Skip to content
BY 4.0 license Open Access Published by De Gruyter Open Access December 20, 2023

Nondestructive detection of potato starch content based on near-infrared hyperspectral imaging technology

  • Jingxiang Zhao , Panpan Peng and Jinping Wang EMAIL logo
From the journal Open Computer Science


The traditional method of determining potato starch content is not only time-consuming and labor-intensive, but also very aggressive and destructive, which also causes serious pollution to the environment. Therefore, it is necessary to study the fast, efficient, and environment-friendly detection technology. Although near-infrared technology can solve these problems well, it cannot detect potato starch because of its dot shape, invisibility, and other shortcomings. Hyperspectral imaging technology has a new technology of near-infrared, which can simultaneously detect surface defects and internal physical and chemical components. In this article, the method of nondestructive testing of potato starch using near-infrared hyperspectral technology was studied. In thisarticle, successive projection algorithm, random frog, and genetic algorithm were used to predict the content of potato starch. The experimental results in this article showed that in random frog, the root mean square error (RMSEC) of correction set and the root mean square error of prediction (RMSEP) model R C 2 and R P 2 have become 0.87 and 0.84, respectively, and RMSEC and RMSEP have become 0.33 and 0.30%, respectively. Therefore, the best method to select the characteristic wavelength of potato starch is the random frog algorithm.

1 Introduction

Potato, commonly known as yam egg, is a kind of crop with strong cold resistance, drought resistance, and barren resistance. So far, more than 150 countries and regions in the world are planting and producing potatoes. China’s potato production and area rank first in the world. In 2015, the Chinese government listed potatoes as the fourth major grain after rice, wheat, and corn, and vigorously developed potatoes as a staple food. In recent years, with the rapid development of computer technology, non-destructive testing technology based on image processing technology has been widely used in agricultural production. At present, there have been many research results on non-destructive testing of potato defects, but most of them are limited to specific environments and lighting conditions. For example, this method cannot meet the needs of modern potato processing enterprises in the test under sealed light conditions. On this basis, by combining the research results of various countries and absorbing the relevant experience and methods of predecessors, the nondestructive testing of potato starch using near-infrared hyperspectral technology is proposed.

In recent years, with the rapid development of China’s potato industry, non-destructive testing of potato starch has become a research hotspot. Banerjee et al. studied the method of non-destructive testing of potatoes using piezoelectric sensors. When evaluating the freshness of products, there were different research reports. However, the surface hardness was a good indicator, which was widely used in practice [1]. Bantadjan et al. [2] used two portable spectrometers to collect the short-wave near-infrared spectra of cassava root and cassava meat in the interaction mode at 720–1,050 and 850–1,150 nm, respectively. The results showed that there was no significant difference between the predicted value and the actual value obtained by the standard method under the 95% confidence interval [2]. Wang et al. joined hyperspectral imaging with chemometrics to anticipate potato starch content. Hyperspectral information was gathered from three testing focuses (top, umbilicus, and center locales), giving a technique to the fast nondestructive location of starch content in potatoes and a decent reason for potato quality checking and evaluating [3]. However, these nondestructive testing methods have certain drawbacks.

Near-infrared hyperspectral imaging technology is one of the main development directions of nondestructive testing and classification technology. Wang et al. made a quantitative analysis of potato flour in steamed bread by combining near-infrared spectroscopy with chemometric technology. The outcomes demonstrated the way that close infrared spectroscopy could be utilized for the fast location of potato flour [4]. Huang et al. made a quantitative prediction of potato flour in potato-wheat mixed flour by using near-infrared spectroscopy technology [5]. Zhang et al. expected to lay out an expectation model of starch content in rice husks for 87 different rice assortments in China by utilizing the hyperspectral imaging framework. The planning of starch content could be got by moving the quantitative model to every pixel. Through visual analysis of rice starch, its distribution in rice could be understood, to realize real-time monitoring of starch content [6]. Su et al. utilized partial least squares regression (PLSR), support vector machine relapse, nearby weighted fractional least squares relapse and back-proliferation fake brain organization to choose the trademark frequencies in unearthly sets I, II and III utilizing relapse coefficients and successive projection algorithm (SPA), separately. To work with modern applications, a bunch of trademark factors were chosen for all tubers from each ghostly district. At last, six gatherings of trademark frequencies chosen from unearthly sets I, II, and III were utilized to advance the model. The outcomes showed that hyperspectral innovation had extraordinary potential for constant estimation of tubers in the food business [7]. However, the accuracy and performance of these methods need to be improved.

