Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Crop Disease Detection Using

Deep Learning Models
Aryan Chaudhary1 Mohit Gupta2 Upasana Tiwari3
AIT-CSE AIT-CSE Assistant Professor
Chandigarh University Chandigarh University Chandigarh University
Punjab,India Punjab,India Punjab, India

Abstract:- Detecting plant diseases during the growth of The neuron architecture in the human brain serves as
plants is a critical challenge in agriculture, as late inspiration for deep learning techniques. These methods use
detection can lead to reduced crop yields and lower Artificial Neural Network and their discrepancies, i.e.,
profits for farmers. To tackle this issue, researchers have Convolutional neural network and Recurrent Neural
developed advanced frameworks based on Neural Networks, to find hidden frameworks in the data. Compared
Networks[1]. However, many of these methods suffer to machine learning techniques, deep-learning methods have
from limited prediction accuracy or require a vast two major advantages [3]. To begin with, they do so
number of input variables. This project comprises of automatically, doing away with the requirement for a
CNN and LSTM models, the CNN component of the completely separate feature extraction module. Second,
project has demonstrated remarkable accuracy, processing large datasets with numerous dimensions takes
achieving a 98.4% success rate in identifying plant less time when using Deep Learning techniques. Deep
diseases from static images. Learning methods are consequently suggested. Due to their
efficiency with image data, DL techniques like CNN and
Keywords:- CNN Architecture, VGG Architecture, Fully LSTM are frequently used in applications involving
Connected Layers, VGG-19, Neural Networks, CNN, LSTM, computer vision.
Convolutional Layers.

I. INTRODUCTION

Agriculture forms the foundational pillar of the Indian

economy. Agriculture accounts for 10% of total GDP, with
India accounting for 16%. In India, the agricultural sector
either directly or indirectly supports nearly 70% of the
population [2]. Therefore, the production of high-quality,
disease-free crops are crucial for the development of the
nation's economy. Similar to humans, different ripening
stages or growth stages of plants are identified to different
diseases. As a result, the overall production of crops and
consequently the farmer's net profit are negatively impacted.
The initial diagnosis of plant viruses is required to address
this problem. Plant disease is manually detected by either
farmers or agricultural scientists. But doing this requires a
lot of effort and time.

These cutting-edge systems employ a wide range of

training variables. As a result, the training and
prognostication times of these processes are extremely long,
or they necessitate the use of a machine with greater
computational power. Figure-1 shows the images of various
plant leaf diseases. This research proposes a novel crop
disease detection system that leverages the power of deep
learning, combining CNNs for image analysis and LSTM
networks for sequential data processing. The integration of
CNNs and LSTMs allows us to take advantage of their
respective strengths to create a more robust and accurate Fig 1 Various Types of Plant Leaf Diseases
detection model. This, in turn, substantially reduces the
number of input variables, resulting in a reduction in
training and detection accuracy.

