Plant Leaf Disease Detection Using

Machine Learning

Jay Trivedi(B) , Yash Shamnani, and Ruchi Gajjar

Department of Electronics and Communication Engineering, Institute of Technology,

Nirma University, Ahmedabad, Gujarat, India
[email protected], [email protected],
[email protected]

Abstract. In this paper, a Convolutional Neural Network (CNN) archi-

tecture for plant leaf disease detection using techniques of Deep Learning
is proposed. A CNN model is trained with the help of the Plant Village
Dataset consisting of 54,305 images comprising of 38 different classes of
both unhealthy and healthy leaves. The disease classification accuracy
achieved by the proposed architecture is up to 95.81% and various obser-
vations were made with different hyperparameters of the CNN architec-
ture. The experiment results achieved are comparable with other existing
techniques in literature.

Keywords: Plant disease detection · Deep Learning · Convolutional

Neural Network · Image classification

1 Introduction
Crop diseases, however, the small issue it may sound, holds an immense potential
to bring famines and is the main reason for the world’s food insecurity. Loss of
30 to 50% of yield for major crops due to disease is not uncommon. History in
recent times has witnessed many severe crop disease outbreaks such as the late
blight of potato in Ireland (1845–1860), Coffee rust in Sri Lanka (Ceylon then,
1870), and in Central and South America (1960, 2012 to present). The effects
are worse for the underdeveloped countries where there are fewer facilities and
access to plant-disease control methods [12].
Diagnosis of crop disease is done by analyzing a large number of samples
by means of Microscopy, DNA sequencing-based methods that provide detailed
information regarding disease-causing pathogens such as bacteria, viruses, fungi,
etc. and many more [3]. However, the problem is, most of the farmers don’t have
access to these diagnosing methods. According to the research of the World Bank
in 2016, 7 out of 10 of the poorest 20% of developing countries are exposed to
mobile communication and 40% of the whole world has internet access [10].
Due to the recent advancement and heavy research being made in the field of
deep learning and Artificial Intelligence (AI), applied to agriculture and vegeta-
tion, a lot of problems like yield detection, crop disease identification, intelligent
c Springer Nature Singapore Pte Ltd. 2020
S. Gupta and J. N. Sarvaiya (Eds.): ET2ECN 2020, CCIS 1214, pp. 267–276, 2020.
https://doi.org/10.1007/978-981-15-7219-7_23
268 J. Trivedi et al.

farming are solved. To address the problem of poor yield and loss due to crop
disease, if an automated system can identify the type of disease in the crop,
timely aid can be provided in terms of pesticides or suitable measures can be
taken, which will be beneficial for agricultural dependent countries like India,
where the agricultural sector contributes 375.61 billion dollars towards the Gross
Domestic Product of the country. Farmers can be provided with the technology
in their hands through an application in their smartphones that can detect the
health of the crop at an early stage and can alert them to proceed further to
treat those crops.
A Convolutional Neural Network Architecture, followed by a fully connected
network had been implemented which exhibits higher accuracy in the detection
of plant disease as compared to other similar researches, and to begin with,
correct identification of the type of disease in a crop is the first step. Once the
disease in a crop is detected, corrective measures to reduce further damage can
be taken. In this paper, Machine Learning approach is employed to detect the
disease in crop. The main contributions of this paper are listed as below:
– A Convolutional Neural Network Architecture is proposed for the detection
of disease in plants from leaf images.
– The proposed architecture gives a higher accuracy as compared to existing
models.
– The classification accuracy is achieved and validated by using different com-
binations of hyperparameters.
Rest of the paper is organized as follows: Sect. 2 gives the Literature Survey
and describes various techniques for plant disease detection. Section 3 describes
the proposed CNN Architecture for disease classification. Section 4 presents the
classification results and comparison with other method. Section 5 gives the
conclusion.

