Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
ScienceDirect
ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2023) 000–000
Procedia Computer Science 00 (2023) 000–000 www.elsevier.com/locate/procedia
ScienceDirect www.elsevier.com/locate/procedia
Procedia Computer Science 235 (2024) 3468–3478

International
InternationalConference onMachine
Conference on MachineLearning
Learningand
and Data
Data Engineering
Engineering (ICMLDE
(ICMLDE 2023)
2023)
International Conference on Machine Learning and Data Engineering (ICMLDE 2023)
Early Disease Detection in Plants using CNN
Early Disease Detection in Plants using CNN
Tejaswinia,*,Priyanka Rastogia, Swayam Duaa, Manikantaa, Vikas Dagara
Tejaswinia,*,Priyanka Rastogia, Swayam Duaa, Manikantaa, Vikas Dagara
a
Manav Rachna International Institute of Research & Studies, Faridabad, Haryana, India
a
Manav Rachna International Institute of Research & Studies, Faridabad, Haryana, India

Abstract
Abstract
The Indian economy benefits from early disease detection in plant leaves. According to reports, 10-30% of crops suffer harm from
diseases
The Indian thateconomy
are not discovered during
benefits from thedisease
early curing process.
detectionRecent
in plant advancements in deeptolearning,
leaves. According particularly
reports, 10-30% Convolutional
of crops suffer harm Neural
from
Networks
diseases that (CNNs),
are nothave paved during
discovered the waytheforcuring
transformative solutions
process. Recent in disease identification
advancements in deep learning, for particularly
agriculture. Convolutional
This study addresses
Neural
the critical(CNNs),
Networks issue of have
earlypaved
disease thedetection by harnessing the
way for transformative powerinofdisease
solutions deep learning models,
identification forspecifically
agriculture.CNNs. Different
This study leaf
addresses
disease detection
the critical issue technologies
of early diseaseare used for different
detection crops. The
by harnessing the pre-trained
power of deep deeplearning
learningmodels,
model isspecifically
used in thisCNNs.
study toDifferent
identify and
leaf
categorize leaf diseases.
disease detection A dataset
technologies of tomato,
are used potato,crops.
for different and bell
Thepepper leaf pictures
pre-trained from the
deep learning plantisvillage
model used inrepository
this study was employed
to identify and
for the current
categorize leaf investigation. The developed
diseases. A dataset of tomato,model can
potato, detect
and 12 plantleaf
bell pepper diseases in from
pictures normaltheleaf tissue.
plant Therepository
village quantitative
wasassessment
employed
of
forour
theCNN-based technique reveals
current investigation. an impressive
The developed model accuracy
can detectrate12of 86.21%.
plant Thisinnotable
diseases normalaccuracy underscores
leaf tissue. the efficacy
The quantitative of our
assessment
approach in the challenging
of our CNN-based techniquedomain
revealsof anplant diseaseaccuracy
impressive detection. Our
rate offindings
86.21%.show potentialaccuracy
This notable for the larger agricultural
underscores environment
the efficacy of our
beyond
approach theinimmediate quantitative
the challenging domainadvantages. This technical
of plant disease detection.advance not only
Our findings paves
show the way
potential forfor
theimproved crop output
larger agricultural and lower
environment
losses,
beyondbut thealso brings in
immediate a new era of
quantitative data-drivenThis
advantages. sustainability in agriculture.The
technical advance bottom
not only paves theline
wayisfor that our research
improved crop "offers
output aand
tangible
lower
pathway
losses, butfor alsoleveraging
brings in aAI-powered solutions to
new era of data-driven address the
sustainability long-standing challenges
in agriculture.The bottom lineof plantourdisease
is that researchdetection, thereby
"offers a tangible
significantly
pathway for contributing
leveraging to the well-being
AI-powered of farmers
solutions and the the
to address sustenance of the Indian
long-standing agricultural
challenges of plantsector.
disease detection, thereby
significantly
© 2024 Thecontributing to the well-being
Authors. Published by ELSEVIERof farmers
B.V.and the sustenance of the Indian agricultural sector.
© 2024 The Authors. Published by ELSEVIER B.V.
© 2024
This is anTheopen Authors.
accessPublished
article under by ELSEVIER B.V.
the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review
This is an openunder responsibility
access article of
under the
thescientific
CC
Peer-review under responsibility of the scientific committee committee
BY-NC-ND of of
license the International Conference on Machine
the(https://creativecommons.org/licenses/by-nc-nd/4.0)
International Conference on Machine Learning
Learning and and
Data
Data Engineering
Peer-review
Engineering under responsibility of the scientific committee of the International Conference on Machine Learning and
Keywords: Deep Learning; classification; CNN; leaf disease detection.
Data Engineering
Keywords: Deep Learning; classification; CNN; leaf disease detection.