During the growth, harvest, storage, and transportation of potatoes, there are various adverse reactions such as sprouting, green skin, disease spots, decay, and black heart. The combination of these defective potatoes with high-quality potatoes greatly affects the economic value of the whole potato. The innovation of this paper is the combination of near-infrared spectroscopy technology and nondestructive testing technology of potato starch. The spectral pretreatment, model construction, pattern recognition, and other processing methods of potato starch have been systematically and deeply analyzed.

2 Near infrared hyperspectral imaging technology

2.1 Hyperspectral imaging system

Potatoes have a high nutritional value. In addition to food, vegetables, feed, and other functions, they also have certain health functions. Potatoes are rich in nutrients that other crops, vegetables, fruits, and other crops cannot reach. They also have the effects of preventing cardiovascular disease, preventing atrophy of liver, kidney, and other tissues, strengthening stomach and spleen, preventing stomach disease, enteritis, constipation, and keeping the digestive tract lubricated [8].

It is exactly a direct result of the nourishment and numerous impacts of potatoes that they have been generally utilized in food, starch, feed, medication, and different fields. Although China is already the largest potato production and consumption in the world, the entire potato industry is still in a relatively extensive state from farmers’ planting to enterprise processing. The processing utilization rate and added value are very low. About 50% is limited to simple food and feed, and the overall level is generally behind other developed countries. The low level of potato production in China is mainly due to the uneven quality of potatoes and the quality grade of potatoes. At present, most of the manual screening technologies rely on manual work. However, due to subjective factors such as artificial vision, long-term observation leads to fatigue and fatigue, reducing work efficiency [9].

The Charge Coupled Device (CCD) camera is used to obtain the hyperspectral image, which is two-dimensional to the naked eye, but actually three-dimensional. The hyperspectral image obtained is the superposition of 512 dimensional images. Figure 1 is a three-dimensional schematic diagram of a hyperspectral image. The plane composed of X and Y axes represents the image pixel information. As shown in Figure 1, in a hyperspectral image, its three-dimensional information is not only the spectral information of a single wavelength, but also includes image information at various wavelengths. Hyperspectral imaging technology is a new technology combining spectral technology and imaging technology, which overcomes the poor detection effect caused by traditional computer vision technology relying only on images and solves the problems of random and one-sided information [10,11]. Therefore, the application of hyperspectral imaging technology can quickly, accurately, and real-time non-destructive detection of potato quality.

Figure 1 
                  Three-dimensional schematic diagram of hyperspectral image.
Figure 1

Three-dimensional schematic diagram of hyperspectral image.

Hyperspectral imaging equipment is a high-precision instrument. Before use, it must be fully tested to ensure its working state in normal operation, to avoid hyperspectral imaging caused by system failure does not meeting the test requirements. Because the electronically controlled mobile platform moves too fast, the image of the potato sample becomes blurred, so the exposure time of the CCD camera is set at 20 ms, and the motion speed of the workbench is adjusted to 1.25 mm/s. The potato sample is evenly spread on a round, moderately sized transparent plastic cover, and then placed on a transmission device to obtain a hyperspectral image [12].

Due to the characteristics of the CCD camera, a large amount of dark current and noise appear under long exposure [13]. Therefore, the black and white calibration test of the hyperspectral image acquisition system can effectively remove the impact of noise, external light, etc., on the image, and obtain the corrected image P.

(1) P = P 0 P q P v P q × 100 % .

In the formula, P v is an all-white calibration image (reflectivity is 100%). P q is an all-black calibration image (reflectivity is 0%). P 0 is the original hyperspectral image of the sample.

