Development of Two Smart Acoustic Yam Quality Detection Devices Using a Machine Learning Approach

doi:10.21203/rs.3.rs-2194989/v1

Quality detection has been a major problem in the agriculture and food industries. This operation is mostly done by a subjective sensory method which is prone to high error and food destruction. Therefore, there is a need to apply artificial intelligence using a machine learning approach. This study developed two intelligent acoustic yam quality detection and classification devices using two sound-generating techniques. The software (multi-wave frequency generator) sound-generating technique generated sound from a laptop to a speaker inside a detecting chamber. This sound passes through the yam and was received on the opposite side by a microphone, into another laptop for analysis using visual analyzer software. The impact sound-generating technique used sound generated from a gentle impact of the yam on a flat surface placed inside the detection chamber. The sound produced was picked up by a microphone into a laptop for analysis. Acoustic properties considered were amplitude, frequency, sound velocity, wavelength, period and sound intensity. Discriminant analysis algorithm only was used in this first stage of the study to prove the applicability of machine learning. Three qualities (good, diseased damaged and insect-damaged) of two yam varieties (white and yellow yam) were tested. The device's performance of white yam was 79 % and 68.7 %, yellow yam was 82.3 % and 68.7 % for the software sound generation-technique and surface impact sound-generating technique, respectively. The study shows that the software sound-generating technique performed better in terms of overall yam quality detection and also proves the applicability of machine learning.

Artificial intelligence

Machine learning

Yam quality

Quality detection

Acoustic

Yam, a perennial tuber crop that is commonly cultivated in South America, Oceania, the Caribbean, Asia, and West Africa belongs to a family called Dioscoreaceae and a genus called Dioscorea (Kalloo and Bergh 1993). Saranraj et al. (2019) reported that out of the more than 800 species of yam documented only a few species are found to be edible by man. Archeological pieces of evidence had shown that the origins of different species of yam had been traced to Africa, Asia and America (Saranraj et al. 2019; Kennedy et al 2019; Nabeshima et al 2020). Yam tuber contains varying amounts (percentages) of carbohydrates, proteins and vitamins. A single tube of yam had been estimated to contain an average of about 494 kJ (118 kcal) of energy. Currently, yam is estimated to be cultivated in about 61 countries with a global production of about 73 million metric tonnes. Yam Market Report (2022) reported that the United States, India, Belgium, Japan and Jamaica in 2019 were the global major exporters of yam. These countries exported a total of 50,593.4 tonnes valued at $ 81,321,244 million. The world's largest producer of yam is Nigeria, accounting for over 70 to 76% of the world's production (Kalloo and Bergh 1993; Saranraj et al. 2019; Kennedy et al 2019; Nabeshima et al 2020; Yam Market Report. 2022). Though, most of the yams produced in Nigeria and other developing countries are mostly consumed locally. Therefore, the need to increase the export of this crop to generate a large foreign market to create jobs and foreign exchange to improve the world economy becomes necessary. To increase the world export market for yam exportation, the export quality of the product must be guaranteed to the importer. So, to achieve this, a non-destructive and intelligent method of detecting yam quality needs to be developed, using the engineering properties of the product to be detected.

One of the fastest ways to detect quality in an industry setting is by employing the knowledge of artificial intelligence through machine learning. According to Pathan et al (2020) the ability to make any machine behave and think like a human to achieve or solve a particular task is known as artificial intelligence. To achieve this intelligence in a machine, then the machine needs to be trained. Machine learning is a process of manipulating computers through programming with mathematical algorithms from data to imitate the way human reasons and make decisions (Swamynathan, 2017; Russell, 2018, Iwar et al., 2021a, Iwar et al., 2021b). In this study, an intelligent device was developed using a machine-learning algorithm to detect the quality of two species of yam using its acoustic properties. Though, only one machine learning algorithm was used in this study. The goal of this first stage of the project was to prove the applicability of the machine learning algorithm. After this initial study, more machine-learning algorithms will be used to evaluate the devices further. The use of acoustic properties of a material to detect material changes has been in existence since the 1800s.

According to Camastra and Vinciarelli (2015) and Knorr et al (2011), the use of acoustic properties of materials in research starts in the US in the 1830s. Though, these researches were not on food materials. Knorr et al (2011) stated that research on the acoustic properties of food begins in the 1920s. Many researchers had used the acoustic properties of crops to successfully detect moisture, defects, insect infestations and other activities affecting harvested crops (Figura and Teixeira 2007). This study reviewed the general application of acoustic properties of all frequency ranges, which had been applied to biomaterials to detect changes in nutrients, constituents or defections. This was done due to little or scarcity of information on the application of audible sound range being used in detecting post-harvest qualities of yam tubers. So, this review was done base on studies of different researchers on a particular crop. Friesen et al (1991) used sound pressure levels of dropped wheat grains to detect their moisture contents. This study developed moisture models with R² ranging from 0.4 to 0.9. De Belie et al (2000), Mengxing et al (2018), Liu et al (2021), Ekramirad et al (2021) and Zhao et al (2021), used acoustic properties of apples to automatically detect fruit firmness, codling moth larvae infestation, fruit firmness and Moldy core fungal infection, respectively. These studies achieved a detection accuracy range of 80 to 100%. Duprat et al (1997) carried out a study to ascertain if the firmness quality of apples and tomatoes fruits can be automatically detected using audible sound range frequency. The study shows a correlation value of 0.8 between the acoustic properties of apples and tomatoes with their fruit's firmness. Mizrach (2004) and Morrison and Abeyratne (2014) used ultrasonic sound frequency range to detect the firmness of plum and orange fruit respectively. Both studies established a strong relationship between acoustic properties and fruit firmness. Fleurat-Lessard (2006) and Guo et al (2019) used acoustic properties of wheat grains to automatically detect the presents of insects, larvae, damaged and undamaged grains. These studies achieved detection accuracy of 95 and 92% respectively. Siriwardena et al (2010) and Caladcada et al (2020) used audible sound properties obtained from the sound of coconut insects and knocking on coconut fruit to automatically detect the presence of insect pests and fruit maturity respectively. These studies obtained detection accuracies ranking of 83 to 97%. Lu et al (2009) used acoustic vibration properties to detect tomato fruit qualities (mass, density, pH, sugar content and firmness) during storage. The study developed a strong correlation between acoustic properties and tomato quality during storage. Pan et al. (2010) detect beans pod and seeds qualities such as skin damage, burst, split, broken and moisture content using audible sound frequency range. The study shows a high correlation between acoustic properties and beans pod and seeds qualities. Liu et al. (2015) study the relationship between the sensory crispness of a carrot with its acoustic properties at audible range frequency. The study shows that there is a strong correlation of 0.7 between the sensory crispness of the carrot to its acoustic property. Sisodiya and Singh (2016) used ultrasonic sound and infrared heat properties to automatically detect insects present in stored coconut, groundnut and mustard seeds, using GSM (Global System for mobiles) module to send detection information to farmers on their mobile phones. This study achieves a detection accuracy of 80%. Zhang et al 2014 and Zhang et al (2021) used the acoustic vibration properties of pear to automatically detect texture and core browning in pear fruits respectively. These studies achieved a detection accuracy range of 85 to 93%. Hussain and Saleh (2017) developed an intelligent monitoring system using CO₂ bust sound produced by white grubs (larvae of scarab beetles) that attach maize and soybeans seeds. Detection results are transmitted with GSM (Global System for mobiles) and GPS (Global positioning system) to the farmer’s phone. Zhang et al (2019) and Przybył et al (2020) used audible sound frequency range to detect the ripeness and texture of watermelon and strawberry respectively. This study achieved a detection accuracy of 92% for watermelon and a detection error of 0.09 for strawberries. The acoustic properties of yam tubers have never been used to detect their quality and that is what this study is aimed at achieving.

2.1 Design and Construction of two acoustic Yam quality detection devices

The two techniques used for developing the two yam quality devices were based on sound generation. These two sound generation techniques were:

i. Software sound generation technique

ii. Surface impact sound generation technique

2.1.1 Design of the yam detecting devices

The design information and calculations of the two intelligent yam quality detectors were displayed in Tables S1 and S2 displayed in the supplementary document provided. These design values were used for the construction of the two detecting devices.

