A Powerful Plant Disease Classification based on Ensemble Learning

Noura Ouled Sihamman

, Assia Ennouni

, My Abdelouahed Sabri

and Abdellah Aarab

Computer Science Department, Faculty of Sciences Dhar el Mahraz, University Sidi Mohamed Ben Abdellah,

Fez, Morocco.

Physique Department, Faculty of Sciences Dhar el Mahraz, University Sidi Mohamed Ben Abdellah, Fez, Morocco

Keywords: Ensemble Learning, Deep Learning, Plant Disease Classification, Smart Agriculture, Image Processing.

Abstract: Intelligence in agriculture is becoming more and more necessary. Several tools have been proposed and used

to make farmers' tasks automatic. Plant disease detection is a tough and challenging step, especially for

inexperienced farmers. In several countries, the production quality is significantly reduced due to various

plant diseases, which has a negative impact on the economic sector. To reduce the impact of these diseases,

early plant disease diagnosis is seen as a way to treat them in time. In this context, several solutions using

artificial intelligence and image processing have been proposed. In this paper, we propose an Ensemble

Learning-based approach for the detection and classification of plant diseases. Thus, we propose to combine

the performance of 5 Deep Learning architectures to design a robust system for plant disease classification.

Simulation results proved the efficiency of the proposed approach, comparing it first with the results obtained

for each different DL architecture and with also other approaches from the literature.

1 INTRODUCTION

The economy of several African countries such as

Morocco depends heavily on agriculture. In order to

meet the huge market demand, countries need to

significantly increase their production. The quality of

production is essentially linked to the management of

plant diseases. Thus, if these diseases are not

diagnosed and treated in time, they will affect

production. According to the FAO (Food and

Agriculture Organization of the United Nations),

plant diseases have increased considerably in recent

years and this is due to climate change, globalization,

... Generally, plant diseases are identified by visual

examination by the farmer and the quality of the

diagnosis is strongly linked to the expertise and

professional knowledge of the farmer. This expertise

is acquired after several years of close experience

with plants and the diseases that affect them.

Plant diseases occur when external actors infect

the plant and cause changes in physiological and

biochemical behaviour. The symptoms of most plant

diseases appear on the leaves. Careful examination of

the shape, colour and even texture of plant leaves

plays a signicant role in the diagnosis of these

diseases. In order to overcome this problem, artificial

intelligence and image processing have been used to

propose systems for plant disease recognition and

detection. The aim of such applications is to use

machine learning or deep learning algorithms to

efficiently detect whether a plant is diseased or not

from a plant photo (Ennouni et al., 2021c) (Prakash et

al., 2017). The Internet of Things is also used to

design efficient systems in smart agriculture

(Muhammad et al., 2019) (Gonzalez et al., 2007).

Deep and automatic learning have become assets for

the detection and classification of medical images,

where several approaches have proven to be very

efficient (Richard et al., 2005) (Wäldchen et al.,

2018).

A large number of DL architectures can be used

to design a powerful classification model for plant

disease detection. CNN, MobileNet, AlexNet,

Inception V3, and VGG16 will be used and tested to

identify the behaviour of each architecture. Each

architecture allows a relatively different behaviour

from the others, but they are complementary. Thus, it

has been noticed that some architectures fail to

correctly classify some plants that other architectures

can classify correctly. Based on this observation, we

will propose in this paper a classification approach

based on the Learning Set that combines the power of

the five architectures. And to evaluate our proposed

approach, we will propose a comparative study

Ouled Sihamman, N., Ennouni, A., Sabri, M. and Aarab, A.

A Powerful Plant Disease Classiﬁcation based on Ensemble Learning.

DOI: 10.5220/0010733500003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 321-326

ISBN: 978-989-758-559-3

321

between a SOFT and HARD Ensemble-based

classification as well as with other approaches from

the literature.

The rest of this paper is as follows: the next

section will present plant diseases in detail. Section 3

presents the different DL architectures that we will

use to design our classification model. Then we will

present our proposed approach. The Dataset used will

be detailed as well as the evaluation measures

afterwards. The results of experiments will be

presented before concluding this paper.