As indicated by the quality of the imaging gear, it tends to be separated into two classifications: one depends on channels, and the other depends on imagers. By using the filter slice imaging technology, 2D images are collected at various wavelengths to obtain high-resolution 3D images with less data and fast data processing speed. However, due to the incomplete information provided, it is difficult to find the characteristic wavelength. The scanning spectral method is used to push and scan the hyperspectral image, which not only has a large amount of data but also has a high accuracy. Therefore, the chemometric method can be used to obtain the characteristic wavelength [14]. In Table 1, the hyperspectral images collected by push broom and filter are compared.

Table 1

Comparison of push broom and filter plate acquisition light patterns

Characteristic Push broom (GPL Zolix) Tunable filter
1 Environmental effect Very small More
2 Spectral “purity” Very nice Not good enough
3 Luminous flux High Low
4 Measuring speed FAST < 5 s SLOW > 120 s
5 Field application Fit Unsuited

Hyperspectral imager is a kind of hyperspectral imager, imaging lens, computer, electronic translation platform, etc. [15], as shown in Figure 2.

Figure 2 
                  Composition of hyperspectral imager.
Figure 2

Composition of hyperspectral imager.

The core of hyperspectral imaging technology is data extraction, which is of great significance for subsequent research. If the original data cannot be extracted correctly, the results also have a large deviation [16]. In the accurate extraction of hyperspectral data, it is essential to select the region of interest (ROI).

ROI means that when analyzing hyperspectral images, the measured object must be analyzed to extract the object of interest from the hyperspectral image. The recognition between potatoes and transparent plastic containers containing potatoes is not high, and the threshold of potatoes is close to the background. Therefore, it is impossible to directly use image masks to obtain samples, and the reflection coefficient method is used. In the obtained hyperspectral image, the high-reflectivity sample image and low-reflectivity background image are selected, and the high-reflectivity sample image is divided by the low-reflectivity background image. After division, the image and background of the potatoes are significantly different. Finally, the threshold segmented sample image is used for masking to obtain the target image [17]. This article selects the whole sample in the contour as ROI for data extraction.

2.2 Characteristic wavelength extraction method

2.2.1 Continuous projection method

Under the background of multivariate correction, researchers put forward the concept of SPA and a method to project the wavelength to other wavelengths. It is contrasted with different frequencies, which makes the frequency with the biggest projection vector the chosen frequency. Then, the last trademark frequency is chosen by the changed model [18].

The basic principle of the SPA algorithm is: Supposing X o × l is the spectral matrix, the spectral matrix is k columns, and then, the number of variables to be selected is o:

(2) c 1 = x l , x k 1 = x k , k = 1 , 2 , .. . , l .

The matrix A o projected onto the orthogonal subspace c o is calculated as:

(3) A o = O c o ( c o ) T ( c o ) T c o .

The projection vector x k o + 1 is calculated as:

(4) x k o + 1 = A o x k o , k = 1 , 2 , .. . , l .

The maximum projection vector is determined, that is:

(5) k = arg max k = 1 , 2 , .. . , l x k o + 1 .

The vector for the projection operation for the next iteration is defined:

(6) c o + 1 = x k o + 1 .

Let o = o + 1 . If o < n , the second step is back.

2.3 Random leapfrog algorithm

Random frog is a very effective method. Its main function is to construct a model through a few iterations. Its operation process is similar to Monte Carlo’s reversible jump Markov chain. Markov bind is utilized to lay out a steady dissemination in the model space, to get the determination likelihood of every variable and select factors.

In the experiment, this method is used to select the wavelength; that is, the spectral data is the independent variable. It is assumed that X ( a × b ) is the sampling spectrum matrix. The sample number is a, and the variable is b. Y ( a × 1 ) is the sampling density matrix.

2.3.1 Genetic algorithm (GA)

GA uses the genetic iteration method to minimize the root mean square (RMSECV) of cross validation of the PLSR model constructed by selected variables through selection, exchange, mutation, and other operations. At the end of this step, each parameter is reordered and selected again to obtain more suitable parameters. Finally, the characteristic wavelength variable of the minimized PLSR model RMSECV is obtained.

The evolutionary algorithm is used to produce the wavelength parameters chosen by the PLSR model calibration, choose the feature wavelength of the hyperspectral image, and finish the prediction operation. The study’s relevant GA parameters are as follows: the original subset is set at 30, the penalty rate is set at 0.001, the variation probability is set at 0.01, the convergence rate is set at 50%, the window size is 5, and the fitness function is taken to be the RMSECV associated with the PLSR model. There are three more variables after the GA is finished.