1. Introduction
1. Introduction
India is a developing nation where agriculture supports around 70% of the people. Plant disease results in a significant
India is aindeveloping
decline nation
the quantity andwhere agriculture
quality supports
of agricultural around
goods. 70% of the
Therefore, topeople.
ensure Plant
food disease results
security, bothinearly
a significant
disease
decline inand
detection thethe
quantity and quality
development of crop
of better agricultural
varietiesgoods. Therefore, to ensure food security, both early disease
is essential.
detection and the development of better crop varieties is essential.
* Corresponding author. Tel.: + 91 7986799578
E-mail address: tejaswinigupta606@gmail.com
* Corresponding author. Tel.: + 91 7986799578
E-mail address: tejaswinigupta606@gmail.com

1877-0509 © 2024 The Authors. Published by ELSEVIER B.V.

This is an open
1877-0509 access
© 2024 Thearticle
Authors.under the CC by
Published BY-NC-ND
ELSEVIERlicense
B.V. (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under
This is an open responsibility
access of the
article under thescientific committee
CC BY-NC-ND of (https://creativecommons.org/licenses/by-nc-nd/4.0)
license the International Conference on Machine Learning and Data
Engineering
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and Data
Engineering
1877-0509 © 2024 The Authors. Published by ELSEVIER B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the International Conference on Machine Learning and
Data Engineering
10.1016/j.procs.2024.04.327
2 Author name / Procedia Computer Science 00 (2019) 000–000
Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478 3469

The traditional technique of disease identification has been manual examination by farmers or professionals, but this
may be time-consuming and expensive, making it impracticable for millions of small and medium-sized farms
worldwide. The development of automated systems for plant disease identification by checking for telltale signals on
leaves has been sped up by the construction of computer vision models. These methods attempt to make the detecting
process as straightforward and error-free for farmers as possible.

Studies have concentrated on employing image processing and feature extraction to build methods for disease
detection. In order to recognise and categorise plant diseases, numerous machine learning models have been used.
Since deep learning (DL) models have been developed, this area of research currently seems to offer a lot of potential
for increased accuracy. Numerous customised DL architectures are used with various visualisation approaches to
detect and classify the symptoms of plant diseases. In addition, various performance metrics are used to evaluate these
frameworks and procedures. Teaching a neural network to understand patterns and symptoms shared by various
ailments is one step in the deep learning process for detecting plant diseases. The deep learning model gains the ability
to distinguish between healthy and unhealthy plants by examining plant photos, making automated detection possible.
By doing away with manual inspection, this method yields quicker and more precise results. Convolutional neural
networks and other deep learning algorithms extract pertinent elements from the photos, allowing the model to
categorise the state of the plant's health. By enabling early identification and intervention, this technology has the
potential to revolutionise agricultural practices, resulting in higher crop yields and decreased economic losses.

Thus, this research paper intends to build a plant disease classifier for three plant species. These include the tomato
(Solanum Lycopersicon), the pepper bell (Capsicum annuum), and the potato (Solanum tuberosum). The model is
trained to recognize a small set of diseases or health conditions specific to each species. The model predicts the disease
in these plants by using the proposed model. The treatment to stop the spread of the disease in other parts of the plant
can be stopped by the farmer, and further corrective measures can be taken by him in a timely manner. The research
objectives of this paper are: i) The model's general suitability for categorizing diseases using leaf images. ii) Evaluate
the model performance iii) Use the trained model to deploy a user-friendly mobile application.

The rest of the paper is structured into five sections: the literature survey, proposed methodology and experimental
setup, results and discussion, conclusion and future scope. Section II reviews existing research on plant disease
detection techniques and identifies the need for more accurate and efficient methods. Section III outlines the use of
CNN for disease detection in plants, explaining the architecture and training process. Section IV The experimental
setup describes the dataset used, data preprocessing techniques, and evaluation metrics employed. Section V The
results and discussion section presents the performance of the CNN model in terms of accuracy, precision, and recall,
along with a detailed analysis of the findings. Section VI summarizes the key contributions of the study, highlighting
the effectiveness of CNN for plant disease detection and its potential impact on agriculture. Finally, the future scope
suggests further research directions, including exploring transfer learning, real-time disease detection, and integrating
other advanced machine learning techniques. This research paper aims to advance the field of plant disease detection
and provide valuable insights for future studies in this area.

