Deep Learning for Early Disease Detection: A CNN Approach to Classify Potato, Tomato, and Pepper Leaf Diseases

Karan Sarawagi; Hardik Dhiman; Ashutosh Pagrotra; Navjot Singh Talwandi

doi:10.20944/preprints202406.0986.v1

Submitted:

12 June 2024

Posted:

14 June 2024

You are already at the latest version

Abstract

Early disease detection is crucial for maximizingcrop yields and minimizing financial losses in agriculture. Inorder to categorise leaf diseases in pepper, tomato, and potato plants, this study suggests a revolutionary CNN design. The algorithm collects disease-specific features from input photos by using data augmentation techniques to artificially enlarge the training dataset. ReLU-activated convolutional layers gradually capture features at different levels, from low to high, while max pooling layers minimise spatial dimensionality. After thecharacteristics are extracted, a fully-connected layers with a softmax activation function classify them into illness groups. A well chosen dataset containing labelled healthy and sick leaves is used to train and validate the model. Generalizability to untesteddata is ensured via performance evaluation on an independent testing set. The field of deep learning for agricultural applications benefits from this research. This strategy has the ability to rev- olutionise agricultural practices by enabling automated disease identification. This might result in early interventions, better crop health, and eventually sustainable agricultural production.

Keywords:

convolutional neural network (CNN)

;

data augmentation

;

plant disease detection

;

ReLU activation

;

deep learning

;

softmax activation

Subject:

Computer Science and Mathematics - Computer Vision and Graphics

I. Introduction

The development of effective and precise techniques for early plant disease detection is required by the rapidly expanding field of precision agriculture. For farmers to minimise financial losses, maximise yields, and maintain ideal crop health, early diagnosis is essential. Conventional disease identification techniques usually entail an agricultural expert’s manual visual assessment. Unfortunately, this method has several built-in drawbacks, such as the fact that it takes a lot of time, is prone to subjectivity, and requires a lot of labour, particularly in large-scale farming operations.

Plant disease classification tasks can now be automated thanks to recent developments in the field of deep learning, especially with Convolutional Neural Networks (CNNs) [1]. CNNs are remarkably good at extracting useful characteristics from image input. Because of this feature, they are especially suitable for identifying the visual cues that distinguish healthy plant leaves from diseased ones. This study looks into how well a CNN-based method works for identifying good and unhealthy potato, tomato, and pepper plant leaves.

Our work makes numerous significant contributions to the development of deep learning applications in agriculture. First, as described in [2], we present a CNN architecture made especially to classify the illnesses in pepper, tomato, and potato leaves. This design makes use of data augmentation tech niques, which entail creating variations by artificially manipulating training data (e.g., flipping, rotating). This procedure reduces the likelihood of overfitting by increasing the diversity of the training data and fortifying the model’s capacity to generalise to unobserved real-world scenarios. Subsequently, we assess the model’s effectiveness using a benchmark image dataset that includes labelled pictures of the target plant species’ healthy and diseased leaves. In-depth classification accuracy measures will be presented in the article, offering a numerical assessment of the model’s performance. We will also provide a thorough visual representation of the training and validation procedures, providing more understanding of the convergence properties and learning behaviour of the model.

The goal of study is to add to the growing corpus of research on deep learning applications in agriculture. The information in this research may open the door to the creation of automatedand intelligent field systems for the identification of plantdiseases.

II. Literature Review

A highly effective CNN architecture was presented by the author [3] for the detection of potato illnesses. This is crucial because fungal infections have a big effect on potato harvests, which are a major source of food. They used Adam and cross- entropy to optimise the model and produced an image-based training dataset in order to accomplish this. The model used much less resources (99.39 percent parameter reduction) to achieve high accuracy (99.53 percent) in disease identification. Using image processing and machine learning, the authors [4] suggest an automated system to recognise and categorise potato leaf illnesses (Early Blight and Late Blight). With a 97 percent accuracy rate, the Random Forest classifier emerged as the most promising technique for automatically classifying the disease.