2 Related Work
Crop Disease detection at an early stage can drastically reduce the economic
losses caused by it. Several methods are used for the same and more research is
being done in order to bring out the most accurate and economical way that can
reach each and every farmer. So basically plant disease detection can be carried
out by two methods: 1) Direct and 2) Indirect. The direct method includes
analyzing a large number of samples through molecular and serological methods.
Direct methods need to have a lot of samples to provide results and are time-
consuming. However, they are accurate as well. Commonly used direct methods
for plant disease detection are Polymerase Chain Reaction (PCR), Fluorescence
In-Situ Hybridization (FISH), Enzyme-Linked Immunosorbent Assay (ELISA),
Immunofluorescence (IF) and Flow Cytometry (FCM) [3].
The advancement in the field of image processing brought out several optical
sensor-based solutions which we call indirect methods for plant disease detection.
Detailed information is provided by these optical sensors in the form of different
 Plant Leaf Disease Detection Using Machine Learning 269

electromagnetic spectra which helps in determining the plant health. Commonly

used indirect methods are thermography, fluorescence imaging, hyperspectral
techniques, gas chromatography [3]. Hyperspectral techniques are used by farm-
ers on a large scale and are one of the most robust and rapid methods to detect
plant diseases [11].
Nanoparticles providing high electronic and optical properties are also under
the research which can provide robust detection by means of biosensors, also it
can be used to treat the diseases as well. These methods involve bio-recognition
elements like DNA, enzymes, antibodies, etc. [3].
In recent times, researches in Artificial Intelligence and Deep Learning are
also carried out to solve agriculture-related problems. A number of solutions
are proposed for the plant disease detection that involves the use of Artificial
Neural Networks (ANN), Support Vector Machines (SVM), K-means Clustering.
A LeNet architecture is used by [1] to efficiently detect diseases of Banana such
as Banana Sigatoka, Banana Speckle. [5] Disease detection in Tomato using three
main detectors Faster Region-based Convolutional Neural Network (Faster R-
CNN), Region-based Fully Convolutional Network (R-FCN), and Single Shot
Multibox Detector (SSD) is explained by [5]. [8] shows the results of their two
proposed architectures VGG-FCN-VD16 and VGG-FCN-S on the Wheat Disease
Database 2017. They received an accuracy of 97.95% and 95.12% with those
architectures respectively. A 9 layer CNN was trained with the help of the images
from the Plant Village dataset in [6]. CNN is also used by [9] to detect diseases
of the rice plants using 500 natural images collected from rice experimental
field. Almost all of these papers have tried to solve the problem using a similar
approach. This approach for disease detection is a four-step process. 1) Image
Acquisition, 2) Image preprocessing, 3) Feature extraction and at last, passing
the image through 4) Recognition and Identification techniques.
The performance of various CNN model on the database of 87,848 images
of healthy and unhealthy plant leaves, consisting of 25 different plants in a set
of 58 distinct classes was investigated by [4]. The various architectures such as
AlexNet, AlexNetOWTBn, VGG, GoogLeNet, Over-feat with parameters such
as batches, epochs, size, momentum, weight decay, and learning rate are exam-
ined. The most successful rates were 99.53% observed in VGG.
Deploying a convolution neural network on hardware was demonstrated by
[2]. Two different deep learning architectures were used which were AlexNet
and SqueezNet for detecting unhealthy leaves in tomato plants in fields. These
architectures were trained and tested on the tomato images from the Plant
Village dataset. The training and validation of both networks were done on
hardware by Nvidia namely Nvidia Jetson Tx1. In this work trained models are
tested on the validation set with using GPU.
A new model named INAR-SSD that uses deep CNN along with GoogLeNet
Inception and Rainbow concatenation to detect apple leaf disease is proposed by
[7]. The CNN model mentioned in the paper [7] was trained with 26,377 images
of a plant leaf with different diseases. The results showed a high-performance
solution with 78.80% mAP.
270 J. Trivedi et al.

