Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Crop Disease Detection using CNN

Ajay Rasave1
PG. Student,
ME Artificial Intelligence,
Department of Computer Engineering,
KJ Somaiya Institute of Engineering and Information Technology, Sion, Mumbai

Dr. Sunita R. Patil2

Professor, Department of Computer Engineering,
KJ Somaiya Institute of Engineering and Information Technology, Sion, Mumbai

Abstract:- Good health is a first priority of all human of main reason is that wrong diagnosis of disease which
being. And to maintain health the first need is a healthy results into wrong treatment and wastage of money. Many
food, which we get from our agriculture resources. times, it becomes difficult to identify unknown diseases with
According to the United Nations’ World Population traditional ways. So, it’s very crucial to use new and smart
Prospects (WPP), India is projected to surpass China as ways of disease detection for early and perfect remedy.
the world’s most populous country in 2023. And to
maintain the appetite and health of such huge population Now a days Computer Vision technology are so
it’s very important for agriculture field to grow the more efficient that they can play a vital role in identifying diseases
and healthy crops. Now a days there are number of on crop with just an image. Classification algorithms such as
diseases which are causing crop underproduction. Delay Convolution Neural Network (CNN), K-Means, Residual
in identifying the actual disease will result into more loss Neural Networks, etc. are famous for classifying different
and cost of remedy. So, it’s very important to have the kinds of images, based on different parameters such as color,
knowledge of diseases for early detection. In today’s date shape, etc.
Artificial intelligence technologies such as Classification
algorithms using Deep Learning are so efficient and are Convolution Neural Networks have shown their
performance and accuracy in image classification activities.
capable of predicting the disease based on the images of
the crops. Hence, in this paper have reviewed the most So, in this paper we are going to consider different types of
efficient classification algorithms and proposed few most CNN algorithms.
suitable models for image-based disease detection. A. Convolution Neural Networks:
Keywords:- CNN, Inception V3, VGG16, VGG19, ResNet18. CNN which also referred as Convolution Networks
(ConvNet) is one of the Deep Learning algorithms. It
I. INTRODUCTION isconsisting of different layers such as pooling Layers and
convolution layers and Fully Connected layersas shown in
India is known as a Krishi Pradhan i.e.an agrarian Fig.1. The term Convolution in CNN stands for a
country. Economy is majorly dependent on agriculture mathematical operation [1]. Mathematically Convolution
sector. Farmers are tending towards newer ways of farming. can be defined as an operation on two functions which gives
They are now trying to use the new technologies for better an output as a new function which is inspired by the
crop production. In recent days there are many new and old previous functions.
types of diseases which are affecting crop production. One

Fig. 1: Convolutional Neural Network [2]

