- Email: [email protected]
Evaluation of tomato maturity by electronic nose
Computers and Electronics in Agriculture 54 (2006) 44–52
Evaluation of tomato maturity by electronic nose Antihus Hern´andez G´omez a,b , Guixian Hu a,c , Jun Wang a,∗ , Annia Garc´ıa Pereira a,b a
Department of Agricultural Engineering, Zhejiang University, 268 Kaixuan Road, Hangzhou 310029, China b Agricultural Mechanization Faculty, Havana Agricultural University, Cuba c Zhejiang Academy of Agricultural Science, Hangzhou 310021, China Received 3 July 2004; received in revised form 21 April 2006; accepted 16 July 2006
Abstract Over the past years, electronic nose (E-nose) technology opened has enhanced the possibility of exploiting information on behavior aroma to assess fruit ripening stage. The objective in this study was to evaluate the capacity of electronic nose to monitor the change in volatile production of ripeness states for tomato, using a specific electronic nose device with 10 different metal oxide sensors (portable E-nose, PEN 2). Principal component analysis (PCA) and linear discriminant analysis (LDA) were used to investigate whether the electronic nose was able to distinguishing among different ripeness states (unripe, half-ripe, full-ripe and over-ripe). The loadings analysis was used to identify the sensors responsible for discrimination in the current pattern file. The results prove that the electronic nose PEN 2 could differentiate among the ripeness states of tomato. The electronic nose was able to detect a clearer difference in volatile profile of tomato when using LDA analysis than when using PCA analysis. Using LDA analysis, it was possible to differentiate and to classify the different tomato maturity states, and this method was able to classify 100% of the total samples in each respective group. Some sensors in E-nose have the highest influence in the current pattern file for electronic nose PEN 2. A subset of a few sensors in E-nose can be chosen to explain all the variance. This result could be used in further studies to optimize the number of sensors. © 2006 Elsevier B.V. All rights reserved. Keywords: Electronic nose; Non-destructive method; Monitoring; Maturity; Tomato
1. Introduction The quality concept in foods is mainly related to consumer perception and preference. Consumer perception is based on the application of the five senses and for this reason, the instrument “par excellence” to determine food quality is the human senses. Actually, panels of trained people are used to fix and label the criteria of quality, to assess the quality of food, and to help in the development of new products. From an instrumental point of view, there is an obvious correlation between the human senses and the application of optical, chemical and tactile sensors. For several years, the instrumental measure of fruit quality has been mostly based on the basis of rheological properties such as texture and firmness (Wang et al., 2004). The main disadvantage of the majority of these techniques is that they are not practical for cultivars or storage stations. Moreover, most of the techniques require the destruction of the samples used for analysis. Currently, optimal ∗
Corresponding author. Tel.: +86 571 86971881; fax: +86 571 86971139. E-mail address: [email protected] (J. Wang).
0168-1699/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2006.07.002
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
45
harvest dates and predictions of storage life are mainly based on practical experience, but learning these critical decisions to subjective interpretation implies that large quantities of fruit are harvested too soon or too late and reach consumer markets in poor condition. Extensive research has been focused on the development of non-destructive techniques for measuring quality attributes of fruit. Aroma sensing is a particularly promising method of assessing fruit quality. An alternative strategy for determining the state of ripeness consists of sensing the aromatic volatiles emitted by fruit using electronic olfactory systems (Benady et al., 1995; Gomez et al., 2006). These systems are concerned with the exploitation of the information contained in the headspace of fruits; they have been studied in the recent past with conventional analytical chemistry equipment; and the correlation between the state of over-ripening and the fruit aroma has also been found both in quantitative and qualitative terms. Some specific compounds have been identified as being responsible for the aroma of particular fruit. In the last decade, the electronic nose technology has opened the possibility to exploit, from a practical point of view, the information contained in the headspace in many different application fields. Among them, food analysis is certainly one of the most often practiced. The electronic nose offers a fast and non-destructive alternative to sense aroma, and, hence, may be advantageously used to predict the optimal harvest date. Commercially available electronic noses use an array of sensors combined with pattern recognition software. There have been several reports on electronic sensing in environmental control, medical diagnostics and the food industry (Keller et al., 1995; Schaller et al., 1998; Hai and Wang, 2006). Some authors reported positive applications of electronic nose technology to the discrimination of different fruits quality, and many experiments were performed, such as: testing orange (Di Natale et al., 2001a), melons (Benady et al., 1992, 1995), blueberries (Simon et al., 1996), pears (Oshita et al., 2000; Correa et al., 2001), peaches (Molto et al., 1999; Brezmes et al., 2000; Di Natale et al., 2001b), bananas (Llobet et al., 1999), apples (Hines et al., 1999; Brezmes et al., 2000, 2001; Saevels et al., 2003), and nectarines (Di Natale et al., 2001b). The objectives in this research were: (1) to evaluate the capacity of the electronic nose for monitoring tomato maturity, using a specific electronic nose device (PEN 2) based on a 10-sensor array, (2) to study principal component analysis (PCA) and linear discriminant analysis (LDA) techniques to obtain whether the electronic nose would be able to distinguish different ripeness, (3) to identify each sensor responsible for a discrimination in the current pattern file, using loading analysis. 2. Materials and methods 2.1. Experimental procedure and sample Tomato ‘Heatwave’ (Lycopersicum esculentum) was selected for the experiment. The samples were hand harvested on 9th November 2003 from the experimental orchard in the Department of Horticulture, Zhejiang University. Tomatoes were sorted and selected according to a uniform size and weight approximately. All measurements were made on the same date and following the same procedure carefully. Because fruit were harvested from different plants, then randomized, the experimental design was completely randomized with each fruit as an experimental unit. All fruits of each sample were individually numbered. Before non-destructive nose measurements, the samples were profiled for two sensory descriptors referring to color (intensity, tone, whiteness; Berna et al., 2004), and the deformation measured at four orientation-points in fruit equator with a 10 N force, was used as a general term to describe the mechanical properties of the fruit. The compression measurements were carried out in a Universal Testing Machine with parallel plate (Model 5543 Single Column, Instron Corp., Canton, MA, USA). The greater deformations (>6% of its “equatorial” diameter) were classified as over-ripe tomatoes and the smaller deformations (<2% of its “equatorial” diameter) were classified as unripe tomatoes. The sensory panel classified the fruit by color (Green: >10% but not more than 30% of the surface is not green; in the aggregate, shows a definite change from green to tannish-yellow, pink, red, or a combination thereof. Pink: >30% but not more than 60% of the surface is not green, in the aggregate, shows pink or red color. Light-red: >60% of the surface is not green; in the aggregate, shows pinkish-red or red, provided that not >90% of the surface is red color. Red: more than 90% of the surface is not green; in the aggregate, shows red color). The sensor panel was a selected and trained profile panel of five assessors with previous experience in tomato assessment. Four groups, each with 20
46
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 1. Schematic diagram of the electronic nose measurements.
