectral Preprocessing Methods

Five pretreatment methods were used to process the spectra of 174 walnut samples, and the back propagation neural network fat content prediction model was established using the pretreated spectral data. The corresponding modeling results are shown in

Table 2. Modeling results for different preprocessing methods.

Modeling Methods	Preprocessing Methods	Wavelength Variable	Correction Set		Test Set		RPD
			R2	RMSEC	R2	RMSEP
BP neural network	MSC	86	0.66	2.40	0.67	2.30	1.43
	FD	99	0.89	1.34	0.82	2.23	2.69
	SD	129	0.86	1.56	0.89	1.58	2.57
	SNV	99	0.62	2.54	0.55	2.65	1.28
	MSC+FD	106	0.86	1.51	0.82	1.69	2.50
SVR	MSC	86	0.60	2.70	0.64	2.04	1.12
	FD	99	0.88	1.81	0.86	2.08	1.44
	SD	129	0.90	1.76	0.83	1.70	1.70
	SNV	99	0.64	2.53	0.65	2.20	1.23
	MSC+FD	106	0.88	1.72	0.84	1.44	1.81

. In contrast to other pretreatment methods, the calibration set coefficient R² values of the FD, SD and MSC + FD methods exceed 0.8, but the MSC + FD combination method has a poorer modeling effect than the SD and FD methods, as indicated by the smaller coefficient of determination for the correction and verification sets, and the larger RMSEC and RMSEP. The R² of the correction and verification sets is larger for SD than for FD, and the RMSEP is smaller, indicating that the model performance is better.

Filtering by Competitive Adaptive Reweighted Sampling (CARS) is a simple and effective way to quickly realize the optimal subset selection method [28]. Figure 3 shows the characteristic wavelength screening results for walnut kernel fat after applying the second derivative method. When performing variable filtering by CARS, the number of samples is set to 100. Figure 3a shows the trend of variable selection during band screening, and the number of variables decreases with the increase in the number of runs. Figure 3b shows the trend of RMSECV change obtained by cross-validation during the process of screening wavelength variables, and a decreasing RMSECV value indicates an increase in the effective information in the spectral variables. Conversely, an increase in RMSECV values indicates there is a decrease in the effective information in the spectral variables. RMSECV decreased when less than 38 samples were processed, and then began to increase, leaving 129 effective wavelength variables. Figure 3c shows the trend of regression coefficients for each wavelength variable as the number of runs increases, where “*” represents the minimum RMSECV at the corresponding location. Following the principle of the minimum RMSECV value, the subset of wavelength variables obtained at the 38th sampling is the optimal result. The number of characteristic wavelength variables screened at this point was 129, and the specific wave numbers are shown in

Table 3. Results of CARS characteristic wavelength screening based on SD combinations.

Pre-Processing Method	Wavelength Variable	Characteristic Wavelength/cm−1
SD + CARS	129	4065	4077	4308	4432	4636	4648	4651	4663	4675	4736	4991	5149	5257	5272	5280	5334
		5407	5577	5681	5704	5708	5758	5782	5824	5863	5866	5886	5994	6013	6052	6079	6090
		6160	6163	6171	6495	6557	6665	6750	6811	6969	6996	7182	7224	7239	7255	7351	7417
		7587	7617	7625	7629	7641	7652	7783	7853	7887	7918	7965	7980	8026	8107	8134	8219
		8227	8273	8312	8354	8373	8416	8431	8458	8501	8508	8539	8616	8647	8659	8713	8724
		8751	8782	8790	8832	8852	8867	8902	8906	8917	8929	8940	8952	8964	8987	9029	9033
		9064	9068	9072	9199	9276	9307	9318	9326	9345	9368	9411	9434	9476	9484	9507	9523
		9530	9538	9542	9631	9665	9669	9673	9696	9773	9777	9797	9808	9835	9866	9947	9955
									9962

.

3.3. Model Validation

Verification is required before a model is put into use, and the 42 samples of the validation sample set are used to test the established calibration model and evaluate the model stability and accuracy. Figure 4 and Figure 5 show the verification results for the back propagation neural network and SVR model, respectively. The verification results demonstrate that the model prediction results share a similar undulating trend to the real values, and the fat content in walnut kernels can be accurately predicted. Regarding the results for the prediction of fat, the validation set coefficient R² for the back propagation neural network was 0.89, with a root mean square error of 1.58 and RPD value of 2.86, and the validation set coefficient R² for the SVR model was 0.83, with a root mean square error of 1.70 and RPD value of 1.27. Therefore, it is concluded that both modeling methods are good predictors of walnut kernel fat content.

