A Novel Hybrid Methodology Based on Transfer Learning, Machine Learning, and ReliefF for Chickpea Seed Variety Classification

Abstract

Seed quality is a critical factor in crop production. Therefore, seed classification is required to obtain high-quality seeds and to enhance agricultural sustainability and productivity. This study focuses on the varietal classification of chickpeas, an important source of protein and fiber. Chickpea seed varieties can currently be identified by domain experts; their reliability and efficiency depend on the experience and skills of experts and are prone to human error. The design of classification models with high accuracy to assist in selection mechanisms is required for chickpea varieties. In this study, a novel hybrid methodology is proposed for the chickpea classification problem. This methodology combines three well-suited and robust components: feature extraction using three pre-trained models, feature selection with the ReliefF algorithm, and classification employing classical machine learning methods to enhance classification accuracy and efficiency. Various experiments have been conducted using the four hybrid models developed. Their performance has been compared in terms of accuracy, recall, F1-score, precision, and AUC. TL+SVM and TL+LDA outperformed the other models, with test accuracies of 94.4% and 94%, respectively. These results demonstrate the potential of a powerful model that will be beneficial as a component of computer vision systems in smart agriculture applications.

1. Introduction

The world’s population is growing (consumption is also increasing), the soil is deteriorating, and the climate is changing. Access to food, a fundamental human need and basic motivator, is in jeopardy as a result of all these changes. The first link in the food production chain is agricultural production, which provides raw materials. Agricultural production consists of two parts: animal and crop production. Crop production is crucial because it is the primary aspect in animal production. Legumes obtained from crop production are greatly important for human nutrition and health. Chickpeas are a unique legume that can directly or indirectly support meeting physiological needs. Rhizobium bacteria living symbiotically in their roots affect and improve soil composition by binding nitrogen-free in the air to the soil. Since chickpeas can grow in rain-fed conditions without the need for extra irrigation, this saves water by reducing the use of irrigation in agriculture. As a serious alternative to animal protein, they support the reduction in greenhouse gas emissions from animal production, which has also been on the agenda of the Kyoto Protocol.

Chickpeas, which have around 90 different varieties [1], are widely cultivated worldwide, and their production levels increase with the expanding world population [2,3]. The quality of seeds is important for enhancing chickpea production and productivity. Currently, variety identification of chickpeas can be performed via expert opinion or laboratory analysis. Expert-based classifications are often subjective. It may not always be attainable to elicit expert opinions. This may also cause inaccuracies and challenges when chickpea varieties have highly similar morphological characteristics. Although laboratory-analysis-based classification (especially for certification) is reliable, it has two important disadvantages: (1) the integrity and usability of the analyzed chickpeas can be affected, and (2) the analyzed chickpeas are only a sample and it is debatable how much they reflect the whole. In addition, both methods are time-consuming and impractical when classifying large quantities of chickpeas and/or quick results are desired. To classify chickpea varieties accurately, quickly, and non-subjectively, innovative approaches that can replace the traditional methods or overcome their limitations are needed in the field.

The integration of artificial intelligence into agriculture, particularly crop production, has become an increasing trend. Machine learning, a subset of artificial intelligence, has emerged as a key technology for enabling smart agriculture solutions and has been successfully applied in many areas, such as efficiency [4,5], production quality [6,7,8], crop monitoring [9,10,11], animal welfare [12,13,14,15], disease management [16,17,18], and sustainability [11,19], leading to significant advances in the agricultural sector. Upon collecting data from the field, these data are analyzed with machine learning or deep learning methods to provide actionable insights. Thereby, irrigation requirements can be managed according to the needs of different plant species [4,9,11], the milk yield for each milking animal can be evaluated individually [5], and plant diseases and pests can be detected [17,20]. The use of resources such as farmland can be optimized [10], productivity can be ensured [11], and renewable energy can be integrated [4] for sustainability goals. By accurately and quickly detecting mature crops or plants with small sizes, the specific requirements for smart harvesting can be met [21,22,23]. In addition, product quality and efficiency can be increased by classifying seed samples according to their quality levels and varieties [7,8,24]. This research aims to contribute to addressing the problem of classifying chickpea seed varieties and to support experts during the identification process by proposing a hybrid approach. The major contributions of this research are summarized as follows:

