Design of Efficient Methods for the Detection of Tomato Leaf Disease Utilizing Proposed Ensemble CNN Model

Abstract

Early diagnosis of plant diseases is of vital importance since they cause social, ecological, and economic losses. Therefore, it is highly complex and causes excessive workload and time loss. Within the scope of this article, nine tomato plant leaf diseases as well as healthy ones were classified using deep learning with new ensemble architectures. A total of 18.160 images were used for this process. In this study, in addition to the proposed two new convolutional neural networks (CNN) models, four other well-known CNN models (MobileNetV3Small, EfficientNetV2L, InceptionV3 and MobileNetV2) are used. A fine-tuning method is applied to the newly proposed CNNs models and then hyperparameter optimization is performed with the particle swarm optimization algorithm (PSO). Then, the weights of these architectures are optimized by the grid search method and triple and quintuple ensemble models are created and the datasets are classified with the help of the five-fold cross-validation. The experimental results demonstrate that the proposed ensemble models stand out with their fast training and testing time and superior classification performances with an accuracy of 99.60%. This research will help experts enable the early detection of plant diseases in a simple and quick manner and prevent the formation of new infections.

1. Introduction

Day by day, the world population is increasing. In proportion to this population growth, the need for agriculture increases expeditiously. The products obtained from agriculture can meet many human needs, such as eating, dressing, and heating. These products are also very essential for the import and export issues in a country. Income from agriculture plays an important role in the development and growth of the country’s economy. In this respect, the products obtained from plants must be of high quality, and the plants must be protected from diseases in order to have high-quality products. There are many factors that cause plant diseases, such as adverse environmental conditions, fungi, bacteria, and viruses. Diseases seen in the plant can damage the vital functions of the plant, such as photosynthesis, pollination, fertilization, and germination. Therefore, for treatment, it is very important to detect diseases as early as possible.

Today, instead of relying only on an outside expert, it is possible to use technological devices to determine whether a plant has a disease and what kind of disease it is. Processes such as object detection, classification, image processing, and artificial intelligence algorithms provide extremely good outcomes as the quality of image acquisition by technological devices improves. Machine learning (ML) and deep learning (DL) outperform traditional optimization and prediction methods. First, these methods can automatically learn from large amounts of data, while traditional methods require manual feature extraction and are limited by data size. Second, ML and DL models can generalize well to unseen data, unlike traditional methods. Thirdly, unlike traditional methods, ML and DL models can learn complex, non-linear data relationships. Thus, ML is better at handling multiple variables, especially complex interactions. Especially in recent years, artificial intelligence applications have been widely used in various application areas such as communication, construction, magnetic, physics, and biomedical systems [1,2,3,4,5]. In this context, accurate, timely detection and classification of plant diseases are of great importance [6]. Advances in artificial intelligence research now enable automatic plant disease detection from raw images [7].

A considerable amount of research has been conducted in the field of disease detection in plants so far. The studies were generally carried out using ready-made data sets, ready-made models, and libraries. An algorithm on the image segmentation technique used in the automatic detection and classification of diseases in plant leaves was presented by Singh and Misra [8]. The average success rate of their study was conducted on five different diseases by using a genetic algorithm in the disease detection stage, which is 97.6%. Zhang et al. performed disease detection on the cucumber leaf database [9]. In this study, they extracted shape and color features to detect diseases by segmenting diseased leaves with the k-means clustering algorithm. They obtained an 86% accuracy rate by classifying these diseased leaves with the sparse representation method. Convolutional neural network (CNN) models were developed by Ferentinos to detect and diagnose plant diseases with deep learning methods [10]. The training of the models was performed on the data set containing 87,848 images. In addition, this dataset includes 58 (plant, disease) combinations of 25 different plant species. The results of the study show that it has a success rate of 99.53%.

Lu et al. [11] proposed a disease detection method in rice based on deep CNN models. The training stage was performed using a dataset consisting of 500 images of healthy and diseased rice leaves. The trained model detects 10 different common diseases. According to the experimental results reported, the success rate in detecting the disease was 95.48%.

Cruz et al. [12] submitted a certain bacterial species and proposed a method for detecting diseased olive leaves. It is intended to detect this disease in olive leaves by retraining the previously trained deep learning model with transfer learning. As a result, it was seen that the success rate was approximately 98.60%. Mohanty et al. used the deep CNN model for the analysis of plant leaf diseases [13]. In this study, two models based on deep learning (AlexNet and GoogleNet) were selected. Using the dataset of 54,306 images of diseased and healthy plant leaves, these two CNNs were trained to classify crop types and disease status for 38 different classes, including 14 plant species and 26 diseases, with over 99% success. Hanson et al. presented an approach to detect and classify disease in leaves using deep learning techniques. As a result, success rates between 91% and 98% were obtained for separate class tests in this study. The final overall success of the trained model was 95% [14]. Namin et al. A new deep learning structure has been proposed to recognize the plant phenotype by combining the advantages of the CNN and LSTM (Long Short-Term Memory) algorithms [15]. According to the developed algorithm, deep features were extracted from the images from the CNN model, and the output was loaded into an LSTM unit to create a series of models. The results of the study showed that the sequence model significantly increased the accuracy from 68% to 76.8% when the handcrafted features were used when using CNN, and the addition of temporal information with LSTM increased the accuracy to 93%. Too et al. The use of different deep learning models was evaluated for the classification of 38 different classes, including the VGG-16 network, Inception V4, ResNet with 50, 101, and 152 layers, and DenseNet with 121 layers. The PlantVillage dataset included healthy and diseased leaf images of 14 different plants [16]. The DenseNets 17 model’s success gradually increased as the number of rotations increased, and there were no signs of overtraining or performance degradation. DenseNets used very few parameters and a reasonable computation time to achieve their state-of-the-art performance. This model outperformed other models with a test success rate of 99.75%. Models are trained using Theano and Keras backends.

