Customized Convolutional Neural Networks for Plant Disease Detection on Leaf Images

Abstract

Ensuring agricultural productivity and food security depends on efectively detecting and classifying plant leaf diseases. In this study, we present a tailored Convolutional Neural Network (CNN) approach specifically designed to identify plant leaf diseases. Our methodology is based on a carefully curated dataset that includes 29 distinct disease classes. By using transfer learning and fine-tuning techniques, we carefully optimize the CNN architecture to suit the unique characteristics of the dataset, resulting in improved accuracy and robustness. Through extensive experimentation and evaluation, we demonstrate the efectiveness of our approach in accurately diagnosing plant leaf diseases across multiple classes. Our model achieves impressive performance, with a notable accuracy of 96.53% and a minimal loss of 0.1. Outperforming existing methods on a range of evaluation metrics, we emphasize the superiority of our customized CNN approach through comprehensive comparative analyses. This study represents a significant advancement in computer vision techniques for agriculture by providing a reliable and eficient solution for automated plant disease diagnosis. The proposed methodology holds great promise for practical implementation in agricultural systems, enabling early disease detection and management to reduce crop losses and promote sustainable agricultural practices.

1. Introduction

Plant diseases represent a major threat to global food security [ 1 ], [ 2 ], leading to significant yield losses and imposing heavy economic burdens on agricultural industries worldwide. Timely and accurate detection of these diseases is critical for implementing efective intervention strategies to prevent extensive crop damage. Recent advancements in computer vision and machine learning have shown great promise in automating the detection and diagnosis of plant diseases through the analysis of leaf images [ 3 ], [ 4 ]. Convolutional Neural Networks ( ) have proven to be highly efective for image classification tasks [ 5 ], [ 6 ], including the detection of plant diseases. However, achieving optimal performance in this area involves addressing several challenges, such as dataset heterogeneity, class imbalance, and symptom variability across diferent plant species. Moreover, existing CNN architectures often require customization to efectively manage the specific characteristics of plant disease datasets [29]. In this paper, we introduce a customized CNN approach designed specifically for identifying plant leaf diseases. Our study utilizes a modified dataset encompassing 29 distinct disease classes, representing a wide variety of plant species and pathologies. Our contribution has two primary objectives: first, to develop a robust and precise model capable of accurately distinguishing between multiple disease classes; and second, to address the limitations of existing CNN architectures by tailoring the model to better accommodate the unique characteristics of the plant disease dataset. By utilizing transfer learning and fine-tuning techniques, we optimize the CNN architecture to adapt to the unique features of the dataset, thereby improving both accuracy and robustness. Through extensive experimentation and evaluation, we showcase the efectiveness of our approach in accurately diagnosing plant leaf diseases across multiple classes. Furthermore, we conduct comprehensive comparative analyses to highlight the advantages of our customized CNN approach over traditional methods. Overall, our study aims to advance computer vision techniques in agriculture by ofering a reliable and eficient solution for automated plant disease diagnosis. By tackling the challenges associated with plant disease classification, our customized CNN approach shows significant potential for practical implementation in agricultural systems. This can facilitate early disease detection and management, helping to mitigate crop losses and promote sustainable agricultural practices. We begin with an Introduction section that delineates the motivation, objectives, and primary contributions of our study. Subsequently, the Problem Statement section delineates the challenges in plant disease detection and articulates the specific aims of our research. In the Related Works section, we scrutinize existing methodologies, elucidating how our approach advances the field. The Methodology section elaborates on the specialized dataset, the design of the customized CNN architecture, the training process, and techniques for hyperparameter tuning. The Results and Discussion section presents the performance metrics of our model, deliberates on its generalization and scalability, and confronts the challenges of deploying deep learning models on diverse datasets, including issues pertaining to interpretability and transparency. Finally, the Conclusion and Future Works section recapitulates our findings and delineates potential avenues for further research and practical implementations.

