Apr, 2021 G.

AMBO UNIVERSITY WOLISO CAMPUS

SCHOOL OF TECHNOLOGY AND INFORMATICS
DEPARTMENT OF INFORMATION
TECHNOLOGY

ETHIOPIAN COFFEE DISEASE IDENTIFICATION

USING IMAGE ANALYSIS
A MASTER THESIS SUBMITTED TO THE DEPARTMENT OF
INFORMATION TECHNOLOGY IN PARTIAL FULFILMENT OF THE
REQUIREMENT FOR THE DEGREE OF MASTER OF SCIENCE IN

INFORMATION TECHNOLOGY

BY: AMANUEL TAMIRU URGESSA

MAIN ADVISOR: TEKLU URGESSA (PhD, Asst. Professor)

WOLISO,
ETHIOPIA

Apr, 2021 G.C

 CERTIFICATION SHEET
A thesis research advisor, I hereby certify that I have read and evaluated this thesis

prepared under my guidance by Amanuel Tamiru Urgessa entiteled:’ Ethiopian

Coffee Disease Identification Using Image Classification ’as recommend that it be

submitted as fulfilling the thesis requirement.

__________________
Name of Major Advisor Signature Date

Name of Co-Advisor Signature Date

As mentioned of the Board of Examiners of the M,Sc/MA. thesis open defense examined.

We certified that we have read and evaluated the thesis prepared by and

examined the candidate. We recommend that the thesis be accepted as fulfilling the thesis

requirements for the degree of Master of Science/Art in

Chairperson Signature Date

Declaration
I, the undersigned, declare that the thesis comprises my own original work. In compliance

with internationally accepted practices, I have dually acknowledged and refereed all
 materials used in this work. I understand that non-adherence to the principles of academic

honesty and integrity, misrepresentation/ fabrication of any idea/data/fact/source will

constitute sufficient ground for disciplinary action by the university and can also evoke

penal action from the sources which have not been cited or acknowledged.

Amanuel Tamiru Urgessa ___________ __________

Name of the student Signature Date

 Acknowledgement

Above all, I would like to thank Almighty God for all of his Blessings throughout my life.

Next, I would like to extend my heartfelt thanks to my advisor Dr. Teklu Urgessa

(Assistant Professor, Adama Science and Technology University) for his sincere giving

me constructive, professional advising and guidance in preparing the proposal, research

guidance and finalizing the thesis. Without his encouragement, comments, suggestion and

support the completion of this research work would not have been possible. He initiated

me to do by giving precious comments on necessary points. My thanks go to him again,

since it is difficult to mention his contribution to my achievements in words, it is better to

say my heart has recorded it forever.

I want to express my grateful thanks for my beloved wife and family, for their strong

cooperation by standing with me throughout every step of my academic work. Last but not

least, I would like to give my thanks for those people who had played a role in this

successfulness for their moral support, endless love and encouragement directly or

indirectly during my study.

May God Bless you all!

ii
 Table of Contents
SCHOOL OF TECHNOLOGY AND INFORMATICS DEPARTMENT OF INFORMATION
TECHNOLOGY.............................................................................................................................1
MAIN ADVISOR: TEKLU URGESSA (PhD, Asst. Professor)................................................1
Declaration......................................................................................................................................5
Acknowledgement......................................................................................................................ii
List of Acronym.............................................................................................................................vi
List of Tables................................................................................................................................vii
List of Figures..............................................................................................................................viii
Abstract..........................................................................................................................................ix

CHAPTER ONE .............................................................................................................................

1
1.1 INTRODUCTION ................................................................................................................
1
1.2 Background of the study .......................................................................................................
2
1.3 Statement of the problem .....................................................................................................
3
1.5.1 General objective ........................................................................................................... 6
1.5.2 Specific Objective.......................................................................................................... 6
1.6 Motivation ............................................................................................................................
7
1.7 Scope of the Study ................................................................................................................
7
Chapter Two: Literature Review ....................................................................................................
8
2.1 Introduction ..........................................................................................................................
8
2.2 Overview of Plant Leaf Disease ...........................................................................................
8
2.3 Coffee Leaf Rust .................................................................................................................
11
2.4 Coffee Berry Disease ..........................................................................................................
12

iii
 2.5 Coffee Wilt Disease ............................................................................................................
15
2.6 Knowledge Based Identification of Coffee Disease ...........................................................
16
2.7 Statistical Local Binary Approach ......................................................................................
17
2.8 Coffee Bean Grading Using Visual Sensor ........................................................................
17
2.10 K-Means Clustering Image Classification Approach .......................................................
20
2.11 SVM/Support Vector Machine/ Approach on Image Classification ................................
22
2.12 BPNN/Back Propagation Neural Network/ ......................................................................
25
2.13 CNN/Convolutional Neural Network/ ..............................................................................
27
2.14 HD-CNN/Hierarchical Deep Convolutional Neural Network/ ........................................
29
2.15 Multiple Linear Regression Approach .............................................................................
30
2.16 KNN/K-Nearest Neighboring/ Approach .........................................................................
31
2. 17 Image Classification Architectures .....................................................................................
31
2.17.1 ResNet50 ......................................................................................................................
31
2.17.2 Google Net ....................................................................................................................
32
2.17.3 Alex Net ........................................................................................................................
33
2.13 RELATED WORK ...........................................................................................................
35
2.3.1 SIMILARITY AND DIFFERENCE OF OBJECT DETECTION
AND
IDENTIFICATION ............................................................................................................. 35
CHAPTER THREE ..................................................................................................................... 37

iv
3. METHODS AND METHODOLOGY .....................................................................................
37
3.1. Proposed System Architecture...........................................................................................
37
3.3 Data Collection and Dataset ...............................................................................................
39
3.4 Pre-processing ....................................................................................................................
40
3.5 Classification Techniques ...................................................................................................
41
3.6.1 K-Means Clustering Model .........................................................................................
41
3.5.1 K-Nearest Neighboring/KNN/ model ....................................................................
41
3.5.2 Support Vector Machine/SVM/ model ..................................................................
42
3.5.3 Convolutional Neural Networks ............................................................................
42
3.5.4 Fast Regional Convolutional Neural Network /Fast RCNN/ .................................
43
3.6 Feature Extraction...............................................................................................................
45
3.8 Implementation Tool ..........................................................................................................
47
3.9 Evaluation ..........................................................................................................................
48
CHAPTER FOUR ........................................................................................................................
49
RESULTS ....................................................................................................................................
49
4.1 Introduction ........................................................................................................................
49
4.2 Experiment Steps ...............................................................................................................
50
4.2.2 Dataset Preprocessing .................................................................................................
50

v
 4.3 Fast RCNNs RESULTS ANALYSIS ................................................................................
52
CHAPTER FIVE ......................................................................................................................... 55
CONCLUSION AND FUTURE WORK .....................................................................................
55
5.1 Conclusion .......................................................................................................................... 55
5.2 Future work ........................................................................................................................ 57
REFERENCES ............................................................................................................................ 58

vi
 List of Acronym
• SVM……………………………………………Support Vector Machine

• SOM…………………………………………………………Self Organizing Map

• RBFN…………………………………………………………Radial

• KNN………………………………………………K-Nearest Neighboring

• GSMA…………………………Global System for Mobile Communication Association

• DTL…………………………………………………. Deep Transfer Learning

• CLR…………………………………………………………Coffee Leaf Rust

• CBD…………………………………………………………Coffee Berry Disease

• CWD…………………………………………………………Coffee Wilt Disease

• BPNN…………………………………………. Back Propagation Neural Network

• CNN……………………………………………Convolutional Neural Network

• RCNN…………………………………Regional Convolutional Neural Network

vii
 List of Tables
Table 2.1 summary of literature review

Table 3.1 Tools and Materials

Table 3.1 Tools and Materials

Table 4.1 comparison of different experiments

viii
 List of Figures
• Fig. 1 Coffee leaf rust symptoms and signs.

• Figure 2.1. Radial Base function distribution.

• Fig 2.2 K-Means Clustering System Architecture.