samples, were classified into different maturity states (unripe, half-ripe, full-ripe and over-ripe). A total of 80 nose measurements were performed. For validation, the same variety, tomato ‘Heatwave’, was selected for the experiment. But the samples were hand harvested from the experimental orchard in the Institute of Vegetable Research, Zhejiang Academy of Agricultural Science, which is 4 km from the experimental orchard in the Department of Horticulture. Tomatoes were harvested on 9th November 2003. Each group included 10 tomatoes, and a total of 40 nose measurements were performed for validation. 2.2. Electronic nose data acquisition and analysis Experiments were performed with a portable E-nose (PEN 2) operating with the enrichment and desorption unit (EDU). The system was from Win Muster Airsense (WMA) Analytics Inc. (Schwerin, Germany). PEN 2 consists of a sampling apparatus, a detector unit containing the array of sensors, and pattern recognition software (Win Muster v.1.6) for data recording. The sensor array is composed of 10 metal oxide semiconductor (MOS) type chemical sensor: MOS1 (aromatic), MOS2 (broadrange), MOS3 (aromatic), MOS4 (hydrogen), MOS5 (arom-aliph), MOS6 (broadmethane), MOS7 (sulphur-organic), MOS8 (broad-alcohol), MOS9 (sulph-chlor), MOS10 (methane-aliph). Fig. 1 shows a schematic diagram of the electronic nose measurements and gas flow of PEN 2 during the experiments. Table 1 lists all 10 of the sensors used, and their main applications. Each fruit was placed into an airtight glass jar with a volume of 1 L (concentration chamber). The glass jar was then closed and the headspace inside it was equilibrated for 1 h. Preliminary experiments showed that after 0.5 h of equilibration, the headspace reached a steady state. One luer-lock needle, connected to Teflon tubing (3 mm), was used to perforate the seal (plastic) of the vial and to absorb the air accumulated inside it during the measurements. The headspace gas was pumped over the sensors of the electronic nose; and during the measurements process, three Table 1 Sensors used and their main applications in PEN 2 electronic nose Number in array
Sensor-name
General description
Reference
1 2
W1C W5S
Toluene, 10 ppm NO2 , 1 ppm
3 4 5 6
W3C W6S W5C W1S
7
W1W
8 9 10
W2S W2W W3S
Aromatic compounds Very sensitive, broad range sensitivity, react on nitrogen oxides, very sensitive with negative signal Ammonia, used as sensor for aromatic compounds Mainly hydrogen, selectively (breath gases) Alkanes, aromatic compounds, less polar compounds Sensitive to methane (environment) ca. 10 ppm. Broad range, similar to No. 8 Reacts on sulphur compounds, H2 S 0.1 ppm. Otherwise sensitive to many terpenes and sulphur organic compounds, which are important for smell, limonene, pyrazine Detects alcohols, partially aromatic compounds, broad range Aromatics compounds, sulphur organic compounds Reacts on high concentrations >100 ppm, sometime very selective (methane)
Benzene, 10 ppm H2 , 100 ppb Propane, 1 ppm CH3 , 100 ppm H2 S, 1 ppm
CO, 100 ppm H2 S, 1 ppm CH3 , 10 CH3 , 100 ppm
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
47
different phases can be distinguished: concentration, measurement and stand-by. The electro valves, controlled by a computer program, guided the air though different circuits depending on the measurement phase. Irrespective of the phase, airflow was always kept constant though the measurement chamber. During the measurement phase, the bomb pushed the volatiles though a closed loop that included the measurement and concentration chambers. No air entered, and none exited the loop. The measurement phase lasted 60 s, which was sufficient time for the sensors to reach a stable value. The sample interval was 1 s. When a measurement was completed, a stand-by phase was activated (60 s). The main purpose of the stand-by phase was to clean the circuit and return the sensors to their baseline. During this phase, clean air entered the circuit, crossed the measurement chamber first, then the empty concentration chamber, and pushed the remaining volatiles out of the circuit. Experiments were conducted at the temperature of 20 ◦ C and 50–60% RH during all experiments, and the temperature was maintained constant with an accuracy of ±1 ◦ C. During the measurement phase, the computer recorded the resistance changes that the sensors experienced. When the measurement was completed, the acquired data were properly stored for later use. The set of signals of all sensors during measurement of a sample is a pattern. Pattern of multiple measurements dealing with the same problem are stored in a pattern file and act as the training set. The pattern data were recorded, checked visually and analyzed using Win Muster (Version 1.5.2.4 June 2003, copyright 1996–2002 WMA Airsense Analysentechnik GmbH 2003). 2.3. Principal component analysis, linear discriminant analysis and loadings analysis Principal component analysis is a chemometric linear, unsupervised and pattern recognition technique used for analyzing, classifying and reducing the dimensionality of numerical datasets in a multivariate problem. This method permits extraction of useful information from the data, and exploration of the data structure, the relationship between objects, the relationship between objects and variables, and the global correlation of the variables. The main features of PCA are the coordinates of the data in the new base (scores plot) and the contribution to each component of the sensors (loads plot). The score plot is usually used for studying the classification of the data clusters, while the loads plot can provide information on the relative importance of the array sensors to each principal component and their mutual correlation. The linear discriminant analysis calculates the discriminant functions and similar to the PCA—a two- or threedimensional display of the training set data. The difference between PCA and LDA is that PCA does not consider the relation of a data point to the specified classes, while the LDA calculation uses the class information that was given during training. The LDA utilizes information about the distribution within classes and the distances between them. Therefore, the LDA is able to collect information from all sensors in order to improve the resolution of classes. The loadings analysis is well correlated to the PCA. Using this analysis, the sensors can be investigated for their responsibility for the discrimination given by the trained patterns. Sensors, located near the center of the diagram (0, 0), have a minor responsibility for the distribution of pattern in the PCA plot. They may be switched off because they may have a negative influence on the pattern resolution, when particular normalizations are selected. The loadings analysis helps to identify the sensor responsible for discrimination in the current pattern file. The sensor may be switched off (the response signal not used) for analysis if it has no positive influence on the identification process. 3. Results and discussion 3.1. Electronic nose response to fruit aroma Fig. 2 shows a typical response of 10 sensors during measurement of tomato. The data obtained are the changing ratio of conductivity between G and G0 (the conductivity of the sensors when the sample gas or zero gas blows over). Each curve represents a different sensor transient. The curves represent sensor conductivity of one sensor of the array against time due to electro valve action when the volatiles from the fruit reach the measurement chamber. In that transition, the clean airflow that reached the measurement chamber was substituted by airflow that came from the concentration chamber, closing a loop circuit between both chambers. It can be seen (Fig. 2) that conductivity increases sharply and then stabilizes after 30 s after an initial period of low and stable conductivity (when only clean air is crossing the
48
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 2. Ten sensor’s typical responses to tomato aroma.