2. Problem Statment

In the fields of agriculture and forestry, detecting plant diseases is vital for ensuring healthy crop production and forest management [26]. However, traditional methods of disease detection have significant disadvantages [27]. They rely on manual inspection and semi-automated processes, which are costly, time-consuming, prone to error, and lack precision and specificity [28]. Furthermore, these methods are not exhaustive and struggle to predict disease outbreaks accurately, resulting in delayed responses and potential agricultural losses (figure 1). The current detection methods involve visual inspection and manual data collection, which are labor-intensive, time-consuming, and prone to human error. These methods do not provide real-time results and often fail to accurately predict disease outbreaks, leading to delayed responses and significant agricultural losses. Given these drawbacks, there is an urgent need for an advanced solution that can overcome these limitations. The proposed solution should be automated, cost-efective, provide real-time results, be precise, and ofer accurate predictions with a low error rate. Additionally, it should be exhaustive in its analysis, covering a wide range of potential diseases and environmental stressors. Artificial Intelligence (AI) presents a promising alternative to traditional detection methods. By using AI technologies, it is possible to develop systems that meet these criteria, thereby enhancing disease detection and management in agriculture and forestry. AI can process large amounts of data quickly and accurately, providing realtime insights and predictions that enable proactive measures to protect plant health against various threats. In conclusion, transitioning to AI-based detection methods can address the critical shortcomings of traditional approaches, ofering a comprehensive and eficient solution for managing plant diseases and adapting to environmental changes.

3. Background of the study

3.1. Deep learning theory

The concept of Deep Learning (DL) was introduced in a seminal paper by Hinton et al., published in Science in 2006 [ 8 ]. Deep learning involves the use of neural networks for data analysis and feature learning [ 9 ]. In this approach, multiple hidden layers are used to extract features from the data. Each hidden layer acts as a perceptron, extracting low-level features that are then combined to form more abstract, high-level features. This approach efectively addresses the issue of local minima. Unlike traditional algorithms that rely on manually designed features, deep learning automates feature extraction, which has generated significant interest among researchers. It has been successfully applied in various domains, including computer vision [ 10 ], pattern recognition [ 11 ], speech recognition, natural language processing, and recommendation systems. Conventional methods for image classification and recognition rely on manually designed features, which only capture basic features and struggle to extract deep, complex information from images. Deep learning overcomes this limitation by directly learning from raw images, obtaining multi-level feature information that ranges from low-level to high-level semantic features [ 12 ]. Traditional algorithms for detecting plant diseases and pests primarily rely on manually designed features, which is a challenging and experience-dependent process that cannot automatically learn from images. On the other hand, deep learning models, with their multiple layers, possess robust autonomous learning and feature representation capabilities, allowing for automatic feature extraction in image classification and recognition tasks. This makes deep learning highly promising in the field of plant disease and pest image recognition[ 32]. Currently, several well-known deep neural network models have been developed within the realm of deep learning, including deep belief networks (DBN), deep Boltzmann machines (DBM), stacked de-noising autoencoders (SDAE), and deep CNNs. These models ofer significant advantages over traditional manual feature extraction methods by automating feature extraction from high-dimensional feature spaces in image recognition tasks [21]. As the volume of training data increases and computational power improves, the representational power of deep neural networks continues to grow. The rise of deep learning is transforming both industry and academia, with deep neural network models consistently outperforming traditional approaches. Among these models, deep convolutional neural networks have become the most popular framework in recent years.

3.2. Convolutional neural network

Convolutional Neural Networks (CNNs) are a type of deep learning model [22] that is particularly well-suited for image classification tasks, such as detecting leaf diseases. The architecture of a CNN consists of multiple layers, including fully connected layers, max pooling, and normalization layers. The initial layer is the input layer, followed by convolutional layers, which apply various 2D filters to the image to extract features [23]. These features are then downsampled through pooling layers, creating a more compact representation. Fully connected (FC) layers, which are considered learnable features, process these extracted features to learn and optimize weights [24]. These FC layers are also crucial for making classifications, such as recognizing diferent plant diseases. The CNN [ 30],[31]learning process begins with training, using labeled images as input. Once trained, the model can accurately identify diferent types of diseases.