IJISRT23DEC614 www.ijisrt.com 1189

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
II. LITERATURE REVIEW architectures such as AlexNet, GoogLeNet, Inception, and
ResNet are explored in the context of plant disease
 Survey on Crop Pest Detection using Deep Learning and detection. The study highlights that deep learning
Machine Learning Approaches [4] approaches have become increasingly popular in recent
In this research paper by M. Chithambarathanu and M. years, but identifies certain research gaps. Most existing
K. Jeyakumar, The authors explore several methods for work primarily relies on datasets from PlantVillage, which
applying deep learning and machine learning techniques to predominantly contain images from controlled laboratory
identify plant diseases and crop pests. This paper includes environments with simple backgrounds. The paper suggests
the analysis of various existing methods such as DT the need for incorporating images from field environments
(Decision Tree), SVM(Support Vector Machine), NB(Naïve and emphasizes the importance of adapting deep learning
Bayes). These techniques are based on machine learning. models to account for changes in the severity of plant
The authors also discover that models based on deep diseases over the entire life cycle of plants.
learning perform better than those based on conventional
machine learning methods.  Crop Leaf Disease Detection and Classification using
Machine Learning and Deep Learning Algorithms by
The authors point out that there is still space for Visual Symptoms [7]
improvement in the forecast accuracy of present methods, The research paper reviews the use of machine
which led to the creation of a deep learning/machine learning and deep learning algorithms for crop leaf disease
learning hybrid approach. Better categorization outcomes detection and classification based on visual symptoms. It
and enhanced illness prediction performance are the goals of highlights that deep learning techniques outperform machine
this suggested methodology. For even greater improvement, learning techniques in accuracy, especially when using
the article also recommends integrating metaheuristic modified CNNs, optimized deep learning models, and
techniques. Additionally, the study predicts the integration transfer learning. Notably, a multi-channel model achieved
of mobile terminal processors and other Internet-based the highest accuracy of 99.5% in deep learning, while SVM
technologies in the future to allow real-time monitoring and with a linear kernel reached 99% in machine learning. The
pest identification in grain storage warehouses, such as the paper discusses the five key steps in developing a crop
Agricultural Internet of Things (IoT). This connection is disease identification system: image acquisition, pre-
considered a first step towards modernising and intelligently processing, segmentation, feature extraction, and
transforming agriculture. classification. It emphasizes the need for more research and
datasets to enhance disease detection in large-scale crops,
 Machine Learning for Detection and Prediction of Crop underlining the importance of integrating computer vision
Diseases and Pests: A Comprehensive Survey [5] and machine learning into agricultural automation
The research paper authored by Tiago Domingues, technologies such as UAVs and smartphones.
Tomás Brandão, and João C. Ferreira provides a
comprehensive survey for the administration machine  An Advanced Deep Learning Models-based Plant
learning (ML) approaches as the observation and prognostic Disease Detection [8]
of crop infection and pests. It highlights the importance of The research paper provides a comprehensive review
long-term datasets encompass weather, diseases, and pests’ of recent advancements in using Machine Learning (ML)
data for accurate forecasting using time-series ML models and Deep Learning (DL) techniques for plant disease
like Recurrent Neural Networks (RNN) and the relevance of detection. It highlights the significance of gathering diverse
NDVI measurements for crop development insights. images from different plant growth stages, seasons, and
regions to enhance the model's robustness and
Additionally, the paper talk about the effectiveness of generalization. The paper emphasizes the importance of
deep learning and Convolutional Neural Network (CNN) incorporating meteorological and plant health data for
models in the observation and categorization of diseases and efficient disease identification and prevention. It also
pests in crop images, emphasizing the challenges associated suggests the use of unsupervised learning and knowledge
with data scarcity in real-life conditions. The study also from human visual cognition to improve DL model training.
underscores the need for further research to explore the The study acknowledges the challenges, such as data
combination of different data modalities in pest and disease availability and distinguishing healthy plants from diseased
forecasting. In summary, the paper serves as a valuable ones, while showcasing the substantial improvements in
overview of ML appeal in crop disease and pest plant disease identification achieved through DL and ML
management, highlighting the potential of various ML methods. Overall, this research contributes valuable insights
techniques while identifying areas for future research and to both researchers and industry professionals in advancing
development in this field. plant disease detection and prevention.

 Deep CNN Models in Plant Disease Identification [6]  Plant Disease Prediction using Hybrid Model [9]
The research paper discusses the application of deep The research paper discusses the effectiveness of a
learning models, particularly Convolutional Neural hybrid model for plant disease prediction, which combines
Networks (CNNs), for the identification of plant diseases. machine learning, deep learning, and image processing
The reseach paper is proposed by Harjeet Kauri, Deepak techniques to accurately identify diseases in plants like
Prashar and Vipul Kumar Various. Well-known CNN pepper, tomato, and potato. The hybrid model's ability to