There are numerous studies on the disease detection of tomato plants in the literature [17,18,19]. Rajasekaran Thangaraj et al. used transfer learning and fine tuning to modify the Xception architecture to yield good results on a smaller dataset [20]. As the dataset, 50,306 images from the PlantVillage dataset, which contains a total of 14 different disease classes and healthy classes, were used. Before the classification step, the model Adam, SGD, and RMSProp optimization methods were tested, and the best accuracy of 99.55% was achieved with Adam.

Patrick M. K. et al. [21] utilized the AlexNet, GoogleNet, and Gabor CapsNet CNN architectures for the classification of both tomato leaf diseases and citrus plant diseases. They found that the model presented gave poor results in the classification of images with mixed backgrounds.

Aliyu Muhammad A. et al. used ML to analyze pathologically local symptom lesions in tomato and potato leaves [22]. The authors experimented with KNN, naive Bayes, and SVM as classifiers and obtained the highest accuracy of 99.67% using SVM.

Automatic identification of tomato plant diseases is the main motivation for this work. In terms of filling a void in the field, the contributions of this paper include the following:

• Diagnosing plant diseases is essential for determining the quality and health of a crop, as it evaluates numerous parameters such as crop yield, yield capacity, grain quality, and nutrient retention.
• Suggesting new CNN models and creating new ensemble structures with the proposed CNNs models.
• The stratified 5-fold cross-validation method is used for classification problems, and the folds are selected so that each fold contains roughly the same ratios as the target class.
• Obtaining a higher accuracy ratio compared to the studies carried out in the literature. In particular, the architectures commonly used in the literature for the detection of tomato disease are selected and proposed CNN models and ensemble models are presented. After these systems are designed, it is intended to contribute to the literature by increasing the accuracy of the models.
• Automatically identifying tomato plant diseases.
• Achieving fast training and testing times and superior classification performance.

The structure of the paper is provided below. Section 2 provides a comprehensive explanation of the deep learning methods, proposed CNN models, and other CNN models available in the literature, as well as a summary of the employed algorithms. In addition, evaluation parameters for the utilized deep learning models are described in detail. In Section 3, the work and outcomes of deep learning models are detailed, along with evaluation findings such as ROC, accuracy, recall, precision, and F1-scores, as well as cross-validation outcomes. Additionally, a comparison is made between the proposed methods and those of other studies employing the same dataset, and results are presented. Section 4 contains the conclusion.

2. Material and Methods

In this section, the methods used to perform the classification of tomato plant leaf disease detection, optimization methods, evaluation parameters, and dataset are briefly explained. The procedure of the study is shown in Figure 1.

2.1. Deep Learning

Deep learning is a methodology for ML that employs numerous nonlinear information processing layers and is a method for feature extraction, pattern analysis, and classification in supervised and unsupervised learning [23,24].

The greatest benefit of this method is that the features are taken directly from the raw data, and the classification algorithm trains itself automatically based on the collected characteristics [25]. Raw pictures are incompatible with conventional image categorization and processing methods. This necessitates several preprocessing techniques and feature extraction algorithms. This is an extremely difficult and experience-required process. In the beginning, researchers aimed to eliminate human feature extraction in favor of the most effective automatic feature extraction. Despite the fact that this learning process was implemented using artificial neural networks with several layers, there were substantial shortcomings in areas such as image processing. To accomplish this, scientists developed the convolutional neural network (CNN), which is compatible with the way images are constructed.

2.1.1. Proposed Convolutional Neural Networks (CNN)

In this part, several new CNN models are designed, and the two models with the highest accuracy among these models are selected and used in the classification process. The two CNN models proposed within the scope of this study contain fewer parameters than the existing models in the literature and also have less storage space. This is because the proposed CNN models are designed with fewer layers. An example of CNN architecture is shown in Figure 2.

The architectural structure of the proposed CNN1 and CNN2 models’ layers and parameter values used are given in

Table 1. Architecture details of CNN1 and CNN2.