2.1.2 Construction of the yam detecting devices

2.1.2.1 Software sound generation-technique device

A 459 mm × 464 mm × 397 mm double-layer Mild density fibreboard (MDF) was used to construct the body of the detection chamber as shown in Fig. 1. The inside of the detection chamber was covered with a 12.5 mm thick Styrofoam for soundproofing. Attached inside the chamber was 25 watts, 4 Ω subwoofer speaker to one end of the chamber. A multi-function Condenser Microphone having a 3.5 mm Jack for computer input was attached at the opposite end of the speaker in the chamber. A sample holder of dimension 192 mm × 140.5mm was placed between the speaker and the microphone. The speaker was then connected to a Hp laptop installed with software called multi-wave frequency generator version 1.2 for sound generation. The microphone was then connected to a Focusrite Solo Soundcard which in turn was connected to a Vibes Cor VC AMP 12 BT Home Amplifier. The Amplifier was then connected to a Hp laptop installed with software called visual analyzer 2011 for sound analysis and collection of data.

2.1.2.2 Sample impact sound generation-technique device

A 382.8 mm × 522 mm × 478.42 mm double-layer Mild density fibreboard (MDF) was used to construct the body of the detection chamber as shown in Fig. 2. The inside of the detection chamber was covered with a 12.5 mm thick Styrofoam for soundproofing. The detecting chamber consists of a sliding panel with an impact surface directly opposite it. Above the impact surface, was attached a multi-function condenser microphone having a 3.5 mm Jack for the computer input. The microphone was then connected to a Focusrite Solo Soundcard which in turn was connected to a Vibes Cor VC AMP 12 BT Home Amplifier. The Amplifier was then connected to a Hp laptop installed with software called visual analyzer 2011 for sound analysis and collection of data.

2.2 Evaluation of the developed yam quality detection devices

2.2.1 Sample acquisition and preparation

Six hundred tubers consisting of 300 white yams (Dioscorea rotundata) and 300 yellow yams (Dioscorea cayanensis) were acquired from National Root Crops Research Institute, Umudike, Nigeria.

These acquired tubers were taken to the agronomy laboratory at the federal university of Agriculture, Makurdi, Nigeria for inspection and verification. The quality of yam tubers acquired was one hundred (100) Good yams, one hundred (100) diseased yams and one hundred (100) insect-destroyed yams of white yams and yellow yams, respectively. Typical yam quality samples are displayed in Fig. 3.

2.2.2 Experimental procedures for acoustic data acquisition

2.2.2.1 Software sound generation-technique device

The Yam sample was placed on the sample holder and the detection chamber was closed, to make the chamber airtight. An audible sound wave of 200 Hz was generated using the multi-wave frequency generator version 1.2 installed on the laptop computer. This value (200 Hz) was used because it was recommended by the WHO report (2015) to be within the safe hearing frequency range. This sound wave was passed into the detecting chamber through the speaker. The sound passes through the yam sample and is then collected by the microphone into the sound card. The sound card then passes the sound wave to the amplifier which in turn passes it into the laptop computer into the visual analyzer 2011 software. Typical screen displays of this sound-generating and analyzing software were shown in Figs. 4a and b, respectively. Then acoustic properties of the sound waves like amplitude, frequency, velocity, wavelength, period and intensity were recorded for each yam quality sample examined. Recorded data are taken for statistical analysis using a machine learning algorithm called multiple discriminant analysis. The reason for using only one machine learning algorithm and only six acoustic properties at this first stage of the project was, to prove the applicability of the machine learning approach, using acoustic properties of yam quality. When this approach has been proven, more machine-learning algorithms and acoustic properties will be used in a future evaluation study to improve detection and classification accuracy.

2.2.2.2 Sample impact sound generation-technique device

The Yam sample was placed at the top of the sliding platform and then allows sliding until it hit the impact platform. Then, the sound wave produced by this impact was received by the microphone into the sound card. The sound card then passes the sound wave to the amplifier which in turn passes it into the laptop computer into the visual analyzer 2011 software. A typical reading display of this software was shown in Fig. 4b. Then acoustic properties of the sound waves like amplitude, frequency, velocity, wavelength, period and intensity were recorded for each yam quality sample examined. Recorded data are taken for statistical analysis using a machine learning algorithm called multiple discriminant analysis.

2.3 Data Analysis

Acoustic data collected from the two detecting devices were analyzed using SPSS software version 23. A machine learning algorithm called multiple linear discriminant analysis is used to correctly classify yam quality categories. Leave-one-out cross-validation analysis was used to validate the classification results obtained.

2.3.1 Discriminate analysis

The formula used by the discriminant algorithm for classification was displayed as Eq. 1

$P\left({C}_{i}|DSF\right)=\left[\frac{\left[P\left(DSF|{C}_{i}\right)P\left({C}_{i}\right)\right]}{?P\left(DSF|{C}_{i}\right)P\left({C}_{i}\right)}\right]$ 1

Where

$P\left({C}_{i}|DSF\right)$ =Posterior probability that a case is in yam quality class i, given that it has a specific discriminant score DSF

DSF = Discriminant Score for Function (Model)

$\left(DSF|{C}_{i}\right)$ = Conditional probability that a case has a discriminant score of DSF, given that it is in yam quality class i

$P\left({C}_{i}\right)$ = prior probability that a case is in group i, which would be equal to (n_i / N)

N = Total number of samples

n_i = Number of cases correctly classified

2.3.2 Classification and validation

The classification was done using the confusion matrix table. There were no hold-out samples before classification. Leave one cross-validation method was used to validate the developed classification models by reclassifying the samples by leaving one case out. To achieve cross-validation by predicting an output (Y) at point (X), given input data in matrix X₂ and output Y₂, Eq. 2 was used:

$output \left(Y\right)=\widehat{\mu } +R\left(X, {X}_{2}\right)R{\left({X}_{2}\right)}^{-1}({Y}_{2}-\mu {1}_{n})$ 2

Where

$\widehat{\mu }$ = overall mean, R = correlation matrix, µ = mean of samples under consideration, n = number of observations under consideration.

So, for leave-one-out predictions, the matrix X₂ will have all the design points except for the one we are predicting.

3.1 White yam (Discorea rotundata)

Results of the entire discriminant analysis for detection evaluation of white yam, for the developed devices on software generation-technique and surface impact generation-technique, are displayed in Table S3 and S4 (tables too large for manuscript) submitted as a supplementary document. These tables also show predicted white yam quality by the discriminant software using the discriminant score functions (equations). These score functions (equations) were used to determine the probability that the predicted quality belongs to either quality class 1 (good), class 2 (diseased damaged), or class 3 (insect-damaged). Table 1 shows the mean and standard deviation of acoustic properties obtained during the evaluation of white yam for the developed devices. The mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), intensity (db), period (s), velocity (m/s) and wavelength (m) were 270.621 ± 14.125, 200 ± 0, 4333 ± 21906, 5 ± 0, 25 ± 0, 0.125 ± 0 for good yams respectively using the software sound generation technique. Similarly, 147.573 ± 27.765, 155.714 ± 11.506, 1448 ± 247, 2.32 ± 0.49, 16 ± 0, 0.101 ± 0.013 respectively were values obtained for good white yam using the surface impact sound generation technique. These results show that the acoustic properties of good white yams were higher in values for the software sound generation-technique than for the surface impact sound generation technique. This phenomenon can be attributed to the sound generation frequency of the software sound generation-technique being constant; while that of the surface impact sound generation-technique was not fixed but depended on sample impact. Looking at the statistical behavior of the obtained results, the acoustic property having the highest deviation from the mean for both techniques is the intensity of the sound. This is because according to Figura and Teixeira (2007) the intensity of the sound measured through solid material depends on the area of the sample that sound must travel through. So, since these white yam samples were not of the same size, the amount of sound going through them will defer notwithstanding the technique used. The mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), intensity (db), period (s), velocity (m/s) and wavelength (m) were 307.764 ± 8.24, 200 ± 0, 2462 ± 65.916, 5 ± 0, 25 ± 0, and 0.125 ± 0 for diseased damaged white yams respectively using the software sound generation technique. Similarly, 72.907 ± 9.309, 83.54 ± 9.446, 800.876 ± 314.904, 1.01 ± 0.1, 8 ± 0 and 0.094 ± 0.012 respectively were values obtained for diseased damaged white yams using surface impact sound generation technique. These results show that the acoustic properties of diseased damaged white yams were higher for the software sound generation-technique than for the surface impact sound generation technique. This same explanation given for good white yam can also be used to explain this phenomenon. The acoustic property having the highest deviations from the mean for both techniques is the intensity of the sound for diseased damaged white yams. Again, the same explanation given for good white yams can also be used to explain this phenomenon in diseased damaged yams. The insect-damaged result mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), intensity (db), period (s), velocity (m/s) and wavelength (m) were 285.566 ± 11.599, 200 ± 0, 2290 ± 92.681, 5 ± 0, 25 ± 0 and 0.125 ± 0 for software sound generation technique. Similarly, 143.055 ± 9.576, 153.772 ± 9.415, 1369 ± 225.5, 2.24 ± 0.429, 16 ± 0 and 0.102 ± 0.009 respectively were values obtained for insect-damaged white yams using surface impact sound generation technique. The same explanation given for white good yams is applicable here as well. The descriptive statistic alone can not give us an in-depth explanation of how these acoustic properties affect the choice for detecting these white yam qualities. So, we take a look at the test for equality of variance, group means and covariance matrices for the two acoustic techniques.