IJISRT23DEC614 www.ijisrt.com 1190

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
capture complex relationships between plant features and hyperparameters in deep learning and identifies key
symptoms leads to precise disease diagnosis, and its real- challenges and research gaps in the field of plant disease
time predictions can help farmers take timely preventive detection, underscoring the need for cost-effective and
measures, improving agricultural practices and crop yields. robust automated solutions.
The study highlights the potential benefits of the CNN-
LSTM model for plant disease prediction, emphasizing that  A Review of Plant Disease Detection and Classification
its performance depends on factors like training data quality Methods" [12]
and model architecture, with a reported best accuracy of This research paper provides an extensive review of
0.9325. Overall, this research offers a promising solution to recent advances in the domain of crop leaf disease detection
enhance food security and agricultural productivity by and classification using image processing, machine learning,
effectively managing and preventing plant diseases. and deep learning techniques. The authors highlight the
significance of accurate disease detection for improving
 Plant Disease Detection using CNN" [10] agricultural yields sustainably. The paper surveys various
The research paper by Nishant Shelar, Suraj Shinde, methodologies, focusing on dataset characteristics, the
Shubham Sawant, Shreyash Dhumal, and Kausar Fakir number of images and classes, algorithms employed, and the
presents a deep learning model for plant disease detection performance of convolutional neural network (CNN)
using Convolutional Neural Networks (CNNs). They models. Additionally, it offers insights into the suitability of
achieved an impressive accuracy rate of 95.6% by these algorithms for deployment in different settings such as
employing early stopping during training over 50 epochs. standard systems, mobile/embedded devices, drones, robots,
The paper demonstrates the successful classification of 38 and unmanned aerial vehicles. The paper also discusses
different plant diseases across 13 plant species, including performance metrics, identifies limitations, and suggests
tomato, strawberry, soybean, raspberry, potato, corn, and areas for future research to enhance real-time automated
others. The authors also visualized the training and crop leaf disease detection systems.
validation accuracy and showcased the model's capability to
distinguish between healthy and diseased plant leaves for  Plant Disease Detection using Image Processing and
various plant species. The research showcases the Machine Learning Algorithm [13]
effectiveness of their approach, including the use of the The research paper presents a novel approach for plant
VGG-19 model and the deployment of the model on an disease detection using image processing and machine
Android app, with ongoing efforts to further improve learning techniques. The study utilizes MATLAB to process
accuracy in both the app and the model. and segment diseased leaf images into clusters based on
their diseases. It calculates the functionality of co-
 Systematic Study on Deep Learning-based Plant occurrence after thresholding the RGB components of the
Diseasedetection or Classification" [11] images and employs K-Means clustering for classification.
The research paper by Sunil, Jaidhar, and Patil The system relies on image processing, K-Means, and
addresses the critical issue of plant disease detection, neural networks to predict diseases in 25 leaves belonging to
emphasizing the economic impact of such diseases on 5 different categories. The results demonstrate an overall
agricultural production. Traditional methods relying on accuracy of approximately 89.8% for pomegranate leaf
domain experts are limited by their availability, travel costs, disease detection and 91% for potato leaf disease detection,
and consultation fees. To overcome these challenges, the showcasing the effectiveness of the proposed hybrid
paper conducts a systematic study that reviews 160 research algorithms and segmentation techniques.
works related to plant disease detection, focusing on Deep
Learning-based and Machine Learning-based approaches. In conclusion, the research introduces a robust
The study categorizes these approaches into single network approach for automatic disease segmentation and
models, hybrid models, and real-time detection methods, classification, leveraging K-Means clustering and neural
covering 50 different plant leaf disease datasets. networks.
Additionally, the research highlights the significance of

