Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).

IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

Deep Learning model for early prediction of plant

2021 Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV) | 978-1-6654-1960-4/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICICV50876.2021.9388538

disease

Rubini PE Dr.Kavitha P
Ph.D. Research Scholar, Department of Computer Science Associate Professor, Department of Computer Science
Engineering (VTU RC), CMR Institute of Technology, Engineering, CMR Institute of Technology, Bengaluru
Bengaluru [email protected]
[email protected]

Abstract— Agriculture is the main activity in many parts of the validation. Each of these groups contains completely different
countries. Agriculture acts as a vital part of the economic system images. For example, the images contained in the training
of every country. Agriculture not only provides food and raw dataset will not be contained in the validation dataset. Transfer
material but also acts as a source of livelihood for farmers. learning emphasizes saving evidence gained while resolving
Today, farmers are facing many challenges in agricultural land.
one problem and applying it to a different but interrelated
This research work focuses on one of the main challenges in
agricultural land i.e., disease prediction. The disease in crop problem. The model architecture used here is VGG-16 and
plants affects agricultural production, so a model is proposed to Dense Net.
automate a method for the prediction of disease in the plants and
intimating the farmers to take appropriate action beforehand. In
this work, a deep learning model is proposed that accurately II. RELAT ED WORK
classifies any leaf images is having a disease or not, in addition to In paper [1], Hyeon Park et al., projected a method to
providing a type of disease. The images of tomato plants were identify the strawberry disease. The farmer has to take a picture
drawn from the Plant Village dataset and trained using pre- of a leaf or a fruit through a smartphone and once the picture is
trained models like VGG16 and Dense Net using transfer uploaded in the analysis engine system, it will find out whether
learning and their accuracy is compared. S o, the proposed system it is a healthy or diseased image. If it’s a diseased image it will
in blending with its measured interpretability and accuracy may classify the type of disease. The author used a convolution and
effectively aid farmers in the early diagnosis of the diseased leaf. fully connected network for identifying the diseased strawberry
Keywords—disease prediction, CNN, pre-trained model, deep
fruits. They used the dataset with four types of diseased images
learning, agriculture
and healthy images of strawberries. Also, they generated the
dataset by intentionally causing diseases to the strawberry
I. INT RODUCT ION plants. In another paper [2], Halil Darmus et al., anticipated a
Agriculture is the main area that needs to be concentrated for deep learning approach to identify a variety of diseases on the
the growth of the economy. Farmers are held up with many leaves of tomato plants. Two different pre-trained deep
issues like weather forecasting, disease prediction, phenology learning network architectures namely Alex Net and Squeeze
identification, and crop harvesting. This research work Net were used. Nvidia Jetson TX1 was employed to train and
validate the architecture. Plant Village dataset with tomato leaf
concentrates on one of the problems i.e., disease prediction.
images has been used for training. It is shown that Squeeze Net
The identification of disease in plants is a very tough task, and
is a good architecture for the mobile deep learning
if it’s not monitored properly it leads to a decrease in the yield. classification. [3] Amara J., et al., proposed a deep learning-
To resolve the issues caused by unhealthy crops which affects based technique of CNN. LeNet architecture was used to
agriculture demands an approach for disease identification and classify the banana leaf diseases. The effectiveness of this
prediction. The need and challenge of disease detection in approach is that it works well with many complicated
crops are solved by a deep learning approach. In our work, conditions.
deep learning mechanism is used to classify leaf images as
diseased or healthy based on the input patterns of the image. [4] Guan Wang et al., had taken the Plant Village dataset
The most important work done in this research is to provide with the apple black rot images of four severity stages and used
the type of disease in which a plant is affected, it is totally deep convolutional neural networks to identify the complexity
of the disease. The performance of both the shallow and deep
irrespective of binary classification of the neural network
network model is evaluated. The author mentioned that the best
where the farmer can understand the type of disease and
model is the deepVGG16 model which provides an overall
provide an appropriate fertilizer. The dataset used in the accuracy of 90.4%. [5] Aditya Khamparia et al., proposed a
proposed work contains around 17000 images of tomato hybrid approach called a Convolutional encoder network to
leaves. They are split into three groups test, train, and identify crop leaf diseases. In this approach, 900 image datasets

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1104

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 08,2021 at 10:59:14 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