measurement chamber). Each sensor signal generally stabilized and was considered for use in analysis of the electronic nose. In this research, the signal of each sensor at 42 s after electro valve action was used in analysis of the electronic nose. Fig. 3 shows the response value of each sensor in Cartesian coordinate for an example at 42 s. 3.2. Signal analysis Fig. 4 shows the evolution of the signals generated by the sensor array. Each line represents the average signal variation of 20 tomatoes for one sensor of the array (10 sensors), linking to the measurements of conductance increase or decrease experienced by the sensor as maturity state that vapors from the fruit reached the measurement chamber. The tomato results were similar to those obtained by Brezmes et al. (2000), when testing peaches and pears. It is possible to infer that, in all cases, sensor 2 exhibited higher values that were different from the rest of the sensors. 3.3. Classification of mandarin using PCA and LDA To investigate whether the electronic nose was able to distinguish between different ripeness state, PCA and LDA analysis were applied to 80 samples. PCA and LDA analysis results are shown in Figs. 5 and 6. These figures show the analysis results on a two-dimensional plane, principal component 1 (PC1) and principal component 2 (PC2) in Fig. 5 and first and second linear discriminant LD1 and LD2 in Fig. 6. PCA is a linear combinatorial method, which reduces the complexity of the dataset. The inherent structure of the dataset is preserved while its resulting variance is maximized. PCA has been performed to describe the aroma changes during the picking process. Fig. 5 shows a clearer discrimination among the various clusters representing the tomato
Fig. 3. Typical relative conductivity (G/G0 ) vs. sensors at 42 s.
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 4. Relative conductivity of each sensor vs. tomato maturity class.
Fig. 5. PCA analysis for tomato ripeness.
Fig. 6. LDA analysis for tomato ripeness.
49
50
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
ripeness state. Each group was clearly distinguishable from the other groups by using PCA analysis, except half-ripe or unripe groups are overlapped lightly, and there is greater distinction among half-ripe or unripe group, full-ripe group and over-ripe group. The processed data show a shift of the different maturity state coinciding with the classification by the trained profile panel. The first principal component, PC1, explains 95.79% of the total variation, while 2.79% of the total variance is explained by PC2. The system has enough resolution to explain the tomato ripeness state. PCA analysis showed the variation of each group along the abscissa (PC1) with a trend, however, the over-ripe group showed an advance in negative direction on the function 1 in relation with full-ripe group. The full-ripe and over-ripe groups show a clear upward and downward displacement in negative and positive direction on the ordinate axis (PC2), respectively, moving these groups away from the other two groups. The LDA analysis was applied to the same dataset, and it showed a very clear discrimination among the various clusters representing different tomato ripeness state, all tomatoes were perfectly classified (Fig. 6). In this plot, about 88.3% of the total variance of the data is displayed. LDA function 1 (LD1) and function 2 (LD2) accounted for 76.6 and 11.7% of the variance, respectively, but the LDA analysis showed an erratic variation of each group along the abscissa (LD1) and ordinate (LD2). Using PCA and LDA analysis, it is possible to classify the fruit into four maturity states. When the electronic nose was performed with LDA, better classification rates were observed. 3.4. Validation analysis for tomato data using LDA Validation analysis was performed using 40 samples divided into four groups with 10 tomatoes in each group. Tomatoes were of the same variety, from the experimental orchard in the Institute of Vegetable Research, Zhejiang Academy of Agricultural Science. These datasets of samples were presented, in correspondence with the four different maturity states, and classified using LDA shown in Fig. 7. The results show that all samples were consequently classified into their respective groups. 3.5. Loadings plot analysis The loadings analysis will help to identify the sensors responsible for discrimination in the current pattern file. The sensor might be switched off for analysis (the response signal was not used) if it has a rather smaller influence on the identification process. The sensor with loading parameters (by loading analysis) near to zero for a particular principal component have a low contribution to the total response of the array, this sensor could be switched off, and response signal was not used for discrimination analysis, whereas high values indicates a discriminating sensor. The loading analysis was performed, and a loading plot of the loading factors associated for tomatoes is shown in Fig. 8. The plot shows the relative importance of the sensors in the array. The loading factor associated to the first and second principal components for each sensor is represented.
Fig. 7. LDA validation analysis for tomato.
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
51
Fig. 8. Loading analysis related to tomato.