IJISRT23MAY1463 www.ijisrt.com 1816

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CNN has two main sections, first Feature Extraction II. LITERATURE SURVEY
model which is combined of Convolution Layers and
Pooling Layers. And the second one is Classification model Due to huge scope in agriculture domain and advance
which mainly consist of Fully Connected Layers. commuting technologies there are lots of research done on
crop diseases detection with deep learning. In this section we
B. Feature Extraction Model have done the literature survey on recent research done for
In this part the network tries to learn detecting the crop diseases classification with different machine learning
features from the input data. Here the convolution and algorithms.
pooling are sequenced as shown in Fig.1.
In the research [4], Sachin B. Jadhav, Vishwanath R.
 Convolution Udupi and Sanjay B. Patilproposed a crop disease detection
This process is mainly responsible for applying different mechanism using Convolutional Neural Network (CNN)
learnable filters and extract the features from the given classification algorithms namely AlexNet and GoogleNet.
image. Filters are applied on the image pixels and output They trained and compared these two models on around 649
given is nothing but 2-D map i.e. a feature map. This unhealthy images containing diseases like Bacterial blight ,
convolution is be formulated as equation 1. Brown spot and Frogeye leaf spot. and 550 healthy images
of soyabean crop. They observed that AlexNet performed
𝑀𝑖 = 𝑏𝑖 + ∑𝑘 𝑊𝑖𝑘 ⋆ 𝑋𝑘 (1) more accurately over GoogleNet. AlexNet gave accuracy
around 98% whereas GoogleNet achieved accuracy around
Where, 96%.
𝑀𝑖 = feature map R. Hu et al. in [5] proposed their own CNN model for
⋆ = convolution operator, corn leaf disease identification model based on the data
𝑋𝑘 = k th input channel, augmentation using GoogleNet. And compared the results
𝑊𝑖𝑘 = sub kernel of that channel, with other pre trained algorithms like ResNet18, VGG16,
bi = bias term and VGG19. They achieved the accuracy of around 97.6%
with their own model, which was more than compared with
Now, the ReLu function is applied on each convolution
others. They used around 4354 corn leaf images from Plant
to make it non-Linear.
Village datastore containing 4 kinds of diseases. They first
 Pooling: trained their own model with proper fine tuning. Then for
This is a Max pooling layer which responsible for same dataset they used transfer learning models of VGG16,
reducing the feature map size. It reduces the convolution VGG19, ResNet18. And compared the accuracies,
map size by sub sampling layer. performances, precisions, Recall, F1 scores.

C. Classification model In the research [6] proposed by Selvaraj et al. they

The output given buy feature extraction layer, which is proposed a system to detect banana leaf disease and pets
nothing, but a feature map is processed in this model in such based on ResNet50 and Inception v2. They achieved the
a way that each neuron will be connected to all the features accuracy of around 90% which was comparatively better
from this feature map i.e. in a fully connected networks than MobileNetV1. They collected banana leaf images from
manner. In this classification process SoftMax activation Africa and south part of India having around 18 different
function is used. And it can be formulated as in equation 2, kinds of diseases.
given below.
As per the research [7] proposed by Godliver
𝑒 𝑧𝑐 Owomugisha et al., Linear SVC classifier gave the accuracy
𝑦𝑐 = ς(z)𝑐 = ∑𝐶 𝑧𝑑 for c = 1···C (2) of around 99% for the ORB features. In this research they
𝑑=1 𝑒
implemented a system to measure the severity of a disease
Where, on the leaf of cassava plants. They used three classifiers
namely Linear SVC, KNN and Extra Trees to extract
ς is Soft Max function different features from the diseased images. They trained the
modules on around 7386 images containing healthy and 4
input a C-dimensional vector z outputs a C-
different classes of diseases. And the severity was measured
dimensional vector y of real values between 0 and 1.[3]
in a range from 1 to 5. Where severity 1 representing healthy
image and severity 5 as unhealthy.

IJISRT23MAY1463 www.ijisrt.com 1817

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
III. THE PROPOSED MECHANISM our proposed system is based on the CNN classifiers.
Proposed solution is as shown in Fig. 2. In order to make
Considering the similar research works done in image our system fast we use transfer learning mechanism i.e.
classification, Convolution Neural Network i.e. CNN has Pretrained CNN modules will be used. That will save the
proved its capability, performance and accuracy, which is time required to train initial common steps.
comparatively far better than other algorithms. And hence

Fig. 2: Proposed System

Proposed system is composed of four commonly used transformations, Fourier transform and Image restauration,
steps in CNN modules. First step is image acquisition which Geometric Transformations, etc.
is a image capturing layer. This step is responsible to avail
image data ready for further processing. Second stage is Images from plant Village datastore are collected from
Image Pre-processing, where images will be filtered, resized many kinds of diseases. Images for same plant for same
as per the requirement. Third one is Feature Extraction disease are captured in multiple ways like different
which find out the different features from images required angle/orientation [10]. It can consist of images of different
for the classification. The fourth step is classification stage resolutions, quality, brightness. So, it’s crucial to pre-
where images will be classified based on the features process images in order to make them in proper size,
extracted in previous step. And a final fifth step which tells brightness, signal level, etc. In the proposed system Images
us about health of crop. will be resized to 256 × 256 pixels.

