Discrimination of 14 olive cultivars using morphological analysis and machine learning algorithms

Discrimination of 14 olive cultivars using morphological analysis and machine learning algorithms Konstantinos N. Blazakis 1 *, Danil Stupichev 1 , Maria Kosma 1 , Mohamad Ali Hassan El Chami 1 , Anastasia Apodiakou 1 , George Kostelenos 2 and Panagiotis Kalaitzis 1 * 1 Department of Horticultural Genetics and Biotechnology, Mediterranean Agronomic Institute of Chania (MAICh), Chania, Greece, 2 Kostelenos Olive Nurseries, Poros, Greece Traditional morphological analysis is a widely employed tool for the identification and discrimination of olive germplasm by using morphological markers which are monitored by subjective manual measurements that are labor intensive and time-consuming. Alternatively, an automated methodology can quantify the geometrical features of fruits, leaves and endocarps with high accuracy and efficiency in order to define their morphological characteristics. In this study, 24 characteristics for fruits, 16 for leaves and 25 for endocarps were determined and used in an automated way with basic classifiers combined with a meta-classsifier approach. This resulted to the discrimination of 14 olive cultivars utilizing data obtained from two consecutive olive growing periods. The cultivar classification algorithms were based on machine learning techniques. The 95% accuracy rate of the meta-classifier approach indicated that was an efficient tool to discriminate olive cultivars. The contribution of each morphological feature to cultivar discrimination was quantified, and the significance of each one was automatically detected in a quantitative way. The higher the contribution of each feature, the higher the significance for cultivar discrimination. The identification of most cultivars was guided by the features of both endocarps and fruits, while those of leaves were only efficient to identify the Kalamon cultivar. The combined use of morphological features of three olive organs might have an additive effect leading to higher capacity for discrimination of cultivars. The proposed methodology might be considered a phenomics tool for olive cultivar identification and discrimination in a wide range of applications including breeding. KEYWORDS olive cultivar identification, morphological analysis, machine learning, image analysis, olive fruit, olive leaf, olive endocarp Frontiers in Plant Science frontiersin.org01 OPEN ACCESS EDITED BY Svein Øivind Solberg, Inland Norway University of Applied Sciences, Norway REVIEWED BY Flemming Yndgaard, Nordic Genetic Resource Centre (NordGen), Sweden Raul De La Rosa, Spanish National Research Council (CSIC), Spain *CORRESPONDENCE Panagiotis Kalaitzis [email protected] Konstantinos N. Blazakis [email protected] RECEIVED 31 May 2024 ACCEPTED 15 July 2024 PUBLISHED 08 August 2024 CITATION Blazakis KN, Stupichev D, Kosma M, El Chami MAH, Apodiakou A, Kostelenos G and Kalaitzis P (2024) Discrimination of 14 olive cultivars using morphological analysis and machine learning algorithms. Front. Plant Sci. 15:1441737. doi: 10.3389/fpls.2024.1441737 COPYRIGHT ©2024Blazakis,Stupichev,Kosma,ElChami, Apodiakou, Kostelenos and Kalaitzis. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. TYPE Original Research PUBLISHED 08 August 2024 DOI 10.3389/fpls.2024.1441737 1 Introduction Olive (Olea europaea L. subsp. europaea var. europaea)isan important fruit tree crop in the Mediterranean Basin while its cultivation has expanded across the rest of the world, including the United States of America, Oceania, South Africa and Asia (Torres et al., 2017;Koubouris et al., 2019;Mousavi et al., 2019). Olive oil originating in the Mediterranean region accounts for more than 90% of global production and has significant socioeconomic importance for the European Union, since Spain, Italy, Greece and Portugal produce approximately 75% of the world’s olive oil supply (Paredes et al., 2019). The ability to identify and discriminate olive cultivars is important for the efficient management and exploitation of the available olive genetic resources and for breeding programs (Beiki et al., 2012;De Ollas et al., 2019). The identification, cataloguing and exploitation of germplasm collections comprising cultivars and accessions is performed by using morphological and molecular markers (Rallo et al., 2018). Molecular markers and new, advanced biotechnological platforms have been used for genetic diversity assessment and cultivar discrimination (D’Imperio et al., 2011;Muleo et al., 2016; Sebastiani and Busconi, 2017;Zhu et al., 2019;Gomez-Rodrıguez et al., 2021) while morphological markers continue to constitute the main approach for describing and discriminating olive germplasm, despite limitations such as the variability of environmental conditions, the age of trees, agronomical practices and phenological stage of trees (Trujillo et al., 2014;Blazakis et al., 2017). Molecular markers and new, advanced biotechnological platforms have been used for genetic diversity assessment and cultivar discrimination (Diaz et al., 2004;D’Imperio et al., 2011; Muleo