Layer	CNN1	CNN2
input	224 × 224 × 3	224 × 224 × 3
Conv. + ReLU	Filter size: 3 × 3	Filter size: 3 × 3
	Number of filters: 32	Number of filters: 16
Max-pooling	Filter size: 2 × 2	Filter size: 2 × 2
Conv. + ReLU	Filter size: 3 × 3	Filter size: 3 × 3
	Number of filters: 64	Number of filters: 32
Max-pooling	Filter size: 2 × 2	Filter size: 2 × 2
Dropout	Dropout rate: 0.3
Conv. + ReLU	Filter size: 3 × 3	Filter size: 3 × 3
	Number of filters: 128	Number of filters: 64
Max-pooling	Filter size: 2 × 2	Filter size: 2 × 2
Dropout	Dropout rate: 0.4
Conv. + ReLU	Filter size: 3 × 3	Filter size: 3 × 3
	Number of filters: 256	Number of filters: 64
Max-pooling	Filter size: 2 × 2	Filter size: 2 × 2
Conv. + ReLU		Filter size: 3 × 3
		Number of filters: 32
Max-pooling		Filter size: 2 × 2
Flatten	36,864	800
Fully connected	128	512
Dropout	Dropout rate: 0.5
Dense	10	10
Output	C0–C9	C0–C9

. We used a typical CNN architecture with 4 (CNN1) and 5 (CNN2) sequential Convolution/Max-pooling modules, followed by a flattening layer, a fully connected layer, a dropout layer, and a SoftMax layer.

In general, CNNs are able to achieve good performance with relatively small model sizes due to their ability to learn features directly from the input data and their use of shared weights in the convolutional layers. This can allow them to be trained faster and require less storage compared to models with a larger number of parameters, such as fully connected deep neural networks. Thus, the objectives of this section are to design new CNN models that require less storage space and training time than the other models proposed in the literature.

2.1.2. CNN Models

Convolutional neural networks (CNNs) are the type of deep learning models that are particularly well-suited for image classification tasks. They are able to learn features directly from the input data and have been successful in a wide range of applications, including object recognition, face detection, and image generation.

MobileNet

Howard et al. developed the MobileNet architecture to run vision applications on embedded and mobile platforms [26]. The proposed algorithm is based on convolutional layers that are depth-wise distinct. Even though the networks have fewer parameters than other deep networks, they are still capable of producing successful results. This model is powered by convolutional layers with granular separation. Deeply distinct convolution layers map spatial and inter-channel correlations in input image feature maps. In this study, MobileNetV3Small and MobileNetV2 architectures are used for classification.

EfficientNet

EfficientNet is a network architecture for deep learning that was developed in 2019 [27]. This architecture illustrates the relationship between three terms that significantly impact the performance of architectures for deep networks. These are defined as depth, width, and resolution. This architecture uses the composite scaling method. The grid search algorithm is this method’s first step. This algorithm enables the network to correlate between scaling sizes of varying sizes.

Inception

The InceptionV3 model is an architecture that was created by enhancing previously developed Inception architectures. It is designed to accomplish more with fewer resources than its predecessors. The GoogleNet model is a deep CNN network that was proposed by Google researchers in 2014 and achieved 93.3% accuracy in the top five in ILSVRC [28]. The GoogleNet architecture, which consists of 22 layers and utilizes a network at the network layer as opposed to conventional sequential processing, is based on the Inception model. Parallel computing is used to calculate the pooling, major convolution, and minor convolution layers. Individual convolutions are used to reduce the dimensionality of features in the Inception architecture. Due to this architecture’s size reduction and parallelism, the number of parameters and operations has been drastically reduced; as a result, these features conserve memory and reduce computational costs [29].

2.2. Dataset

The PlantVillage dataset is a well-known dataset in the field of plant disease detection and consists of a large number of images of plant leaves along with labels indicating the type of disease (if any) present in the image. Using this dataset to train a CNN model for tomato plant disease detection can be a useful way to demonstrate the effectiveness of the model in detecting various types of diseases and can help establish a baseline for comparison with other approaches.

In this study, the dataset composed of images of tomato plants gathered from the PlantVillage dataset “https://data.mendeley.com/datasets/tywbtsjrjv (accessed on 3 March 2022)” was used [30]. This dataset consists of images of plant leaves classified into ten categories as healthy and diseased for the tomato plant species and includes a total of 18,160 leaf images. The images are in the RGB color space and have 256 × 256 pixels. Example images of each class are shown in Figure 3. The total numbers of training, validation, and test images used for each class are given in

Table 2. The Total Number of training, validation, and test images are used for each class.

Class Label	Total Data	Training and Validation Data	Test Data
Bacterial spot	2127	1585	542
Early blight	1000	766	234
Healthy	1591	1122	469
Late blight	1909	1662	247
Leaf mold	952	544	408
Septoria leaf spot	1771	1306	465
Spider mites two-spotted spider mite	1676	1323	353
Target spot	1404	69	1335
Tomato mosaic virus	373	269	104
Tomato yellow leaf curl virus	5357	4974	383

.

2.3. Tuning

Identifying the optimal (hyperparameter) settings for each model is essential for a bias-free evaluation of a model’s predictive ability. In this study, the algorithms used for tuning are given below.

2.3.1. Fine Tuning

Fine tuning refers to the process of training and applying a model that has been trained on one set of data to another set of data. Fine tuning is a concept in transfer learning. To accelerate learning, CNN models are fine-tuned to identify and classify nine plant disease categories using four pre-trained models on the ImageNet dataset. Fine tuning a model involves adjusting its layers in order to improve its performance. The goal of fine tuning is to find the optimal model that gives the best performance on the planned task, and this method is used in our study to adjust our proposed models.

