2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 13 – 14 May 2021 | Coimbatore

Plant disease classification using deep learning

Akshai KP J.Anitha
Karunya Institute of Technology and Karunya Institute of Technology and
Sciences Sciences
Coimbatore, India Coimbatore, India
2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 978-1-6654-2864-4/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICSPC51351.2021.9451696

[email protected] [email protected]

Abstract—Agriculture plays a crucial role in the Indian trained using the plant dataset. Among them, the DenseNet
economy. Early detection of plant diseases is very much model achieved the highest accuracy of 98.27%.
essential to prevent crop loss and further spread of diseases.
Most plants such as apple, tomato, cherry, grapes show visible
symptoms of the disease on the leaf. These visible patterns can
be identified to correctly predict the disease and take early II. RELATED WORKS
actions to prevent it. The conventional method is the farmers
or plant pathologists manually observe the plant leaf and In [1] authors proposed a method to classify different
identify the type of disease. In this project, a deep learning plant diseases using the EfficientNet deep learning model.
model is trained to classify the different plant diseases. The The model was trained using the Plant Village dataset having
convolutional neural network (CNN) model is used due to its 55,448 images and the augmented version of the dataset has
massive success in image-based classification. The deep 61,486 images. The proposed EfficientNet model's
learning model provides faster and more accurate predictions performance is compared with other state-of-the-art CNN
than manual observation of the plant leaf. In this work, the models such as ResNet, VGG, AlexNet, and Inception
CNN model and pre-trained models such as VGG, ResNet, and model. In transfer learning, all the layers were set to
DenseNet models are trained using the dataset. Among them, trainable. Among all the models, the EfficientNet B5 and B6
the DenseNet model achieves the highest accuracy. models achieved the highest results.
Keywords—Plant disease detection, Convolutional Neural M. Akila and P. Deepan used Region-based Fully
Networks, Transfer Learning, VGG, ResNet, DenseNet. Convolutional Network (R-FCN), Faster Region-based CNN
(Faster R-CNN), and Single Shot Multibox Detector (SSD)
to classify different plant leaf diseases [2]. The dataset
contains plant leaf disease images of commercial crops such
I. INTRODUCTION as banana, sugarcane, cotton, potato, brinjal, carrot, chilly,
Plant disease detection is a significant challenge in the rice, wheat, and guava. The images were collected from the
agriculture sector. Some of the plants show visible symptoms internet and by manually taking the photo. The image
on the plant leaf. These leaf patterns can be used to identify augmentations such as affine and perspective transformation,
different diseases and take immediate action to prevent the rotations, and image intensity transformations were
spread. Most of these plant diseases are difficult to detect performed to expand the dataset and to prevent the model's
through naked eyes, and even experienced persons end up overfitting.
wrong. The accuracy of the manual prediction depends upon In [3], the authors used Caffe deep learning framework
the experience and knowledge of the person [1]. There are with Imagenet weights to classify plant diseases. The model
much research works done for plant leaf disease consists of eight learning layers and five convolution and
identification using machine learning and deep learning fully connected layers. The dataset was prepared by
models. This study proposes a deep learning model to downloading images from the internet. Image augmentation
classify different plant diseases. In standard machine was performed to increase the dataset and prevent overfitting
learning models, the features are extracted manually, and the during the training stage. The proposed Caffe model
algorithm learns from that data. Thus it is a two-step process. achieved an accuracy of 96%.
The deep learning model uses an artificial neural network
that can learn features from an input image and make Plant disease detection can be done by targeting the
intelligent decisions on its own. Thus deep learning models disease-affected places using image processing techniques.
are far more capable than the standard machine learning Sanjay and Shrikant used simple threshold methods and
models for image-based classification. triangle thresholding methods to segment the leaf and lesion
area, respectively [4]. The diseases are classified by
The working model uses convolutional neural networks calculating the lesion and leaf area. The proposed system
and transfer learning to classify different plant leaf diseases. achieved an accuracy of 98.60%.
CNN is a type of deep learning neural network and has good
success in image-based classification. The proposed system Revathi and Hemalatha used particle swarm optimization
is faster and more accurate than the conventional way of (PSO) for feature extraction and Cross Information Gain
manual observation of each plant leaf. Deploying such a Deep forward Neural Network to classify different cotton
model into a mobile application can help farmers detect leaf diseases with an overall accuracy of 95% [5]. The color,
different plant diseases using mobile cameras and take shape, and texture features are extracted using particle swarm
necessary actions to avoid disease spread. For this study, the optimization to classify different diseases. The feature
images of grape plant diseases from the Plant Village dataset extraction method helps to identify the disease leaf spots and
are used to train the model. The CNN (Convolutional neural increases the model's overall accuracy.
network), VGG19, ResNet-152v2, and DenseNet models are

978-1-6654-2864-4/21/$31.00 ©2021 IEEE 407

Authorized licensed use limited to: National University of Singapore. Downloaded on July 03,2021 at 21:10:55 UTC from IEEE Xplore. Restrictions apply.
 2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 13 – 14 May 2021 | Coimbatore

Kulkarni and Patil proposed a methodology for detecting

plant leaf diseases using image processing techniques and
artificial neural networks (ANN) [6]. The images were
filtered and segmented using the Gabor filter. The extracted
features that could classify the healthy and diseased samples
were used to train the ANN model and achieved an accuracy
of 91%.