• A novel image dataset “TRCS_8_SET” consisting of eight different certified chickpea seeds grown in Türkiye;
• A novel hybrid methodology that integrates pre-trained deep learning models for feature extraction, feature selection, and classical machine learning methods for classification to achieve superior performance in chickpea seed variety classification;
• Chickpea seed variety classification with a success rate of 94% and above to increase productivity in agricultural production and contribute to seed purity.

This research is organized as follows. The next section summarizes the related works on seed classification. The third section provides a detailed description of the dataset prepared, the methodology proposed, the transfer learning and machine learning methods employed, the feature selection technique applied, and the performance metrics used. In the fourth section, the research findings are reported, exhaustively assessed, and discussed. The final section concludes this study and outlines potential future research on the topic.

2. Related Work

Seed classification in smart agriculture has attracted attention as an important research area related to the powerful integration of image processing, computer vision, and machine learning techniques. This integration accelerates the evaluation process, enhances accuracy (minimizing human error and subjectivity) [24], improves the quality of agricultural products, and serves farmers/consumers as well. Recent research highlights the potential benefits of machine learning techniques in the agricultural industry. In 2021, Elmas [25] applied Convolutional Neural Networks (CNNs) to identify species of mushrooms. The most successful model was presented with an accuracy of 97.62%. In 2022, Adige et al. [26] conducted a study on a total of 120 images with a 512 × 512 resolution from six apple varieties. The experiments were performed with AlexNet and GoogleNet pre-trained models. GoogleNet (91.67%) demonstrated better performance than AlexNet (83.33%). The same year, Jaithavil et al. [27] carried out their study on an image dataset (balanced) of paddy seed varieties. They achieved overall accuracy values of 80%, 83.33%, and 83.33%, respectively, with the VGG16, InceptionV3, and MobileNetV2 pre-trained models. In [28], pistachio images were classified according to their outer shell color. The image dataset was first balanced and experiments were then conducted for different training, validation, and test subsets. AlexNet presented the best results, with an accuracy of 98.44%. In [29], images of five different corn types were classified with various machine learning algorithms. The highest accuracy rates were 96.26% with Multi-Layer Perceptron (MLP) and 96.46% with Support Vector Machine (SVM). In a research study conducted on bread wheat seeds [30], InceptionV3, MobileNetV2, and ResNet-18 produced similar results, with ResNet-18 achieving the highest classification accuracy of 97.7%. Taşpınar et al. [31] employed the InceptionV3, VGG16, and VGG19 models to classify images of dry bean types and achieved a classification success rate of 84.48% with the InceptionV3 model. Secondly, they extracted image features from the layer immediately before the classification layer of each model and then used them as inputs for the Logistic Regression (LR) and SVM algorithms. The images were classified using six different hybrid models, and the best results were obtained using VGG19+LR (83.54%), VGG16+LR (83.71%), and InceptionV3+LR (82.35%). Çakmak and Boyacı [6] proposed an Artificial Neural Network (ANN) integrated computer vision system and classified chickpea quality as good, acceptable, or defective.

Chickpea seed classification, in particular, presents special challenges and opportunities owing to the variety of shapes, sizes, and colors of chickpea seeds. Researchers aim to accurately classify chickpea seeds utilizing CNNs, pre-trained models, and machine learning methods. In 2019, Pourdarbani et al. [1] automatically classified five different types of chickpea seeds using two approaches. In the first approach, chickpea images were segmented, color and texture features were extracted, feature selection was conducted, and classification was performed using a hybrid method called ANN-PSO. In the second approach, patches of chickpea images were directly used as inputs to the ANN. Accuracies of 99.35% and 98.04% were achieved with the ANN and ANN-PSO methods, respectively. In 2021, Taheri-Garavand et al. [32] studied four types of chickpea seeds and used the pre-trained VGG16 model by changing the last three layers in a problem-specific manner. The new model presented an average accuracy of 94.21% as a result of the training. In 2023, Golcuk et al. [33] performed classification experiments on the images of six certified chickpea (cicer arietinum) seed varieties. Accuracy values of 92.3% (MobileNetV2) and 92.97% (MobileNetV2+LSTM) were achieved. The same year, Saha et al. [34] classified chickpea varieties with NasNet-A, MobileNetV2, and EfficientNet-B0. MobileNetV2 demonstrated a significantly high classification accuracy (99%).

