School of Computer Science Engineering and

Information Systems

Fall Semester 202425

Course Name: Soft Computing – J Component

Course Code: SWE1011

Review 3 Final Report

Project Title: Plant Leaf Disease Classification

Team Members :

1. Kaushik K 21MIS0332
2. Somasekhar N 21MIS0220
3. Naresh R 21MIS0354
1. Introduction :

Plant diseases pose significant threats to global food security and agricultural
sustainability. Early detection and accurate diagnosis of these diseases are essential
for effective disease management and ensuring crop yield and quality. Traditional
methods of disease detection often rely on visual inspection by trained experts,
which can be timeconsuming, subjective, and prone to human error. Recent
advancements in machine learning (ML) and deep learning (DL) techniques offer
promising solutions for automating the process of plant disease detection. ML
algorithms can learn patterns and features from large datasets of plant images, while
DL algorithms, particularly convolutional neural networks (CNNs), excel at
automatically extracting relevant features from raw image data, making them
wellsuited for imagebased tasks like disease detection. We aim to develop a system
that can accurately identify and classify diseases based on images of plant leaves,
enabling early detection and timely intervention to prevent widespread crop damage.
The proposed approach involves several key steps. First, highresolution images of
plant leaves are acquired using digital cameras or smartphones. These images are
then preprocessed to enhance their quality and remove any noise or distortions that
could affect the accuracy of disease detection. Next, features are extracted from the
preprocessed images, capturing important characteristics such as color, texture, and
shape of the leaf lesions. ML and DL models are trained using labeled datasets
consisting of images of healthy and diseased plant leaves. During the training phase,
the models learn to differentiate between healthy and diseased leaves and to classify
the specific type of disease present if applicable. Once trained, the models can be
deployed to analyze new leaf images and provide realtime feedback on the presence
and severity of diseases. By leveraging ML and DL techniques, our proposed system
aims to revolutionize the process of plant disease detection, enabling farmers to
monitor the health of their crops more efficiently and accurately. This has the
potential to significantly improve agricultural productivity, reduce crop losses, and
contribute to global food security in the face of increasing challenges posed by plant
diseases and climate change.

2. Problem Statement :

Develop a machine learning model capable of accurately classifying plant conditions

based on leaf images. The goal is to train a model that can automatically classify an
unseen image of a plant leaf into one of these three categories. Performance on the
test set will evaluate the model's realworld applicability in early detection of plant
diseases for improved agricultural outcomes.

3. Objectives:

• With the increasing prevalence of plant diseases and the growing complexity of
agricultural systems, there is a critical need for innovative technologies that can
streamline disease diagnosis and enable timely interventions.

• By harnessing the power of ML and DL, we aim to develop a system that not only
automates the detection of plant diseases but also provides farmers with
actionable insights to mitigate crop losses and optimize yield.
4. Dataset Description (Max. 100 words)

In our project we are using Kaggle dataset. The dataset comprises 1530 images
categorized into three labels: "Healthy", "Powdery", and "Rust", representing different
plant conditions. These images are divided into train, test, and validation sets,
facilitating the development and evaluation of machine learning models for plant
disease detection. Each image depicts the foliage of various plant species affected
by either healthy conditions or specific diseases like powdery mildew ("Powdery") or
rust fungus ("Rust"). This dataset serves as a valuable resource for researchers and
practitioners in the field of agriculture and computer vision, enabling the creation of
robust algorithms for automated plant disease diagnosis and management. There is
a total of 1530 images divided into train, test, and validation sets. The total dataset is
divided into 70/30 ratio of training and validation set.

Source:
https://www.kaggle.com/datasets/rashikrahmanpritom/plant-disease-
recognition-dataset/data

5. Fundamentals of ML/DL techniques used:

In this project, four machine learning and deep learning techniques were used to
classify plant diseases. These techniques include a Deep Neural Network (DNN), a
Convolutional Neural Network (CNN), and Transfer Learning using two pretrained
models: MobileNetV2 and VGG16.:

1. Deep Neural Network (DNN):

Architecture:
• The DNN is a basic fully connected neural network consisting of an input
layer, three hidden layers, and an output layer.
• Each of the hidden layers uses the ReLU activation function to introduce
nonlinearity, allowing the network to learn more complex patterns.
• Dropout regularization is used between layers to avoid overfitting by randomly
setting some neuron outputs to zero during training.
• The output layer consists of 3 neurons corresponding to the three possible
plant disease classes (e.g., Healthy, Powdery, Rust), and it uses the Softmax
activation function to compute the probabilities of each class.