2.4 Model performance evaluation indicators

At present, the commonly used mathematical model evaluation indicators are determination coefficient ( R 2 ), root mean square error (RMSE), prediction residual sum of squares, and discrimination accuracy. In general, the larger the R-value of the model is, the smaller the RMSE value is, and the prediction effect of the model is correspondingly improved. Ideally, its value R 2 is 1 and RMSE is 0.

In this article, the determination coefficient of correction set R C 2 , the determination coefficient of prediction set R P 2 , the root mean square error of correction set RMSEC, and the root mean square error of prediction set RMSEP are used to analyze potatoes.

In regression analysis, R 2 is a very common data and a good predictor. In this paper, the determination coefficient R 2 is used to express the correlation between the starch content of potato samples predicted by the model and the actual yield measured by the chemometric method. The closer the numerical value R 2 is to 1, the closer the predicted result of this model is to the actual value. The formula of R 2 is:

(7) R 2 = 1 o = 1 m ( y o y ˆ o ) 2 o = 1 m ( y o y ¯ o ) 2 .

In the formula, y o is the actual value and y ˆ o is the predicted value. y ¯ o is the average value of the actual sample.

In this article, RMSE is also called standard error, which reflects the difference between the starch content of potato samples and the actual yield. The lower the RMSE value is, the better the model can handle the tested sample. The value of RMSE is expressed by the following formula:

(8) RMSE = o = 1 m ( y o y ˆ o ) 2 m .

In the formula, y o is the actual value, and y ˆ o is the predicted value. m is the number of samples.

3 Physical and chemical experiments of potato starch

3.1 Sample acquisition and preparation

About 158 fresh potato samples were chosen for this investigation, and they were all bought from market A. Fresh appearance, hard solidity, good maturity, good potato type, no epidermal and mechanical damage, and no internal or exterior faults are required by the potato grade specification requirement. The surface of the fresh potato that was purchased was dried with dirt, cleaned, and removed, and then marked in order to prepare for the experiment. All 158 samples underwent morphological testing following the aforementioned procedure. Each sample’s equatorial diameter and height were tested with a vernier, each sample’s quality was weighed with a balance, and the results were recorded. All samples were maintained at 22°C at room temperature for 12 h in order to establish the hyperspectral image acquisition conditions and remove the impact of temperature and other factors on the experimental results. The samples of potatoes are displayed in Figure 3.

Figure 3 
                  Potato samples.
Figure 3

Potato samples.

3.2 Physical experiments

The following techniques are frequently used to measure the starch content of potatoes: enzyme hydrolysis, acid hydrolysis, light rotation, and ferin reagent-based techniques. To determine the amount of potato starch and make the necessary modifications, we use a random leapfrog method. The more the dry matter in the tuber, the bigger the percentage of it and water, and the more water content in the tuber, the closer the proportion is to pure water, according to the different principles of the proportion of pure water and starch. To determine the sample’s specific gravity, the fresh potato sample was weighed before immersing it in water, that is, 22°C and drying the surface moisture.

The random leapfrog algorithm was used to check the range of potato nondestructive starch used in near-infrared hyperspectral imaging, as shown in Table 2. Table 2 lists value range of potato starch under near-infrared hyperspectral imaging.

Table 2

Value range of potato starch under near-infrared hyperspectral imaging

Data set Number of samples Starch content (%)
Maximum value Minimum value Average value Standard deviation
Training sample 70 23.87 18.44 21.71 1.09
Verification sample 88 24.05 18.23 21.83 1.04

3.3 Chemical experiments on starch content

It is required to utilize a chemical method to determine the actual value of potato samples in order to serve as a benchmark while researching the nondestructive detection of potato starch content based on hyperspectral image technology. The acid hydrolysis method of national food starch is the accepted chemical analytical method for determining the amount of potato starch. The basic idea behind the acid hydrolysis method is that the sample reacts with the dissolved solution to remove fat and soluble sugars and then hydrolyzes starch in a solution of hydrochloric acid to form reducing monosaccharides, from which the starch content may be inferred.