The use of deep learning for very accurate automated identification of potato and tomato leaf diseases was suggested by the authors [5]. For this aim, two CNN architectures were created especially. The model was optimised using Adam and Categorical Crossentropy on an image dataset. With the least amount of resources used, the model was able to detect diseases with excellent accuracy (98.14 percent for potatoes and 91.34 percent for tomatoes).

A multi-level deep learning model was presented by the authors [6] for the analysis of the disease. Due to regional training data, existing approaches frequently lack generalizability. The suggested model isolates potato leaves initially by using YOLOv5 for picture segmentation. Next, in the separated leaves, an innovative CNN finds both early andlate blight. This model beat previous approaches on the PlantVillage dataset in terms of accuracy and computational efficiency, achieving 99.75 percent accuracy on a Pakistani dataset.

A CNN architecture was proposed by the authors [7] to detect fungal-caused late and early blight infections in potatoes. An important source of food, potato harvests are severely impacted by this illness. The model has a Softmax final judgement function and analyses data using categorical cross- entropy and the Adam optimizer. With a parameter count of 10,089,219, the model minimises resource use while achieving excellent accuracy (99 percent) in disease identification.

A convolutional neural network architecture was presented by the authors [8] to categorise potato leaf blight (healthy, early blight, late blight). Their testing accuracy with a 14- layer model was 98 percent on average. They employed data augmentation to grow their dataset from 1,722 to 9,822 photos in order to increase accuracy.

A deep learning model was presented by the authors [9] to identify potato, tomato, and pepper leaf diseases, such as early and late blight. To obtain different characteristics from pictures, they combined transfer learning with a pre-trained VGG19 model. In the test dataset, logistic regression achieved a classification accuracy of 97.8 percent, outperforming other classifiers.

Using PySpark dataframe, the authors [10] proposed MCIP, a framework for analysing crop image data. By employing K- means for clustering and PCA for feature selection, MCIP overcomes the drawbacks of existing techniques. It can be used on other crops and identified potato, tomato, and pepper leaf diseases with near-perfect accuracy (almost 100 percent).

The authors [11] suggest an automated machine learning- based approach for detecting potato diseases in order to enhance potato yield and agricultural digitization in Bangladesh. Their approach uses machine learning and image processing to identify disease from the potato photos. The model, which was trained on more than 2000 healthy and diseased potato leaves, demonstrated a testing accuracy of 99.23 percent, indicating that machine learning presents a viable option for the diagnosis of potato diseases.

Four potato leaf diseases can be classified, according to the authors’ [12]. Leaf photos of potato were analysed using the VGG16 and VGG19 convolutional neural network architectures for the purpose of disease classification. With an average accuracy of 91 percent, this method proved that deep learning is a viable method for detecting potato diseases.

For the classification of plant diseases, the authors [13] suggest a two-step deep learning method. To determine which CNN architecture performs the best, they first compare a number of them, including modified and mixed models. Based on training epochs, F1-score, validation accuracy, and loss, they assess these models. Second, they use various deep learning optimizers to optimise this best model. A dataset comprising 26 illnesses from 14 different plant species is used to train all of the models. Outperforming earlier techniques, the Xception architecture with the Adam optimizer produced the best accuracy (99.81 percent) and F1-score (0.9978). A CNN-based method was suggested by the authors [14] for the detection of tomato leaf diseases. Their algorithm detects a number of illnesses, including septoria blight and early blight, using image processing.

In an effort to increase potato production, the authors [15] suggest an advanced machine learning system for identifying and categorising potato leaf diseases. High agricultural yields depend on early disease diagnosis, yet manual inspection is labor-intensive and skill-intensive.