• Figure2.3. Coffee Plant Disease Identification Model (Debasu, 2018)

• Fig 2.4 Google Net Architecture

• fig 2.5 Alex net architecture

• Fig 3.1 faster RCNN diagram (towardai.net)

• fig 3.2: Coffee wilt/spot/ disease

• Fig 3.3: coffee leaf rust

ix
 Abstract

The prevention and control of plant disease have always been widely discussed because plants

are exposed to outer environment and are highly prone to diseases Coffee Berry Disease, Coffee leaf

rust and Coffee wilt disease are the three common types of coffee leaf diseases. Early diagnosis and

accurate identification of coffee leaf diseases can control the spread of infection and ensure the healthy

development of the coffee plants. The existing research uses an approach that cannot guarantee high

recognition rates for coffee leaf diseases. To address this challenge on this thesis We have implemented

a more accurate approach for identification of coffee leaf diseases based on deep convolutional neural

networks. It includes generating sufficient pathological images and designing a novel architecture of a

deep convolutional neural network based on Fast RCNN to detect coffee leaf diseases. For testing

purposes, we have collected infected and healthy leaf images from ministry of Agriculture of south west

shoa zone, This paper provides survey a classification technique that can be used for plant leaf diseases

classification. Identification of symptoms of disease by naked eye is difficult for farmer. Crop

protection in large frames is done by using computerized image processing technique that can detect

diseased leaf using color information of leaves. There are so many classification techniques such as k-

Nearest Neighbor Classifier, Probabilistic Neural Network, Genetic Algorithm, Support Vector

Machine, and Principal Component Analysis, Artificial neural network, Fuzzy logic. Selecting a

classification method is always a difficult task because the quality of result can vary for different input

data. Plant leaf disease classifications have wide applications in various fields such as in biological

research, in Agriculture etc. This paper provides an overview of a classification techniques used for

plant leaf disease classification. using deep convolutional neural network model and Fast RCNN

approach We have identified three common types of disease of coffee leaves. As experimental result

shows, the automated system has achieved 97.0% of accuracy using Fast RCNN approach.

x
 CHAPTER ONE

1.1 INTRODUCTION

Ethiopia is one of the major producers of coffee Arabica ranked as the seventh largest

producer of coffee over the world, and top producer of the coffee Arabica in Africa (Degaga,

2020). In Ethiopia, coffee is produced in different forest, semi-forest, garden and plantation

production methods. More than half of the coffee is consumed by Ethiopians, and the country

leads the Africa’s domestic consumption. The revenue from coffee exports account for 10% of the

annual government revenue, because of the large share the industry is given very high priority

(Nation, 2006). However, Ethiopia has not yet fully exploited its position as the producer of some

of the best coffees in the world.so this research has planned to bring a better way of detecting

disease using different image classification model. Different Approaches are developed in recent

years; one of the best image classification approaches till now is the convolutional neural network

for identification. Early disease images have become a new research hotspot in agricultural

informatization in the field of crop disease recognition. The prevention and control of plant

disease have always been widely discussed because plants are exposed to outer environment and

are highly prone to diseases. Normally, the accurate and rapid diagnosis of disease plays an

important role in controlling plant disease, since useful protection measures are often

implemented after correct diagnosis. This system is based on image processing technology and

uses MATLAB as the main processing tool in this thesis for identifying infected coffee leaf WE

implemented Fast Regional convolutional neural network Approach (Fast RCNN) which is a

powerful approach to classify plant leaves disease.

 1.2 Background of the study

Coffee is one of the world’s most important crops in terms of production

volume and trade. In Ethiopia coffee grows in every region of the country but

majority are produced in the Oromia Region (63.7%) and in the Southern Nations,

Nationalities (34.4%), with lesser amounts in the Gambela Region and around the

city of Dire Dawa.

Generally, in Ethiopia much of the coffee are produced in altitudes between 1,000

and 2,000 meters. Several coffee diseases have caused significant yield losses across

production. Landscapes are a significant threat to global coffee production.

Therefore, early detection of pests and diseases in the field is a first crucial step.

AI with different models of deep learning that can help to identify plant diseases

by the plant’s appearance and visual symptoms should be considered. AI apps

specially Smartphone could alert farmers and perform disease diagnosis, using this

apps prevents the possible outbreak of pests and diseases. Even though many farmers

of developing countries like Ethiopia do not have access to these advanced tools,

internet infiltration and smartphone penetration offer new outfits for in-field crop

disease detection. (GSMA, n.d.) Deep learning is a novel method for image

processing and object detection with greater accuracy in the classification of various

crop diseases. Transfer earning is one such popular approach in deep learning, where

pre-trained models are adapted to do a new task. Deep transfer learning (DTL)

generates a fresh framework for digital image processing and predictive analytics,

with greater accuracy and has huge potential in crop disease detection. DTL approach

also offers a promising avenue for in-field disease recognition using large trained

image datasets and bids a shortcut to the developed models to meet the restrictions
2
that are offered by different platform application. This would have a distinct practical

value for real field environment. The objective of this study was to apply state-of-the

art deep learning techniques for the detection of visible coffee disease and pest

symptoms on different parts of the coffee Plant.

1.3 Statement of the problem

In agriculture disease recognition is a major challenge especially in coffee plant.

Therefore, suitable actions have to be taken to control diseases. (Tylor, 2020) Coffee

diseases cause considerable losses when not treated. , 57% yield loss was observed by

the infection of disease-causing organisms on coffee crop. The most economically

important pathogenic coffee diseases are coffee berry disease (CBD), coffee wilt

disease (CWD) and coffee leaf rust (CLR), and physiological disorder like coffee

branch die back is caused by pseudomonas syringe and non-pathogenic agents.

Similarly, CBD and branch dieback were causing high yield loss of coffee production.

The assessment carried out in Eastern Ethiopia indicated that diseases and insect pests

are causing considerable crop losses (Teferi, 2015 ). Similarly, major insect pest that

affects coffee production in Eastern Ethiopia were coffee stem borer and coffee berry

borer. So, to overcome the decrease of coffee production is one or in other way it

building up the country’s economy as reported (Tylor, 2020) 57% of coffee

production is decreased because of disease.so to control this its highly recommended

that early control of pests to stop spreading from one leaf to another.as the need of

coffee production increases every time it’s not an easy task to control and identify

diseases with accuracy and within minimum time.

3
 Traditionally, plant disease inspection is undertaken with visual inspection of

plant tissues by trained expertise which leads to expensive cost and low efficiency.

The (Talaviya, 2020) automation in agriculture is the main concern and the emerging

subject across the world. The population is increasing tremendously and with this

increase the demand Agricultural products. The traditional methods which were used

by the farmers were not sufficient enough to fulfill these requirements. Thus, new

automated methods were introduced. These new methods give a good promising result

to satisfy the needs and also provided employment opportunities to billions of people.

Artificial Intelligence in agriculture has brought an agriculture revolution. This

technology has protected the coffee yield from various factors like the climate

changes, population growth, employment issues and the food security problems. With

the popularization and advance of information technology in agriculture, cultivation

and management expert systems developed have been widely used, greatly improving

the production capacity of plants. With the popularity of machine learning algorithms

in computer vision different methods and models, such as random forest, k-nearest

neighbor, and Support Vector Machine (SVM) are started to be implemented, and

shows a good improvement on detection of plant disease using machine learning.

However, because the classification features are selected and adopted based on human

experience, these approaches improved the recognition accuracy, but the recognition

rate is still not satisfactory to artificial feature selection. Deep learning-based models

have had great success in object detection, but the state-of-the-art models have not yet

been widely applied to identification of coffee disease. we have used Faster Region

based Convolutional Neural Network (Faster R-CNN), one of the top performing

object detection models in recent years, pre-trained on ImageNet but fine-tuned with

4
my data. No previous work done to identify Ethiopian coffee leaf disease using Fast

RCNN which helps us to make good sprit for better identification result.

1.4 Research questions:

1. Which identification approach has a better performance to identify coffee diseases

compared to other approaches?

2. which convolutional Neural Network architecture best suit for the identification of Coffee
Leaf Rust, Coffee Berry Disease and Coffee Wilt Disease?

5
 1.5 Objectives of Research:

1.5.1 General objective

The aim of this research is to develop automatic coffee disease detection system that

provide identification of coffee leaf disease and diagnose in a very minimal time

compared to manual way by implementing Fast R-CNN.

1.5.2 Specific Objective

1. The specific objective of the research is to identify three types of coffee disease Coffee

Berry Disease (CBD), Coffee Leaf Rust (CLR) and Coffee Wilt Disease.

2. The other plan is to collect relevant Dataset of images of diseased coffee leaves.

3. The goal of this study was to apply Fast RCNN of deep learning techniques for

the identification of coffee leave disease and pest symptoms coffee plant.

4. To review more relevant literatures related to this topic.

5. Developing some implementation on MATLAB that Could identify diseased leaf of at

least one type of disease.