Fig. 8 shows that sensors 2, 6 and 8 have a higher influence in the current pattern file, while sensors 7 and 9 have minor influence. There are sensor groups that have almost identical loading parameters and these could be represented by just one sensor. For example, sensors 1, 3 and 5 have similar loading factors, the same can be observed for sensors 4 and 10, and both of these groups have a low influence in the current pattern file. This is identical with the results shown in Fig. 4. Sensors with loading parameters near to the dilution factor for a particular principal component also make little contribution to the total response of the array. Hence, a subset of few sensors can be chosen to explain nearly all variance. This result could be used in further studies to optimize the number of sensors. 4. Conclusions It was evaluated that the capacity of electronic nose to monitor the change in volatile production of ripeness states for tomato, using a specific electronic nose device with 10 different metal oxide sensors (PEN 2). Principal component analysis and linear discriminant analysis were used to investigate whether the electronic nose was able to distinguishing among different ripeness states. The loadings analysis was used to identify the sensors responsible for discrimination in the current pattern file. The results prove that the electronic nose PEN 2 could differentiate among the ripeness states of tomato. The electronic nose was able to detect a clearer difference in volatile profile of tomato using LDA analysis than using PCA analysis. Using LDA analysis is possible to differentiate and to classify the different tomato maturity states, and this method was able to classify 100% of the total samples in each respective group. Some sensors in E-nose have the highest influence in the current pattern file for electronic nose PEN 2. A subset of few sensors in E-nose can be chosen to explain all the variance. This result could be used in further studies to optimize the number of sensors. Acknowledgements The authors acknowledge the financial support of Chinese National Foundation of Nature and Science (30571076) and the financial support of Program for New Century Excellent Talents in Chinese University (NCET-04-0544). References Benady, M., Simon, J.E., Charles, D.J., Miles, G.E., 1992. Determination Melon Ripeness by Analyzing Headspace Gas Emission. ASAE Paper No. 92-6055, ASAE, St. Joseph, MI. Benady, M., Simon, J.E., Charles, D.J., Miles, G.E., 1995. Fruit ripeness determination by electronic sensing of aromatic volatiles. Trans. ASAE 38, 251–257.
52
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Berna, A.Z., Lammertyn, J., Saevels, S., Di Natale, C., Nicola¨ı, B.M., 2004. Electronic nose systems to study shelf life and cultivar effect on tomato aroma profile. Sens. Actuators B 97, 324–333. Brezmes, J., Llobet, E., Vilanova, X., Orts, J., Saiz, G., Correig, X., 2001. Correlation between electronic nose signals and fruit quality indicators on shelf-life measurements with pinklady apples. Sens. Actuators B 80, 41–50. Brezmes, J., Llobet, E., Vilanova, X., Saiz, G., Correig, X., 2000. Fruit ripeness monitoring using an electronic nose. Sens. Actuators B 69, 223–229. Correa, E., Barreiro, P., Ruiz-Altisent, M., Lopez, M.L., Miro, J., Graeli, J., 2001. An aroma sensor for assessing peach quality. In: Proceedings of the Sixth International Symposium on Fruits, Nuts, and Vegetables Production Engineering, Potsdam, Germany, 11–14 September 2001, 162 pp. Di Natale, C., Macagnano, A., Martinelli, E., Paolesse, R., Proietti, E., D’Amico, A., 2001a. The evaluation of quality of post-harvest orange and apples by means of an electronic nose. Sens. Actuators B 78, 26–31. Di Natale, C., Macagnano, A., Martinelli, E., Proietti, E., Paolesse, R., Castellari, L., Campani, S., D’Amico, A., 2001b. Electronic nose based investigation of the sensorial properties of peach and nectarines. Sens. Actuators B 77, 561–566. Gomez, A.H., Wang, J., Hu, G.X., Pereira, A.G., 2006. Electronic nose technique potential monitoring mandarin maturity. Sens. Actuators B 113 (1), 347–353. Hai, Z., Wang, J., 2006. Detection of adulteration in camellia seed oil and sesame oil using an electronic nose. Eur. J. Lipid Sci. Tech. 108 (2), 116–124. Hines, E.L., Llobet, E., Gardner, J.W., 1999. Neural network based electronic nose for apple ripeness determination. Electron. Lett. 35, 821–823. Keller, P.E., Kangas, L.J., Liden, L.H., Hashem, S., Kouzes, R.T., 1995. Electronic noses and their applications. In: IEEE Technical Applications Conference (TAC’95) at Northcon’95, Portland, Oregon, 10–12 October 1995. Llobet, E., Hines, E.L., Gardner, J.W., Franco, S., 1999. Non-destructive banana ripeness determination using a neural network-based electronics nose. Sci. Technol. 6, 538–548. Molto, E., Selfa, E., Ferriz, J., Conesa, E., Gutierrez, A., 1999. An aroma sensor for assessing peach quality. J. Agric. Eng. Res. 72, 311–316. Oshita, S., Shima, K., Haruta, T., Seo, Y., Kagawoe, Y., Nakayama, S., Kawana, S., 2000. Discrimination of odors emanating from ‘La France’ pear by semi-conducting polymer sensors. Comput. Electron. Agric. 26, 209–216. Saevels, S., Lammertyn, J., Berna, A.Z., Veraverbeke, E.A., Di Natale, C., Nicola¨ı, B.M., 2003. Electronic nose as a non-destructive tool to evaluate the optimal harvest date of apples. Postharvest Biol. Technol. 30, 3–14. Schaller, E., Bosset, J.O., Escher, F., 1998. “Electronic noses” and their application to food: a review. Food Sci. Technol.—Lebensm.-Wiss. Technol. 31, 305–316. Simon, J.E., Hetzroni, A., Bordelon, B., Miles, G.E., Charles, D.J., 1996. Electronic sensing of aromatic volatiles for quality sorting of blueberries. J. Food Sci. 61, 967–969. Wang, J., Teng, B., Yu, Y., 2004. Pear dynamic characteristics and firmness detection. Eur. Food Res. Technol. 218 (3), 289–294.
