1. Introduction

The quality of agricultural products is of critical importance to ensure the sustainable success of agricultural businesses and to meet consumers’ demands for higher product standards [1]. Various stakeholders in the agricultural product supply chain, including suppliers, processors, and quality inspection departments, must implement quality measures to ensure product quality and safety [2]. The assurance of quality and safety in agricultural goods across the complete food supply chain is of utmost importance for human well-being and underscores the necessity for a thorough monitoring framework [3]. Prominent agricultural organizations prioritize enhancing quality and safety control protocols to align with consumer expectations and safeguard the integrity of agricultural products [4]. Phenotype classification is an important method for maintaining the quality and health of the agricultural production line.

Walnut is one of the agricultural products that require rapid phenotype classification during processing. The nutritional values of walnuts vary significantly depending on the type of walnut. Different types of walnuts exhibit variations in terms of moisture content, oil content, fatty acid composition, protein, carbohydrate, and mineral content [5]. Different studies have highlighted the significance of classification based on morphological, biochemical, and molecular markers [6]. Techniques like low-field nuclear magnetic resonance (LFNMR) have been utilized to classify walnut varieties based on water and oil content, aiding in determining storage conditions [7]. Fatty acid profiling has been identified as a key indicator for distinguishing walnuts from different geographical locations, providing insights into quality assessments and possible adulterations in the market [8]. Implementing walnut classifying and screening production lines can enhance sorting efficiency, size, weight, and quality identification in walnut sorting processes [9]. Among these methodologies, “computer vision” and “machine learning” applications in agriculture are gaining interest due to their nondestructive evaluation capabilities and lower costs compared to traditional manual methods [10]. These technologies offer a way to efficiently analyze and monitor agricultural products without causing any damage, making them a valuable tool for enhancing productivity and quality control in the industry. In the literature, deep learning methods are widely used for phenotype prediction and classification across a broad spectrum. For instance, the classification of rice varieties can be achieved using artificial intelligence methods [11].

When the literature is examined, machine learning methods have been extensively researched in the classification of walnut phenotypes. Substantial advancements and valuable insights have been achieved within this domain. Diverse machine learning algorithms have been employed to effectively categorize different traits of walnuts. In various studies in the literature, it has been seen that artificial neural networks (ANNs), support vector machines (SVMs), and random forests (RFs), which are well-known methods, are used in developing classification models to distinguish walnuts with certain characteristics, especially shrunken kernels [12, 13]. These approaches have exhibited notable levels of accuracy, notably with SVM attaining a classification accuracy (CA) rate of 97% in the identification of walnuts with shriveled kernels [13].

Moreover, machine learning techniques have been instrumental in enhancing the precision (Prec) of phenotypic profiling in walnuts. By leveraging these methods, researchers have been able to improve the accuracy of suture strength phenotyping in walnuts, enabling the application of quantitative trait loci (QTL) mapping and genome-wide association studies (GWAS) to dissect the genetic basis of walnut traits [14]. Additionally, machine learning models, such as deep convolutional neural networks (CNNs) (DCNNs) and SVM, have been employed to automatically classify phenotypic traits in walnuts based on various characteristics, including appearance and coloration [15].

Furthermore, the integration of machine learning with advanced technologies like X-ray computed tomography has enabled the 3D characterization of walnut morphological traits, showcasing the versatility of machine learning in analyzing complex walnut structures [16]. Additionally, machine learning algorithms have been utilized in conjunction with hyperspectral imaging technology for nondestructive detection of moldy walnuts, achieving an impressive overall CA of 93%.

The application of machine learning methods in the classification of walnut phenotypes has proven to be a powerful tool in enhancing accuracy, efficiency, and Prec in phenotypic profiling. By leveraging a diverse range of machine learning algorithms and technologies, researchers have made significant strides in understanding and classifying walnut traits, paving the way for advancements in walnut cultivation and agricultural research. The fact that deep learning methods, which are among the machine learning methods, can directly receive images as input has caused deep learning methods to attract attention and be seen as advantageous in this regard.

Deep learning methods stand out as an important tool in determining plant phenotypes. These methods can help plant breeders quickly and accurately identify plant traits, diagnose diseases, and increase productivity [17, 18]. Therefore, deep learning methods are considered an important tool for maintaining quality and health in determining plant phenotypes. Deep learning methods enable image-based high-throughput phenotype analysis to identify plant phenotypes [19]. These methods can automatically identify plant traits and help plant breeders make informed decisions.