III. LITERATURE SURVEY OF THE PAPERS

Table 1 Literature Survey of the Papers

Year and
Article/ Author Technique Dataset Source Evaluation Parameter
Citation
Random Forest, Accuracy using CNN 94.96%
M. Chithambarathanu and M.
2023 SVM, Decision Google Scholar Accuracy using DCNN 95.28%
K. Jeyakumar
Tree, CNN. Accuracy using SVM 92.65%
F1-Score (0.93),
SVM, Random
Tiago Dominguez, Tomas Accuracy using SVM 94.6%
2023 Forest, Google Scholar
Brando, and Joao C. Ferreira Accuracy using RF 95.5%
VGG-16.
Accuracy using VGG-16 98%
Harjeet Kauri, Deepak Prashar CNN, Random Accuracy using CNN 92.8%
2023 Data Mendeley
and Vipul Kumar Forest, Accuracy using RF 94.56%

IJISRT23DEC614 www.ijisrt.com 1191

 Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
RestNet Accuracy using ResNet 97.28%
Pallepati Vasavi, Arumugam
Digitial cameras, Accuracy using KNN 99%,
2023 Punitha and Venkat Narayana SVM, KNN
drones, UAV Accuracy using SVM 99.5%
Rao[7]
Muhammad Shoaib, Babar Complied dataset
Accuracy using CNN 98%
2023 Shah, Shanker al Sapagh, CNN, SVM from Google
Accuracy using SVM 94%
Akhtar Ali, And Asad Ullah Scholar, Kaggle
Mr. Gopinath V, Ilakiya V,
Kaggle , Accuracy 93.2%
2023 Nandhini R, Monikasri B, CNN and LSTM
Mendeley Data
Shalene V
Nishant Shelar, Suraj Shinde,
2022 Shubham Sawant and Shreyansh CNN Data Mendeley Accuracy 95.6%
Dheyal
C.K. Sunil, C.D. Jailandar And R-CNN, Inception Accuracy using R-CNN 99.7%
2023 Own Dataset
Nagamma Patil V3 Accuracy using Inception v3 99.4%
Nauman Qadaeer, Thabit Wheat Fields Val Accuracy 93%
2023 Sabbah and Muhammad attique CNN From Kotli Precison-0.94
khan Kashmir F1-Score-0.93,Recall-0.93
S. Nandhini AlexNet, Data Mendley,
2022 Accuracy 99.35%
and K. Ashokkumar GoogleNet, Kaggle

IV. CONCLUSION FROM REVIEWING THE PAPERS includes image classification and feature extraction. Growth
of pre-trained models are given below.
The key accuracy result among these Research Papers
is that Support Vector Machine (SVM) and KNN 99% and
99.5% respectively [7]. This accuracy level is exceptionally
high and indicates that the model was highly effective in
correctly identifying and classifying different plant diseases
within the dataset. The only drawback that occurred while
researching into that paper is that its dataset.

The primary constraints often encountered in crop

disease detection systems relying on visual symptom
analysis include a notable deficiency of extensive and
publicly accessible datasets. While Plant Village stands as a
valuable resource, it remains the solitary widely available
dataset generated within controlled environments. Some
researchers have crafted their proprietary datasets; however,
they are typically unwilling to share access with the broader
community for comparative analysis. Moreover, it's
important to note that the Plant Village dataset may not
encompass imagery of commercial crops such as chili,
which exhibit a diverse array of diseases.

The accuracy level achieved in this research is

exceptionally high as the dataset available on the web has
not more than 50k images, plus the model is no longer
focusing on the sequential learning of the crop disease
detection. So the top most priority is to build and implement
a model which can give high accuracy in predicting the Fig 2 Summary of the Evolution of Various Deep Learning
crop’s disease as well as the disease’s progression analysis. Models from 2012 Until Now
Although in the market various projects have been V. DOMAIN KNOWLEDGE
built with the motive to achieve high accuracy in the crop
disease detection but they have been trained on the small A. Overview of Proposed System
size dataset due to the availability and the quality of the
dataset. The dataset which we have complied compresses of  Problem Definition:
68.9k images that are labelled with the disease name and The proposed project aims to develop a Crop Disease
also categorised with two parameters i.e. Healthy and Detection System by using CNN networks for image
Unhealthy. Moreover the model which we have built is classification and feature extraction and LSTM networks for
using latest pre-trained model VGG-19 whose expertise disease progression analysis. This system will integrate