Table 1

Descriptive statistic on the performance of two acoustic techniques used in detecting white yam quality
White Yam			Software Sound generation technique		Surface impact sound generation technique
Yam Quality	Acoustic Property	N	Mean	Std. Deviation	Mean	Std. Deviation
Good	Amplitude (db)	100	270.621	14.125	147.573	27.765
	Frequency (Hz)	100	200.000	0.000	155.714	11.506
	Intensity (db)	100	4333.156	21906	1448	247
	Period (s)	100	5.000	0.000	2.320	0.490
	Velocity (m/s)	100	25.000	0.000	16.000	0.000
	Wavelength (m)	100	0.125	0.000	0.101	0.013
Disease damaged	Amplitude (db)	100	307.764	8.240	72.907	9.309
	Frequency (Hz)	100	200.000	0.000	83.540	9.446
	Intensity (db)	100	2462	65.916	800.876	314.904
	Period (s)	100	5.000	0.000	1.010	0.100
	Velocity (m/s)	100	25.000	0.000	8.000	0.000
	Wavelength (m)	100	0.125	0.000	0.094	0.012
Insect Damaged	Amplitude (db)	100	285.566	11.599	143.055	9.576
	Frequency (Hz)	100	200.000	0.000	153.772	9.415
	Intensity (db)	100	2290	92.681	1369	225.561
	Period (s)	100	5.000	0.000	2.240	0.429
	Velocity (m/s)	100	25.000	0.000	16.000	0.000
	Wavelength (m)	100	0.125	0.000	0.102	0.009
Total	Amplitude (db)	300	287.984	19.150	121.178	38.558
	Frequency (Hz)	300	200.000	0.000	131.009	35.125
	Intensity (db)	300	3028	12639	1206	391.775
	Period (s)	300	5.000	0.000	1.857	0.710
	Velocity (m/s)	300	25.000	0.000	13.333	3.778
	Wavelength (m)	300	0.125	0.000	0.099	0.012
N is the experimental observation number

The test for equality of variance, group means and covariance matrices are statistical properties that show if the acoustic data obtained can be discriminated against using discriminant analysis. These tests are displayed in Table S5 submitted in the supplementary document. The first test on the table is the test of equality of yam quality groups means. This test is a univariate ANOVA test, using ‘Wilks' Lambda’ and ‘Fisher (F)’ values. Lower wilks’ lambda and F values mean that the variable had effects on the detection classification. In the software sound generation-technique, the test shows that among the five acoustic properties tested only amplitude and Intensity can be tested while the remaining variables (frequency, period, velocity and wavelength) can not be tested because they do not show any variation among it data sets. Now, between amplitude and intensity that were tested, only amplitude was found to have an effect while intensity does not have any effect on detecting white yam quality. This decision was taken because, the amplitude has a low wilks’ lambda value of 0.363 and f value of 4.24E-66, while Intensity has a higher wilks’ lambda value of 0.995 and f value of 4.49E-01. This means that in using the software sound generation-technique only one acoustic property or variable (amplitude) is required. Therefore, this makes the software-technique less expensive and an easy acoustic form of detecting yam quality. In the surface impact sound generation technique, the test of equality shows that among the five acoustic properties tested, variables like amplitude, frequency, period, wavelength and Intensity can be tested; while only sound velocity can not be tested because velocity data do not have many variations. In this impact technique, it was found that: amplitude, frequency, intensity, period and wavelength were affecting the detection of white yam quality. The decision was taken because their wilks’ lambda values were 0.211, 0.083, 0.456, 0.285 and 0.927 while their F values were 6.07E-101, 4.89E-161, 2.45E-51, 1.11E-81 and 1.21E-05 respectively. This means that the surface impact sound generation-technique requires five acoustic properties (variables) to detect white yam quality, unlike the software sound generation-technique that requires just one. Another statistical test conducted was the box's tests of equality of covariance matrices. It tests the homogeneity of variances and covariances within the white yam quality groups. So, to show that the covariance matrices are the same for data of all-white yam quality groups the log determinant was found. The equal log determinant shows equal covariance matrices among the quality groups. Log determinants for good, diseased damage and insect damage in the software sound generation-technique were 25.283, 3.781 and 12.435. These values are not equal, therefore, violating the assumption that the quality group covariance matrices must be equal. Nevertheless, ‘Box M; value can still be calculated to see if it has an effect. Non-significant 'Box M; value with an unequal log determinant still proves that the quality group covariance matrices are equal, this situation always occurs when data are large. The ‘Box M value calculated was 2956.271 with an F value of 487.918. This also violates the assumption that the quality group covariance matrices are equal but the analysis can continue because these can also occur when data are large. Log determinants for good, diseased damage and insect damage in the surface impact sound generation-technique were 9.486, 1.574, and 1.344. The calculated ‘Box M value was 1244 with an F value of 40.497. This also violates the assumption that the quality group covariance matrices are equal but the analysis can continue because these can also occur when data are large. Despite the drawback of the 'Box M' test, the test of the discriminant score functions will give a better insight into the discriminant analysis.

