A Hybrid Method for Automatic Plant Leaf Disease Identification Using Whale Optimization Algorithm and Convolutional Neural Networks
الموضوعات :
Zahra Akeshteh
1
,
Parvaneh Asghari
2
,
Seyyed Hamid Haji Seyyed Javadi
3
,
Hamidreza Navidi Ghaziani
4
1 - گروه مهندسی کامپیوتر، واحد بروجرد، دانشگاه آزاد اسلامی، بروجرد، ایران
2 - گروه مهندسی کامپیوتر، واحد تهران مرکزی، دانشگاه آزاد اسلامی، تهران، ایران
3 - گروه مهندسی کامپیوتر، دانشگاه شاهد، تهران، ایران
4 - گروه ریاضی و علوم کامپیوتر، دانشگاه شاهد، تهران، ایران
الکلمات المفتاحية: leaf disease, classification, convolutional neural network, whale optimization algorithm.,
ملخص المقالة :
This study introduces a combined approach using deep learning and optimization to accurately and efficiently classify plant leaves based on disease and health. By optimizing hyper parameters with a whale optimization algorithm and utilizing a convolutional neural network for disease classification, the model achieves high accuracy. The Plant Village dataset is used, and data augmentation is applied to improve the model's performance. The optimized network achieves a classification accuracy of 95.22% for the test set and 99.57% for the training set, with precision and recall values of 95.24% and 95.22% respectively. The performance and efficiency of the proposed method are proven to be superior when compared to other models and pre-trained networks. This study has potential applications in various image classification tasks and can be valuable in agriculture, horticulture, and plant disease identification. Furthermore, the proposed network achieves higher accuracy with fewer trainable parameters and computations compared to similar works.
A Hybrid Method for Automatic Plant Leaf Disease Identification Using Whale Optimization Algorithm and Convolutional Neural Networks
Abstract
This study introduces a combined approach using deep learning and optimization to accurately and efficiently classify plant leaves based on disease and health. By optimizing hyper parameters with a whale optimization algorithm and utilizing a convolutional neural network for disease classification, the model achieves high accuracy. The Plant Village dataset is used, and data augmentation is applied to improve the model's performance. The optimized network achieves a classification accuracy of 95.22% for the test set and 99.57% for the training set, with precision and recall values of 95.24% and 95.22% respectively. The performance and efficiency of the proposed method are proven to be superior when compared to other models and pre-trained networks. This study has potential applications in various image classification tasks and can be valuable in agriculture, horticulture, and plant disease identification. Furthermore, the proposed network achieves higher accuracy with fewer trainable parameters and computations compared to similar works.
Keywords: leaf disease, classification, convolutional neural network, whale optimization algorithm.
1- Introduction
Agriculture is one of the most important sectors that drive development. It serves as a unique source of wealth generation, driven by farmers. One of the needs and necessities of strong countries is agricultural development in the global market. The population is growing in the world, making increased food production essential. In the future years, changes in various types of products and diseases will occur at every level, and analyzing this information will play a crucial role in the early stages and contribute dynamically to the agricultural sector. The occurrence of various diseases in crops poses a significant challenge for farmers. Disease analysis and classification are fundamental concerns in agriculture. The technology for identifying diseases in agricultural products and classifying them is important and will be of great interest in the future for ensuring food security and increasing production efficiency. The approach of identifying diseases in agricultural products when symptoms appear on plant leaves is considered a rapid, efficient, cost-effective, and highly effective method [1, 2]. The manual observation of plant diseases is indeed time-consuming and requires more effort. Moreover, this method may only be applicable in limited areas and becomes limited in cases where a large number of plants need to be examined. On the other hand, using automated detection techniques allows for obtaining results with minimal effort and more precise timing. By utilizing algorithms and deep neural networks, plant leaf images can be analyzed, and diseases can be automatically identified. These methods can be employed for rapid and accurate detection of plant diseases in extensive areas, assisting farmers in promptly addressing disease-related issues in their crops. In general, the use of automated detection techniques for identifying plant diseases not only reduces time and effort but also provides more accurate results [3].
Currently, due to the technological progress for theoretical and practical solutions, many image analysis methods that are fully or partially automated have been developed. A new era of smart agriculture has begun, using emerging technologies such as the Internet of Things, big data, and machine learning. As a result of artificial neural network innovations, great achievements have been observed in the field of agriculture. The use of convolutional neural networks (CNNs) has led to human-level accuracy in diagnosing plant diseases, identifying weeds, and grading fruits. With the gradual increase in the number of architectural layers in CNNs, the ability to solve multiple problems multiclass has also been created. [4]. Diseases such as leaf burn, rust, rot, powdery mildew, etc., reduce the productivity and number of products. Therefore, one of the important needs of the agricultural field is to identify the disease in its early stages. In addition, it is necessary to determine the amount of damage caused by these diseases to the product. Diseases cannot be identified by humans until symptoms appear. There is also a need for sampling, testing, and estimating the amount of damage. These things are expensive and the effectiveness of these traditional methods is low. To avoid this problem, automatic methods are needed [5].
Identification of plant disease is necessary for optimal agriculture and plant phenotype. These fields require data and information. Traditional methods, to identify plant diseases, require visual inspection and therefore, are expensive, time-consuming, and rely on humans [6-8]. These methods are not efficient and in addition, due to their dependence on humans, there is a possibility of errors due to fatigue or bias. To solve this problem, image processing methods have been proposed. In 1983, one of the first studies in this field was conducted[9, 10]. In this study, using black and white images and films of tomato leaves, blackened fern disease was identified. Additionally, in another research endeavor, an image-processing method was employed to quantify the severity of streak disease in corn. The findings indicated that image processing outperformed conventional techniques in terms of accuracy [11].
The main idea of artificial intelligence is that machines behave like humans and it enables the development of intelligent systems. One of the most popular artificial intelligence techniques is artificial neural networks. This type of neural network is based on the inspiration of the human biological brain. This network can be used to model big data and find complex relationships [12-14]. Deep learning is a subset of machine learning. This method has become a high-priority approach in disease detection due to advantages such as high computing power, storage capacity, and the availability of large data sets. After the ImageNet competition in 2012[15], many researchers in different fields used deep learning in their research. One of the common techniques in deep learning is the use of CNNs to recognize objects, and segment and classify images. However, deep learning seriously requires big data that includes hundreds and thousands of images[16]. Since 2015 and after the creation of the Plant Village dataset, deep learning using CNN has become one of the prominent research topics [17]. This database is used to identify diseases[18], estimate disease severity[19], and develop management systems[20]. In recent years, several other disease datasets have been made available to the public, such as the Digipathos dataset [21], the PlantDoc dataset [22], the Northern Leaf Blight dataset [23], the rice disease dataset [24] and cassava disease dataset[25]. Many studies have used this data to train deep-learning models [26]. Using deep learning to diagnose plant diseases has advantages such as the isolation of disease symptoms, more accurate identification, and the possibility of estimating the severity of the disease, suggesting and providing cost-effective solutions [8, 26, 27]. Using convolution layers in deep learning is a promising approach. Using this method, the important characteristics of the plant such as colors, the texture of lesions, and waste are automatically identified. In addition, similar performance could be achieved while removing 75% of the parameters [28]. This omission shows the applicability and usefulness of deep learning. In addition, CNN can identify the features in the images and perform classification at the same time. However, this approach is limited and time-consuming. Because it requires big data. More researchers are using this technique. They have provided their data for further research for public use and research [27-30].