2. Literature Survey

A preliminary literature survey was conducted for plant disease detection in which existing research and studies related
to the detection and diagnosis of plant diseases were reviewed. The survey aimed to gather information on the methods,
techniques, and algorithms used by researchers in this field. Researchers have used various techniques, including
spectral data for hyperspectral imaging, machine learning algorithms like support vector machines (SVM) and deep
learning models like CNNs, and image processing techniques like segmentation and feature extraction. These
techniques were used to precisely diagnose and categorize illnesses by analysing spectral data or plant image analysis.

Saleem et al. (2019) [2] proposed plant disease detection and classification using deep learning methods. It was a
pioneering effort that used hyperspectral imaging to identify plant disease automatically. The experimental study
aimed to detect Tomato Spotted Wilt Virus (TSWV) infection in capsicum plants. After inoculation, a routine
automated system was developed to collect hyperspectral hypercubes of plant leaves in the VNIR and SWIR spectrum
3470 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478
Author name / Procedia Computer Science 00 (2019) 000–000 3

regions. Additionally, the technology was used to collect pertinent meta-data about the plants. After processing the
data, it was possible to see how effectively plant leaves could be automatically differentiated using machine learning
and image processing techniques. The results showed an accuracy of over 90%. They took advantage of the VNIR
(400–1000 nm) and SWIR (900–2500 nm) portions of the electromagnetic spectrum.

Durmuş et al. (2017) [3] proposed disease detection on the leaves of tomato plants by using deep learning. The dataset
of tomato leaf images was downloaded from the Plant Village Dataset, and the proposed research was based on the
six different types of tomato leaf diseases. The error-correcting output code was utilised to train a deep convolution
neural network model, which was employed for detection and classification. The classification results were computed
using parameters including accuracy, specificity, sensitivity, F-Score, and true negative rate. The trained model was
compared to the basis paper and had a 98.8% accuracy rate.

Sharma et al. (2020) [4] proposed a performance analysis of deep-learning CNN models for plant disease detection
using image segmentation. This study has demonstrated the viability of training a CNN model instead of whole images
using segmented and annotated images. Model performance on independent data increases from 42.3% to 98.6% when
the same CNN model was trained using segmented pictures (S-CNN) instead of full images (F-CNN). With 82% of
the test dataset displaying an increase in confidence, quantitative examination of self-classification confidence
demonstrated a considerable improvement. Before CNN model training, picture preprocessing has proven invaluable
in achieving excellent real-world performance.

Khirade et al. (2015) [5] proposed an approach to Plant disease detection using image processing. Various novel
approaches to quickly identify and categorise any illness are essential for the successful production of crops and for
ensuring their sustainability. Image Processing is one of the modern technologies that paved the path for this. This
study investigated the segmentation of the sick plant components utilising several techniques. This study also
examined the methods for feature extraction and classification and the classification of plant diseases. Plant diseases
can be effectively classified using SVMs, backpropagation methods, self-organizing feature maps, and other ANN
techniques. These techniques enable accurate diagnosis and classification of numerous plant diseases using image
processing tools.

Sood et al. (2020) [6] describe a hybrid system for detecting and classifying plant disease using qualitative texture
features analysis. In this work, the diseased affected capsicum piece is first subdivided, and then the features of the
infected region are extracted using feature extraction. The classification of the capsicum illnesses is completed using
an SVM classifier. In order to distinguish between healthy and diseased capsicum and its leaves, the proposed solution
is tested for the five illnesses anthracnose, bacterial spot, powdery mildew, Cercospora leaf spot, and grey leaf spot.
An SVM classifier can differentiate between healthy and sick capsicum with almost 100% accuracy. Using multi-
class classifiers, diseased capsicum can be further categorised based on the type of disease, either bacterial or fungal.
According to the suggested method, there are two types of healthy and diseased capsicums. This investigation is
limited to identifying five capsicum infections at this time.