A. Comparative Analysis of Studies

Table 1 summarizes recent advancements in machine learning (ML) for potato, tomato, and pepper leaf disease detection. All studies achieved impressive accuracy, ranging from 97.8percent to near-perfect (almost 100 percent). [8] utilized a 14-layer CNN, achieving 98 percent accuracy. However, their approach necessitates a large training dataset. [9] explored transfer learning with VGG19, attaining 97.8 percent accuracy, suggesting efficient feature extraction capabilities. [10] introduced MCIP, a framework leveraging K-means clustering and PCA for feature selection. This framework achieved near- perfect accuracy and demonstrates potential generalizability to other crops. [11] prioritized high accuracy (99.23 percent) to support agricultural digitization initiatives. Finally, [7] proposed a CNN specifically designed for late/early blight detection, achieving 99 percent accuracy with an efficient model architecture (Softmax, categorical cross-entropy, Adam optimizer). These findings collectively underscore the efficacy of ML for potato leaf disease detection. While all approaches exhibited strong performance, further research should investigate methods for enhancing efficiency and generalizability across diverse datasets and disease types.

III. Methodology

This paper presents a novel CNN architecture for categoriz ing various leaf diseases. The model leverages the TensorFlow and Keras libraries to construct an efficient image recognition pipeline. The following sections delve into the technical aspects of the model, including data preparation, augmentation, network architecture, training configuration, and evaluation metrics.

1): >1) Data Acquisition and Preprocessing: The cornerstone of training a robust image classification model lies in a well- curated dataset. The implemented approach utilizes the tf.keras.preprocessing image dataset from directory function for data access and preprocessing.

Following data loading, essential preprocessing steps are introduced. The get dataset partitions tf function partitionsthe dataset into training, validation, and testing sets. This segregation ensures the model is trained on the representative subset, its performance is evaluated on a separate unseen hold- out set (validation set), and its generalizability is assessed on another unseen set (testing set). The hyperparameter configuration (train split, val split, and test split) controls the data allocation proportions, ensuring balanced representation for training and evaluation.

To further enrich the training data and mitigate overfitting, data augmentation techniques are employed. Overfitting occurs when a model excessively memorizes training data specifics, resulting in poor performance on unseen examples. The data augmentation function introduces random horizontal and vertical flips along with rotations within the training images.

2): >2) Model Architecture Design: The proposed system’s core is a meticulously crafted CNN architecture specifically designed for various disease classifications of leaves. The model employs a sequential stack of convolutional and dense layers to extract informative features from input images and map them to the corresponding disease categories.

The architecture commences with preprocessing layers. The resize and rescale block performs two crucial steps. First, it resizes all input images to a uniform dimension of 128x128 pixels using a resizing layer. This ensures consistency in the input data format for subsequent convolutional layers. Second, it rescales the pixel intensities to a range of 0 to 1 using a rescaling layer. This normalization step facilitates the training process and promotes faster convergence.

The first convolutional layer utilizes 32 filters with a size of (3, 3) and uses the activation function, the rectified linear unit (ReLU). Subsequent convolutional layers progressively increase the number of filters (64, 128, and 256) to capture increasingly complex features. After each convolutional layer, we join a max pooling layer with a size of (2, 2) to reduce the matrix dimensionality by half.

Following feature extraction, the model employs fully- connected (dense) layers for disease classification. A dense layer with 64 neurons and a ReLU activation function serves as a hidden layer, further processing the features. The lastlayer has the same number of outputs as the disease classes(15 in this case) and employs a softmax activation function. The softmax function outputs a probability distribution across all disease categories, with the highest value indicating the predicted class for the input image.