4. Related Works

We’ll examine various studies that have utilized machine learning and deep learning methods for plant disease detection. Throughout this review, we’ll focus on their methodologies, findings, and pinpoint the gaps our approach seeks to address. So, In the literature we detect several works related to our study. We illustrate some of them as follows: ∙ Xu et al. [ 7 ] tackled the task of identifying corn leaf diseases (healthy, leaf blight, rust) in intricate ifeld settings with a scarcity of data. They introduced a CNN model leveraging VGG16 and transfer learning. By utilizing weight parameters pre-trained on ImageNet, they attained an average recognition accuracy of 95.33%. ∙ In (Hatuwal, B., and Thapa, D. (2020)) [ 15 ], the authors aimed to tackle significant crop losses in developing nations like Nepal, stemming from the delayed identification of plant diseases. They proposed a method to classify and predict plant diseases utilizing machine learning models, including Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest Classifier (RFC), and CNN. The study utilized Haralick texture features such as contrast, correlation, entropy, and inverse diference moments for SVM, KNN, and RFC models, while CNNs were directly fed with images. Their findings revealed that CNN achieved the highest accuracy at 97.89%, surpassing RFC (87.43%), SVM (78.61%), and KNN (76.96%) across sixteen distinct image categories. These results underscore the superior eficacy of CNNs in plant disease classification. ∙ In (Roy, A., and Patel, D. (2024))[ 16 ], the authors tackled the prevalent issue of leaf diseases in agriculture through disease classification employing deep learning techniques. They evaluated the performance of VGG-16, VGG-19, InceptionV3, and DenseNet-121 architectures, providing a comparative analysis. DenseNet-121 achieved the highest accuracy at 91.75%. The study encompassed training a dataset containing 13 diferent diseases, accompanied by an analysis based on validation accuracy, loss, and the number of epochs. This research underscored DenseNet121’s superiority over other deep learning models, afirming its eficacy in accurately classifying leaf diseases. ∙ The work by (Lingwal et al. (2023)) [ 17 ] addresses the critical need for early detection and classification of plant diseases to prevent their spread and minimize crop damage. By using deep learning techniques, specifically CNNs, the study focuses on classifying tomato leaf diseases using the PlantVillage dataset. This dataset includes nearly 16,000 leaf images, which were divided into training, test, and validation sets with ratios of 70%, 20%, and 10%, respectively[ 17 ]. The research compares a custom-developed CNN model with four transfer learning models: DenseNet121, ResNet50, Inception-V3, and VGG-16. Performance evaluations based on accuracy and crossentropy loss indicated that both VGG-16 and the custom CNN model achieved impressive results, with validation accuracies of 90% and 83% on the test set, respectively[ 17 ]. These findings underscore the efectiveness of CNNs, especially transfer learning approaches, in accurately classifying tomato leaf diseases, thereby providing a valuable tool for early disease detection in agriculture[ 17 ]. ∙ The study by (Zheng et al. (2023)) [18]emphasizes the critical role of computer vision in detecting plant diseases, particularly the necessity for accurate pattern recognition. A CNN was trained using a dataset comprising 22,930 tomato leaf images. The baseline model achieved a training accuracy of 90% [18]. Comparative analysis of various architectures, including VGG16, MobileNet, and InceptionV3, revealed that MobileNet had the highest training accuracy of 91% and was the most eficient[ 18]. Despite MobileNet’s superior performance, the proposed CNN architecture ofers faster training due to its shallower design. This research lays the foundation for future eforts in developing lightweight, fast, and accurate algorithms for classifying plant diseases, ensuring their practical application in agriculture[18].