Evaluation of tomato maturity by electronic nose Antihus Hern´andez G´omez a,b , Guixian Hu a,c , Jun Wang a,∗ , Annia Garc´ıa Pereira a,b a
Department of Agricultural Engineering, Zhejiang University, 268 Kaixuan Road, Hangzhou 310029, China b Agricultural Mechanization Faculty, Havana Agricultural University, Cuba c Zhejiang Academy of Agricultural Science, Hangzhou 310021, China Received 3 July 2004; received in revised form 21 April 2006; accepted 16 July 2006
Abstract Over the past years, electronic nose (E-nose) technology opened has enhanced the possibility of exploiting information on behavior aroma to assess fruit ripening stage. The objective in this study was to evaluate the capacity of electronic nose to monitor the change in volatile production of ripeness states for tomato, using a specific electronic nose device with 10 different metal oxide sensors (portable E-nose, PEN 2). Principal component analysis (PCA) and linear discriminant analysis (LDA) were used to investigate whether the electronic nose was able to distinguishing among different ripeness states (unripe, half-ripe, full-ripe and over-ripe). The loadings analysis was used to identify the sensors responsible for discrimination in the current pattern file. The results prove that the electronic nose PEN 2 could differentiate among the ripeness states of tomato. The electronic nose was able to detect a clearer difference in volatile profile of tomato when using LDA analysis than when using PCA analysis. Using LDA analysis, it was possible to differentiate and to classify the different tomato maturity states, and this method was able to classify 100% of the total samples in each respective group. Some sensors in E-nose have the highest influence in the current pattern file for electronic nose PEN 2. A subset of a few sensors in E-nose can be chosen to explain all the variance. This result could be used in further studies to optimize the number of sensors. © 2006 Elsevier B.V. All rights reserved. Keywords: Electronic nose; Non-destructive method; Monitoring; Maturity; Tomato
1. Introduction The quality concept in foods is mainly related to consumer perception and preference. Consumer perception is based on the application of the five senses and for this reason, the instrument “par excellence” to determine food quality is the human senses. Actually, panels of trained people are used to fix and label the criteria of quality, to assess the quality of food, and to help in the development of new products. From an instrumental point of view, there is an obvious correlation between the human senses and the application of optical, chemical and tactile sensors. For several years, the instrumental measure of fruit quality has been mostly based on the basis of rheological properties such as texture and firmness (Wang et al., 2004). The main disadvantage of the majority of these techniques is that they are not practical for cultivars or storage stations. Moreover, most of the techniques require the destruction of the samples used for analysis. Currently, optimal ∗
Corresponding author. Tel.: +86 571 86971881; fax: +86 571 86971139. E-mail address: [email protected] (J. Wang).
0168-1699/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2006.07.002
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
45
harvest dates and predictions of storage life are mainly based on practical experience, but learning these critical decisions to subjective interpretation implies that large quantities of fruit are harvested too soon or too late and reach consumer markets in poor condition. Extensive research has been focused on the development of non-destructive techniques for measuring quality attributes of fruit. Aroma sensing is a particularly promising method of assessing fruit quality. An alternative strategy for determining the state of ripeness consists of sensing the aromatic volatiles emitted by fruit using electronic olfactory systems (Benady et al., 1995; Gomez et al., 2006). These systems are concerned with the exploitation of the information contained in the headspace of fruits; they have been studied in the recent past with conventional analytical chemistry equipment; and the correlation between the state of over-ripening and the fruit aroma has also been found both in quantitative and qualitative terms. Some specific compounds have been identified as being responsible for the aroma of particular fruit. In the last decade, the electronic nose technology has opened the possibility to exploit, from a practical point of view, the information contained in the headspace in many different application fields. Among them, food analysis is certainly one of the most often practiced. The electronic nose offers a fast and non-destructive alternative to sense aroma, and, hence, may be advantageously used to predict the optimal harvest date. Commercially available electronic noses use an array of sensors combined with pattern recognition software. There have been several reports on electronic sensing in environmental control, medical diagnostics and the food industry (Keller et al., 1995; Schaller et al., 1998; Hai and Wang, 2006). Some authors reported positive applications of electronic nose technology to the discrimination of different fruits quality, and many experiments were performed, such as: testing orange (Di Natale et al., 2001a), melons (Benady et al., 1992, 1995), blueberries (Simon et al., 1996), pears (Oshita et al., 2000; Correa et al., 2001), peaches (Molto et al., 1999; Brezmes et al., 2000; Di Natale et al., 2001b), bananas (Llobet et al., 1999), apples (Hines et al., 1999; Brezmes et al., 2000, 2001; Saevels et al., 2003), and nectarines (Di Natale et al., 2001b). The objectives in this research were: (1) to evaluate the capacity of the electronic nose for monitoring tomato maturity, using a specific electronic nose device (PEN 2) based on a 10-sensor array, (2) to study principal component analysis (PCA) and linear discriminant analysis (LDA) techniques to obtain whether the electronic nose would be able to distinguish different ripeness, (3) to identify each sensor responsible for a discrimination in the current pattern file, using loading analysis. 2. Materials and methods 2.1. Experimental procedure and sample Tomato ‘Heatwave’ (Lycopersicum esculentum) was selected for the experiment. The samples were hand harvested on 9th November 2003 from the experimental orchard in the Department of Horticulture, Zhejiang University. Tomatoes were sorted and selected according to a uniform size and weight approximately. All measurements were made on the same date and following the same procedure carefully. Because fruit were harvested from different plants, then randomized, the experimental design was completely randomized with each fruit as an experimental unit. All fruits of each sample were individually numbered. Before non-destructive nose measurements, the samples were profiled for two sensory descriptors referring to color (intensity, tone, whiteness; Berna et al., 2004), and the deformation measured at four orientation-points in fruit equator with a 10 N force, was used as a general term to describe the mechanical properties of the fruit. The compression measurements were carried out in a Universal Testing Machine with parallel plate (Model 5543 Single Column, Instron Corp., Canton, MA, USA). The greater deformations (>6% of its “equatorial” diameter) were classified as over-ripe tomatoes and the smaller deformations (<2% of its “equatorial” diameter) were classified as unripe tomatoes. The sensory panel classified the fruit by color (Green: >10% but not more than 30% of the surface is not green; in the aggregate, shows a definite change from green to tannish-yellow, pink, red, or a combination thereof. Pink: >30% but not more than 60% of the surface is not green, in the aggregate, shows pink or red color. Light-red: >60% of the surface is not green; in the aggregate, shows pinkish-red or red, provided that not >90% of the surface is red color. Red: more than 90% of the surface is not green; in the aggregate, shows red color). The sensor panel was a selected and trained profile panel of five assessors with previous experience in tomato assessment. Four groups, each with 20
46
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 1. Schematic diagram of the electronic nose measurements.