In the context of plant phenotyping, deep learning models have been successfully applied in various applications, such as weed species classification and disease detection [20]. Specifically, deep learning has been used to establish ANN models for screening differentially expressed genes in walnut endocarps during the hardening period, overcoming the limitations of traditional algorithms [21]. Li et al. [22] developed a YOLOv8-based method for fast classification of walnut material. In this study, it was reported that walnut classes were accurately separated using a lightweight and fast CNN model [22].

Similarly, in a study by Guo et al. [23], deep learning was combined with phenotypic data for walnut maturity levels and oil content estimation. This approach is important not only for classification but also for walnut quality estimation [23].

Working on a more specific problem such as the separation of walnut shell and kernel, Ni et al. optimized a deep learning model that provided high accuracy in walnut shell-kernel separation with machine vision [24]. It was observed that image processing and data preprocessing techniques increased model performance. Moreover, the backpropagation ANN classification model has been employed to identify regional variations in walnuts with a high mean accuracy of 89.7% [25]. Karadeniz et al. created a walnut dataset and achieved successful classification using VGG-16, with an accuracy of 85.52% on the original dataset and 90.55% on the augmented dataset [26]. Additionally, Chen et al. proposed a deep-learning model combining spatial attention and SE-network structures, achieving a test set accuracy of 92.2% for walnut kernel grading, and outperforming VGG-19 and EfficientNet [27]. These findings collectively demonstrate the success of deep learning methods, particularly VGG-16, VGG-19, and novel models, in accurately classifying walnut phenotypes with impressive accuracy rates. Machine learning methods, including deep learning, have been instrumental in developing classification models for authenticating walnuts from different regions worldwide [28]. Table 1 represents brief information about the mentioned studies.

The application of deep learning architectures such as InceptionV3, SqueezeNet, VGG-16, and VGG-19 in walnut image categorization has yielded promising results. Research indicates that employing pre-existing CNN architectures proves to be proficient in accurately categorizing walnut leaf visuals [26]. Additionally, studies with pretrained CNN models such as VGG-16, DenseNet, MobileNet, InceptionV3, ResNet50, and Xception have highlighted advances in image classification tasks [29]. Thus, the integration of these pre-existing architectures can substantially enhance the Prec and effectiveness of walnut image categorization tasks while offering valuable insights for individuals involved in the domains of computer vision and deep learning.

Overall, the integration of deep learning methods in the classification of walnut phenotypes holds significant potential for accurate and efficient phenotypic profiling. By leveraging the capabilities of deep ANNs and machine learning algorithms, researchers can enhance the CA and speed of identifying walnut phenotypes, contributing to advancements in agricultural research and walnut cultivation practices.

In this study, the effectiveness of various deep learning methods for classifying walnut species was evaluated. Two different approaches were tested for this purpose. The first approach involved generating features using deep learning methods and then identifying the most effective features. These features were subsequently used for classification with various machine learning methods. The second approach directly utilized various deep learning methods for classification.

The main reason for choosing InceptionV3, VGG-16, and VGG-19 models is their proven success in image classification tasks. InceptionV3 stands out with its ability to efficiently extract detailed features and its low computational cost due to its innovative architecture. VGG-16 and VGG-19 offer powerful feature extraction capabilities with their deeper but simpler structures and are known for their versatility in different classification tasks. These models were selected to explore the performance of different architectures on the walnut dataset in order to provide a balance between computational efficiency and accuracy.

Extracting features with CNN models and classifying these features with traditional machine learning methods allowed us to benefit from the strengths of both methods. CNNs are very good at detecting complex patterns and features that are difficult to extract manually from image data. By removing FC layers, the resulting features represent high-level abstractions of the dataset. These features increase CA and interpretability when used with classifiers such as LR, SVM, and k-NN. This hybrid approach allowed us to examine the dataset in detail and explore the performance of feature-based classification compared to end-to-end deep learning models.

When selecting machine learning models, our main criteria were CA, computational efficiency, and the ability to process feature vectors extracted by CNNs. LR was chosen due to its simplicity and robust performance on linearly separable data. SVM was chosen due to its ability to effectively process high-dimensional data and its success in achieving high accuracy. k-NN was included to evaluate its performance as a nonparametric method. These models were chosen to comparatively analyze the strengths and weaknesses of different algorithms in walnut species classification based on extracted features. Another purpose of choosing these methods in the study is to adapt this model to industrial applications after obtaining an effective classification model.

The organization of this study is as follows: In the second section, the dataset, performance metrics, cross-validation, and methods used in the study are described. In the third section, the experimental results obtained in the study are presented. In the final section, the experimental results are interpreted, and recommendations are provided.

