Study on the estimation of β-glucan content in oats based on hyperspectral imaging system

Oats are also called brome and wild wheat in the Compendium of Materia Medica. At present, oats are mainly divided into hulled oats and naked oats. They are planted in 42 countries around the world, and their total output ranks fifth among food crops. my country mainly grows naked oats, accounting for more than 90% of the total oats. Qinghai, Inner Mongolia, Gansu and other places are the main planting areas. Oats are low-sugar, high-nutrition crops. The β-glucan rich in its dietary fiber has the effects of lowering cholesterol, lowering blood lipids, stabilizing blood sugar and preventing cardiovascular diseases. The monitoring methods for the β-glucan content in oats include Congo red detection method, enzyme detection method, modified enzyme detection method, etc. Although these methods have high accuracy, the detection speed is slow and the cost is high, which is not suitable for rapid detection of β-glucan content in oats.

Hyperspectral imaging technology can simultaneously obtain the spatial distribution and spectral information of each pixel of the detection object, extract the appearance information and internal quality information of the tested sample, and can realize rapid and non-destructive detection of samples. At the same time, it does not require complex pre-treatment and can perform multi-component detection at the same time. At present, the common ones include grating spectrometry, acousto-optic tunable filter spectrometry, prism spectrometry and aluminum film coating technology. Hyperspectral imaging technology has been widely studied and applied in the fields of agriculture and food. For example, Ma Brom et al. pointed out that hyperspectral imaging technology can obtain rich images and other information of fruits for analysis, and has broad prospects in the non-destructive detection of the comprehensive quality of fruits.

1 Materials and methods

1.1 Materials and instruments

Oats: grown around Xining, Qinghai, shelled, free of mold, ungerminated, intact, stored indoors at 20°C for future use.

A 400-1000nm hyperspectral camera was used, and the FS13 product of Hangzhou Caipu Technology Co., Ltd. can be used for related research. The spectral range is 400-1000nm, the wavelength resolution is better than 2.5nm, and up to 1200 spectral channels. The acquisition speed can reach 128FPS in the full spectrum, and up to 3300Hz after band selection (supports multi-region band selection).

1.2 Experimental method

1.2.1 Sample preparation

After repeated mixing of oats, the oats were taken out by quartering method and spread in a culture dish with a diameter of 10 cm and a thickness of 1 cm. The oats were kept covering the entire culture dish. The smaller the gaps between the oats, the better. A total of 200 samples were made, which were randomly divided into 150 modeling groups and 50 prediction groups and numbered.

1.2.2 Hyperspectral data acquisition

The spectral imaging data of all samples were collected using a hyperspectral imaging system, as shown in Figure 1. The selected hyperspectral acquisition system conditions were a camera height of 33 cm, an exposure time of 25 ms, and a platform movement speed of 27 mm/s.

1.2.3 Noise correction

The calibration images in all white and all black environments were scanned by standard whiteboard and blackboard, and the calibration images in all white and all black environments were obtained. Then, the corrected images were calculated. The noise effects caused by uneven light source strength and dark current of the instrument in different bands can be eliminated by calculation.

1.2.4 Data preprocessing

Since the test sample is a plane paved with oatmeal particles, the surface of the sample is uneven and has gaps. Therefore, it is necessary to maximize the effective information in the detected spectral data and remove the instrument noise, sample background and stray light, which have a greater impact on the subsequent analysis results. Through the experiment, a fixed area of 25×25 pixels was finally taken as the test area of the sample, and its mean was calculated. The test effects caused by surface unevenness were corrected by variable standardization, multivariate scattering correction, etc., and 20 bands before and after were removed within the wavelength range of the test.

2 Results and discussion

2.1 Determination and statistical analysis of β-glucan content in samples

The samples after hyperspectral data collection were crushed and then sieved through a 0.2 mm sieve to prepare the test samples. The content of β-glucan in oats was determined using NY/T 2006-2011 "Determination of β-glucan content in cereals and their products". The test results are shown in Table 1.

As can be seen from Table 1, the β-glucan content in oats of the modeling group and the prediction group has good consistency and uniformity.

2.2 Spectral characteristics of samples

The average spectral reflectance of 150 samples at each wavelength was calculated from the original spectral curve of 150 modeling group samples, and the average spectral curve of oats in the modeling group was obtained, as shown in Figure 2.

As can be seen from Figure 2, oats have obvious spectral absorption near the wavelengths of 640 and 840 mm, resulting in a trough in spectral reflectance.

2.3 Hyperspectral data preprocessing and extraction

The data was processed by parametric numerical analysis method of piecewise linear regression to reduce the dimension, and finally a wavelength of 640 nm was selected. At this wavelength, the recognition speed can be improved without reducing the detection accuracy. The scatter plot of the spectral reflectance obtained in this study and the β-glucan content value to be estimated is shown in Figure 3.

2.4 Establishment of prediction model based on BP neural network

Based on BP neural network, a prediction model for the content of B-glucan in oats is established. The schematic diagram of the model structure is shown in Figure 4. The data vector of the input layer is the characteristic spectrum information Pi obtained by piecewise linear regression. The transfer function of the hidden network layer in the middle is the tangent function logsig. The output layer is the content of β-glucan in oats a. There is no clear theoretical guidance for the selection of the hidden network layer. This experiment combines the sample situation to continuously try and optimize to finally determine the number of nodes in the hidden layer. The function used is logsig(Q+X), where Q is the weighted calculation weight of different characteristic spectra in the experiment, X is the residual error of the value during calculation, and the function of the training algorithm selects the Gauss-Newton method. The maximum number of training times is 1000 times, and the target error is 0.02%.

Through the training experiment of the model, the model and estimation results of the content of β-glucan in oats are obtained, and finally the expected output value of the model is close to the estimated output value, as shown in Figure 5, and the average absolute deviation is within 0.02%, realizing a quantitative model for estimating the content of β-glucan in oats by observing characteristic spectra.

2.5 Verification of model output results

The established BP neural network model was used to estimate the samples of 50 prediction groups. The smallest difference was 0 and the largest difference was 0.56. The overall accuracy was high and the error was small. See Table 2 for details.

As can be seen from Table 2, there is only one sample with a difference greater than 0.5 between the estimated value and the detected value, and 60% of the samples have a difference less than 0.2. By analyzing the data of the modeling group and the prediction group, the results are shown in Table 3. It can be found that the R of the modeling group is 0.97, the RMSE is 0.64, and the R of the prediction group is 0.98, the RMSE is 0.58, and the model budget performance is relatively accurate.

Figure 6 is a scatter plot of the estimated values obtained using the BP neural network model and the test values obtained using NY/T 2006-2011. It can be seen from Figure 6 that all points in the scatter plot are basically near the straight line, indicating that the β-glucan content in oats estimated by the model in the budget group is basically consistent with the value detected by NY/T 2006-2011. Therefore, it is feasible to predict the β-glucan content in oats using the established BP neural network model.

3 Conclusions

This paper uses a hyperspectral imaging system to scan the spectral reflectance of oats in the range of 400~1000mm, uses a piecewise linear regression analysis method to process the data, establishes a 3-layer BP neural network model, and predicts the β-glucan content in oats with satisfactory accuracy. This experiment provides a new method for rapid and non-destructive detection of β-glucan content in oats, provides a reference and basis for the subsequent construction of a multispectral imaging system, and has a broad prospect in oat quality monitoring.