Figure 4 shows the implementation of a fine-tuned model for CNN. In this section, firstly, the weights of the CNN model are fixed. Then, the updating and fine tuning of the weights of the last two layers is carried out using transfer learning with the VGG-16 network.

2.3.2. Hyperparameter Optimization with Particle Swarm Optimization (PSO)

The different parameters utilized to train the neural network model in a deep learning model are referred to as “hyperparameters”. In order to improve the performance of a neural network, these hyperparameters are tuned as if they are buttons, resulting in an optimized model. The hyperparameters include the number of hidden layers, the number of units or nodes in the hidden layer, the learning rate, the dilution rate, the epochs or iterations, adam, nadam, adagrad, rmsprop, etc., and optimizers such as ReLu, sigmoidal, etc.

A deep learning model with fine-tuned hyperparameters will provide the most effective model performance optimizers, activation functions, learning coefficients, etc. It is the procedure of determining the value of hyperparameters, such as in manual search, grid search, random search [32,33], and Bayesian optimization [34,35]. The scheme of PSO is shown in Figure 5. In this paper, the PSO algorithm is utilized for hyperparameter tuning.

2.3.3. Grid Search

In the grid search approach, there is a grid space where different possible values of hyperparameters are defined. After testing all combinations in the search space, the best-performing point is selected. Therefore, certain ranges are determined for these parameters. The model is trained with values in the determined ranges, and the best combination group obtained by observing the results is determined as the hyperparameters of the model [30]. Weight optimization with a grid search algorithm is given in Figure 6.

2.4. Evaluation Metrics

The metrics described in the equations given below are commonly used to evaluate the performance of a classification model, such as a convolutional neural network (CNN) used for plant disease detection. These metrics provide different insights into the performance of the model and can help identify strengths and weaknesses in its ability to classify images correctly.

The confusion matrix is used in this study to assess the performance of the proposed systems.

• The true-positive (tp) value represents the number of correctly identified infected image samples.
• False-positive (fp) indicates the number of samples incorrectly identified as infected.
• True-negative (tn) identifies correctly classified non-infected plants.
• False-negative (fn) identifies individuals who are incorrectly classified as healthy.

Depending on the earlier statuses, statistical measures are used to have a complete vision of the model’s performance. Performance measures such as accuracy, recall, precision, and F1-score are used using the confusion matrix. Mathematical expressions related to these performance criteria are given below [37]:

$$Accuracy=\frac{tn+tp}{tn+tp+fn+fp}$$

(1)

$$Recall=\frac{tp}{tp+fn}$$

(2)

$$Precision=\frac{tp}{tp+fp}$$

(3)

$$F1\mathit{-Score}=2\times \frac{Precision\times Recall}{Precision+Recall}$$

(4)

3. Experimental Results

In this section, detailed modeling studies with deep learning architectures are given for each problem, using the data in the original data set created for tomato plant disease prediction and classification problems. The configuration of the hardware and size of the used models are given in

Table 3. Configuration of the hardware.

Name	Parameter
Memory	64 GB
Processor	Intel® Xeon(R) Silver 4114 CPU @ 2.20 GHz × 40
Server model	Hp z6 g4
Graphics	NVIDIA Corporation GP102 [GeForce GTX 1080 Ti]
OS	Ubuntu 20.04.5 LTS
Language	Python 3
Framework	Jupyter Notebook

and

Table 4. Size of used models.

Models	Size
MobileNetV3Small	224 × 224 × 3
EfficientNetV2L	224 × 224 × 3
InceptionV3	224 × 224 × 3
MobileNetV2	224 × 224 × 3
CNN1	224 × 224 × 3
CNN2	224 × 224 × 3

, respectively.

In this study, we use the cross-validation methods to eliminate the over-learning issue [38], and the dataset is separated into training and testing groups. The data utilized for training and testing groups are distinct. Thus, in the testing phase of the algorithm, data distinct from those used in the training phase are utilized. In other words, the algorithm is tested with data that it does not know. Thus, a more realistic testing procedure and accuracy rate are obtained.

Tuning a model involves adjusting its hyperparameters in order to improve its performance. This can be achieved by manually adjusting the hyperparameters or by using an optimization algorithm such as the particle swarm optimization (PSO) algorithm described in the study. The goal of tuning is to find the optimal set of hyperparameters that gives the best performance on the planned task.

Using two CNN models and four ready-made models proposed in this study, the detection of tomato plant disease is carried out by the 5-fold cross-validation method. The proposed CNN1 and CNN2 models were fine-tuned in order to improve their accuracy, as shown in Figure 4. In addition, the hyperparameter values of the models are optimized with the help of the PSO algorithm (See Figure 5) for the hyperparameter setting of these two models, as shown in

Table 5. Hyperparameters of the models.