III. METHODOLOGY OF THE CLASSIFICATION MODEL

The convolutional neural network is used in this work to
classify different plant diseases. Fig.1 shows the diagram of
the working of the classification model. The first layer in the
model is the convolution layer that is used to extract features
from an input image. The convolution layer applies different
filters on the image to create a feature map and help to
extract different features. The convolution layers are
followed by pooling layers. The pooling layers reduce the
image size and the no of parameters. Common types of
pooling operations are max pooling, average pooling, and
sum pooling. After convolution and pooling operations, the
matrix is flattened into a vector and is passed into the fully Fig 1. Diagram of the classification system
connected layer. In this study, rectified linear activation
function or ReLU activation function is used for the
convolution layer and a Softmax activation function for the The DenseNet model is similar to ResNet and is
output layer. Rectified linear activation function returns the introduced to solve the vanishing gradient problem in large
output value if the input is greater than zero and returns zero neural networks. The ResNet adds the output of one layer to
if the input value is zero or less (1). the next layer, whereas in the DenseNet model, the output
feature maps are concatenated with the following future
feature maps. The model has different versions such as
DenseNet121, DenseNet160, and DenseNet201. The model
f(x) = max[0, x] (1)
is divided into different blocks called Dense Blocks, in
which the dimensions of the feature maps remain the same
and the number of filters changes. Between every dense
Transfer learning is a method in which pre-trained block, there are transition layers that perform convolution,
models are reused for a new task instead of developing pooling, and batch normalization. In this study, the
models from scratch. These models are trained on a large DenseNet-201 model with Imagenet weight is used.
number of images and can improve the accuracy of
prediction. In this study, VGG, DenseNet, and Resnet
models are trained using the transfer learning technique. The IV. IMPLEMENTATION DETAILS
results of these models are compared with each other, and the
performance is analyzed. A. Dataset
The data for this study is gathered from the Plant village
The VGG is one of the popular models in ILSVRC 2014. dataset. The Plant Village dataset consists of more than
It is made by making improvements on the AlexNet by 55,000 images with 38 classes of 14 different plant species.
replacing large kernel filters with multiple filters. The input Out of 38 classes, 12 are healthy, and 26 are diseased leaf
of the VGG model is 224x224 RGB image. The input image classes. In this study, the grape plant leaf images are taken,
is passed through a stack of convolution layers of 3x3 filters. containing 4,062 images with four classes. The classes are
The convolution layers are followed by 2x2 pixel window healthy, black rot, Esca (Black Measles), and Leaf blight
Max pooling layers. In this study, the VGG19 model is taken (Isariopsis Leaf Spot). The dataset consists of 423 healthy
from Keras application with Imagenet weights, and the leaves, 1,180 black rot affected leaves, 1,383 esca affected
layers were made non-trainable. leaves, and 1,076 Leaf blight affected images. The sample
Researchers at Microsoft research in 2015 propose the images from each class are shown in Fig 2-5. Deep learning
Resnet model. They have introduced a new architecture models require a large dataset to achieve a good
called Residual Network. After the AlexNet model, the performance. Using the Image data generator in Keras, the
winner of the ImageNet 2012 competition, all other proposed images are resized into 224x224 pixel size, and
architecture uses more layers to reduce the error rate. One of augmentations such as rotation, zoom, and shift were
the main problems faced is the vanishing/exploding gradient. applied. The primary purpose of the augmentation process is
To solve this problem, the residual network has been to expand the dataset and prevent overfitting during the
introduced. In this architecture, few layers are skipped from training stage. To perform the experiments, the dataset is
training and are connected directly to the output class. It split into training and validation set in the ratio of 80:20. The
helps to reduce the vanishing/exploding gradient problem. training set contains 3,258 images, and the validation set
contains 812 images.

408

Authorized licensed use limited to: National University of Singapore. Downloaded on July 03,2021 at 21:10:55 UTC from IEEE Xplore. Restrictions apply.
 2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 13 – 14 May 2021 | Coimbatore

Fig. 2. Healthy grape leaves

Fig. 6. Schematic representation of CNN model

The literature studies reveal that pre-trained models using

transfer learning are an efficient strategy for plant disease
Fig. 3. Blight affected leaves classification. Popular pre-trained models such as VGG,
ResNet, and DenseNet are used and the experimental results
are compared. The images are resized to 224x224, which is
the default size accepted by the VGG, ResNet, and Densenet
models. The loss and accuracy plotted during different
epochs are shown in Fig 7. The experimental results
presented in Table 1 show that the DenseNet model
outperformed all other models with an accuracy of 98.27%.
The classification performance of each model is analyzed
using metrics such as confusion matrix, precision, accuracy,
Fig. 4. Black rot affected leaves recall, and F1-score. The confusion matrix of the model is
generated for the test set and is shown in Fig 8. The
confusion matrix gives a summary of the predictions against
the actual values.