2 PLANT DISEASES

Plants are living beings, the majority of which are

attached to the earth by their roots. This characteristic

makes that the behaviour of these plants is strictly

related to their environment. Plants are subject to

various diseases and these diseases can be grouped by

regions (El-Sayed et al., 2020). Moreover, as for

humans, diseases that can affect plants disrupt and, or

modify their vital functions. The different diseases

that can affect plants can be categorized into three

classes: fungal, bacterial, or viral (El-Sayed et al.,

2020) (Vijai, 2020).

 Viral diseases: Viral diseases are transmitted to

plants mainly by insects or worms. Typical

symptoms of viral diseases include mosaic

patterns, yellowing, stripes, and leaf rolling

(Fig. 1).

Figure 1: Necrotic spot virus on chilli leaves.

 Fungal diseases: Fungal diseases are diseases

that are harmful to plants but not with great

risk. Fungi, mould, mildew, and others cause

these diseases. The symptoms of these diseases

are leaf rust, especially for corn, Stem rust,

white mould (Sclerotinia). Figure 2 shows an

example of a plant affected by Sclerotinia

fungus.

Figure 2: Sclerotinia Infected by Soybeans.

 Bacterial diseases: Plant infections of bacterial

origin have almost the same symptoms as

fungal diseases. They can be identified by the

presence of rot, scab, scorch, wilt and leaf

spots. These types of diseases, if not treated in

time, can cause serious and disastrous diseases

(Sujeet et al., 2016). Two examples of bacterial

diseases are shown in figure 3.

Figure 3: Bacterial strands on stems and Bacterial blight of

wheat leaves.

The three different categories of diseases

presented below are identifiable based on

specifications and characterizations extracted after

visualization of plants, stems, and leaves. Therefore,

one can use image processing algorithms in

combination with artificial intelligence to develop a

robust system for automatic detection of plant

diseases.

3 DEEP LEARNING BASED

APPROACHES FOR PLANT

DISEASES CLASSIFICATION

Automatic images classification is an area of research

that is very focused in recent years. A distinction is

made between supervised classification,

unsupervised classification and reinforcement

classification. In supervised image classification,

machine learning, deep learning and image

processing have been widely used in recent years. The

main objective is to design a powerful classification

model capable of predicting the class of a new image.

In machine learning we proceed to a primordial step

to extract relevant features before using them as input

for a classification algorithm (Ennouni et al., 2021a).

In contrast, in deep learning, the image is generally

used directly as input to design a classification model.

Several DL architectures have been proposed for the

case of supervised image classification. Most of these

approaches are based on convolutional neural

networks (CNN).

The deep learning architecture CNN, also called

ConvNet, is generally composed of a sequence of

convolution layers, pooling layers (Max, min, …

pooling), and at the end one or more fully connected

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

322

layers. The difference between the Neural Network

Deep Learning architectures is essentially in the

prearrangement of the neural network components.

Also, for a given architecture we should fixe a

number of parameters (Kafi et al., 2015) (Shin et al.,

2016):

- The number of layers

- The number of neurons in each layer

- The filters mask size

- The neurons weight

- The activation function

- The learning rate

- ....

The following table lists the most used

architectures from the literature the authors, the year,

the number of parameters and the depth.

Table 1: The most successful CNN architectures.

Reference Parameters depth

LeNet-5 (LeCun, 1998) 60,000 5

AlexNet (Krizhevsky, 2012) 60 M 8

VGG

(

Simon

an, 2016

)

138 M 19

Goo

leNet

(

Sze

, 2015

)

4 M 22

Inception V3 (Szegedy, 2015) 23 M 159

ResNet (He, 2016) 25 M 152

MobileNet V2 (Sandler, 2019) 3.47 M 53

4 ENSEMBLE LEARNING-

BASED APPROACHES FOR

PLANT DISEASE

CLASSIFICATION

Several classification algorithms can be used for plant

disease classification. Each classification algorithm

has strengths and weaknesses. Also, it is often noticed

that diseases are misclassified by some classifiers

while another has correctly classified them and vice

versa. Based on this principle, it is possible to

combine the performance of a set of classification

algorithms to obtain a single robust classification

model (Yawen et al., 2018). The Ensemble Learning

is a meta-classifier allowing to combine similar or

conceptually different machine learning classifiers

(by majority or plurality voting) (Sabri et al., 2020).

Ensemble Learning is often used in professional

applications and has helped win several competitions.

Several Ensemble Learning techniques can be

used (Mao et al., 2019):

1- Hard voting: The predicted label is defined as

the class label most frequently predicted by the

classification models.