Layers	Hyperparameters	CNN1	CNN2
Convolution	Number of convolution layers	4	5
Pooling	Number of pooling layers	4	5
Dropout	Number of dropout layers	3	0
Flatten	Number of flatten layers	1	1
Fully connected	Number of fully connected layer	1	1
	Number of filters	32, 64, 128, 256	16, 32, 64, 64, 32
Convolution	Filter size	3 × 3	3 × 3
Pooling	Filter size	2 × 2	2 × 2
Dropout	Dropout rate	0.3, 0.4, 0.5
Flatten	Vector length	36,864	800
Fully connected	Number of neurons	128	512
Training optimizers	Learning rate	0.0001	0.0001
	Optimizer	Adam	Adam
	Batch size	4	4
	Epoch	50	50

.

In this study, ensemble models are proposed for the methods given above to classify 10 classes of tomato plant diseases. In the ensemble models proposed here, a method based on weighted average ensemble learning is used and given in Figure 6. The best combinations of base models in the proposed system are found using the grid search method, and the effect of each base model on the average recall score is presented as weighted, unweighted, and ideal. In addition, the accuracy results are given in

Table 6. Performance metrics of deep learning architectures.

Deep Learning Architectures	Parameters	Storage Space	Epochs Required to Train the Model	Training Time (in Hours)	Testing Time (in Seconds)	Training Accuracy	Validation Accuracy	Testing Accuracy	Precision	Recall	F1-Score
Mobile NetV3Small	1,234,547	55.6 MB	15	2.76	5.70	98.91%	99.9%	98.30%	0.98	0.98	0.982
EfficientNet V2L	117,911,203	1.4 GB	15	15.39	38.68	99.17%	99.71%	98.79%	0.99	0.99	0.988
InceptionV3	22,065,443	266 MB	15	3.51	10.35	97.06%	95.82%	96.1%	0.964	0.96	0.96
MobileNetV2	2,422,339	29.7 MB	15	2.32	7.0	98.34%	96.94%	96.99%	0.974	0.974	0.974
CNN1	5,108,426	5.1 MB	50	1.38	3.69	96.39%	99.03%	95.45%	0.956	0.954	0.96
CNN2	494,218	6 MB	50	1.37	3.49	97.61%	99.48%	96.87%	0.97	0.968	0.97

with other used models.

When examining Table 6 for the application conducted within the scope of this study, four distinct CNN models (from the literature) are offered for comparison in addition to the two proposed models. The results for all classes of tomato plant leaf disease in this study are given in Figure 7 separately. When Figure 7 is examined, performance results (accuracy, precision, recall, and F1-Score) are given on a class basis. In addition, it is seen that each class produces values very close to the average success. It is seen that ROC and AUC values are generally one. When these metrics are evaluated, it is seen that the different sample sizes on a class basis do not negatively affect the results of the study.

In this part, models (CNN1–CNN2) are proposed that take up significantly less space (5.1 and 6 MB) than models in the literature for two parameters (space and training/testing time). These models reduce training time from approximately 16 h to nearly 1.5 h, and the testing times of the two CNN models are faster than other models. The low value of the store parameter reveals the significance of the offered models, particularly when portable systems are considered in daily life. In addition, the accuracy of the presented models is roughly equivalent to that of the existing models. Compared to other models, the CNN2 model presented within the scope of this study has a relatively small number of total parameters. The performance plot of the CNN models is shown in Figure 7.

The values in the tables given above are created by calculating the arithmetic average of the classification reports obtained as a result of the five-fold cross-validation of each CNN model. The accuracy and loss graphs of the CNN models proposed within the scope of the study are obtained for k = 5 and shown in Figure 8 and Figure 9. In addition, the confusion matrix of the models is given in

Table 7. Confusion matrix of used CNN models.

	C0	C1	C2	C3	C4	C5	C6	C7	C8	C9
Mobile NetV3Small	98%	0%	0%	0%	0%	0%	0%	0%	0%	0%	C0
	0%	94%	3%	0%	0%	0%	0%	0%	0%	0%	C1
	0%	1%	99%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	99%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	99%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	99%	1%	0%	0%	0%	C5
	0%	0%	0%	0%	0%	1%	98%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
EfficientNetV2L	100%	0%	0%	0%	0%	0%	0%	0%	0%	0%	C0
	0%	95%	3%	0%	0%	0%	0%	0%	0%	0%	C1
	0%	2%	97%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	99%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	100%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	100%	0%	0%	0%	0%	C5
	0%	0%	0%	0%	1%	0%	98%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
CNN1	97%	0%	0%	0%	0%	0%	1%	0%	0%	0%	C0
	2%	77%	4%	0%	8%	0%	4%	1%	0%	0%	C1
	0%	2%	94%	0%	1%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	97%	2%	0%	0%	0%	0%	0%	C3
	1%	0%	0%	1%	94%	0%	1%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	97%	2%	0%	0%	0%	C5
	0%	0%	0%	0%	2%	3%	92%	0%	0%	1%	C6
	0%	0%	0%	0%	0%	0%	0%	99%	0%	0%	C7
	0%	0%	0%	0%	0%	1%	0%	0%	98%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
CNN2	99%	0%	0%	0%	0%	0%	0%	0%	0%	0%	C0
	1%	89%	3%	0%	2%	0%	1%	1%	0%	0%	C1
	0%	2%	96%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	98%	0%	0%	0%	0%	0%	0%	C3
	2%	1%	1%	1%	93%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	97%	1%	0%	1%	0%	C5
	0%	0%	0%	0%	0%	3%	94%	0%	0%	1%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	1%	0%	1%	0%	1%	95%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
InceptionV3	98%	0%	0%	0%	1%	0%	0%	0%	0%	0%	C0
	0%	92%	2%	0%	3%	0%	1%	0%	0%	0%	C1
	0%	4%	94%	0%	1%	0%	0%	0%	0%	0%	C2
	0%	1%	2%	93%	2%	0%	0%	0%	0%	0%	C3
	0%	1%	0%	0%	95%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	97%	1%	0%	0%	0%	C5
	0%	0%	0%	0%	1%	1%	96%	0%	0%	0%	C6
	1%	0%	0%	0%	0%	0%	0%	98%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	99%	C9
MobileNetV2	95%	1%	0%	0%	1%	0%	1%	0%	0%	0%	C0
	0%	90%	5%	0%	3%	0%	1%	0%	0%	0%	C1
	0%	0%	99%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	99%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	99%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	95%	4%	0%	0%	0%	C5
	0%	0%	0%	0%	0%	0%	98%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9