et al., 2016;Sebastiani and Busconi, 2017;Zhu et al., 2019; Gomez-Rodrıguez et al., 2021). However, morphological markers continue to constitute the main approach for describing and discriminating olive germplasm, despite limitations such as the variability of environmental conditions, the age of trees, agronomical practices, and the phenological stage of trees (Trujillo et al., 2014;Blazakis et al., 2017). Most morphological studies are based on a simplified scheme that has been adopted by the International Union for the Protection of New Varieties of Plants (UPOV) which focuses on the morphological characteristics of leaves, fruits and endocarps (Trujillo et al., 2014). These characteristics have been widely used for descriptive purposes to distinguish olive cultivars (Barranco et al., 2000;Trujillo et al., 2014; Beyaz et al., 2017;Blazakis et al., 2017;Rallo et al., 2018;Koubouris et al., 2019). Currently, most morphological characterization of olive organs is performed by either time-consuming, laborintensive, manual measurements or by one of several stand-alone software-based methodologies (Riquelme et al., 2008;Rodrıguez et al., 2010;Orrù et al., 2013;Beyaz et al., 2017). Despite the extensive use of morphological characteristics of olive fruits, leaves and endocarps for the identification of cultivars, there is a lack of automated methodologies to further assist in the development of this field. Working in this direction, (Blazakis et al., 2017) an integrated image-based tool on automated methodology was developed in order to describe olive fruit, leaf and endocarp morphologies. The methodology quantifies many features of these organs based on strictly mathematical morphological parameters and provides accurate, objective numerical measurements of the olive organ morphology attributes in a more robust and efficient way. Currently, there are several olive cultivar identification methods that use chemical and genetic fingerprinting techniques. However, these methods require a high level of specialization, specific expensive infrastructure, and laboratory work. In contrast, userfriendly image-based methodologies could enable fast, accurate identification of olive cultivars that could be valuable for farmers, food inspection authorities or researchers (Beiki et al., 2012;Trujillo et al., 2014;Ponce et al., 2019;Gomes et al., 2020;Khadivi et al., 2022;Gago et al., 2024). The effectiveness and accuracy of imagebased methodologies for morphological characterization of different crop species is usually linked to cultivar discrimination and classification (Fuentes et al., 2018;Ishikawa et al., 2018;Wäldchen et al., 2018;Dheer and Singh, 2019). Traditionally, a rapid cultivar identification is a non-automated process involving visual inspection: the user tries to identify the organ characteristics that will be considered the discriminating keys for each cultivar (Wäldchen et al., 2018). Discriminating morphological characteristics of fruits, leaves, endocarps or other organs is commonly used for a quick cultivar identification based on appearance, but visual observations require experience and sometimes appear to be very subjective, inconsistent and inaccurate (Grinblat et al., 2016;Fuentes et al., 2018;Wäldchen and Mäder, 2018). Automated methodologies for plant cultivar identification based only on morphological characteristics are still in the very early stages (Ishikawa et al., 2018;Miho et al., 2024). This report focusses on the design of an automated methodology for olive cultivar discrimination based on the calculation of different morphological features of fruits, leaves and endocarps through geometrical feature extraction and cultivar classification. The cultivar classification algorithms are based on established machine learning techniques, while the morphological analysis was based on a previously developed methodology (Blazakis et al., 2017). This new automated methodology takes a further step in the development of an integrated automated tool to characterize, identify and discriminate a large set of olive cultivars using machine learning approaches. The relative contribution of each morphological feature for olive cultivar discrimination was also determined in a quantitative way. 2 Materials and methods 2.1 Plant material Fourteen Greek and international olive cultivars were discriminated in this study: Arbequina, Arbosana, Asprolia Alexandroupolis, Kalamon, Karidolia Chalkidikis, Koroneiki, Kothreiki, Koutsourelia, Mastoidis Gigas, Mavrolia Serron, Megaron, Ntopia Atsiholou, Thiaki and Tragolia. All fruit, leaf and endocarp samples of the olive cultivars were collected in 2016 and 2017 from the olive germplasm collection of the commercial nursery “Kostelenos”in Poros, Trizinia in Greece. All the trees were grown under identical conditions, and the fruit samples were Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org02 collected at the breaker stage. Fruits were collected from the middle part of one year-old shoots around the canopy at approximately 1.5 m height from the ground. They were collected from fruitbearing branches avoiding irregular fruits and taking into account the fruit load of each tree. Moreover, mature, healthy leaves were collected from the most representative one year-old shoots located on the southern part of the tree. Finally, the endocarps were extracted from the sample fruits, and the pulp was removed by a coarse fabric. All kernels were soaked in 10% bleach for 5 minutes and stored in a dry place for later usage. At least 100 samples from fruits, leaves or endocarps were used for the morphological analysis and the classification algorithms. 