- Email: [email protected]
Identification of some spanish olive cultivars using image processing techniques
Scientia Horticulturae 225 (2017) 286–292
Contents lists available at ScienceDirect
Scientia Horticulturae journal homepage: www.elsevier.com/locate/scihorti
Identification of some spanish olive cultivars using image processing techniques
MARK
⁎
Abdullah Beyaza, , Mücahit Taha Özkayab, Duygu İçenc a b c
Department of Agricultural Machinery and Technologies Engineering, Faculty of Agriculture, Ankara University, Ankara, Turkey Department of Horticulture, Faculty of Agriculture, Ankara University, Ankara, Turkey Department of Statistics, Faculty of Science, Hacettepe University, Ankara, Turkey
A R T I C L E I N F O
A B S T R A C T
Keywords: Artificial vision Image analysis Olive fruit Olive stone Varietal identification
The aim of this research was to identify some Spanish olive cultivars using image processing techniques. For this purpose, Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre Olive cultivars were identified utilizing the image processing and analysis techniques. Therefore, images of olives taken as 300 dpi with the 2896 × 1944 pixels, were captured using a DSLR camera, and evaluations of pixels were used for considering the pixel distribution and dimension measurements. LabVIEW Vision Assistant v2013 (NI) and Image j (NIH) software were used for image analysis procedures. Artificial Neural Network analysis were used to assess information of the length, width and color data results obtained from the fruits and stones (olive stones). All cultivars were identified. In addition, different classification techniques were applied to the olive stone and fruit data with the help of SPSS v22. Clementine v12 was used as a data mining software package from SPSS. The cultivars were identified 90% from dimensions with Artificial Neural Networks.
1. Introduction Olive is an important agricultural product for Turkey and the world. According to the FAO statistics, Turkey ranked as the 4th country in terms of olive production with an income of 2783 million US Dollars. It is known that there are 171,992 olive trees (TSI, 2015) in Turkey. In addition to this, Turkey has an olive production capacity of 1,700,000 tons in a year (TSI, 2015). Olive (Olea europaea L.) production is of a great economic importance for Turkey as it is located in olive production zone. The geographical positions of Turkey have been proved to be suitable in terms of olive production as the overall climate is dominated greatly by Mediterranean climate (i.e. Aegean, Marmara, Mediterranean, Southeast Anatolia, Black Sea regions). The remainders can be observed in olive cultivars due to elements such as widespread region, differences between the climate conditions of the regions and crosspollination. Therefore, it is of significant importance to define the type of the surface patterns of olive cultivars that are widespread in vast parts of the world. In this respect, many studies and researches needs to be conducted for improving olive production. A common problem encountered in studies conducted is that the forms of obtained data are insufficient for the determination of olive cultivars. The reason for this is that the characteristics of olive cultivars depend on the ecological conditions. The phenotypic and genotypic origin, molecular marker
⁎
Corresponding author. E-mail addresses: [email protected], [email protected] (A. Beyaz).
http://dx.doi.org/10.1016/j.scienta.2017.06.041 Received 29 March 2017; Received in revised form 28 May 2017; Accepted 14 June 2017 0304-4238/ © 2017 Elsevier B.V. All rights reserved.
researches cannot provide accurate olive cultivar determination data (Sakar Çakır, 2009) as the full genome sequences of all olive cultivars have not been defined yet. The cultivar determination process has started with the identification of cultivars pomological information. As a consequence of these reasons, a new study was conducted on some Spain cultivars after conducting some Turkish cultivars evaluations. Many studies and researches were conducted on the identification of olive cultivars around the world. For example, Diaz et al. (2000) produced four different algorithms for a machine vision system and applied these algorithms for the identification of olives. In addition to this, they compared human selected olive cultivars with computer vision algorithms. Bari et al. (2003) concentrated on identifying the characteristics of the olives in their studies and they stated that structural characteristics are significant factors in the identification of olives. Diaz et al. (2004) worked on the classification of olives based on the fruit surface defects in different quality categories. Mendoza et al. (2006) worked on different color spaces like RGB, HSV, and L * a * b *, and they stated that the colors of fruits and vegetables were measured by a computerized imaging system that can be applied for the determination of product status. Riquelme et al. (2008) conducted studies on the color defects and structural parameters of olive cultivars. Al Ibrahem et al. (2008) conducted studies on different olive
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
2.2. Imaging system
cultivars (19) using manual measurement and image analysis techniques. Moreda et al. (2012) worked on the shape determination of horticultural products using computer vision technology. They also stated that the computer vision measurements have become a proven and reliable tool for describing the product shape. Vanloot et al. (2014) emphasized that image analysis and metric evaluations of agricultural products like olive stones for determining the varietal origin require specialists, who cannot always conclude with certainty, due to the large number of cultivars identified. Olive production is also significant for many industrial applications as olive identification methods will affect the output characteristics of the olives. In this regard, the aim of this study was to distinguish the characteristics of some olive cultivars in Spain’s World Collection using image processing and analysis techniques. In addition to this, the efficiency of image analysis methods was also investigated in this study. Olives are generally divided into two groups; olive oil and table olive. Olives have a highly variable chemical and physical composition, but these variabilities mostly depend on four major factors. These four factors are cultivars, ecology, cultural practices, and processing. For example, a cultivar which is called dual purpose can be used as both table olive and olive oil. Therefore, cultivar is one of the most important factors for determining the quality of these two products (Gurdeniz et al., 2007). Mono variety is important and indispensable factor for table olive processing. In recent years, mono varietal olive oil has gained importance because of the health benefits and sensory qualities of each cultivar (Gurdeniz et al., 2007). Table olive and olive oil produced from olives of just one cultivar (mono varietal) obtained form one geographical region are becoming increasingly important in the market (Gurdeniz et al., 2007). The documentation of the geographical source of table olive and olive oil is becoming increasingly important as the authenticity and quality issues can be linked with a special region and cultivar. The Denomination of Origin (Protected Denomination of Origin (PDO) and Protected Geographical Indication (PGI)) is a certification program established in Europe for the protection of high-quality agricultural products (Babcock and Clemens, 2004). There are many analytical methods used to determine the varietal origin of olive. Olive stone (pit) is one of the fingerprinting organ used for varietal identification (Ozen et al., 2005). The aim of this study was to identify olive cultivars from their characteristics using image processing and analysis techniques. In this study, the methods of International Olive Council (IOC) and European Union (EU) were adopted for determining olive cultivar and,then, stone, fruit, and leaf data were used for identifying olive cultivars. Moreover, image processing and analysis techniques were employed to accomplish this aim.