3): >3) Training Configuration and Evaluation: For the model compilation, the optimizer Adam has been used, which is a beneficial choice for models because of its efficiency and fast learning rate adjustments, which lead to faster convergence. The loss function employed is the sparse categorical cross- entropy, well-suited for multi-class classification problems.It measures the difference between the predicted probability distribution and the true one-hot encoded label for each image. The model is trained for several epochs, where epoch defines one complete pass of the entire training dataset. During each epoch, the model changes its parameters to reduce the loss, progressively improving its classification accuracy.
3): A. Proposed Model

A custom independently trained neural network CNN architecture is used to classify tomato, potato, and pepper leaf diseases using deep learning. Below is a detailed explanation of each step:

Step I Data Gathering: Get a potato, tomato, and pepper leaf disease-specific picture collection. Although publicly available datasets are useful, they need be tailored to target diseases like early and late blight. Choose a dataset that uses JPEG or PNG and labels each image with a disease status.

Step II: Preprocessing data prepares images for model training after data gathering. This phase accelerates CNN learning and ensures data format uniformity. An outline of preprocessing workflow:

Data Loading: Use tf.keras.preprocessing.image dataset from directory to load image data from a directory. Image- label pairings are automatically extracted using the directory’s subfolder organisation.

A Resizing layer in the model design scales all input images to a consistent size. The appropriate size (128x128 pixels) can be calculated from the dataset’s attributes and computer power. Image Normalisation: Apply a rescaling layer to normalise photo pixel intensities to 0–1. Normalisation ensures equal contribution of all images during training, preventing bias towards brighter or darker shots.

Step III: Carefully partition the loaded dataset. The instruction set, which contains 80 percent of the data, trains the model. Adjusting hyperparameters and avoiding overfitting requires the validation set, which makes up about 10 percent. The remaining 10 percent—the testing set—is a fully held-out set used for final performance evaluation on unseen data after training. Additional shuffles, such shuffle or shuffle(buffer size), ensure the model is trained on a random order of images, preventing biases towards specific sequences.

Step IV: Optional Data Enhancement Dataset size affects deep learning model performance. Lack of data might induce overfitting. This risk is reduced by data augmentation. By controlling training picture alterations and transformations, these approaches create new synthetic instances and artificially enlarge the training dataset. Define an Augmentation Strategy using the data augmentation function or a related mechanism. This technique could use random rotations within a specific range (e.g., +/- 20 degrees), horizontal and vertical flips, and cropping to manage training image disparities.

Step V Model Creation: The system is based on a refined CNN architecture designed for potato, tomato, and pepper leaf disease categorization. CNNs are great at extracting spatial properties from image data because they learn from local patterns.

Convolutional layers apply learnable filters that scan the input image and activate when patterns or features are found. These filters capture edges and other low-level features.

Pooling layers reduce convolutional layer feature maps. This reduces data spatial dimensionality, boosting computational performance and preventing overfitting. Max pooling, which selects the most activations within a window, is popular.

Classification Layer Configuration: Convolutional layer characteristics must be converted into sickness categories. How the final step goes:

Flattening layers convert convolutional layer multi- dimensional feature maps into one-dimensional vectors that may be input into dense layers.

As a hidden layer, the retrieved characteristics are processed by a dense layer consisting of the activation function ReLU. The ReLU activation function gives the model non-linearity, allowing it to learn more complex feature correlations.

Last dense layer employs softmax activation function and has as many neurons as disease classes (15 for late blight, early blight, and healthy leaves). The output probability distribution maximum value from the softmax functionreflects the projected class for the input image across all sickness categories.

Step VI Training Models: After designing the model architecture, train it with ready-made data:

The compilation of the CNN model is done using the optimizer Adam, which is popular for its efficiency and learning rate flexibility. Use the sparse categorical cross-entropy loss function for multi-class classification.

Configure how many epochs the model will undergo training. Twenty epochs is a usual starting point, however it can be altered depending on dataset size, validation effectiveness, and computer capability.

Start training by fitting the model to the training set using the learning rate, optimizer, loss function, and epochs. Each iteration, the model updates its filter weights and biases to minimise the training and validation set loss function.

Validation Monitoring: Monitor the model’s validation set performance throughout training. Check validation loss and accuracy. Early stopping may be used to cease training if validation loss does not improve after a given number of epochs. To avoid overfitting, pause the training process when the model becomes too familiar with the training data and loses its ability to generalise to new data.