2.2 Morphological analysis To generate the imaging data for the olive samples, we followed the methodology described in (Blazakis et al., 2017). The morphological analysis of the fruit and endocarp samples was performed using the imaging positions adopted by UPOV and the International Olive Council (IOC). However, for the development of the classification algorithms we used all the numerical values corresponding to both positions of fruits and endocarps. A meopta copy imaging stand was used to create the imaging data of fruits and endocarps. All the samples were placed on a 2mm-thick piece of elevated glass to eliminate shadows, and the camera was installed above them, on a fixed solid arm. An HP DeskJet Ink Advantage 3636 scanner was used to scan the leaves at a resolution of at least 600 dots per inch (dpi). All the photographs were saved as jpeg or png files, and a scaler was placed next to them. Next, we separated the items in an image from their background (a process known as segmentation). This resulted in a binary image of each shape. The morphological analysis of olive fruits, leaves and endocarps was performed using OliveID, a set of state-of-the-art automatic algorithms for object contour extraction from imaging data that was developed in MATLAB (The Mathworks Inc., Natick, MA, USA) (Guide, 1998;Blazakis et al., 2017). OliveID is a computational methodology for olive morphological analysis that identifies various geometrical characteristics which are assigned to different morphological traits. The outcome of the algorithm is the representation of each shape by a discrete sequence with all its boundary points that enables us to quantitatively and qualitatively analyze the morphology of the olives, leaves and endocarps of each cultivar. For the morphological analysis of fruit, we used 24 parameters that describe fruit morphological characteristics, while we used 16 for the leaf. Finally, for the stone’s morphology description in two positions (A and B) we used 22 morphological characters. All the morphological characters were purely mathematically defined, and we refer to (Blazakis et al., 2017) and the references therein, for further details. 2.3 Data processing and analysis The statistical analysis of datasets and the creation and study of classifiers were implemented in a Jupyter Notebook in Python (Kluyver et al., 2016). Principal component analysis (PCA) was executed to visualize the initial classification of the olive cultivars based on quantitative data retrieved from the analysis that corresponds to the morphological traits of olive fruits, leaves and endocarps. It was implemented using the scikit-learn library (Pedregosa et al., 2011) in Python. For a more comprehensive view of the data set collected and calculated during the two consecutive years, split violin plots were built which allowed a direct comparison between the years. Violin plots can graphically represent the data distribution of a set of data by combining a box plot and a rotated density plot. By inspecting the shape of the violin plot representing the density estimate of the data points, regions with a higher frequency of particular values can easily be highlighted. Within a split violin plot, the left side represents the numerical data regarding the specific morphological characteristic relative to the first year, while the right side is for data extracted from the next year’s morphological analysis. 2.4 Classification algorithm The ability to explain in understandable terms why a machine learning model makes a certain classification is becoming immensely important, as it ensures trust and transparency in the decision-making process of the model. Therefore, in order to explore the morphological features’importance in the classification process, a general model-agnostic method for model interpretation, Shapley values, was used. Shapley values can provide accurate explanations, as they assign each morphological feature an importance value for olive cultivar prediction and determination (Kumar et al., 2020;Wang et al., 2021). It was implemented using the scikit-learn library (Pedregosa et al., 2011) in Python. 