2. Materials and Methods

2.1. Dataset

In this study, phenotypic classification was conducted by examining four different types of walnuts. The dataset comprised walnuts of the Chandler, Fernor, Howard, and Oguzlar varieties. General information about these walnut types is as follows.

Chandler walnuts are recognized for their favorable fatty acid composition, with high levels of polyunsaturated fatty acids (PUFAs) [30]. Chandler walnuts are characterized by specific phenological and pomological traits, such as nut dimensions larger than traditional Chandler varieties, light kernel color, and easy kernel removal [31].

Fernor walnuts exhibit specific characteristics based on various studies. In terms of sensory evaluation, Fernor walnuts were noted to have a rough shell, dark pellicle, and less brightness in the kernel, making them one of the least crispy cultivars [32].

The appearance features of the Howard-type walnut were studied in various contexts. Lampinen et al. conducted a trial on pruned and unpruned Howard walnut trees, observing canopy growth and nut quality characteristics [33].

Oguzlar-type walnut, as studied in various contexts, exhibits specific characteristics related to storage and quality. Research by Yalçin et al. highlights that Oguzlar-77 walnuts maintain their quality under 0°C storage conditions, emphasizing the importance of suitable storage conditions for preserving freshness (Yalçin et al.).

The dataset comprises various images of each class, captured using a specialized image acquisition system designed specifically for this type of task. The system consists of a sealed box equipped with dimmable internal lighting, a camera, and interchangeable backgrounds. For this study, a black background was utilized. The internal lighting was provided by 4000 K, 160 Lumen LED lights. Figure 1 illustrates the image acquisition system.

Figure 1a represents the general view of the system, Figure 1b represents the camera system, which can be changed due to task, and Figure 1c represents the upper section of the system with images of walnuts captured by the computer. There are no moving parts in the mechanism. Walnuts were placed in various positions in front of the camera, and their images were captured.

Images obtained from the image acquisition system were used to generate a dataset with four classes which are named Chandler, Fernor, Howard, and Oguzlar. While creating the dataset, each image was obtained separately. The images shown in Figure 2 are the first 20 images of each class in this dataset, combined into a single image.

In Figures 2a, 2b, 2c, and 2d, each class is shown as Chandler, Fernor, Howard, and Oguzlar. The dataset consists of 952 images. Table 2 shows the number of images contained in each class in the dataset.

Table 2. Amount of images contained in each class in the dataset.

Total	Chandler	Fernor	Howard	Oguzlar
952	234	258	238	222

The dataset is used to train and test various machine learning models. In order to obtain a sufficient number of image data and to share the amount of data for each class in the same proportion as possible, a data set was created by selecting 3 kg of each walnut type.

2.2. Experiment Design

In this study, two different experimental approaches were adopted. In both approaches, deep learning models such as InceptionV3, VGG-16, and VGG-19 were preferred. InceptionV3, introduced by Szegedy et al. [34], is known for its efficiency in extracting detailed information from images while reducing training parameters. VGG-16 and VGG-19, on the other hand, are recognized for their simplicity and effectiveness in image classification tasks. The architectures examined in this study can be summarized as follows.

2.2.1. VGG-Net Architecture

VGG-Net, developed by the Visual Graphics Group in Oxford, made a breakthrough idea to reinforce the value of depth in deep ANNs. Featuring a simple architecture and stacking cascades of 3 × 3 convolution (Conv.) layers one after the other, VGG-Net models are readily interpretable and implementable. These characteristics have contributed to the popularity of VGG-Net in image feature extraction [35]. Figure 3 shows the VGG-16 architecture, which is characterized by sequential Conv. layers to capture hierarchical spatial features in walnut images. This structure plays a critical role in distinguishing walnut varieties in the dataset by effectively extracting visual patterns such as texture and shape.

VGG-16 and VGG-19 share a similar architecture, with the primary distinction being the number of layers. VGG-19, based on CNN principles, consists of 16 Conv. layers and 3 FC layers. These layers work together to transform image data into feature vectors. Figure 4 presents the VGG-19 architecture, which is an extension of VGG-16. The added Conv. layers enhance the extraction of more complex features, especially subtle differences in walnut shell texture or kernel coloration, and improve the CA.