IJISRT23DEC614 www.ijisrt.com 1192

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
advanced dl techniques which are used to foresee and  Model Selection:
classify the crop disease based on the symptoms captured
through images from the available dataset. Using CNNs to  CNNs are the most popular choice for image
extract relevant information from pictures and the ability of classification tasks, so a CNN-based model is a good
LSTMs to simulate temporal relationships in the course of starting point. Some popular CNN architectures include
disease, the system will provide farmers with timely and VGG19, ResNet, and MobileNet.
precise recommendations for disease management,  Further the selection of deep learning model is the main
ultimately enhancing crop yield and food security task for the image classification. Here we are using
VGG-19 for the implementation of the model.
 Problem Architecture:  LSTM networks can also be used for plant disease
The Proposed system comprises of using some basic detection, especially for tasks such as disease
deep learning algorithm like CNN and LSTM. Although for progression analysis.
the implementation of the crop disease detection system we
have chosen the pre trained model VGG-19. Moreover the  Data Splitting:
VGG-19 is the latest model available on the web, its Splitting the dataset and categorizing it into training,
compatibility for the image’s feature extraction is strong. validation and testing sets has already been done and
uploaded on the Kaggle Datasets.
Hence the proposed system will use VGG-19 for the
classification or categorizing them into the Healthy and  Transfer Learning:
Diseased plant crop. The system will analyze linguistic cues Once a pre trained model is selected, it needs to be
and extrinsic information to determine the plant’s disease. applied on the preprocessed dataset. This involves feeding
This project can help in early detection and management of the images to the model and allowing it to learn the features
crop diseases, which is crucial for agriculture. that are associated with different crop diseases. The model
will then be able to use these obtained features from the
 The Proposed System will Include these Steps: images to classify the new images into category like healthy
or diseased.
 Data collection:
 Image’s Feature Extraction using Convolutional Neural
 Accumulate a comprehensive dataset with images Network(CNN):
containing healthy and diseased plant crops. Ensure that Executing the CNN model in order to enable it to
the dataset is well-labeled with information about the extract features from the input photos. Several layers are
type of disease and the plant crop’s name. built into the CNN's architecture, including:
 Preprocessing of the images i.e. resizing them to a
particular uniform size and normalizing the pixel values.  Multiple Convolutional layers
Augmenting the dataset if needed (e.g., rotation,  Pooling Layers
flipping, brightness adjustments).  Fully Connected Layers
 Resizing the images to a consistent size (e.g., 224x224)
for compatibility with VGG19.  LSTM for Sequence Learning:

 Convert the feature vectors extracted by the CNN into

sequences.
 Implement the LSTM network to learn the temporal
dependencies in the feature sequences. This is useful for
capturing how diseases evolve over time.

 Training the Model:

 Train the model (CNN & LSTM) on the training dataset.

Use an appropriate loss function for classification, such
as categorical cross-entropy.
 Monitor training with validation data to prevent
overfitting. Apply techniques like dropout, batch
normalization, and early stopping (if needed).

 Evaluating the model:

Analyze the model's performance with measures such
as these on the testing dataset:

 Accuracy (Both Training and validation)