, and the values obtained are the average values taken by calculating k = 5-fold. When the accuracy and loss graphs are examined, over- or under-learning did not occur in the two proposed CNN models, and early stopping was not completed in the applications within the scope of this study. The values in the table are rounded, with values less than one percent disregarded.

3.1. The Framework of Proposed Ensemble Models

In this study, ensemble models are presented using newly proposed CNN models and existing CNN models available in the literature. The ensemble models with the highest accuracy rate within the scope of the study are listed below, along with their results.

A voting ensemble is a deep-learning ensemble model that combines the predictions of numerous other models. We use the voting method for the proposed ensemble models.

It is a technique that can be used to enhance model performance, ideally achieving a higher level of performance than any individual model used in the ensemble. A voting ensemble functions by combining the forecasts of multiple models. It is applicable to both classification and regression.

In hard voting, also known as majority voting, each classifier votes for a class, and the class with the most votes wins. In statistical terms, the ensemble label prediction is the mode of the distribution of individual label predictions. Max-voting, which is typically employed for classification problems, is one of the most straightforward methods for combining predictions from multiple ML algorithms. Each base model makes a prediction and votes for each sample in maximum voting. The final predictive class is composed of only the sample class with the most votes.

Maximum and hard voting methods are employed by using CNN models to generate ensemble models for this study.

3.1.1. Model I, 3-EnsCNNs (MobileNetV3Small, EfficientNetV2L, CNN2)

In this paper, a new ensemble model, Model I, is presented using the MobileNetV3Small, EfficientNetV2L, and CNN2 models. The results of training the ensemble structure created within the scope of the study using the proposed dataset are given below. The classification report obtained for the proposed ensemble model is presented in

Table 8. Classification report of proposed 3-EnsCNNs (Model I).

	Precision	Recall	F1-Score	Support
C0	98	100	99	447	3-EnsCNNs (MobileNet V3Small, EfficientNetV2L, CNN2)
C1	99	95	97	183
C2	99	99	99	362
C3	99	100	99	193
C4	98	98	98	333
C5	99	99	99	374
C6	99	97	98	283
C7	100	100	100	1061
C8	100	100	100	79
C9	100	100	100	317
accuracy			99	3632
Macro avg	99	99	99	3632
Weighted avg	99	99	99	3632

. In addition, the confusion matrix and ROC diagram of the model are given in Figure 10.

3.1.2. Model II, 3-EnsCNNs (MobileNetV3Small, EfficientNetV2L, CNN1)

In this section, a new ensemble model II is presented using the MobileNetV3Small, EfficientNetV2L, and CNN1 models. The classification report obtained for the proposed ensemble model is presented in

Table 9. Classification report of proposed 3-EnsCNNs (Model II).

	Precision	Recall	F1-Score	Support
C0	99	98	99	447	3-EnsCNNs (MobileNet V3Small, EfficientNetV2L, CNN1)
C1	97	98	98	183
C2	98	100	99	362
C3	99	99	99	193
C4	99	98	98	333
C5	98	100	99	374
C6	100	97	98	283
C7	100	100	100	1061
C8	99	100	99	79
C9	100	100	100	317
accuracy			99	3632
Macro avg	99	99	99	3632
Weighted avg	99	99	99	3632

. In addition, the confusion matrix and ROC diagram of the model are given in Figure 11.

3.1.3. Model III, 3-EnsCNNs (InceptionV3, CNN2, and MobileNetV2)

In this section, a new ensemble model III is presented using the InceptionV3, CNN2, and MobileNetV2 models. The classification report obtained for the proposed ensemble model is presented in

Table 10. Classification report of proposed 3-EnsCNNs (Model III).