2.5 Machine learning classifier algorithms for olive cultivar identification Supervised and unsupervised machine learning techniques provide a powerful tool for agriculture due to their wide range of applications, such as detection of crop disease, crop management and plant phenotyping (Rehman et al., 2019). The task of discriminating olive cultivars based on morphological parameters of olive fruits, leaves and endocarps relies on a data mining problem in which several classification supervised learning methods were tested for an appropriate classification model (Gomes et al., 2020). The proposed methodology uses classifiers of two different types: •basic classifiers that determine the cultivar by using the quantitative data revealed by the morphological analysis of either fruit, leaf, or endocarp; •meta-classifiers that determine the cultivar by using all data from the morphological analysis of olive fruit, leaf and endocarp. The Random Forest Classifier (Breiman, 2001), k-Nearest Neighbor (Venables and Ripley, 2002) and Support Vector Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org03 Classifier (Boser et al., 1992) and the scikit-learn library (Pedregosa et al., 2011) were used as basic classifiers for olive fruits, leaves, and endocarps. Of all the classification algorithms tested, the XGBClassifier had an advantage due to the high accuracy of the classifier analysis. In order to perform the olive cultivar identification based on the morphological analysis of a plant’s organs, we use the stacking method (Dou et al., 2020), in which not raw data are used for training, but probability matrices from the basic classifiers. The meta-classifier learning process can be described as follows (Supplementary Figure 1): 1. Initially, the entire dataset is divided into two different sets. 9/10 of the whole dataset is used for training; and the other 1/10 is used at the end to test the meta-classifier. 2. Then the first set which was intended for training of the algorithm is sub-divided into 5 parts. 4/5 is used to train the three individual basic classifiers of fruits, leaves and endocarps. 3. After training, the remaining 1/5 of the data is used as input for the basic classifiers, and the result of the classification is 3 matrices of the probability of attribution to the cultivar by fruit, leaf and endocarp. Obtained matrices of the probabilities are used as a training sample for the meta-classifier. 4. The meta-classifier is trained on the new data set, and the cross-validation algorithm is performed to obtain the metric of such classification. 3 Results 3.1 Morphological characterization A group of 14 olive cultivars was characterized at the morphological level and a wide range of diversity was determined (Figure 1). This group was comprised of 12 Greek cultivars, Asprolia Alexandroupolis, Kalamon, Karidolia Chalkidikis, Koroneiki, Kothreiki, Koutsourelia, Mastoidis Gigas, Mavrolia Serron, Megaron, Ntopia Atsiholou, Thiaki, and Tragolia and two Spanish cultivars, Arbosana and Arbequina (Figure 1). The indicative fruit, leaf and endocarp samples were acquired from the first growing period (Figure 1). According to visual judgement, Karidolia-Chalkidikis and Mavrolia-Serron have larger fruits. The leaf morphology of Kalamon was also distinguished compared to the other cultivars, as it is substantially larger (Supplementary Figure 2). The endocarps of Karidolia-Chalkidikis and MavroliaSerron were the largest in size. The quantitative data of morphological features were acquired from the consecutive growing periods of 2016 and 2017. The distribution of the quantitative morphological features was visualized using split violin plots to take into consideration the variation due to climatic conditions, which showed the kernel density estimation of the data as well as the median and the upper and lower quartiles. Figures 2–4show the fruit, leaf and endocarp split violin plots for the most remarkable morphological features, respectively. All the data are presented in the Supplementary Figure 2. The shape of fruits appeared to be similar for most of the cultivars in the two consecutive growing periods, with the exception of Arbosana, which showed a different shape distribution each year. The fruit shape index indicated that Koroneiki and Kalamon had more elongated fruits, whereas Asprolia-Alexandroupolis and Mavrolia-Serron had more spherical fruits. Furthermore, the Koroneiki, Tragolia and Thiaki fruits had a more pointed apex than Mavrolia-Serron and Karidolia-Alexandroupolis fruits, in which the apex curvature was close to zero. The height and length of the minor axis of a fitted ellipse to the fruits followed the same tendency in the two growing seasons. The leaves of Megaron, Asprolia-Alexandroupolis and MavroliaSerron were the only ones with a lanceolate shape, while the maximum transverse diameter of blades appeared to be medium sized in most cultivars. Kalamon appeared to be the only cultivar with a wide leaf blade, while Koutsourelia, Asprolia-Alexandroupolis and Megaron cultivars had narrow leaves. The thickness of the upper part of the petiole was also determined quantitatively, with AsproliaAlexandroupolis and Arbequina displaying contrasting profiles in two consecutive years. The vertical symmetry of the leaf clearly illustrated that the maximum transverse diameter was located towards the apex, and this remained stable in time. The most striking observation regarding the endocarps was that the kernel density estimation of the morphological data remained stable between the two years of growth in most of