After the harvest, 100 samples were taken randomly from each olive cultivar. These samples were photographed from 4 different views − from the front, handle hole, left and tip sides (Fig. 2). These imaging sides are also employed as standard classification views by Kemalpaşa Olive Research Institute for the identification of national collection of Turkey’s olive cultivars. In aggregate, 4400 digital images were captured from 1100 olive fruits for evaluations. A macro capture tripod was set at a distance of 40 centimeter away from the olives to obtain these digital pictures. A Nikon D300 s body with an 18–140 mm zoom lens was used for general purpose imaging and a Nikon D800 with a 105 mm macro lens was used for macro captures. The captured images were stored as JPEG files. All images were captured in 2896 × 1944 pixel dimension and at 300 dpi resolution. Olive fruit and stone images were captured one by one for evaluations. 5 × 5 mm calibration plate was used with a white backdrop for making precision calibration (Fig. 3). First, the primary image capturing problems such as shaded regions and image focus clarity were worked out. Different lenses and flashes were employed for working out the problems encountered during experimentation. The problem of shaded regions was worked out by using a ring flash. Image focus problem was also worked out by using a 105 mm macro lens. After overcoming these problems related to digital imaging, olive cultivars were harvested from the national collection. Then, each olive fruit was photographed, put in numbered plastic bags (5 × 5 cm) and prepared for the following step (olive stone removal step). After completing the picture capturing process, olive stones were removed from fruits. The olive stones were removed from the fruits with an extractor and, then, the olive stone pattern structures were determined. After cleaning the olive stones with the help of a knife, the stones were washed. Then, the stones were stored in plastic containers holding in a 10% bleach solution for 15 h. As a final step, olive stones were stored at − 4 °C to prevent cracking because of physiological activity. Removed olive stones were put in numbered plastic bags (5 × 5 cm). The processes applied to olives are demonstrated in Fig. 4.
2.3. Image processing and determination of sample proportions Image J (NIH) and LabVIEW Vision Assistant v2013 (NI) software were used for evaluations. Segmentation process was not applied to the images of olives during the image analysis of the fruit and stones as location as LabVIEW Vision Assistant v2013 software gave us a chance for precision measurement without the need for any segmentation process. After completing all surface segmentation processes, LabVIEW Vision Assistant v2013 (NI) software was used for measuring the dimensions of olives. The length and width data were collected from the digital pictures of olive stones (Fig. 5). LabVIEW Vision Assistant v2013 (NI) was used for measuring the length and width dimensions of olive stones and fruits (Fig. 6). Image J (NIH) software was used for determining pixel counts. The monochrome pixel values were counted from the images. Therefore, these counts were changed into ‘%’ values for evaluating the olive cultivars. The aim of converting counts into ‘%’ was the standardization of pixel counts for comparison of each picture. In this study, the IOC and EU determination methods were used for the experimentations. Özilbey, 2011 provided information on manual olive identification methods in his book and outlined the characterstics of Turkey’s olive cultivars. These methods are based on morphological and pomological measurements methods, i.e. olive tree measurements, leaf measurements, flower measurements, fruit dimension and olive stone evaluations. In addition to this, other studies in literature were found and analyzed. For the stones’ monochrome pixel distributions, the length and width of stones were investigated (Özilbey, 2011).
2. Material and method 2.1. Harvest of fruit samples In this study, Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre olive cultivars were used for the experiments (Fig. 1). All cultivars were obtained from Instituto Andaluz de Investigación y Formación Agroalimentaria y Pesquera de Andalucía (IFAPA), Cordoba, Spain (37°51′42.7″N, 4°45′56.7″W). Olive samples were harvested from 5 different trees randomly for each olive cultivars in November 2015. A total of 100 olive fruit samples were harvested from each cultivar. During the period of the experimentations, olive cultivars placed in cold storage (+4C°, 80% humidity) in The Faculty of Agriculture, Department of Horticulture in Ankara University, Turkey.
287
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 1. Olive cultivars obtained from Instituto Andaluz de Investigación y Formación Agroalimentaria y Pesquera de Andalucía (IFAPA), Cordoba, Spain.
Fig. 2. Front side (a), left side (b), handle hole (c) and tip side (d) of images of olive fruit.
comparison between all images. Thus, monochrome color values were used for determining stone patterns for each olive cultivar.
2.4. Olive stones removing process and stone measurements Data histograms of stone images were evaluated. For this aim, new images were obtained from the original images in the same resolution but without calibration plates. After obtaining new images, Image J (NIH) software was used for determining the pixel frequency of color values between 0 and 255 (Fig. 7). Then, these frequencies were converted into ‘%’ values for the morphological evaluation of olive cultivars. The aim of this conversion was to make a substantial and equal
2.5. Statistical analysis Artificial Neural Network analysis was employed to evaluate the data and determine the length, width and color data results obtained from different fruits and stones. The H0 hypothesis was: There is no difference between the length, 288
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 3. Images of olive fruit and stone, shown with a calibration plate.
width, and color data of whole fruits and stone populations according to variance analysis. The H1 hypothesis was: At least one difference exists between the length, width and color data from the whole fruit and stone population according to variance analysis. In addition, different classification techniques were applied to the olive stone color value data with the help of SPSS v22. Clementine v12 was used as a data mining software package from SPSS. Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre Olive were coded as olive types from 1 to 7 respectively for the statistical analysis process. The effects are described and detailed below. Artificial Neural Networks: Six different training methods were used in Clementine v12 for building neural network examples. These methods are given below:
Fig. 5. Obtained information from the digital pictures of fruit and olive stones.