	Precision	Recall	F1-Score	Support
C0	99	98	99	447	3-EnsCNNs (InceptionV3, CNN2, MobileNetV2)
C1	98	92	95	183
C2	99	99	99	362
C3	98	96	97	193
C4	94	100	97	333
C5	98	99	99	374
C6	100	94	97	283
C7	100	100	100	1061
C8	93	100	96	79
C9	99	100	100	317
accuracy			98	3632
Macro avg	98	98	98	3632
Weighted avg	98	98	98	3632

. In addition, the confusion matrix and ROC diagram of the model are given in Figure 12.

In this part,

Table 11. Accuracy results of 3-EnsCNNs models.

Test Accuracy (%)
3-EnsCNNs	Unweighted Average (Acc)	Weighted Average (Acc)	Ideal Weighted (Acc)	F1-Score	Accuracy
MobileNetV3Small EfficientNetV2L CNN2	99.51	99.40	97.35	99	99.12
MobileNetV3Small EfficientNetV2L CNN1	99.44	99.29	95.92	99	99.12
InceptionV3 CNN2 MobileNetV2	99.44	99.09	97.64	98	98.43

display the values obtained for the proposed triple and quintuple ensemble models. Examining Table 11, the unweighted average (hard voting) produced better results than the weighted average (maximum voting).

3.1.4. Model IV, 5-EnsCNNs (MobileNetV3Small, EfficientNetV2L, CNN2, CNN1, InceptionV3)

In this section, a new ensemble model IV is presented using the MobileNetV3Small, EfficientNetV2L, CNN2, CNN1, and InceptionV3 models. The classification report obtained for the proposed ensemble model is presented in

Table 12. Classification report of proposed 5-EnsCNNs (Model IV).

	Precision	Recall	F1-Score	Support
C0	99	98	99	447	5-EnsCNNs (MobileNetV3Small, EfficientNetV2L, CNN2, CNN1, InceptionV3)
C1	98	93	96	183
C2	97	99	98	362
C3	99	99	99	193
C4	99	100	100	333
C5	99	99	99	374
C6	98	99	98	283
C7	100	100	100	1061
C8	100	100	100	79
C9	100	100	100	317
accuracy			99	3632
Macro avg	99	99	99	3632
Weighted avg	99	99	99	3632

. In addition, the confusion matrix and ROC diagram of the model are given in Figure 13.

In this section,

Table 13. Accuracy results of 5-EnsCNNs model.

Test Accuracy (%)
5-EnsCNNs	Unweighted Average (Acc)	Weighted Average (Acc)	Ideal Weighted (Acc)	F1-Score	Accuracy
MobileNetV3Small EfficientNetV2L InceptionV3 CNN1 CNN2	99.60	99.42	95.74	99	99.12

display the values obtained for the proposed triple and quintuple ensemble models. Examining Table 13, the unweighted average (hard voting) produced better results than the weighted average (maximum voting).

The confusion matrices of 3-EnsCNNs models and 5-EnsCNNs models are demonstrated in

Table 14. Confusion matrix of 3-EnsCNNs models and 5-EnsCNNs model.

	C0	C1	C2	C3	C4	C5	C6	C7	C8	C9
Ensemble model 1 (MobileNet V3Small, EfficientNetV2L, CNN2)	100%	0%	0%	0%	0%	0%	0%	0%	0%	0%	C0
	0%	95%	2%	0%	1%	0%	0%	0%	0%	0%	C1
	0%	0%	100%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	100%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	99%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	100%	0%	0%	0%	0%	C5
	1%	0%	0%	0%	0%	0%	98%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
Ensemble model 2 (MobileNet V3Small, EfficientNetV2L, CNN1)	98%	0%	0%	0%	0%	0%	0%	1%	0%	0%	C0
	0%	98%	2%	0%	0%	0%	0%	0%	0%	0%	C1
	0%	0%	100%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	99%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	99%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	100%	0%	0%	0%	0%	C5
	0%	0%	0%	0%	0%	1%	98%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
Ensemble model 3 (InceptionV3, CNN2, MobileNetV2)	98%	0%	0%	0%	1%	0%	0%	0%	0%	0%	C0
	0%	92%	1%	0%	4%	0%	0%	0%	1%	0%	C1
	0%	0%	99%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	97%	1%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	100%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	100%	0%	0%	0%	0%	C5
	1%	0%	0%	0%	2%	2%	95%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9
Ensemble model 4 (MobileNet V3Small, EfficientNetV2L, CNN2, CNN1, InceptionV3)	99%	0%	0%	0%	0%	0%	0%	0%	0%	0%	C0
	0%	94%	3%	0%	0%	0%	1%	0%	0%	0%	C1
	0%	0%	100%	0%	0%	0%	0%	0%	0%	0%	C2
	0%	0%	0%	99%	0%	0%	0%	0%	0%	0%	C3
	0%	0%	0%	0%	100%	0%	0%	0%	0%	0%	C4
	0%	0%	0%	0%	0%	100%	0%	0%	0%	0%	C5
	0%	0%	0%	0%	0%	0%	99%	0%	0%	0%	C6
	0%	0%	0%	0%	0%	0%	0%	100%	0%	0%	C7
	0%	0%	0%	0%	0%	0%	0%	0%	100%	0%	C8
	0%	0%	0%	0%	0%	0%	0%	0%	0%	100%	C9

. With its quick training duration and superior classification performance, the proposed study stands out.