Accuracy = (TN + TP) / (TN+TP+FN+FP) (2)

Recall = TP / (TP + FN) (3)

Fig. 5. Esca affected leaves

Precision = TP / (TP + FP) (4)

B. Training Neural Network F1 score = 2 * [(precision*recall) / (precision+recall)] (5)

In this study, a convolutional neural network and popular
pre-trained models such as VGG, ResNet, and DenseNet are
trained using the plant dataset, and results are compared. The
CNN model used contains pairs of convolution and pooling TABLE I. CLASSIFICATION PERFORMANCE OF DIFFERENT MODELS
layers. The convolution layers perform filters on the input
image and extract the features. The pooling layers reduce the Overall
Model Precision Recall F1-score
size of the input image and reduce the computation accuracy
performed in the network. Fig 6 shows the schematic
representation of the CNN model. CNN Model 94.60 94.58 94.56 94.58

The CNN model is trained using the training set for 20

VGG Model 95.54 95.32 95.32 95.32
epochs. The optimizer used is Adam, and the loss function
selected is categorical cross-entropy. Adam is an algorithm RESNET
for stochastic gradient optimization which works well with 97.11 97.04 97.05 97.04
Model
sparse data. The loss and accuracy plotted during different DENSENET
epochs are shown in Fig 7a and 7b. The CNN model 98.31 98.27 98.28 98.27
Model
achieves an overall accuracy of 94.58%.

409

Authorized licensed use limited to: National University of Singapore. Downloaded on July 03,2021 at 21:10:55 UTC from IEEE Xplore. Restrictions apply.
 2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 13 – 14 May 2021 | Coimbatore

Fig 7 a-b) Loss and Accuracy curves of the CNN model. c-d) Loss and Accuracy curves of the VGG model e-f) Loss and Accuracy
curves of the ResNet model g-h) Loss and Accuracy curves of the DenseNet model

410

Authorized licensed use limited to: National University of Singapore. Downloaded on July 03,2021 at 21:10:55 UTC from IEEE Xplore. Restrictions apply.
 2021 3rd International Conference on Signal Processing and Communication (ICPSC) | 13 – 14 May 2021 | Coimbatore

Fig 7 a) Confusion matrix of CNN Model b) Confusion matrix of VGG Model c) Confusion matrix of ResNet Model d) Confusion
matrix of DenseNet Model

website/application as it will help farmers/pathologists to

V. CONCLUSION identify different diseases using their mobile cameras.
This study facilitates the early diagnosis of plant
diseases to prevent crop loss and the spread of diseases. REFERENCES
The CNN model is used to predict different plant diseases [1] Umit Atilaa, Murat Uçarb, Kemal Akyolc, and Emine Uçar, "Plant
correctly. The performance of various pre-trained CNN leaf disease classification using EfficientNet deep learning model,"
models such as VGG, ResNet, and DenseNet is observed, Ecological Informatics, vol.61, Article 101182, March 2021
and then based on performance metrics, the DenseNet [2] M.Akila and P.Deepan, "Detection and Classification of Plant Leaf
model is found to be more accurate. The model's testing is Diseases by using Deep Learning Algorithm," International Journal
done using performance evaluation metrics such as of Engineering Research & Technology (IJERT), SSN: 2278-0181,
2018
accuracy, precision, recall, and F1 score. The DenseNet
model achieved the highest accuracy of 98.27%. One of [3] Srdjan Sladojevic, Marko Arsenovic, Andras Anderla, Dubravko
Culibrk, and Darko Stefanovic, "Deep Neural Networks Based
the main problems faced in a larger neural network is the Recognition of Plant Diseases by Leaf Image Classification,"
vanishing gradient problem. The ResNet model adds the Hindawi Publishing Corporation Computational Intelligence and
output of one layer to the next layer, whereas in the Neuroscience, vol. 2016, Article ID 3289801, 2016
DenseNet model, the output feature maps are concatenated [4] S. B. Patil and S. K. Bodhe, “Leaf disease severity measurement
with the following future feature maps. Each layer will using image processing,” International Journal of Engineering and
Technology, vol.3, no.5, pp.297–301, 2011.
receive features from the previous layers and pass its
features maps to all subsequent layers. Due to that [5] P. Revathi and M. Hemalatha, “Identification of cotton diseases
based on cross information gain deep forward neural network
DenseNet model will have features of all complexity and classifier with PSO feature selection,” International Journal of
performs well in small datasets. Future works include Engineering and Technology, vol.5, no.6, pp.4637–4642, 2014
expand the dataset and increase the number of classes. [6] Kulkarni Anand H and Ashwin Patil RK, “Applying image
Another future work is deploying the model into a processing technique to detect plant diseases,” International Journal
of Modern Engineering Research (IJMER), vol. 2, pp. 3661-3664,
2012

411

Authorized licensed use limited to: National University of Singapore. Downloaded on July 03,2021 at 21:10:55 UTC from IEEE Xplore. Restrictions apply.