samples, were classified into different maturity states (unripe, half-ripe, full-ripe and over-ripe). A total of 80 nose measurements were performed. For validation, the same variety, tomato ‘Heatwave’, was selected for the experiment. But the samples were hand harvested from the experimental orchard in the Institute of Vegetable Research, Zhejiang Academy of Agricultural Science, which is 4 km from the experimental orchard in the Department of Horticulture. Tomatoes were harvested on 9th November 2003. Each group included 10 tomatoes, and a total of 40 nose measurements were performed for validation. 2.2. Electronic nose data acquisition and analysis Experiments were performed with a portable E-nose (PEN 2) operating with the enrichment and desorption unit (EDU). The system was from Win Muster Airsense (WMA) Analytics Inc. (Schwerin, Germany). PEN 2 consists of a sampling apparatus, a detector unit containing the array of sensors, and pattern recognition software (Win Muster v.1.6) for data recording. The sensor array is composed of 10 metal oxide semiconductor (MOS) type chemical sensor: MOS1 (aromatic), MOS2 (broadrange), MOS3 (aromatic), MOS4 (hydrogen), MOS5 (arom-aliph), MOS6 (broadmethane), MOS7 (sulphur-organic), MOS8 (broad-alcohol), MOS9 (sulph-chlor), MOS10 (methane-aliph). Fig. 1 shows a schematic diagram of the electronic nose measurements and gas flow of PEN 2 during the experiments. Table 1 lists all 10 of the sensors used, and their main applications. Each fruit was placed into an airtight glass jar with a volume of 1 L (concentration chamber). The glass jar was then closed and the headspace inside it was equilibrated for 1 h. Preliminary experiments showed that after 0.5 h of equilibration, the headspace reached a steady state. One luer-lock needle, connected to Teflon tubing (3 mm), was used to perforate the seal (plastic) of the vial and to absorb the air accumulated inside it during the measurements. The headspace gas was pumped over the sensors of the electronic nose; and during the measurements process, three Table 1 Sensors used and their main applications in PEN 2 electronic nose Number in array
Sensor-name
General description
Reference
1 2
W1C W5S
Toluene, 10 ppm NO2 , 1 ppm
3 4 5 6
W3C W6S W5C W1S
7
W1W
8 9 10
W2S W2W W3S
Aromatic compounds Very sensitive, broad range sensitivity, react on nitrogen oxides, very sensitive with negative signal Ammonia, used as sensor for aromatic compounds Mainly hydrogen, selectively (breath gases) Alkanes, aromatic compounds, less polar compounds Sensitive to methane (environment) ca. 10 ppm. Broad range, similar to No. 8 Reacts on sulphur compounds, H2 S 0.1 ppm. Otherwise sensitive to many terpenes and sulphur organic compounds, which are important for smell, limonene, pyrazine Detects alcohols, partially aromatic compounds, broad range Aromatics compounds, sulphur organic compounds Reacts on high concentrations >100 ppm, sometime very selective (methane)
Benzene, 10 ppm H2 , 100 ppb Propane, 1 ppm CH3 , 100 ppm H2 S, 1 ppm
CO, 100 ppm H2 S, 1 ppm CH3 , 10 CH3 , 100 ppm
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
47
different phases can be distinguished: concentration, measurement and stand-by. The electro valves, controlled by a computer program, guided the air though different circuits depending on the measurement phase. Irrespective of the phase, airflow was always kept constant though the measurement chamber. During the measurement phase, the bomb pushed the volatiles though a closed loop that included the measurement and concentration chambers. No air entered, and none exited the loop. The measurement phase lasted 60 s, which was sufficient time for the sensors to reach a stable value. The sample interval was 1 s. When a measurement was completed, a stand-by phase was activated (60 s). The main purpose of the stand-by phase was to clean the circuit and return the sensors to their baseline. During this phase, clean air entered the circuit, crossed the measurement chamber first, then the empty concentration chamber, and pushed the remaining volatiles out of the circuit. Experiments were conducted at the temperature of 20 ◦ C and 50–60% RH during all experiments, and the temperature was maintained constant with an accuracy of ±1 ◦ C. During the measurement phase, the computer recorded the resistance changes that the sensors experienced. When the measurement was completed, the acquired data were properly stored for later use. The set of signals of all sensors during measurement of a sample is a pattern. Pattern of multiple measurements dealing with the same problem are stored in a pattern file and act as the training set. The pattern data were recorded, checked visually and analyzed using Win Muster (Version 1.5.2.4 June 2003, copyright 1996–2002 WMA Airsense Analysentechnik GmbH 2003). 2.3. Principal component analysis, linear discriminant analysis and loadings analysis Principal component analysis is a chemometric linear, unsupervised and pattern recognition technique used for analyzing, classifying and reducing the dimensionality of numerical datasets in a multivariate problem. This method permits extraction of useful information from the data, and exploration of the data structure, the relationship between objects, the relationship between objects and variables, and the global correlation of the variables. The main features of PCA are the coordinates of the data in the new base (scores plot) and the contribution to each component of the sensors (loads plot). The score plot is usually used for studying the classification of the data clusters, while the loads plot can provide information on the relative importance of the array sensors to each principal component and their mutual correlation. The linear discriminant analysis calculates the discriminant functions and similar to the PCA—a two- or threedimensional display of the training set data. The difference between PCA and LDA is that PCA does not consider the relation of a data point to the specified classes, while the LDA calculation uses the class information that was given during training. The LDA utilizes information about the distribution within classes and the distances between them. Therefore, the LDA is able to collect information from all sensors in order to improve the resolution of classes. The loadings analysis is well correlated to the PCA. Using this analysis, the sensors can be investigated for their responsibility for the discrimination given by the trained patterns. Sensors, located near the center of the diagram (0, 0), have a minor responsibility for the distribution of pattern in the PCA plot. They may be switched off because they may have a negative influence on the pattern resolution, when particular normalizations are selected. The loadings analysis helps to identify the sensor responsible for discrimination in the current pattern file. The sensor may be switched off (the response signal not used) for analysis if it has no positive influence on the identification process. 3. Results and discussion 3.1. Electronic nose response to fruit aroma Fig. 2 shows a typical response of 10 sensors during measurement of tomato. The data obtained are the changing ratio of conductivity between G and G0 (the conductivity of the sensors when the sample gas or zero gas blows over). Each curve represents a different sensor transient. The curves represent sensor conductivity of one sensor of the array against time due to electro valve action when the volatiles from the fruit reach the measurement chamber. In that transition, the clean airflow that reached the measurement chamber was substituted by airflow that came from the concentration chamber, closing a loop circuit between both chambers. It can be seen (Fig. 2) that conductivity increases sharply and then stabilizes after 30 s after an initial period of low and stable conductivity (when only clean air is crossing the
48
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 2. Ten sensor’s typical responses to tomato aroma.