A CNN is known as a deep learning architecture, which is inspired by biological vision mechanisms [31]. This network uses integration and convolution to identify information and is used in numerous cases [32]. Its applications such as identifying objects and classifying images [33]. The LetNet-5[34] network is a classic CNN that is often used to test algorithms [35]. Before training a neural network, some parameters should be specified including settings like batch size, learning rate, and dropout rate, named hyperparameters. These values are not learned during the training process but are predetermined by the user or researcher. They play a crucial role in determining the behavior and performance of the neural network during training. The selection of appropriate hyperparameter configurations has a significant impact on the performance of deep learning models. However, determining the optimal structure and hyperparameter values for a deep learning model can be time-consuming and a complex process. In a model, hyperparameters have a crucial role, as they influence the performance of the model [35].
There are various methods such as manual search [36], random networks, and Bayesian to optimize hyperparameters. The manual search method can determine hyperparameters for simple networks. Before the age of big data, the complexity of neural networks was low and its simple models were processed. For this reason, it was possible to determine hyperparameters manually by experts. With the emergence of big data, it is not possible to manually determine hyperparameters, and their automatic determination has become a new research field [37]. Lorachelle used a network search algorithm[38]. The network search algorithm is a calculation method that creates a balance between system performance and calculation efficiency by determining hyperparameters. When dealing with large datasets, there is a challenge known as the "exponential explosion" problem, which refers to the exponential increase in search space and the subsequent reduction in search efficiency. As the dataset size grows, the number of possible combinations and patterns within the data increases exponentially. This makes it more difficult and time-consuming to search for optimal solutions or patterns within the dataset. Consequently, it becomes increasingly challenging to efficiently analyze and extract meaningful insights from large datasets due to the exponential increase in search complexity [35, 39]. Bergstra [40], proposed an easy method for a random search. The blind search method, which involves trying out different combinations of hyperparameters without considering specific characteristics of the dataset, can indeed be inefficient and lack adaptability. This approach may not take advantage of the specific properties or patterns present in the dataset, potentially leading to suboptimal performance. To improve efficiency and adaptability, it is recommended to incorporate techniques such as automated hyperparameter optimization algorithms or utilize domain knowledge to guide the search process. These methods can help narrow down the search space and select hyperparameter configurations that are more likely to yield high-performance models, thus improving efficiency and adaptability in the hyperparameter tuning process. In sequential models [41, 42], Bayesian optimization [43-46] is one of the classic methods for setting hyperparameters. This method can use all the information from the previous searches [47], but it only samples around the optimal point. Therefore, it can be used in complex problems and performs better than manual and random searches. Heuristic algorithms work based on experience [47]. These algorithms can provide a solution with an acceptable cost by considering space and location calculations. These methods have been widely used to adjust hyperparameters. Sun[48] has used the Simulated Annealing high-speed algorithm for neural network clustering. Zhang [49] used Genetic Algorithm to optimize the deep learning network, with high convergence. Heuristic algorithms commonly face a shared challenge. This problem is their inapplicability when they fall into a locally optimal point [35]. Training a model and optimizing it is a time-consuming operation, difficult and professional. For this purpose, a powerful processing unit and thousands of training examples are needed. Therefore, transfer learning reduces many of these problems [50]. Pre-trained CNNs, for this purpose, can transfer knowledge for optimal transmission. For the transfer of learning, the initial preparation of the image is necessary. To address this issue, one approach is to utilize a multi-crop image set with a size of 224 ˟ 224. This involves taking multiple crops of the original image and resizing them to the desired size. By considering multiple crops, the network can capture different perspectives and variations in the data, potentially improving accuracy. However, depending on the specific requirements of the task or dataset, this layer can be modified or replaced with a different pooling or aggregation technique to better suit the problem at hand. By implementing these adjustments, the network can potentially achieve better performance and adaptability for the given task [2].
This research focuses on using a CNN to automatically learn features instead of relying on human detection for classifying leaf diseases. This approach not only saves time and money but also achieves promising accuracy and performance in disease prediction. To compare different CNN architectures and datasets, the study aims to analyze the performance differences, abilities, and errors of different methods. The authors introduce a novel meta-heuristic optimization algorithm named Whale Optimization Algorithm (WOA) that imitates the hunting behavior of humpback whales. The authors did not find any previous studies that simultaneously addressed detection in different plants using CNN and WOA; existing cases were limited to specific plants. The unique aspect of this work compared to other papers by the authors is the use of the hunting behavior of whales. The optimization results reveal that WOA is competitive compared to other optimization methods. The remaining sections of the paper are structured as follows: Section 2 provides a literature overview, Section 3 describes the CNN method and the Plant Village database, Section 4 presents experimental results and analysis, and Section 5 discusses the research findings and concludes the paper.
2- Literature review
In [51], Rangarajan Aravind used differently to classify 10 types of plant diseases. The results showed that the performance of GooleNet is better compared to other methods. In study[52], a pre-trained VGG-16 network was used to extract features, and a multi-class supported vector machine was used to classify eggplant diseases. The results of RGB photos showed the highest classification accuracy. Atila et al. [53] investigated an automatic model to identify and classify diseased parts in plants. Advanced networks based on convolution neural networks were investigated and their performance accuracy was compared with different models. The obtained model had an accuracy higher than 90% and it was more accurate than other architectures. Ji et al. [54] used the unified convolutional network to identify grape plant diseases. The demonstration capability of the unified model and the use of special techniques is an advantage for these models and makes them perform better than the competitors. In one of the studies [55], both complete images and segmented images have been used to identify the disease. The results showed that when the trained CNN is applied to segmented images; it has higher performance accuracy compared to when it is applied to a whole image. In a study, Azimi et al. [56] investigated a CNN model that has 23 layers and compared it with other models. The feature of CNN can be easily classified with the proposed network. The disease causes a decrease in production and economic damage.
Gadekallu et al. [57], extracted image features using the combined Principal Component Analysis technique and whale optimization. The results of his model were compared with other models in terms of accuracy and preference. Sinha et al. [58], investigated two diseases, Rust and Cercospora. These diseases affect the quality and productivity of the coffee plant. Feature extraction is done using K-Means and threshold segmentation algorithms. In the next step, by analyzing the plant tissue, the relationship between the infected and healthy parts of the plant is determined. In [59], wormholes, insects, and plant-related pests were classified. Different techniques were used in this study. Models such as AlexNet, GoogleNet, and ResNet50 have been used for this task, and the results showed that ResNet50 has a better performance compared to other techniques. Franczyk et al.,[60], used a CNN model and showed that its performance is better than machine learning techniques. Identifying plant diseases with traditional methods requires high cost, human supervision, and maintenance. Intelligent methods of data collection and classification that reduce costs and increase accuracy are well explained by Kundu et al. [61]. Some plant diseases have symptoms on the leaves. Diseases such as bacteria, fungi, and nematodes are common diseases that appear on the leaves of plants. To identify these diseases in the early stages, it is necessary to perform the analysis by an expert. In traditional methods, the common features in diseases are extracted from color images and suggested to the classifier. In the study of Oyewola [25], deep learning and pre-processing are used to balance the data that are not balanced. For color stretching, correlation, and gamma corrections are used. The results showed that unbalanced data has more accuracy. Abayomi-Alli [62] used image histogram transformation techniques for detection. Basavaiah [63] discussed the identification of diseases that cause yield reduction. Using Plant Village data, various techniques have been used to classify plant diseases. Abdu et al. [64], used machine learning to compare with support vector machine. Both models were deployed under normal environmental conditions and the same processing capacity and training data.
The study [65] introduced a novel model named "PlantXViT" that integrates CNNs with Vision Transformers (ViTs) to achieve accurate identification of diverse plant diseases across multiple crops. The proposed PlantXViT model exhibits a lightweight architecture, boasting a mere 0.8 million trainable parameters. This characteristic renders it highly suitable for deployment in the Internet of Things (IoT)-based smart agriculture systems. To evaluate its performance, PlantXViT is subjected to rigorous testing on five publicly available datasets. The experimental results showcase the exceptional capabilities of the model, with average accuracy levels exceeding 93.55% for Apple, 92.59% for Maize, and an impressive 98.33% for Rice in the realm of plant disease recognition.