4. Discussion

In this study, the fat content in ‘Wen 185’ walnut samples from southern Xinjiang was predicted based on NIR. The results indicate that there is little difference between NIR and chemical detection methods for fat content, demonstrating that NIR can be utilized for the rapid detection of walnut quality, which is consistent with the conclusions of Ma Wenqiang et al. [12], Tang Wentao et al. [29] and Yi J et al. [10]. During spectral scanning, factors such as sample status, the number of repeats, light scattering and instrument response can cause interference, leading to baseline drift and the appearance of noise in the walnut kernel spectra [30]. To effectively reduce the influence of these factors, it is necessary to improve the signal-to-noise ratio of the spectrum [31] and thereby establish a more reliable model. Preprocessing is particularly critical in the modeling process for laying a solid foundation for the subsequent selection of characteristic bands and model construction. From our experimental results comparing the effects of five different spectral pretreatments, it can be seen that there is a better spectral modeling effect after SD treatment. This is inconsistent with the results of Zhang Hailiang et al. [32] and Song Xuejian et al. [33], possibly due to differences in the modeling methods utilized. The former used principal component regression (PCR) and partial least squares (PLS), whereas the latter used partial least squares (PLS), and in our study, the BP neural network and SVR methods were used to establish a predictive model. Feature extraction is a crucial step in the modeling process, and feature band filtering can be used to reduce the number of input parameters, shorten model operation time and partially eliminate redundant interference information to improve the accuracy of model prediction. In this paper, there were 1557 near-infrared spectral data points but only 174 samples, leading to increased collinearity in the modeling process. In addition, the near-infrared spectrum contained substantial amounts of redundant information, resulting in unsatisfactory modeling outcomes [34]. Thus, the CARS method was used to extract the characteristic wavelengths, and 129 effective wavelength variables for fat were finally extracted from 1557 data points. The back propagation neural network quantitative model and SVR model were established from the extracted 129 feature wavelengths, and the coefficient of determination R² of the verification set of both models exceeded 0.83, which is higher than that of the model established in the full spectrum. Wang Cuixiu et al. [35] used the CARS method to screen for characteristic bands, greatly reducing the number of wavelength variables from 1981 to less than 100, with a variable compression ratio of greater than 94.95%. Duan Lingyao et al. [36] used the DWT method to process spectral data for the rapid detection of proteins, fats and polysaccharides in privet based on NIR, compressing the initial 1050 spectral data points to 133, which greatly improved model performance. Zhu Jinyan et al. [37] used the synergy interval partial least squares (SiPLS) algorithm to screen for characteristic bands, namely, the wavelengths of soluble solids, vitamin C and anthocyanins from 1751 wavelength variables to 701, 351 and 701, respectively, and the SiPLS-ELM model demonstrated more accuracy than the model established in the full spectrum.

A back propagation neural network and SVR prediction models for walnut kernel fat content were established on the basis of spectral preprocessing and characteristic band screening, and the coefficient of determination of both models exceeded 0.83. However, this is slightly lower than the values in Li Lu et al. [38] and Wang Cuixiu et al. [35], and our inferior results may be attributed to the use of fewer sample varieties in this study. The coefficient of determination for the fat content prediction model established using the SVR method is lower than that of Ma Wenqiang et al. [12] because different numbers of wavelength variables were screened, with the former screening out six wavelength variables, and fewer feature bands were extracted. In our study, a higher number of 129 wavelength variables were extracted, resulting in lower model performance. Another possible reason is the different number of near-infrared spectrum scans collected by the two, 64 times vs. 32 in our study. Lu Hui et al. [39] used 126 rice samples to establish an NIR analysis model for fat content, obtaining a validation set coefficient of 0.83, which is slightly lower than the value of 0.89 for the BP neural network prediction model in this study; this may have led to the low model performance they observed, owing to the smaller number of samples. Although the NIR prediction model for fat content proposed in this study can meet the requirements for application, a limited number of samples were collected; thus, the representativeness of samples needs to be strengthened and further improved. In follow-up studies, our team should collect a large number of samples to enhance the sample representativeness and thereby establish a more reliable near-infrared spectroscopy model.

5. Summary

In this paper, ‘Wen 185’ walnut in Nanjiang was taken as the research object, and a BP neural network prediction model and SVR prediction model of walnut kernel fat content was established following five data pretreatment methods. We verified that the SD pretreatment method was the most suitable for the pretreatment of walnut kernel near-infrared spectrum information. After selecting the characteristic wavelengths using the CARS algorithm, 129 wavelength variables of fat were extracted, greatly reducing the computational complexity and improving model accuracy. Both models showed better prediction of fat content of ‘Wen 185’ walnuts in southern Xinjiang.