MONITORING OF HYDROPONICS

PLANT AND PREDICTION OF

LEAF DISEAES USING IOT

PROJECT REPORT

Submitted by

DEEPIKA V 17BEC017

DHIVYAARTHI G 17BEC088

GOWTHAM K 17BEC130

In partial fulfilment for the award of the degree

of

BACHELOR OF ENGINEERING

in

ELECTRONICS AND COMMUNICATION ENGINEERING

KUMARAGURU COLLEGE OF TECHNOLOGY

(An Autonomous Institution Affiliated to Anna University, Chennai)

MAY 2021

1
 KUMARAGURU COLLEGE OF TECHNOLOGY

COIMBATORE

BONAFIDE CERTIFICATE

Certified that this Report titled “MONITORING OF

HYDROPONICS PLANT AND PREDICTION OF LEAF DISEASE
USING IOT” is the bonafide work of “DEEPIKA V [17BEC017],
DHIVYAARTHI G [17BEC088], Mr GOWTHAM K [17BEC130]”, who
carried out the project work under my supervision during the 8th semester.

SIGNATURE SIGNATURE OF SUPERVISOR

Dr S.A. PASUPATHY A. KALAISELVI
PROJECT SUPERVISOR
Professor & Head Assistant Professor II,
Department of ECE Department of ECE
Kumaraguru College of Technology, Kumaraguru College of Technology
Coimbatore - 641 049 Coimbatore - 641 049

Submitted for the viva-voice examination held at KUMARAGURU COLLEGE

OF TECHNOLOGY, held on 24/05/2021

Internal Examiner External Examiner

2
 ACKNOWLEDGEMENT

First, we would like to express our praise and gratitude to the Lord, who has
showered his grace and blessings enabling us to complete this project in an
excellent manner.
We express our sincere thanks to the management of Kumaraguru
College of Technology and Joint Correspondent Shri. Shankar Vanavarayar,
for his kind support and for providing necessary facilities to carry out the work.
We would like to express our sincere thanks to our beloved Principal Dr.
D. Saravanan Ph.D., Kumaraguru College of Technology, who encouraged us
with his valuable.
We would like to thank Prof [Link] M.E.,PhD (UK) Head
of the Department, for his kind support and for providing necessary facilities to
carry out the project work.
We wish to thank everlasting gratitude to the project coordinator Mr. Timothy
D Paul, ME., (Ph.D.), Department of Electronics and Communication
Engineering, for her consistent support throughout the course of this project
work.
We are greatly privileged to express our heartfelt thanks to our project
guide Mrs Kalaiselvi A, Assistant Professor II, Department of Electronics and
Communication Engineering, for his expert counselling and guidance to make
this project to a great deal of success and we wish to convey our deep sense of
gratitude to all teaching and non-teaching faculties of ECE Department for their
help and cooperation.
Finally, we thank our parents and family members for giving us the moral
support and abundant blessings in all of our activities and friends who helped us
to endure our difficult times with their unfailing support and warm wishes.

3
 ABSTRACT

In India, numerous People are doing the agribusiness, and it is a spine of our
country, in farming there are numerous issues like little land property,
excrements, pesticides, synthetic compounds utilized by rancher for horticulture
and so on, in addition Clients requesting for the quality food diet which is
acceptable in quality also, liberated from horticultural synthetic compounds and
pesticides. Our proposed framework is satisfying the above necessities and we
are developing the plants naturally. It is feasible to controlled climate, in any
spot like room porch, overhang and so on Moreover; we are developing huge
number of plants around there. This is one of the sorts of agribusiness which
can be more profitable whenever observed, controlled effectively. Hydroponics
framework is a developing plant framework that doesn't utilize soil however
utilizes water with supplements to save space planting and not debased with
synthetic substances in the dirt. The computerized aqua-farming framework
dependent on IOT was created to work with the development and the
framework can change and control significant ecological components that
influence plant development including temperature, humidity, water and
recognition of illness in the leaf. The objective of this venture is to keep up the
Hydroponic plant with required measure of temperature, humidity, dampness
and to identify the status of the leaves in the plant. For, detecting the infection
in the leaf we use Convolutional Neural Networks (CNN) algorithm. These
Parameters are at that point, moved to the Cloud utilizing the Raspberry Pi3
Board.

4
 TABLE OF CONTENTS

CHAPTER TITLE PAGE NUMBER

1 INTRODUCTION 9
1.1 Plant Leaf Diseases 9
1.2 Hydroponics Plants 9
1.3 Types of diseases 10
1.4 Statistics 11
1.5 Objective 11
2 LITERATURE REVIEW 13
3 METHODOLOGY 17
3.1 Data Base 17
3.2 Deep Learning 18
3.3 Transfer Learning 20
3.3.1 Working of Transfer 21
Learning
3.3.2 Advantages of Transfer 21
Learning
3.3.3 Approaches to Transfer 22
Learning
a) Training a model to reuse it 23
b) Using a pre-trained model 23
c) Feature extraction 23
3.4 Convolutional Neural Network 24
3.4.1 Input layer 25
3.4.2 Convolutional layer (CONV) 25
3.4.3 Pooling layer (POOL) 26
3.4.4 Fully connected layer 26
3.4.5 Output Layer 27
3.4.6 RELU 27
3.4.7 ResNet 27
a) Advantages of ResNet 28

3.5 VGG 29
3.5.1 VGG-16 29
3.5.2 Advantages of VGG-16 29
3.5.3 VGG-19 30
3.5.4 Advantages of VGG-19 31
3.5.5 Inception V3 31
3.6 Google Colaboratory 32

5
4 Hardware Requirements 32
4.1 Raspberry Pi3 32
4.2 DTH11 Sensor 33
4.3 pH Sensor 34
4.4 Thonny Python IDE 35
4.5 Block Diagram 36

5 Working and Setup 36 36

5.1 Working 36 36
5.2 Environmental Setup 37 37

6 Results & Discussions 38

6.1 Results 38
6.2 Graph & Output 38
6.2.1 CNN with VGG-16 38
6.2.2 CNN with VGG-19 40
6.2.3 CNN with ResNet 41
6.2.4 CNN with Inception V3 42
6.2.5 IDE &Web Results 43
6.2.6 Discussions 43