4 Nondestructive test of potato starch content

4.1 Selection of hyperspectral characteristic wavelength based on SPA

When using SPA method to optimize the hyperspectral characteristic wavelength of the sample, the minimum value is 1 and the maximum value is 40. SPA technology is mainly developed to optimize multiple linear regression (MLR). The SPA method is used to model the characteristic wavelengths. In general, MLR is superior to PLSR.

The SPA method is used to select the full band spectral parameters of the sample, and the specific selection is shown in Figure 4. It can be seen from Figure 4(a) that when the total wavelength parameter is about 10, the corresponding minimum RMSECV value is obtained. Figure 4(b) is the distribution of 10 selected characteristic wavelength variables. It can be seen from Figure 4(b) that the characteristic wave of the starch content pattern is between 410–520 and 970 nm. Using 10 characteristic wavelength parameters, the corresponding MLR model is established for the prediction of correction set and verification set. According to the picture and formula calculation, R C 2 , R P 2 , RMSEC, and RMSEP of the model are 0.84, 0.80, 0.35, and 0.33% respectively.

Figure 4 
                  Characteristic wavelength of total starch content selected by SPA: (a) cross validation RMSE of selected variables; (b) characteristic wavelength distribution.
Figure 4

Characteristic wavelength of total starch content selected by SPA: (a) cross validation RMSE of selected variables; (b) characteristic wavelength distribution.

Figure 5 shows the data results predicted by MLR model, including the relevant starch content after SPA selection. Figure 5(a) is a correction set, while Figure 5(b) is a prediction set. According to Figure 5(a), the difference between the two is not very large, which indicates that the model prediction has a high accuracy.

Figure 5 
                  MLR model prediction for SPA wavelength selection: (a) correction set; (b) prediction set.
Figure 5

MLR model prediction for SPA wavelength selection: (a) correction set; (b) prediction set.

4.2 Selection of hyperspectral characteristic wavelength based on random frog

In this article, the characteristic wavelength of starch is extracted by using the method of random frog 20 times, and the average calculation result is obtained. Figure 6 shows the hyperspectral characteristic wave.

Figure 6 
                  Selection of medium starch sample variables based on Random frog algorithm: (a) variable selection probability; (b) wavelength distribution.
Figure 6

Selection of medium starch sample variables based on Random frog algorithm: (a) variable selection probability; (b) wavelength distribution.

Figure 6(a) shows the possibility of selecting all relevant parameters, while Figure 6(b) shows the layout of the wavelength parameters selected by the algorithm. As shown in Figure 6(a), it is not difficult to see that most of the sensitivity bands are concentrated around 550, 640, and 815 nm. The test sets the threshold value as 0.2 and finally selects 13 characteristic wavebands including 503, 510, 525, 530, 545, 578, 597, 616, 621, 623, 625, 692, and 815 nm. On this basis, the corresponding PLSR model is established using 13 characteristic wave parameters, and the prediction operation is realized in the correction set and verification set.

As shown in Figure 7, Figure 7(a) is a starch correction set, while Figure 7(b) is a starch prediction set. It can be seen that the values in the starch correction set are different, which indicates that the accuracy of the model prediction needs to be further improved. According to Figures 6 and 7, and Random frog algorithm, when the number of main factors is 7, the model is optimal. The R C 2 , R P 2 , RMSEC and RMSEP of the obtained model are 0.87, 0.84, 0.33, and 0.30% respectively.

Figure 7 
                  PLSR model estimates of total starch content: (a) starch correction set; (b) starch prediction set.
Figure 7

PLSR model estimates of total starch content: (a) starch correction set; (b) starch prediction set.

4.3 Selection of hyperspectral characteristic wavelength based on GA

The characteristic wavelength of the hyperspectral image is selected by GA, and the wavelength parameters selected by PLSR mode are established and predicted. In this article, the relevant parameters of GA are set as follows. The initial subset is 30. The penalty rate is 0.001. The mutation rate is 0.01. The convergence rate is 50%, and the window size is 5. Assuming that the genetic iteration number is 100, RMSECV related to the PLSR model is an adaptive function.