Step VII Model Assessment: Analysing the model’s performance on unknown data after training determines its generalizability:

Testing Set: Evaluate the model’s output on a private testing set. Various parameters such as recall, precision, F1-score, and accuracy can be calculated to assess the model’s potato, tomato, and pepper leaf disease classification accuracy.

Visualisation and Interpretation: Visualise the model’s predictions using test photographs to understand its pros and cons. Confusion matrices can identify misclassifications and improvement opportunities.

Step VIII: Optional Model Refinement: If the original model performance is unacceptable, consider these refining methods: Adjustable hyperparameters include optimizer selection, learning rate, dense layer neurons, convolutional layers, and filter sizes. Hyperparameter setups can improve model performance.

If data augmentation is utilised to improve model robustness and generalizability, experiment with different procedures or amounts.

In conclusion, A structured approach that includes data gath ering, preprocessing, augmentation (optional), model construction, training, and evaluation is the focus of the methodology. Image datasets that are publicly accessible or domain-specific datasets specifically designed to target illnesses (late blight and early blight) can be utilised. By scaling and normalising the data, data preparation guarantees consistency in image format and makes model training easier. When data augmentation techniques are used, the training dataset is artificially expanded in order to reduce overfitting.

The carefully constructed CNN architecture, which is intended to extract useful information from the images, is the central component of the system. A sequential stack of convolutional and dense layers is used in the architecture. Features are retrieved by convolutional layers with pooling layers, and then processed for illness classification by dense layers with activation functions.

Figure 1. Flowchart of the model.

B. Pseudocode for potato, tomato, and pepper leaf disease Classification

1:procedure CLASSIFYPOTATOLEAF(image)

2: preprocessed_image ← Preprocess(image)

3: predictions ← ForwardPass(CNN_model, preprocessed_image)

4: disease_class ← ExtractDiseaseClass(predictions)

5: return disease_class

6: function PREPROCESS(image)

7: resized_image ← Resize(image, target_size)

8: normalized_image ← Normalize(resized_image)

9: return normalized_image

10: function FORWARDPASS(model, picture)

11: result ← model(picture)

12: return result

13: function EXTRACTDISEASECLASS(predictions)

14: disease_class ← argmax(predictions)

15: return disease_class

IV. Results

Using a well-defined deep learning pipeline, the CNN architecture intended for the separation of various diseases in leaves was trained and assessed. The outputs are shown in this section.

Figure 2 and Figure 3 are showing the sample training and sample test respectively that are received from the model, the accuracy scores are shown above each output. sample predictions where the model performed as intended. The ”Actual” labels (ground truth) on the left-hand side precisely match the ”Predicted” labels on the right-hand side. Additionally, the associated probabilities for the predicted classes are high (e.g., 98.48 percent for ”Tomato Late blight”), indicating a strong degree of confidence in the classifications. These results are promising and demonstrate the model’s ability to accurately classify some unseen test set images.

Model Performance: Using the visual elements that are retrieved from the photos, the proposed CNN architecture attempts to automatically categorise potato leaf images into several disease categories (e.g., late blight, early blight, healthy leaves). During the training phase, the model acquires these attributes and creates a mapping between the image data and the associated illness diagnoses. Analysis of Training and Validation Performance: Several parameters such as correctness and loss are majorly used to assess the model’s performance. Figure 4 shows the correctness and loss curves for training and validation against several epochs. The model learns from more training data and improves its internal parameters across the number of epochs, improving its accuracy in classifying potato leaf illnesses.

With more epochs, the training accuracy curve shows a rising trend. This shows that the training image classification is being learned by the model efficiently. Similar to the training accuracy, the validation accuracy curve likewise shows a positive tendency, albeit maybe more slowly. This is a promising finding, indicating that the model is effectively generalising and picking up significant patterns outside of the particular training set.