2- Soft voting: This type of technique can be

divided into two sub-elements:

a. Progressive voting: we predict the class

labels by averaging the class probabilities

(recommended only if the classifiers are

well calibrated).

b. Weighted progressive voting: this is an

extension of progressive voting where we

assign different weights to the classifiers

used. These weights are defined according

to the importance and quality of the

classifier.

3- Bagging: Bagging consists of creating random

samples of the training dataset with substitution

(subsets of the training dataset) and then

applying the same classifier (Used usually with

Decision Tree) for each sample and applying a

hard vote.

4- Boosting: Is considered as progressive voting

with adaptation of the weights of the classifiers

in an iterative way.

4.1 Majority and Hard Voting

Majority voting is a widely used approach that

combines many classification algorithms where each

algorithm makes its prediction and the predicted class

is the one obtained by majority voting

ỹ  𝑚𝑜𝑑𝑒 𝐶





𝑥



,𝐶





𝑥



,…,𝐶





𝑥





where ỹ is the predicted class and Cj (i= 1 to

m) is the j classifier

4.2 Weighted Majority Voting

Weighted voting is a special case of the majority

voting where each prediction algorithm is used with a

predefined weight.

ỹ  𝑎𝑟𝑔𝑚𝑎𝑥 𝑤



𝐶



𝑥





4.3 Soft Voting

Soft voting is also a special case of majority voting

where each prediction algorithm is used with a

probability p. This approach is only effective if the

classification algorithms are calibrated.

ỹ𝑤



𝑝







where wj is the weight assigned to the jth classifier.

A Powerful Plant Disease Classiﬁcation based on Ensemble Learning

323

5 THE PROPOSED APPROACH

The objective of this work is to propose a powerful

approach for plant disease classification based on

Ensemble Learning. Indeed, we have opted in a first

step to the design of a classification model of plant

diseases using 5 Deep Learning architectures,

namely: VGG16, AlexNet, Inception V3, CNN and

MobileNet.

When we design classification models using the

five architectures distinctively. Each of the 5

classification models have different behaviour. We

noticed after a judicious study of the classification

errors for each of the five architectures that they are

complementary. Thus, we noticed that for each image

in the database at least one of the five architectures

manages to classify it correctly. From there, we

thought of combining the performances of the five

architectures in an Ensemble Learning architecture to

propose a powerful approach for the classification of

plant diseases.

Figure 4 presents the proposed approach



Figure 4: The proposed approach.

We will implement the two Ensemble Learning

techniques; the Hard Voting approach and the Soft

Voting approach. We will compare these two

approaches and also compare them with recent

approaches in the literature.

6 DATASET AND EVALUATION

6.1 Plant Village Dataset Description

The Plant Village (PV) is a well-known dataset used

to evaluate the plant disease classification algorithms.

It contains infected and healthy crop leaves images.

The dataset contains 87K RGB images classified into

38 subsets. The dataset is provided with training

dataset containing 70295 images and validation

dataset with 17572 images (Ennouni et al., 2021b).

6.2 Performance Measures

In images classification, precision, recall and

accuracy are the most evaluations measures used. We

evaluating multiclass classification algorithm, their

values are the mean of all the values for each class:

𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 

∑











(1)

𝑟𝑒𝑐𝑎𝑙𝑙 

∑











(2)

𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 

∑











(3)

𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛

















(4)

𝑟𝑒𝑐𝑎𝑙𝑙

















(5)

𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦





























(6)

where n is the number of classes, i the current class,

TP is The True Positive, FN is the False Negatives,

TN is the True Negatives and FP is the False Positives

numbers.

7 SIMULATION RESULTS

All simulations in this article were performed on an

Intel Core i7 3.6 GHz processor, 16 GB of RAM and

a Windows 10 operating system.

In our study, we will test each of the five DL

architectures to evaluate the behaviour of each one.

Then, we will test the two Ensemble learning

proposed approaches. For the CNN architecture, we

will use three layers depth. And for the VGG16,

AlexNet, Inception V3 and MobileNet the number of

epochs is 25.

Table 2 presents the accuracy of the five DL

architectures, the accuracy of the Soft Ensemble

Learning and the accuracy of the hard ensemble

learning

Table 2: Classification results of the proposed approach.