The primary aim of the Alshammari study [66] was to detect and classify diseases that have the potential to affect olive leaves. The outcomes of the research were highly encouraging, showcasing the effective fusion of CNN and vision transformer models. Through a comprehensive analysis of the experimental results detailed in their investigation, the authors highlighted the exceptional performance of their proposed model when compared to alternative approaches. The accuracy achieved by the model in disease classification was approximately 95%, further reinforcing its superiority in the domain.
The authors of the [67] presented an advanced approach for feature computation in plant disease classification. Their method combines Squeeze and Excitation Networks (SE networks) with Capsule Networks (CapsNet). Specifically, the authors employed two SE networks, one based on AlexNet and another based on ResNet, to evaluate the severity of Late Blight disease in Tomato crops. The leaf images used for analysis were obtained from the PlantVillage dataset and categorized into four severity stages: healthy, early, middle, and end. Before inputting the images into the SE networks, they were downscaled and enhanced. The resulting feature maps from the two networks were then separately fed into the Capsule Network for classification purposes. To compare the performance of these models, the original CapsNet was also evaluated using two different image sizes: 32˟32 and 64˟64. The experimental results revealed that the SE-Alex-CapsNet achieved the highest accuracy of 92.76%, while the SE-Res-CapsNet achieved the highest accuracy of 94.4% when using the 64˟64 image size. These accuracies surpassed the original CapsNet's accuracy of 85.53%. The findings highlight the potential of the proposed techniques for assessing disease severity in various crops. Furthermore, these techniques can be extended to other applications such as plant species classification and weed identification.
In the research conducted by Thakur et al.[68], a lightweight CNN named "VGG-ICNN" was introduced to identify plant diseases using leaf images. The VGG-ICNN architecture comprises approximately 6 million parameters, which is notably lower than the number of parameters of most high-performing deep learning models. To evaluate the effectiveness of the proposed model, it is tested on publicly available datasets encompassing a diverse range of agricultural products. Specifically, the datasets consist of apple, maize, and rice, which are categorized into four, four, and five disease categories respectively. The experimental results demonstrate the superior performance of the VGG-ICNN approach in accurately identifying diseases in apple, maize, and rice, surpassing the performance of recent deep learning methods. The achieved accuracies for apple, maize, and rice are 94.24%, 91.36%, and 96.67% respectively. Furthermore, the VGG-ICNN model consistently exhibits excellent performance across all the datasets utilized in this study, outperforming several lightweight CNN models that have been recently proposed.
In their research, Joshi and Bhavsar [69] introduced a novel model that enhances the accuracy and efficiency of plant leaf disease identification when compared to existing approaches. The study involved training standard models such as AlexNet, VGG, and GoogleNet, alongside the proposed model, using the Night shed plant leaf dataset available in the plant village. This dataset comprises nine distinct categorical classes, representing various diseases as well as healthy plant leaves. Several parameters, including batch size, dropout, learning rate, and activation function, were carefully examined to assess the performance of the models. The proposed model exhibited remarkable disease classification accuracy rates ranging from 93% to 95%. These accuracy results indicate the promising potential of the proposed model to significantly improve the speed and accuracy of identifying disease-infected leaves.
In the research conducted by Singh et al. [70], the focus was on utilizing transfer learning with three pre-trained deep learning models, namely MobileNetV2, EfficientNetB6, and NasNet, on a dataset consisting of bean leaf images. The dataset itself contains 1295 images that are categorized into three distinct classes. To evaluate the performance of various CNN models, several optimization techniques are employed. The experimental findings indicate that EfficientNetB6 surpasses the other models, achieving an impressive accuracy rate of 91.74%. This outcome underscores the importance of carefully selecting an appropriate model for disease classification tasks. Furthermore, the study explores the impact of different optimizers on CNN models, providing valuable insights for future research endeavors in this domain.
In the research conducted by Ulutaş and Aslantaş [71], a novel approach is proposed to enhance the performance of CNN models. The proposed models undergo fine-tuning and subsequent hyperparameter optimization using the particle swarm optimization algorithm (PSO). Additionally, the weights of these architectures are optimized using the grid search method, resulting in the development of both triple and quintuple ensemble models. The datasets utilized in the study are classified using a five-fold cross-validation technique. The experimental findings showcase the exceptional performance of the proposed ensemble models, characterized by fast training and testing times as well as a remarkable classification accuracy of 96.87%. This research significantly contributes to the early detection of plant diseases simply and efficiently, providing valuable support to experts in effectively preventing the spread of infections.
This research focuses on a new area of study that gained attention from researchers in recent years due to its use of advanced technology. Tugrul et al.[72], conducted a review of 100 studies that explored the use of CNNs for detecting plant diseases within the last five years. The study examined various aspects such as approaches, limitations, model performance, and comparisons of the studies. The research revealed a significant increase in studies related to this area in 2022. The authors conducted a keyword search in various databases and found 23 studies related to leaf disease and using CNN in 2023. The number of studies conducted in different years is presented in Figure 1.
Figure 1. Distribution of the year in 100 reviewed studies [72].
Studies in which whale optimization algorithm has been used to diagnose plant diseases are limited. In them, only one specific plant has been investigated. In 2022, Sai Satyanarayana Reddy et al.[73], conducted a study to diagnose rice diseases. They improved their network using a whale optimization algorithm. In 2023, Kiran Pandir et al. [74], diagnosed potato leaf diseases. They provided a network named Pot-Net for this purpose. In 2022 Suresh et al.[75], investigated Groundnut diseases. The researchers have developed a model that utilizes the Whale Optimization Algorithm (WOA) in conjunction with the Long Short-Term Memory (LSTM) method to identify and classify different types of plant diseases. By integrating the WOA, which is an optimization algorithm, with LSTM, a type of recurrent neural network, the model aims to enhance the accuracy and effectiveness of disease recognition in plants. This combined approach enables the model to capture long-term dependencies in sequential data and optimize the parameters to improve disease classification performance across various plant diseases.
3- Research methodology
Human supervision and expertise in feature extraction are highly effective. Machine learning methods often struggle to find useful information among high-dimensional and irrelevant data. However, deep learning models like CNNs have a better capacity to make broad conclusions and withstand disturbances caused by noise. They can perform feature extraction without the need for human supervision, making them more reliable when used as unbiased classifiers. CNNs can extract features throughout convolution and pooling operations. This research suggests an automated method for classifying leaf diseases using a CNN. Additionally, an optimization strategy for hyperparameters of the CNN is employed, which accelerates convergence to the optimal solution. Furthermore, it improves the accuracy and confidence in the used hyperparameters within the CNN framework. This framework has four main stages: 1-data preprocessing and augmentation, 2- designing the architecture of the CNN, 3- hyperparameter optimization, and 4- classification. The optimized CNN's performance is calculated by various evaluation metrics and compared with VGG-16 and VGG-19, to assess the model's performance. The image of plant leaf disease classification using the CNN along with hyperparameter optimization is illustrated in Figure 2.
Figure 2. Schematic of leaf disease image classification.
3.1. Plant Village dataset
The first step to good agricultural yield is protecting the plants from diseases. Early plant disease recognition and prevention is the first step in this regard. However manual disease recognition is time-consuming and costly. This is one of those use cases, where deep learning can be used proactively for great benefit. Using deep learning, plant diseases can be recognized very effectively. Large-scale plant disease recognition using deep learning can cut costs to a good extent. It consists of approximately 87,000 images of healthy and diseased leaves. These images are classified into 38 different classes based on the type of disease or health condition. The entire dataset is then divided into a training set and a validation set, maintaining an 80/20 ratio. During this division process, the directory structure of the dataset is preserved, ensuring that the organization of the original dataset is maintained in the recreated version [76].