Figure 8 shows GA-based selection conditions. As shown in Figure 8(a), it is not difficult to see that when the value of RMSECV related to GA is the least, the total number of relevant parameters is 70. Therefore, this method finally selects 70 points as the characteristic wavelength. It is not difficult to see from Figure 8(b) that some wavebands fluctuate greatly, reaching the highest value of 37.5 at about 180.

Figure 8 
                  Selection of hyperspectral characteristic wavelength of total potato starch based on GA: (a) RMSECV changes with the number of selected variables; (b) variable selection frequency.
Figure 8

Selection of hyperspectral characteristic wavelength of total potato starch based on GA: (a) RMSECV changes with the number of selected variables; (b) variable selection frequency.

The arrangement of characteristic wavelengths obtained by this method is shown in Figure 9. It can be seen from Figure 9 that the preferred distribution range of characteristic wavelengths of starch samples selected by this method is 412–470, 503–577, 650–685, 720–818, and 895–980 nm. Using 70 characteristic wavelength parameters, the corresponding PLSR model is established to predict the information of the correction set. According to Figures 8 and 9 and GA calculation, the number of main factors is 13, and the number of R C 2 and R P 2 for check set and verification set is 0.84 and 0.82, respectively. The RMSEC and RMSEP for check set and verification set are 0.34 and 0.31%, respectively.

Figure 9 
                  Hyperspectral characteristic wavelength distribution obtained by GA algorithm.
Figure 9

Hyperspectral characteristic wavelength distribution obtained by GA algorithm.

Figure 10 shows the data predicted by the relevant PLSR model, which is the characteristic wavelength selected by GA. Figure 10(a) is the value of the test set, while Figure 10(b) is the estimated set value. It can be seen from Figure 10(a) that the values of the check set are inconsistent, and the values are basically concentrated between 13.5 and 16.15. However, the values of the prediction set are relatively scattered.

Figure 10 
                  Selecting the characteristic wavelength GA: (a) checkset value; (b) estimated set value.
Figure 10

Selecting the characteristic wavelength GA: (a) checkset value; (b) estimated set value.

4.4 Comparison of hyperspectral characteristic wavelength selection results of various methods

Modeling applications are where cross-validation is most frequently employed. The majority of the samples in the construction appearance book are removed for model construction, but a tiny portion of the samples are kept for forecasting with the newly developed model. The prediction error of this small number of samples is then calculated, and its squared sum is recorded. Referring to the cross-validation approach and combining it with reality will help you separate the sample set into the check set and the validation set. Table 3 displays the results of the division.

Table 3

Statistical results of starch content sample set

Sample classification Ingredients Numerical value
Training sample Sample number 70
Scope/% 10.22–18.63
Standard deviation 14.82
Sample number 1.05
Verification sample Sample number 88
Scope/% 10.88–19.02
Standard deviation 15.34
Sample number 1.07

Table 4 shows the prediction results of the hyperspectral characteristic wavelength selection and the correlation mode establishment of the characteristic wavelength of starch content using SPA, Random frog, and GA methods. It can be seen from Table 4 that these selection methods can simplify the model, but the change in the model has a great relationship with the selected wavelength. More feature parameters can be obtained through GA, while the total number of feature parameters obtained by the random frog method and SPA method is the least.

Table 4

Comparison of results and models using different methods

SPA Random-frog GA
Number of variables 10 13 70
Factor number 7 13
R C 2 0.84 0.87 0.84
R P 2 0.80 0.84 0.82
RMSEC 0.35% 0.33% 0.34%
RMSEP 0.33% 0.30% 0.31%

It can be seen from Table 4 that GA changes the total parameters of the whole wavelength to 70, and the performance of the corresponding model is worse than that of the full band model, which indicates that there is some noise in the wavelength parameters selected by this method. Although the SPA method only selects 10 parameters, it still has obvious shortcomings compared with other methods. The key is that in the process of extracting starch by the SPA method, most of the key parameters related to starch are ignored, which leads to the decline of the performance of the model. The Random frog method changed the wavelength parameters in the model to 13, but its performance is better. The values of RMSEC and RMSEP models are 0.87 and 0.84, respectively, and those of RMSEC and RMSEP models are 0.33 and 0.30%, respectively. Therefore, the best method to select the characteristic wavelength of starch is the random frog algorithm.