A summary of canonical discriminant score functions of white yam for both acoustic techniques is shown in table S6 (submitted in the supplementary document). The discriminant score functions are equations that will be used to determine the probability of a case belonging to a quality group. In this analysis, two discriminant score functions (equations) were developed for the software sound generation-technique and another two for the surface impact sound-generating technique. In the software sound generation technique, the first and the second discriminant score functions (equations) developed had Eigenvalue, % of the variance, cumulative % and canonical correlation of 1.766, 99.87, 99.87, 0.799 and 0.002, 0.13, 100, 0.048 respectively. The eigenvalue is the ratio between and within quality groups sum of squares and a higher value makes a discriminant score function (equation) a better predictor. So, for the software sound generation-technique, the first discriminant score function (Eq. 3) was chosen because it has a higher eigenvalue and higher canonical correlation value than the second function (Eq. 4). These equations (functions) are also displayed in Table S6 (submitted in the supplementary document). In the surface impact sound-generating technique the first and the second discriminant score functions (equations) developed had Eigenvalue, % of the variance, cumulative % and canonical correlation of 21.669, 99.91, 99.91, 0.978 and 0.02, 0.09, 100, 0.139 respectively. So, for the impact sound-generating technique the first discriminant score function (Eq. 5) was chosen because it has a higher eigenvalue and higher canonical correlation value than the second function (Eq. 6). These equations (functions) are also displayed in table S6 (submitted in the supplementary document). Wilks’ lambda test was also done to the discriminant score functions (equations) for both techniques. The test shows that for the software sound generation-technique the wilks' lambda value, chi-square value, degree of freedom and P-value for the two developed discriminant score functions (equations) were 0.361, 302.369, 4, 3.34E-64 and 0.998, 0.68, 1, 0.41 respectively. The first discriminant score function (Eq. 3) was again chosen over the second because it has a lower wilks' lambda value. Also, this same wilks’ lambda test for the surface impact sound-generating technique shows the wilks' lambda value, chi-square value, degree of freedom and P-value for the two developed discriminant score functions (equations) were 0.043, 926.42, 10, 1.31E-192 and 0.981, 5.722, 4, 0.221 respectively. The first discriminant score function (Eq. 5) was again chosen over the second (Eq. 6) because it has a lower wilks' lambda value. Another test called the standardized discriminant score function coefficients test” was done for functions (equations) developed for both techniques. This test shows the importance of the coefficients to the developed functions (equations). In the software sound generation technique, only the amplitude and intensity of sound variables were considered. The choices of these two variables were the same as explained for the test of equality of variance. The importance of the values of the coefficients for amplitude (db) and intensity (db) were 1, 0.042 and − 0.078, 0.998 for the first and second discriminant score functions (equations 3 and 4) developed respectively. These values show that only the coefficient of amplitude and intensity in the first developed function (Eq. 3), was important to it. In the surface impact sound-generating technique only the amplitude, frequency, intensity, period and wavelength variables were considered. The choices of these five variables were the same as explained for the test of equality of variance Their importance of the values of the coefficients for amplitude (db), frequency (Hz), intensity (db), period (s) and wavelength (m) were 0.042, 1.318, 0.094, -0.104, 0.948 and 0.266, -0.822, 1.152, 0.542, 0.368 for the first and second discriminant score function (Eqs. 4 and 5) developed respectively. These values show that for the first developed function (Eq. 5), only the coefficient of the period has negative importance while in the second discriminant score function (Eq. 6) developed, only the coefficient of frequency has negative importance ( see table S6 submitted in the supplementary document). These phenomena can be attributed to the difference in weights of the different yam samples used for the surface impact experiments. Also, a further statistical test was carried out on the functions (equations) developed. This further test was called the “structure matrix test” and was carried out for the function developed for both techniques. The structure matrix test tells the correction between the acoustic properties variable and the developed functions (equations). In the software sound generation technique, only the intensity and amplitude of sound variables were considered as explained by the equality of variance test, for the two developed functions (equations 3 and 4). The values of intensity (db) and amplitude (db) for the test obtained were − 0.042, 0.997 and 0.999, 0.078 respectively for the two developed functions (equations). This again shows that the intensity of the sound did not contribute to the prediction of yam quality. This phenomenon was caused by using a constant frequency of sound. In the surface impact sound-generating technique only the amplitude, frequency, intensity, period and wavelength variables were considered as explained by the equality of variance test, for the two developed functions (equations 5 and 6). The values of amplitude (db), frequency (Hz), intensity (db), period (s) and wavelength (m) for the test obtained were 0.713, 0.233, 0.415, 0.34, 0.06 and 0.154, 0.746, 0.509, 0.423, -0.265 respectively for the two developed functions (equations 5 and 6). This again shows the superiority of the first function (Eq. 5) over the second function (Eq. 6). The developed discriminant score functions (equations) for both techniques are displayed from equations 3–6 (also see table S6 in the supplementary document).

Software sound generation technique

DSF1 = 0.086 A − 6.149 I − 24.855 3

DSF2 = 0.004 A + 7.888 E-5 I − 1.284 4

Surface impact sound generation technique

DSF1 = 0.002 A + 0.130 F + 3.55E-4 I − 0.273 T + 82.855 W − 25.374 5

DSF2 = 0.015 A − 0.081 F + 0.004 I + 1.425 T + 32.153 W − 2.287 6

Where

DSF1 is the discriminant score function (equation) for the first developed function (equation)

DSF2 is the discriminant score function (equation) for the second developed function (equation)

A = amplitude (db), F = frequency (Hz), I = intensity (db), T = period (s) and W = wavelength (m)

The first discriminant score function (equations 3 and 5) was chosen in either technique for white yam quality detection classification

The detection classification or confusion matrix results for the discriminant analysis detection of white yam quality, for both acoustic techniques used in the study are displayed in Table 2. This table shows that two experimental classifications were carried out for both acoustic techniques used in the study. These detection classifications were called original classification and validated classification. In the software sound generation technique, the actual detection classification result shows that out of the 100 good white yams used for the detection. A total of 69% of them were correctly detected and classified as good yams, 31% were wrongly detected and classified as insect damage yams while no good yams were wrongly classified as diseased damaged yams. Still on the original classification, out of the 100 diseased damaged white yams used for the detection. 97% were correctly detected and classified as diseased damaged yams while 1% were wrongly detected and classified as good yams and 2% were wrongly detected and classified as insect-damaged yams. Again, still on the original classification, out of the 100 insect-damaged white yams used for the detection. A total of 73% were correctly detected and classified as insect-damaged yams. While 19% were wrongly detected and classified as good yams and 8% were wrongly detected and classified as diseased damaged yams. The overall original performance of correctly detecting the classification of all-white yam quality considered in this study using the software sound generation-technique for the original detection classification experiment was 79.7%. A cross-validation detection classification experiment was done again to confirm the original result. In the cross-validation detection and classifications done, out of the 100 good white yams used for the detection and classifications. A total of 68% of them were correctly detected and classified as good yams, 31% were wrongly detected and classified as insect damage yams and 1% were wrongly classified as diseased damaged yams. Still on the validation classification, out of the 100 diseased damaged white yams used for the detection. A total of 97% were correctly detected and classified as diseased damaged yams. In comparison, 1% were wrongly detected and classified as good yams and 2% were wrongly detected and classified as insect-damaged yams. Again, still on the validation classification, out of the 100 insect-damaged white yams used for the detection. A total of 73% were correctly detected and classified as insect-damaged yams. In comparison, 19% were wrongly detected and classified as good yams and 9% were wrongly detected and classified as diseased damaged yams. The overall cross-validated performance of correctly detecting all-white yam quality considered in this study using the software sound generation-technique was 79%. In the surface impact sound-generating technique, the actual detection classification result shows that out of the 100 good white yams used for the detection. A total of 37% of them were correctly detected and classified as good yams, 62% were wrongly detected and classified as insect damage yams and 1% were wrongly classified as diseased damaged yams. Still on the original classification, out of the 100 diseased damaged white yams used for the detection. A total of 100% was correctly detected and classified as diseased damaged yams while none was wrongly detected and classified as neither good yams nor insect-damaged yams. Again, still on the original classification, of the 100 insect-damaged white yams used for the detection. 74% were correctly detected and classified as insect-damaged yams while 26% were wrongly detected and classified as good yams and none were wrongly detected and classified as diseased damaged yams. The overall original performance classification of correct detections of all-white yam quality considered in this study using the surface impact sound-generating technique was 70.3%. Cross-validation detection was done again to confirm the original result. In the cross-validation detection and classifications done, out of the 100 good white yams used for the detection and classifications. A total of 35% were correctly detected and classified as good yams, 64% were wrongly detected and classified as insect damage yams and 1% were wrongly classified as diseased damaged yams. Still on the cross-validation classification, out of the 100 diseased damaged white yams used for the detection. A total of 100% was correctly detected and classified as diseased damaged yams while none was wrongly detected and classified as neither good yams nor insect-damaged yams. Again, still on the cross-validation classification, out of the 100 insect-damaged white yams used for the detection. A total of 71% were correctly detected and classified as insect-damaged yams. In comparison, 29% were wrongly detected and classified as good yams and none were wrongly detected and classified as diseased damaged yams. The overall cross-validated performance of correct detections of all-white yam quality considered in this study using the surface impact sound-generating technique was 68.7%. Comparing both acoustic techniques, the software sound generation-technique detected more good yams while the surface impact sound-generating technique detects more diseased damaged yams. Both acoustic techniques detect insect-damaged yams equally well. The reason for the high detection of good yams by the software sound generation-technique could be due to the lack of interference with the transmission of a sound wave through the yam tubers. Also, the uniform texture of the sound transmission material can account for the high detection of good yams. The poor detection of diseased and insect damage in the sound generation-technique could be due to differences in texture within the yam tubers. These differences in texture hinder the smooth transmission of sound through the yam tubers. The reason for the high detection of diseased yams than good yams in the surface impact sound-generating technique could be attributed to the difference in yam tuber tissue texture (deformation) as well. Tissue deformation could have softened the yam tubers; therefore making them produce low-frequency sounds upon impact with any solid surface than good yam tubers. Though, the overall performance of white yam quality detection was higher in the software sound generation technique. The surface impact sound-generating technique is the best to use when the goal of the detection was to detect diseased damaged white yams.