Finally, complete content and organizational editing C. Feature Extraction:

before formatting. Please take note of the following items The output of a deep learning-based prediction system is
when proofreading spelling and grammar: based on the number of features/factors it has. Having more
numbers of features will make it even harder or complex to
A. Image Acquisition predict result [11]. Feature extraction layer is responsible to
Image Acquisition [8] is the first step of the proposed produce new features from existing one. These new features
system. The image data collected from any sources can be are reduced features, and they represent the same
consist of images of different size, lightening, background. information from previous features.
Or it can be impacted because of any corrupt format. So, this
stage will make sure that image data provided to the next As our proposed system uses the Transfer Learning
step is in appropriate format. mechanism, we will use pretrained models which will save
the time required for few time-consuming steps. Here the
We are going to use image data from Plant Village proposed CNN module needs to be fine-tuned in order to get
data source. Our dataset is consisting of around 6354 images highest accuracy.
of Tomato Leaf containing different kinds of disease such as
Bacterial Spot, Late Blight, Leaf Mold, Septoria Leaf Spot, D. Classification:
Yellow Leaf Curl Virus, Mosaic Virus, etc. This is the main step of the CNN module. This stage is
responsible to predict out the most suitable class for input
B. Image Pre-processing: image by comparing the features captured by previous layer.
In deep learning Image pre-processing [9] is one of the Or it gives the probability that the input belongs to particular
important steps. It also referred as the data cleaning. The class [12].
aim of the image pre-processing is to modify data in such a
E. Disease Detection:
way that it will be more efficient and suitable for specific
This is the final stage of the proposed system. It takes
application. Modification here means that reduce any
output of classification as the input. This stage in general
distortion in images, adjust colour, brightness, or improve
referred image post-processing layer. In this stage we can get
any specific feature which will help to improve the
further information regarding the health of the crop such as if
performance of the application. To do this we can different
leaf is diseased then how much portion is infected or which
kinds of image pre-processing techniques such as Image
area is infected. Also this layer can further help us to suggest
Filtering and Segmentation, Pixel brightness corrections or
remedies for respective diseases.