A comparison of the literature is provided in

Table 1. Comparison of this work with the literature on the classification of chickpea seeds.

Reference	Chickpea Varieties	Dataset	Total	Methods	Accuracy
[1]	Adel, Arman,	Balanced	1020	ANN+PSO	99.35
	Azad, Bevanij,			ANN	98.04
	Hashem
[32]	Adel, Arman,	Balanced	400	VGG16	94.21
	Azad, Saral
[33]	Atabey, Aydoğan,	Unbalanced	5192	MobileNetV2	92.3
	Bahadır, Goktürk,			MobileNetV2+LSTM	92.7
	Karlı, Tunç
[34]	Alma, Orion,	Balanced	2200	NasNet-A	96.8
	Consul, Cory			MobileNetV2	99.9
				EfficientNet-B0	98.1
[35]	Jam, ILC,	Balanced	400	ANN	79
	Piroz, Kaka
[36]	Alma, Leader,	Balanced	2400	ResNet-50	100
	Frontier, Orion,			MobileNetV2	100
	Palmer, Luna,			GoogleNet	99
	Consul, Cory
[37]	Nihatbey, Çiftçi,	Balanced	1500	CNN-1	94
	Sarı98, Aslanbey,			CNN-2	98
	Aksu
[38]	Nihatbey, Çiftçi,	Balanced	1500	VGG19	97
	Sarı98, Aslanbey,			VGG16	96.7
	Aksu
This	Nihatbey, Çiftçi,	Balanced	7200	TL+SVM	94.4
study	Sarı98, İnci,			TL+LDA	94
	Hisar, Aslanbey,
	Akça, Aksu

. Although the current research has made progress in chickpea seed variety classification, there are still some gaps. More comprehensive datasets (larger in scale and/or including different chickpea types) are needed, and a novel dataset has been introduced in our study. Furthermore, four different hybrid models, namely TL+SVM, TL+LDA, TL+NB, and TL+KNN, have been proposed to address a significant gap in the use of hybrid models for the varietal classification of chickpea seeds. These models leverage the proven benefits of using pre-trained models and machine learning methods together. In addition, the experiments were performed on more images (7200 images in total) compared to the literature. While the reported accuracy may appear lower than that of some studies in Table 1, it is important to note that our dataset includes a larger number and greater variety of images, and this introduces additional complexity into the classification process.

3. Methodology

We introduce a novel hybrid methodology for the successful classification of different chickpea seed varieties through this paper. The methodology adopted integrates three tasks: transfer learning for feature extraction, ReliefF algorithm for feature selection, and classical machine learning methods for classification. This approach combines the power of transfer learning models in automated feature extraction and the generalization ability of classical machine learning methods. It also incorporates ReliefF algorithm to identify/select the most discriminative features and thus provides a more robust and durable consistent structure. Furthermore, this methodology can provide a more effective solution, particularly in agricultural applications with limited training data that have complex traits such as seed varieties, compared to directly using transfer learning models. The utilization of classical machine learning methods instead of the final layers of transfer learning models can reduce the computational cost in the classification stage.

In this study, four different hybrid models have been developed, and a novel dataset has been created. The dataset includes pre-processed images of eight different chickpea types. ResNet-50, EfficientNet-B0, and DenseNet-201 models are used to extract features from the images. Then, the most relevant features have been selected by ReliefF algorithm. The selected features have been fed into the classifier, for example, Support Vector Machine (SVM), K Nearest Neighbor (KNN), Linear Discriminant Analysis (LDA), and Naïve Bayes (NB). Each hybrid model has been called “TL + classifier name”, where TL stands for transfer learning. The performance of the models has been evaluated using accuracy, recall, F1-score, precision, and AUC metrics. Figure 1 illustrates the sequential steps of this research. Each task is elucidated step by step in this section.