6. The other objective of this research to set future direction of research on this title for

other researchers.

6
 1.6 Motivation

Over the last 40 years breeding activities have been carried out to combat

CLR, CBD and CWD. Breeding for resistance against CLR in Arabica coffee has

successfully utilized single or combinations of major genes designated as SH genes.

Major gene resistance has also been deployed in breeding for resistance against CBD,

though this biological way of defending coffee disease brings good way of combating

disease. but it still has lack of fast way of identification of disease.so I believe if we

identify disease faster, we can easily diagnose diseases early. The inspiration for

conducting this research has arisen from this and many computers versioned

approaches for automated crop disease detection and classification.

1.7 Scope of the Study

In this paper putting scope and limitation to the boundaries of the research

will be a key for getting accurate and precise result as the end of our research. Coffee

Arabica which is type of coffee that grows in Ethiopia. There are many kinds of

disease that affect coffee leaves but we focused only on identification of the three

common disease Those are Coffee berry disease, coffee leaf rust and coffee wilt

disease. There could be more kinds of disease related with geo location and climate

condition as well but for this paper our scope will be in the identification of three

kinds of coffee leave diseases. This research going to address to solve the time,

7
energy, cost consumed to identify coffee leaf disease but preparing the dataset for all

kinds of coffee leaf disease need more budget and time to prepare the dataset.

Chapter Two: Literature Review

2.1 Introduction

This section discusses research works related to the field of AI for different diseases

that affects coffee leaves to classify and detect what type of diseases that worked done are

reviewed. This chapter focuses on the review of the literature of many researchers

undertaken on a different kind of approaches. It presents some of the very prominent

research conducted concerning on coffee leaf and plant disease identification using

different image classification methods and approaches. In this section describes various

approaches for detecting the disease in plant leaf using image processing technique

2.2 Overview of Plant Leaf Disease

Agricultural sector has been the priority of Ethiopia since the early 1990s,

when the Agricultural Development-Led Industrialization and related policy

frameworks were adopted. More of Ethiopia’s population is still employed in the

agricultural sector. The share of agriculture in the GDP has declined over five years

from 44.7% in 2010/11 to 37.2% in 2015/16. Despite its declining contribution to

GDP over five years, agriculture remains the leading sector in terms of contribution to

the country’s overall economy. It is the major source of food for domestic

8
consumption, of raw materials for the domestic manufacturing industries and of

primary commodities for export. (Degaga, 2020) The main coffee producing areas in

Ethiopia are west and south west, southern, eastern, and central regions. On the basis

of management level, vegetation, structural complexity, and agronomic practices,

coffee production systems in Ethiopia can be categorized into four; namely: forest

coffee, semi-managed forest coffee, garden coffee and plantation. Forest Coffee: The

forest coffee system uses wild stands of coffee, which exists naturally within the

forest, and the farmer undertakes minimal management and intervention. The local

communities living in and around the forest simply pick the wild coffee berries from

naturally growing coffee plants and there is no management to improve coffee

productivity. Coffee is the most important agricultural commodity, with an estimated

retail value of 70 billion US dollars. It is crucial for the economy of more than 60

countries and is the main source of income for more than 100 million people. Coffee

leaf rust (CLR) causes losses of one to two billion US dollars annually and is one of

the main limiting factors of Arabica coffee (Coffee arabica) production worldwide.

CLR was first recorded by an English explorer in 1861 near Lake Victoria on wild

Coffee species. The disease symptoms and signs include large orange spore masses on

the lower leaf surface, leading to premature leaf fall. The two main cultivated coffee

species, C. canephora (Robusta coffee) and C. arabica, account, on average, for

40% and 60%, respectively, of the world’s coffee production (ICO, 2016). Coffee

arabica is native to the relatively dry and high-altitude areas of Ethiopia and northern

Kenya and its genetic pool is considered to have low diversity (Steiger et al., 2002).

Arabica coffee was domesticated in Yemen, and its cultivation subsequently spread to

Asia, America and other parts of Africa. Severe genetic bottlenecks during its
9
domestication have narrowed even further the genetic diversity of the crop: it is

believed that a single coffee plant from the Botanical Garden of Amsterdam was one

of the progenitors of most of the current coffee cultivars. These genetic

Fig. 1 Coffee leaf rust symptoms and signs.

the world, led to the absence of selection pressure towards rust resistance (Rodrigues

et al., 1975). Coffee germplasm disseminated from Yemen was most probably free of

rust. Further selection and adaptation to other regions and climates in different parts

of Asia and America during the 17th and 18th centuries occurred in the absence of the

pathogen, but nonetheless under disease-favorable conditions. The 19th century

epidemic in Ceylon was the outcome of such genetic, biological and agronomic. Since

then, rust has spread to most coffee-growing countries worldwide, in the second half

of the 20th century, the identification and characterization of ‘Hıbrido de Timor’

populations provided the basis for a breeding programmed that enabled the release of

rust-resistant cultivars in different coffee growing countries, including the Americas

(Rodrigues et al., 1975; Silva et al., 2006). Recently CLR has regained notoriety

because of a severe and widespread epidemic throughout Central America, Colombia,

Peru and Ecuador, as a result of the convergence of several agronomic, climatic and

10
economic factors Yield losses were up to 35%, with a direct impact on the income and

livelihood of hundreds of thousands of farmers and laborers.

2.3 Coffee Leaf Rust

CLR infections seldom kill the host plant, although severe infections affect the

yield in subsequent years because they hamper vegetative development and can

generate polyetic epidemics over successive seasons. Climate (including the altitude

effect), shade, soil fertility and canopy architecture influence disease severity.

Although a wealth of knowledge has been accumulated on epidemics and modeling,

disease management strategies are still frequently ineffective. This gap may be

explained by the very large number of environmental variables, their interaction at a

given moment and their cumulative effect over time (Avelino et al., 2004).

Fluctuations in the price of coffee influence producer decisions regarding crop

management, which, in turn, condition rust incidence and severity. This process seems

to have become more acute in recent decades with the liberalization of prices and of

cultivation quotas, and a sharp decrease in support of research and agricultural

extension (PEDRO TALHINHAS 1, 2016).

A direct link between agricultural intensification and disease severity was reported

during the. Since then, epidemiology has gained several analytical tools, and

multivariate studies at local, national and regional levels have contributed to a better

understanding of the role of the relevant variables in the disease outcome. Usually, the

peak of CLR epidemics occurs during fruit harvest; therefore, primary yield losses are

frequently of low importance. Secondary losses, arising from low yield as a result of

11
reduced vegetative growth caused by the previous epidemic, tend to be more important

than primary losses (Avelino et al., 1991). Modeling of rust epidemics according to

agro ecological variables has revealed the relevance of local agronomic factors,

including shade, canopy density and soil fertility, interacting with regional

environmental factors, such as rainfall (Avelino et al., 2006; Boudrot et al.,2016). At

the global level, climate change scenarios have also been analyzed in the context of

CLR, with shorter incubation periods being forecasted (Ghini et al., 2011) and disease

favoring scenarios mapped (Alves et al., 2011). Indeed, this disease has been

increasingly reported at higher altitudes in recent years (Boudrot et al., 2016; Rozo et

al., 2012).

2.4 Coffee Berry Disease

Unlike CLR, CBD is still restricted to the continent of Africa where it causes

considerable loss in Coffee arabica. Reported first in Kenya (McDonald, 1926), the

disease is now present virtually in all Arabica coffee-growing countries in the

continent. Chemical methods, resistant varieties, and, to an extent, cultural practices

can control the disease. Chemical control is very expensive. The control could run up

to a colossal sum of money in the country. Further, losses up to 40 percent can still

occur in inclement weather, despite chemical control. Consequently, on a long-term

basis, breeding for CBD resistance is the only economically viable option in Arabica

coffee. The main coffee breeding programs dedicated to CBD resistance are found in

Kenya, and to some extent in Ethiopia and Tanzania. (Teferi, 2015) CBD is a major

cause of crop loss of Arabica coffee in Africa and a dangerous threat to production

elsewhere. The disease is an anthracnose of green and ripe berries induced by

Colletotrichum Kahnawake.

12
McDonald first detected CBD in 1922 in Kenya causing about 75% crop loss (Gibbs,

1969). Since then, the disease was found in many estates of the Rift valley in Kenya.