• RBFN method • Exhaustive prune method
• Quick method • Dynamic method • Multiple method • Prune method
We have compared six different types of training methods to predict olive classification from independent variables (color codes and fruit dimensions). The implementation of Artificial Neural Network with Fig. 4. The removal of an olive stone using a hand tool (a), cleaning of the removed olive stone using a knife (b), cleaning of the removed olive stones with water (c), stone samples cleaned with a solution (d).
289
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 6. LabVIEW Vision Assistant v2013 (NI) software used for olive cultivars determination.
3. Results and discussion Today, modern techniques and approaches eliminate these problems. The widely used image analysis techniques in agricultural areas are the solution to this matter. In this way, olive cultivar determination was done without the need for expensive processes or expert input. In this study, image processing and analysis techniques were employed for creating an olive database. As our study showed, this database can also be used for biotechnological researches since the morphological data about olives we gathered (such as monochrome color sequences, or olive stone, fruit width − length) give 90% accuracy and the results are confirmed. Diaz et al. (2000) used four algorithms for comparing olive selection methods when analyzing human and machine vision. According to the answers they obtained from the first algorithm, the machine vision system failure was observed as 53%, and human failure was observed as 52.5%; for the second algorithm 63% and 42.5%; for the third algorithm 57% and 11. 4%; and for the fourth algorithm 14% and 15%, respectively. Bari et al. (2003) worked on olive characteristics and they stated that these features are 90% accurate in identifying olives. Diaz et al. (2004) worked on olive classification and they stated that, according to the results, it is possible to classify olives at a rate of 90% based on artificial neural networks. Vanloot et al. (2014) worked on picture analysis and metric evaluations of agricultural products such as olive stones and they emphasized that the best model considers all the data obtained from the front and profile pictures and gives 100% correct classification. In our study, we also took pictures from four different sides of olives as seen in Fig. 2.
Fig. 7. The determination of pixel frequency values of an olive stone using Image J (NIH) software.
Clementine v12 module is given below (Fig. 8). In order to prevent over-training problems that can occur within the neural networks, a randomly selected proportion of the training data was used to train the network. We selected 90% cases for training and, then, 10% cases were used for testing. Thus, six different neural network training methods in Clementine v12 were applied for the classification of olive types. Preparation and testing result according to different training methods presented by Clementine v12 can be seen in the results and discussion section of this study. The results of validation items are summarized in the results and discussion section.
Fig. 8. The implementation of Artificial Neural Network with Clementine v12.
290
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Table 1 Preparation and testing solutions.
Table 3 General classification accuracy (74.81%) with confusion matrix for Dynamic Method.
Method Quick
Correct Wrong Correct Wrong Correct Wrong Correct Wrong Correct Wrong Correct Wrong
Dynamic Multiple Prune RBFN Exhaustive Prune
Preparation
Testing
N- Variety
478 (77.10%) 142 (22.90%) 461 (74.35%) 159 (25.65%) 113 (18.23%) 124 (12.51%) 538 (86.77%) 82 (13.23%) 472 (76.13%) 148 (23.87%) 560 (90.32%) 60 (9.68%)
60 (75.00%) 20 (25.00%) 62 (77.50%) 18 (22.50%) 66 (82.50%) 14 (17.50%) 68 (85.00%) 12 (15.00%) 60 (75.00%) 20 (25.00%) 71 (88.75%) 9 (11.25%)
Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
71 16 0 0 0 0 0
18 83 0 0 4 0 1
0 0 67 13 0 0 0
0 0 31 71 0 17 0
0 1 0 0 64 15 0
4 0 2 15 32 68 0
7 0 0 1 0 0 99
Table 4 General classification accuracy (81.66%) with confusion matrix for Multiple Method.
The effects of different sides of olive images as used for olive identification can be seen in Table 1 according to Artificial Neural Network classification results. Al Ibrahem et al. (2008) emphasized that the width − length ratio of olive stones could be used for identifying olive cultivars with a 60% success rate. We also presented the identification results of the width − length measurements of the experimented olives in Tables 2–7. Riquelme et al. (2008) stated that they achieved a 75% – 97% success rate in their study according to the results of color defects and structural characteristics. Mendoza et al. (2006) worked on different color spaces such as RGB, HSV, and L * a * b *, and they stated that the colors of fruits and vegetables were measured by a computerized imaging system. They also stress that L * a * b * color space is the best color space. According to Table 1, the most successful classification results are held when the exhaustive prune training method in artificial neural networks is applied. This result supports Diaz et al.’s (2004) study on olive classification. Furthermore, we obtained a higher accuracy classification percentage in our study. We classified olives testing data with a rate success of 90% by using the exhaustive prune training method in artificial neural networks. In order to calculate the general identification accuracy, confusion matrixes were calculated for each training method (Tables 2–7). According to Confusion matrix for Exhaustive Prune Method, olive type = 1 is classified correctly at the level 83%, olive type = 2 (88%), olive type = 3 (98%), olive type = 4 (91%), olive type = 5 (89%), olive type = 6 (83%), olive type = 7 (99%), respectively. When combined with the extant literature, our study shows that this identification method is cheap, fast, and reliable with a high degree of accuracy and, therefore, an alternative to genetic applications for identifying olive cultivars.
N- Variety
4
5
6
7
1 2 3 4 5 6 7
46 12 0 1 0 0 0
50 86 0 0 0 0 0
0 0 92 12 0 0 0
0 0 6 78 0 22 0
3 2 1 1 93 35 0
1 0 1 8 7 43 0
0 0 0 0 0 0 100
4
5
6
7
1 2 3 4 5 6 7
76 17 0 0 0 1 0
20 81 0 0 1 0 0
0 0 95 13 0 0 0
0 0 3 71 0 12 0
0 2 0 1 80 17 0
4 0 2 15 19 70 0
0 0 0 0 0 0 100
Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
84 13 0 1 0 2 0
16 86 0 0 0 1 0
0 0 97 2 0 0 0
0 0 3 92 1 24 0
0 1 0 1 93 18 1
0 0 0 4 6 55 0
0 0 0 0 0 0 99
N- Variety Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
70 34 0 0 0 3 0
26 64 0 0 0 0 0
0 0 87 18 0 0 1
0 0 12 74 1 23 0
0 2 0 0 79 15 0
3 0 1 7 20 59 0
1 0 0 1 0 0 99
Faculty of Agriculture, Ankara University and, then, included the other cultivars around world (Beyaz and Öztürk, 2016). The samples of the first study were obtained from Turkey’s National Olive Gene Bank in Izmir-Turkey for the PhD thesis. The results obtained from the first study have been already published in Turkish Journal of Agriculture and Forestry. This study is the second step of the study conducted on world olive cultivars. Therefore, Cordoba-Spain World Olive Gene Bank was selected for conducting study. By selecting this gene bank, it was aimed to make the results of this study international. In addition to these, Spanish olive varieties contribute to the international status of this study. However, in future, this study will be conducted for the other cultivars in Spain World Olive Gene Bank as some cultivars yield
N- Variety 3
3
Table 6 General classification accuracy (76.00%) with confusion matrix for RBFN Method.