The data available in Plant Village consists of two parts: Train and Validation. Each of these sections has 38 subcategories for leaf diseases (healthy leaves for each plant are also considered as a category). For the final evaluation of the obtained models, the validation data were divided into two categories. The first category, named the Test, contains 10% of the total data (about half of the validation data). These data are not used in any of the stages of training and optimization of the network and are used only after the finalization of the network to evaluate it and compare it with other networks under equal conditions. The number of images in the second category, which is called the Validation, is the same as the previous category. These data were used for validation. These data were used to optimize hyperparameters. In addition, in the evaluation of the networks, the performance of each of them will be used for all three categories of data, Train, Test, and Validation, and the results will be evaluated for all of them. In the optimization of the hyperparameters stage, the goal is to obtain the maximum value for accuracy. The Plant Village database, like many other datasets in the field of plant pathology, often requires augmentation before use. This is because the availability of training data for plant diseases is usually limited or insufficient. Augmentation techniques such as rotation, flipping, zooming, and adding noise are applied to increase the size of the dataset. By augmenting the dataset, researchers could improve the performance of machine learning models. CNN can detect relevant minor features in an image while being robust to large uncorrelated changes in the image [77]. For image datasets, this enhancement can be performed by changing a few pixels in the images to improve generalizability and avoid overfitting.
Some of the transformations used to enhance data include zooming, scaling, and rotating the image. Augmentation is beneficial as it improves the accuracy of classification regardless of the image's size, position, or level of distortion. In the next step, preprocessing and data augmentation are performed on the input images. For this purpose, rescale, shear, flip and zoom are used. The values of these methods are shown in Table 1.
Table 1. Preprocessing and augmentation data.
No. | Method | Value |
1 | Rescale | 1/255 |
2 | Shear range | 0.2 |
3 | Zoom range | 0.2 |
4 | Horizontal flip | True |
When utilizing the augmentation approach, it is crucial to exercise caution to avoid altering the correct class by employing inappropriate transformations. It is essential to select transformations that preserve the integrity and characteristics of the original class, ensuring that the augmented data remains representative of the intended class. This helps to maintain the accuracy and reliability of the classification process and prevents any unintended misclassification due to incorrect transformations [78].
Resized images consist of three dimensions: color depth, width, height, and. The color depth is shown by three: red, green, and blue channels. However, when converting RGB images to grayscale, the color channels are combined into a single monochrome channel with varying shades of gray. To handle the computational workload of networks, which can be challenging or slow for a CPU, the model is implemented using graphics processing units (GPUs). Originally developed for graphics applications in video games, GPU hardware has been custom-made for use in network computing. In terms of hardware specifications, the system has 12.681GB RAM and is equipped with an NVIDIA GeForce RTX 3090 GPU. This powerful GPU enables efficient and high-performance processing for deep learning tasks.
3.2. Convolutional Neural Networks
CNNs have sparse connections within their structure, making them more efficient. CNNs are specifically designed for datasets with structured topologies, such as two-dimensional images. They are widely used for segmentation, pattern recognition, and classification. CNNs excel at detecting spatial and temporal relationships in images, focusing on the most relevant information while disregarding less critical details. CNNs learn by weights and biases adjusting. One key advantage of CNNs is their ability to recognize patterns in various locations of an image. Once a pattern is learned in certain areas, CNN can identify it elsewhere without the need for relearning.[79]
The presence of convolutional layers is one of the key distinguishing features of CNNs compared to traditional machine-learning methods. While dense layers learn global patterns, convolutional layers in CNNs focus on learning local patterns within an image. These local patterns can include edges, textures, and other visual features. CNNs can deeply learn complex visual concepts. [79]
3.2.1. Convolution operation
Convolution can be defined as a mathematical operation that combines two functions by measuring their overlap as one function is shifted over the other. In the context of CNNs, convolution is used to extract features from input data. When applying convolution in CNNs, one function (often referred to as a filter or kernel) is flipped concerning the independent variable before the convolution operation takes place. Convolutional layers in CNNs utilize this operation to merge local information, and reduce the feature map size. This reduction in dimensionality makes CNNs well-suited for classification tasks.
The output of a convolutional layer is a vector that includes predictions for every individual class. Images contain various features and information. Complex geometric filters can be used to detect specific features like edges, circles, squares, etc. These filters require more detailed networks in subsequent layers to detect specific objects based on the extracted features. In a CNN, the input is represented as a matrix, with individual pixels from the image serving as values in the matrix. The stride determines how many pixels the filter should shift in each movement. An activation function is applied to each block of convolved outputs, resulting in computed activation values. Subsequently, the output predictions are generated by pooling layers, which subsample from these activated outputs, automatically reducing the dimensions of the input data.
3.2.2. Max pooling operation
Pooling layers play a crucial role in reducing the dimensions of feature maps in convolutional neural networks (CNNs). This reduction in dimensions helps to decrease the number of parameters and computational requirements in the network. After the convolutional layer generates the feature map, the pooling layer summarizes the features. By summarizing features, the pooling layer can capture the most salient information while discarding unnecessary details.
3.3. Proposed CNN architecture
The The framework used in this study is Keras, which is a popular deep-learning library. In this study, a convolutional neural network (CNN) is designed from scratch, taking inspiration from the VGG-19 network. VGG-19 is a well-known CNN architecture that has achieved excellent performance in image classification tasks. It consists of 19 layers, including convolutional layers, pooling layers, and fully connected layers. The network has a deep structure that allows it to learn complex features and patterns from images. Simonian and Zisserman [80], introduced and developed Visual Geometry Group Network (VGG) based on CNN architecture at Oxford Robotics Institute[81]. In a global image recognition competition, VGG19 won first place in 2014[80, 82]. The dimensions of the input images in the designed network are set to 64˟64˟3. Choosing the appropriate dimensions for input images depends on the trade-off between feature recognition and execution speed. Higher image dimensions allow for more recognizable features, while reducing the image size may result in the loss of certain features. However, it's important to note that higher dimensions also increase the number of calculations and therefore decrease the execution speed of the network. Thus, the selection of input image dimensions should be based on the specific conditions and experience.
The proposed network consists of 4 blocks, each with different layers. The specifications of these blocks can be found in Table 2. The network concludes with an output layer, which includes a dense layer and utilizes the SoftMax function for feature classification and disease prediction. The Adam optimizer is used with a learning rate of 0.001, and the loss function employed is categorical cross-entropy. Overall, the network architecture and configuration, including the choice of input image dimensions, optimizer, learning rate, and loss function, are designed to optimize the performance and accuracy of the network in classifying and predicting the disease type.
Table 2. Parameter values of blocks in the designed CNN.
Parameter | Block 1 | Block 2 | Block 3 | Block 4 | ||
Number of filters | 128 | 256 | 256 | 512 | ||
Size of the kernels | 2 | 2 | 2 | 2 | ||
Activation function | Relu | Relu | Relu | Relu | ||
Padding | Same padding | Same padding | Same padding | Same padding | ||
Stride | 2 | 1 | 1 | 1 | ||
Size of Maxpooling | 2 | 2 | 2 | - | ||
Dropout rate | 0.1 | 0.1 | 0.1 | 0.5 | ||
Batch normalization | No | No | No | Yes | ||
Flatten | No | No | No | Yes | ||
Dense layer size | - | - | - | 512 |
Hyperparameter | Range/ Type |
Number of filters |
|
Size of the kernels | 2,3,4,5,7,9 |
Activation function | tanh, sigmoid, Relu, elu |
Pooling size | 2,3,4 |
Dropout rate | n/100 , n= in range 0 and 100, with stride=3 |
Dense layer size |
|
3.4.1. Whale Optimization Algorithm (WOA)
The WOA algorithm is a technique that mimics the behavior of whales [83]. WOA idea is taken from the gigantesque humpback whale hunting method [103] and is known as the bubble-net feeding method[84]. Groups of whales emit high-pitched calls and dive beneath the school of fish. The sounds produced by the underwater fish prompt them to swim towards the surface. The whales release distinct bubbles in a circular pattern resembling the number 9. These bubbles form a swirling spiral that constricts around the fish, preventing them from escaping. The experimental setup for this process is depicted in Figure 3[85].