(600: training and 300: test set) were used from Plant Village. results are analyzed with the same dataset. The dataset taken
They have considered three crops (Potato, Tomato, and Maize) for implementation is the Plant Village dataset which contains
and five kinds of crop disease (early blight, late blight, leaf 17000 training images and 200 testing images
mold, yellow leaf curl, rust disease) for their implementation.
The proposed approach works well to identify crop disease D. VGG16 Architecture
given the input as a leaf image. Also, two kinds of The proposed work focusses on VGG16 architecture.
Convolutional filters like 2X2 and 3X3 were used in the VGG16(Visual Geometry Group) is a CNN model presented
implementation part which produces an accuracy of 97.50% by K. Simonyan and A. Zisserman. It is trained on the
and 100 % respectively. ImageNet dataset, a dataset containing 1.4 Million pictures of
1000 categories. It has a total of 16 CNN layers, hence the
III. M ET HODOLOGY name VGG-16, it consists of 13 CNN layers and 3 fully
The implementation is done in three stages: dataset connected layers. The character extraction layer of VGG-16
collection, pre-processing, Training and evaluating the model. will be frozen and the changes are made to the last three fully
connected layers. VGG-16 will use the weights of the already
A. Dataset Collection trained images from ImageNet as a reference and classify the
leaf images as diseased and not diseased. The architecture is
The dataset was taken from the Plant Village database. It
was created for use in Plant Disease Detection System, depicted in the below diagram Fig.2
containing 4 different species of plant varieties. The proposed
research work concentrated only on tomato leaf images. The
dataset has two categories of images corresponding to diseased
and healthy leaf images. The diseased images are labeled into 9
different classes as tomato bacterial spot, tomato early blight,
tomato leaf mold, tomato Septoria leaf spot, tomato spider
mites, tomato target spot, tomato yellow leaf curl virus, and
tomato mosaic virus.

B. Data Pre-Processing
In data preprocessing, the raw image taken from the
database had gone through preprocessing before fed into the
CNN model. The images are reconstructed and normalized to Fig. 2. VGG16 Architecture
establish a base size for all images and remove noise. Each
image is characterized as a three-dimensional vector of P, Q, The algorithmic step of VGG16 as described below:
and R. P and Q represents the width and height of an image,
and R represents a number of RGB channel. The images are
1. The fixed size of 224 x 224 RGB image is
resized to 224 x 224 pixels to be compatible with the transfer
forwarded as input to Conv 1.1. Then the image
learning models. The images obtained after pre-processing is
shown in Fig.1 proceeds through a pile of convolutional (Conv.)
layers, where the filters or the kernel used is of size
3*3 with a stride fixed to 1 pixel.
2. Spatial pooling is proceeded by 5 max-pooling
layers over a 2×2-pixel window, with stride 2.
3. Three Fully-Connected (FC) layers trail a pile of
convolutional layers:
a. The first two dense layers have 4096
channels.
Fig. 1. Pre-processing of images b. The third performs 2-way classification
and contains two channels (one for diseased
C. CNN Model and another for healthy).
4. The fourth and the last layer is the soft-max layer
Convolutional Neural Network (CNN) is a category of feed- For model creation, the VGG-16 is able to classify 1000
forward ANN in which the shape of connection between the different labels, but here just needed 2 labels i.e., Healthy or
neurons or nodes is the same as the working principle of the diseased. In order to do that the last fully connected layer of the
visual cortex of the human brain. There are three layers in model is replaced with a new one with 2 output features instead
CNN: Convolution layer, Pooling layer, and Fully Connected of 1000. The two labels used here are 0 for diseased and 1 for
layer. Training the model from scratch is a tedious and time- healthy.
consuming process so there are many pre-trained models are
The loss function is used to compute the loss of the model
like VGG 16, VGG-19, ResNet, Inception, Squeeze Net, Alex
so that the weights can be restructured to minimize the loss in
Net and DenseNet can be directly used. In this proposed work
the next evaluation. The choice of loss function plays
VGG 16 and DenseNet architecture were implemented and the

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1105

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 08,2021 at 10:59:14 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

an important role because it must go with the appropriate concatenates all the feature map. The architecture of DenseNet
modeling problem. The equation (1) is used to compute loss is depicted in the below diagram Fig.3
function.

(1)
The loss function used is Cross -Entropy which is more
suitable for binary classification problems where the target
values are in the set {0, 1}. Cross -entropy will estimate a value
that precises the average difference between the actual and
predicted probability distributions for predicting class 1. The
score is minimized and a perfect cross-entropy value is 0. The
number of epochs chosen here is 70. The equation (2) is used
to calculate Cross-Entropy. Fig. 3. DenseNet Architecture

In the above diagram Fig.3, the input layer is connected

with a dense block. There are three layers in the dense block.
The first Conv layer receives the feature map from the input
(2) layer. The second Conv layer receives the feature map from
Conv1 and the input layer. Similarly, the third conv layer
The two phases involved are training the model and testing. receives the feature map from Conv2, Conv,1 and the input
For every epoch, all the training batches are iterated, compute layer. Then the output of dense block is passed to the transition
the loss, and adjust the network weights. Then the performance layer which actually performs the task of max pooling. The
over the validation set is evaluated. At the end of every epoch, transition layer receives the feature map from Conv3, Conv2,
the network progress (loss and accuracy) is evaluated. The Conv1 and input layer then concatenates the feature map and
accuracy will convey how many predictions were correct. reduces the dimensionality. To avoid loss, the output of each
Training is done only for the classifier part of the model and layer feature map is fixed. Each conv layer in dense block
freeze of the feature extraction layer. For testing, an image contains Batch Normalization, ReLu, and Convolution layer.
which isn’t present in any of the training and validation dataset The number of feature maps in each layer is called growth
is used as an input. Using an entirely different image would test factor. Before the input reaches the dense block there is an
the efficiency of the model and will check if the model can initial convolution layer that produces the output size of
predict correctly even if a completely different image is given 112x112, followed by a pooling layer which reduces the output
as an input. size to 56x56.