7 Conclusions & Future Scope 44

7.1 Conclusion 44
7.2 Inference 44
7.3 Future Scope 44

6
 TABLE OF FIGURES

CHAPTER TITLE PAGE NUMBER

1 1.1 Bacterial Spot 11

1.2 Tomato Mosaic Virus 11
1.3 Fungal White Spot 11
1.4 Pythium Plant 11

3 3.1 Tomato Mosaic Virus 18

3.2 Healthy Plants 18
3.3 Bacterial Spot 18
3.4 Late blight 18
3.5 Leaf Mould 19
3.6 Flowchart of CNN 19
3.6 Simple Deep Learning Network 21
3.7 Transfer Learning 22
3.8 Feature Extraction 24
3.10 Layers of Convolutional Neural 25
Network 25
3.11 Schematic Diagram of ResNet 28
3.12 VGG 16 & VGG19 30

6 6.1 Output of CNN with VGG-16 38

6.2 Output of CNN with VGG-19 40
6.3 Output of CNN with ResNet 41
6.4 Output of CNN with Inception V3 42
6.5 Python IDE Outputs 43
6.6 Web Server Outputs 43

7
 TABLE OF TABLES

CHAPTER TITLE PAGE NUMBER

3 Different types of ResNet 27

LIST OF ABBREVIATIONS

ABBREVIATIONS NOMENCLATURE
CNN Convolutional Neural Network
SVM Support Vector Machine
GPU Graphics Processing Unit
ANN Artificial Neural Network

8
 CHAPTER 1

INTRODUCTION

1.1) Plant Leaf Diseases:

In India, Farmers have a splendid range of vegetation. Various pathogens are

gift within side the surroundings which severely have an effect on the
vegetation and the soil wherein the plant is planted, thereby affecting the
manufacturing of vegetation .Various diseases are discovered at the flowers and
vegetation the main identity of the affected plant or crop are its leaves. The
numerous colored spots and styles at the leaf are very beneficial in detecting the
disease. This system has no adverse effects on environment or quality on crops.
The beyond scenario for plant leaf ailment detection taken direct eye
observation, recollect the unique set of ailment as in step with the climate,
season etc. India is a cultivated country and approximately 80% of the populace
relies upon agriculture. Plant leaf ailment results in the discount in each the
exceptional and amount of agricultural products. The diseased plant leaf refers
back to the research of visually observable styles at the plants. Health of plant
leaf and ailment on plant leaf performs an critical function in successful
cultivation of plants within side the farm. In early days, evaluation of plant leaf
sicknesses had been achieved manually via way of means of the proficiency
man or woman in that discipline only. This calls for a large quantity of paintings
and additionally calls for immoderate processing time. Diseases at the plant leaf
have become a significant trouble as it is able to purpose extreme discount and
losses in each exceptional and amount of agricultural products. A vast majority
of the developing countrywide populace relies upon on agriculture yields. But
the cultivation of those plants for foremost yield and exceptional product is
tremendously technical & challenging.

1.2) Hydroponics Plants:

Hydroponics is a way of developing plants without soil. Plants are grown in

rows or on trellises, similar to in a conventional garden; however they have got
their roots in water instead of in dirt. Most folks confuse soil with nutrients. In
fact, soil gives structure, now no longer the real meals itself, for plant roots. The
9
meals come from different substances combined with inside the soil, together
with compost, broken-down plant waste or fertilizers. Plants grown
hydroponically can surely develop quicker and healthier than plant life in soil
due to the fact they do not need to combat soil borne diseases; in addition, all of
the meals and water they want are given without delay to their roots across the
clock. Hydroponics is the technological know-how of developing plant life
without soil. Learn how pretty much any plant, in particular vegetables, may be
grown the use of this method. Growing plant life hydroponically does not need
to be performed on a massive scale, and it is simpler than you may think. Now
there are kits, do-it-your self-structures or even completely computerized
developing tables, all designed for domestic gardeners. Hydroponics is quite
simple in lots of ways; it is less difficult than developing plant life in soil. Plants
want meals, water and air. When you destroy it right all the way down to the
ones 3 things, it will become easy to provide plant life most effective what they
want. Hydroponics is the technological know-how of developing plant life
without soil. The plant lives thrive at the nutrient answer alone; the medium
simply acts as a assist for the plant life and their root system.

1.3) Types of Plant Disease:

There are different types of Plant diseases that are classified by how severe the
disease is. They are Bacterial, Viral, Pythium, Fungal Etc., Plants are vulnerable
to viruses, diseases, fungi, and Pythium. Once a pathogen enters your vegetation
you‟ll observe an improvement of galls, swellings or leaf curls, yellowing or
stunted leaf, fruit, or root boom, or plant tissue die-off.

I. Bacteria: If your plant has been infected via way of means of bacteria,
you'll observe a slimy, brown-coloured coat at the plant‟s roots. It may be
visible in hydroponic reservoir‟s walls, tubes, and water pump filter.

II. Fungi: Evidence of a fungal pathogen is whilst you observe powdery or

fuzzy growths at the plant‟s roots or leaves. These growths are generally
grey, blue, white, or inexperienced colours.

III. Pythium: This kind of sickness is understood for being the maximum
lethal for hydroponic developing systems. Pythium will blacken the
plant‟s roots, ooze a foul-smelling odour, and halt the boom of the plant.

10
IV. Virus: While viruses in vegetation are rare, they may be deadly after they
do occur. Viruses are generally introduced approximately via way of
means of out of doors insects.

Fig 1.1 Bacterial Spot Fig 1.2 Tomato Mosaic Virus

Fig 1.3 Fungal White Spot Fig 1.4 Pythium Diseased Plant

1.4) STATISTICS:

The Automated Hydroponics System from C-DAC includes calculated go with

the drift of vitamins and water in a 5 layered blending tank. The answer is
pumped into the gullies in which flora are grown with appropriate assisting
cups. The roots of the flora without delay take in water and minerals.
Hydroponics is the developing of flora the use of substrates at the same time as
handing over water, vitamins and oxygen. Hydroponic develop structures are a
sustainable approach of developing with water with financial savings

11
everywhere among 70% and 90% pending at the form of plant and your
develop set-up.

1.5) OBJECTIVES

The proposed work contributions are:

 To build an effective system in cost wise, for the predicting and

classifying the Plant Leaf disease.
 To design a deep learning model this can predict and detect this disease.
 To get improved accuracy rate than the existing methods by training more
number of datasets and by using various Architectures.
 To minimise the training time duration with the help of pre-trained model
with greater accuracy rate.