Optimising the algorithms by reducing the number of parameters and

to improve the classification accuracy of maize leaf disease using improved
GoogLeNet and Cifar10 models was aimed in [15]. These 2 models test and train
nine kinds of maize leaf images by changing a few parameters such as adding
dropouts to reduce overfitting, altering the pooling combination and some linear
functions. In this process, both the models were able to give an average classi-
fication accuracy of 98.8%. It also shows that altering the parameters like the
addition of activation function, dropout layers with proper parameters, and few
pooling combinations can improve the accuracy of the model.
Convolutional Neural Network and Learning Vector Quantization (LVQ) on
500 images of tomato leaves classified in 4 different disease of tomato with major
focus on the color of the leaf is used by [13]. The average accuracy that they
achieved on tomato leaves dataset is around 86.0%. A few variations in the LeNet
is implemented by [14] on a dateset of 18100 tomato images with 10 classes. The
validation accuracy achieved here was 94.8%.
This paper address the issue of crop disease detection using similar approach
and have employed a Convolutional Neural Network for the disease identification
task with higher accuracy.

3 Methods and Materials

The proposed CNN architecture for crop disease detection using leaf images
of different plants is discussed in detail here. The various stages included in
the process are discussed in the subsections from collecting the images to the
classification of the images.

3.1 Data Pre-processing

The images in the dataset comprises of the healthy and diseased plant leaf images
from the Plantvillage Dataset. Few samples of the leaf images from the dataset
are shown in Fig. 1. The dataset for classification consists of a total of 54,305
images which are divided into 38 different classes of healthy and diseased plant
leaf images, which are identified by the given label for each class [6]. Table 1
lists the classes of the plants and related diseases with number of images in each
category. The image transformation is done in order to decrease the size of the
image to limit the parameters of the CNN network. The original images are of
the size 256 × 256, which is reduced to a size of 64 × 64, both the sizes being in
the RGB channel.

3.2 Training the CNN Network

The Convolutional Neural Network plays a significant role in feature extraction
from the images which helps in the process of classification. Initially, the pro-
posed CNN architecture is trained with images from the dataset mentioned in
the above Sect. 3.1.
 Plant Leaf Disease Detection Using Machine Learning 271

Table 1. Classes of plant leaf disease dataset

Class name Number of images

Apple with scab 630
Apple with Black Rot 621
Apple with Cedar Apple Rust 275
Healthy Apple 1645
Healthy Blueberry 1502
Cherry with Powdery Mildew 1052
Healthy Cherry 854
Corn with grey leaf spot 513
Corn with common rust 1192
Corn with northern leaf blight 985
Healthy Corn 1162
Grape with black rot 1180
Grape with black measles 1383
Grape with leaf blight 1076
Healthy grape 423
Orange with Huanglongbing 5507
Peach with bacterial spot 2297
Healthy peach 360
Pepper with bacterial spot 997
Healthy pepper 1478
Potato with early blight 1000
Healthy potato 152
Potato with late blight 1000
Healthy raspberry 371
Healthy soybean 5090
Squash with powdery mildew 1835
Healthy strawberry 456
Strawberry with leaf scorch 1109
Tomato with bacterial spot 2127
Tomato with early blight 1000
Healthy tomato 1591
Tomato with late blight 1909
Tomato with leaf mold 952
Tomato with septoria leaf spot 1771
Tomato with two spotted spider mite 1676
Tomato with target spot 1404
Tomato with mosaic virus 373
Tomato with yellow leaf curl virus 5357
272 J. Trivedi et al.

(a) (b)

(c) (d)

(e) (f)

Fig. 1. Sample images of the dataset: (a) Apple with Black Rot. (b) Apple with Scab.
(c) Corn with Leaf Blight. (d) Healthy Cherry. (e) Tomato Healthy. (f) Tomato Leaf
with Yellow Curl Disease. (Color figure online)

The proposed convolutional network comprises of various layers in which dif-

ferent convolutions take place. The layers generate various forms of the training
data which become deeper with the layers in the network. Initially, the convo-
lutional layers work as feature extractors and the dimensionality of the image
is reduced by the pooling layers. The convolutional layers being the fundamen-
tal building blocks of the network, they extract various lower level features into
additional distinct features.
The pooling layer of the network is used to reduce the dimensions of the
images, also known as down-sampling. The max-pooling is used in the proposed
architecture for downsampling the input image. Another important layer used in
architecture is the dropout layer. It used as a regularization technique to prevent
overfitting. Finally, the dense layer classifies the image using the output of con-
volutional and pooling layers. There are other hyperparameters that determine
the performance of the neural network like an epoch, which is a hyperparameter
 Plant Leaf Disease Detection Using Machine Learning 273