Devi et al. (2019) [7] proposed IoT-enabled efficient detection and classification of plant diseases for agricultural
applications. The proposed IoT-enabled sickness detection and categorization system was implemented in the
openCV-python environment. Eighty images of the hill banana plant were used to test and show the process. Two
diseases—bunchy top of the banana and sigatoka leaf spot disease—are chosen for demonstration and evaluation.
During the image acquisition stage, the plant images were shrunk to 256 x 256, and the histogram was levelled using
a histogram equalisation technique. Once the segmentation process had been finished using the K-means clustering
method, the GLCM features were uploaded to the cloud. The experts gathered the information, looked through it, and
provided the farmers with solutions. RFC generated a classification accuracy of approximately 99% when compared
to the current k-NN classifier in terms of system performance.

Singh et al. (2021) [8] presented disease detection and classification of potato plant leaves using machine learning
methodologies. Utilising machine learning techniques, the proposed model was used to identify and categorise potato
leaves that were damaged and those that weren't. The entire procedure involved several processes, including image
collecting, image preprocessing, image segmentation, feature extraction, evaluation of the impacted region,
 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478 3471
4 Author name / Procedia Computer Science 00 (2019) 000–000

processing, training and testing data, and assessment through metrics. The Grey Level Co-occurrence Matrix was
utilised for feature extraction, the K-means approach was employed for image segmentation, and a multi-class support
vector machine algorithm with a linear kernel function was used to classify potato leaves. The proposed framework
was assessed using specific assessment criteria, such as precision, recall, F1-score, and accuracy. The proposed
model's F1-score is approximately 96.16%, its precision is roughly 96.12%, its recall is approximately 96.25%, and
its total accuracy is approximately 95.99%.

3. Proposed Methodology

Figure 1 shows the entire process of the proposed model. The three main steps of the technique are the data
preprocessing stage, model training using the training set, and prediction using the test set. The following subsections
cover the specifics of each phase.

Fig. 1. Overall Proposed Methodology.

3.1. Data Preprocessing Step

The data preprocessing step involved the normalization of images. Image normalisation is a preprocessing procedure
used in the dataset for deep-learning plant disease detection to guarantee that all images have uniform sizes and ranges
of pixel values. This stage is crucial for boosting the model's convergence during training and for enhancing its
generalisation abilities. Image normalisation frequently involves the following steps:
Rescaling: Images are scaled or resized to 224x224 pixels. Usually, this is done to ensure that all of the images are
3472 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478
Author name / Procedia Computer Science 00 (2019) 000–000 5

the same size so that feeding them to the neural network is simpler.
Pixel Value Normalisation: The image's pixel values are altered to fall between a range of values, frequently between
0 and 1 or -1 and 1. This normalisation aids in minimising the effects of variations in pixel intensity and enhances the
neural network's convergence during training.
Mean Subtraction: Each pixel has its mean pixel value from the entire dataset is subtracted. This process prevents
neural network activation functions from becoming saturated and centres the data distribution around zero.
Standardisation: The pixel values are further scaled by dividing the pixel values by the dataset's standard deviation.
The unit variance of the data is ensured using this procedure, which also speeds up neural network learning.

The dataset is divided into three sets: the training set, the validation set, and the test set once it has been normalised.

Training Set: Most of the dataset, known as the training set, is used to train the deep learning model. Through forward
and backward passes throughout the training phase. It is used to update the model's parameters.
Validation Set: Using the validation set, the validation performance of the model is assessed, and its hyperparameters
are modified. As it enables tracking of the model's performance on data it hasn't seen during training, it is crucial for
preventing overfitting.
Test Set: The test set is a distinct dataset that is only utilized following the conclusion of model training. It measures
the trained model's ultimate performance and ability to generalize to new data.

Data augmentation generates new images that are altered versions of our input data by applying image modification,
contrast adjustments, and image blurring without affecting the labels on the original photos. The training set images
were augmented for the model's robustness. The data augmentation steps are depicted in Figure 2.
Even the issue of over-adjustment is also decreased by the expansion of the entire “train”. To boost data, the “Keras
Image data Generator” class was used, and the following options produced the best results:

(a) Rotation range: 20

(b) Width and height shift ranges: 0.2
(c) Horizontal flip: True
(d) Vertical mode: True

Fig. 2. Data Augmentation Steps.