12
 CHAPTER 2

LITERATURE REVIEW

1. Leaf Disease Detection Using Raspberry Pi LAYA YESUDAS 1,

SANTHIYA.S 2 , PARIMALA.R 3 , MOHAMMAD HARRIS S.

Green plants are particularly critical to the human climate; they structure the
reason for the maintainability and long haul wellbeing of ecological
frameworks. In this way it is vital to develop sound plants. The plant sickness
could be relieved in the event that it is known in the prior stage. In this paper,
we have proposed a framework utilizing raspberry pi to identify solid and
unfortunate plants and caution the rancher by sending email. We have utilized
tensor stream apparatus for mathematical calculation. It very well may be
utilized in an controlled climate ranches with the end goal that it identifies the
signs of infection at whatever point they show up on the leaves of the plant.

2. PLANT LEAF DISEASE DETECTION USING CNN AND

RASPBERRY Pi G. Rama Mohan Reddy, Nettam Sai Sumanth, N. Sai
Preetham Kumar

For recognizing sicknesses for plants we utilize bare eye. In any case, it isn't
precise so SVM is utilized for expectation which is less exact 70 % regardless
of whether dataset is good. So, this paper we use Raspberry pi for ID of the
illnesses. Here we utilize the convolutional neural organizations for preparing of
pictures. CNN utilizes different central focuses to distinguish the infections in
the plants. In research purposes it is exceptionally valuable on the grounds that
by this we can recognize infections of leaf accurately utilizing the CNN
calculation. Illness alludes to sort of the plant affected. Here we utilize the best
calculation to distinguish furthermore, foresee the plant lead illness. Effectively
distinguishing proof of infections will be extremely useful on the grounds that it
will expand the efficiency and lessen synthetics on the plants. It will increment
horticulture yield with least expense which is very accommodating now a days
and we can get precision of 90% if dataset is acceptable.

3. Monitoring of Hydroponics System using IoT Technology, Nivesh Patil1,

Shubham Patil1, Animesh Uttekar1, A. R. Suryawanshi2

13
Agriculture all throughout the planet plays significant part in the advancement
of rural countries. In India practically 68% of individuals rely on cultivating and
1/3 of the public capital comes from agrarian. Issues related horticulture have
been continually forestalling the advancement of the country. The answer for
this issue can be settled by savvy horticulture and modernizing the present
customary techniques of cultivating. Thus the point of the task is to execute
aquaculture framework utilizing IoT advances utilizing Raspberry pi. The
significant highlights of this undertaking incorporate water driven horticulture
framework that will dispense with need for soil. With this aqua-farming
computerized framework, the yields are provided with water and supplements
relying upon the sensors criticism like temperature and mugginess (DHT 11),
pH sensor and electric conductivity circuits.

4. Plant Disease Detection using CNN & Remedy, Adnan Mushtaq Ali
Karol1, Drushti Gulhane2, Tejal Chandiwade3

The proposed framework helps in distinguishing proof of plant sickness and

gives cures that can be utilized as a guard component against the illness. The
information base got from the Internet is appropriately isolated and the unique
plant species are distinguished and are renamed to shape an appropriate
information base at that point get test-data set which comprises of different
plant illnesses that are utilized for checking the exactness and certainty level of
the venture .Then utilizing preparing information we will prepare our classifier
and afterward yield will be anticipated with ideal exactness. We use
Convolution Neural Network (CNN) which contains various layers which are
utilized for expectation. A model robot model is too planned which can be
utilized for live inclusion of enormous farming fields to which a high goal
camera is appended furthermore, will catch pictures of the plants which will go
about as contribution for the product, based of which the product will advise us
if the plant is sound. With our code and preparing model we have accomplished
an exactness level of 78% .Our programming gives us the name of the plant
species with its certainty level and furthermore the cure that can be taken as a
fix.

4. How Convolutional Neural Networks Diagnose Plant Disease, Yosuke

Toda1, 2 and Fumio Okura1,3

Profound learning with convolutional neural organizations (CNNs) has made

incredible progress in the arrangement of different plant sicknesses.

14
Nonetheless, a set number of studies have clarified the interaction of deduction,
leaving it as an unapproachable black box. Uncovering the CNN to extricate the
learned element as an interpretable structure guarantees its unwavering quality
as well as empowers the approval of the model validness and the preparation
dataset by human mediation. In this examination, an assortment of neuron-wise
and layer-wise perception techniques were applied utilizing a CNN, prepared
with an openly accessible plant illness picture dataset. We showed that neural
organizations can catch the shadings and surfaces of sores explicit to individual
infections upon analysis, which looks like human dynamic. While a few
perception strategies were utilized as they are, others must be upgraded to focus
on a particular layer that completely catches the highlights to produce
noteworthy yields. Additionally, by deciphering the produced consideration
maps, we distinguished a few layers that were not adding to deduction and
eliminated such layers inside the organization, diminishing the quantity of
boundaries by 75% without influencing the characterization precision. The
outcomes give a catalyst to the CNN discovery clients in the field of plant
science to all the more likely comprehend the determination cycle and lead to
additional proficient utilization of profound learning for plant infection
conclusion.

5. Plant Disease Detection and Classification using CNN Model with

Optimized Activation Function Publisher: IEEE

This paper manages the streamlined constant recognition of illnesses that

influence the plant and the territory influenced utilizing Convolutional Neural
Networks (CNN) calculations so that proper composts can be utilized to keep
further harm to plants from pathogenic infections. The initiation work is the
centre of the CNN model as it joins the non - linearity to have a legitimate man-
made reasoning framework for order. ReLu is one among the best actuation
capacities; however has a detriment that the subsidiary of the capacity is zero
for negative qualities and prompts neuronal putrefaction. New numerical
initiation work is created and contrasted with existing enactment capacities with
improve the precision and execution of the framework on a Tensor Flow
system. The trial results on prepared data sets show that the created enactment
work has improved the CNN model exactness and execution for example 95%.
The preparation speed of the CNN model is improved by 83% when carried out
in ARM processor utilizing the proposed enhancer. Further region influenced by

15
illness is determined by utilizing K - implies bunching calculation for
advancement of manure use.