Fig 3 Samples Images from the Dataset which are  Precision
Preprocessed  Recall

IJISRT23DEC614 www.ijisrt.com 1193

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Hyperparameters Tunning: that are often difficult or impossible to detect with
Fine-tune hyperparameters of both the CNN and traditional methods.
LSTM models, such as:
Convolutional layers play a crucial role in
 Batch Size Convolutional Neural Networks (CNNs). They operate by
 Number of layers/units in LSTM utilizing a collection of filters to process the input image.
These filters, which are applied to different areas of the
VI. METHODOLOGY picture, are tiny weight matrices. Convolutional layer output
is a collection of feature maps, or more specifically, a
A. CNN (Convolutional Neural Network)- collection of pictures that highlight various aspects of the
An artificial neural network (ANN) with specific input image.
capabilities for processing and analysing visual information
is called a convolutional neural network (CNN). Tasks By lowering the spatial resolution of these feature
requiring object identification, categorization, and picture maps, pooling layers help to lower the network's
recognition are especially well-suited for them. CNNs excel computational cost and lessen the chance of overfitting.
at capturing the spatial relationships between pixels in an Fully connected layers, in essence, resemble the layers
image, allowing them to learn and identify intricate patterns commonly found in traditional neural networks. They take
the output of the pooling layers and connect them to a layer
of neurons that produces the final output of the network.

Fig 4 Basic Architecture of CNN Model based on Plant Leaf.

Convolutional Neural Networks (CNNs) undergo able to learn temporal features from sequences of images.
supervised learning, wherein a dataset of labelled images is This is useful for tasks such as disease progression analysis.
provided to the network. Through this process, the network For example, a CNN can be used to track the changes in the
gains the ability to associate image features with their size and shape of a diseased leaf over time.
corresponding labels. Once trained, the network uses its
learned features to predict new image classifications.  CNNs are Typically used for Plant Disease Detection in
these ways:
In several image-related tasks such as object
recognition, semantic segmentation, and image  Using photos of plant illnesses identified with the
classification, CNNs have continuously shown outstanding associated plant disease to train the CNN model. CNN
performance. They find applications beyond image tasks, gains the ability to identify various plant diseases by
extending to domains like natural language processing, identifying spatial and temporal features in the images.
machine translation, and video analysis.  Once the model is trained on the CNN, it is ready to
classify the diseases in the crops.
The suitability of CNNs for plant disease detection is  The CNN gives the probability score for every crop
attributed to their capacity to discern spatial features within disease as an output, and the class that gets the highest
images. The identification of spatial features is crucial in probability score will get predicted and marked as the
plant disease detection, as they enable the recognition of correct class.
patterns within the distribution of affected plant tissue.
B. VGG-19-
For instance, a CNN can be trained to recognize VGG19 is a CNN architecture initially crafted by
distinct patterns like the circular lesions characteristic of researchers from the University of Oxford, renowned for its
fungal leaf spot diseases. Additionally, the network can robust performance in image classification tasks. While it
learn to distinguish the irregular shape and texture offers impressive results, it demands significant
associated with bacterial leaf blight diseases. CNNs are also computational resources for training. Nonetheless, a

IJISRT23DEC614 www.ijisrt.com 1194

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
multitude of pre-trained VGG19 models are accessible  Block 1:
online, making it an ideal candidate for transfer learning. In
utilizing VGG19 for plant disease detection, the primary  Two 2D convolutional layers with "same" padding and
step entails removing the pre-trained model's classifier layer. 64 filters of 3x3.
Subsequently, a novel classifier layer is appended, equipped  An additional layer called max-pooling with a 2x2
with the required number of classes (e.g., healthy and window and a 2x2 stride comes next..
diseased). This adapted model is then fine-tuned using a
plant disease dataset. Once training is complete, the model  Block 2:
becomes proficient at categorizing new plant images as Two 3x3 2D convolutional layers with 128 filters per
either healthy or diseased. Each class receives a probability and "same" padding make up this block. The max-pooling
score; the projected label belongs to the class with the layer that follows has a 2x2 window and a 2x2 stride. Two
highest score. 3x3 2D convolutional layers with 128 filters per and "same"
padding make up this block. The layer beyond that is a max-
It belongs to the VGG model family, celebrated for pooling layer with a window and stride of two by two
their straightforward yet effective approach to image dimensions.
classification. With 19 layers total—16 convolutional layers
and 3 fully linked layers—VGG19 in particular stands out  Block 3:
for its depth. K Simonyan and A. Zisserman’s Publication, In this block, there are four 2D convolutional layers,
"Very Deep Convolutional Networks for Large-Scale Image each with 256 filters of size 3x3, employing 'same' padding.
Recognition," introduced the concept. In the year 2014. Subsequently, a max-pooling layer with a 2x2 window and a
stride of 2x2 is applied.
 Here's a Detailed Overview of the Architecture of
VGG19:  Block 4:
This block consists of four 2D convolutional layers,
 Input Layer: each employing 512 filters of size 3x3 and 'same' padding.
An RGB picture with a fixed size of 224 by 224 pixels After the convolutional layers, there is a max-pooling layer
is inputted into VGG19. Red, Green, and Blue are the three with a 2x2 window and a stride of 2x2.
colour channels of the input picture.
 Block 5:
 Convolutional Layers (Blocks): In this block, there are four 2D convolutional layers
VGG19 is composed of five series of convolutional with 512 filters of size 3x3, utilizing 'same' padding.
blocks, each of which is accompanied by a subsequent max- Following these convolutional layers, a max-pooling layer is
pooling layer. Multiple convolutional layers precede applied with a 2x2 window and a stride of 2x2.
Rectified Linear Unit (ReLU) activation functions inside
each block. The quantity of filters progressively grows as
you delve deeper into the network.