2.2.2. InceptionV3 Architecture

The architecture of InceptionV3, depicted in Figure 5, employs different filter sizes within the same Conv. block. This design enables the extraction of features at multiple scales from the input volume. For instance, it processes the input using 1 × 1, 2 × 2, 3 × 3, and 5 × 5 filters, and max-pooling (same padded) within a single block, resulting in feature volumes of varying depths. These volumes are then concatenated to produce a final output that is fed into the next layer. Computational complexity is reduced using 1 × 1 Conv., which decrease the channel depth of the volumes without altering their height and width. The simultaneous Conv. allow the network to expand in width. InceptionV3 further reduces computational complexity by factorizing Conv. into smaller and asymmetric operations [29]. Figure 5 shows the InceptionV3 architecture, which uses the multiscale approach. This architecture processes the inputs by passing walnut images through various filter sizes in a single Conv. layer. This design captures detailed features and broader contextual patterns, which increases the ability to distinguish walnut varieties.

The education of these models frequently includes extensive datasets like ImageNet, as referenced by [36]. For example, models such as InceptionV3 undergo training on millions of images and subsequently undergo assessment on validation sets to guarantee their Prec and generalization abilities. The process of training usually encompasses the enhancement of the models’ parameters through methods such as backpropagation and gradient descent to reduce prediction inaccuracies.

In the first approach, the classification task was carried out only with deep learning approaches, and the “FC” layer was not removed in the last classification step. In the second approach, three deep learning models were used to extract deep learning features from the dataset. In this approach, the “FC” layer is removed from the models. Then, the features obtained with these models have been presented to classification algorithms which are LR, k-NN, and SVM. Figure 6 represents both experimental designs simultaneously.

Figure 6 represents the experimental designs of the classification tasks. Figure 6a represents the only deep learning-based classification Figure 6b represents the deep learning features and classifier-based classification. The classification algorithms used in the first can be explained as follows:

LR is one of the most widely used models. In LR, the independent variable is predicted using one or more independent variables. This method is used to combine arguments and independent variables. An important advantage of LR is that the variables are not normally distributed. Since predicted values are interpreted as probabilities, values of LR can be between 0 and 1. This is not a direct consequence of LR, but a consequence of estimating probability [11].

As a machine learning algorithm, k-NN is highly effective for classification tasks. Originally introduced by Fix and Hodges in 1951 [37], the method works by identifying k objects from the training set that closely match the test data. Based on the classes of these neighboring objects, the algorithm assigns the test data to the most common class among them [9].

SVM is a kernel-based method used for lights and regression. Thanks to its separation program, SVM can classify data as linear in two-dimensional space, plane in three-dimensional space, and hyperplane in multidimensional space. SVM involves the process of finding the best differentiating hyperplanes of data [13, 16].

2.4. Cross Validation

Cross-validation is a fundamental technique in machine learning used to evaluate the performance and generalization ability of predictive models. It involves partitioning the dataset into subsets for training and testing, iteratively switching between different subsets for training and validation. This method helps in assessing how well a model will perform on unseen data, guarding against overfitting and providing a more accurate estimation of the model’s predictive capabilities [38]. By repeating cross-validation, the variability in model performance due to different dataset splits can be accounted for, aiding in the selection of optimal models [38]. Additionally, cross-validation allows for the estimation of prediction errors, which is particularly useful when dealing with limited data, as it maximizes the use of available information [39]. It is a versatile tool that can be applied in various scenarios, even when precise likelihood derivation or parameter counting is challenging [40]. Moreover, cross-validation results in more stable evaluations compared to simply splitting data into training and validation sets [41]. Figure 7 represents the cross-validation process.

Overall, cross-validation is a crucial method in machine learning that ensures robust model evaluation and selection, significantly contributing to the reliability and generalizability of predictive models. In this study, each model is trained with the 10-fold cross-validation method.

3. Results

In the experiments carried out within the scope of this study, three different deep learning models (InceptionV3, VGG-19, and VGG-16) and the features extracted from these models were compared with various machine learning methods. In the first experiment, the direct classification performance of the models was evaluated and InceptionV3 stood out as the most successful model with a high AUC (0.995) and CA (0.947). In the second experiment, the “FC” layers of the deep learning models were removed and used as a feature extraction tool, and these features were classified with logistic regression (LR), SVM, and k-NN algorithms. The LR method showed superior performance in both accuracy and other metrics (CA: 0.951–0.961). In addition, the errors and correct predictions in the classification process were detailed by analyzing the confusion matrices. The results reveal the potential of the proposed approaches to increase CA and the opportunities for future studies to eliminate the deficiencies.