Precision Recall Accuracy

VGG16 0.89 0.96 0.92

AlexNet 0.86 0.96 0.95

Ince

tion V3 0.89 0.95 0.91

CNN 0.95 0.95 0.95

MobileNet 0.94 0.97 0.95

Soft EL 0.95 0.96 0.96

Hard EL 0.99 0.99 0.99

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

324

We notice from the classification results presented in

Table 1 that each of the five architectures has a

different behaviour and with different values of

precision, recall and accuracy. We notice that the

Inception V3 architecture had the lowest values in

terms of accuracy followed by AlexNet. The

AlexNet, CNN and MobileNet architectures had the

same accuracy score. But in terms of Recall we can

say that the MobileNet architecture is the best.

However, the proposed Ensemble Learning

approaches have significantly improved the

classification results and especially using the HARD

technique with a value equal to 0.99. We can say that

the use of the vote of the 5 architectures allowed us to

have almost 100% of classification rate.

To evaluate our proposed approach in regards to

the most relevant approaches from literature, we

conducted a comparative study. Table 3 presents the

classification results of our proposed approach and

the approaches proposed by (Srdjan, 2016) and

(Sibiya, 2019). in (Srdjan, 2016) authors proposed to

use Deep CNN. Authors in (Sibiya, 2019) used also

CNN with 50 hidden layers.

Table 3: A comparison results of the classification of the

approaches we propose with recent approaches.

(Srdjan, 2016) (Sibiya, 2019). Hard EL

Accuracy (%) 94.60 95.81 99.21

As stated at the outset, the Ensemble Learning has a

higher classification accuracy than the other

approaches. Although the work presented in [25] uses

a large number of layers, it is still less efficient than

our proposed approach.

8 CONCLUSION

Ensemble learning consists in combining a number of

classification models in order to make them

complementary in terms of classification accuracy.

Thus, assuming that at least one of the classification

models can correctly classify an image that the other

classification models have misclassified, the

objective of this work is to propose an ensemble

learning classification approach based on 5 Deep

Learning architectures namely; VGG16, AlexNet,

Inception V3, CNN and MobileNet. We have

implemented two techniques of ensemble learning;

SOFT and HARD. The ensemble learning using

HARD allowed us to significantly improve the

classification rate of plant diseases. With a rate of

almost 100% we can say that our proposed approach

can be effectively used in smart agriculture to

automatically classify plant diseases.

REFERENCES

El-Sayed, A.M.A., Rida, S.Z., Gaber,Y.A. : Dynamical of

curative and preventive treat-ments in a two-stage plant

disease model of fractional order, Chaos, Solitons &

Fractals, Volume 137, 109879, ISSN 0960-0779,

(2020).

Ennouni A., Sabri M.A., Aarab A. (2021) Plant Diseases

Detection and Classification Based on Image

Processing and Machine Learning. In: Gherabi N.,

Kacprzyk J. (eds) Intelligent Systems in Big Data,

Semantic Web and Machine Learning. Advances in

Intelligent Systems and Computing, vol 1344. Springer,

Cham. https://doi.org/10.1007/978-3-030-72588-4_20

Ennouni A., Sihamman N.O., Sabri M.A., Aarab A. (2021)

Analysis and Classification of Plant Diseases Based on

Deep Learning. In: Motahhir S., Bossoufi B. (eds)

Digital Technologies and Applications. ICDTA 2021.

Lecture Notes in Networks and Systems, vol 211.

Springer, Cham. https://doi.org/10.1007/978-3-030-

73882-2_12.

Ennouni, A., Sihamman, N. O., Sabri, M. A. & Aarab, A.

(2021). Early Detection and Classification Approach

for Plant Diseases based on MultiScale Image

Decomposition. Journal of Computer Science, 17(3),

284-295. https://doi.org/10.3844/jcssp.2021.284.295

Gonzalez RC, Woods RE (2007) Digital image processing,

3rd edn. Prentice Hall.

He K, Zhang X, Ren S, Sun J : Deep Residual Learning for

Image Recognition. Multimed Tools Appl 77:10437–

10453, 2015.

Kafi M, Maleki M, Davoodian N : Functional histology of

the ovarian follicles as deter-mined by follicular fluid

concentrations of steroids and IGF-1 in Camelus

dromedarius. Res Vet Sci 99:37–40. (2015).

Krizhevsky A, Sutskever I, Hinton GE. : ImageNet

Classification with Deep Convolutional Neural

Networks. Adv Neural Inf Process Syst 1–9, (2012).