Figure 3. Bubble-net feeding behavior of humpback whales [83].
The whales conclude the hunting process by ascending to the surface, with their mouths open, using a helix-shaped movement triggered by the hunting call emitted by the leader whale[86]. The WOA was formulated using the following steps:
(1) Encircling prey
The first step is to encircle the prey in the best possible position. Whales search for the optimal position of the prey and adjust their positions accordingly. This adjustment is represented by Equations (1) and (2).
| (1) |
| (2) |
Where 
and 
are two position vectors. indicates the optimal solution, and t is the current iteration. Furthermore, 
specify the coefficient vectors that are given by Equations (3) and (4):
| (3) |
| (4) |
Here, "a" represents a variable that decreases linearly, while "r" is a random vector ranging from 0 to 1.
(2) Bubble-net attacking.
The exploitation phase includes two processes, namely:
(a) Shrinking encircling prey
The values of and 
decrease in Equations (7) and (8). The vector has a random value between [-a, a] and declines from 2 to 0. The new position can be reached using the original and current optimal agent positions.
(b) Spiral position updating
Let 
and 
indicate the positions of the whale and prey and 
the distance between them; the whales mimic the helix-shaped movement according to the spiral equation in Equation (5):
| (5) |
Where is constant, and l is a random variable between [-1, 1], and b is constant. The updating formula of the whales’ position is as Equation (6):
| (6) |
Where p, is a random variable ranging from 0 to 1.
(3) Search for prey (exploration phase)
This phase can perform a global search when the whales search for a randomly chosen agent. This mechanism is used when has a random value larger than 1 or less than -1. The search for prey can be modeled by the Equations (7) and (8):
| (7) |
| (8) |
Where the position vector is chosen randomly from a whale between the current populations.
To configure the WOA, the population number is 10, the maximum repetition is 50, and the dimensions of the problem are equal to the number of selected hyperparameters (18). In addition, the lower band is equal to zero and the upper band is considered according to the following vector.
[5, 5, 3, 2, 5, 5, 3, 2, 5, 5, 3, 2, 5, 5, 3, 33, 11, 3].
In the WOA algorithm, the individual vector is represented in an n-dimensional space, where the parameters are defined in this higher-dimensional space rather than in a single space. This allows the WOA algorithm to align with the Convolutional Neural Network (CNN) model. The objective fitness of an individual is evaluated using the forward network, which takes into account the obtained values of hyperparameters. The efficiency of the configured CNN architectures for plant leaf disease detection is tested using a fitness function, which can be calculated using Equation (9).
| (9) |
Which,
is the last 5 values of accuracy in the runs. The flowchart of the WOA algorithm is illustrated in Figure 4.
Figure 4. Flowchart of the Whale Optimization Algorithm (WOA) algorithm.
3.4.2. Optimized hyperparameters
The results obtained by the network depend on the change of the hyperparameter. This has been shown in many studies. There are many modes to find the best performance of a network. In this study, the whale optimization algorithm was used to find the optimal values of hyperparameters. After setting the optimal values for the hyperparameters, the network was run for 100 epochs. In Table 4, the optimized hyperparameters are shown. The CNN uses these values. In this study, 18 parameters have been optimized: the number of filters, size of the kernels, type of activation function, size of pooling in blocks 1 to 3, and the dropout rate, dense layer size, and activation function in the output layer. In the WOA, the optimization ranges for each parameter are determined based on past studies and experience. The WOA is known for its ability to quickly find the optimal solution, which helps in reducing the computational load of the system. On average, the WOA algorithm takes approximately 11 minutes to complete 100 epochs.
Table 4. Initial and optimized hyperparameters value.
Block/output | Hyperparameter | Initial Value | Optimized Value |
Block 1 | Number of filters | 128 | 128 |
Size of the kernels | 2 | 9 | |
Activation function | Relu | elu | |
Size of pooling | 2 | 4 | |
Block 2 | Number of filters | 256 | 128 |
Size of the kernels | 2 | 4 | |
Activation function | Relu | elu | |
Size of pooling | 2 | 2 | |
Block 3 | Number of filters | 256 | 128 |
Size of the kernels | 2 | 9 | |
Activation function | Relu | elu | |
Size of pooling | 2 | 2 | |
Block 4 | Number of filters | 512 | 128 |
Size of the kernels | 3 | 2 | |
Activation function | Relu | sigmoid | |
Output | Dropout rate | 0.5 | 0.24 |
Dense layer size | 512 | 2048 | |
Activation function | softmax | elu |
4- Experimental results and analysis
In the proposed CNN, the input consists of a training image that represents different plant leaf diseases. The CNN's layers allow it to learn and recognize different features at various levels of complexity. The first layer learns basic features. The second layer recognizes more complex objects, and the third layer can identify specific details. The images are split into three subsets for training, validation, and testing. Specifically, 80% (70,295 images) are used for training, 10% (8,795 images) for validation, and another 10% (8,777 images) for testing the performance of the CNN.
4.1. Accuracy and loss plots
Plots depicting accuracy and loss are generated during the training and testing phases, with the number of epochs plotted on the x-axis. These plots give valuable information about datasets. The accuracy curve illustrates the performance of the CNN on both sets throughout the training process. A significant gap between the accuracy of the training phase and the test phase indicates overfitting. If the training accuracy is much higher than the test accuracy, it suggests that the model has memorized the training data and may not generalize well to new, unseen data. The size of this gap can indicate the severity of overfitting. Figure 5 illustrates that in most cases, the training curve closely follows the test curve, which is desirable. However, after the initial few epochs, the test accuracy starts to become lower than the training accuracy, indicating a potentially small amount of overfitting. To address this issue, a dropout rate, determined during the optimization stage, can be added. Dropout helps in preventing overfitting by encouraging the network to learn more robust and generalizable features that randomly set a fraction of the input units to 0 during training.
(a) |
(b) |
(c) |
(d) |
Figure 5. Graph of accuracy vs. epochs for the training and test sets (a) CNN Base, (b) VGG-16, (c) VGG-19 and (d) Optimized CNN.
In the CNN model, the accuracy of the train achieved 99.6%, indicating that the model was able to learn and classify the training images with a high degree of accuracy. The test accuracy reached 94.7% at epoch 100, which is also a relatively high percentage. These high percentages confirm the effectiveness of the model in identifying plant leaf diseases in real-world conditions. The loss function is utilized to evaluate the performance of the network on the training set. It quantifies the discrepancy between the predicted output and the actual output, providing a measure of how well the network is learning during training. In this case, the categorical cross-entropy loss function is used, which is suitable for segmenting images into multiple classes, such as detecting different types of diseases or healthy leaves.
Figure 6 illustrates that as the optimization iterations progress, the models' loss curve improves, with the loss decreasing after each or several iterations. This indicates that the models' predictions are becoming more accurate and the network is converging towards a better solution.
(a) |
(b) |
(c) |
(d) |
Figure 6. Graph of losses vs. epochs for the training and test sets (a) CNN Base, (b) VGG-16, (c) VGG-19 and (d) Optimized CNN.