Training:
• The DNN model is trained using the Adam optimizer, which is a popular
optimization algorithm combining the benefits of Adagrad and RMSProp.
• The categorical crossentropy loss function is used to measure the difference
between the predicted probabilities and the actual class labels for multiclass
classification.
• Batch normalization is applied to stabilize and speed up the training process
by reducing internal covariate shift.
2. Convolutional Neural Network (CNN):
Architecture:
• CNNs are specifically designed for processing gridlike data, such as images,
and are known for their ability to automatically learn spatial hierarchies.
• In this model, multiple convolutional layers are used to convolve the input
image with filters (kernels) to extract lowlevel features such as edges and
textures.
• After convolution, maxpooling layers are applied to reduce the spatial
dimensions of the feature maps, helping the network focus on the most
important features and reducing computational cost.
• The final output of the convolutional layers is flattened into a 1D vector, which
is passed through fully connected layers (Dense layers) for classification.

Activation Functions:
• ReLU (Rectified Linear Unit) is used as the activation function for all hidden
layers, as it helps the model learn nonlinear patterns and speeds up training
by avoiding issues like vanishing gradients.
• Softmax activation is used in the output layer to classify the image into one of
the three disease categories (Healthy, Powdery, Rust).

Training:
• The CNN is trained using the Adam optimizer and the categorical
crossentropy loss function.
• Dropout is applied to prevent overfitting and ensure the model generalizes
well to unseen data.

3. MobileNetV2 (Transfer Learning):

Architecture:
• MobileNetV2 is a lightweight deep learning model designed to perform well on
mobile and edge devices. It uses a novel architecture called depthwise
separable convolutions, which separates the convolution process into two
steps to reduce computation and improve efficiency.
• MobileNetV2 has fewer parameters compared to traditional CNNs, making it
ideal for edge devices, but it still maintains high accuracy on image
classification tasks.
• Global Average Pooling is applied to reduce the dimensionality of the output
feature maps before passing them to the final classification layer.

Transfer Learning:
• The MobileNetV2 model is pretrained on a largescale dataset like ImageNet,
which enables it to learn general features from millions of images.
• For this project, the model was finetuned by replacing the final classification
layer with one suited to the plant disease classification task (3 output classes).
• The model's weights are frozen during training except for the new
classification layers, allowing the model to leverage its learned features while
adjusting to the new dataset.
 Training:
• The model is trained with categorical crossentropy as the loss function and
the Adam optimizer.
• Transfer learning allows the model to generalize better on the plant disease
dataset, especially when labeled data is limited.

4. VGG16 (Transfer Learning):

Architecture:
• VGG16 is a deep CNN architecture that is known for its simplicity and depth.
It consists of 16 layers, with 13 convolutional layers and 3 fully connected
layers. The convolutional layers use small 3x3 filters, and the network utilizes
maxpooling after every few convolutional layers.
• VGG16 was designed for object recognition tasks and has been widely used
as a feature extractor in various image classification problems.
• Similar to MobileNetV2, Global Average Pooling is applied to reduce the
output dimensions of the convolutional layers before passing the result to the
dense layer for classification.

Transfer Learning:
• Pretrained weights from ImageNet are used to initialize the VGG16 model.
The model is then finetuned for the plant disease classification task by
adjusting the last fully connected layers to suit the specific number of classes
(Healthy, Powdery, Rust).
• By leveraging the learned features from ImageNet, the model performs better
with fewer training samples.

Training:
• The model is trained using the categorical crossentropy loss function and the
Adam optimizer.
• Transfer learning helps the VGG16 model achieve high accuracy with fewer
epochs, making it an effective method for solving image classification
problems, especially when data is scarce.

Summary of Techniques:

DNN: A fully connected neural network that learns hierarchical representations from
the raw input features.

CNN: A convolutional neural network that automatically extracts spatial features from
images, leading to high accuracy in image classification tasks.

MobileNetV2: A lightweight CNN that uses depthwise separable convolutions for

efficient processing, finetuned for plant disease classification using transfer learning.