measurement chamber). Each sensor signal generally stabilized and was considered for use in analysis of the electronic nose. In this research, the signal of each sensor at 42 s after electro valve action was used in analysis of the electronic nose. Fig. 3 shows the response value of each sensor in Cartesian coordinate for an example at 42 s. 3.2. Signal analysis Fig. 4 shows the evolution of the signals generated by the sensor array. Each line represents the average signal variation of 20 tomatoes for one sensor of the array (10 sensors), linking to the measurements of conductance increase or decrease experienced by the sensor as maturity state that vapors from the fruit reached the measurement chamber. The tomato results were similar to those obtained by Brezmes et al. (2000), when testing peaches and pears. It is possible to infer that, in all cases, sensor 2 exhibited higher values that were different from the rest of the sensors. 3.3. Classification of mandarin using PCA and LDA To investigate whether the electronic nose was able to distinguish between different ripeness state, PCA and LDA analysis were applied to 80 samples. PCA and LDA analysis results are shown in Figs. 5 and 6. These figures show the analysis results on a two-dimensional plane, principal component 1 (PC1) and principal component 2 (PC2) in Fig. 5 and first and second linear discriminant LD1 and LD2 in Fig. 6. PCA is a linear combinatorial method, which reduces the complexity of the dataset. The inherent structure of the dataset is preserved while its resulting variance is maximized. PCA has been performed to describe the aroma changes during the picking process. Fig. 5 shows a clearer discrimination among the various clusters representing the tomato
Fig. 3. Typical relative conductivity (G/G0 ) vs. sensors at 42 s.
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Fig. 4. Relative conductivity of each sensor vs. tomato maturity class.
Fig. 5. PCA analysis for tomato ripeness.
Fig. 6. LDA analysis for tomato ripeness.
49
50
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
ripeness state. Each group was clearly distinguishable from the other groups by using PCA analysis, except half-ripe or unripe groups are overlapped lightly, and there is greater distinction among half-ripe or unripe group, full-ripe group and over-ripe group. The processed data show a shift of the different maturity state coinciding with the classification by the trained profile panel. The first principal component, PC1, explains 95.79% of the total variation, while 2.79% of the total variance is explained by PC2. The system has enough resolution to explain the tomato ripeness state. PCA analysis showed the variation of each group along the abscissa (PC1) with a trend, however, the over-ripe group showed an advance in negative direction on the function 1 in relation with full-ripe group. The full-ripe and over-ripe groups show a clear upward and downward displacement in negative and positive direction on the ordinate axis (PC2), respectively, moving these groups away from the other two groups. The LDA analysis was applied to the same dataset, and it showed a very clear discrimination among the various clusters representing different tomato ripeness state, all tomatoes were perfectly classified (Fig. 6). In this plot, about 88.3% of the total variance of the data is displayed. LDA function 1 (LD1) and function 2 (LD2) accounted for 76.6 and 11.7% of the variance, respectively, but the LDA analysis showed an erratic variation of each group along the abscissa (LD1) and ordinate (LD2). Using PCA and LDA analysis, it is possible to classify the fruit into four maturity states. When the electronic nose was performed with LDA, better classification rates were observed. 3.4. Validation analysis for tomato data using LDA Validation analysis was performed using 40 samples divided into four groups with 10 tomatoes in each group. Tomatoes were of the same variety, from the experimental orchard in the Institute of Vegetable Research, Zhejiang Academy of Agricultural Science. These datasets of samples were presented, in correspondence with the four different maturity states, and classified using LDA shown in Fig. 7. The results show that all samples were consequently classified into their respective groups. 3.5. Loadings plot analysis The loadings analysis will help to identify the sensors responsible for discrimination in the current pattern file. The sensor might be switched off for analysis (the response signal was not used) if it has a rather smaller influence on the identification process. The sensor with loading parameters (by loading analysis) near to zero for a particular principal component have a low contribution to the total response of the array, this sensor could be switched off, and response signal was not used for discrimination analysis, whereas high values indicates a discriminating sensor. The loading analysis was performed, and a loading plot of the loading factors associated for tomatoes is shown in Fig. 8. The plot shows the relative importance of the sensors in the array. The loading factor associated to the first and second principal components for each sensor is represented.
Fig. 7. LDA validation analysis for tomato.
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
51
Fig. 8. Loading analysis related to tomato.