Table 2 General classification accuracy (76.85%) with confusion matrix for Quick Method.
2
2
N- Variety
The first step of this study was conducted as a PhD thesis in Department of Agricultural Machinery and Technologies Engineering in
1
1
Table 5 General classification accuracy (86.57%) with confusion matrix for Prune Method.
4. Conclusion
Variety
Variety
291
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
starting point of a web-based identification system. In addition to these, panoramic images of stones can provide more information about olive cultivars because of the surface rise. It can be said that different statistical analysis may provide more useful outcomes for the identification of the various olives around world.
Table 7 General classification accuracy (90.14%) with confusion matrix for Exhaustive prune Method. N- Variety Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
83 11 0 0 0 1 0
16 88 0 0 0 0 0
0 0 98 1 0 0 0
0 0 1 91 0 8 0
1 0 0 0 89 8 1
0 1 1 8 11 83 0
0 0 0 0 0 0 99
References Özilbey, N., 2011. Zeytin Çeşitlerimiz. Sidas Medya Ltd., İzmir, Turkey. Al Ibrahem, A., Bari, A., Rashed, M.M., 2008. Olive genetic diversity of palmyra under threat. Acta Hortic. 791, 143–148. Babcock, B.A., Clemens, R., 2004. Geographical indications and property rights: protecting value-added agricultural products. MATRIC Briefing Paper 04-MBP 7. Iowa State University, Ames, IA (USA). Bari, A., Martın, A., Boulouha, B., Gonzales-Andujar, J.L., Barranco, D., Ayad, G., Padulosı, S., 2003. Use of fractals and moments to describe olive cultivars. J. Agr. Sci. 141, 63–71. BEYAZ, A., Öztürk, R., 2016. Identification of olive cultivars using image processing techniques. Turk. J. Agric. For. 40, 671–683. http://dx.doi.org/10.3906/tar-150495. © TÜBİTAK. Diaz, R., Faus, G., Blasco, M., Blasco, J., Molto, E., 2000. The application of a fast algorithm for the classification of olives by machine vision. Food Res. Int. 33, 305–309. Diaz, R., Gil, L., Serrano, C., Blasco, M., Molto, E., Blasco, J., 2004. Comparison of three algorithms in the classification of table olives by means of computer vision. J. Food Eng. 61, 101–107. Gurdeniz, G., Tokatli, F., Ozen, B., 2007. Differentiation of mixtures of monovarietal olive oils with mid-infrared spectroscopy and chemometrics. Eur. J. Lipid Sci. Technol. 109, 1194–1202. Mendoza, F., Dejmek, P., Aguilera, J.M., 2006. Calibrated color measurements of agricultural foods using image analysis. Postharvest Biol. Technol. 41, 285–295. Moreda, G.P., Muñoz, M.A., Ruiz-Altisent, M., Perdigones, A., 2012. Shape determination of horticultural produce using two-dimensional computer vision. J. Food Eng. 108, 245–261. Ozen, B., Tokatli, F., Korel, F., 2005. Emerging topics in olive oil research: determination of geographical origin and adulteration. In: Olive Oil and Olive-Pomace Oil Symposium & Exhibition. Izmir (Turkey). Riquelme, M.T., Barreiro, P., Ruiz-Altisent, M., Valero, C., 2008. Olive classification according to external damage using image analysis. J. Food Eng. 87, 371–379. Sakar Çakır, E., 2009. Selection Based Breeding and Genetic Characterization of Selected Olive (Olea Europaea L.) Genotypes from Adıyaman, Mardin, Siirt, Şanlıurfa and Şırnak Provinces. Ph.D. Ankara University, Ankara, Turkey (thesis in Turkish with an abstract in English). Turkish Statistical Institute Olive production, 1988–2014. Web page: http://www.tuik. gov.tr/, Access date: 29 January 2017. Vanloot, P., Bertrand, D., Pinatel, C., Artaud, J., Dupuy, N., 2014. Artificial vision and chemometrics analyses of olive stones for varietal identification of five French cultivars. Comput. Electron. Agric. 102, 98–105.