6. Plant Disease Detection and Growth Monitoring Using IoT, Anudeep

Department of Computer science and Engineering, Satyabhama Institute of
Science and Technology, Chennai, India

As we as a whole realize that horticulture assume an imperative part for

building up rural society or country. Farming has been the vocation for
individuals past a century anyway in view of natural part of climatic condition
ranchers are going through a lot of issues which may likewise also bring about
address basic issues in the general public. So IoT can be a pivotal key to assist
the ranchers with mechanizing the entire cycle and profound inclining
calculations like CNN can be executed for plant illness recognition. This gadget
will help to oversee the different natural circumstances which incorporate soil
dampness, temperature and mugginess for the development of the plant and
identify the plant infection. This paper presents the observing of the plant
through Thing speak and illness recognition through CNN.

7. Environmental monitoring and disease detection of plants in smart

greenhouse using internet of things, Fazeel Ahmed Khan1, Adamu Abubakar
Ibrahim1 and Akram M Zeki1

This exploration executes the possibility of robotization utilizing Internet of

Things (IoT) in a nursery climate. The improvement is centres around
arrangement of agricultural nurseries into limited scope level changing it into a
keen nursery. They are to help in observing the nursery climate conditions,
water system the executives, picture assortment utilizing introduced cameras
just as foreseeing illnesses in the plants on gathered leaf datasets. This
examination centre around advancement to approve a proposed framework plan
and engineering for an appropriate IoT based observing for climate conditions,
overseeing water system framework and a compelling technique for
distinguishing leaf sicknesses on the plants inside a nursery climate.

8. PLANT DISEASE IDENTIFICATION USING DEEP LEARNING

CLASSIFICATION MODEL CNN, Sakshi Mangal1 and Pratiksha Meshram2

Horticulture in India isn't tied in with getting nourishment for occupation from
agribusiness. Huge part of Indian economy comes from agribusiness as India is
perhaps the biggest exporter of food, grains and other agrarian items. Over 70%

16
of rustic populace of India is subject to farming of their work. 42% of yield gets
harmed each year which prompts gigantic misfortune to Indian ranchers, out of
which 15.7% of yield is harmed due to bother. In this manner, early discovery
of plant illness is vital to shield total plant from getting harmed. The more
established method in which sicknesses are distinguished when changes are
recognizable on plant leaf isn't effective as till that time practically entire plant
gets harmed in this way new method like picture handling and PC vision
calculations ought to be utilized to distinguish illness at beginning phase. The
procedure which is carried out ought to give precise and exact identification of
sickness, so the pesticide and bug spray which are being utilized doesn't harm
nature of soil and furthermore spreading of huge measure of pesticides and
insect spray harms straightforwardly or in a roundabout way the soundness of
yield. Harm in crop quality or amount prompts misfortunes in this way early
and exact location of plant sickness are very significant. In this way, for early
illness location we utilized Laplacian channel and UN sharp veiling strategy as
picture preparing method and Canny edge discovery as picture division
procedure. The arrangement model which we are carrying out in this venture is
convolution neural organization which is a profound learning grouping model.

CHAPTER 3

METHODOLOGY

3.1) DATABASE

Recently deep learning plays an important role in Medical Science, Plant

Pathology, etc., by using CNN. The neural networks trained with larger number
of datasets prove more accuracy than earlier methods. The network is trained to
predict the disease. The Models had been trained on a publicly to be had dataset
referred to as Plant Village, which contains a complete of 54,306 pictures
containing 38 distinctive healthy/diseased leaves associated with their 14 plant
species. The length of the pictures become modified to 224 × 224 × 3 and
normalization become taken into consideration with the aid of using dividing
the values of pixel with the aid of using 255 for making it appropriate for the

17
preliminary values of the fashions. The dataset become divided with the aid of
using 70%, 20%, and 10% into 3 classes to keep away from over fitting:
training, validation, and checking out datasets, respectively .The dataset is in the
form of [Link] which is an effective way of implementation. With this
datasets, Deep learning would work more quick and efficient. Some of the
sample images of datasets are given below.

Fig 3.1 Tomato Mosaic Virus Fig 3.2 Healthy Plant

Fig 3.3 Bacterial Spot Fig 3.4 Late Blight

18
 Fig 3.5 Leaf Mould

Display plant
Status and
Disease
Name

Fig 3.6 Flow Chart of CNN

19
3.2) DEEP LEARNING

The deep learning process is same as the machine learning process that is a
collection of several levels ordered in pecking order. Initially, the networks
learn something at that level in the order and then this information is sent to the
following level. The upcoming level takes the information and combines it into
something, and it is passed on the third level. This process proceeds as each
level in the pecking order builds more complex from input takes from the
preceding level. Deep learning is easy for learning that has the prospective to be
useful for a more real-time world. It is used for detecting objects, language
translation, making decisions and so on. It finds its application in big data with a
wide extension of knowledge-based prediction. So, deep learning will be a
powerful engine for pragmatic results production. The figure 2.5 shows the flow
of deep learning network layers.

The main advantage of deep learning over machine learning is its capacity to
execute the feature engineering. Deep learning will scan the dataset for
searching the features that will correlate and combine for faster learning. It can
be trained with different datasets. Deep learning is flexible to get adapted with
new problems to solve them in future. Deep learning is widely used in Image
caption generation, Character text generation, Colorization of black and white
images and so on.

20
 Fig 3.6 Simple deep learning Network

3.3) TRANSFER LEARNING

Transfer learning is one of the machine learning techniques, which is used to

develop a model for the task. It is also a recommended one in deep learning
method where the pre-trained models are used as a starting point for the
computer vision. In this technique, a model which is trained on a single task is
remodelled on another related task. It is an enhancement which allows a quicker
progress of tasks while modelling the new task. In other words, it is related to
multi-task learning and concept drift.

21
 Learning
Dataset 1 System Task 1

Knowledge

Learning
Dataset 2 System Task 2

Fig 3.7 Transfer Learning

Transfer learning is not like a machine learning technique, but it can be used as
a “Design methodology”. It is not a complete study-area of machine learning
only a part of the area is applicable for transfer learning. In spite of that, it is
famous in deep learning given the expansive resources needed to train the huge
amount and challenging data sets in that models are trained. This learning works
only in deep learning if the model features learned from the initial tasks are
common which means that they can be used for a related task as well. This way
of transfer learning used in deep learning is called inductive transfer. Also, the
model input should be of the same size as it was early trained with. Otherwise, a
pre-processing step is added to resize the input to the required size. The figure
Fig. 2.6 represents how the data set is trained and tested using transfer learning.