Table 2

Classification or confusion matrix results for the discriminant analysis detection of white yam quality for both acoustic techniques
Yam quality			Software Sound generation technique				Surface Impact sound-generating technique
			Predicted Group Membership			Total	Predicted Group Membership			Total
			Good	Disease damaged	Insect Damaged	Total	Good	Disease damaged	Insect Damaged	Total
Original	Count	Good	69	0	31	100	37	1	62	100
		Disease damaged	1	97	2	100	0	100	0	100
		Insect Damaged	19	8	73	100	26	0	74	100
	%	Good	69	0	31	100	37	1	62	100
		Disease damaged	1	97	2	100	0	100	0	100
		Insect Damaged	19	8	73	100	26	0	74	100
Cross-validated	Count	Good	68	1	31	100	35	1	64	100
		Disease damaged	1	97	2	100	0	100	0	100
		Insect Damaged	19	9	72	100	29	0	71	100
	%	Good	68	1	31	100	35	1	64	100
		Disease damaged	1	97	2	100	0	100	0	100
		Insect Damaged	19	9	72	100	29	0	71	100
Overall White yam quality detection			79.7% of original cases were correctly classified.				70.3% of original cases were correctly classified.
Overall White yam quality detection			79.0% of cross-validated cases were correctly classified.				68.7% of cross-validated cases were correctly classified.

3.2 Yellow yam (Dioscorea Cayanensis)

Results of discriminant analysis for detection evaluation of yellow yam for the developed devices for software generation-technique and surface impact generation-technique are displayed in Table S7 and S8 (tables too large for manuscript) submitted as a supplementary document. These tables also show predicted yellow yam quality by the discriminant software using either the first or second discriminant score function (equation). The discriminant score function then was used to determine the probability that the predicted quality belongs to either quality class 1 (good), class 2 (diseased damaged), or class 3 (insect-damaged). Table 3shows the mean and standard deviation of acoustic properties obtained during the evaluation of yellow yam for the developed devices. The mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), velocity (m/s), wavelength (m), period (s), and intensity (db) were 2907 ± 26188, 150.83 ± 7.841, 125666 ± 47947, 49839 ± 213523, 0.005 ± 0, 56447 ± 30494 for good yams respectively using the software sound generation technique. Similarly, 318.391 ± 1726.894, 0.006 ± 0.001, 166.78 ± 39.061, 1045 ± 245.458, 175420 ± 101794 and 45308 ± 221269 respectively were values obtained for good yellow yam using surface impact sound generation technique. These results show that the acoustic properties of good yellow yams were higher in values for the software sound generation-technique than for the surface impact sound generation technique. This phenomenon can be attributed to the sound generation frequency of the software sound generation-technique being fixed; while that of the surface impact sound generation-technique was not fixed but was based on sample impact. The acoustic property having the highest deviations from the mean for both techniques is the intensity of the sound. This is because the intensity of the sound measured depends on the area of the sample. So, since these yellow yam samples were not of the same size, the amount of sound going through them will defer notwithstanding the technique used. The mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), velocity (m/s), wavelength (m), period (s), and intensity (db) was 3429 ± 25953, 156.48 ± 9.215, 1067 ± 951, 152997 ± 22924, 0.005 ± 0, 486381 ± 1808198, for diseased damaged yams respectively using the software sound generation technique. Similarly, 134.399 ± 22.364, 0.006 ± 0.001, 151.81 ± 36.975, 5311 ± 21512, 153308 ± 76012 and 18469 ± 6472 respectively were values obtained for diseased damaged yams using the surface impact sound generation technique. These results show that the acoustic properties of diseased damaged yellow yams were higher for the software sound generation-technique than for the surface impact sound generation technique. This same explanation given for good yellow yam can also be used to explain this phenomenon. The acoustic property having the highest deviations from the mean for both techniques was the intensity of the sound for diseased damaged yellow yams. Again, the same explanation given for good yellow yam can also be used to explain this phenomenon. The mean and standard deviation values of acoustic properties like amplitude (db), frequency (Hz), velocity (m/s), wavelength (m), period (s), and intensity (db) were 296.266 ± 17.016, 175.677 ± 19.82, 12912 ± 117991, 208022 ± 164412, 0.005 ± 0, 178253 ± 934925, for insect-damaged yellow yams respectively using the software sound generation technique. Similarly, 1664 ± 15287, 0.014 ± 0.014, 87.88 ± 15.844, 4933 ± 43863, 146105 ± 649042, and 18523 ± 4896, respectively were values obtained for insect-damaged yellow yams using surface impact sound generation technique. The same explanation given for good yellow yams is applicable here. The descriptive statistic can not give us an in-depth explanation of how these acoustic properties affect the choice for detecting these yellow yam qualities. So, we take a look at the test for equality of variance, group means and covariance matrices for the two acoustic techniques.

Table 3

Descriptive statistic on the performance of two acoustic techniques used in detecting yellow yam quality
White Yam			Software Sound generation technique		Surface impact sound generation technique
Yam Quality	Acoustic Property	N	Mean	Std. Deviation	Mean	Std. Deviation
Good Yam	Amplitude (db)	100	2907	26188.	318.391	1726
	Frequency (Hz)	100	150.830	7.841	0.006	0.001
	Velocity (v/s)	100	125666	47947	166.780	39.061
	Wavelength (m)	100	49839	213523	1045	245.458
	Period (s)	100	0.005	0.000	175420	101794
	Intensity (db)	100	56447	30494	45308	221269
Insect damaged Yam	Amplitude (db)	100	3429	25953	134.399	22.364
	Frequency (Hz)	100	156.480	9.215	0.006	0.001
	Velocity (v/m)	100	1067	951.832	151.810	36.975
	Wavelength (m)	100	152997	22924	5311	21512
	Period (s)	100	0.005	0.000	153308	76012
	Intensity (db)	100	486381	1808198	18469	6472
Diseased damaged Yam	Amplitude (db)	100	296.266	17.016	1664.229	15287.966
	Frequency (Hz)	100	175.677	19.820	0.014	0.014
	Velocity (m/s)	100	12912	117991	87.880	15.844
	Wavelength (m)	100	208022	164412	4933.341	43863.637
	Period (s)	100	0.005	0.000	146105.648	649042.884
	Intensity (db)	100	178253	934925	18523.229	4896.031
Total	Amplitude (db)	300	2211	21260	705.673	8879
	Frequency (Hz)	300	160.996	17.088	0.009	0.009
	Velocity (m/s)	300	46549	92383	135.490	47.070
	Wavelength (m)	300	136953	168918	3763	28178
	Period (s)	300	0.005	0.000	158277	380762
	Intensity (db)	300	240360	1185384	27433	128035

The test for equality of variance, group means and covariance matrices are statistical properties that show if the acoustic data obtained can be discriminated against using discriminant analysis. These tests are displayed in Table S9 for yellow yam (table too large for manuscript) submitted as a supplementary document. The first test on the table is the test of equality of yam quality groups' means of yellow yam. This test is a univariate ANOVA test, using ‘Wilks' Lambda’ and ‘Fisher (F)’ values. Lower wilks’ lambda value and F value means that the variable is affecting the detection. In the software sound generation-technique for yellow yam, the test shows that all six acoustic properties can be tested, unlike the white yam test. So, for all the six acoustic properties tested frequency, intensity, period and wavelength were found to be affecting the detection with low wilks’ lambda values of 0.611, 0.629, 0.849, and 0.977 respectively. This means that only these four acoustic properties were used to decide the quality of yellow yam. This phenomenon occurs because unlike the white yam the frequency of the sound generated was not constant but varied slightly. This variation was responsible for more sound properties being involved in the determination of yellow quality. Acoustic properties like amplitude and velocity were found to not have any effect on the detection with high wilks’ lambda values of 0.996 and 1.000 respectively. So, these two acoustic properties do not play any part in deciding the quality properties of yellow yam. In the surface impact sound generation technique, the test shows that among the six acoustic properties tested, variables like amplitude, frequency, intensity and wavelength were found to be affecting the detection with low wilks’ lambda values of 0.994, 0.844, 0.47 and 0.99 respectively; while velocity and period were found to be no effect on the detection with high wilks’ lambda values of 0.995 and 0.999 respectively. This means that only acoustic properties like amplitude, frequency, intensity and wavelength are important in deciding the quality of yellow yam using the surface impact sound generation-technique while velocity and period were not. This could be attributed to the low sound produced by the yam impacts on the surface. The box's tests of equality of covariance matrices test the homogeneity of variances and covariances within the yellow yam quality groups. So, to show that the covariance matrices are the same for data of all yellow yam quality groups the log determinant was found. The equal log determinant shows equal covariance matrices among the quality groups. Log determinants for good, diseased damage and insect-damaged in the software sound generation-technique were 90.789, 86.235 and 86.494. These values are not equal therefore violating the assumption that the quality group covariance matrices must be equal. Nevertheless, ‘Box M; value can still be calculated to see if it has an effect. A non-significant 'Box M; value with an unequal log determinant still proves that the quality group covariance matrices are equal, this situation always occurs when data are large. ‘Box M value was 3425 with an F value of 111.43. This also violates the assumption that the quality group covariance matrices are equal but the analysis can continue because these can also occur when data are large. Log determinants for good, diseased damage and insect damage in the surface impact sound generation-technique were 60.103, 54.137 and 81.392. ‘Box M value is 5958 with an F value of 137.793. This also violates the assumption that the quality group covariance matrices are equal but the analysis can continue because these can also occur when data are large. Despite this drawback with the ‘Box M’ test, the test of the discriminant score functions will give a better insight into the discriminant analysis.