Fig 4 VGG-19 Architecture

 Fully Connected Layers:  The third fully connected layer serves as the output layer
and features 1000 neurons, representing the 1000 classes
 VGG19's architecture includes three fully connected present in the ImageNet dataset, which was the dataset
layers following the convolutional blocks. initially used to train VGG19..
 The first fully connected layer comprises 4096 neurons  It typically uses a softmax activation function to produce
with a ReLU activation function. class probabilities.
 The second fully connected layer also consists of 4096
neurons with a ReLU activation function.

IJISRT23DEC614 www.ijisrt.com 1195

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Output: time. The network could then use these features to classify
The VGG19 network produces a probability new plant images as healthy or diseased.
distribution across 1000 classes, rendering it well-suited for
image classification tasks. VGG19 is renowned for its  Disease Progression Analysis:
consistent architectural design, featuring a repeated LSTM networks can also be used to predict the
arrangement of convolutional and max-pooling layers. progression of plant diseases. This can be done by training
an LSTM network on a dataset of plant images that are
This uniform and deep structure enabled it to achieve labelled with the corresponding plant disease and the stage
cutting-edge performance in image classification tasks when of the disease. The LSTM network would acquire the ability
it was originally introduced. to identify temporal characteristics in the pictures that
correspond to the illness's progression. The network could
then be used to predict the stage of the disease for new plant
images.

VII. FUTURE SCOPE OF THE PROJECT

 Deploy the trained model to a server or cloud service,

where it can be accessed by users for real-time or batch
processing of images.
 Create a user-friendly interface for users to upload
images and receive disease detection results.
 Remote Sensing and Drones, Integration with remote
sensing technologies and drones can enable large-scale,
automated monitoring of agricultural fields. Drones with
cameras are able to take pictures, which the CNN-LSTM
model processes to identify illnesses.
 The project's future scope may involve fine-tuning the
LSTM model to improve its accuracy and consistency.
Continued research in LSTM architecture,
hyperparameter tuning, and larger training datasets can
contribute to enhanced performance.

VIII. CONCLUSION

In conclusion, A major advancement in contemporary

Fig 5 Model Overview for the Proposed System agriculture and crop management is the creation of a crop
disease detection system employing Convolutional Neural
C. Long Short Term Memory (LSTM)- Networks (CNNs) and Long Short-Term Memory (LSTM)
Plant disease identification is a perfect application for networks.. This project offers a promising solution to the
LSTM networks, a subtype of recurrent neural networks pressing issue of early disease detection and monitoring in
(RNNs), because of its capacity to identify temporal plants, with far-reaching implications for crop yields,
relationships in sequential data. This is important for plant sustainable farming practices, and food security.
disease detection because it allows the network to identify
patterns in the distribution of disease symptoms over time. The project that we had built comprises of CNN and
LSTM models, the CNN component of the project has
For example, an LSTM network can be used to track demonstrated remarkable accuracy, achieving a 98.4%
the changes in the size and shape of a diseased leaf over success rate in identifying plant diseases from static images.
time. This information can then be used to classify the plant This level of accuracy is a testament to the effectiveness of
disease or to predict the progression of the disease. deep learning in image classification tasks.