positive results and some not. Therefore, the study including all these cultivars will shed a real and accurate light on this issue. As monocultivar olive oil is becoming important and the number of high quality olive oil producers is decreasing, cultivar identification is becoming increasingly important for growers, producers and exporters. However, monocultivar is also important for table olive producers due to the different performance of fruits of each cultivars during processing. Therefore, it is important for growers, processors and raw material suppliers to identify the cultivars in olive oil and table olive production. As a result of this experimental study, it is seen that the measurement method adopted in this study (expert method) can be used effectively for olive identification. Genetic identification methods provide more detailed information about olive cultivars, but this expert method is a fast, dependable and cheap alternative to identification by experts and other different identification methods. The regions included in this study are limited to defined local olive cultivars of SPAIN. Furthermore, the primary identifications were evaluated from profile pictures of fruits and stones. Some statistical analyses mentioned in literature were conducted for the local olive cultivar identification. There are many different olive cultivars around the world. This expert method can be used for these various cultivars and, then, an olive cultivar database can be created. This expert method offers a future vision of web-based databases of olive cultivars, while it serves as a
292
Contents lists available at ScienceDirect
Scientia Horticulturae journal homepage: www.elsevier.com/locate/scihorti
Identification of some spanish olive cultivars using image processing techniques
MARK
⁎
Abdullah Beyaza, , Mücahit Taha Özkayab, Duygu İçenc a b c
Department of Agricultural Machinery and Technologies Engineering, Faculty of Agriculture, Ankara University, Ankara, Turkey Department of Horticulture, Faculty of Agriculture, Ankara University, Ankara, Turkey Department of Statistics, Faculty of Science, Hacettepe University, Ankara, Turkey
A R T I C L E I N F O
A B S T R A C T
Keywords: Artificial vision Image analysis Olive fruit Olive stone Varietal identification
The aim of this research was to identify some Spanish olive cultivars using image processing techniques. For this purpose, Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre Olive cultivars were identified utilizing the image processing and analysis techniques. Therefore, images of olives taken as 300 dpi with the 2896 × 1944 pixels, were captured using a DSLR camera, and evaluations of pixels were used for considering the pixel distribution and dimension measurements. LabVIEW Vision Assistant v2013 (NI) and Image j (NIH) software were used for image analysis procedures. Artificial Neural Network analysis were used to assess information of the length, width and color data results obtained from the fruits and stones (olive stones). All cultivars were identified. In addition, different classification techniques were applied to the olive stone and fruit data with the help of SPSS v22. Clementine v12 was used as a data mining software package from SPSS. The cultivars were identified 90% from dimensions with Artificial Neural Networks.
1. Introduction Olive is an important agricultural product for Turkey and the world. According to the FAO statistics, Turkey ranked as the 4th country in terms of olive production with an income of 2783 million US Dollars. It is known that there are 171,992 olive trees (TSI, 2015) in Turkey. In addition to this, Turkey has an olive production capacity of 1,700,000 tons in a year (TSI, 2015). Olive (Olea europaea L.) production is of a great economic importance for Turkey as it is located in olive production zone. The geographical positions of Turkey have been proved to be suitable in terms of olive production as the overall climate is dominated greatly by Mediterranean climate (i.e. Aegean, Marmara, Mediterranean, Southeast Anatolia, Black Sea regions). The remainders can be observed in olive cultivars due to elements such as widespread region, differences between the climate conditions of the regions and crosspollination. Therefore, it is of significant importance to define the type of the surface patterns of olive cultivars that are widespread in vast parts of the world. In this respect, many studies and researches needs to be conducted for improving olive production. A common problem encountered in studies conducted is that the forms of obtained data are insufficient for the determination of olive cultivars. The reason for this is that the characteristics of olive cultivars depend on the ecological conditions. The phenotypic and genotypic origin, molecular marker
⁎
Corresponding author. E-mail addresses: [email protected], [email protected] (A. Beyaz).
http://dx.doi.org/10.1016/j.scienta.2017.06.041 Received 29 March 2017; Received in revised form 28 May 2017; Accepted 14 June 2017 0304-4238/ © 2017 Elsevier B.V. All rights reserved.
researches cannot provide accurate olive cultivar determination data (Sakar Çakır, 2009) as the full genome sequences of all olive cultivars have not been defined yet. The cultivar determination process has started with the identification of cultivars pomological information. As a consequence of these reasons, a new study was conducted on some Spain cultivars after conducting some Turkish cultivars evaluations. Many studies and researches were conducted on the identification of olive cultivars around the world. For example, Diaz et al. (2000) produced four different algorithms for a machine vision system and applied these algorithms for the identification of olives. In addition to this, they compared human selected olive cultivars with computer vision algorithms. Bari et al. (2003) concentrated on identifying the characteristics of the olives in their studies and they stated that structural characteristics are significant factors in the identification of olives. Diaz et al. (2004) worked on the classification of olives based on the fruit surface defects in different quality categories. Mendoza et al. (2006) worked on different color spaces like RGB, HSV, and L * a * b *, and they stated that the colors of fruits and vegetables were measured by a computerized imaging system that can be applied for the determination of product status. Riquelme et al. (2008) conducted studies on the color defects and structural parameters of olive cultivars. Al Ibrahem et al. (2008) conducted studies on different olive
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
2.2. Imaging system
cultivars (19) using manual measurement and image analysis techniques. Moreda et al. (2012) worked on the shape determination of horticultural products using computer vision technology. They also stated that the computer vision measurements have become a proven and reliable tool for describing the product shape. Vanloot et al. (2014) emphasized that image analysis and metric evaluations of agricultural products like olive stones for determining the varietal origin require specialists, who cannot always conclude with certainty, due to the large number of cultivars identified. Olive production is also significant for many industrial applications as olive identification methods will affect the output characteristics of the olives. In this regard, the aim of this study was to distinguish the characteristics of some olive cultivars in Spain’s World Collection using image processing and analysis techniques. In addition to this, the efficiency of image analysis methods was also investigated in this study. Olives are generally divided into two groups; olive oil and table olive. Olives have a highly variable chemical and physical composition, but these variabilities mostly depend on four major factors. These four factors are cultivars, ecology, cultural practices, and processing. For example, a cultivar which is called dual purpose can be used as both table olive and olive oil. Therefore, cultivar is one of the most important factors for determining the quality of these two products (Gurdeniz et al., 2007). Mono variety is important and indispensable factor for table olive processing. In recent years, mono varietal olive oil has gained importance because of the health benefits and sensory qualities of each cultivar (Gurdeniz et al., 2007). Table olive and olive oil produced from olives of just one cultivar (mono varietal) obtained form one geographical region are becoming increasingly important in the market (Gurdeniz et al., 2007). The documentation of the geographical source of table olive and olive oil is becoming increasingly important as the authenticity and quality issues can be linked with a special region and cultivar. The Denomination of Origin (Protected Denomination of Origin (PDO) and Protected Geographical Indication (PGI)) is a certification program established in Europe for the protection of high-quality agricultural products (Babcock and Clemens, 2004). There are many analytical methods used to determine the varietal origin of olive. Olive stone (pit) is one of the fingerprinting organ used for varietal identification (Ozen et al., 2005). The aim of this study was to identify olive cultivars from their characteristics using image processing and analysis techniques. In this study, the methods of International Olive Council (IOC) and European Union (EU) were adopted for determining olive cultivar and,then, stone, fruit, and leaf data were used for identifying olive cultivars. Moreover, image processing and analysis techniques were employed to accomplish this aim.