Fig. 2. Layers of proposed CNN architecture

in which all the images are once passed individually forward and backward to
the network, Batch size, which is the number of training images in the forward
or backward pass and the learning rate which determines the rate at which the
weights of the parameters of the network get updated.
In the proposed CNN architecture for plant leaf disease classification, the
RGB images of leaves of size 64 × 64 are given as an input. The first layer
consists of a convolutional layer with 32 filters each of size 3 × 3 followed by a
max-pool layer of size 2 × 2. The second layer is also a convolutional layer with
64 filters each of size 3 × 3 followed by a max-pool layer of filter size 2 × 2. The
third layer is again a convloutional layer but with 128 filters each of size 3 × 3
and again followed by a max-pool layer of filter size 2 × 2. The 3D structure is
then flattened to a 1D structure of 8192 parameters which is then followed by a
fully connected layer of size 1 × 128 which classifies the input in one of the 38
categories. The visual representation of the layers is shown in Fig. 2.
The model was trained using the optimized hyperparameters which are men-
tion in Table 2 and the dataset mentioned in the Sect. 3.1.

Table 2. Hyper parameters of the architecture

Name of the parameter Value

Training Epochs 10
Batch size 64
Dropout after first layer 0.1
Dropout after second layer 0.3
Dropout after third layer 0.2
Learning Rate 0.1
Training set size 42,242
Validation set size 10,561
274 J. Trivedi et al.

4 Results
4.1 Dataset Details

The RGB images used for training the proposed CNN Architecture are taken
from the Plant Village Dataset [6] which consists of a total of 38 classes of
unhealthy and healthy plant leaf images with each image of size 256 × 256. The
total images of the plant leaves in the dataset are 54,305. A total number of 14
different types of plants are covered by the Plant Village Dataset.

4.2 Data Pre-processing

The data pre-processing plays a vital role in determining the accuracy of the
proposed architecture and time of training. The size of an RGB image in the
Plant Village Dataset is 256 × 256 which is then reduced to 64 × 64 before
feeding the image into the architecture, to decrease the number of parameters
of architecture and time of training of the architecture.

4.3 Simulation Results

The CNN architecture proposed in this paper was trained on a total of 42,242
plant leaf disease images and was validated on 10,561 images. Later the architec-
ture was tested on 441 images unknown to the architecture. As mentioned the
images are to be classified into 38 different classes of leaf diseases. The results
we obtained were compared with the results of the architecture mentioned in
[6]. Table 3 presents the comparison of classification accuracy by proposed CNN
architecture and that by [6].

Table 3. Comparison of results

CNN architecture Validation accuracy Validation loss

Model mentioned in [6] 91.43% 0.4797
Our architecture 95.81% 0.1366

The validation loss observed in the first epoch was high at around 0.5665
and decreased to 0.1366 in the last epoch. The same issue was addressed by [6]
who proposed a CNN architecture with different sets of parameters. The input
it accepts is 128 × 128 × 32 which is large as compared to the size input image
used in the proposed architecture. The validation accuracy obtained by [6] on
the same dataset as used in the present paper came out to be 91.43% and the
validation loss obtained is 0.4797, while the architecture proposed in this paper
gave a validation accuracy of 95.81% and loss of 0.1366 for disease classification
from leaf images.
 Plant Leaf Disease Detection Using Machine Learning 275

Fig. 3. Plot of validation accuracy v/s Epochs

Fig. 4. Plot of validation loss v/s Epochs

Figure 3 gives the plot of Validation Accuracy vs number of Epochs and

from this figure it is observed that accuracy increases with the epochs. Figure 4
presents the plot of Validation Loss vs number of Epochs and it is observed that
the Validation Loss decreases with increase in number of epochs.