VGG16: A deep CNN pretrained on ImageNet and finetuned for plant disease
classification, leveraging transfer learning for high accuracy with fewer data.
6. Methodology:

Our proposed approach involves several key steps aimed at building a robust and
scalable system for plant disease detection. The first step is the acquisition of high
resolution images of plant leaves using digital cameras or smartphones. These
images serve as the input data for subsequent processing and analysis. Next, the
acquired images undergo preprocessing techniques to enhance their quality and
remove any noise or distortions that could affect the accuracy of disease detection.
This preprocessing stage is crucial for ensuring the reliability of the subsequent
analysis steps. Following preprocessing, features are extracted from the
preprocessed images, capturing important characteristics such as color, texture, and
shape of the leaf lesions. These features serve as input to ML and DL models, which
are trained using labeled datasets consisting of images of healthy and diseased
plant leaves. During the training phase, the models learn to differentiate between
healthy and diseased leaves and classify the specific type of disease present if
applicable. This process involves iteratively adjusting the model parameters to
minimize prediction errors and improve overall performance. Once trained, the
models can be deployed to analyze new leaf images and provide realtime feedback
on the presence and severity of diseases. This enables farmers to take timely
corrective actions, such as applying targeted treatments or implementing preventive
measures, to mitigate the spread of diseases and minimize crop losses.

FlowChart:
7. Results:
performance metrics of models:

Model / Training Test Validation Model Size

Technique Accuracy Accuracy Accuracy
CNN Model 0.96 0.90 0.95 136.94 MB
Dense Neural 0.29 0.32 0.32 891.82 MB
Network
VGG16 0.98 0.88 0.89 129.71 MB
Pretrained
Model (Transfer
Learning)
MobileNetV2 0.99 0.97 0.96 192.72 MB
Screenshots:
Convolution Neural Network Model :
VGG16 Pretrained Model (Transfer Learning):
MobileNetV2 model:
Deep Neural Network Model:
8.Conclusion:

In conclusion, the integration of machine learning (ML) and deep learning (DL) techniques holds
immense potential for revolutionizing plant disease detection in agriculture. By leveraging ML
algorithms such as support vector machines (SVM), random forests, and DL architectures like
convolutional neural networks (CNNs), researchers can accurately classify and diagnose plant
diseases based on visual symptoms captured from images of plant leaves. These technologies
offer farmers a proactive approach to disease management by enabling early detection and
intervention, thus minimizing crop losses and improving yield quality. Moreover, the
accessibility of these tools through mobile applications and automated systems empowers
farmers, especially those in remote areas, to make informed decisions about crop health and
treatment strategies. Despite the progress made, challenges such as dataset availability, model
scalability, and realtime implementation remain. Collaborative efforts among researchers,
agronomists, and technologists are essential to address these challenges and refine ML and DL
models for practical deployment in agricultural settings. In the future, continued advancements
in ML and DL techniques, coupled with the integration of IoT sensors and drone technology,
could further enhance the efficiency and accuracy of plant disease detection systems. By
embracing innovation and interdisciplinary collaboration, the agricultural sector can harness
the full potential of ML and DL to ensure global food security and sustainable crop production.

9.References :

[1]Mohanty, S. P., Hughes, D. P., & Salathé, M. (2016). Using Deep Learning for ImageBased Plant
Disease Detection. Frontiers in Plant Science, 7, 1419. DOI: 10.3389/fpls.2016.01419

[2]Ferentinos, K. P. (2018). Deep Learning Models for Plant Disease Detection and Diagnosis.
Computers and Electronics in Agriculture, 145, 311318. DOI: 10.1016/j.compag.2018.01.009

[3]Barbedo, J. G. A. (2019). A Review on the Use of Convolutional Neural Networks in

Agriculture. Computers and Electronics in Agriculture, 147, 142152. DOI:
10.1016/j.compag.2018.02.032

[4]Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., & Stefanovic, D. (2016). Deep Learning
for Plant Diseases: Detection and Diagnosis. Computers in Industry, 100, 103111. DOI:
10.1016/j.compind.2018.01.007

[5]Sharma, A., & Kaur, P. (2020). A Review on Plant Leaf Diseases Detection Using Image
Processing Techniques. Computers and Electronics in Agriculture, 169, 105220. DOI:
10.1016/j.compag.2019.105220

[6]Garg, G., Kaushik, P., & Kaushik, P. (2021). A Comprehensive Survey of Plant Leaf Disease
Detection Techniques Using Machine Learning and Deep Learning Paradigms. Sustainable
Computing: Informatics and Systems, 29, 100481. DOI: 10.1016/j.suscom.2021.100481

[7]Nanni, L., Brahnam, S., & Ghidoni, S. (2021). Deep Learning Techniques for Plant Disease
Detection Using Leaf Images: A Comprehensive Review. Applied Sciences, 11

[8], 3644. DOI: 10.3390/app11083644 [8]Zou, X., Luo, Y., & Liu, K. (2021). A Review of Deep
Learning Methods for Plant Diseases Recognition. Plant Methods, 17(1), 103. DOI:
10.1186/s13007021007970