After the harvest, 100 samples were taken randomly from each olive cultivar. These samples were photographed from 4 different views − from the front, handle hole, left and tip sides (Fig. 2). These imaging sides are also employed as standard classification views by Kemalpaşa Olive Research Institute for the identification of national collection of Turkey’s olive cultivars. In aggregate, 4400 digital images were captured from 1100 olive fruits for evaluations. A macro capture tripod was set at a distance of 40 centimeter away from the olives to obtain these digital pictures. A Nikon D300 s body with an 18–140 mm zoom lens was used for general purpose imaging and a Nikon D800 with a 105 mm macro lens was used for macro captures. The captured images were stored as JPEG files. All images were captured in 2896 × 1944 pixel dimension and at 300 dpi resolution. Olive fruit and stone images were captured one by one for evaluations. 5 × 5 mm calibration plate was used with a white backdrop for making precision calibration (Fig. 3). First, the primary image capturing problems such as shaded regions and image focus clarity were worked out. Different lenses and flashes were employed for working out the problems encountered during experimentation. The problem of shaded regions was worked out by using a ring flash. Image focus problem was also worked out by using a 105 mm macro lens. After overcoming these problems related to digital imaging, olive cultivars were harvested from the national collection. Then, each olive fruit was photographed, put in numbered plastic bags (5 × 5 cm) and prepared for the following step (olive stone removal step). After completing the picture capturing process, olive stones were removed from fruits. The olive stones were removed from the fruits with an extractor and, then, the olive stone pattern structures were determined. After cleaning the olive stones with the help of a knife, the stones were washed. Then, the stones were stored in plastic containers holding in a 10% bleach solution for 15 h. As a final step, olive stones were stored at − 4 °C to prevent cracking because of physiological activity. Removed olive stones were put in numbered plastic bags (5 × 5 cm). The processes applied to olives are demonstrated in Fig. 4.
2.3. Image processing and determination of sample proportions Image J (NIH) and LabVIEW Vision Assistant v2013 (NI) software were used for evaluations. Segmentation process was not applied to the images of olives during the image analysis of the fruit and stones as location as LabVIEW Vision Assistant v2013 software gave us a chance for precision measurement without the need for any segmentation process. After completing all surface segmentation processes, LabVIEW Vision Assistant v2013 (NI) software was used for measuring the dimensions of olives. The length and width data were collected from the digital pictures of olive stones (Fig. 5). LabVIEW Vision Assistant v2013 (NI) was used for measuring the length and width dimensions of olive stones and fruits (Fig. 6). Image J (NIH) software was used for determining pixel counts. The monochrome pixel values were counted from the images. Therefore, these counts were changed into ‘%’ values for evaluating the olive cultivars. The aim of converting counts into ‘%’ was the standardization of pixel counts for comparison of each picture. In this study, the IOC and EU determination methods were used for the experimentations. Özilbey, 2011 provided information on manual olive identification methods in his book and outlined the characterstics of Turkey’s olive cultivars. These methods are based on morphological and pomological measurements methods, i.e. olive tree measurements, leaf measurements, flower measurements, fruit dimension and olive stone evaluations. In addition to this, other studies in literature were found and analyzed. For the stones’ monochrome pixel distributions, the length and width of stones were investigated (Özilbey, 2011).
2. Material and method 2.1. Harvest of fruit samples In this study, Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre olive cultivars were used for the experiments (Fig. 1). All cultivars were obtained from Instituto Andaluz de Investigación y Formación Agroalimentaria y Pesquera de Andalucía (IFAPA), Cordoba, Spain (37°51′42.7″N, 4°45′56.7″W). Olive samples were harvested from 5 different trees randomly for each olive cultivars in November 2015. A total of 100 olive fruit samples were harvested from each cultivar. During the period of the experimentations, olive cultivars placed in cold storage (+4C°, 80% humidity) in The Faculty of Agriculture, Department of Horticulture in Ankara University, Turkey.
287
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 1. Olive cultivars obtained from Instituto Andaluz de Investigación y Formación Agroalimentaria y Pesquera de Andalucía (IFAPA), Cordoba, Spain.
Fig. 2. Front side (a), left side (b), handle hole (c) and tip side (d) of images of olive fruit.
comparison between all images. Thus, monochrome color values were used for determining stone patterns for each olive cultivar.
2.4. Olive stones removing process and stone measurements Data histograms of stone images were evaluated. For this aim, new images were obtained from the original images in the same resolution but without calibration plates. After obtaining new images, Image J (NIH) software was used for determining the pixel frequency of color values between 0 and 255 (Fig. 7). Then, these frequencies were converted into ‘%’ values for the morphological evaluation of olive cultivars. The aim of this conversion was to make a substantial and equal
2.5. Statistical analysis Artificial Neural Network analysis was employed to evaluate the data and determine the length, width and color data results obtained from different fruits and stones. The H0 hypothesis was: There is no difference between the length, 288
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 3. Images of olive fruit and stone, shown with a calibration plate.
width, and color data of whole fruits and stone populations according to variance analysis. The H1 hypothesis was: At least one difference exists between the length, width and color data from the whole fruit and stone population according to variance analysis. In addition, different classification techniques were applied to the olive stone color value data with the help of SPSS v22. Clementine v12 was used as a data mining software package from SPSS. Lechin De Granada, Arbequina, Picual, Verdial De V-M, Picudo, Hojiblanca and Empeltre Olive were coded as olive types from 1 to 7 respectively for the statistical analysis process. The effects are described and detailed below. Artificial Neural Networks: Six different training methods were used in Clementine v12 for building neural network examples. These methods are given below:
Fig. 5. Obtained information from the digital pictures of fruit and olive stones.