3.2. Proposed Model Architecture

The architecture of CNN, which enables the effective processing of image data, is depicted in Figure 3. A deep CNN
architecture is made up of numerous layers of various kinds. It is the most often used technique for sifting through
large datasets and extracting useful information. The multiple layers of the proposed CNN model architecture for plant
disease detection are intended to extract features from the input photos and classify them into several disease
categories. The architecture follows a sequential structure, with each layer performing specific operations to transform
the input data. The details of CNN model layers are described in brief.
 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478 3473
6 Author name / Procedia Computer Science 00 (2019) 000–000

Fig. 3 Convolution Neural Network (CNN) Architecture

Convolutional layers
Convolutional layers make up the first few layers of the CNN. Convolutional layers perform feature extraction by
applying filters to the input image. These filters learn spatial patterns and local features present in the image. From
one block to another, the number of filters gradually increases, but the size of the filter remains fixed at 3 x 3. There
are 32 filters in the first convolutional block, 64 in the second, and 128 in the third. The size of the feature maps was
lowered due to the use of pooling layers in each block, necessitating an increase in the number of filters.
Convolutional layers are particularly effective in capturing hierarchical representations of images, allowing the model
to learn discriminative features at different scales. Each convolutional layer applies a set of filters (kernels) to the
input image. Each filter convolves with the image, producing feature maps highlighting specific patterns. Non-linear
activation functions, such as ReLU (Rectified Linear Unit), are commonly applied after each convolutional layer to
introduce non-linearity and increase the model's ability to capture complex patterns.

Pooling Layer
A max pooling layer with a pool size of 2x2 is added after each convolutional layer. The feature maps created by the
convolutional layers are downsampled using pooling layers. Doing this makes it possible to keep key features while
reducing the feature maps' spatial dimensions. By condensing local data, pooling layers simplify computation and
increase the model's spatial invariance.
The most popular pooling process, max pooling, chooses the highest value possible within a predetermined pooling
frame. Pooling layers increases the model's resistance to changes in object location, scale, and rotation by
downsampling the feature maps.

Flattened Layer
The flattened layer creates a 1-dimensional vector from the output of the preceding levels. This flattening phase is
essential before transferring the data into a fully connected layer.

Fully Connected Layers

The CNN architecture's fully connected layers, which are found near the bottom, are in charge of categorising the
learned features into particular illness categories. Fully connected layers capture global relationships between features
and enable the model to make predictions based on the extracted representations. One or more fully connected layers
receive the output of the final pooling layer after it has been vectorized and flattened. Fully Connected layers provide
thorough feature combining and categorization since every neuron in the layer below is coupled to every neuron in
the layer above.
The first dense layer consists of 64 neurons with the ReLU activation function. It captures global relationships between
features. The second dense layer has 16 neurons with the ReLU activation function, allowing further feature
combinations. The 15 neurons in the last dense layer, which have a softmax activation function and provide a
probability distribution over the 15 illness classifications, are what make up this layer. The softmax activation function
ensures that the predicted probabilities sum up to 1.
3474 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478
Author name / Procedia Computer Science 00 (2019) 000–000 7

4. Dataset and implementation details

4.1 Dataset Creation

The 'Plant Village Dataset' is the source of all Potato, Tomato, and Pepper Bell pictures. The 20,639 leaf images in
the Plant Village dataset are categorised into 15 groups based on species and illness, as shown in Table 1. The leaf
pictures dataset includes both healthy and ill leaves that have been impacted by various biotic causes. Images of every
significant type of leaf disease are included in this dataset. These vegetables are part of the staple diet in many Indian
households. Thus, plant disease detection in such plants was important. We build the model using a subset of 70%
and test it using a subset of 30%. Examples of the "Train" folder images from the adopted distribution are included in
Table 2.

Table 1. Dataset Used for Classification

Species Class No. of Images
Pepper Bell Healthy 1478
Pepper Bell Bacterial Spot 997
Potato Early Blight 1000
Potato Late Blight 1000
Potato Healthy 152
Tomato Bacterial Spot 2172
Tomato Early Blight 1000
Tomato Late Blight 1909
Tomato Leaf Mold 952
Tomato Septoria leaf Spot 1771
Tomato Spider mites 1676
Tomato Target Spot 1404
Tomato Yellow Leaf Curl Virus 3209
Tomato Healthy 1591
Tomato Mosaic Virus 373

Table 2. Images from Dataset

Pepper Bell
Healthy Leaf

Pepper Bell
Bacterial Spot

Tomato Healthy
Leaf
 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478 3475
8 Author name / Procedia Computer Science 00 (2019) 000–000