The training loss curve shows a trend that decreases as the number of epochs increases. This indicates that the model can minimise its loss function using the training set of data. Additionally, there is a diminishing trend in the validation loss curve, which suggests that the model is learning generalizable characteristics for disease classification rather than just over- fitting to the training set.

V. Discussion

In order to classify illnesses of the pepper, tomato, and potato leaves, this study employed deep learning. A well- defined deep learning methodology was used to construct and train a convolutional neural network (CNN) architecture. Training and validation performance curves showed positive results. The rising training accuracy curve revealed the model’s ability to learn from training data. Increasing the validation accuracy curve also showed some generalizability to new validation data. Training and validation loss curves showed a decreased trend, indicating the model’s capacity to minimise loss function.

However, a restricted reading of these plots does not rule out overfitting. Further testing on a larger unseen testing set and thorough validation performance monitoring throughout training are needed to ensure robust and generalizable model performance. Early halting or data supplementation may reduce overfitting and increase the model’s illness presentation adaptability. Studying sample test data provided more insights. The model’s capacity to classify unknown test set photos showed its practicality. However, misclassification occurrences highlighted the need for generalizability improvement. Adding a larger, more diverse dataset or adopting data augmentation can help solve these challenges.

Future research will examine a statistically sound unseen testing set. Future research could examine the effects of domain-specific knowledge in model architecture and transfer learning methodologies employing pre-trained models on larger image datasets. Our goal is to refine the deep learning pipeline utilising these methods to create a highly accurate and generalizable model for early and effective potato, tomato, and pepper leaf disease detection. This will help farmers make better decisions that boost crop health and yield.

Conclusion

This study used a independently trained CNN architecture to demonstrate a deep learning approach for classifying tomato, potato and pepper leaf diseases. A well-organized pipeline comprising data collection, preprocessing (with optional augmentation), model construction, training, and evaluation was part of the methodology. By efficiently learning discriminative characteristics from photos of potato leaves, the applied CNN architecture made it possible to classify a variety of illnesses, including early and late blight. Positive findings were obtained from the examination of the training and validation performance curves. The model’s capacity to understand the training data and reach a certain degree of generalizability to unobserved validation data was demonstrated by the growing accuracy curves and declining loss curves. To guarantee robust and generalizable performance in real-world circumstances, further testing on a bigger, unknown testing set and careful observation of validation performance throughout training are necessary. It is possible to investigate methods such as data augmentation and early halting to reduce overfitting and improve the model’s capacity to accommodate changes in illness presentation. The deep learning methodology that has been suggested presents a viable strategy for classifying potato, tomato, and pepper leaf diseases.