3.1. Data Collection and Preprocessing

Chickpeas, the second most consumed legume crop in the world, are cultivated in over 50 countries [39]. They are rich in proteins, vitamins, fiber, and minerals and comprise a high-quality food resource that benefits human health [40,41]. This research, addressing chickpea classification, has conducted experiments on the original dataset named “TRCS_8_SET”. The dataset we created is available at https://github.com/ibrahimkilic/TRCS_8_SET/ [42] (accessed on 31 December 2024). Our dataset includes images (captured in three different directions) of eight different chickpea types registered in Türkiye (visible in Figure 2). These are called Nihatbey, Çiftçi, Sarı98, İnci, Hisar, Aslanbey, Akça, and Aksu.

Certified chickpea seed samples (verified by domain experts) have been obtained from various agricultural research institutes. A controlled environment has been set up to capture images of the samples (see Figure 3). The chickpea images have been obtained with the Arducam Hawk-Eye Camera module with autofocus features integrated into the Raspberry Pi 4–8 GB system from a height of 15 cm under a white LED light of 1050 candela intensity.

The TRCS_8_SET dataset contains a total of 7200 (900 per variety) images of chickpea seed varieties and has a balanced distribution. Each digital image has a resolution of 9152 × 6944. The chickpea images have been pre-processed to improve the quality of the images and isolate chickpea seeds before the classification process. Firstly, the images have been cropped to a resolution of 1200 × 1100 pixels to locate chickpea seeds in the center. Thus, the computational and classification costs are reduced. Then, the background from the cropped images has been removed and the smoothing technique has been applied. The background-removed and smoothed images have been converted to grayscale. Otsu’s method has been used to segment the grayscale images. Python code has been written and executed for image processing steps. OpenCV and Scikit-Image open-source libraries have also been used.

3.2. Transfer Learning and Feature Extraction

Deep learning methods are effective for extracting information from visual data. Transfer learning is often preferred, especially in deep learning studies where the size of the dataset is relatively small. In this study, ResNet-50, DenseNet-201, and EfficientNet-B0 have been used for the extraction of significant features from the original dataset. These transfer learning models, which are widely recognized to demonstrate effective and reliable performance [35,36,44,45,46] for image classification tasks, have strong feature extraction and non-linear fitting capabilities. The 1000 most discriminative features have been extracted from the final fully connected layer of each respective model.

ResNet is a deep learning architecture introduced by K. He, X. Zhang, S. Ren, and J. Sun in their paper titled “Deep Residual Learning for Image Recognition” that addresses the degradation problem [47]. In order to reduce the vanishing/exploding gradient problem and facilitate the learning of deep networks, ResNet includes residual blocks connected by shortcuts (skipping one or more layers) that transfer the convolution block input directly to the output before activation [43,47]. This network can be used to extract more information from the original data [47,48]. The ResNet-50 (a 50-layer ResNet) architecture has been employed in this study. It is one of the most widely used deep neural networks [49,50] and has been proven to be highly accurate for seed classification [35,51,52].

DenseNet was introduced in 2017 with the paper titled “Densely Connected Convolutional Networks” [53] and is a stack of dense and transition blocks [54]. DenseNet contains dense connections between convolutional layers to increase network performance and helps to transmit crucial data from the first layer to the final layer [55]. The numerical definitions in the name of DenseNet network architectures indicate the model depth, in other words, the number of layers [38]. The DenseNet-201 architecture has been used in this study for its capability of deep feature extraction and remarkable performance [50,55,56].