Tomato Leaf
Mold

Tomato Early
Blight

Potato Healthy
Leaf

Potato Early
Blight

Potato Late
Blight

4.2 Implementation Details

The Adam optimizer, a stochastic gradient descent (SGD) variant that modifies the learning rate during the course of
training, is used to train the model. The Adam optimizer's default setting of a negligibly low number, such as 0.001,
is utilized as the learning rate. A loss function suitable for multi-class classification applications is used, categorical
cross-entropy. The training process lasts for 20 epochs, and to improve the model's capacity for generalization, data
augmentation is applied during training using the ImageDataGenerator object. Accuracy, which measures the
percentage of accurate predictions overall predictions, and categorical cross-entropy loss, which gauges the
discrepancy between the actual and predicted probability distributions, are the performance metrics considered for
evaluating model performance. The model's effectiveness is evaluated after training, and the final accuracy and loss
are given using testing data. Additionally, the trained model is stored in a h5 file for later usage. The hyperparameters
like batch size and learning rate were finetuned using both grid search and then coarse to fine search.

a b

Fig. 4. Plot of (a) Model Accuracy (b) Model Loss with Epochs.

The model is transformed into a format that can be used on mobile devices to deploy the trained model to produce a
mobile app. One typical method is to transform the model into the TensorFlow Lite format, which was created
3476 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478
Author name / Procedia Computer Science 00 (2019) 000–000 9

especially for embedded and mobile devices. On mobile devices, TensorFlow Lite offers effective inference and
supports a number of optimisations that both shrink the model's size and boost speed.
Using the cross-platform mobile app development framework Flutter, we will integrate the model into the mobile app
after converting it to TensorFlow Lite format. With the help of these frameworks, you can load and execute the
TensorFlow Lite model inside of a mobile application and use it to forecast the contents of pictures the app has taken
or that you've chosen from the device's photo library. The app can then show the outcomes of the predictions made
by the algorithm, such as classifying the type of plant disease from a plant photograph. The model would be integrated
into a mobile app using the Flutter mobile app development technology and delivered to a mobile app by converting
it to TensorFlow Lite format.

5. Results and Discussion

The suggested convolution neural network achieved an overall testing accuracy of 86.21% using data augmentation
and transfer learning. The improvement in accuracy and loss over time from epoch 13 to 20 is seen in Table III. The
accuracy improves, and the loss is significantly reduced with each epoch. The results have been reported for 20 epochs
due to system limitations. Although the loss value continues to drop significantly after iteration 10, there has been no
appreciable gain in our model's accuracy.

Table 3. Variation in Accuracy and Loss value with Epochs

Epoch Number Accuracy (%) Loss
Value
13 85.09 4.515
14 85.09 4.579
15 86.06 4.504
16 86.02 4.343
17 85.86 4.295
18 86.38 4.340
19 85.48 4.420
20 86.21 4.344

Furthermore, table 3 demonstrates that as more iterations are performed, the value of loss and accuracy gets closer to
being constant with different epochs.
Over 20 training epochs, the maximum validation accuracy of 74.94% was recorded, while the best training accuracy
of 86.21% was also noted. A validation accuracy of 86% on average has been found. This is a helpful measure of how
accurately the Deep Learning algorithm has classified various objects. A method of visualising and indicating the rate
of model convergence is provided by the plots of train and test accuracy and loss against the epochs in Fig. 4. It is
clear that the model has stabilised after about 20 epochs, and the metrics do not indicate that the model has much
improved in the last ten epochs. In Figure 5, The images are given as input to the trained model, and the predicted
labels match with the actual label. The outcomes demonstrate the model's effectiveness on the dataset and its
applicability for classifying leaf diseases.

The proposed CNN model architecture leverages the hierarchical and translation-invariant nature of CNNs to extract
meaningful features from plant disease images. The model can effectively detect and classify plant diseases by
combining convolutional, pooling, and fully connected layers. However, further experimentation and optimization
may be necessary to finetune the architecture for specific datasets and disease detection tasks.
1 Author name / Procedia Computer Science 00 (2019) 000–000
0 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478 3477

Fig. 5. Prediction of diseases using CNN model. The label category denotes the predicted disease of the respective plant