The summarized studies encompass a variety of crops and diseases, including those afecting corn, wheat, tomatoes, apples, and various fungal infections. They employ a range of methods, from traditional machine learning algorithms such as SVM, KNN, and RFC, to advanced deep learning architectures like VGG16, CNN, DenseNet-121, MobileNet, and SqueezeNet. The dataset sizes vary considerably, with image counts ranging from a few thousand to nearly 37,000. The reported accuracy metrics generally demonstrate high efectiveness in disease detection. Studies utilizing advanced deep learning methods like CNNs and DenseNets typically achieve higher accuracy levels, highlighting their eficacy in image-based plant disease detection. Table 1 summarizes these studies, including the accuracy metrics used for model evaluation.

5. Methodology

Our methodology (see in Figure2) is designed to ensure accurate classification of plant diseases through a systematic process. First, we curate a customized dataset and preprocess it, carefully splitting it into training and test sets. Data augmentation techniques are then applied to enhance the dataset, improving the model’s robustness. We construct a CNN model with precision, selecting convolutional and pooling layers to efectively extract key features while maintaining eficiency and accuracy. The model is optimized to balance performance with minimized parameters. After constructing the model, it is rigorously trained using advanced optimization techniques, followed by performance evaluation on the test set to assess its efectiveness in real-world scenarios. The final model is deployed for swift and accurate plant disease prediction, providing a reliable tool for early crop health detection. Throughout Reference Object Xu et al. (2020) [ 7 ] Corn Kiruthika et al. Leaf (2019)[ 14 ] Hatuwal and Lee Leaf (2020)[ 15 ] Roy et al. Leaf (2024)[ 16 ] Lingwal et al. (2023) Tomato VGG-16 [ 17 ] Zheng et al. Tomato (2023)[18] Shin et (2021)[19] Zhong et al. (2020) Apple [20]

Method VGG16 ANN SVM KNN RFC CNN DenseNet-121

Dataset – 6108 images 36,958 images –

Performance Accuracy=95.33% Accuracy=93.33% Accuracy=78.61% Accuracy=76.96% Accuracy=87.43% Accuracy=97.89%

Accuracy=91.75% 16,000 images

Accuracy=90% al. Fungal SqueezeNet-MOD2 11,600 images

Accuracy=92.61% MobileNet 22,930 images

Accuracy=91% DenseNet-121 2462 images

Accuracy=92.29% the process, we carefully manage the dataset, model architecture, and parameter optimization to ensure precise and eficient disease classification.

5.1. Customized Dataset

The database contains a collection of plant leaf images, consisting of 39 diferent classes (figure 3). These classes include healthy leaves as well as leaves afected by various diseases or health issues. Each class represents a specific type of plant disease or health condition, such as apple scab, common corn rust, or tomato leaf mold. The dataset consists of a total of 61,486 images. After modifying and cleaning the database, we have created a new dataset with 29 diferent classes, representing various plant diseases and health states. The total number of images in the new dataset is now 31,573. To better manage our model training, we have divided this dataset into two parts: a training set and a test set. The training set contains 26,830 images, while the test set has 4,743 images (table 2). The purpose of modifying and cleaning the database is to personalize it for our needs. We have used an existing database that includes 16 types of plants and their associated diseases. This optimization helps to streamline processing on the server by avoiding the need to manage two separate processes.

5.2. CNN architecture

The CNN architecture, outlined in (table 3 and figure4), contains a total of eight layers. It starts with an input layer that accommodates images of dimensions 256x256 pixels with three color channels (RGB). The next three layers are Conv2D layers for feature extraction. The first Conv2D layer contains 32 iflters of size 5x5, followed by a Rectified Linear Unit (ReLU) activation function. This is followed by MaxPooling2D layers with pool sizes of 3x3 and 2x2 to downsample the spatial dimensions of the feature maps. The subsequent two Conv2D layers contain 32 and 64 filters of size 3x3, respectively, each with ReLU activation. These are followed by additional MaxPooling2D layers to further downsample the feature maps. The seventh layer flattens the resulting feature maps into a one-dimensional array. This is followed by two Dense layers with ReLU activation, containing 512 and 128 units, respectively. Dropout regularization is applied to the first Dense layer to mitigate overfitting. The final layer is the output layer, which contains 29 units with a Softmax activation function, facilitating multi-class classification by outputting class probabilities.