• RBFN method • Exhaustive prune method
• Quick method • Dynamic method • Multiple method • Prune method
We have compared six different types of training methods to predict olive classification from independent variables (color codes and fruit dimensions). The implementation of Artificial Neural Network with Fig. 4. The removal of an olive stone using a hand tool (a), cleaning of the removed olive stone using a knife (b), cleaning of the removed olive stones with water (c), stone samples cleaned with a solution (d).
289
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Fig. 6. LabVIEW Vision Assistant v2013 (NI) software used for olive cultivars determination.
3. Results and discussion Today, modern techniques and approaches eliminate these problems. The widely used image analysis techniques in agricultural areas are the solution to this matter. In this way, olive cultivar determination was done without the need for expensive processes or expert input. In this study, image processing and analysis techniques were employed for creating an olive database. As our study showed, this database can also be used for biotechnological researches since the morphological data about olives we gathered (such as monochrome color sequences, or olive stone, fruit width − length) give 90% accuracy and the results are confirmed. Diaz et al. (2000) used four algorithms for comparing olive selection methods when analyzing human and machine vision. According to the answers they obtained from the first algorithm, the machine vision system failure was observed as 53%, and human failure was observed as 52.5%; for the second algorithm 63% and 42.5%; for the third algorithm 57% and 11. 4%; and for the fourth algorithm 14% and 15%, respectively. Bari et al. (2003) worked on olive characteristics and they stated that these features are 90% accurate in identifying olives. Diaz et al. (2004) worked on olive classification and they stated that, according to the results, it is possible to classify olives at a rate of 90% based on artificial neural networks. Vanloot et al. (2014) worked on picture analysis and metric evaluations of agricultural products such as olive stones and they emphasized that the best model considers all the data obtained from the front and profile pictures and gives 100% correct classification. In our study, we also took pictures from four different sides of olives as seen in Fig. 2.
Fig. 7. The determination of pixel frequency values of an olive stone using Image J (NIH) software.
Clementine v12 module is given below (Fig. 8). In order to prevent over-training problems that can occur within the neural networks, a randomly selected proportion of the training data was used to train the network. We selected 90% cases for training and, then, 10% cases were used for testing. Thus, six different neural network training methods in Clementine v12 were applied for the classification of olive types. Preparation and testing result according to different training methods presented by Clementine v12 can be seen in the results and discussion section of this study. The results of validation items are summarized in the results and discussion section.
Fig. 8. The implementation of Artificial Neural Network with Clementine v12.
290
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
Table 1 Preparation and testing solutions.
Table 3 General classification accuracy (74.81%) with confusion matrix for Dynamic Method.
Method Quick
Correct Wrong Correct Wrong Correct Wrong Correct Wrong Correct Wrong Correct Wrong
Dynamic Multiple Prune RBFN Exhaustive Prune
Preparation
Testing
N- Variety
478 (77.10%) 142 (22.90%) 461 (74.35%) 159 (25.65%) 113 (18.23%) 124 (12.51%) 538 (86.77%) 82 (13.23%) 472 (76.13%) 148 (23.87%) 560 (90.32%) 60 (9.68%)
60 (75.00%) 20 (25.00%) 62 (77.50%) 18 (22.50%) 66 (82.50%) 14 (17.50%) 68 (85.00%) 12 (15.00%) 60 (75.00%) 20 (25.00%) 71 (88.75%) 9 (11.25%)
Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
71 16 0 0 0 0 0
18 83 0 0 4 0 1
0 0 67 13 0 0 0
0 0 31 71 0 17 0
0 1 0 0 64 15 0
4 0 2 15 32 68 0
7 0 0 1 0 0 99
Table 4 General classification accuracy (81.66%) with confusion matrix for Multiple Method.
The effects of different sides of olive images as used for olive identification can be seen in Table 1 according to Artificial Neural Network classification results. Al Ibrahem et al. (2008) emphasized that the width − length ratio of olive stones could be used for identifying olive cultivars with a 60% success rate. We also presented the identification results of the width − length measurements of the experimented olives in Tables 2–7. Riquelme et al. (2008) stated that they achieved a 75% – 97% success rate in their study according to the results of color defects and structural characteristics. Mendoza et al. (2006) worked on different color spaces such as RGB, HSV, and L * a * b *, and they stated that the colors of fruits and vegetables were measured by a computerized imaging system. They also stress that L * a * b * color space is the best color space. According to Table 1, the most successful classification results are held when the exhaustive prune training method in artificial neural networks is applied. This result supports Diaz et al.’s (2004) study on olive classification. Furthermore, we obtained a higher accuracy classification percentage in our study. We classified olives testing data with a rate success of 90% by using the exhaustive prune training method in artificial neural networks. In order to calculate the general identification accuracy, confusion matrixes were calculated for each training method (Tables 2–7). According to Confusion matrix for Exhaustive Prune Method, olive type = 1 is classified correctly at the level 83%, olive type = 2 (88%), olive type = 3 (98%), olive type = 4 (91%), olive type = 5 (89%), olive type = 6 (83%), olive type = 7 (99%), respectively. When combined with the extant literature, our study shows that this identification method is cheap, fast, and reliable with a high degree of accuracy and, therefore, an alternative to genetic applications for identifying olive cultivars.
N- Variety
4
5
6
7
1 2 3 4 5 6 7
46 12 0 1 0 0 0
50 86 0 0 0 0 0
0 0 92 12 0 0 0
0 0 6 78 0 22 0
3 2 1 1 93 35 0
1 0 1 8 7 43 0
0 0 0 0 0 0 100
4
5
6
7
1 2 3 4 5 6 7
76 17 0 0 0 1 0
20 81 0 0 1 0 0
0 0 95 13 0 0 0
0 0 3 71 0 12 0
0 2 0 1 80 17 0
4 0 2 15 19 70 0
0 0 0 0 0 0 100
Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
84 13 0 1 0 2 0
16 86 0 0 0 1 0
0 0 97 2 0 0 0
0 0 3 92 1 24 0
0 1 0 1 93 18 1
0 0 0 4 6 55 0
0 0 0 0 0 0 99
N- Variety Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
70 34 0 0 0 3 0
26 64 0 0 0 0 0
0 0 87 18 0 0 1
0 0 12 74 1 23 0
0 2 0 0 79 15 0
3 0 1 7 20 59 0
1 0 0 1 0 0 99
Faculty of Agriculture, Ankara University and, then, included the other cultivars around world (Beyaz and Öztürk, 2016). The samples of the first study were obtained from Turkey’s National Olive Gene Bank in Izmir-Turkey for the PhD thesis. The results obtained from the first study have been already published in Turkish Journal of Agriculture and Forestry. This study is the second step of the study conducted on world olive cultivars. Therefore, Cordoba-Spain World Olive Gene Bank was selected for conducting study. By selecting this gene bank, it was aimed to make the results of this study international. In addition to these, Spanish olive varieties contribute to the international status of this study. However, in future, this study will be conducted for the other cultivars in Spain World Olive Gene Bank as some cultivars yield
N- Variety 3
3
Table 6 General classification accuracy (76.00%) with confusion matrix for RBFN Method.