the cultivars for all the phenotypic characteristics, according to the split violin plots of the endocarp morphological analysis. Koroneiki and Tragolia had small endocarps, while Karidolia-Chalkidikis and Kalamon had larger ones. The Tragolia endocarps had an ovoid shape, whereas Koroneiki and Koutsourelia were more elliptic based on the shape index. The endocarp shape of Kalamon and Karidolia-Chalkidikis appeared to be elongated. According to the shape index, cultivars were clustered in two groups: one with an ovoid shape, and another with an elliptic shape. This methodology also successfully detected the endocarp apex length. Comparing the area of the apex curve and the endocarp size in both years, the cultivars of Mastoidis and Koutsourelia had a mucro present, whereas it was absent in Kothreiki. Figure 5 shows a dendrogram indicating the clustering of the 14 cultivars based on the entire set of morphological characteristics. Four main groups were identified, all but one consisting of three to five cultivars. Only the cultivar of Ntopia Atsiholou is grouped distantly from the other clusters. It is interesting that Thiaki and Tragolia showed higher similarity to Koroneiki compared to the other cultivars. Moreover, the two Spanish cultivars, Arbequina and Arbosana, are grouped together in one sub cluster, separately from the Greek cultivars. Finally, the cultivars of Kalamon, Mavrolia Serron and Karidolia Chalkidikis form the cluster with the largest fruits. 3.2 The effect of environment conditions on morphological parameters Wide variation in the morphological characteristics was detected due to variations in the environmental conditions in the two sampling years (Supplementary Table 1). In fruits, higher Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org04 variability was observed in the first growing period compared to the second. The size of the fruit nipple showed the greatest variability while the nipple index (presence or absence of the nipple) the lowest (Supplementary Table 1). Kothreiki and Mastoidis showed the highest variation among all cultivars in fruit area in 2017 and 2016, 0.20 and 0.18, respectively. The fruit shape index was one of the most stable traits over the two growing periods, showing only a 6% coefficient variation (Supplementary Table 2), while the fruit area exhibited a higher variability, as expected, compared to the other morphological characteristics of fruit. Moreover, Arbequina showed the highest stability in fruit morphological characteristics, while Koutsourelia was the least stable (Supplementary Table 2). In leaf characteristics, the average curvature of the Kothreiki leaf tip had the highest variability in 2016 (1.74), while the shape index of the Megaron leaf blade had the lowest (0.084). In addition, the leaf petiole of all cultivars was the most variable feature over the two FIGURE 1 Morphological features of olive cultivars. Fruits, leaves and endocarps of Arbequina, Arbosana, Asprolia Alexandroupolis, Kalamon, Karidolia Chalkidikis, Koroneiki, Kothreiki, Koutsourelia, Mastoidis Gigas, Mavrolia Serron, Megaron, Ntopia Atsiholou, Thiaki, Tragolia are presented with a similar scale. The images were modified to indicate the actual sizes for comparative purposes. The red line represents 1cm. Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org05 years, whereas the vertical symmetry and the leaf circularity were the most stable (Supplementary Table 2). The observed variation in fruit size among cultivars was accompanied by similar variation in endocarp size. Cultivars with greater fruit size showed also greater endocarp size. The coefficient of variation for all traits of endocarps varies in the lower range of 0.006 to 0.35 (Supplementary Table 1). The highest variation appeared in the minimum distance between the transversal diameter and endocarp’s contour of Kalamon in 2016 (CV=0.35), while the lowest appeared in the average circularity of Arbequina endocarps in 2016 (CV=0.006) (Supplementary Table 1). The traits related to circularity and the shape index of the endocarp exhibited high stability compared to those related to endocarp area (Supplementary Tables 1,2). 3.3 Principal component analysis Principal Component Analysis (PCA) was used to determine morphological features which differentiate among cultivars, indicating which characteristics are more reliable for discrimination. Figures 6–8 show principal component analysis performed using the numerical data of morphological characteristics of fruits, leaves and endocarps, respectively. The first two components account for approximately 66%, 70%, and 77% of the total variance numerical data, respectively. PCA revealed significant variation within cultivars which might also be attributed to the differences in environmental conditions between the growing periods. Olive cultivars are dispersed along both principal component axes and are hardly distinguishable from each other, illustrating the complexity of the classification. The two PCA scatter plots of fruits and endocarps illustrate a more even distribution of cultivars, providing higher discriminatory