3.3.1) WORKING

In computer vision, the neural networks of the initial layer find edges of the
image, the center layers for the shapes and the later layers for the specific
features of several tasks. In transfer learning, initial and center layers are used
and the later layers are retrained. It helps to hold the labeled data of the task it
was early trained on. In transfer learning, we try to fetch more information from
the previous task for the new task at hand. This information can be in different
forms and it depends only on the problem and data.

3.3.2) ADVANTAGES OF TRANSFER LEARNING

Transfer learning has some benefits but the main advantage is train time
limitation, better performance in neural networks and no need for more data for
getting more accuracy. More data are required to train a neural network but the

22
ingress to the data is not available always. This is the main reason for transfer
learning is more users friendly. With transfer learning, a little training data for a
machine learning model can be built with an already pre-trained model. This is
hardly obtainable in neural network processing because training time is reduced
on complex tasks also. This is a method to speed up the training of neural
networks as a feature extraction method. Also, this is a reusing method for a
model trained on a related analysing modelling problem.

3.3.3) APPROACHES TO TRANSFER LEARNING

a) TRAINING A MODEL TO REUSE IT

Consider one particular task is wanted to solve but it does not have much
information (data) to train a deep neural network. One way for this is problem is
to detect a related task with a large amount of data. Now, the deep neural
network had been trained on the second task and this method is used as a
starting point to solve the first task. The layers used depend on the problem. If
the inputs are the same in both the tasks, conclusively reusing the model and
making predictions with the same data for the new input is an option. On the
other hand, changing and retraining dissimilar task-specific layers and the
output layer is a method to consider.

b) USING A PRE-TRAINED MODEL

The second approach uses a pre-trained model. The number of layers to reuse
and how to retrain depends on the task of this model. For example, nine pre-
trained models are provided by Keras that can be used for prediction in
comparison with other available data, transfer learning, fine-tuning and feature
extraction of data. This type is mostly used in deep learning.

c) FEATURE EXTRACTION

Another approach is to use deep learning, which discovers the important

features for a better representation of the problem. This can often result in a
better performance that can be acquired with the representation of hand-
designed. Domain experts and researchers, manually hand-crafted the features
using machine learning. Providentially, deep learning can automatically extract
features. Neural networks have the dexterity to decide which features are large.
A good combination of features can be discovered by representation learning
algorithms within a short time frame, even though for complex tasks. The figure
2.7 shows the extraction of feature from the original data to reduced data.
23
 Features New features

Data Original Data Reduced Data

points

Fig 3.8 Feature extraction

Then the learned representation can also be used for other problems. The first
layers are used to find the right representation of features, but the output of the
network does not use because of too task specific. Instead, data is fed into the
network and one of the intermediate layers is used for the feature extraction.
This approach is commonly used in computer vision. Because the size of the
dataset is reduced, this decreases the run time and makes suitable for traditional
algorithms as well.

3.4) CONVOLUTIONAL NEURAL NETWORK (CNN)

A Convolutional Neural Network (CNN or ConvNet) is one of the special types

of artificial neural network (ANN), and is used in image processing techniques,
recognition and is designed for processing the pixel data. The CNN is designed
for bi-dimensional image data, although they can be designed for one
dimensional and three-dimensional data. CNN's are most powerful for image
processing; Artificial Intelligence (AI) to perform generative and descriptive
tasks uses deep learning along with some recommended systems and natural
language processing (NLP). A system of hardware or/and software patterned
after the operation of the human brain's neurons is called a neural network.
Traditional neural networks must feed images in compressed resolution pieces
and these are not ideal for image processing. This network has its "neurons"
which are arranged the frontal lobe, the area responsible for processing visual
stimuli in humans and the other animals. The layers are organized to cover the
whole visual field avoiding the piece by piece image processing problem of
traditional neural networks.

24
 Fig 3.9 Layers of Convolutional Neural Network

It uses a system like a multilayer perceptron (an artificial neuron) that has been
designed for lower requirements for processing. Increase in efficiency and
removal of limitations for image processing results in a system, which is
simpler to train limited for image processing. CNN has different layers such as
the input layer, an output layer and the hidden layers (multiple convolutional
layers, pooling layers, fully connected layers and normalization layers).

3.4.1) INPUT LAYER

The input layer or volume is an image and the dimensions of the image dataset
are represented in three dimension as [width x height x depth]. This dimension
represents the matrix of pixel values. For example input is [64x32x2]. So
width=64, height=32 and depth=2. Here the depth represents Red, Green, Blue
(R, G, B) channels. And the input layer or volume should be divisible number
of times by 2.

3.4.2) CONVOLUTIONAL LAYER (CONV)

This layer is the main building block of CNN. The main objective of this
CONV layer is a feature extraction from the input layer. A small part of the
image and the convolutional layer is joined for reducing the expense. For this
purpose, dot products had been applied between a filter and a receptive field on

25
all the entire dimensions. After the dot product, the output volume along with
single integer will be obtained at the end. This one is called a feature map. This
process is done for the all the entire input image. The output of the present layer
is given as input to the next layer. For feature detection, filter, kernel or feature
detector, a small matrix is used. The size of the typical filter on the first layer of
CNN is [5x5x3]. After the computation of dot product and by sliding the filter
over the image dataset, a convolved feature, activation map or feature map will
be formed (output volume). The total number of filters is called depth. The size
of the filter and the receptive field that is the local region of the input volume
both are the same. The set of neurons that all points towards the receptive field
is represented as the depth column or fibre. Stride is used to spatially produce
the smaller output volumes. Zero-padding is very important in the convolutional
layer because it increases the performance and reduces the loss of the border
information by adding zeros around the outside of the input layer. So the same
number of outputs and inputs are maintained when the convolutions end up. A
filter with spaces between its cells is called dilation. This is a new hyper
parameter that is introduced to the convolutional layer. This process is repeated
until it goes through the entire image. The Convolutional layer has ReLu
activation to replace all negative values to zero.

3.4.3) POOLING LAYER (POOL)

The function of this layer is to reduce the computational complexity of the

model and spatial dimensions of the given input data. This layer is in between
two convolutional layers. Also, over fitting can be controlled by this layer. It
does not depend on the depth slice of the input. Different functions such as Max
pool, Average pool or L2 norm-pool are in this part. Max pooling is considered
as the most important part of the input layer, because it is a way to reduce the
spatial dimensions of input data. So, it is mostly used. Here the most important
part is the value of the brightest pixel.