The performance of the Convolutional Neural Network (CNN) and K-Nearest Neighbors (KNN) on the Plant Village
dataset was evaluated. The accuracy achieved by both models was initially evaluated. The proposed Convolutional
Neural Network (CNN) achieved an accuracy of 86.21%, and K-Nearest Neighbors (KNN) achieved an accuracy of
82.5%. The CNN's higher performance appears to be significantly influenced by its capacity to capture complex
characteristics and patterns in the images of plant disease. The KNN classifier, in comparison, may perform poorly
due to its dependency on the selection of K and sensitivity to the distribution of the dataset.
Its applicability for challenging picture classification tasks, such as plant disease identification, is further supported
by the CNN's generalisation capabilities, which enables it to adapt to new and unexplored data. In conclusion, our
comparison analysis unequivocally shows that, in terms of accuracy and overall performance, the CNN-based
technique is superior to the KNN classifier for plant disease identification. This emphasises how crucial it is to use
deep learning methods, such as CNNs, for reliable and precise disease identification in agriculture.

In summary, Convolutional layers, pooling layers, and fully connected layers are used to extract and classify disease-
related information from input images in the proposed CNN model architecture, which, in turn, offers a framework
for precise disease detection in plants.

6. Conclusion and Future Scope

The majority of the Indian people continue to rely heavily on the agricultural industry, which is still one of the most
significant sectors. Therefore, spotting diseases in these crops is essential to expanding the economy. Therefore, this
research aims to discover and characterize distinct illnesses in the tomato, potato and pepper bell crops. Using a
Convolutional Neural Network model, the suggested method classifies tomato leaf diseases, potato leaf diseases, and
pepper bell leaf diseases from the Plant Village dataset. To categorize tomato, potato and pepper bell leaf illnesses
into different classes, a straightforward Convolutional Neural Network with a minimal number of layers was utilized
as the architecture.

Future Scope: Future work may also involve experimenting with the suggested model using various learning rates and
optimizers. It can also involve experiments with more modern architectures to improve the model's functionality on
the train set. As a result, farmers can utilize the aforementioned model as a decision-making tool to assist and support
them in recognizing diseases that can affect tomato, potato and pepper bell plants.

References

[1] Agricultural Staistics at a Glance, Department of Agriculture & Farmers welfare (2023) [Online] Available:
http://agricoop.nic.in/en/Agricultural_Statistics_at_a_Glance [Accessed August 07, 2023]
[2] Saleem, M.H., Potgieter, J. and Arif, K.M., 2019. Plant disease detection and classification by Deep Learning. Plants, 8(11), p.468.
3478 Tejaswini et al. / Procedia Computer Science 235 (2024) 3468–3478
Author name / Procedia Computer Science 00 (2019) 000–000 11
[3] We Durmuş, H., Güneş, E.O. and Kırcı, M., 2017, August. Disease detection on the leaves of the tomato plants by using Deep Learning. In 2017
6th international conference on agro-geoinformatics (pp. 1-5). IEEE.
[4] Sharma, P., Berwal, Y.P.S. and Ghai, W., 2020. Performance analysis of Deep Learning CNN models for disease detection in plants using
image segmentation. Information Processing in Agriculture, 7(4), pp.566-574.
[5] Khirade, S.D. and Patil, A.B., 2015, February. Plant disease detection using image processing. In 2015 International conference on computing
communication control and automation (pp. 768-771). IEEE.
[6] Sood, M. and Singh, P.K., 2020. Hybrid system for detection and classification of plant disease using qualitative texture features analysis.
Procedia Computer Science, 167, pp.1056-1065
[7] Devi, R.D., Nandhini, S.A., Hemalatha, R. and Radha, S., 2019, March. IoT enabled efficient detection and classification of plant diseases for
agricultural applications. In 2019 International Conference on Wireless Communications Signal Processing and Networking (WiSPNET) (pp. 447-
451). IEEE.
[8] Singh, A. and Kaur, H., 2021. Potato plant leaves disease detection and classification using machine learning methodologies. In IOP Conference
Series: Materials Science and Engineering (Vol. 1022, No. 1, p. 012121). IOP Publishing.
[9] Singh, A. and Kaur, H., 2021. Potato plant leaves disease detection and classification using machine learning methodologies. In IOP Conference
Series: Materials Science and Engineering (Vol. 1022, No. 1, p. 012121). IOP Publishing.
[10] Parikh, A., Raval, M.S., Parmar, C. and Chaudhary, S., 2016, October. Disease detection and severity estimation in cotton plant from
unconstrained images. In 2016 IEEE international conference on data science and advanced analytics (DSAA) (pp. 594-601). IEEE.