By the 1950s CBD had established in the east, the main coffee growing areas

(Rodrigues et al., 1992). Apparently, the free movement of coffee plant materials from

CBD infected areas has been the main factor in distribution of this disease throughout

all important Arabica growing areas in Africa. The disease was reported in Angola

around 1930, Zaire in 1937, Cameroon 1955-1957, Uganda in 1959, Tanzania in 1964,

Ethiopia 1971(Van der Graaff, 1981) and in Malawi in 1985 (Lutzeyer et al., 1993).

CBD was also confirmed in Malawi, Zimbabwe and Zambia in 1985 (Masaba and

Waller, 1992). It is not known outside of Africa, although a leaf spot and ripe berry

anthracnose caused by related Colletotrichum species has been reported from

Guatemala and Brazil (Griffiths et al., 1991). In Ethiopia CBD first reported in 1971.

Then spread to all major coffee producing regions within very short period except to

the lower altitude less than 1500. All coffee production systems like plantations,

garden, Semi-forest and forest coffee, with and without shade all were infested with

the CBD in all extent. Merdassa Ejetta (1985) reported yield losses of 51% at Melko

and 81% at Wondo Genet due to

CBD. In 1994 crop season prevalence of CBD was conducted in Oromiya Region and

Southern Nations Nationalities and Peoples Region (SNNPR) and the result indicated

38.8 and 17.2% of mean percent prevalence of the disease, respectively (IAR, 1997).

According to the result, CBD pressure was very high at higher altitudes in the southwest

region, while severe disease was recorded in valleys of Sidama zone. The CBD/Coffee

Berry Disease/ severity varied from year to year and among areas. In Amhara region of

Northern Ethiopia where CBD occurs, survey result showed that an average severity for

the 1996/97-crop season was 38% (TesfayeAlemu and Ibrahim Sokar, 2000). Survey

13
conducted in 1997 and 1998 in six major coffee growing zones (in 32 Districts) of

Oromiya region showed an average of 31% and 32% disease severity for the respective

years (Melaku Jirata and Samuel Assefa, 2000). CBD incidence and severity assessment

in 10 zones and 31 woredas of Southern Nations Nationalities and Peoples Region

(SNNPR), conducted in September 1998, resulted with 40% and 22.8% mean incidence

and severity of the disease, respectively (Tesfaye Negash and Sinedu Abate, 2000).

Following the advent of CBD and modernization of the crop production system leads to

the replanting of limited number of CBD resistant cultivars, which brought deforestation

and rehabilitation of diverse coffee population. Side by side modern cultural practices

are widely employed. Coffee berry disease alone is known to reduce coffee yields

between 2530 %. Hararghe coffee is susceptible to CBD; as a result, it is under threat of

genetic erosion mainly because of the losses caused by the disease and farmers prefer

growing alternative cash crops such as chat to planting coffee. Climatic factor

significantly influences the occurrence and distribution of CBD. Adugna (1995) stated

that the prevalence of low temperature in conjunction with high rainfall total and longer

number of rainy days could trigger CBD infection. The existing knowledge also

revealed that the developing coffee berry is most susceptible in the course of rapid

expansion stage. Nutman and Roberts (1960) reported that temperature between 17 and

22 0C in the presence of water or completely saturated atmosphere are responsible for

moderate CBD infection under laboratory condition. Similarly, Masaba and Vander

Vosen (1982) observed formation of scabs on some CBD resistance coffee cultivars

under such weather conditions favoring CBD. A partial regression-correlation analysis

on disease parameters (severity and incidence of CBD recorded for ten years on

progenies of thirteen CBD resistant selections and a susceptible standard at Gera (CBD

hot spot area) as dependent variable against major weather factors such temperature,

14
rainfall and relative humidity revealed prevalence of low temperature about 21.50c

accompanied by high rainfall (>700mm) extended over long period (>74days) favored

CBD development and increased disease intensity (Adugna, 1995). Analytical result of

the relationship between climatic variable and percentage of CBD infected berries of

two coffee cultivars 741 and 7440 grown at Gera show that the variability of infection

function of seasonal and special weather condition. April to July mean daily maximum

temperature with r2 of 37% and

36%.

2.5 Coffee Wilt Disease

Coffee wilt disease is a common wilt that results in complete death of coffee

trees it infects. This vascular disease is induced by the fungal pathogen. In 1927 coffee

wilt disease was first observed in the Central African Republic where it developed

slowly and went on to cause two epidemics between the 1930s and the 1960s. Hosts

of coffee wilt disease include Coffee arabica (Arabica coffee), Coffee canephora

(Robusta coffee), Coffee liberica (Liberian coffee), and Coffee excels (Excelsa

coffee). Currently, the disease is limited to Eastern and Central Africa; however,

studies have shown that most Coffee species are likely to be susceptible to the disease,

which can potentially lead to more worldwide problems in the coffee industry. Due to

the nature of coffee wilt disease, coffee plants often exhibit symptoms of disruption to

vascular systems. Internal symptoms are disturbances to conduction of water in the

plant. External symptoms include loss of moisture on leaves, discoloration, leaf loss,

dieback of the infected region, swelling of trunks, cracks in mature trees and lastly

plant death. Signs of coffee wilt disease include small blackish brown perithecia

15
caused by the sexual stage of the fungus, and cracks in the bark which cause an

observable bluish-black stain on the wood.

2.6 Knowledge Based Identification of Coffee Disease

From each of the three selected district, three coffee growing peasant associations

were selected. From each of the three peasant associations, three coffee fields were

randomly selected at intervals of 1-2 km along the main and accessible rural roads and

assessed for the disease survey. The assessment was conducted on a total of 10 randomly

taken plants within each field. Random selection of the plants for assessment was done by

moving in ‘X’ fashion within the field by taking plants at a distance of about 6-8 m intervals

based on the size of the field. Sample size consisting of all the berries in four simple

branches of 4-6 plants was reported to have least coefficient of variation in a plot of 100

plants. Based on these, stratified sampling We.e., three pairs of branches from each upper,

middle and lower canopy layers of the coffee plant were selected to assess CBD severity.

The occurrence and severity of CBD was recorded from each sample plants. The incidence

of the disease within each coffee field was then determined as the percentage of diseased

plants out of the 10 assessed plants. Likewise, CBD severity was computed as the average of

10 plants by estimating the percentage of diseased berries (damaged barriers over on all

barriers of sampled branch times 100) from each of the plants assessed. The scores were

changed into Percentage Severity Index (PSI) for the analysis using the formula of Wheeler.
16
For each field assessed, other diseases of coffee, variety type, and plant age and growth

habit were noted. The experiment was structured as nested design (tree was nested under

location). Statistical analysis: The mean incidence and severity of CBD were calculated for

each district, peasant association and field. Data were then subjected to nested random

effects analysis of variance. (Jambo, 2015)

2.7 Statistical Local Binary Approach

(Lucas ximenes Boa Sorte, 2019) This research aims to apply computational

methods to recognize main diseases in coffee leaves, with the purpose to implement

an expert system to assist coffee producers in disease diagnosis during its initial

stages. Since these two diseases are shapeless, it inspires a texture attribute extraction

approach for pattern recognition. Two texture attributes were considered in this work:

statistical attributes and local binary patterns. Surprisingly, this second approach

showed better results than the texture extraction method. It could be explained by the

small number of diseases he aimed to recognize and a sufficient number of training

samples used during the deep learning process.

2.8 Coffee Bean Grading Using Visual Sensor

(Faridah*1, 2011) The objective of the research is to develop a system to determine

the coffee bean grading based on SNI No: 01-2907-1999. A visual sensor, a webcam

connected to a computer, was used for image acquisition of coffee bean image

17
samples, which were placed under uniform illumination of 414.5+2.9 lux. The

computer performs feature extraction from parameters of coffee bean image samples

in the term of texture (energy, entropy, contrast, homogeneity) and color (R mean, G

mean, and B mean) and determines the grade of coffee bean based on the image

parameters by implementing neural network algorithm. The accuracy of system testing

for the coffee beans of grade WE, II, III,IVA, IVB, V, and VI have the value of 100,

80, 60, 40, 100,

40, and 100%, respectively.

But, it’s not an easy task to implement on disease identification because on bean

grading mostly color and some feature extraction matters if we directly implement to

identification of disease we could not come up with good result. To get a better result

WE should have to work with unsupervised approaches by manipulating huge datasets

diseased coffee leaves.