5 Conclusion

Today, most of the research work is focused on solving the common problems
using various techniques of Machine Learning. In this paper, a solution to the
most common problem of untimely or wrongly identified crop diseases faced by
Indian farmers and the agricultural sector is discussed.
Deep Learning is widely used in the field of image classification and pattern
recognition. The proposed CNN architecture helps to classify 38 different classes
276 J. Trivedi et al.

of healthy and diseased plant leaves effectively. The CNN architecture with opti-
mized parameters gave us a validation accuracy of 95.81%. It can be seen that
convolutional neural networks provides a high-end performance and hence they
are highly suitable for automated detection of the plant diseases through simple
plant leaf images. Also, the hyperparameters play a great deal of role in deter-
mining the accuracy of the architecture. Alteration in the parameters can bring
up changes to the accuracy of the architecture.

References
1. Amara, J., Bouaziz, B., Algergawy, A., et al.: A deep learning-based approach for
banana leaf diseases classification. In: BTW (Workshops), pp. 79–88 (2017)
2. Durmuş, H., Güneş, E.O., Kırcı, M.: Disease detection on the leaves of the tomato
plants by using deep learning. In: 2017 6th International Conference on Agro-
Geoinformatics, pp. 1–5. IEEE (2017)
3. Fang, Y., Ramasamy, R.P.: Current and prospective methods for plant disease
detection. Biosensors 5(3), 537–561 (2015)
4. Ferentinos, K.P.: Deep learning models for plant disease detection and diagnosis.
Comput. Electr. Agri. 145, 311–318 (2018)
5. Fuentes, A., Yoon, S., Kim, S., Park, D.: A robust deep-learning-based detector for
real-time tomato plant diseases and pests recognition. Sensors 17(9), 2022 (2017)
6. Geetharamani, G., Pandian, A.: Identification of plant leaf diseases using a nine-
layer deep convolutional neural network. Comput. Electr. Eng. 76, 323–338 (2019)
7. Jiang, P., Chen, Y., Liu, B., He, D., Liang, C.: Real-time detection of apple leaf
diseases using deep learning approach based on improved convolutional neural
networks. IEEE Access 7, 59069–59080 (2019)
8. Lu, J., Hu, J., Zhao, G., Mei, F., Zhang, C.: An in-field automatic wheat disease
diagnosis system. Comput. Electron. Agric. 142, 369–379 (2017)
9. Lu, Y., Yi, S., Zeng, N., Liu, Y., Zhang, Y.: Identification of rice diseases using
deep convolutional neural networks. Neurocomputing 267, 378–384 (2017)
10. Nikola, M., Trendov, S.V., Zeng, M.: Digital technologies in agriculture and rural
areas briefing paper (2019). http://www.fao.org/3/ca4887en/ca4887en.pdf
11. Patil, J.K., Kumar, R.: Advances in image processing for detection of plant dis-
eases. J. Adv. Bioinform. Appl. Res. 2(2), 135–141 (2011)
12. Shurtleff, M.C., Pelczar, M.J., Kelman, A., Pelczar, R.M.: Plant disease. Ency-
clopædia Britannica (2019). https://www.britannica.com/science/plant-disease
13. Sardogan, M., Tuncer, A., Ozen, Y.: Plant leaf disease detection and classification
based on CNN with LVQ algorithm. In: 2018 3rd International Conference on
Computer Science and Engineering (UBMK), pp. 382–385. IEEE (2018)
14. Tm, P., Pranathi, A., SaiAshritha, K., Chittaragi, N.B., Koolagudi, S.G.: Tomato
leaf disease detection using convolutional neural networks. In: 2018 Eleventh Inter-
national Conference on Contemporary Computing (IC3), pp. 1–5. IEEE (2018)
15. Zhang, X., Qiao, Y., Meng, F., Fan, C., Zhang, M.: Identification of maize leaf dis-
eases using improved deep convolutional neural networks. IEEE Access 6, 30370–
30377 (2018)