In general, it is expected to observe more loss in both the test and train datasets during the initial epochs of training. This is because the model is still in the process of learning and adjusting its parameters to minimize the loss. As the number of epochs increases, the training loss continues to decrease, indicating that the model is improving its ability to fit the training data. However, the loss plot of the test set may initially decrease before starting to increase or fluctuate. This can be an indication of overfitting. Overfitting occurs when the model becomes too specialized in the training data and starts to capture random noise or variations specific to the training dataset. While the model performs fine on the training dataset, its ability to specify new, unseen data may decrease slightly, resulting in test errors. In the case of the CNN base and optimized network, there is a greater loss in the test dataset, telling a potential overfitting issue. This can be due to the model capacity or flexibility that necessary for the given problem. It is worth noting that in the CNN base, the CNN directly operates on images, and its performance on the test dataset achieves 94% accuracy compared to the 99% accuracy on the training dataset. The difference in performance can be attributed to downsizing the images, which may result in a loss of information.
4.2. Confusion matrix and performance metrics
The confusion matrix, as shown in Figure 7, provides a detailed analysis of how well the CNN model detects different disease types in the images. It presents information about both correct classifications and incorrect classifications made by the model. The matrix helps identify the confusion or misclassifications between different disease types. Each row in the matrix represents the actual disease type, while each column represents the predicted disease type by the CNN model. The elements of the matrix show the number or percentage of images that were classified into each disease type. By examining the confusion matrix, it is possible to identify which disease types are frequently confused or misclassified by the CNN model. This information can help understand the strengths and weaknesses of the model and potentially improving its performance in accurately detecting and classifying different disease types in plant leaves.
(a) |
(b) |
(c) |
(d) |
Figure 7. Confusion matrix for the test sets (a) CNN Base, (b) VGG-16, (c) VGG-19 and (d) Optimized CNN.
Table 5 summarizes the performance measures. The metrics include accuracy, F1-score, recall, precision, sensitivity, specificity, and others. These metrics provide insights into how well the classifier can differentiate between different classes and make accurate predictions.
Accuracy is calculated as the share of images in which the network produces the correct labels and can be predicted by measuring the error rate. Recall is calculated as the ratio of true positives to the sum of true positives and false negatives. It indicates the ability of the classifier to correctly identify positive cases. Precision is the ratio of true positives to the sum of true positives and false positives. It indicates the ability of the classifier to correctly classify positive cases without misclassifying negative cases.
To evaluate the proposed CNNs, values of metrics are calculated which provide a comprehensive evaluation of the CNN model's ability to classify and diagnose different types of leaf diseases accurately.
Table 5. Performance measures of Networks on the test set.
| CNN Base | VGG-16 | VGG-19 | Optimized CNN |
Accuracy | 0.94 | 0.9352 | 0.9068 | 0.9529 |
Precision | 0.947 | 0.9397 | 0.9123 | 0.955 |
Recall | 0.94 | 0.9362 | 0.9073 | 0.9531 |
F1_score | 0.94 | 0.9354 | 0.9064 | 0.953 |
Time | 7 | 29 | 37 | 11 |
Figure 8 illustrates the process of searching for the optimal solution using the WOA. The convergence curve is used to reflect the performance of the algorithm over iterations. The convergence curve shows how the WOA progresses in its search for the optimal solution over time. It demonstrates the changes in the objective function value or fitness score as the algorithm iteratively updates and improves its solutions. The curve typically starts with a higher value and gradually decreases as the algorithm explores and exploits the search space. As the iterations progress, the algorithm converges towards the optimal solution, and the curve tends to flatten out. By analyzing the convergence curve, it is possible to assess the performance of the WOA. A steeper decrease in the curve indicates faster convergence, while a slower decrease suggests slower progress towards the optimal solution.
The convergence curve provides a visual representation of how the WOA performs in terms of finding the optimal solution and can be used to evaluate and compare the performance of different optimization algorithms.
Figure 8. The convergence curve of fitness value vs. iterations.
Based on the information provided, Figure 8 demonstrates that the convergence trend of the Whale Optimization Algorithm (WOA) does not fall into local optima and results in reducing the classification error. The figure shows that the algorithm improves and achieves better results over 50 iterations. This behavior of the WOA is desirable as it indicates that the algorithm is effective in exploring the search space and avoiding getting stuck in local optima. Local optima are suboptimal solutions that are found within a limited region of the search space but are not the global optimal solution. By continuously improving over iterations, the WOA demonstrates its ability to refine its solutions and reduce classification error. This suggests that the algorithm is capable of finding better and more accurate solutions as it progresses. The fact that the convergence trend of the WOA shows continuous improvement over 50 iterations is a positive outcome, indicating that the algorithm is effective in optimizing the classification process and reducing errors.
4.3. Comparison with other works
Table 6 displays a comparison of the performance of the proposed model with other related works on the PlantVillage dataset.
Table 6. Performance comparison of the proposed method and other works.
Author | Year | Accuracy | Precision | Recall | F1score | Trainable parameters |
Shradha Verma[67] | 2022 | 0.9375 | 0.9286 | 0.9274 | 0.928 | 16.4 M |
Barkha M. Joshi[69] | 2023 | 0.9523 | 0.94 | 0.94 | 0.93 | 26 M |
This work | 2023 | 0.9529 | 0.955 | 0.9531 | 0.953 | 5.8 M |
This method shows an improved accuracy compared to other state-of-the-art methods. For this study, the optimized CNN has 5.791 million trainable parameters, which is less than most previous studies. In the case where the number of trainable parameters is less than other works in Table 6 this comparison shows the superiority of this model. Also, compared to the study of Barkha M. Joshi [69], the proposed model with about 10.6 M fewer parameters has obtained similar and slightly better results.
5- Discussion and conclusion
This study presents an overview of various approaches used for classifying plant leaf disease images, with a specific focus on evaluating the performance of recent CNN models. The proposed method combines CNN architecture with WOA for hyperparameter optimization. The results of this hybrid approach are highly promising, achieving 99% accuracy for the training dataset and 95% accuracy for the test dataset. Precision and Recall values are also impressive, at 99.59% and 99.58% respectively. These metrics demonstrate the effectiveness of the proposed optimized CNN architecture in accurately classifying plant leaf diseases.
One of the key advantages of this approach is that it eliminates the need for separate feature extraction and image segmentation steps. In traditional methods, feature extraction involves manual design and extraction of relevant features from images, which can be time-consuming and complex. Similarly, image segmentation aims to separate different objects or regions within an image, which can be challenging and computationally expensive. However, the proposed approach integrates these steps within the model itself, enabling it to simultaneously learn relevant features and perform segmentation. This end-to-end learning approach simplifies the overall pipeline and allows the model to automatically learn and adapt to the most relevant features and segmentation boundaries for the given task. Consequently, this can lead to improved performance and efficiency in image classification tasks. The optimized CNN model also demonstrates improved processing time and computational efficiency. Reducing the number of parameters in the model, it enables faster processing of each image and quicker convergence to the optimal solution. This enhanced efficiency and reduced training time make the model more suitable for agricultural applications, where timely disease classification is crucial for effective treatment and outcomes. Overall, the proposed optimized CNN model exhibits higher accuracy, faster convergence, and increased efficiency compared to baseline CNN models. These qualities make it a valuable tool for agricultural applications, where accurate and timely disease classification is essential.
References
1. Kwon, Y., et al., Electromyography-based decoding of dexterous, in-hand manipulation of objects: Comparing task execution in real world and virtual reality. IEEE Access, 2021. 9: p. 37297-37310.
2. Paymode, A.S. and V.B. Malode, Transfer learning for multi-crop leaf disease image classification using convolutional neural network VGG. Artificial Intelligence in Agriculture, 2022. 6: p. 23-33.