EfficientNet model family was developed by M. Tan and Q. V. Le in their 2019 paper “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks”. EfficientNet network architecture proposes a compound scaling method to achieve much better performance with fewer parameters and lower computational costs (efficiency). This method uniformly scales the resolution, width, and depth dimensions of CNNs [57]. In this study, we have experimented with the EfficientNet-B0 scaled-up version, which was particularly chosen for its suitability for resource-constrained environments [57,58] considering the practical needs of agricultural applications.

3.3. ReliefF Feature Selection

ReliefF algorithm was preferred to handle noisy and redundant features in the dataset and to select the most important features. Thus, it is aimed to increase the classification accuracy while reducing the computational complexity. ReliefF is a feature selection technique particularly designed to handle multi-class problems and deal with high-dimensional data [59,60]. The pseudocode of the ReliefF technique [61] is presented below.

• Inputs: Dataset with instances and features, k.
• Output: Weight features.
• Initialize weights.
• Randomly select a subset of instances.
• Find k nearest hits and misses for each selected instance.
• Update the weights of features using the differences between instances and feature values.
• Normalize the weights to fall within a certain range.

3.4. Classification Using Machine Learning

Four well-known machine learning methods have been used alongside transfer learning models: (1) KNN, which is based on similarities between data points, (2) SVM, which is based on mathematical model inference and boundary lines, (3) NB, which is based on conditional probability principles, and (4) LDA, which is based on statistical relationships. These methods have been used in this study because of their simplicity, effectiveness, and wide applicability. They have been applied in the classification experiments carried out on the final dataset obtained after the feature extraction and selection phases.

KNN [62] classifies an unknown instance point according to the classifications of the nearest k points from a set of previously classified instances. It uses various distance measures such as Manhattan, Euclidean, and Cosine to determine these k neighbor points.

SVM [63] is designed to find the optimal hyperplane separating the data into different classes and maximizing the margin between the classes [64]. Data points closest to the hyperplane are defined as support vectors. It can be efficiently used for both linear and non-linear classification tasks.

NB classifier [65] is based on Bayes’ probability theorem with the conditional independence assumption. It calculates the probability of an instance belonging to a specific class.

LDA [66,67] can be used for identifying the most significant features with their level of significance and also classifying. It aims to maximize the separability of instances by finding linear combinations of features, in other words, by computing linear discriminant functions. The best separation of classes can be achieved by maximizing the distance between the means of the classes and minimizing the variance in each class [68,69].

3.5. Performance Evaluation

Accuracy, precision, recall, F1-score, and area under the characteristic curve (AUC) metrics have been used to evaluate the classification performance of the pre-trained models. These metrics include the definitions of true positive (TP, the number of positive instances correctly predicted as positive), true negative (TN, the number of negative instances correctly predicted as negative), false positive (FP, the number of negative instances correctly predicted as positive), and false negative (FN, the number of positive instances correctly predicted as negative).

Accuracy is a widely used metric to assess classification performance. It is the ratio of correct predictions to all predictions and is calculated by Equation (1).

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(1)

Precision is the ratio of correct positive predictions to all positive predictions. It is formulated as Equation (2) and used to measure the accuracy of positive predictions.

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(2)

Recall (True Positive Rate, TPR) is the ratio of correct positive predictions to all positive instances. It measures the ability to correctly expose true positive instances, and its formula is as follows:

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(3)

F1-score is the harmonic mean of the Precision and Recall values, as provided in Equation (4). This composite metric is a robust evaluation metric for especially imbalanced datasets.

$$F1-score=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(4)

AUC is the area under the ROC (Receiver Operating Characteristic) curve. ROC curve plots the TPR (y-axis) against the False Positive Rate (x-axis). A larger area (close to 1) indicates better classification performance.

$$FalsePositiveRate\left(FPR\right)={\displaystyle \frac{FP}{FP+TN}}$$

(5)

4. Results and Discussion

In this study, four different hybrid models (TL+SVM, TL+LDA, TL+NB, and TL+KNN) were proposed for the multi-classification of chickpea seeds. Considering a commonly used split in machine learning applications, the dataset is divided into two subsets: 80% for training and 20% for testing. Z-score normalization was used for standardizing the data. Via this powerful technique, the data were standardized to have a mean of 0 and a standard deviation of 1. The training and test experiments were performed on a notebook with an AMD Ryzen 7 3750H processor, an RTX2060 GPU, and 16 GB DDR4 2400 RAM running at 2.3 GHz. The training and testing of the hybrid models were performed via MATLAB R2022b. The 5-fold cross-validation technique was used to evaluate and validate the performance of the models during model training. The parameters of the models are provided in

Table 2. Parameters used in the models.