2.9 RBFN/Radial Base Function Network/ Approach and

SOM

“Recognition of Facial Expression by Image Processing and Radial Basis Function

Network”. In this work he reported the application of a neural network-type Radial Basis

Function in the recognition of facial expressions. The facial images were previously

segmented to reduce the amount of data. The results were satisfactory for the description of

seven basic facial expressions of the human facial recognition to be applied in the automation

processes. The facial expression recognition is a complex field of investigation and was

subject of considerable research [Hager and Eckman 1983, Faigin 1990, Freitas Magalhães

18
2007]. There are many types of facial expressions, but there are six basic facial expressions

for facial analysis

[Ekman and Friesen 1978, Gil 1993]: happiness, fear, sadness, surprise, disgust, and anger.

These six expressions together with the neutral emotion expression serve as means of

communication in most of the cultures. After the image processing, the total pixel number

were reduced to around 20-30%. A preliminary study of the application described above was

conducted by using only two output channels (y1, y2). Figure 1 show the results of training in

about 50 images. WE observed that the outputs are grouped in clusters associated with facial

expressions. This result, although preliminary, has given support to consider this a promising

method

Figure

1. Radial Base function distribution.

Disadvantage of RBF: though the training is faster in RBF network but classification

is slow in comparison to Multi-layer Perceptron due to fact that every node in hidden

layer have to compute the RBF function for the input sample vector during

classification.

The various methodologies were proposed earlier for identification and detection of

agriculture pests. Mostly work was done for identification of whitefly pest on sticky

19
traps in greenhouse environments and in real fields. Diagnosis of agricultural pests in

the field is very critical and difficult. We propose a decision support system which

exposes advance computing technology that has been developed to help the farmer to

identify agricultural pests and take proper decision about preventive or control

measure on it. In my proposed work, we will collect images of pests from various

fields. These images will be preprocessed for enhancement. Then segmentation will be

carried out for extraction of pest from foreground of image. After segmentation,

various features of pests including color and shape will be extracted. Using radial

basis function neural network classifier, we would be identified name of pest and give

preventive and control measures to farmer.

2.10 K-Means Clustering Image Classification Approach

(Feng Qin1, 2016) In this study, the identification and diagnosis of the four types of alfalfa

leaf diseases were investigated using pattern recognition algorithms based on image

processing technology. A sub-image with one or multiple typical lesions was obtained by

artificial cutting from each acquired digital disease image. Then the sub images were

segmented using twelve lesion segmentation methods integrated with clustering

algorithms (including K-Means clustering, fuzzy C-means clustering and KMedian

clustering) and supervised classification algorithms (including logistic regression analysis,

Naive Bayes algorithm, classification and regression tree, and linear discriminant

analysis). After a comprehensive comparison, the segmentation.

(Maity)In this paper, the researcher have addressed that problem and proposed an efficient

method to detect leaf disease. Leaf diseases can be detected from simple images of the

leaves with the help of image processing and segmentation. Using k-means clustering and

20
Otsu’s method the faulty region in a leaf is detected which helps to determine proper

course of action to be taken. Further the ratio of normal and faulty region if calculated

would be able to predict if the leaf can be cured at all. In this paper, he has addressed the

major problem of different types of leaf diseases of the crops. With the help of image

processing, he successfully segmented the leaf images and classified them into the faulty

and normal region. Due to this segmentation, it becomes easier to detect if any leaf

diseases affect the crops just with the help of an digital image. he used the unsupervised

learning method by implementing the k-means clustering algorithm for segmentation of

the faulty region in leaves which is way more accurate than the existing methods. This

paper throws several new challenges to us, on which we are currently working. WE have

worked to improve the accuracy further and determine a ratio of the normal and the faulty

regions. This would be helpful to diagnose the disease and cure process to be effective and

quick. It would also show if the disease not curable from the ratio and will help to set the

further course of action. • (Rupali Patil1, Apr-2016) this paper present survey on different

classification techniques that can be used for plant leaf disease classification. A

classification technique deals with classifying each pattern in one of the distinct classes. A

classification is a technique where leaf is classified based on its different morphological

features. There are so many classification techniques such as k-mean clustering, Support

Vector Machine. Selecting a classification method is always a difficult task because the

quality of result can vary for different input data. Plant leaf disease classifications have

wide applications in various fields such as in biological research, in Agriculture etc. Plant

disease detection is emerging field in India as agriculture is important sector in Economy

and Social life. Earlier unscientific methods were in existence. Gradually with technical

and scientific advancement, more reliable methods through lowest turnaround time are

developed and proposed for early detection of plant disease. Such techniques are widely
21
 used and proved beneficial to farmers as detection of plant disease is possible with

minimal time span and corrective actions are carried out at appropriate time. The detection

of plant disease is significantly based on type of family plants and same is carried out in

two phases as segmentation and classification

Fig 2.2 K-Means Clustering System Architecture.

2.11 SVM/Support Vector Machine/ Approach on

Image Classification
‘Detection of Plant Leaf Diseases Using Image Segmentation and Soft Computing Techniques’ (Vijai

Singh, 2017) This paper presents an algorithm for image segmentation technique which is used for

automatic detection and classification of plant leaf diseases. It also covers survey on different

diseases classification techniques that can be used for plant leaf disease detection. Image

segmentation, which is an important aspect for disease detection in plant leaf disease, is done by

using genetic algorithm.

22
In this paper steps used to identify disease image is illustrated as below:

1. Image acquisition is the very first step that requires capturing an image with the help of

a digital camera.

2 Preprocessing of input image to improve the quality of image and to remove the

undesired distortion from the image. Clipping of the leaf image is performed to get

the interested image region and then image smoothing is done using the smoothing

filter.

To increase the contrast Image enhancement is also done.

3. Mostly green colored pixels, in this step, are masked. In this, we computed a

threshold value that is used for these pixels. Then in the following way mostly

green pixels are masked: if pixel intensity of the green component is less than the

pre-computed threshold value, then zero value is assigned to the red, green and

blue components of this pixel.

4. In the infected clusters, inside the boundaries, remove the masked cells.

5. Obtain the useful segments to classify the leaf diseases. Segment the components

using genetic algorithm

6. Computing the features using color co-occurrence methodology

(Vijai Singh, 2017) For experiment he performed all experiments in MATLAB. For

input data disease, samples of plant leaves like rose with bacterial disease, beans leaf

with bacterial disease, lemon leaf with Sun burn disease, banana leaf with early scorch

disease and fungal disease in beans leaf are considered. Segmented image can be

classified into different plant diseases. In the same manner classification of diseases of

other input plant leaves. The co-occurrence features are calculated after mapping the

R, G, B components of the input image to the thresholded images. The co-occurrence

features for the leaves are extracted and compared with the corresponding feature
23
values that are stored in the feature library. First, he has done classification using the

Minimum Distance Criterion with KMean Clustering and shows its efficiency with

accuracy of

86.54%. The detection accuracy is improved to 93.63% by proposed algorithm which

is Support vector machine approach. In the second phase classification is done using

SVM classifier and shows its efficiency with accuracy of 95.71%. Now the detection

accuracy is improved to 95.71% by SVM with proposed algorithm. From the results it

can be seen that the detection accuracy is enhanced by SVM with proposed algorithm

compared to other approaches reviewed in 2.6(K-Means Clustering Approach)

Gupta proposed an autonomously modified SVM-CS (Cuckoo Search) model to

Identify the healthy portion and disease. Using a dataset of diseases containing plant

leaves suffering from Alternaria Alternata, Cercospora Leaf Spot, Anthracnose, and

Bacterial Blight, along with healthy leaf images, the proposed model was trained and

optimized using the concept of a cuckoo search. However, identification and

classification approaches of these studies are semiautomatic and complex, and deal

with a series of image processing technologies. At the same time, it is very difficult to

accurately detect the specific disease images without extracting and designing the

appropriate classification features depending heavily on expert experience proposed a

novel identification approach for rice diseases based on deep convolutional neural

networks. Using a dataset of 500 natural images of diseased and healthy rice leaves and

stems, (Monzurul Islam, 2017) presented an approach that integrated image processing

and machine learning to allow the diagnosis of diseases from leaf images. In this

research has done Experimental simulation of combination of RBF and SOM and he

has found a better performance than the other classifiers. But when he sees the training

time of the combination of RBF and SOM, it takes longer time in training. This
24
automated method classifies diseases on potato plants from ‘Plant Village’, which is a

publicly available plant image database. The segmentation approach and utilization of

an SVM demonstrated disease classification in over 300 images, and obtained an

average accuracy of 95%.