3. Singh, V. and A.K. Misra, Detection of plant leaf diseases using image segmentation and soft computing techniques. Information processing in Agriculture, 2017. 4(1): p. 41-49.
4. Malvade, N.N., et al., A comparative analysis of paddy crop biotic stress classification using pre-trained deep neural networks. Artificial Intelligence in Agriculture, 2022. 6: p. 167-175.
5. Kundu, N., et al., Disease detection, severity prediction, and crop loss estimation in MaizeCrop using deep learning. Artificial Intelligence in Agriculture, 2022. 6: p. 276-291.
6. Barbedo, J.G.A., A review on the use of unmanned aerial vehicles and imaging sensors for monitoring and assessing plant stresses. Drones, 2019. 3(2): p. 40.
7. Bock, C., et al., Plant disease severity estimated visually, by digital photography and image analysis, and by hyperspectral imaging. Critical reviews in plant sciences, 2010. 29(2): p. 59-107.
8. Kim, W.-S., D.-H. Lee, and Y.-J. Kim, Machine vision-based automatic disease symptom detection of onion downy mildew. Computers and Electronics in Agriculture, 2020. 168: p. 105099.
9. Barbedo, J. and J. Garcia, Digital image processing techniques for detecting, quantifying and classifying plant diseases. SpringerPlus 2: 660. 2013.
10. Lindow, S. and R. Webb, Quantification of foliar plant disease symptoms by microcomputer-digitized video image analysis. Phytopathology, 1983. 73(4): p. 520-524.
11. Martin, D.P. and E.P. Rybicki, Microcomputer-based quantification of maize streak virus symptoms in Zea mays. Phytopathology, 1998. 88(5): p. 422-427.
12. Alamri, N.M.H., M. Packianather, and S. Bigot, Deep learning: Parameter optimization using proposed novel hybrid bees Bayesian convolutional neural network. Applied Artificial Intelligence, 2022. 36(1): p. 2031815.
13. De Filippis, L.A.C., et al., ANN modeling to optimize manufacturing process. Advanced applications for artificial neural networks, 2018: p. 201-226.
14. Li, B.-h., et al., Applications of artificial intelligence in intelligent manufacturing: a review. Frontiers of Information Technology & Electronic Engineering, 2017. 18: p. 86-96.
15. Krizhevsky, A., I. Sutskever, and G.E. Hinton, Imagenet classification with deep convolutional neural networks. Communications of the ACM, 2017. 60(6): p. 84-90.
16. Marcus, G., Deep learning: A critical appraisal. arXiv preprint arXiv:1801.00631, 2018.
17. Hughes, D. and M. Salathé, An open access repository of images on plant health to enable the development of mobile disease diagnostics. arXiv preprint arXiv:1511.08060, 2015.
18. Mohanty, S.P., D.P. Hughes, and M. Salathé, Using deep learning for image-based plant disease detection. Frontiers in plant science, 2016. 7: p. 1419.
19. Liang, Q., et al., PD2SE-Net: Computer-assisted plant disease diagnosis and severity estimation network. Computers and electronics in agriculture, 2019. 157: p. 518-529.
20. Singh, K.K. An artificial intelligence and cloud based collaborative platform for plant disease identification, tracking and forecasting for farmers. in 2018 IEEE international conference on cloud computing in emerging markets (CCEM). 2018. IEEE.
21. Barbedo, J.G.A., et al., Annotated plant pathology databases for image-based detection and recognition of diseases. IEEE Latin America Transactions, 2018. 16(6): p. 1749-1757.
22. Singh, D., et al., PlantDoc: A dataset for visual plant disease detection, in Proceedings of the 7th ACM IKDD CoDS and 25th COMAD. 2020. p. 249-253.
23. Wiesner-Hanks, T., et al., Image set for deep learning: field images of maize annotated with disease symptoms. BMC research notes, 2018. 11(1): p. 1-3.
24. Prajapati, H.B., J.P. Shah, and V.K. Dabhi, Detection and classification of rice plant diseases. Intelligent Decision Technologies, 2017. 11(3): p. 357-373.
25. Oyewola, D.O., et al., Detecting cassava mosaic disease using a deep residual convolutional neural network with distinct block processing. PeerJ Computer Science, 2021. 7: p. e352.
26. Ahmad, A., D. Saraswat, and A. El Gamal, A survey on using deep learning techniques for plant disease diagnosis and recommendations for development of appropriate tools. Smart Agricultural Technology, 2023. 3: p. 100083.
27. Ahmad, A., D. Saraswat, and A. El Gamal, A survey on using deep learning techniques for plant disease diagnosis and recommendations for development of appropriate tools. Smart Agricultural Technology, 2022: p. 100083.
28. Azlah, M.A.F., et al., Review on techniques for plant leaf classification and recognition. Computers, 2019. 8(4): p. 77.
29. Jeon, W.-S. and S.-Y. Rhee, Plant leaf recognition using a convolution neural network. International Journal of Fuzzy Logic and Intelligent Systems, 2017. 17(1): p. 26-34.
30. Toda, Y. and F. Okura, How convolutional neural networks diagnose plant disease. Plant Phenomics, 2019.
31. Wu, X., D. Sahoo, and S.C. Hoi, Recent advances in deep learning for object detection. Neurocomputing, 2020. 396: p. 39-64.
32. Trappey, C.V., A.J. Trappey, and S.C.-C. Lin, Intelligent trademark similarity analysis of image, spelling, and phonetic features using machine learning methodologies. Advanced Engineering Informatics, 2020. 45: p. 101120.
33. Escalante, H.J., Automated machine learning—a brief review at the end of the early years. Automated Design of Machine Learning and Search Algorithms, 2021: p. 11-28.
34. LeCun, Y., et al., Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. 86(11): p. 2278-2324.
35. Fan, Y., et al., A Hybrid Sparrow Search Algorithm of the Hyperparameter Optimization in Deep Learning. Mathematics, 2022. 10(16): p. 3019.
36. Van Rijn, J.N. and F. Hutter. Hyperparameter importance across datasets. in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018.
37. Wang, Z. and J. Xuan, Intelligent fault recognition framework by using deep reinforcement learning with one dimension convolution and improved actor-critic algorithm. Advanced Engineering Informatics, 2021. 49: p. 101315.
38. Larochelle, H., et al. An empirical evaluation of deep architectures on problems with many factors of variation. in Proceedings of the 24th international conference on Machine learning. 2007.
39. Lerman, P., Fitting segmented regression models by grid search. Journal of the Royal Statistical Society Series C: Applied Statistics, 1980. 29(1): p. 77-84.
40. Bergstra, J. and Y. Bengio, Random search for hyper-parameter optimization. Journal of machine learning research, 2012. 13(2).
41. Hutter, F., H.H. Hoos, and K. Leyton-Brown. Sequential model-based optimization for general algorithm configuration. in Learning and Intelligent Optimization: 5th International Conference, LION 5, Rome, Italy, January 17-21, 2011. Selected Papers 5. 2011. Springer.
42. Talathi, S.S. Hyper-parameter optimization of deep convolutional networks for object recognition. in 2015 IEEE international conference on image processing (ICIP). 2015. IEEE.
43. Cui, J., et al., A novel framework of graph Bayesian optimization and its applications to real-world network analysis. Expert Systems with Applications, 2021. 170: p. 114524.
44. Lee, M., J. Bae, and S.B. Kim, Uncertainty-aware soft sensor using Bayesian recurrent neural networks. Advanced Engineering Informatics, 2021. 50: p. 101434.
45. Kong, H., et al., Energy management strategy for electric vehicles based on deep Q-learning using Bayesian optimization. Neural Computing and Applications, 2020. 32: p. 14431-14445.
46. Jin, N., et al., Highly accurate energy consumption forecasting model based on parallel LSTM neural networks. Advanced Engineering Informatics, 2022. 51: p. 101442.