Using the SPA method and GA to study the selection of potato starch found that they all can improve the nondestructive detection of potato starch content, and the model established by these two methods has high accuracy for the prediction of potato starch content, but comparing them to the random frog algorithm; values in the starch correction set are different, This indicates that the prediction accuracy of both models needs to be further improved. The potato starch content detection model was constructed by using the random frog algorithm, regardless of the prediction accuracy, or is it better for feature extraction, For the majority of the key parameters associated with starch, the model performance has been improved.

5 Conclusions

In this article, GA, SPA, and random frog were used to select the characteristic wavelength of starch content, and the advantages and disadvantages of each method were compared. The results showed that GA, Random frog, and other methods could reduce the total number of parameters to a certain extent, thereby simplifying the model. However, the impact of different methods on the detection results is very different. The SPA method can significantly reduce the number of variables required for modeling and minimize the number of characteristic variables. The starch content test model established by the SPA method has a good effect, but the corresponding prediction results are poor. The effect is best when the characteristic wavelength of starch content is optimized by random frog. Random frog method was used to select the characteristic wavelength of starch content, which not only has reduced the total parameters of the model, but also has significantly improved the performance of the model. The quantitative analytical model of starch content’s characteristic wavelength selection method was investigated. Analysis was done on the impact of GAs and random leapfrog on starch content. According to the test findings, Random-frog can provide the best starch model, and it does it with fewer variables than other techniques of selection. Because of this, Random-frog is superior to the other five methods and may be used to choose the starch content’s high spectral characteristic wavelength. However, due to time constraints, there are some shortcomings in this paper. For example, only the internal starch content of potato varieties in one market was studied; the origin and variety can be expanded to further investigate the method of detecting starch content. The nondestructive examination of potato dry material and water content can also be included in its scope.

  1. Conflict of interest: There is no potential conflict of interest in our article, and all authors have seen the manuscript and approved it to submit to your journal. We confirm that the content of the manuscript has not been published or submitted for publication elsewhere.

  2. Data availability statement: The data that support the findings of this study are available from the corresponding author upon reasonable request.


[1] R. Banerjee, A. Pal, I. Ganguly, G. Bej, T. Sutradhar, T. Dey, et al., “Rapid and nondestructive assessment of freshness of potatoes using a piezo based sensor,” Int. J. Chem. Environ. Sci., vol. 1, no. 2, pp. 12–18, 2020.10.15864/ijcaes.1202Search in Google Scholar

[2] Y. Bantadjan, R. Rittiron, K. Malithong, and S. Narongwongwattana, “Establishment of an accurate starch content analysis system for fresh cassava roots using short-wavelength near infrared spectroscopy,” ACS Omega, vol. 5, no. 25, pp. 15468–15475, 2020.10.1021/acsomega.0c01598Search in Google Scholar PubMed PubMed Central

[3] F. Wang, C. Wang, S. Song, S. Xie, and F. Kang, “Study on starch content detection and visualization of potato based on hyperspectral imaging,” Food Sci. Nutr., vol. 9, no. 8, pp. 4420–4430, 2021.10.1002/fsn3.2415Search in Google Scholar PubMed PubMed Central

[4] H. Wang, D. Lv, N. Dong, S. Wang, and J. Liu, “Application of near-infrared spectroscopy for screening the potato flour content in Chinese steamed bread,” Food Sci. Biotechnol., vol. 28, no. 4, pp. 955–963, 2019.10.1007/s10068-018-00552-xSearch in Google Scholar PubMed PubMed Central

[5] Y. Huang, F. Xu, H. Hu, X. Dai, and H. Zhang, “Development of a predictive model to determine potato flour content in potato-wheat blended powders using near-infrared spectroscopy,” Int. J. Food Prop., vol. 21, no. 1, pp. 2030–2036, 2018.10.1080/10942912.2018.1502199Search in Google Scholar

[6] Z. Zhang, X. Yin, and C. Ma, “Development of simplified models for the nondestructive testing of rice with husk starch content using hyperspectral imaging technology,” Anal. Methods, vol. 11, no. 46, pp. 5910–5918, 2019.10.1039/C9AY01926JSearch in Google Scholar