E. DenseNet Architecture IV. RESULT

In CNN, the neural network is very deep, even though there The main intent of the proposed work is to recognize and
is a statement saying “deeper network provides greater identify whether a leaf is diseased or healthy and mention the
accuracy”, but it is very hard to train the model due to type of disease to the farmer. Trained models are tested on the
vanishing gradient. This process can be overcome by DenseNet validation set using GPU. There are 14000 training images and
which improves the gradient propagation by connecting all 3000 validation images. All of them are labeled. The deep
layers directly with each other. If there are L layers in the learning model was able to classify with an accuracy of 95-97
network then DenseNet will be having L(L+1)/2 connections, percent. The accuracy can be increased when trained with a
where a typical network with L layers will have L connections. vast number of images and by adopting pre-trained CNN
models.
DenseNet consists of dense blocks , and transition blocks. In
dense blocks all the layers are densely connected together. The The output of the VGG-16 and DenseNet model is shown
layers in the dense block get the input from previous layer below in Fig.4 Upon giving a healthy leaf image as input the
output feature maps. For example, if there are 10 layers and model predicts the output as healthy.
each layer provide the output feature map of different size, then
each layer receives more supervision from the previous layer
which causes the loss function so the DenseNet model fixes the
feature map of each layer.
The dense block contains a batch normalization, ReLu
activation and 3x3 convolution. The transition layer acts like a
max pooling layer to reduce the dimensionality, it

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1106

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 08,2021 at 10:59:14 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV 2021).
IEEE Xplore Part Number: CFP21ONG-ART; 978-0-7381-1183-4

Fig. 4. Prediction result for the test image.

The comparison of accuracy between VGG 16 and

DenseNet is tabulated in Table I and also a graph is drawn to
find the model accuracy and loss with respect to each epoch
count. It is observed that the accuracy of DenseNet is slightly Fig 7. Accuracy of VGG16 in training and testing dat a
higher than the VGG16 model in terms of crop disease
detection.
V. CONCLUSION
The research work has discussed the development of a
TABLE I. COMP ARISON OF VGG 16 AND DENSE NET
machine learning model that identifies whether a plant is
S. No Algorithm Accuracy diseased or healthy. The accuracy achieved with this model is
1. VGG16 92 satisfactory. To improve this model, we can adopt other
machine learning algorithms and try to obtain a more efficient
2. DenseNet 98.25
classifier. The main drawback of VGG-16 is its fully connected
nodes and is over 533MB. This makes deploying VGG a
The model accuracy of DenseNet and VGG16 for the tiresome task. The DenseNet architecture proves better
training and testing data is shown in Fig.5,6 & 7. The model accuracy than VGG16 because of its more diversified features.
also proves better accuracy for the test image given by the user. The dataset used here includes only one type of plant and can
The model loss is evaluated and it shows less than 5% detect the type of disease with which the leaf is affected with.
In addition to this, we can include different varieties of plant
species and train the model to detect different kinds of plant
diseases.

REFERENCES

[1] Hyeon Park, Eun JeeSook, Se-Han Kim, “Crops Disease Diagnosing
using Image-based Deep Learning Mechanism,” IEEE International
Conference on Computing and Network Communications (CoCoNet),
October 2018.
[2] Halil Durmus,Ece Olcay Guneu, Murvet Korco,“Disease Detection on
Fig. 5. DenseNet Accuracy the Leaves of the Tomato Plants by Using Deep Learning”, International
Conference on Agro-geoinformatics, 2017.
[3] J.Amara, B.Bouaziz, Algergawy, “A Deep Learning-Based Approach
for Banana Leaf Diseases Classification”, BT W workshop, Stuttgart,
2017, pp.79–88.
[4] Guan Wang, Yu Sun, and Jianxin Wang, “Automatic Image-Based Plant
Disease Severity Estimation Using Deep Learning”, Computational
Intelligence and Neuroscience, 2017.
[5] Aditya Khamparia, Gurinder Saini, Deepak Gupta, Ashish Khanna,
Shrasti T iwari, Victor Hugo C., de Albuquerque, “ Seasonal Crops
Disease Prediction and Classification Using Deep Convolutional
Encoder Network”, Circuits, Systems, and Signal Processing, Springer
Publication,2019.
Fig 6. DenseNet Loss

978-1-6654-1960-4/21/$31.00 ©2021 IEEE 1107

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 08,2021 at 10:59:14 UTC from IEEE Xplore. Restrictions apply.