IJISRT23MAY1463 www.ijisrt.com 1818

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
IV. PERFORMANCE EVALUATION In the proposed approach we will use transfer learning
mechanism in order to improve performance of the system.
We are going to use accuracy, precision, recall, and The dataset is consist of 6354 tomato leaf images containing
F1-score as performance evaluation metrics [13]. six types of diseases. These algorithms will be trained on
same set of data. And the performance of each module will
A. Accuracy be compared with the other one. Performance will be
Accuracy is one metric for evaluating classification evaluated based on Accuracy, Precision, Recall and F1 score
models. Informally, accuracy in percentage is the percentage etc.
of correct predictions over the total number of predictions
done by model. Formally, accuracy has the following REFERENCES
definition:
[1.] Ian Goodfellow, Yoshua Bengio and Aaron Courville.
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟𝑠 𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
(3) "Deep Learning", MIT Press, p. 326, 2016.
[2.] Pérez, Dalila & García-Vázquez, Juan & Salomon-
For binary classification, accuracy can also be Torres, Ricardo. (2021). Evaluation of Convolutional
calculated in terms of positives and negatives predictions as Neural Networks' Hyperparameters with Transfer
follows: Learning to Determine Sorting of Ripe Medjool
Dates. Agriculture. 11. 1-12.
TP+TN 10.3390/agriculture11020115.
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = TP+TN+FP+NP
(4)
[3.] Jihen Amara, Bassem Bouaziz, and Alsayed
Where, Algergawy, “A Deep Learning-based Approach for
Banana Leaf Diseases Classification”, BTW 2017–
TN – True Positive Workshopband, pp. 83-85, 2017
TN – True Negative [4.] Sachin B. Jadhav, Vishwanath R. Udupi, Sanjay B.
FP – False Positive Patil, "Identification of plant diseases using
FN – False Negative convolutional neural networks", International Journal
of Information Technology, 2020.
B. Precision: [5.] Rongjie Hu, Shan Zhang, Peng Wang, Guoming Xu,
Precision is nothing but the fraction of Positives over Daoyong Wang and Yuqi Qian. 2020. The
total positives i.e. False positives and True Positives. It can identification of corn leaf diseases based on transfer
be formulated as, learning and data augmentation. In Proceedings of
2020 3rd International Conference on Computer
[True Positives]
Precision = [True Positives + False Positives] (5) Science and Software Engineering (CSSE’20).
Beijing, China, 8 pages.
Recall: https://doi.org/10.1145/3403746.3403905
[6.] M. G. Selvaraj, A. Vergara, H. Ruiz, N. Safari, S.
Recall is portion of correctly identified positives. It is Elayabalan, W. Ocimati, and G. Blomme, “AI-
formulated as, powered banana diseases and pest detection,” Plant
Methods, vol. 15, no. 1, pp. 1–11,2019
[True Positives]
Recall = (6) [7.] Godliver Owomugisha, Ernest Mwebaze, “Machine
[True Positives + False Negatives]
Learning for Plant Disease Incidence and Severity
C. F1 Score: Measurements from Leaf Images,” 2016 15th IEEE
F1 Score is useful to get measure of wrongly classified International Conference on Machine Learning and
cases compare to Accuracy Metric. It can be formulated as Applications, 2016 IEEE
below. [8.] Ivan Zhou. “Designing Image Acquisition for
Machine Vision.” LANDING AI.
Recall∗ Precision https://landing.ai/designing-image-acquisition-for-
𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 ∗ Recall+ Precision
(7)
machine-vision/#open-overlay (accessed Mar. 30,
V. CONCLUSION 2023).
[9.] Great Learning Team. “Introduction to Image Pre-
Due to huge variety of diseases with huge variety of processing | What is Image Pre-processing?.” Great
symptoms its usually difficult for farmers to identify the Learning.
correct disease on the crop in order to take correct remedy https://www.mygreatlearning.com/blog/introduction-
action. Which further lead to more production loss and the to-image-pre-processing/ (accessed Mar. 30, 2023).
wasted cost on wrong remedy because of wrong diagnosis. [10.] Mohanty SP, Hughes DP and Salathé M (2016),
Hence it is very crucial to find out solution which can “Using Deep Learning for Image-Based Plant Disease
identify crop disease with fast and easier way. Artificial Detection,”. Front. Plant Sci. 7:1419, Volume 7 –
Intelligence with variety of technology can help to solve this 2016
problem with efficient ways. Deep Learning modules such [11.] Dishaa Agarwal. “Guide For Feature Extraction
as Inception V3, VGG, ResNet, etc. are very popular deep Techniques.” Analytics Vidhya.
learning algorithms in image classification field. So, our https://www.analyticsvidhya.com/blog/2021/04/guide
proposed system will be implemented with these algorithms.

IJISRT23MAY1463 www.ijisrt.com 1819

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
-for-feature-extraction-techniques/ (accessed Mar. 30,
2023).
[12.] Anne Bonner. “The Complete Beginner’s Guide to
Deep Learning: Convolutional Neural Networks and
Image Classification.” Towards Data Science.
https://towardsdatascience.com/wtf-is-image-
classification-8e78a8235acb (accessed Mar. 30,
2023).
[13.] Koo Ping Shung. “Accuracy, Precision, Recall or
F1?.” https://towardsdatascience.com/accuracy-
precision-recall-or-f1-331fb37c5cb9 (accessed Mar.
30, 2023).

IJISRT23MAY1463 www.ijisrt.com 1820