The parameters of the deep learning algorithms used in all experiments are listed in Table 4. The table provides a comparative overview of the three deep learning models used in the experiments, InceptionV3, VGG-16, and VGG-19. The table includes the basic features such as the number of layers in the architecture of the models, the total number of parameters (which determines the complexity of the model), and the input size they are designed to process. All models use the ReLU activation function for nonlinear transformations and support optimization techniques such as Adam. In addition, these models are pretrained on the ImageNet dataset and are quite effective for transfer learning in image recognition tasks. Structural differences such as the number of layers and parameters affect the performance and computational requirements of these models, providing flexibility for different usage scenarios.

In Table 4, Conv. are abbreviated as FC layers. The number of layers, parameter sizes, input sizes, and optimization methods of the models were taken from the sources that describe the original model designs. In addition, the information that these models were pretrained on the ImageNet dataset was included in the table as a basic feature commonly mentioned in the literature.

4. Discussion

It is of great importance to classify agricultural products during production to evaluate them, ship them correctly, and protect the supply chain. Classifying walnuts in this way provides a great advantage in the production and packaging process. Classification methods have been developed by looking at the leaves of the trees while they are still growing before the fruits are collected [26], and these methods are very successful. In addition, there is a need for rapid classification during packaging on the production line. In this context, a method is needed to classify the fruits of walnuts.

After obtaining the image of the walnut fruit, it is possible to extract various features and then classify and analyze them. Apart from this, another method is to classify it directly with deep learning methods or to classify it by presenting the features obtained from deep learning methods to classification algorithms.

In this study, a deep learning-based classification method was developed to classify walnut fruits. In the beginning, four types of walnut images were obtained with a special image-obtaining system with inner lightning. The walnut types were named Chandler, Fernor, Howard, and Oguzlar. By using these images, a dataset has been generated. The dataset consists of 952 images.

In the first experiment, this dataset was used to train and test three different deep learning methods, which are InceptionV3, VGG-19, and VGG-16. Considering the CA values, InceptionV3 showed the highest success with 0.947, while VGG-19 and VGG-16 reached CA values of 0.944 and 0.939, respectively.

In the second experiment, InceptionV3, VGG-19, and VGG-16 are used as feature extraction algorithms. The features obtained by these methods represented the LR, SVM, and k-nearest neighbor (k-NN) algorithms. It has been observed that overall success rates increase with this approach. Among these methods, LR was the most successful method in all trials. The most successful combination was obtained with the combination of InceptionV3 and LR methods. Considering the CA values, VGG-16 + LR achieved 0.951, VGG-19 + LR achieved 0.951, and InceptionV3 + LR achieved 0.961 success rates.

The results show that deep learning-based classification methods serve as effective methods that can be used to develop a classification system using images. Table 18 represents the results in a tabular format to provide an immediate understanding of model performance.

The heatmap graph in Figure 8 provides a comparative view of the metrics for all models and classifiers. It highlights the superior performance of the InceptionV3 + LR combination, achieving the highest CA of 0.961.

As can be seen in Figure 8, the combination of InceptionV3 and LR achieved the highest success in all metrics and offers great potential for increasing automation in the agricultural sector. This method can contribute to economic and environmental sustainability while increasing the efficiency of agricultural processes.

5. Conclusion

This study investigated the application of deep learning methodologies for the classification of walnut varieties based on phenotype images. The research is aimed at enhancing the accuracy and efficiency of automated classification systems, which hold significant potential for agricultural production, quality control, and supply chain management. Two experimental approaches were employed: (i) direct deep learning–based classification using CNN architectures and (ii) feature extraction using deep learning models, followed by classification with traditional machine learning algorithms.

The first set of experiments demonstrated that the InceptionV3 model achieved the highest CA (94.7%) among the three tested CNN architectures (InceptionV3, VGG-16, and VGG-19). The second experimental approach, in which features were extracted using deep learning models and then classified using LR, SVM, and k-NN, resulted in improved classification performance. The best-performing model was InceptionV3 combined with LR, achieving a CA of 96.1%, the highest recorded in this study.

The findings emphasize the effectiveness of deep learning models in agricultural classification tasks, particularly when combined with traditional machine learning techniques. The results also suggest that integrating CNN-extracted features with conventional classifiers can enhance accuracy and computational efficiency, making the approach suitable for industrial applications. The successful classification of walnut varieties using deep learning could facilitate automated sorting, improved quality control, and more efficient agricultural processing pipelines.

Future research directions may include (i) expanding the dataset to include a broader range of walnut varieties, (ii) implementing real-time classification on production lines using edge computing or IoT-based solutions, and (iii) incorporating additional data types, such as spectral or biochemical properties, to further refine CA. The integration of AI-driven classification models into agricultural workflows has the potential to significantly improve automation, reduce manual labor, and ensure consistency in walnut grading and distribution.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding

This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.