3.4.4) FULLY CONVOLUTED LAYER (FC LAYER)

One of the main functions of this FC layer is that it connects each neurons of
one layer to the neurons of another layer. It operates on a flattened input and
each input is connected to all neurons. This layer generally uses the softmax
activation function. This function is used to classify the generated features of
the input images into various classes, flatten the result set and this classification
process is based on the training dataset.

26
3.4.5) OUTPUT LAYER

After all the process, output is needed. The hidden layers of convolution,
pooling can only extract the features and reduce the number of parameters but
fully convoluted layer is required to provoke the output. The output layer
contains different loss functions like binary cross-entropy, categorical cross-
entropy, in order to determine the error at prediction.

3.4.6) RELU

Rectified Unit (ReLu) is the most commonly deployed activation function for
CNN neuron‟s outputs. It is a non-linearity, which is applied similar to neural
networks.

3.4.7) RESNET

ResNet represents Residual Network which is an ANN kind. This builds on

pyramidal cells in the cerebral cortex. ResNet does this by short-cuts to jump
over several layers. ResNet was allowed to train extraordinary deep neural
networks with more than 500 layers and it still achieves better powerful
performance. According to number of layers in the network, the number of
parameters will be selected. There are many variants in this Architecture, that is
same concept but with different number of layers ResNet-v2, ResNet-18,
ResNet-34, ResNet-50, ResNet-101, ResNet-110, ResNet-152, ResNet-164,
ResNet-1202 etc., Table 2.1 represents the number of layers and the parameters
used in different ResNet layers.

Number of Layers Number of Parameters

ResNet 18 11.17M
ResNet 34 21.282M
ResNet 50 23.521M
ResNet 101 42.513M
Inception ResNet v2 54.33M
ResNet 152 58.45M

Table 3.1 Different types of ResNet

ResNet uses batch normalization in the core. The input layer is adjusted by
batch normalization in order to improve the performance of the network. The

27
problem of covariate shift will also be reduced. The identity connection helps to
preserve the network from deleting the gradient problem. This connection is
widely used in ResNet. Deep Residual Network uses Bottle design residual
network for increasing the network performance furthermore. A ResNet-101 is
one of the models in ResNet, implemented in Keras library with Tensor flow
backend. ResNet 101 is more accurate than many other states of models and it
has a vanishing gradient. Figure 3.9 shows the schematic diagram of ResNet-
101.

A) Advantages of ResNet:

 Res-Net connects faster than plain counterpart and it also reduces the
degradation problem.
 Only the dimension of layer changes and also Shortcuts are identical for
all ResNet blocks.
 Inception ResNet-101 architecture offers bigger batch size which helps us
to reduce the training time.

Fig 3.10 Schematic Diagram of ResNet 101

28
3.5) VGG:

The simple running precept of VGG Net is to apply deeper layer with smaller
filters. The enter layer measurement of the VGG structure is about for an image
length of 224 × 224. Pre-processing includes subtraction of the suggest RGB
cost from every pixel of the enter photo. Pre-processing is accompanied with
the aid of using a stack of five convolutional layers, every of that is
accompanied with the aid of using a MaxPool layer, i.e., every set of
convolutional layers is accompanied with the aid of using a MaxPool layer. The
very last MaxPool layer precedes 3 absolutely connected (FC) layers. The first
FC layers have 64 × 64 (4096) channels, while the remaining FC layer has one
thousand channels, that is accompanied with the aid of using a softmax
activation function. VGG community has a couple of flavours, drastically VGG-
sixteen and VGG-19. VGG-sixteen and VGG-19 use the identical structure with
distinctive range of layers. VGG-sixteen makes use of sixteen layers, while
VGG-19 makes use of 19 layers.

3.5.1) VGG-16:

VGG16 is a variation of VGG model with 16 convolution layers and we have

investigated the VGG16 design top to bottom. VGGNet-16 comprises of 16
convolutional layers and is engaging a result of its very uniform Architecture.
Like Alex Net, it has just 3x3 convolutions, yet bunches of channels. It tends to
be prepared on 4 GPUs for 2–3 weeks. It is presently the most favoured
decision locally for removing highlights from pictures. The weight design of the
VGG Net is freely accessible and has been utilized in numerous different
applications and difficulties as a standard component extractor. However, VGG
Net comprises of 138 million boundaries, which can be somewhat difficult to
deal with. VGG can be accomplished through transfer learning. In which the
model is pretrained on a dataset and the boundaries are refreshed for better
exactness and you can utilize the boundaries esteems.
3.5.2) Advantages of VGG-16:

 It only has Conv and pooling layers in it.

 It has an aggregate of around 138 million boundaries.

 It has been trained on Image Net information.

29
Fig 3.11 VGG 16 & 19 Architecture
30
3.5.3) VGG-19:

VGG19 is comparative model architecture as VGG16 with three extra

convolutional layers, it comprises of an aggregate of 16 Convolution layers and
3 thick layers. Following is the engineering of VGG19 model. In VGG
organizations, the utilization of 3 x 3 convolutions with step 1 gives a viable
responsive documented identical to 7 * 7. This implies there are fewer
boundaries to prepare. Another variety of VGG Net has 19 weight layers
comprising of 16 convolutional layers with 3 completely associated layers and
same 5 pooling layers. In both variety of VGG Net there comprises of two Fully
Connected layers with 4096 channels every which is trailed by another
completely associated layer with 1000 channels to anticipate 1000 names. Last
completely associated layer utilizes softmax layer for arrangement reason.
3.5.4) Advantages of VGG-19:

• It is an excellent design for benchmarking on a specific errand.

• Also, pre-prepared organizations for VGG are accessible uninhibitedly on the

web, so it is generally utilized out of the case for different applications.

3.5.5) Inception V3:

In contrast with VGG Net, Inception Networks (GoogLeNet) have ended up

being all the more computationally effective, both as far as the quantity of
boundaries created by the organization and the affordable expense caused
(memory and different assets). In the event that any progressions are to be made
to an Inception Network, care should be taken to ensure that the computational
benefits aren't lost. In this manner, the variation of an Inception network for
various use cases ends up being an issue because of the vulnerability of the new
organization's effectiveness. In an Inception v3 model, a few procedures for
upgrading the organization have been put recommended to extricate the
limitations for simpler model variation. The procedures incorporate factorized
convolutions, regularization, measurement decrease, and parallelized
calculations.