A summary of canonical discriminant score functions of yellow yam for both acoustic techniques is shown in table S10 (table too large for manuscript) and submitted as a supplementary document. The discriminant score functions (equations) are equations that will be used to determine the probability of a case belonging to a quality group. In this analysis, two discriminant score functions (equations) were developed for the software sound generation-technique and another two for the surface impact sound-generating technique. In the software sound generation technique, the first and the second discriminant score functions (equations) developed had Eigenvalue, % of the variance, cumulative % and canonical correlation of 1.134, 82.2, 82.2 0.73 and 0.246, 17.8, 100, 0.44 respectively. The eigenvalue is the ratio between and within quality groups sum of squares and a higher value makes a discriminant score function (equation) a better predictor. So, for the software sound generation-technique first, the discriminant score function (equation) was chosen because it has a higher eigenvalue and higher canonical correlation value than the second function (equation). These equations (functions) are displayed in table S10. In the surface impact sound-generating technique the first and the second discriminant score functions (equations) developed had Eigenvalue, % of the variance, cumulative % and canonical correlation of 1.32, 98.969, 98.969, 0.754 and 0.014, 1.031, 100, 0.116 respectively. The eigenvalue is the ratio between and within quality groups sum of squares and a higher value makes a discriminant score function (equation) a better predictor. So, for the impact sound-generating technique the first discriminant score function (equation) was chosen because it has a higher eigenvalue and higher canonical correlation value than the second function (equation). These equations (functions) are also displayed in table S10 (table too large for manuscript) submitted as a supplementary document. Wilks’ lambda test was also done to the discriminant score functions (equations) for both techniques. This test shows that for the software sound generation-technique the wilks' lambda value, chi-square value, degree of freedom and P-value for the two developed discriminant score functions (equations) were 0.38, 288.421, 10, 4.35E-56 and 0.8, 64.833, 4, 2.79E-13 respectively. The first discriminant score function (equation) was chosen over the second because it has a lower wilks' lambda value and is affect the detection. Also, this test for the surface impact sound-generating technique shows the wilks' lambda value, chi-square value, degree of freedom and P- value for the two developed discriminant score functions (equations) was 0.425, 251.814, 12, 5.72E-47 and 0.986, 4.023, 5, 0.546 respectively. The first discriminant score function (equation) was chosen over the second because it has a lower wilks' lambda value and is affecting the detection classification. The standardized discriminant score function coefficients test was done for functions (equations) developed for both techniques. This test shows the importance of the coefficients to the developed functions (equations). In the software sound generation-technique five acoustic properties, amplitude, frequency, velocity, wavelength and intensity were considered. Their importance of the coefficients values for were − 0.013, 0.701, -0.662, 0.247, 0.047and − 0.107, 0.646, 0.655, -0.008, -0.292 for the first and second discriminant score function (equation) developed respectively. These values show that only five of the coefficient of all-acoustic properties studied in the first developed function (equation) was important to it. In the surface impact sound-generating techniques all six acoustic properties studied which were amplitude (db), frequency (Hz)), velocity (m/s), wavelength (m), intensity (db) and period (s) were considered. These coefficients values were − 0.062, 0.938, -0.135, -0.052, 0.063, -0.34 and 0.184, 0.158, 0.134, -0.478, 0.658, 0.512 for the first and second discriminant score function (equation) developed respectively. These values show that for the first developed function (equation) some of the values of the coefficients have negative importance while in the second discriminant score function (equation) developed only the coefficient of wavelength has negative importance (table S10) (tables too large for manuscript) submitted as a supplementary document. These phenomena can be attributed to the difference in weights of the different yam samples used for the surface impact experiments. A structure matrix test was also carried out for the function developed for both techniques. The structure matrix test tells the correction between the acoustic properties variable and the developed functions (equations). In the software sound generation technique, only five of the six acoustic properties were important to the two functions (equations) developed. In the surface impact sound-generating technique, all six acoustic properties were important to the two functions (equations) developed. The developed discriminant score functions (equations) for both techniques are displayed in equations 7–10 (see table S10 in the supplementary document).

Software sound generation technique

DSF1 = − 6.038E-07 A + 5.227E-02 F − 8.996E-06 V + 1.584E-06 W + 3.962E-08 I − 8.222 7

DSF2 = -5.045E-06 A + 4.817E-02 F + 8.912E-06 V − 4.866E-08 W − 2.484E-07 I − 8.093 8

Surface impact sound generation technique

DSF1 = − 7.034E-06 A − 4.033E + 01 T + 2.898E-02 F − 1.838E-06 W − 3.546E-07 V + 4.931E-07 I − 3.517 9

DSF2 = 2.072E-05 A + 6.062E + 01 T + 4.890E-03 F − 1.695E-05 W + 3.510E-07 V + 5.146E-06 I − 1.345E 10

Where

DSF1 is the discriminant score function (equation) for the first developed function (equation)

DSF2 is the discriminant score function (equation) for the second developed function (equation)

A = amplitude, F = frequency, I = intensity, T = period and W = wavelength, V = velocity

The first discriminant score function (equation) was chosen in either technique for yellow yam quality detection classification.

The detection classification or confusion matrix result for the discriminant analysis detection of yellow yam quality, for both acoustic techniques used in this study is displayed in Table 4. This table shows that two experimental classifications were carried out for both acoustic techniques used in this study. These detection classifications were called original classification and validated classification. In the software sound generation technique, the actual detection classification result shows that of the 100 good yellow yams used for the detection. 87% of them were correctly detected and classified as yellow good yams, 10% were wrongly detected and classified as diseased damaged yellow yams and 3% were wrongly classified as insect-damaged yellow yams. Still on the original classification, of the 100 diseased damaged yellow yams used for the detection. 75% were correctly detected and classified as diseased damaged yellow yams. In comparison, none was wrongly detected and classified as good yellow yam and 25% were wrongly detected and classified as insect-damaged yellow yams. Again, still on the original classification, of the 100 insect-damaged yellow yams used for the detection. 87% were correctly detected and classified as insect-damaged yellow yams while 2% were wrongly detected and classified as good yellow yams and 11% were wrongly detected and classified as diseased damaged yams. The overall original classification performance of correct detections of all yellow yam quality considered in this study using the software sound generation-technique for the original detection classification experiment was 83%. A validation detection experiment classification was done again to confirm the original result. In the validation detection and classifications experiment, of the 100 good yellow yams used for the detection and classifications. 86% were correctly detected and classified as good yellow yams, 11% were wrongly detected and classified as diseased damaged yellow yams and 3% were wrongly classified as insect-damaged yellow yams. Still on the validation classification, of the 100 diseased damaged yellow yams used for the detection. 74% were correctly detected and classified as diseased damaged yellow yams. In comparison, 1% were wrongly detected and classified as good yellow yams and 25% were wrongly detected and classified as insect-damaged yellow yams. Again, still on the validation classification, of the 100 insect-damaged yellow yams used for the detection. 87% were correctly detected and classified as insect-damaged yellow yams. In comparison, 2% were wrongly detected and classified as good yellow yams and 11% were wrongly detected and classified as diseased damaged yellow yams. The overall validated performance of correct detections and classification of all yellow yam quality considered in this study using the software sound generation-technique was 82.3%. In the surface impact sound-generating technique, the actual detection classification result shows that of the 100 good yellow yams used for the detection. 47% of them were correctly detected and classified as good yellow yams, 48% were wrongly detected and classified as diseased damaged yellow yams and 5% were wrongly classified as insect-damaged yams. Still on the original classification, of the 100 diseased damaged yellow yams used for the detection. 61% were correctly detected and classified as diseased damaged yellow yams, 31% were wrongly detected and classified as good yellow yams and 8% were wrongly classified as insect-damaged yams. Again, still on the original classification, of the 100 insect-damaged yellow yams used for the detection. 100% were correctly detected and classified as insect-damaged yams while none were wrongly detected and classified as neither good yams nor diseased damaged yellow yams. The overall original performance classification of correct detections of all yellow yam quality considered in this study using the surface impact sound-generating technique was 69.3%. A validation detection experiment was done again to confirm the original result. In the validation detection and classifications experiment, of the 100 good yellow yams used for the detection and classifications. 47% were correctly detected and classified as good yellow yams, 48% were wrongly detected and classified as diseased damaged yellow yams and 5% were wrongly classified as insect-damaged yams. Still on the validation classification, of the 100 diseased damaged yellow yams used for the detection. 61% were correctly detected and classified as diseased damaged yellow yams, 31% were wrongly detected and classified as good yellow yams and 8% were wrongly classified as insect-damaged yams. Again, still on the validation classification, of the 100 insect-damaged yellow yams used for the detection. 98% were correctly detected and classified as insect-damaged yams while 1% were wrongly detected and classified as good yellow yams and 1% were wrongly detected and classified as diseased damaged yellow yams. The overall validated performance of classification of correct detections of all yellow yam quality considered in this study using the surface impact sound-generating technique was 68.7%. Comparing both acoustic techniques, the software sound generation-technique detected more good yellow yams while the surface impact sound-generating technique detects more insect-damaged yellow yams. Both acoustic techniques detect diseased damaged yams equal well. Though, the overall performance of yellow yam quality detection was higher in the software sound generation technique. The surface impact sound-generating technique is the best to use when the goal of the detection was to detect insect damage to yellow yams.