47. del Rio Chanona, E.A., et al., Real-time optimization meets Bayesian optimization and derivative-free optimization: A tale of modifier adaptation. Computers & Chemical Engineering, 2021. 147: p. 107249.
48. Sun, L.-X., et al., Cluster analysis by simulated annealing. Computers & Chemistry, 1994. 18(2): p. 103-108.
49. Zhang, Y. and G. Huang, Traffic flow prediction model based on deep belief network and genetic algorithm. IET Intelligent Transport Systems, 2018. 12(6): p. 533-541.
50. Nevavuori, P., N. Narra, and T. Lipping, Crop yield prediction with deep convolutional neural networks. Computers and electronics in agriculture, 2019. 163: p. 104859.
51. Rangarajan Aravind, K. and P. Raja, Automated disease classification in (Selected) agricultural crops using transfer learning. Automatika: časopis za automatiku, mjerenje, elektroniku, računarstvo i komunikacije, 2020. 61(2): p. 260-272.
52. Krishnaswamy Rangarajan, A. and R. Purushothaman, Disease classification in eggplant using pre-trained VGG16 and MSVM. Scientific reports, 2020. 10(1): p. 1-11.
53. Atila, Ü., et al., Plant leaf disease classification using EfficientNet deep learning model. Ecological Informatics, 2021. 61: p. 101182.
54. Ji, M., L. Zhang, and Q. Wu, Automatic grape leaf diseases identification via UnitedModel based on multiple convolutional neural networks. Information Processing in Agriculture, 2020. 7(3): p. 418-426.
55. Sharma, P., Y.P.S. Berwal, and W. Ghai, Performance analysis of deep learning CNN models for disease detection in plants using image segmentation. Information Processing in Agriculture, 2020. 7(4): p. 566-574.
56. Azimi, S., T. Kaur, and T.K. Gandhi, A deep learning approach to measure stress level in plants due to Nitrogen deficiency. Measurement, 2021. 173: p. 108650.
57. Gadekallu, T.R., et al., A novel PCA–whale optimization-based deep neural network model for classification of tomato plant diseases using GPU. Journal of Real-Time Image Processing, 2021. 18: p. 1383-1396.
58. Sinha, A. and R.S. Shekhawat, Olive spot disease detection and classification using analysis of leaf image textures. Procedia Computer Science, 2020. 167: p. 2328-2336.
59. Raikar, M.M., et al., Classification and Grading of Okra-ladies finger using Deep Learning. Procedia computer science, 2020. 171: p. 2380-2389.
60. Franczyk, B., et al., Deep learning for grape variety recognition. Procedia Computer Science, 2020. 176: p. 1211-1220.
61. Kundu, N., et al., IoT and interpretable machine learning based framework for disease prediction in pearl millet. Sensors, 2021. 21(16): p. 5386.
62. Abayomi‐Alli, O.O., et al., Cassava disease recognition from low‐quality images using enhanced data augmentation model and deep learning. Expert Systems, 2021. 38(7): p. e12746.
63. Basavaiah, J. and A. Arlene Anthony, Tomato leaf disease classification using multiple feature extraction techniques. Wireless Personal Communications, 2020. 115(1): p. 633-651.
64. Aliyu, M.A., M.M.M. Mokji, and U.U.U. Sheikh, Machine learning for plant disease detection: An investigative comparison between support vector machine and deep learning. IAES International Journal of Artificial Intelligence, 2020. 9(4): p. 670.
65. Thakur, P.S., et al., Explainable vision transformer enabled convolutional neural network for plant disease identification: PlantXViT. arXiv preprint arXiv:2207.07919, 2022.
66. Alshammari, H., et al., Olive disease classification based on vision transformer and CNN models. Computational Intelligence and Neuroscience, 2022. 2022.
67. Verma, S., et al., SE-CapsNet: Automated evaluation of plant disease severity based on feature extraction through Squeeze and Excitation (SE) networks and Capsule networks. Kuwait Journal of Science, 2022. 49(1).
68. Thakur, P.S., T. Sheorey, and A. Ojha, VGG-ICNN: A Lightweight CNN model for crop disease identification. Multimedia Tools and Applications, 2023. 82(1): p. 497-520.
69. Joshi, B.M. and H. Bhavsar, Deep Learning Technology based Night-CNN for Nightshade Crop Leaf Disease Detection. International Journal of Intelligent Systems and Applications in Engineering, 2023. 11(1): p. 215-227.
70. Singh, V., A. Chug, and A.P. Singh, Classification of Beans Leaf Diseases using Fine Tuned CNN Model. Procedia Computer Science, 2023. 218: p. 348-356.
71. Ulutaş, H. and V. Aslantaş, Design of Efficient Methods for the Detection of Tomato Leaf Disease Utilizing Proposed Ensemble CNN Model. Electronics, 2023. 12(4): p. 827.
72. Tugrul, B., E. Elfatimi, and R. Eryigit, Convolutional neural networks in detection of plant leaf diseases: A review. Agriculture, 2022. 12(8): p. 1192.
73. Reddy, S.S.S., et al. Deep CNN Based Whale Optimization for Predicting the Rice Plant Disease in Real Time. in Artificial Intelligence and Data Science. 2022. Cham: Springer Nature Switzerland.
74. Kiran Pandiri, D., et al., POT-Net: solanum tuberosum (Potato) leaves diseases classification using an optimized deep convolutional neural network. The Imaging Science Journal, 2023: p. 1-17.
75. Suresh, G. and K. Seetharaman. Automated Groundnut Leaf Disease Recognition Using Whale Optimization Algorithm with Deep Learning Model. in 2022 International Conference on Augmented Intelligence and Sustainable Systems (ICAISS). 2022.
76. Bhattarai, S., New Plant Diseases Dataset. 2019, kaggle.
77. Shifat‐E‐Rabbi, M., et al., Cell image classification: a comparative overview. Cytometry Part A, 2020. 97(4): p. 347-362.
78. Goodfellow, I., Y. Bengio, and A. Courville, Deep learning. 2016: MIT press.
79. Chollet, F., Deep learning with Python. 2021: Simon and Schuster.
80. Simonyan, K. and A. Zisserman, Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
81. Hemdan, E.E.-D., M.A. Shouman, and M.E. Karar, Covidx-net: A framework of deep learning classifiers to diagnose covid-19 in x-ray images. arXiv preprint arXiv:2003.11055, 2020.
82. Sukegawa, S., et al., Deep neural networks for dental implant system classification. Biomolecules, 2020. 10(7): p. 984.
83. Mirjalili, S. and A. Lewis, The whale optimization algorithm. Advances in engineering software, 2016. 95: p. 51-67.
84. Hof, P.R. and E. Van der Gucht, Structure of the cerebral cortex of the humpback whale, Megaptera novaeangliae (Cetacea, Mysticeti, Balaenopteridae). The Anatomical Record: Advances in Integrative Anatomy and Evolutionary Biology: Advances in Integrative Anatomy and Evolutionary Biology, 2007. 290(1): p. 1-31.
85. Chen, W., et al., Spatial prediction of landslide susceptibility using gis-based data mining techniques of anfis with whale optimization algorithm (woa) and grey wolf optimizer (gwo). Applied Sciences, 2019. 9(18): p. 3755.
86. Watkins, W.A. and W.E. Schevill, Aerial observation of feeding behavior in four baleen whales: Eubalaena glacialis, Balaenoptera borealis, Megaptera novaeangliae, and Balaenoptera physalus. Journal of Mammalogy, 1979. 60(1): p. 155-163.
سند
الروابط
المراكز ذات الصلة
دعامة
الصفحات الرسمية
حقوق هذا الموقع الإلكتروني مملوكة لنظام SANAD Press Management. © 1442-1446