 LSTM Networks can be used in Crop Disease Detection Although the LSTM model has achieved a
Systems in the following ways: comparatively lower accuracy of 70%, its role in analyzing
sequential data, such as disease progression, is crucial. The
 Classifying the Diseases: LSTM provides valuable insights into the temporal aspect of
LSTM networks can be used to classify plant diseases disease development, enhancing the overall system's ability
by analysing sequences of plant images. For example, A to understand disease dynamics.
collection of crop photos annotated with the corresponding
crop disease might be used to train an LSTM network. An
example of a temporal characteristic that the LSTM network
would be trained to extract from the photos would be the
variations in the size and form of illness symptoms over

IJISRT23DEC614 www.ijisrt.com 1196

Volume 8, Issue 12, December – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES

[1]. "Survey on crop pest detection using deep learning

and machine learning approaches" by M.
Chithambarathanu1, M. K. Jeyakumar published in
Springler in April 2023.
[2]. "Machine Learning for Detection and Prediction of
Crop Diseases and Pests: A Comprehensive Survey"
by Domingues Tiago, Tomad Brando and C. Ferreria
published in MDPI in August 2023.
[3]. "Deep CNN models in plant Disease Identification"
by Harjeet Kauri, Deepak Prashar and Vipul Kumar
published in Mili Publications in March 2023.
[4]. “Crop leaf disease detection and classification using
machine learning and deep learning algorithms by
visual symptoms: a review” by P. Vasavi, Arumugam
P. and V. Narayana Rao published in IJECE in April
2023.
[5]. "An advanced deep learning models-based plant
disease detection" by Muhammad Shoaib, B.
Shah, S. Sapagh, Akhtar Ali, and A. Ullah, published
in Frontiers in March 2023.
[6]. "Plant Disease Prediction Using Hybrid Model" by
Mr. Gopinath Vinketha, Ilakiya Venkat, Nandhini
Rethinan, Monikasri Bhaduri, Shalene V published in
IJERT in June 2023.
[7]. "Plant Disease Detection Using Cnn" by Nishant
Shelar, S. Shinde, S. Sawant and Shreyansh Dheyal
in ITM conference in May 2022.
[8]. "Systematic study on deep learning-based plant
diseasedetection or classification" by
C. Karan. Sunil, C. D. Jaidhar and Nagamma Patil
published in Mili Publications in June 2023.
[9]. "A review of plant disease detection and
classification methods" by Mohan M, M. Akila and
Prasath B published in Tandf in August 2023.
[10]. "Plant Disease Detection using image processing
and machine learning algorithm" by Rajeni Deepika
sai chowdhry, Yogeshwari V, Varsha and M.R Banu
Priya published in Research Gate in January 2023.
[11]. “A Convolutional Neural Network Model for Wheat
Crop Disease Prediction” by Nauman Qadaeer,
Thabit Sabbah and Muhammad attique khan in
Research Gate in March 2023.
[12]. “Crop Disease detection with Convolutional Neual
Network (CNN) Augumented with cellular automata”
by Kiran Sree Pokkulri and Usha Devi Nedunnuri in
IAJIT in September 2022.
[13]. “Machine Learning Technique for Crop Disease
Prediction Through Crop Leaf Image” by S.
Nandhini and K. Ashokkumar in Naturals publishing
in Janurary 2022.

IJISRT23DEC614 www.ijisrt.com 1197