Table 4

Classification or confusion matrix results for the discriminant analysis detection of yellow yam quality for both acoustic techniques

Yam quality			Software Sound generation technique				Surface Impact sound-generating technique
			Predicted Group Membership			Total	Predicted Group Membership			Total
			Good	Disease damaged	Insect Damaged	Total	Good	Disease damaged	Insect Damaged	Total
Original	Count	Good	87	10	3	100	47	48	5	100
		Disease damaged	0	75	25	100	31	61	8	100
		Insect Damaged	2	11	87	100	0	0	100	100
	%	Good	87	10	3	100	47	48	5	100
		Disease damaged	0	75	25	100	31	61	8	100
		Insect Damaged	2	11	87	100	0	0	100	100
Cross-validated	Count	Good	86	11	3	100	47	48	5	100
		Disease damaged	1	74	25	100	31	61	8	100
		Insect Damaged	2	11	87	100	1	1	98	100
	%	Good	86	11	3	100	47	48	5	100
		Disease damaged	1	74	25	100	31	61	8	100
		Insect Damaged	2	11	87	100	1	1	98	100
Overall White yam quality detection			83.0 % of original cases were correctly classified.				69.3 % of original cases were correctly classified.
Overall White yam quality detection			82.3 % of cross-validated cases were correctly classified.				68.7 % of cross-validated cases were correctly classified.

This study developed two intelligent acoustic yam quality detection and classification devices. One uses a software sound generation-technique and the other uses a surface impact sound-generating technique. The machine learning algorithm used in the study called multiply discriminant analysis shows that fewer acoustic (sound) properties are needed for detecting yam quality in the software sound generation-technique than in the surface impact sound-generating technique. Three qualities (good, diseased damaged and insect-damaged) of two yam varieties (white and yellow yam) were tested. The Overall validated detection and classification performance of white yam were 79% and 68.7% for the software sound generation-technique and surface impact sound-generating technique respectively. The overall validated detection and classification performance of yellow yam was 82.3% and 68.7% for the software sound generation-technique and surface impact sound-generating technique respectively. The study shows that the software sound generation-technique performed better than the surface impact sound-generating technique in terms of overall yam quality detection and classification. Nevertheless, the surface impact sound-generating technique detects and classified damaged yams better while the software sound generation-technique detected and classifies good quality yams better. Therefore, the choice of the technique to be recommended for yam quality detection will depend on the goal or final expected outcome of the yam processor. These two quality detecting technologies can also be incorporated into a yam planting machine, yam export quality detection line, yam storage warehouses and storage facilities, etc.

Funding: This study was not funded by any external organization. The study was a contributory research work, self funded by the authors and their students across various Nigeria tertiary institutions.

Authors contributions:

Audu J.: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Writing - Original Draft, Visualization.

Alonge, A. F: Project administration, financial contributor, Conceptualization, Methodology

Dinrifo R. R.: Formal analysis, Investigation, Resources, Software, financial contributor

Adegbenjo A: Writing - Review & Editing, Supervision, Financial contributor

Anyebe, S. P.: Financial contributor, Formal analysis, Investigation.

Conflict of Interest: There is no conflict of interest what so ever among the authors of this study

Ethics approval: Not applicable

Consent to participate: Not applicable

Consent for publication: Not applicable

Availability of data and material included: The authors confirm that the data supporting the findings of this study are available within the article [and/or] its supplementary materials.