LeCun Y., Bottou L., Bengio Y, : Haffner P Gradient-based

learning applied to document recognition. Proc IEEE

86:2278–2324, (1998).

Mao, S., Lin, W., Jiao, L., Gou, S., & Chen, J.-W. End-to-

End Ensemble Learning by Exploiting the Correlation

Between Individuals and Weights. IEEE Transactions

on Cybernetics, 1–12. doi:10.1109/tcyb.2019.2931071.

(2019).

Muhammad Hammad Saleem, Johan Potgieter and Khalid

Mahmood Arif. "Plant Disease Detection and

Classification by Deep Learning". Plants 2019, Vol. 8,

Issue 11, 468. doi:10.3390/plants8110468.

Richard N Strange and Peter RScott. 2005. Plant disease:a

threat to global food security. Annu. Rev.

Phytopathol.43(2005),83–116.

Prakash, R. M. Saraswathy, P. G. Ramalakshmi, K. H.

Mangaleswari and T. Kaviya.: Detection of leaf

A Powerful Plant Disease Classiﬁcation based on Ensemble Learning

325

diseases and classification using digital image

processing, 2017 International Conference on

Innovations in Information, Embedded and

Communication Systems (ICIIECS), pp. 1-4,

Coimbatore (2017).

Sabri, M. A., Filali Y., Khoukhi , H. and Aarab, A.:Skin

Cancer Diagnosis Using an Im-proved Ensemble

Machine Learning model, 2020 International

Conference on Intelligent Systems and Computer

Vision (ISCV),pp. 1-5, Fez, Morocco (2020).

Sandler Mark, Andrew Howard, Menglong Zhu, Andrey

Zhmoginov, Liang-Chieh Chen. MobileNetV2:

Inverted Residuals and Linear Bottlenecks.

arXiv:1801.04381,(2019).

Shin H-CC, Roth HR, Gao M, et al Deep Convolutional

Neural Networks for Comput-erAided Detection: CNN

Architectures, Dataset Characteristics and Transfer

Learning. IEEE Trans Med Imaging 35:1285–1298,

(2016).

Sibiya M, Sumbwanyambe M (2019) A computational

procedure for the recognition and classification of

maize leaf disease out of health leaves using

convolutional neural networks. Agric Eng 1:119–131

Simonyan K, Zisserman: A Very deep convolutional

networks for large-scale image recog-nition. ICLR

75:398–406,(2015).

Srdjan Sladojevic, Marko Arsenovic, Andras Anderla,

Dubravko Culibrk, Darko Stefanovic, "Deep Neural

Networks Based Recognition of Plant Diseases by Leaf

Image Classification", Computational Intelligence and

Neuroscience, vol. 2016, Article ID 3289801, 11 pages,

2016. https://doi.org/10.1155/2016/3289801.

Sujeet V and Tarun D. : A Novel Approach for the

Detection of Plant Diseases International Journal of

Computer Science and Mobile Computing Vol. 5 Iss. 7

p. 44 - 54 ISSN: 2320–088X, (2016).

Szegedy C, Vanhoucke V, Ioffe S, et al : Rethinking the

Inception Architecture for Comput-er Vision. In:

Proceedings of the IEEE Computer Society Conference

on Computer Vision and Pattern Recognition. IEEE, pp

2818–2826, (2016).

Szegedy C, Wei Liu, Yangqing Jia, et al : Going deeper

with convolutions. In: 2015 IEEE Conference on

Computer Vision and Pattern Recognition (CVPR).

IEEE, pp 1–9, (2015).

Vijai S,. Namita S., Shikha S.: A review of imaging

techniques for plant disease detection, Artificial

Intelligence in Agriculture, Volume 4, Pages 229-242,

ISSN 2589-7217, (2020).

Wäldchen, J., Mäder, P. Plant Species Identification Using

Computer Vision Techniques: A Systematic Literature

Review. Arch Computat Methods 25, 507-543 (2018).

https://doi.org/10.1007/s11831-016-9206-z

Yawen Xiao, Jun Wu, Zongli Lin, Xiaodong Zhao. A deep

learning-based multi-model ensemble method for

cancer prediction. Computer Methods and Programs in

Biomedicine. Volume 153, Pages 1-9. ISSN 0169-

2607. https://doi.org/10.1016/j.cmpb.2017.09.005,

(2018).

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

326