5.3. Hyperparameters configuration

The hyperparameters presented in the table 4 have been carefully selected and fine-tuned through a series of preliminary tests in order to optimize the performance of the model. These values were determined using a systematic approach.

• Image Dimensions: The image dimensions were chosen based on the typical size and aspect ratio of the input images in the dataset. We experimented with various dimensions and found that 256 × 256 pixels struck a good balance between capturing important features and computational eficiency. • Batch Size: We tried diferent batch sizes and settled on 32 because it allowed for eficient training without excessive memory consumption or compromising model performance. • Number of Classes: The number of classes in the dataset was predetermined based on the nature of the classification problem. In this case, there are 29 distinct classes representing diferent types of plant diseases.

Layer Type

Input (InputLayer)

Conv2D MaxPooling2D

Conv2D MaxPooling2D

Conv2D MaxPooling2D

Flatten

Dense Dropout

Dense Dense

These specific values were selected through an iterative process of experimentation and evaluation, ensuring optimal performance and generalization ability of the model.

6. Results and discussion

6.1. Results

The customized CNN architecture applied to the customized dataset achieved an outstanding performance with an accuracy of 96.53% and a loss value of 0.1. These results indicate that the model is highly efective at classifying plant diseases, demonstrating both precision and reliability in its predictions.

The figure 5 illustrates the performance of our customized CNN model on the training and validation datasets.

The confusion matrix (Figure 6) provides a detailed breakdown of the performance of our customized CNN model on the test dataset. Each cell in the matrix represents the number of predictions made by the model for each class, allowing us to see where the model performs well and where it might be making errors. The confusion matrix is a crucial tool for evaluating the performance of a classification model. By examining the diagonal and of-diagonal values, we can identify which classes the model predicts well and which ones it struggles with. This can guide further improvements in the model or dataset. It allows us to calculate these metrics for each class: • Precision: It tells us how many of the predicted positive instances are actually correct. • Recall: It indicates how many actual positive instances are correctly identified by the model. • F1-Score: The harmonic mean of precision and recall, providing a single metric that balances both.

By leveraging the insights from the confusion matrix, we can iteratively refine our model, making targeted adjustments to enhance its overall performance and reliability.

6.2. Discussion

In this study, we have developed a customized CNN architecture tailored for a specialized dataset containing images of various plant diseases. Our primary objective was to create a model capable of accurately identifying diferent plant diseases, making it a vital tool for agricultural management and disease control.

6.2.1. Model Performance

Our CNN model achieved an impressive accuracy of 96.53% and a loss value of 0.1 on the test dataset. These metrics underscore the model’s ability to learn and recognize intricate patterns and features associated with each plant disease class efectively. The robustness of our model is further evidenced by the training and validation accuracy and loss graphs, as well as the confusion matrix, which illustrates the model’s proficiency in correctly classifying the majority of plant disease images with minimal misclassifications.

6.2.2. Hyperparameter Tuning

Finding the ideal hyperparameters for training a deep learning model, such as the learning rate, batch size, and number of layers, can be a challenging and time-consuming task. It often requires extensive experimentation.

6.2.3. Interpretability and transparency

Interpretability and transparency are essential when deploying machine learning models, particularly in scenarios where understanding and justifying decisions is critical. Deep learning models like CNNs are often considered "black boxes" due to their complexity, making it hard to interpret their internal workings. This is a significant issue in fields like medical diagnosis or autonomous driving, where trust in the model’s decisions is paramount. In our study, we tackled this challenge by creating a custom CNN architecture for plant disease classification. Unlike pre-trained models with many layers and parameters, our model is simpler, with fewer parameters, reducing complexity and enhancing interpretability. This streamlined design makes it easier to understand how features are processed and which ones are most influential in predictions. Additionally, the architecture incorporates domain-specific knowledge, which aligns the model’s decisions with biological factors behind plant diseases, further increasing transparency. 6.2.4. Scalability and Adaptability • Adaptation to New Data: A significant challenge is ensuring that the model can be easily adapted to new data and diferent contexts, such as new plant diseases or variations, without requiring extensive retraining. • Model Scalability: The model should be able to handle a growing dataset and the addition of new classes without the need for extensive reengineering, while still maintaining its performance.

