Abstract
Visual recognition of disease symptoms in plant leaves involves observing changes in leaf appearance that can indicate disease or plant stress. These symptoms include spots, discoloration, wilting, leaf rust, abnormal growth, and abnormal leaf drop. These symptoms can be identified and managed using advanced visual tools, such as remote sensing and drone imagery, and artificial intelligence-based algorithms. Therefore, present paper aimed to identify plant diseases in different crops using image processing and deep learning algorithms. The models including RegNet, ShuffleNet, and DenseNet were evaluated based on criteria such as accuracy, precision, recall, F1 score, and ROC and accuracy-recall curves. The overall accuracy of the RegNet model is about 96%, indicating it successfully identifies most rice and wheat diseases. The overall accuracy of the ShuffleNet model was 94.4%, a promising result, though slightly lower than that of the previous model. The highest accuracy, 97.2%, was achieved with the DenseNet model.
Introduction 
The life cycle of agricultural products undergoes significant changes due to diseases. Visual recognition of disease symptoms in plant leaves involves observing changes in leaf appearance that can indicate disease or plant stress. These symptoms include spots, discoloration, wilting, leaf rust, abnormal growth, and abnormal leaf drop. These symptoms can be identified and managed using advanced visual tools, such as remote sensing and drone imagery, and artificial intelligence-based algorithms.
Meshram et al. (2021) reviewed the applications of machine learning algorithms in agriculture in three phases: pre-harvest, during-harvest, and post-harvest. This review found that machine learning algorithms have achieved significant results in solving agricultural problems. Their study showed that there is a need to follow the path of machine learning with standard empirical methods. Researchers should create their own datasets and make them available on different platforms so others can use them to test and validate their models. Rahmn et al. (2025) developed a real-time leaf disease detection system for 8 plant species (potato, tomato, bell pepper, apple, corn, grape, peach, and rice) using deep learning techniques, including Inception, VGG, MobileNet, and DenseNet. Their dataset consisted of 30,945 images and 35 disease classes. The highest accuracy for disease detection was 98% for tomatoes with the Inception model, 100% for bell peppers with the MobileNet model, 100% for apples with the MobileNet model, 99% for grapes with the VGG model, 100% for peaches with the VGG model, and 98% for rice with the DenseNet model.
 
Material and Methods 
A total of 13,324 images of diseased leaves from 5 different crops, namely, cotton, potato, sugarcane, wheat, and rice (Figure 1), with different image sizes but a total of 4.79 GB, were obtained from the Kaggle site. Table 1 shows the setup and implementation features of the proposed model training.
Proposed Models
1. ShuffleNet Model
The ShuffleNet model is a convolutional neural network architecture with a lightweight computational structure and is less complex than models such as VGG and ResNet. Its disadvantages include its possible lower performance compared to models such as EfficientNet.
2. RegNet Model
The RegNet model is another convolutional neural network architecture that is highly flexible for computer vision applications such as image classification, object recognition, and segmentation. Compared to complex architectures such as VGG, the network is simple and efficient, and offers a better balance between accuracy and computational cost. 
3- DenseNet Model
Unlike standard architectures that only connect between consecutive layers, the DenseNet model connects all layers together. This means that because each layer adds new learned features to subsequent layers, the number of parameters is reduced, and the model consumes less memory. 
To evaluate the efficiency of the classifiers, the criteria of precision, accuracy, recall, and F-score were extracted from the confusion matrices, the operating characteristic curve diagram, and the precision-recall diagram.
Results and Discussion 
The ResNet model does not overfit; both the training and validation datasets have good accuracy, and the training and validation errors have decreased and then stabilized. This means the model has been able to train on new data effectively. In the ShuffleNet classifier, it was observed that the training accuracy increased rapidly. The validation accuracy also increased but did not reach the training accuracy, indicating mild overfit. The DenseNet model also showed an increase in accuracy, but validation accuracy fluctuated and did not reach the desired value. This model was not stable in learning new data.
In all models, most values lie on the main diagonal of the matrices, indicating that the models make correct predictions across classes. The model performs well across most classes but may be biased in some.
In all three proposed models, the lines for many classes are close to the upper right corner, indicating high precision and recall. This means that the model performs very well in recognizing most classes. Classes whose lines are not close to the upper corner may need improvement, which may be due to class similarity or data imbalance.
The overall accuracy of the RegNet model is about 96%, indicating it successfully identifies most rice and wheat diseases. The overall accuracy of the ShuffleNet model was 94.4%, a promising result, though slightly lower than that of the previous model. The highest accuracy, 97.2%, was achieved with the DenseNet model.
Conclusions 
This study aimed to identify plant diseases in different crops using image processing and deep learning algorithms. The results showed that different models, including RegNet, ShuffleNet, and DenseNet, performed well at classifying plant diseases. The models were evaluated based on criteria such as accuracy, precision, recall, F1 score, and ROC and accuracy-recall curves.
• The RegNet model performed stably and showed the best balance between accuracy and recall.
• The ShuffleNet model had a slight overfit, but overall provided acceptable performance.
• The DenseNet model, despite its dense connections, had fluctuations in validation accuracy in some cases and needs improvement.
• The ROC and accuracy-recall curves showed that the models performed very well in identifying most classes. Still, some classes need improvement, which may be due to feature similarity or data imbalance.
Finally, the results of this research confirm that the use of deep learning algorithms in plant disease identification can effectively increase agricultural productivity and be an important step towards smart agriculture.
Author Contributions
Sajad Sabzi: Conceptualization 
Razyieh Pourdarbani: Writing
Mahan Masoumzade: Methodology
Data Availability Statement
"Not applicable".
Ethical Considerations
It is confirmed adherence to ethical standards, including avoidance of data fabrication, falsification, plagiarism, and misconduct.
Conflict of Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper
Funding Statement
The author(s) received no specific funding for this research