References

S. G. Durai, T. S. S. G. Durai, T. S. Sujithra, and M. M. Iqbal, “Image classification for potato plant leaf disease detection using deep learning,” 2023 International Conference on Sustainable Computing and Smart Systems (ICSCSS), pp. 154–158, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:259364231.
Falaschetti, L.; Manoni, L.; Di Leo, D.; Pau, D.; Tomaselli, V.; Turchetti, C. A CNN-based image detector for plant leaf diseases classification. HardwareX 2022, 12, e00363. [Google Scholar] [CrossRef] [PubMed]
Lee, T.-Y.; Lin, I.-A.; Yu, J.-Y.; Yang, J.-M.; Chang, Y.-C. High Efficiency Disease Detection for Potato Leaf with Convolutional Neural Network. SN Comput. Sci. 2021, 2, 1–11. [Google Scholar] [CrossRef]
M. A. Iqbal and K. H. Talukder, “Detection of potato disease using image segmentation and machine learning,” 2020 International Conference on Wireless Communications Signal Processing and Networking (WiSPNET), pp. 43–47, 2020. [Online]. Available: https://api.semanticscholar. 2218.
Mori, Y.; Limbasia, B.; Gupta, R.K.; Bharti, S.K. AI Based Potato and Tomato Leaf Disease Detection Using CNN and Pre-Trained Model. 2023 7th International Conference On Computing, Communication, Control And Automation (ICCUBEA). LOCATION OF CONFERENCE, IndiaDATE OF CONFERENCE; pp. 1–6.
Rashid, J.; Khan, I.; Ali, G.; Almotiri, S.H.; AlGhamdi, M.A.; Masood, K. Multi-Level Deep Learning Model for Potato Leaf Disease Recognition. Electronics 2021, 10, 2064. [Google Scholar] [CrossRef]
Lee, T.-Y.; Yu, J.-Y.; Chang, Y.-C.; Yang, J.-M. Health Detection for Potato Leaf with Convolutional Neural Network. 2020 Indo-Taiwan 2nd International Conference on Computing, Analytics and Networks (Indo-Taiwan ICAN). LOCATION OF CONFERENCE, IndiaDATE OF CONFERENCE; pp. 289–293.
N. E. M. Khalifa, M. H. N. N. E. M. Khalifa, M. H. N. Taha, L. M. A. El-Maged, and E. Hassanien, “Artificial intelligence in potato leaf disease classification: A deep learning approach,” 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:230567631.
Tiwari, M. Ashish, N. Gangwar, A. Sharma, S. Patel, and S. Bhardwaj, “Potato leaf diseases detection using deep learning,” 2020 4th International Conference on Intelligent Computing and Control Systems (ICICCS), pp. 461–466, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:219990325.
Chaudhary, Y.; Pathak, H. MCIP: Mining Crop Image Data On pysparkdataframe Using Feature Selection and Cluster Based Techniques. Int. Acad. Publ. House 2023, 34, 106–119. [Google Scholar] [CrossRef]
M. I. Tarik, S. M. I. Tarik, S. Akter, A. A. Mamun, and A. Sattar, “Potatodisease detection using machine learning,” 2021 Third InternationalConference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV), pp. 800–803, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:233135786.
Sholihati, R.A.; Sulistijono, I.A.; Risnumawan, A.; Kusumawati, E. Potato Leaf Disease Classification Using Deep Learning Approach. 2020 International Electronics Symposium (IES). LOCATION OF CONFERENCE, IndonesiaDATE OF CONFERENCE; pp. 392–397.
Saleem, M.H.; Potgieter, J.; Arif, K.M. Plant Disease Classification: A Comparative Evaluation of Convolutional Neural Networks and Deep Learning Optimizers. Plants 2020, 9, 1319. [Google Scholar] [CrossRef] [PubMed]
M. K. R. Asif, M. A. M. K. R. Asif, M. A. Rahman, and M. H. Hena, “Cnn based disease detection approach on potato leaves,” 2020 3rd International Conference on Intelligent Sustainable Systems (ICISS), pp. 428–432, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:231683443.
G, G.B.; N, N.; S, N.; D, M.M.; Patil, R.V. Advanced Machine Learning Technique to Detect Disease in Potato. Int. J. Adv. Res. Sci. Commun. Technol. 2022, 253–256. [Google Scholar] [CrossRef]

Figure 2. Sample Training.

Figure 3. Sample testing.

Figure 4. Accuracy and Loss of Testing and Validation.

Table 1. Comparative analysis.

Literature	Methodology	Results	Gaps
[7]	CNN forlate/early blight detection	99 percent	High accuracy,efficient resource usage
[8]	14-layer CNNwith data augmentation	98 percent	High accuracy, butrequires large dataset
[9]	Transferlearning with VGG19 model	97.8 percent(Logistic Regression)	Efficient featureextraction, good accuracy
[10]	MCIPframework with K-means andPCA	Near-perfect (almost 100 percent)	Highly accurate,potentially generalizable, fastexecution
[11]	Machinelearning withimage processing	99.23 percent	Excellent accuracy,supportsagricultural digitization

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.