capacity compared to the PCA of leaves. In the leaf PCA, only Kalamon was clearly discriminated from the rest of the cultivars, probably because this cultivar has larger leaves. KaridoliaChalkidikis in both PCA of fruits and endocarps appears to be more discernible from the other cultivars, which might be attributed to the size of its fruits. Similarly, the fruit size and shape of Kothreiki and MavroliaSerron might be responsible for the classification in different groups in the fruit Principal component analysis. The first principal component of fruits and endocarps was predominantlydeterminedbyareaandheight,whereasthesecond principal component was generally determined by features responsible for shape elongation (Supplementary Table 3). In the first main component of leaves, the greatest contribution was made by features that characterize the geometric size, such as perimeter, area, and height. FIGURE 2 Morphological analysis. Violin graphs for fruits. Violin plots representing trait variation in two consecutive years. Each plot shows the distribution of data from the minimum to the maximum level, with white inner dot showing the data median for 14 cultivars. The black boxplots represent the lower and upper limits of the first and third quartiles. The outliers were removed. The horizontal width of the violin depends on the data density. Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org06 The shape index, which is responsible for the leaf blade elongation, was predominantly a positive effector for the second principal component. 3.4 Machine learning algorithm - classification accuracy The average values of morphological quantitative data for fruits, leaves and endocarps showed no adequate capacity to reliably discriminate olive cultivars according to Principal Component Analysis. In fact, discrimination between cultivars could be observed only if the morphological data of plant organs of individual plants was taken into consideration using multivariate statistical analysis. Therefore, machine learning algorithms were used to classify and discriminate olive cultivars. The Extreme Gradient Boost (XGBoost) algorithm performed better than three other basic classifiers (Random Forest Classifier, k-Nearest Neighbor (ΚNN) FIGURE 3 Morphological analysis. Violin graphs for leaves. Violin plots representing trait variation in two consecutive years. Each plot shows the distribution of data from the minimum to the maximum level, with white inner dot showing the data median for 14 cultivars. The black boxplots represent the lower and upper limits of the first and third quartiles. The outliers were removed. The horizontal width of the violin depends on the data density. Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org07 and Support Vector Classifier (SVM)) (Supplementary Table 4). Figure 9 illustrates the mean accuracy of the classifier with different numbers of cultivars. As expected, the number of olive cultivars determined the success of the classification algorithm. The higher the number of cultivars, the lower the classification accuracy (Figure 9). Up to 6 cultivars could be classified with higher than 85% accuracy, while 14 cultivars were classified with 76% and 80% accuracy using fruit and endocarp morphological data, respectively (Figure 9). The mean accuracy of leaf classifier was nearly 55%. These results indicate that endocarp morphological features have the greatest capacity for cultivar discrimination, since these traits are less affected by the variability of environmental conditions and the training systems. Single classifiers can be used to classify data for relatively simple tasks. Olive cultivar discrimination is a complex task, and the combination of simple classifiers can significantly improve performance. A meta-classifier approach which uses three different combination methods (KNN, RandomForest, XGBoost) was implemented in a similar way to the combination of simple classifiers. This approach used the numerical data of the morphological features of fruits, leaves and endocarps. Table 1 indicates the performance measures (precision, recall and F1-score values) of the meta-classifier approach on the data set. It is remarkable that the average accuracy of an identification was approximately 95%. 3.5 Importance of each morphological feature and the contribution of plant organs to cultivar classification Machine learning models are increasingly used to replace human decision-making. A concept from cooperative game theory could be used to detect the fair contribution of each morphological feature to olive cultivar identification. The FIGURE 4 Morphological analysis. Violin graphs for endocarps. Violin plots representing trait variation in two consecutive years. Each plot shows the distribution of data from the minimum to the maximum level, with white inner dot showing the data median for 14 cultivars. The black boxplots represent the lower and upper limits of the first and third quartiles. The outliers were removed. The horizontal width of the violin depends on the data density. Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org08 FIGURE 5 Dendrogram indicating the clustering of the 14 cultivars based on the entire set of morphological characteristics. FIGURE 6 Principal Component Analysis (PCA). PCA for fruits. PCA shows a score plot of the morphological analysis of the fruits of 14 cultivars in two consecutive years. Different colors represent different cultivars. Squares represent the 2016 data, whereas circles represent 2017. Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org09 Dheer, P., and Singh, R. (2019). Identification of Ιndian rice varieties using machine learning classifiers. Plant Arch. 19, 155–158. Diaz, R., Gil, L., Serrano, C., Blasco, M., Molto, E., and Blasco, J. (2004). Comparison of three algorithms in the classification of table olives by means of computer vision. J. Food Eng. 61, 101–107. doi: 10.1016/S0260-8774(03)00191-2 Dou, J., Yunus, A. P., Bui, D. T., Merghadi, A., Sahana, M., Zhu, Z., et al. (2020). Improved landslide assessment using support vector machine with bagging, boosting, and stacking ensemble machine learning framework in a mountainous watershed, Japan. Landslides 17, 641–658. doi: 10.1007/s10346-019-01286-5 Fuentes, S., Hernandez-Montes, E., Escalona, J. M., Bota, J., Viejo, C. G., PobleteEcheverrıa, C., et al. (2018). Automated grapevine cultivar classification based on machine learning using leaf morpho-colorimetry, fractal dimension and near-infrared spectroscopy parameters. Comput. Electron. Agric. 151, 311–318. doi: 10.1016/ j.compag.2018.06.035 Gago, P., Boso, S., Santiago, J. L., and Martınez, M. C. (2024). Identification and characterization of relict olive varieties (Olea europaea L.) in the Northwest of the Iberian Peninsula. Horticulturae 10, 175. Ganino, T., Bartolini, G., and Fabbri, A. (2006). The classification of olive germplasm. J.Hortic.Sci.Biotechnol.81, 319–334. doi: 10.1080/ 14620316.2006.11512069 Gomes, L., Nobre, T., Sousa, A., Rei, F., and Guiomar, N. (2020). Hyperspectral reflectance as a basis to discriminate olive varieties—a tool for sustainable crop management. Sustainability 12, 3059. doi: 10.3390/su12073059 Gomez-Rodrıguez, M. V., Beuzon, C., Gonzalez-Plaza, J. J., and Fernandez-Ocaña, A. M. (2021). Identification of an olive (Olea europaea L.) core collection with a new set of SSR markers. Genet. Resour. Crop Evol. 68, 117–133. doi: 10.1007/s10722-020-00971-y Grinblat, G. L., Uzal, L. C., Larese, M. G., and Granitto, P. M. (2016). Deep learning for plant identification using vein morphological patterns. Comput. Electron. Agric. 127, 418–424. doi: 10.1016/j.compag.2016.07.003 Guide, M. U. S. (1998) (Natick, MA: The Mathworks. Inc.) 5 (333), p.4. Ishikawa, T., Hayashi, A., Nagamatsu, S., Kyutoku, Y., Dan, I., Wada, T., et al. (2018). Classification of strawberry fruit shape by machine learning. Int. Arch. Photogrammetry Remote Sens. Spatial Inf. Sci. 42, 463–470. doi: 10.5194/isprs-archives-XLII-2-463-2018 Kaya, H. B., Akdemir, D., Lozano, R., Cetin, O., Sozer Kaya, H., Sahin, M., et al. (2019). Genome wide association study of 5 agronomic traits in olive (Olea europaea L.). Sci. Rep. 9, 1–14. doi: 10.1038/s41598-019-55338-w Khadivi, A., Mirheidari, F., Moradi, Y., and Paryan, S. (2022). Identification of the promising olive (Olea europaea L.) cultivars based on morphological and pomological characters. Food Sci. Nutr. 10, 1299–1311. doi: 10.1002/fsn3.2767 Kluyver, T., Ragan-Kelley, B., Perez, F., Granger, B., Bussonnier, M., Frederic, J., et al. (2016). “Jupyter Notebooks–a publishing format for reproducible computational workflows”in Positioning and power in academic publishing: Players, agents and agendas, eds. F. Loizides and B. Schmidt, (Clifton, VA, USA: IOS Press), 87–90. Koubouris, G. C., Avramidou, E. V., Metzidakis, I. T., Petrakis, P. V., Sergentani, C. K., and Doulis, A. G. (2019). Phylogenetic and evolutionary applications of analyzing endocarp morphological characters by classification binary tree and leaves by SSR markers for the characterization of olive germplasm. Tree Genet. Genomes 15, 1–12. doi: 10.1007/s11295-019-1322-0 Kumar, I. E., Venkatasubramanian, S., Scheidegger, C., and Friedler, S. (2020). “Problems with Shapley-value-based explanations as feature importance measures”in Proceedings of the Proceedings of the 37th International Conference on Machine Learning; PMLR, Salt Lake City, UT, USA, 13 July 2020, eds. H. DaumeIII and A. Singh, 119, 5491–5500. Mancuso, S., and Nicese, F. P. (1999). Identifying olive (Olea europaea) cultivars using artificial neural networks. J. Am. Soc. Hortic. Sci. 124, 527–531. doi: 10.21273/ JASHS.124.5.527 Martınez, S. S., Gila, D. M., Beyaz, A., Ortega, J. G., and Garcıa, J. G. (2018). A computer vision approach based on endocarp features for the identification of olive cultivars. Comput. Electron. Agric. 154, 341–346. doi: 10.1016/j.compag.2018.09.017 Miho, H., Pagnotta, G., Hitaj, D., De Gaspari, F., Mancini, L. V., Koubouris, G., et al. (2024). OliVaR: Improving olive variety recognition using deep neural networks. Comput. Electron. Agric. 