Fig. 8 shows that sensors 2, 6 and 8 have a higher influence in the current pattern file, while sensors 7 and 9 have minor influence. There are sensor groups that have almost identical loading parameters and these could be represented by just one sensor. For example, sensors 1, 3 and 5 have similar loading factors, the same can be observed for sensors 4 and 10, and both of these groups have a low influence in the current pattern file. This is identical with the results shown in Fig. 4. Sensors with loading parameters near to the dilution factor for a particular principal component also make little contribution to the total response of the array. Hence, a subset of few sensors can be chosen to explain nearly all variance. This result could be used in further studies to optimize the number of sensors. 4. Conclusions It was evaluated that the capacity of electronic nose to monitor the change in volatile production of ripeness states for tomato, using a specific electronic nose device with 10 different metal oxide sensors (PEN 2). Principal component analysis and linear discriminant analysis were used to investigate whether the electronic nose was able to distinguishing among different ripeness states. The loadings analysis was used to identify the sensors responsible for discrimination in the current pattern file. The results prove that the electronic nose PEN 2 could differentiate among the ripeness states of tomato. The electronic nose was able to detect a clearer difference in volatile profile of tomato using LDA analysis than using PCA analysis. Using LDA analysis is possible to differentiate and to classify the different tomato maturity states, and this method was able to classify 100% of the total samples in each respective group. Some sensors in E-nose have the highest influence in the current pattern file for electronic nose PEN 2. A subset of few sensors in E-nose can be chosen to explain all the variance. This result could be used in further studies to optimize the number of sensors. Acknowledgements The authors acknowledge the financial support of Chinese National Foundation of Nature and Science (30571076) and the financial support of Program for New Century Excellent Talents in Chinese University (NCET-04-0544). References Benady, M., Simon, J.E., Charles, D.J., Miles, G.E., 1992. Determination Melon Ripeness by Analyzing Headspace Gas Emission. ASAE Paper No. 92-6055, ASAE, St. Joseph, MI. Benady, M., Simon, J.E., Charles, D.J., Miles, G.E., 1995. Fruit ripeness determination by electronic sensing of aromatic volatiles. Trans. ASAE 38, 251–257.
52
A.H. G´omez et al. / Computers and Electronics in Agriculture 54 (2006) 44–52
Berna, A.Z., Lammertyn, J., Saevels, S., Di Natale, C., Nicola¨ı, B.M., 2004. Electronic nose systems to study shelf life and cultivar effect on tomato aroma profile. Sens. Actuators B 97, 324–333. Brezmes, J., Llobet, E., Vilanova, X., Orts, J., Saiz, G., Correig, X., 2001. Correlation between electronic nose signals and fruit quality indicators on shelf-life measurements with pinklady apples. Sens. Actuators B 80, 41–50. Brezmes, J., Llobet, E., Vilanova, X., Saiz, G., Correig, X., 2000. Fruit ripeness monitoring using an electronic nose. Sens. Actuators B 69, 223–229. Correa, E., Barreiro, P., Ruiz-Altisent, M., Lopez, M.L., Miro, J., Graeli, J., 2001. An aroma sensor for assessing peach quality. In: Proceedings of the Sixth International Symposium on Fruits, Nuts, and Vegetables Production Engineering, Potsdam, Germany, 11–14 September 2001, 162 pp. Di Natale, C., Macagnano, A., Martinelli, E., Paolesse, R., Proietti, E., D’Amico, A., 2001a. The evaluation of quality of post-harvest orange and apples by means of an electronic nose. Sens. Actuators B 78, 26–31. Di Natale, C., Macagnano, A., Martinelli, E., Proietti, E., Paolesse, R., Castellari, L., Campani, S., D’Amico, A., 2001b. Electronic nose based investigation of the sensorial properties of peach and nectarines. Sens. Actuators B 77, 561–566. Gomez, A.H., Wang, J., Hu, G.X., Pereira, A.G., 2006. Electronic nose technique potential monitoring mandarin maturity. Sens. Actuators B 113 (1), 347–353. Hai, Z., Wang, J., 2006. Detection of adulteration in camellia seed oil and sesame oil using an electronic nose. Eur. J. Lipid Sci. Tech. 108 (2), 116–124. Hines, E.L., Llobet, E., Gardner, J.W., 1999. Neural network based electronic nose for apple ripeness determination. Electron. Lett. 35, 821–823. Keller, P.E., Kangas, L.J., Liden, L.H., Hashem, S., Kouzes, R.T., 1995. Electronic noses and their applications. In: IEEE Technical Applications Conference (TAC’95) at Northcon’95, Portland, Oregon, 10–12 October 1995. Llobet, E., Hines, E.L., Gardner, J.W., Franco, S., 1999. Non-destructive banana ripeness determination using a neural network-based electronics nose. Sci. Technol. 6, 538–548. Molto, E., Selfa, E., Ferriz, J., Conesa, E., Gutierrez, A., 1999. An aroma sensor for assessing peach quality. J. Agric. Eng. Res. 72, 311–316. Oshita, S., Shima, K., Haruta, T., Seo, Y., Kagawoe, Y., Nakayama, S., Kawana, S., 2000. Discrimination of odors emanating from ‘La France’ pear by semi-conducting polymer sensors. Comput. Electron. Agric. 26, 209–216. Saevels, S., Lammertyn, J., Berna, A.Z., Veraverbeke, E.A., Di Natale, C., Nicola¨ı, B.M., 2003. Electronic nose as a non-destructive tool to evaluate the optimal harvest date of apples. Postharvest Biol. Technol. 30, 3–14. Schaller, E., Bosset, J.O., Escher, F., 1998. “Electronic noses” and their application to food: a review. Food Sci. Technol.—Lebensm.-Wiss. Technol. 31, 305–316. Simon, J.E., Hetzroni, A., Bordelon, B., Miles, G.E., Charles, D.J., 1996. Electronic sensing of aromatic volatiles for quality sorting of blueberries. J. Food Sci. 61, 967–969. Wang, J., Teng, B., Yu, Y., 2004. Pear dynamic characteristics and firmness detection. Eur. Food Res. Technol. 218 (3), 289–294.