Table 2 General classification accuracy (76.85%) with confusion matrix for Quick Method.
2
2
N- Variety
The first step of this study was conducted as a PhD thesis in Department of Agricultural Machinery and Technologies Engineering in
1
1
Table 5 General classification accuracy (86.57%) with confusion matrix for Prune Method.
4. Conclusion
Variety
Variety
291
Scientia Horticulturae 225 (2017) 286–292
A. Beyaz et al.
starting point of a web-based identification system. In addition to these, panoramic images of stones can provide more information about olive cultivars because of the surface rise. It can be said that different statistical analysis may provide more useful outcomes for the identification of the various olives around world.
Table 7 General classification accuracy (90.14%) with confusion matrix for Exhaustive prune Method. N- Variety Variety
1
2
3
4
5
6
7
1 2 3 4 5 6 7
83 11 0 0 0 1 0
16 88 0 0 0 0 0
0 0 98 1 0 0 0
0 0 1 91 0 8 0
1 0 0 0 89 8 1
0 1 1 8 11 83 0
0 0 0 0 0 0 99
References Özilbey, N., 2011. Zeytin Çeşitlerimiz. Sidas Medya Ltd., İzmir, Turkey. Al Ibrahem, A., Bari, A., Rashed, M.M., 2008. Olive genetic diversity of palmyra under threat. Acta Hortic. 791, 143–148. Babcock, B.A., Clemens, R., 2004. Geographical indications and property rights: protecting value-added agricultural products. MATRIC Briefing Paper 04-MBP 7. Iowa State University, Ames, IA (USA). Bari, A., Martın, A., Boulouha, B., Gonzales-Andujar, J.L., Barranco, D., Ayad, G., Padulosı, S., 2003. Use of fractals and moments to describe olive cultivars. J. Agr. Sci. 141, 63–71. BEYAZ, A., Öztürk, R., 2016. Identification of olive cultivars using image processing techniques. Turk. J. Agric. For. 40, 671–683. http://dx.doi.org/10.3906/tar-150495. © TÜBİTAK. Diaz, R., Faus, G., Blasco, M., Blasco, J., Molto, E., 2000. The application of a fast algorithm for the classification of olives by machine vision. Food Res. Int. 33, 305–309. Diaz, R., Gil, L., Serrano, C., Blasco, M., Molto, E., Blasco, J., 2004. Comparison of three algorithms in the classification of table olives by means of computer vision. J. Food Eng. 61, 101–107. Gurdeniz, G., Tokatli, F., Ozen, B., 2007. Differentiation of mixtures of monovarietal olive oils with mid-infrared spectroscopy and chemometrics. Eur. J. Lipid Sci. Technol. 109, 1194–1202. Mendoza, F., Dejmek, P., Aguilera, J.M., 2006. Calibrated color measurements of agricultural foods using image analysis. Postharvest Biol. Technol. 41, 285–295. Moreda, G.P., Muñoz, M.A., Ruiz-Altisent, M., Perdigones, A., 2012. Shape determination of horticultural produce using two-dimensional computer vision. J. Food Eng. 108, 245–261. Ozen, B., Tokatli, F., Korel, F., 2005. Emerging topics in olive oil research: determination of geographical origin and adulteration. In: Olive Oil and Olive-Pomace Oil Symposium & Exhibition. Izmir (Turkey). Riquelme, M.T., Barreiro, P., Ruiz-Altisent, M., Valero, C., 2008. Olive classification according to external damage using image analysis. J. Food Eng. 87, 371–379. Sakar Çakır, E., 2009. Selection Based Breeding and Genetic Characterization of Selected Olive (Olea Europaea L.) Genotypes from Adıyaman, Mardin, Siirt, Şanlıurfa and Şırnak Provinces. Ph.D. Ankara University, Ankara, Turkey (thesis in Turkish with an abstract in English). Turkish Statistical Institute Olive production, 1988–2014. Web page: http://www.tuik. gov.tr/, Access date: 29 January 2017. Vanloot, P., Bertrand, D., Pinatel, C., Artaud, J., Dupuy, N., 2014. Artificial vision and chemometrics analyses of olive stones for varietal identification of five French cultivars. Comput. Electron. Agric. 102, 98–105.
positive results and some not. Therefore, the study including all these cultivars will shed a real and accurate light on this issue. As monocultivar olive oil is becoming important and the number of high quality olive oil producers is decreasing, cultivar identification is becoming increasingly important for growers, producers and exporters. However, monocultivar is also important for table olive producers due to the different performance of fruits of each cultivars during processing. Therefore, it is important for growers, processors and raw material suppliers to identify the cultivars in olive oil and table olive production. As a result of this experimental study, it is seen that the measurement method adopted in this study (expert method) can be used effectively for olive identification. Genetic identification methods provide more detailed information about olive cultivars, but this expert method is a fast, dependable and cheap alternative to identification by experts and other different identification methods. The regions included in this study are limited to defined local olive cultivars of SPAIN. Furthermore, the primary identifications were evaluated from profile pictures of fruits and stones. Some statistical analyses mentioned in literature were conducted for the local olive cultivar identification. There are many different olive cultivars around the world. This expert method can be used for these various cultivars and, then, an olive cultivar database can be created. This expert method offers a future vision of web-based databases of olive cultivars, while it serves as a
292