31
3.6) GOOGLE COLAB:

Google Colaboratory (Google Colab) is one of the free cloud services, a Jupyter
notebook environment and one advantage is, it supports free Graphics
Processing Unit (GPU). By using this python programming language, coding
skill is improved and deep learning applications are developed using popular
libraries like Keras, Tensor flow, Pytorch, and Open CV. It allows us to write an
arbitrary python code and execute it via the browser. It provides free
computational power, supports Python 2.7 and Python 3.6, and is also integrated
with Git Hub. The most important feature that differentiates from some other
free cloud services is, it provides GPU and free of cost. Using this, default
hardware is easily altered from Central Processing Unit (CPU) to GPU or vice
versa.

CHAPTER 4

Hardware Requirements

4.1) Raspberry Pi3:

The Raspberry Pi is a low cost, credit-card estimated computer that connects to

a PC screen or TV, and utilizations a standard console and mouse. It is a fit little
gadget that empowers individuals, all things considered, to investigate figuring,
and to figure out how to program in dialects like Scratch and Python. It can do
all that you'd anticipate that a desktop computer should do, from perusing the
web and playing top quality video, to making accounting pages, word-
preparing, and messing around. RASPBERRY PI 3 is an improvement board in
PI arrangement. It very well may be considered as a solitary board PC that deals
with LINUX working framework. The board not just has huge loads of
highlights it additionally has spectacular handling speed making it reasonable
for cutting edge applications. PI board is explicitly intended for specialist and
architects who are keen on LINUX frameworks and IOT (Internet of Things).

32
 Fig 4.1 Raspberry PI3 Model

4.2) DTH11 Sensor:

Dampness estimation decides the measure of dampness present in the gas that
can be a combination of water fume, nitrogen, argon or unadulterated gas and so
forth… Humidity sensors are of two sorts dependent on their estimation units.
They are an overall dampness sensor and Absolute stickiness sensor. DHT11 is
a computerized temperature and stickiness sensor. DHT11 is a minimal expense
computerized sensor for detecting temperature and humidity. This sensor can
be effectively interfaced with any miniature regulator, for example, Arduino,
Raspberry Pi and so forth… to gauge mugginess and temperature quickly.

Fig 4.2 DTH11 Sensor

33
4.3) pH Sensor:

A pH Meter is a gadget utilized for potentiometrically estimating the pH, which

is either the fixation or the action of hydrogen particles of a fluid arrangement.
It contains an AtMega8 IC. It ordinarily has a glass cathode in addition to a
calomel reference anode, or a blend terminal. The pH meters are normally used
to quantify the pH of fluids, however uncommon tests are in some cases used to
gauge the pH of semi-strong substances.

Specialized Specifications of pH Sensor:

• Meter: 5 x 8 x 3.5cm [2 x 3.2 x 1.5″]

• Probe length: 21cm [8″]

• Bronze test measurement: 4.8mm

• Silver test width: 5.1mm

• Distance between tests: 1.2 cm

• Color: blue

• Item size: 330x100x30mm

• Net weight: 87g

• Package weight: 99g

34
 Fig 4.3: pH Sensor

4.4) Thonny Python IDE:

Thonny is an incorporated improvement climate for Python that is intended for

fledglings. It upholds various methods of venturing through the code, bit by bit
articulation assessment, itemized representation of the call stack and a mode for
clarifying the ideas of references and heap. The program deals with Windows,
Mac OS and Linux. It is accessible as double group including on-going Python
mediator or pip-installable bundle. It very well may be introduced by means of
working framework bundle supervisor on Debian, Raspberry Pi, Ubuntu and
Fedora.

35
4.5) Block Diagram:

Using Thing Speak Software

Fig 4.4 Block Diagram

CHAPTER 5

Working and Setup

5.1) Working:

Once, the Hardware Setup and Coding is over we will switch on the Raspberry
Pi to get Started. The Raspberry Pi connected with Camera captures images
first and the CNN Algorithm is executed. Now, the CNN Algorithm runs and
checks the status of the plant and displays the disease name. We have prepared
the Convolution Neural Network application with in excess of 1000 solid leaves
and sickness contaminated leaves. Preparing the tensor stream takes time
around 45 minutes relying upon the quantity of dataset that we give. The
exactness of recognizing the illness is corresponding to the quantity of datasets.
Subsequent to preparing, the order for catching the genuine time leaf is

36
executed through the terminal window. The raspberry pi camera subsequent to
catching the leaf contrasts the picture and the dataset that we have given and
gives the outcome. The outcome depends on the dataset gave, which means the
rate of leaf influenced by disease and which are definitely not. We can improve
the precision by increasing the dataset. After all the preparation steps are
finished, the content runs a last test precision assessment. The test assessment
gives the best gauge of how the prepared model will perform on the order task
and by running the content on the picture. Every execution will print a rundown
of leaf. The Values are then sent to the Cloud and the results are displayed in the
Web Server using Thing Speak Software.

5.2) Environmental Setup:

Fig 5.1 Hardware Setup

37
 CHAPTER 6

Results and Discussions

6.1) Results:

In this framework, we are checking and controlling the climate around the
plants. Every one of the sensors are interfaced to the Raspberry Pi; it is utilized
to every one of the sensors are working in simultaneously. The DHT11 is
utilized for detecting the temperature and stickiness from general climate,
while pH sensor is utilized for the detecting the pH level of supplements and
furthermore shows the electrical conductivity of supplements. In this
framework there is sprinklers are utilized to control the mugginess and
temperature around the plants. Sprinklers are utilized for sprinkling the pH
arrangement/water in light of the moistness. The submarine engine is utilized
to fill the pH tank of plants. Furthermore, bug identification on plants and
sprinklers will be utilized to sprinkle natural pesticides. The all yield
information is interfaced with ESP8266 Wi-Fi module so that is all the
information is ceaselessly observed on Blink application and furthermore
information put away on cloud.