Table 15. Comparison of the studies.

Study	Cross-Validation	Method	Number of Diseases	Accuracy (%)
[39]	-	CNN (Full-Color, Grayscale)	6	99.84, 95.56
[16]	-	DenseNets-121	14	97.75
[40]	5 k	AlexNet, GoogleNet	9	97.35, 97.71
[41]	-	AlexNet	10	95.65
[10]	-	GoogleNet	58	97.27
[42]	-	AlexNet	9	95.83
[43]	-	MobileNet	10	88.4
[44]	-	VGG16	10	93.5
[45]	10 k	CNN	10	98.2
[46]	-	ResNet50 + SeNet	10	96.81
[47]	-	Wavelet-like Auto-Encoder (WAE)	10	98.1
[48]	-	FWDGAN + B-ARNet	4	98.75
[49]	-	VGG16	10	95.71
[50]	-	DenseNet+ Attention mechanism	5	97.56
[51]	-	Lightweight Attention-Based CNN	11	99.34
[52]	-	Spatial attention with CNN	10	95.20
[53]	-	MobileNetV3	10	99.81
[54]	-	Compact CNN	10	98.49
[55]	10 k	Gray Wolf + MobileNetV2	8	98
[17]	-	ResNet + Discriminative Learning	10	99.51
	5 k	MobileNetV3Small, EfficientNetV2L, InceptionV3, CNN1, CNN2	10	99.60

lists various methods for the classification of tomato leaf diseases and research using datasets from tomatoes. The accuracy rate part of this table displays the individual successes of tomato leaf diseases in the datasets in the literature. The findings of the model with the highest rate of accuracy in investigations using various methodologies, aside from the suggested study, are shown in the table.

The accuracy rates obtained using k-fold cross-validation are nearly the same as those discovered in previous studies. This study, which may serve as a model for further research, could be expanded to use a dataset with more data to achieve a greater accuracy rate. Tomato producers can utilize the suggested approach in this way to identify diseases more quickly. A cost-effective calculation can be used to identify diseases in tomato leaves using the proposed models, which have an accuracy of nearly 99.60%. Additionally, it is anticipated that it will aid in the development of future cooperative decision-making systems. Additionally, the CNN classification performances put forth in the presented study demonstrated that analyses with high accuracy are conceivable and that developing a CNN-based mobile application can facilitate analyses in terms of practicality and adaptability.

For farmers and the agricultural sector, plant disease and insect infestation are risky. By destroying plants and production quality, it has a significant negative economic impact. An autonomous system for spotting plant diseases is crucial to the development and improvement of agricultural output. The image-based automatic disease and pest detection system, which uses a variety of image processing and neural network approaches, provides the foundation for this research. Without compromising the performance evaluation criteria, the presented approach can be expanded to compress other well-known pre-trained models, including MobileNet, EfficentNet, and Inception. On a variety of publicly accessible plant villages, training, validation, and testing are carried out. When compared to conventional methods, the research reported in this study produced more favorable results with improved performance. Future research on the automatic plant disease and pest detection approach is primarily focused on enhancing system efficiency by speeding up computation. In accordance with the degree of the impact, it also focuses on additional treatment and prevention strategies. Since this system can be utilized on embedded systems with autonomous drones, the entire system should be implemented in real-time using mobile and web applications.

4. Conclusions

This study aims to investigate and apply several deep-learning models to identify diseases in tomato plants. Nine distinct diseases of tomato plant leaves were used in this context, and the widely available PlantVillage dataset was chosen. Using this data set, modeling studies are conducted by recommending two novel CNN architectures for tomato plant disease prediction and classification, in addition to MobileNet, Inception, and EfficientNet, which are specialized deep learning architectures. CNN models found in the literature have been seen to be useful for classification tasks. However, because of its intricate and substantial architecture, the storage space and inference time are extremely high. These models cannot be applied to edge computing devices due to storage capacity and inference time limitations. The proposed CNN models are used to overcome this issue, and the results are detailed in Section 3. Each architecture included in the study is initially categorized using the 5-fold cross-validation approach. The PSO method is used to perform hyperparameter adjustments in the CNN models proposed here, and the VGG16 model is used for the fine-tuning method. The proposed model is one of the best models out of all the applied architectures in the initial modeling tests utilizing the data set generated for the classification of plant species. As a consequence of comprehensive experimental studies, four novel ensemble models are suggested for the optimal topologies for each categorization task. According to the ensemble model results, each model successfully passes the test with an accuracy of 99.60%. Within the parameters of the investigation, the obtained ROCs, a confusion matrix, and classification reports are presented. The findings demonstrate that a significant amount of data for deep learning boosts both the model’s learning success and the test success, and it may be employed with excellent results particularly in classification tasks using image data. This disease prediction model’s future applicability can be expanded through the incorporation of mobile agent-based feature selection and classification algorithms, which are capable of selecting the most important features and predicting plant diseases of various types. In addition, ML classifiers can be replaced by the final layers of the proposed CNN model. Additionally, images can be enhanced during pre-processing to improve the results.