216, 108530. doi: 10.1016/j.compag.2023.108530 Mousavi, S., de la Rosa, R., Moukhli, A., El Riachy, M., Mariotti, R., Torres, M., et al. (2019). Plasticity of fruit and oil traits in olive among different environments. Sci. Rep. 9, 16968. doi: 10.1038/s41598-019-53169-3 Muleo, R., Morgante, M., Cattonaro, F., Scalabrin, S., Cavallini, A., Natali, L., et al. (2016). “Genome sequencing, transcriptomics, and proteomics,”in The Olive Tree Genome, eds E. Rugini, L. Baldoni, R. Muleo and L. Sebastiani (Cham: Springer), 141– 161. doi: 10.1007/978-3-319-48887-5_9 Orrù, M., Grillo, O., Lovicu, G., Venora, G., and Bacchetta, G. (2013). Morphological characterisation of Vitis vinifera L. seeds by image analysis and comparison with archaeological remains. Vegetation History Archaeobotany 22, 231–242. doi: 10.1007/ s00334-012-0362-2 Paredes, D., Karp, D. S., Chaplin-Kramer, R., Benıtez, E., and Campos, M. (2019). Natural habitat increases natural pest control in olive groves: economic implications. J. Pest Sci. 92, 1111–1121. doi: 10.1007/s10340-019-01104-w Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., et al. (2011). Scikit-learn: Machine learning in python. J. Mach. Learn. Res. 12, 2825–2830. Ponce, J. M., Aquino, A., and Andujar, J. M. (2019). Olive-fruit variety classification by means of image processing and convolutional neural networks. IEEE Access 7, 147629–147641. doi: 10.1109/Access.6287639 Ponce, J. M., Aquino, A., Millan, B., and Andujar, J. M. (2018). Olive-fruit mass and size estimation using image analysis and feature modeling. Sensors 18, 2930. doi: 10.3390/s18092930 Puerto, D. A., Martınez Gila, D. M., Gamez Garcıa, J., and Gomez Ortega, J. (2015). Sorting olive batches for the milling process using image processing. Sensors 15, 15738– 15754. doi: 10.3390/s150715738 Rallo, L., Barranco, D., Dıez, C. M., Rallo, P., Suarez, M. P., Trapero, C., et al. (2018). “Strategies for olive (Olea europaea L.) breeding: cultivated genetic resources and crossbreeding,”in Advances in Plant Breeding Strategies: Fruits, (New York City, NY, USA: Springer International Publishing), vol. 3, 535–600. Rehman, T. U., Mahmud, M. S., Chang, Y. K., Jin, J., and Shin, J. (2019). Current and future applications of statistical machine learning algorithms for agricultural machine vision systems. Comput. Electron. Agric. 156, 585–605. doi: 10.1016/j.compag.2018. 12.006 Riquelme, M. T., Barreiro, P., Ruiz-Altisent, M., and Valero, C. (2008). Olive classification according to external damage using image analysis. J. Food Eng. 87, 371–379. doi: 10.1016/j.jfoodeng.2007.12.018 Rodrıguez, G. R., Moyseenko, J. B., Robbins, M. D., Morejon, N. H., Francis, D. M., and van der Knaap, E. (2010). Tomato Analyzer: a useful software application to collect accurate and detailed morphological and colorimetric data from two-dimensional objects. JoVE (Journal Visualized Experiments) 37, e1856. doi: 10.3791/1856 Sadok, I. B., Celton, J. M., Essalouh, L., El Aabidine, A. Z., Garcia, G., Martinez, S., et al. (2013). QTL mapping offlowering and fruiting traits in olive. PLoS One 8, e62831. doi: 10.1371/journal.pone.0062831 Sebastiani, L., and Busconi, M. (2017). Recent developments in olive (Olea europaea L.) genetics and genomics: applications in taxonomy, varietal identification, traceability and breeding. Plant Cell Rep. 36, 1345–1360. doi: 10.1007/s00299-017-2145-9 Torres, M., Pierantozzi, P., Searles, P., Rousseaux, M. C., Garcıa-Inza, G., Miserere, A., et al. (2017). Olive cultivation in the southern hemisphere: Flowering, water requirements and oil quality responses to new crop environments. Front. Plant Sci. 8, 1830. doi: 10.3389/fpls.2017.01830 Trujillo, I., Ojeda, M. A., Urdiroz, N. M., Potter, D., Barranco, D., Rallo, L., et al. (2014). Identification of the Worldwide Olive Germplasm Bank of Cordoba (Spain) using SSR and morphological markers. Tree Genet. Genomes 10, 141–155. doi: 10.1007/ s11295-013-0671-3 Vanloot, P., Bertrand, D., Pinatel, C., Artaud, J., and Dupuy, N. (2014). Artificial vision and chemometrics analyses of olive stones for varietal identification of five French cultivars. Comput. Electron. Agric. 102, 98–105. doi: 10.1016/ j.compag.2014.01.009 Venables, W. N., and Ripley, B. D. (2002). “Random and mixed effects,”in Modern applied statistics with S (New York, NY: Springer New York), 271–300. Wäldchen, J., and Mäder, P. (2018). Plant species identification using computer vision techniques: a systematic literature review. Arch. Comput. Methods Eng. 25, 507– 543. doi: 10.1007/s11831-016-9206-z Wäldchen, J., Rzanny, M., Seeland, M., and Mäder, P. (2018). Automated plant species identification—Trends and future directions. PLoS Comput. Biol. 14, e1005993. doi: 10.1371/journal.pcbi.1005993 Wang, Y., Ma, G., Mei, J., Zou, Y., Zhang, D., Zhou, W., et al. (2021). Machine learning reveals the influences of grain morphology on grain crushing strength. Acta Geotechnica 16, 3617–3630. doi: 10.1007/s11440-021-01270-1 Zhu, S., Niu, E., Shi, A., and Mou, B. (2019). Genetic diversity analysis of olive germplasm (Olea europaea L.) with genotyping-by-sequencing technology. Front. Genet. 10, 755. doi: 10.3389/fgene.2019.00755 Blazakis et al. 10.3389/fpls.2024.1441737 Frontiers in Plant Science frontiersin.org16