2.12 BPNN/Back Propagation Neural Network/

“An Automatic Coffee Plant Diseases Identification Using Hybrid Approaches of Image

Processing and Decision Tree‟‟ (Debasu, 2018) This paper presents the identification

of these types‟ diseases using hybrid approaches of image processing and decision

tree. The images are taken from Southern Ethiopia, Jimma and Zegie. In this paper

back propagation artificial neural network (BPNN) and decision tree had been used as

techniques; a total of 9100 images were collected. During his research o train, the

classifiers, a set of training diseased coffee image was given to the model in addition

to the class label of Ethiopian coffee plant image. From the total of 9100 data sets,

6370 were used for model training and 2730 were used for performance testing. In this

research, there are three output classes, because the coffee plant diseases type was

three. The representing features of training were normalized with mean 0 and variance

1 this helps the model to converge. The neural network needs 17 inputs of the

combined feature vectors of GLCM, COLOR and STATISTICAL and 3 neurons in its

output layer to classify the type. The hidden layer has 17 neurons this number was

picked by trialand-error methods, if the network has trouble of learning capabilities,

and then neurons can be added to this layer. There is a significant change when we

25
increase the number of hidden layers neurons until 17 but there is no change when the

number of hidden layer neurons increases above 17.

Each value from the input layer is duplicated and sent to all of the hidden nodes. The

result indicated that there was 94.5% using BPNN with tanh sigmoid activation

function. After conducting the above experiments 94.5% success achieved when back

propagation artificial neural network with tanh activation function is combined. In this

paper the design of coffee plant disease is very good as show on fig But, the actual

performance of back propagation on a specific problem is dependent on the input data

that means the quality of the input image should be good which is very exposed to

wrong result. Back propagation can be quite sensitive to noisy data. You need to use

the matrixbased approach for back propagation instead of mini-batch.

Figure2.3. Coffee Plant Disease Identification Model (Debasu, 2018)

26
 2.13 CNN/Convolutional Neural Network/
(al, 2016) They have presented an approach based on CNN to recognize apple

pathologic images, and employed a self-adaptive momentum rule to update CNN

parameters. The results demonstrated that the recognition accuracy of the proposal

was up to 96.08%, with a fairly quick convergence.

(Kemin Du, 2018) A novel cucumber leaf disease detection system was presented

based on convolutional neural networks. Under the four-fold cross-validation

strategy, the proposed CNN-based system achieved an average accuracy of 94.9% in

classifying cucumbers into two typical disease classes and a healthy class. The

experimental results indicate that a CNN-based model can automatically extract the

requisite classification features and obtain the optimal performance.

‘Going deeper with convolutions’ (Erhan, June 2015) We propose a deep convolutional

neural network architecture codenamed Inception, which was responsible for setting

the new state of the art for classification and detection in the ImageNet LargeScale

Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this

architecture is the improved utilization of the computing resources inside the network.

27
This was achieved by a carefully crafted design that allows for increasing the depth

and width of the network while keeping the computational budget constant. To

optimize quality, the architectural decisions were based on the Hebbian principle and

the intuition of multi-scale processing. One particular incarnation used in our

submission for ILSVRC14 is called Google Net, a 22 layers deep network, the quality

of which is assessed in the context of classification and detection.

(Deep Neural Network Based Recognition of Plant Diseases by Leaf Image

Classification, 2016) This paper is concerned with a new approach to the

development of plant disease recognition model, based on leaf image classification,

by the use of deep convolutional networks. Novel way of training and the

methodology used facilitate a quick and easy system implementation in practice. The

developed model is able to recognize 13 different types of plant diseases out of

healthy leaves, with the ability to distinguish plant leaves from their surroundings.

The researcher has used Artificial Neural Network and Convolutional Neural

Network and he has tested total of 30880 images, for testing 2538 has got promising

result 96.3% of accuracy.

Using Deep Learning for Image-Based Plant Disease Detection‟ (Sharada Prasanna

Mohanty1, 2016) Digital Epidemiology Lab, EPFL, Switzerland; School of Life

Sciences,EPFL, Switzerland; Department of Entomology, College of Agricultural

Sciences, Penn State University, USA; Department of Biology, Eberly College of

Sciences, Penn State University, USA; Center for Infectious Disease Dynamics, Huck

Institutes of Life Sciences, Penn State University, USA; School of Computer and

Communication Sciences, EPFL, Switzerland This manuscript was compiled on

28
April 15, 2016 the approach of training deep learning models on increasingly large

and publicly available image datasets presents a clear path towards smartphone-

assisted crop disease diagnosis on a massive global scale.

2.14 HD-CNN/Hierarchical Deep Convolutional

Neural Network/

“HD-CNN: Hierarchical Deep Convolutional Neural Networks for Large Scale Visual

Recognition”. (Zhicheng Yan†, 2015), In this paper, the researcher introduces

hierarchical deep CNNs (HD-CNNs) by embedding deep CNNs into a two-level

category hierarchy. An HD-CNN separates easy classes using a coarse category

classifier while distinguishing difficult classes using fine category classifiers.

During HDCNN training, component-wise pretraining is followed by global

finetuning with a multinomial logistic loss regularized by a coarse category

consistency term. In addition, conditional executions of fine category classifiers

and layer parameter compression make HD-CNNs scalable for largescale visual

recognition we achieve state-of-the-art results on both CIFAR100 and large-scale

ImageNet 1000-class benchmark datasets. Experimentally, he builds up three

different twolevel HD-CNNs, and they lower the top-1 error of the standard CNNs

by 2.65%,3.1%, and 1.1%.(Sanat Bhandarkar, July,2019) “Deep Learning and

Statistical Models for Detection of White Stem “The researchers explore

Unsupervised Autoencoder models, which can work with limited image datasets.
29
In addition, statistical analysis of long-term climatic factors such as temperature,

rainfall, humidity and luminescence is explored for reliable detection and diagnosis

of the infestation. Based on the encouraging results, a mobile application is

proposed for near real time monitoring of WSB infestation to help the coffee

planters. An image library was created representing different stages of the plant

infestation using camera/mobile devices. The Convolutional Neural Network

(CNN) models use these images of healthy and infested plants for early detection

of white stem borer infestation. The overall methodology included image

processing, machine learning, supervised transfer learning and unsupervised auto-

encoding techniques to solve the problem of early detection and severity of the

infestation. Using the Inception v3 transfer learning model, he obtained average

accuracy of 85.5% which is quite encouraging with limited image dataset. But one

of the limitations of this research is limitation of image dataset. the other drawback

is though the research has chosen a good neural approach he did not implement a

more powerful neural network architecture like resnet50, ANN, Google Net etc.

but it still has put great foundation to look forward for other researchers like me.

2.15 Multiple Linear Regression Approach

“Plant Diseases Recognition Based on Image Processing Technology‟

(GuilingSun, 2018) a new image recognition system based on multiple linear

regression is proposed. Particularly, there are a number of innovations in image

segmentation and recognition system. In image segmentation, an improved

histogram segmentation method which can calculate threshold automatically and

accurately is proposed. Meanwhile, the regional growth method and true color

30
 image processing are combined with this system to improve the accuracy and

intelligence. While creating the recognition system, multiple linear regression and

image feature extraction are utilized. After evaluating the results of different image

training libraries, the system is proved to have effective image recognition ability,

high precision, and reliability.

2.16 KNN/K-Nearest Neighboring/ Approach

(Hossain, 2019) This paper proposed a technique for plant leaf disease detection and

classification using K-nearest neighbor (KNN) classifier. The texture features are extracted

from the leaf disease images for the classification. In this work, KNN classifier will classify

the diseases like alternaria alternata, anthracnose, bacterial blight, leaf spot, and canker of

various plant species. The researcher uses only 200 image of dataset which is very minimum

number of images to get more accurate result the reason for this is its very impractical to

implement KNN for huge amount of dataset because it will be very slow KNN’s main

disadvantage of becoming significantly slower as the volume of data increases makes it an

impractical choice in environments where predictions need to be made rapidly. Moreover,

there are faster algorithms that can produce more accurate classification and regression

results.

2. 17 Image Classification Architectures

2.17.1 ResNet50

ResNet-50 is a convolutional neural network that is 50 layers deep. You can

load a pre-trained version of the network trained on more than a million images from

the Image Net database. The pre-trained network can classify images into 1000 object

31
categories. As a result, the network has learned rich feature representations for a wide

range of images. The network has an image input size of 224-by-224.