Code availability: Not applicable

Caladcada J. A., Cabahugb S., Catamcoa M. R., Villaceranb P. E., Cosgafaa L., Cabizaresb K. N., Hermosillaa M., Piedadc E. 2020. Determining philippine coconut maturity level using machine learning algorithms based on the acoustic signal. Computers and electronics in agriculture 172; 105327. https://doi.org/10.1016/j.compag.2020.105327
Camastra F. and Vinciarelli A. 2015. Machine learning for audio, image and video analysis: theory and applications. Second edition. ISBN 978-1-4471-6735-8. DOI 10.1007/978-1-4471 6735-8. Springer-verlag. London.
De Belie N., S. Schotte, P. Coucke, J. De Baerdemaeker. 2000. Development of an automated monitoring device to quantify changes in firmness of apples during storage. Postharvest biology and technology, 18; 1–8. PII: S0925-5214(99)00063-0. www.elsevier.com:locate:postharvbio
Duprat F . , Grotte M . Pietri E., Loonis, D . and Studman, C.J. 1997. The Acoustic impulse response method for measuring the overall firmness of fruit. J. Agric. Engng Res, 66, 4;251 –259. ISSN 0021-8634. https://doi.org/10.1006/jaer.1996.0143. https://www.sciencedirect.com/science/article/pii/S0021863496901432
Ekramirad N., Khaled A. Y. , Parrish C. A. , Donohue K. D. ,Villanueva R. T., Adedeji A. A., 2021. Development of pattern recognition and classification models for the detection of vibroacoustic emissions from codling moth-infested apples. Postharvest biology and technology, 181; 111633. https://doi.org/10.1016/j.postharvbio.2021.111633
Figura, L. O. and Teixeira, A. A. 2007. Food Physics: Physical properties – measurement and applications. ISBN 978-3-540-34191-8 Springer, Berlin, Heidelberg, New york. doi:10.1007/b107120.
Fleurat-Lessard F., Tomasini B., Kostine L. and Fuzeau B. 2006. Acoustic detection and automatic identification of insect stages activity in grain bulks by noise spectra processing through classification algorithms. 9th international working conference on stored product protection. Pp.476-486.
Friesen T. L., Brusewitz G. H. and Lowery R. L. 1998. An acoustic method of measuring moisture content in grain. J. Agric. Engng Res, 39; 49-56
Guo M., Yuting M., Xiaojing Y. and Richard W. M. 2019. Detection of damaged wheat kernels using an impact acoustic signal processing technique based on Gaussian modeling and an improved extreme learning machine algorithm. Biosystem engineering, 184; 37 – 44. https://doi.org/10.1016/j.biosystemseng.2019.04.022
https://bettersoundproofing.com/drywall-osb-plywood-mdf-for-soundproofing/#:~:text=MDF%20has%20a%20density%20of,the%20densest%20option%20for%20soundproofing. Accessed on 19 February 2022
https://in.pinterest.com/pin/746823550685383649/. Accessed on 19 February 2022.
https://www.soundproofcow.com/. Accessed on 19 February 2022.
Hussain, R. and Saleh, Y. (2017). Intelligent system for white grub monitoring through WSN. International journal of current advanced research, 06(05). 3816-3821.
Ijabo, O. J., Irtwange, S. V. and Uguru, H. 2019. Effects of storage on physical and viscoelastic properties of yam tubers. Direct research journal of agriculture and food science, 7(7), 181 191. http://doi.org/10.5281/zenodo.3262306
Iwar, R.T., Iorhemen, O.T., Ogedengbe, K., Katibi, K.K., 2021a. Novel aluminium (hydro) oxide-functionalized activated carbon derived from Raffia palm (Raphia hookeri) shells: augmentation of its adsorptive properties for efficient fluoride uptake in aqueous media. Environ. Chem. Ecotoxicol. 3, 142–154.
Iwar, R. T., Ogedengbe, K., Katibi K. K. and Oshido, L. E. 2021b. Meso-microporous activated carbon derived from Raffia palm shells: optimization of synthesis conditions using response surface methodology. Heliyon 7 e07301. https://doi.org/10.1016/j.heliyon.2021.e07301
Kalloo G. and Bergh B.O. 1993. Genetic improvement of vegetable crops. eBook ISBN: 9780080984667. Pergamon. Elsevier Ltd.
Kennedy G., Raneri J. E., Stoian D., Attwood S., Burgos G., Ceballos H., Ekesa B., Johnson V., Low J. W. and Talsma E. F. 2019. Roots, tubers and bananas: contributions to food security, Editor(s): Pasquale Ferranti, Elliot M. Berry, Jock R. Anderson, Encyclopedia of Food Security and Sustainability, Elsevier, Pages 231-256, ISBN 9780128126882, https://doi.org/10.1016/B978-0-08-100596-5.21537-0.
Knorr D., Froehling A., Jaeger H., Reineke K., Schlueter O., and . Schoessler K. 2011. Annu. rev. food sci. technol. 2:203–235. doi:10.1146/annurev.food.102308.124129
Liu Y., Huang B., Sun Y., Chen F., Yang L., Mao Q., Liu J. and Zheng M. 2015. Relationship of carrot sensory crispness with acoustic signal characteristics. International journal of food science and technology, 50; 1574–1582. doi:10.1111/ijfs.12808
Liu,Y., Qianwen W., Jialing H., Xinru Z., Yingheng Z., Saimin Z., Huimin L., Lin G. and Mengling C. 2021. Comparison of apple firmness prediction models based on non-destructive acoustic signal. International journal of food science and technology, 1 – 7. doi:10.1111/ijfs.15311
Lu Q., Wang J., Gómez A. H. and Pereira A. G. 2009. Evaluation of tomato quality during storage by acoustic impulse response. Journal of food processing and preservation 33; 356 370. DOI: 10.1111/j.1745-4549.2008.00346.x
Mengxing Li, Ekramirad, Nader, Rady, Ahmed, and Adedeji, Akinbode A. 2018, Application of acoustic emission and machine learning to detect codling moth infested apples. Transactions of the ASABE, 61, 3; 1157-1164. https://doi.org/10.13031/trans.12548
Mizrach, A. 2004. Assessing plum fruit quality attributes with an ultrasonic method. Food research international 37; 627–631. doi:10.1016/j.foodres.2003.12.015
Morrison D.S. and Abeyratne U.R. 2014. Ultrasonic technique for non-destructive quality evaluation of oranges. Journal of food engineering, 141; 107–112. http://dx.doi.org/10.1016/j.jfoodeng.2014.05.018
Nabeshima E. H., Moro T. M.A., Campelo P. H., Sant'Ana A. S. and Clerici M. T. P.S. 2020. Chapter Seven - Tubers and roots as a source of prebiotic fibers, Editor(s): Adriano Gomes da Cruz, Elane Schwinden Prudencio, Erick Almeida Esmerino, Marcia Cristina da Silva. Advances in food and nutrition research. Academic Press, volume 94, pages 267-293, ISSN 1043-4526, ISBN 9780128202180, https://doi.org/10.1016/bs.afnr.2020.06.005.
Pan Z., Griffiths G. A., Lin W. and Ronald H. 2010. Development of impact acoustic detection and density separations methods for production of high-quality processed beans. Journal of food engineering, 97; 292–300. doi:10.1016/j.jfoodeng.2009.10.016
Pathan M., Pate N. , Yagnik H., Shah M. 2020. Artificial cognition for applications in smart agriculture: A comprehensive review. Artificial intelligence in agriculture, 4; 81–95. https://doi.org/10.1016/j.aiia.2020.06.001
Przybył K. , Adamina D. , Krzysztof K. , Jerzy S., Mariusz P. and Łukasz G. 2020. Classification of dried strawberry by the analysis of the acoustic sound with artificial neural networks. Sensors, 20, 499; doi:10.3390/s20020499
Russell R. 2018. Machine learning: step-by-step guide to implement machine learning algorithms with python. ISBN:1719528403, 9781719528405. Create space independent publishing platform. https://books.google.com.ng/books?id=9O-mtwEACAAJ
Saranraj P., Behera S. S. and Ray R. C. 2019. Chapter 7 - Traditional foods from tropical root and tuber crops: innovations and challenges. Editor(s): Charis M. Galanakis, Innovations in traditional foods, Woodhead publishing, pages 159-191, ISBN 9780128148877, https://doi.org/10.1016/B978-0-12-814887-7.00007-1.
Siriwardena, K. A. P.; Fernando, L. C. P.; Nanayakkara, N.; Perera, K. F. G.; Kumara, A. D. N. T and Nanayakkara, T. (2010). Portable acoustic device for detection of coconut palms infested by Rynchophorus ferruginous. Crop protection. 29:25-29.
Sisodiya, K. and Singh, M. (2016). Design and development of ultrasonic and IR insect detectors for oilseeds crop. International journal of electronics & communication technology. 7(4): 2230-7109.
Swamynathan, M. 2017. Mastering machine learning with python in six steps. ISBN-13 (pbk): 978-1-4842-2865-4, ISBN-13 (electronic): 978-1-4842-2866-1, DOI 10.1007/978-1-4842-2866 1. Apress Media LLC, New York, USA.
World health organization (WHO) report. 2015. Make listening safe. 20 avenue appia, 1211 geneva 27, Switzerland. https://www.who.int/pbd/deafness/activities/MLS_Brochure_English_lowres_for_web.pdf
Yam market report. 2022. Yams market - growth, trends, covid-19 impact, and forecasts (2022 - 2027). https://www.researchandmarkets.com/r/z2o12g
Zhang H., Zhihua Z., Don K. and Jie W. 2021. Detection of early core browning in pears based on statistical features in vibro-acoustic signals. Food and bioprocess technology 14:887–897. https://doi.org/10.1007/s11947-021-02613-2
Zhang W., Cui D., Ying Y. 2014. Nondestructive measurement of pear texture by acoustic vibration method. Postharvest biology and technology 96; 99–105. http://dx.doi.org/10.1016/j.postharvbio.2014.05.006
Zhang Y., Deng X., Xu Z., Yuan P. 2019. Watermelon ripeness detection via extreme learning machine with kernel principal component analysis based on acoustic signals. International Journal of Pattern Recognition and Artificial Intelligence, 33, No. 08, 1951002. DOI: 10.1142/S0218001419510029.
Zhao K., Zha Z., Li H. and Wu J. 2021. Early detection of moldy apple core based on time-frequency images of vibroacoustic signals. Postharvest biology and technology, 179; 111589. https://doi.org/10.1016/j.postharvbio.2021.111589

sublimentryforacoustic.docx

Development of Two Smart Acoustic Yam Quality Detection Devices Using a Machine Learning Approach

Status:

Journal Publication

Version 1

Abstract

Figures

1.0 Introduction

2.0 Materials And Methods

2.1 Design and Construction of two acoustic Yam quality detection devices

2.1.1 Design of the yam detecting devices

2.1.2 Construction of the yam detecting devices

2.1.2.1 Software sound generation-technique device

2.1.2.2 Sample impact sound generation-technique device

2.2 Evaluation of the developed yam quality detection devices

2.2.1 Sample acquisition and preparation

2.2.2 Experimental procedures for acoustic data acquisition

2.2.2.1 Software sound generation-technique device

2.2.2.2 Sample impact sound generation-technique device

2.3 Data Analysis

2.3.1 Discriminate analysis

2.3.2 Classification and validation

3.0 Results And Discussion

3.1 White yam (Discorea rotundata)

3.2 Yellow yam (Dioscorea Cayanensis)

4.0 Conclusions

Declarations

References

Supplementary Files

Status:

Journal Publication

Version 1