7. Conclusion

In this study, we have developed a customized Convolutional Neural Network for identifying plant diseases using a specialized dataset. The model achieved an impressive accuracy of 96.53% and a low loss value of 0.1, demonstrating its robustness and potential for practical agricultural applications. Moving forward, expanding the dataset to include more plant species and diseases, as well as images from diferent geographic regions, will improve the model’s generalizability. Additionally, testing the model under real-world conditions and employing advanced hyperparameter tuning techniques can further enhance its performance. Exploring model pruning and quantization techniques can also help reduce computational requirements, allowing for deployment on mobile and low-power edge devices. To make the model accessible to farmers and agricultural professionals, it is crucial to integrate it with Internet of Things (IoT) devices for real-time monitoring. Furthermore, developing user-friendly mobile and web applications will facilitate the use of the model. These eforts will contribute to agricultural sustainability and food security. In conclusion, our customized CNN model holds great promise for plant disease detection. With ongoing development and refinement, it has the potential to become a crucial tool in global agriculture.

Declaration on Generative AI

The authors have not employed any Generative AI tools. Security and Data Sciences: Select Proceedings of 3rd International Conference on MIND 2021.

Singapore: Springer Nature Singapore, 2023. [18] Zheng, Ji, and Minjie Du. "Study on tomato disease classification based on leaf image recognition based on deep learning technology." International Journal of Advanced Computer Science and Applications 14.4 (2023). [19] Shin, Jaemyung, et al. "A deep learning approach for RGB image-based powdery mildew disease detection on strawberry leaves." Computers and electronics in agriculture 183 (2021): 106042. [20] Zhong, Yong, and Ming Zhao. "Research on deep learning in apple leaf disease recognition."

Computers and electronics in agriculture 168 (2020): 105146. [21] Ben Mesmia, Walid, and Kamel Barkaoui. "Production Chain Modeling based on Learning Flow

Stochastic Petri Nets.", Soft Computing (2024), https://doi.org/10.1007/s00500-024-09865-y [22] Krichen, Moez. "Convolutional neural networks: A survey." Computers 12.8 (2023): 151. [23] Hcini, Ghazala, Imen Jdey, and Habib Dhahri. "Investigating Deep Learning for Early Detection and Decision-Making in Alzheimer’s Disease: A Comprehensive Review." Neural Processing Letters 56.3 (2024): 1-38. [24] Hcini, Ghazala, Imen Jdey, and Hela Ltifi. "Improving Malaria Detection Using L1 Regularization

Neural Network." JUCS: Journal of Universal Computer Science 28.10 (2022). [25] Topaloglu, Ismail. "Deep learning based convolutional neural network structured new image classification approach for eye disease identification." Scientia Iranica 30.5 (2023): 1731-1742. [26] Abbas, Aqleem, et al. "Drones in plant disease assessment, eficient monitoring, and detection: a way forward to smart agriculture." Agronomy 13.6 (2023): 1524. [27] Ray, Monalisa, et al. "Fungal disease detection in plants: Traditional assays, novel diagnostic techniques and biosensors." Biosensors and Bioelectronics 87 (2017): 708-723. [28] Ngugi, Lawrence C., Moataz Abelwahab, and Mohammed Abo-Zahhad. "Recent advances in image processing techniques for automated leaf pest and disease recognition–A review." Information processing in agriculture 8.1 (2021): 27-51. [29] Mastour, I., Sliman, L., Charroux, B., Djemaa, R. B., Barkaoui, K. . Privacy-preserving Collaborative Computation: Methods, Challenges, and Directions. In 2023 International Conference on Computer and Applications (ICCA) (pp. 1-6). IEEE. [30] Jdey, Imen. "Trusted smart irrigation system based on fuzzy IoT and blockchain." International

Conference on Service-Oriented Computing. Cham: Springer Nature Switzerland, 2022. [31] Jdey, Imen, et al. "Fuzzy fusion system for radar target recognition." International Journal of

Computer Applications & Information Technology 1.3 (2012): 136-142. [32] Hcini, Ghazala, Imen Jdey, and Hela Ltifi. "HSV-Net: a custom CNN for malaria detection with enhanced color representation." 2023 International Conference on Cyberworlds (CW). IEEE, 2023.