2.17.2 Google Net

The architecture of Google Net, proposed by Szegedy et al (Sultana)., is

different from conventional CNN. They have increased the number of units in each

layer using parallel filters called inception module of size 1 × 1, 3 × 3 and 5 × 5 in

each convolution layer (Conv layer). They have also increased the layers to 22. While

designing this model, they have considered the computational budget fixed. So that

the model can be used in mobile and embedded systems. They have used a series of

weighted Gabor filters of various sizes in the inception architecture to handle multiple

scales. To make the architecture computationally efficient they have used inception

module with dimensionality reduction instead of the naive version of inception

module.

Despite 22 layers, the number of parameters used in Google Net is 12 times lesser

than Alex Net but its accuracy is significantly better. All the convolution, reduction

and projection layers use ReLU nonlinearity. They have used average pooling layer

instead of the fully connected layers. On top of some inception modules, they have

used auxiliary classifiers which are basically smaller CNNs, to combat vanishing

gradient problem and overfitting

32
Fig 2.4 GoogleNet Architecture

2.17.3 Alex Net

With the recent launch of remote sensing satellites around the world, a large volume

of multi-level, multi-angle, and multi-resolution HSR remote sensing images can now be

obtained, where the remote sensing big data brings new understandings for the traditional

definition of big data. These multi-source remote sensing images allow the ground object

observation from multiple perspectives. The rapid development of HSR remote sensing

imaging sensors has provided us with a large number of HSR remote sensing images with

abundant detail and structural information, and a higher spatial resolution. (Han, 2017)

Alex Net, which was first proposed by Alex Krizhevsky et al. in the 2012 ImageNet

Large Scale Visual Recognition Challenge (ILSVRC-2012), is a fundamental, simple, and

effective CNN architecture, which is mainly composed of cascaded stages, namely,

convolution layers, pooling layers, rectified linear unit (ReLU) layers and fully connected

layers. Specifically, Alex Net is composed of five convolutional layers, the first layer, the

second layer, the third layer and the fourth layer followed by the pooling layer, and the fifth

layer followed by three fully-connected layers. For the Alex Net architecture, the

convolutional kernels are extracted during the back-propagation optimization procedure by

optimizing the whole cost function with the stochastic gradient descent (SGD) algorithm.
33
Generally, the convolutional layers act upon the input feature maps with the sliding

convolutional kernels to generate the convolved feature maps, and the pooling layers

operate on the convolved feature maps to aggregate the information within the given

neighborhood window with a max pooling operation or average pooling operation. The

reason why Alex Net is successful can be attributed to some of the practical strategies, for

instance, the ReLU non-linearity layer and the dropout regularization technique. The

dropout technique can be regarded as a kind of regularization by stochastically setting a

number of the input neurons or hidden neurons to be zero to reduce the co-adaptations of

the neurons, which is usually utilized in the fully connected layers in the Alex Net

architecture. f(x) = max(x, 0) (1) The transfer mechanism and the pre-training mechanism

allow the CNN network parameters to be transferred from natural imagery datasets to HSR

remote sensing imagery datasets. The reason why this can succeed can be explained, to

some extent, by the similarities between natural imagery datasets and remote sensing

imagery datasets, and the category compatibility. It can also be easily understood that the

large and complicated Image Net datasets can help to obtain a well-trained Alex Net

architecture, and well-trained network parameters are important for initializing the

subsequent classification framework. Therefore, the pre-training mechanism helps the

AlexNet architecture to perform the HSR remote sensing imagery scene classification task.

Based on the introduction of the convenient and comprehensive representation ability of the

pre-trained Alex Net architecture in dealing with HSR remote sensing imagery scene

classification, the pre training mechanism also makes the Alex Net architecture an end-to

end classification pipeline. The pre-trained Alex Net network architecture is shown in

Figure 1.

34
 .

fig 2.5 Alex net architecture

2.13 RELATED WORK

2.13.1 SIMILARITY AND DIFFERENCE OF OBJECT DETECTION AND

IDENTIFICATION

2.13.1.1 Object Detection/Recognition/

Object recognition is to describe a collection of related computer vision tasks

that involve activities like identifying objects in digital photographs. Image

classification involves activities such as predicting the class of one object in an

image. Object localization is referring to identifying the location of one or more

objects in an image and drawing an abounding box around their extent. Object

detection does the work of combines these two tasks and localizes and classifies one

or more objects in an image. When a user or practitioner refers to the term “objects

recognition “, they often mean “objects detection “. It may be challenging for

beginners to distinguish between different related computers.

2.13.1.2 Object Identification

The aim of object detection is to detect all instances of objects from a known

class, such as people, cars or faces in an image. Information is more detailed and
35
 contains the parameters of a linear or non-linear transformation. For example, for face

detection in a face detector may compute the locations of the eyes, nose and mouth, in

addition to the bounding box of the face

Table 2.1 summary of literature review

N Research Features Machine Learning Dataset Accuracy
o Extraction
Worked by approach used

1 (Isotani, 2009) ANN Radial Base Function 17000 70%

Network/RBFN/

2 (Vijai Singh, color co-occurrence Support Vector 300 95.71%

methodology images
2017 Machine/SVM/

3 K-Means

4 (Hossain, GLCM and image K-Nearest 237 leaf 96.76%

2019) segmentation Neighbor/KNN/ images

5 (Erhan, June GoogleNet Convolutional Neural 100000 93%

Network/CNN/
2015)

36
 6 Debasu, 2018 GLCM Back Propagation Neural 9100 94.5%

Network/BPNN/

7 (GuilingSun, histogram Multiple Linear 40 90%

segmentation Regression
2018 method
Approach

CHAPTER THREE 3. METHODS AND

METHODOLOGY

3.1. Proposed System Architecture

Digital camera or similar devices are used to take images of leaves of different types, and

then those are used to identify the affected area in leaves. Then different types of image

processing techniques are applied on them, to process those images, to get different and

useful features needed for the purpose of analyzing later. Algorithm written below

illustrated the step by step approach for the proposed image recognition and segmentation

processes: For the development of this research, the methodology set forth by Roldán et al

(2019) mentioned above was taken as a reference, which has 5 steps.

• Image acquisition: It is the most important step in the investigation and the most careful,

since the data entered into the model must avoid generating noise within it. At this stage,

the photos are generated or obtained from a repository, ideally in controlled

environments.

37
• Preprocessing: It is the set of techniques that are applied to digital images in order to

avoid noise within the model or highlight characteristics, these techniques facilitate the

search for information for the model. In other words, an image is generated from an

image that contains certain characteristics that are more significant for the model. The

main objectives sought with this phase are: Smooth the image, eliminate noise, enhance

edges and detect edges.

• Segmentation: It is the process of dividing an image into several parts, each pixel is

assigned a category, these vary according to the chosen segmentation (binary, assisted,

semantic, instantiated etc.). The objective of this stage is to locate regions that have

meaning for the model

• Characteristics extraction: This is a phase where a set of raw data is reduced to

characteristics that are more manageable for the human being or significant for the model.

This facilitates the learning and generalization of the models, since it reduces the

resources used to process the information.

• Classification: It is the step where the label of the images is predicted with the help of

the chosen model, this model will be the one that obtains the best results in the

crossvalidation.

38
Fig 3.1 faster RCNN diagram (towardai.net)

3.3 Data Collection and Dataset

These days the digital device which captures images are coming to be very complex

and sophisticated and these technologies are used to take the image of those leaves

infected by disease in the dataset collection. To process those images different types of

image processing methods and models are implemented, to extract different and useful

features needed for the purpose of processing later Infected wilt disease leaves with

typical symptoms and signs. The images are collected from jimma by different

39
 expertise, the place where coffee arabica exists in abandoned. The infected coffee

leaves are gathered and samples from different levels of infection stages. When

capturing images, the leaves were stretched as straight as possible, and the camera lens

was mostly parallel with the plane of the leaves which is effect for analyzing and

processing images. A total of 11340 images with typical disease symptoms were

collected, (jpg format) from by Ministry of Agriculture of south west shoa zone. we

are going to use the collected image for testing purpose.