Method	Parameter(s)
KNN	Distance: Euclidean
	Distance Weight: SquaredInverse
	Number of Neighbors: 10
	Standardize: true
SVM	Kernel Function: polynomial
	Polynomial Order: 2
	Kernel Scale: auto
	Box Constraint: 1
	Coding: onevsone
	Standardize: true
NB	Distribution Names: kernel
LDA	Gamma: 0
	Discrim Type: Linear
	Fill Coeffs: off
ReliefF	k = 10

.

4.1. Experimental Results

The classification reports of the hybrid models are presented in

Table 3. The training and test results of the hybrid models.

	Accuracy		Weighted		Weighted		Weighted
			F1-Score		Precision		Recall
Model	Training	Test	Training	Test	Training	Test	Training	Test
TL+KNN	0.9007	0.8951	0.9012	0.8950	0.9035	0.8959	0.9007	0.8951
TL+SVM	0.9356	0.9438	0.9360	0.9438	0.9365	0.9442	0.9356	0.9438
TL+NB	0.8453	0.8653	0.8445	0.8644	0.8449	0.8651	0.8453	0.8653
TL+LDA	0.9476	0.9403	0.9480	0.9407	0.9488	0.9415	0.9476	0.9403

. The proposed TL+SVM and TL+LDA hybrid models demonstrated remarkable results, achieving high test accuracy rates of 94.4% and 94%, respectively. In the training set, the TL+LDA model provided the highest classification accuracy at 94.8%. Considering seed purity as the main target, the most successful model is the TL+LDA model (precision of 94.9%) on the training data set and the TL+SVM model (precision of 94.4%) on the test set. The TL+KNN model achieved high precision (89.6%), recall (89.6%), F1-score (89.5%), and accuracy (89.5). The weighted precision of the TL+NB model is 86.5% and that of TL+KNN is 89.6%.

As can be observed, the TL+SVM and TL+LDA models outperformed the other hybrid models in terms of training and test accuracy rates. Figure 4 presents four hybrid models’ performance on training and test data in terms of AUC values. In the training phase, the AUC value of TL+SVM is the highest, reaching about 0.997, followed by TL+LDA (0.995), TL+KNN (0.9898), and TL+NB (0.946). In the test phase, again, the highest AUC value was produced by TL+SVM (about 0.998) and the lowest AUC value by TL+NB (0.952). These results indicate that TL+SVM has the highest predictive ability.

Table 4. TL+LDA results.

No.	Class	F1-score	Precision	Recall	AUC
1	Nihatbey	0.826	0.809	0.844	0.9854
2	Çiftçi	0.857	0.848	0.867	0.9903
3	Sarı98	0.924	0.942	0.906	0.9936
4	İnci	0.986	0.983	0.989	0.9974
5	Hisar	0.997	1.000	0.994	0.9997
6	Aslanbey	0.983	0.978	0.989	0.9986
7	Akça	0.983	0.989	0.978	0.9975
8	Aksu	0.969	0.983	0.956	0.9965

and

Table 5. TL+SVM results.