6.2) GRAPH & OUPUTS:

6.2.1) CNN for VGG-16:

38
Fig 6.1 Output of CNN with VGG-16

39
6.2.2) CNN for VGG-19:

Fig 6.2 Output of CNN with VGG-19

40
6.2.3) CNN for ResNet:

Fig 6.3 Output of CNN with ResNet

41
6.2.4) CNN for Inception V3:

Fig 6.4 Output of CNN with Inception V3

42
6.2.5) IDE & Web Server Outputs:

Fig 6.5 Output of Values in Python IDE

43
 Fig 6.6 Web Results showing a Bacterial and a Healthy Plant

6.2.6) DISCUSSIONS

The data sets are processed with pre-trained network to predict the disease. The
final testing stage is done by choosing datasets in the testing image. It is then
tested for accuracy. Prediction of the disease is done with CNN with the help of
Google Colaboratory and Python IDE.

CHAPTER 7

Conclusions and Future Scope

7.1) Conclusions:

The datasets are processed with pre-trained CNN Network, Transfer Learning
Method to predict the Plant‟s disease. The image datasets are taken for training
and testing them. The model is tested for accuracy level. This process is done
using python. By using GPU, the training time to run the dataset had been
reduced to 30-40mins. With the different number of epochs the disease
detection is done to get higher accuracy. Further training with 4400 images
(Train – 3121, Test - 1279) of datasets, accuracy is improved by improving the
epoch values.

In this work, we utilized distinctive convolutional neural network (pretrained)

organizations to programmed recognition and grouping of illnesses in a leaf of a
44
plant. We took vgg-16 and 19 models, to assess their exhibition and we
accomplished the preferable exactness over different models. On, looking at
vgg-16 and vgg-19, vgg-19 gives preferred exactness rate over vgg-16 (i.e.,
92.13% and 94.90% individually. In this manner, Vgg-19 is productive than
ResNet and Inception as a result of its utilization of Lesser number of Filters,
Lesser number of Layers, a basic organization and less preparing misfortune
and approval Loss.

7.2) INFERENCE:

By increasing the epoch value, accuracy of prediction of Disease is also

increased. We can improve the accuracy value, even by increasing epoch value
further.

7.3) FUTURE SCOPE:

In future, this project can be implemented with high performance models and
the disease can be predicted with great accuracy in a short span of time. This
can also be improved further by improving the code. This can further be
implemented as a Drone, so that it can be monitored easily.

Reference:

[1] Dr.S Murugan, et al. “Social Ramification of Fire on Forest using IoT”, CSI
communication, vol. 42, issue.3, pp.32- 34,2018.

[2] G rama mohan reddy “Internet of Things: Power controlling through in

smart mobiles” international journal of pure and applied mathematics (IJPAM))
vol 118 issue 17 [Link] 781-800,2017

[3] Smith J S., Camargo A., "An algorithm based on image processing to
automatically identify visual symptoms of plant disease."

[4] Tapas kanungo, David Mount, "Efficient k means Clustering Algorithm:

Analysis and Implementation."IEEE Transactions on Pattern Analysis and
Machine Intelligence,Vol 24,No 7,July 2002

[5] Color Space Conversions, Ford A., Robert,.August 11,1998(a).

45
[6] C M Procopiuc, P. K. Agarwal, "Clustering Algorithms Exact and Appro
ximate.".Algorithmica(2002) 33;201-226.

[7] Devendra P. Marathe, V. A. Patil, V. D. Chaudhari, "Reliable transfer of

massage to alternative methods, i.e. Wi fiBluetooth, ZigBee, Internet, M edia
GPRS / GSM, IJET, vol 5, issue 7. Pp 2015.

[8] Kiran R. Gavhale, and U. Gawande, “An Overview of the Research on

Plant Leaves International Journal of Pure and Applied Mathematics Special
Issue 882 Disease detection using Image Processing Techniques,” IOSR J. of
Compu. Eng. (IOSRJCE),vol. 16, PP 10-16, Jan. 2014.

[9] Y. Q. Xia, Y. Li, and C. Li, “Intelligent Diagnose System of Wheat

Diseases Based on Android Phone,” J. of Infor.& Compu. Sci., vol. 12,
pp.6845-6852, Dec. 2015.

[10] K. Padmavathi, and K. Thangadurai, “Implementation of RGB and Gray

scale images in plant leaves disease detection – comparative study,” Indian J. of
Sci. and Tech., vol. 9, pp. 1- 6,Feb. 2016.

[11] Sachin D. Khirade and A. B. Patil. “Plant Disease Detection Using Image
Processing.” International Conference on Computing Communication Control
and Automation (ICCUBEA), 2015 International Conference on, pp. 768-771.
IEEE, 2015.

[12] Sannakki, Sanjeev S., Vijay S. Rajpurohit, V. B. Nargund, and Parag

Kulkarni. “Diagnosis and classification of grape leaf diseases using neural
networks.” In computing, communications and Networking Technologies
(ICCCNT), 2013 Fourth International conference on, pp. 1-5 IEEE, 2013

[13] Kutty, Suhaili Beeran, Noor Ezan Abdullah, Habibah Hashim, and Aida
Sulinda. “Classification of Watermelon Leaf Diseases Using Neural Network
Analysis.” In Business Engineering and Industrial Applications Colloquium
(BELAC), 2013 IEEE, pp. 459-464. IEEE, 2013.

[14] Rothe, P. R., and R. V. Kshirsagar. “Cotton Leaf Disease Identification

using Pattern Recognition Techniques.” In Pervasive Computing (ICPC), 2015
International Conference on, pp. 1-6. IEEE, 2015. [15] Pearson, Roger C., and
Austin C. Goheen. Compendium of leaf diseases, American Phytopathological
Society, 1988.

46
[15] Nitrogen fixation‟ available at https:\\[Link]/wiki Nitrogen
fixation, accessed February 2017

[16] Savary, s., ficke, A, Aubertort,j-n.,et al crop losses sue to disaeses and their
implications for global food production losses.

[17] Mian MA. Missaoui, A.M walker D.R .,et al: „frog eye leaf spot of
soybean: a review and proposed race designations for isolates of sercospora
sojina hara.

[18] soybean diseases available at

[Link]
accessed january 2017

[19] Shrivastava. S., Hooda, D.s automatic brown spot and frog eye detection
from the image captured in the field.,[Link].2014

[20] Li,x. yang, X.B., „similarity, pattern and grouping of soybean fungal

diseases in the united states implications for the risk of soybean rust .

47
48
49
50
51
52
53