[7] W. H. Su, S. Bakalis, and D. W. Sun, “Chemometric determination of time series moisture in both potato and sweet potato tubers during hot air and microwave drying using near/mid-infrared (NIR/MIR) hyperspectral techniques,” Dry. Technol., vol. 38, no. 5–6, pp. 806–823, 2020.10.1080/07373937.2019.1593192Search in Google Scholar

[8] G. Kim, H. Lee, I. Baek, B. K. Cho, and M. S. Kim, “Short-wave infrared hyperspectral imaging system for nondestructive evaluation of powdered food,” J. Biosyst. Eng., vol. 47, no. 2, pp. 223–232, 2022.10.1007/s42853-022-00141-1Search in Google Scholar

[9] V. Rasooli Sharabiani and A. Khorramifar, “Quantitative analysis of organic acids in potatoes using NIR spectroscopy with PLS and ANN methods,” J. Environ. Sci. Stud., vol. 7, no. 4, pp. 5541–5549, 2022.Search in Google Scholar

[10] L. Sainni, “Wheat biomass estimation based on neural network and hyperspectral vegetation index,” Acad. J. Environ. Biol., vol. 3, no. 1, pp. 56–64, 2022.10.38007/AJEB.2022.030106Search in Google Scholar

[11] Y. Sun, J. Wen, Z. Chen, S. Qiu, Y. Wang, E. Yin, et al., “Non-destructive and rapid method for monitoring fish freshness of grass carp based on printable colorimetric paper sensor in modified atmosphere packaging,” Food Anal. Methods, vol. 15, no. 3, pp. 792–802, 2022.10.1007/s12161-021-02158-2Search in Google Scholar

[12] H. Lee, “Artificial neural network based Raman spectroscopy system for water quality monitoring,” Water Pollut. Prev. Control. Proj., vol. 3, no. 3, pp. 28–36, 2022.10.38007/WPPCP.2022.030304Search in Google Scholar

[13] Y. He, X. Bai, Q. Xiao, F. Liu, L. Zhou, and C. Zhang, “Detection of adulteration in food based on nondestructive analysis techniques: A review,” Crit. Rev. Food Sci. Nutr., vol. 61, no. 14, pp. 2351–2371, 2021.10.1080/10408398.2020.1777526Search in Google Scholar PubMed

[14] Ů. Kisalaei, G. Shahgholi, A. Moeinfar, and A. Khorramifar, “Close infrared spectroscopy to estimate the sugar content of different potato cultivars,” J. Environ. Sci. Stud., vol. 7, no. 3, pp. 5121–5128, 2022.Search in Google Scholar

[15] L. Cui, X. Wang, Y. Xu, Y. Li, and M. Han, “Hyperspectral reflectance imaging for water content and firmness prediction of potatoes by optimum wavelengths,” J. Consum. Prot. Food Saf., vol. 17, no. 1, pp. 51–64, 2022.10.1007/s00003-021-01343-zSearch in Google Scholar

[16] X. Sun, K. Zhu, and J. Liu, “Nondestructive detection of reducing sugar of potato flours by near infrared spectroscopy and kernel partial least square algorithm,” J. Food Meas. Charact., vol. 13, no. 1, pp. 231–237, 2019.10.1007/s11694-018-9936-8Search in Google Scholar

[17] A. Bhih, “Distributed system design based on image processing technology and resource state synchronization method,” Distrib. Process. Syst., vol. 2, no. 4, pp. 28–35, 2021.10.38007/DPS.2021.020404Search in Google Scholar

[18] F. Tao, H. Yao, F. Zhu, Z. Hruska, Y. Liu, K. Rajasekaran, et al., “A rapid and nondestructive method for simultaneous determination of aflatoxigenic fungus and aflatoxin contamination on corn kernels,” J. Agric. Food Chem., vol. 67, no. 18, pp. 5230–5239, 2019.10.1021/acs.jafc.9b01044Search in Google Scholar PubMed

Received: 2022-12-02
Revised: 2023-04-27
Accepted: 2023-08-04
Published Online: 2023-12-20

© 2023 the author(s), published by De Gruyter

This work is licensed under the Creative Commons Attribution 4.0 International License.

Downloaded on 3.3.2024 from
Scroll to top button