References

[1] Fones , Helen N. , et al. "Threats to global food security from emerging fungal and oomycete crop pathogens . " Nature Food 1.6 ( 2020 ): 332 - 342 .

[2] Singh , Vipin Kumar , et al. "Current status of plant diseases and food security." Food security and plant disease management . Woodhead Publishing , 2021 . 19 - 35 .

[3] Wani , Javaid Ahmad , et al. "Machine learning and deep learning based computational techniques in automatic agricultural diseases detection: Methodologies, applications , and challenges. " Archives of Computational methods in Engineering 29.1 ( 2022 ): 641 - 677 .

[4] Ouhami , Maryam , et al. "Computer vision, IoT and data fusion for crop disease detection using machine learning: A survey and ongoing research." Remote Sensing 13.13 ( 2021 ): 2486 .

[5] Chen , Leiyu , et al. "Review of image classification algorithms based on convolutional neural networks . " Remote Sensing 13.22 ( 2021 ): 4712 .

[6] Hossain , Md Anwar , and Md Shahriar Alam Sajib. "Classification of image using convolutional neural network (CNN)." Global Journal of Computer Science and Technology 19.2 ( 2019 ): 13 - 14 .

[7] Xu , J. H. , et al. "Recognition of corn leaf spot and rust based on transfer learning with convolutional neural network." Transactions of the CSAM 51.2 ( 2020 ): 230 - 236 .

[8] Hinton , Geofrey E. , and Ruslan R. Salakhutdinov . "Reducing the dimensionality of data with neural networks . " science 313 .5786 ( 2006 ): 504 - 507 .

[9] Aggarwal , Charu C. Neural networks and deep learning . Vol. 10 . No. 978 . Cham: springer, 2018 .

[10] Venkateswara , Hemanth, Shayok Chakraborty , and Sethuraman Panchanathan . "Deep-learning systems for domain adaptation in computer vision: Learning transferable feature representations . " IEEE Signal Processing Magazine 34.6 ( 2017 ): 117 - 129 .

[11] Abiodun , Oludare Isaac , et al. "Comprehensive review of artificial neural network applications to pattern recognition." IEEE access 7 ( 2019 ): 158820 - 158846 .

[12] Jdey , Imen, Ghazala Hcini , and Hela Ltifi . "Deep Learning and Machine Learning for Malaria Detection: Overview, Challenges and Future Directions . " International Journal of Information Technology & Decision Making ( 2023 ): 1 - 32 .

[13] Hang , Jie , et al. "Classification of plant leaf diseases based on improved convolutional neural network . " Sensors 19 .19 ( 2019 ): 4161 .

[14] Kiruthika , S. Usha , et al. "Detection and classification of paddy crop disease using deep learning techniques." International Journal of Recent Technology and Engineering 8 .3 ( 2019 ): 4353 - 4359 .

[15] Hatuwal , B. K. , A. Shakya , and B. Joshi . Plant leaf disease recognition using random forest, KNN, SVM and CNN . Polibits 62 , 13 - 19 . doi: 10 .17562. PB- 62 - 2 , 2020 .

[16] Roy , Vikas Kumar , et al. "Leaf disease recognition: Comparative analysis of various convolutional neural network algorithms." Intelligent Prognostics for Engineering Systems with Machine Learning Techniques . CRC Press, 2024 . 59 - 71 .

[17] Lingwal , Surabhi, Komal Kumar Bhatia, and Manjeet Singh . "Deep convolutional neural network approach for tomato leaf disease classification." Machine Learning, Image Processing , Network