No.	Class	F1-score	Precision	Recall	AUC
1	Nihatbey	0.847	0.833	0.861	0.9859
2	Çiftçi	0.867	0.884	0.850	0.9907
3	Sarı98	0.943	0.925	0.961	0.9964
4	İnci	0.981	0.978	0.983	0.9997
5	Hisar	0.997	1.000	0.994	1.0000
6	Aslanbey	0.966	0.977	0.956	0.9990
7	Akça	0.981	0.978	0.983	0.9998
8	Aksu	0.969	0.977	0.961	0.9987

provide a detailed class-by-class analysis of the other performance metrics. The TL+SVM model obtained better classification results for all the varieties of chickpea seeds compared to the TL+LDA model. Hisar exhibits nearly perfect precision, recall, F1-score, and AUC values for these two models. Akça, İnci, and Aslanbey also stand out with high performance across every metric. TL+SVM exhibited considerably high training and test AUC values for the Hisar and İnci classes. The Akça class demonstrated a higher level of performance in terms of AUC. These results also show that both models have low AUC (0.9854 and above) for Nihatbey. Çiftçi and Nihatbey have relatively lower precision and recall compared to the other classes. This may involve examining the data distribution, feature selection, and also further model tuning.

4.2. Discussion

A seed is both the source/starter of agricultural production and the propagation matter of many crops. Access to high-quality seeds is important for the agriculture and food sectors. In the literature, research has been conducted on the quality status of chickpea seeds, the determination of their purity from foreign substances, and the classification of their varieties using artificial intelligence techniques. In studies targeting chickpea variety classification, machine learning algorithms, deep learning approaches, and optimization-based hybrid methods are used. In [35], the ANN method was used, and four types of chickpeas were classified with 79% accuracy. In [1], the ANN and PSO methods were used together and the classification of five chickpea varieties was performed with 98% accuracy. In [32,36], using pre-trained models, an accuracy of 94.21% was achieved by the VGG16 model in [32] and 100% by the ResNet50 model in [36]. In a classification of five chickpea varieties [38], the ResNet-50, DenseNet-201, and EfficientNet models presented training accuracies of 93.9%, 97.2%, and 92.3%, respectively. Their test accuracies were 95.7%, 93%, and 90%, respectively.

In studies using hybrid methods based on pre-trained CNN models, it was stated that an F1-score of 86.2% was obtained for seven types of chickpeas with the SVM-based model [70], accuracy of 92.97% for six types of chickpeas with the LSTM-based model [33], and accuracy of 99.10% for three types of chickpeas with the ensemble-learning-based model [71]. The hybrid models proposed in this research, especially including the SVM and LDA machine learning algorithms, can be utilized for the classification problem of eight chickpea varieties with an accuracy of 94% and over. As a result, a fundamental requirement for increasing the production and use of high-quality chickpea seeds was satisfied.

For more accurately classifying chickpea seeds or improving the efficiency of transferring information (extracting discriminative features), a method can be applied to block the convolution-based progressive training proposed by [72]. On the other hand, testing the proposed hybrid models on different datasets in future work may increase their applicability. This could further increase the value of this study and contribute to the field.

5. Conclusions

In this study, we prepared a novel dataset of chickpea seed images and used it for multiple classifications of chickpea seed varieties. We also developed four different hybrid models named TL+SVM, TL+LDA, TL+NB, and TL+KNN (including the ReliefF algorithm for feature selection). Comprehensive experiments were performed with these hybrid models, and comparative analyses were conducted. The seed varieties were also classified with high accuracy by employing the models. As evident in Table 3, the research results show that the hybrid models, especially those containing the SVM and LDA machine learning methods, can be implemented with an accuracy rate of 94% and above for the chickpea variety classification problem. For future work, we will improve the classification accuracy, achieve superior performance for the dataset, and address binary classification for the selected chickpea variety. This research can be further expanded by employing generative deep learning methodologies like generative adversarial networks. The utilization of transfer progressive training is recommended to improve the models’ ability to extract information pertaining to local details.

The TL+SVM, TL+LDA, TL+NB, and TL+KNN models developed for classification present a robust and scalable solution for overcoming the limitations of manual classification and addressing the challenges of seed classification in agricultural and industrial applications. This study also highlighted these models’ potential in chickpea seed classification tasks. These hybrid models can be used to increase productivity in agriculture and to promote/determine sustainable agricultural policies. They can also be integrated as a component of a mechanical selection system, a mobile environment, or a computer-aided chickpea classification process to aid in the real-time classification of chickpea varieties.