3.4 Pre-processing

Non-regularized Convolutional Neural Networks has a high capacity for learning. This

means every data weather it is noise data or legitimate data it can be learnt equally. Thus, a

cleaned and structured large dataset is required. Data Augmentation should address and fix-

lack of data and presence of noisy data. The input data has a direct relationship with the ability

of any model. Accuracy significantly can Increase an existing dataset with the right instances

of the domain being studied. Sample datasets require growth in terms of both quality and

quantity. For the dataset Augmentation Artificial neural network (ANN) could be

implemented. These models like ANN have the ability to effectively train and learn the model

of data in an unsupervised way. There are two kinds of sub-models of Auto encoders these are

an encoder and a decoder. The encoder submodel plays major role compressing the input

image dimensions by applying compression. This compressed image is then given as input to

the decoder sub-model, which tries to rebuild the original image from the compressed

representation. This representation can entirely be used to generate new, different data from

existing data. This effectively brings the outcome to augmentation of the existing dataset with

new, valid instances of data without the need for gathering and collecting more data from

40
external sources. A small dataset can be increased in size radically. For below

implementations, simple data augmentation methods of rotating, inverting, sheering has been

used, to increase the lesser numbered classes.

3.5 Classification Techniques

3.6.1 K-Means Clustering Model

K-means is an unsupervised classification algorithm, also called cauterization

that groups objects into k groups based on their characteristics. The grouping is done

minimizing the sum of the distances between each object and the group or cluster

centroid.

The distance usually used is the quadratic or Euclidean distance.

The algorithm has three steps:

1. Initialization: once the number of groups, k has been chosen, k centroids are established in

the data space, for instance, choosing them randomly.

2. Assignment of objects to the centroids: each object of the data is assigned to its nearest

centroid.

3. Centroid’s update: The position of the centroid of each group is updated taking as the new

centroid the average position of the objects belonging to said group.

Repeat steps 2 and 3 until the centroids do not move, or move below a threshold distance in each

step

3.5.2 K-Nearest Neighboring/KNN/ model

The k-Nearest Neighbor classifier is by far the simplest machine learning/image

41
classification algorithm. In fact, it’s so simple that it doesn’t actually “learn” anything.

Inside, this algorithm simply relies on the distance between feature vectors, much like

building an image search engine — only this time, we have the labels associated with

each image so we can predict and return an actual category for the image.

Simply put, the k-NN algorithm classifies unknown data points by finding the most

common class among the k-closest examples. Each data point in the k closest

examples casts a vote and the category with the most votes win.

3.5.3 Support Vector Machine/SVM/ model

Support vector machines are supervised learning models with associated learning

algorithms that analyze data used for classification and regression analysis. Yes, you read it

right… It can also be used for regression problems. We will look at the power of SVMs for

classification.

3.5.4 Convolutional Neural Networks

CNNs are Deep Learning models which has the capability the input data to extract

features of the image, without the need for implementation of sophisticated feature

extraction. Since there is no need for advanced feature extraction technique, CNNs are

highly adoptable at handling very complex and conv features, and making them

extremely capable in image classification. Deep Convolutional Neural Networks

(DCNNs) are CNNs with a large number of convolutional layers. These multiple

layers function together to build a huge and complex feature space. The complexity of

feature learning increases in the traversal of the network. At the initial few layers,

42
lower order features for instance like blob or detection of the edge are learnt and

higher order features are step by step learnt over other layers. The final layer features

are fed into the classifier, which may consist of one or more layers

3.5.5 Fast Regional Convolutional Neural Network /Fast RCNN/

In this approach the input image pass to the network, and it is then sent through

various convolutions and pooling layers. Finally, we get the output in the form of the

object’s class. For each input image, we get a corresponding class as an output. So we

can use this technique for detection and identification of objects. Let’s see how it will

go to solve a general object detection problem using a CNN.

1. First, we take an image as input:

2. Then divide the image into various regions selective search algorithm,

3. Then pass it through neural network architectures

4. Extract Features

5. Implement SoftMax Probability.

In Real world implementation of RCNN has high computational time. Instead of

running a CNN 2,000 times per image, we can run it just once per image and get all

the regions of interest (regions containing some object). Ross Girshick, the author of

RCNN, came up with this idea of running the CNN just once per image and then

finding a way to share that computation across the 2,000 regions. In Fast RCNN, we

feed the input image to the CNN, which in turn generates the convolutional feature

maps. Using these maps, the regions of proposals are extracted. We then use a RoI

pooling layer to reshape all the proposed regions into a fixed size, so that it can be fed

43
into a fully connected network.Let’s break this down into steps to simplify the

concept:

As with the earlier two techniques, we take an image as an input.

This image is passed to a ConvNet which in turns generates the Regions of Interest. A

RoI pooling layer is applied on all of these regions to reshape them as per the input of

the ConvNet. Then, each region is passed on to a fully connected network.A SoftMax

layer is used on top of the fully connected network to output classes. Along with the

SoftMax layer, a linear regression layer is also used parallelly to output bounding box

coordinates for predicted classes. So, instead of using three different models like in

RCNN, Fast RCNN uses a single model which extracts features from the regions,

divides them into different classes, and returns the boundary boxes for the identified

classes simultaneously.

Steps to implement Fast RCNN.

1. Taking an image as input

2. Then we apply the RoI pooling layer on the extracted regions of interest to make sure all the regions

are of the same size:

3.This image is passed to a ConvNet which returns the region of interests accordingly:

4. Finally, these regions are passed on to a fully connected network which classifies

them, as well as returns the bounding boxes using SoftMax and linear regression

layers simultaneously: This is how Fast RCNN resolves two major issues of RCNN,

passing one instead of 2,000 regions per image to the ConvNet, and using one instead

of three different models for extracting features, classification and generating

bounding boxes.

44
 3.6 Feature Extraction
Getting information from the raw data is a very essential process to analyze the image.

To do this we have to apply the feature extraction. Selecting the most meaningful

features is a crucial step in the process of classification problems because from the

initial set of data It is necessary to find all possible feature subsets that can be

formed.The idea of R-CNN was to use a Selective Search (SS) approach to propose

around 2000 Regions-Of-Interest (ROI), which were then fed into a Convolutional

Neural Network (CNN) to extract features. For a more detailed explanation, see this

piece. This was quickly followed by Fast R-CNN, a faster and better approach of

object detection, in early 2015. Fast R-CNN used an ROI pooling approach, which

shares the features across the whole image and uses a modified form of spatial

pyramid pooling method to extract features in a computationally efficient way.

3.7. Regional Convolutional Neural Network

Instead of working on a massive number of regions, the RCNN algorithm

proposes a bunch of boxes in the image and checks if any of these boxes contain any

object. RCNN uses selective search to extract these boxes from an image these boxes

are called regions. Let’s first understand what selective search is and how it identifies

the different regions. There are basically four regions that form an object: varying

scales, colors, textures, and enclosure. Selective search identifies these patterns in the

image and based on that, proposes various regions. Here is a brief overview of how

selective search works:

1. It first takes an image as input:

45
2. Then, it generates initial sub-segmentations so that we have multiple regions from this image:

3. The technique then combines the similar regions to form a larger region (based on color

similarity, texture similarity, size similarity, and shape compatibility):

• Finally, these regions then produce the final object locations (Region of Interest).
Below is a succinct summary of the steps followed in RCNN to detect objects:

1. We first take a pre-trained convolutional neural network.

2. Then, this model is retrained. We train the last layer of the network based on the

number of classes that need to be detected.

3. The third step is to get the Region of Interest for each image. We then reshape all
these

Regions so that they can match the CNN input size.

4. After getting the regions, we train SVM to classify objects and background. For
each

Class, we train one binary SVM.

5. Finally, we train a linear regression model to generate tighter bounding boxes for
each

Identified object in the image.

3.8 Implementation Tool

46
 For image processing and analyzing the dataset and for identification of coffee plant

leaf disease using machine learning approach by implementing the proposed Fast RCNN

method We will test on MATLAB 2018r.

N Material and Tools Capacity

o
1 Computer RAM size 8GB,

Processor minimum of core i5 with

2.2GHZ

HDD 1TB

Graphics G-force/NVidia

2. Software MATLAB 2018r a

Table 3.1 Tools and Materials

3.9 Evaluation

47
 To evaluate the performance of algorithms for identification coffee leaf

diseases WE have used different algorithms and different approaches. In this phase

have tested each algorithm and approaches by passing same training dataset and test

dataset through all approaches and evaluate the result by accuracy and time elapsed

for processing images.

CHAPTER FOUR:
RESULTS

48
 4.1 Introduction

In this chapter, we describe the experiments we have conducted along with

the results obtained. When we discussed in chapter one, the main objective of this

study is to experiments on the possibility of developed detection and classification of

coffee lead diseases. In this paper on implementation phase, we will perform image

classification using Convolutional Neural Network (CNN) in MATLAB R2018a.

Convolutional Neural Network (CNN) is a powerful machine learning technique from

the deep learning domain. A collection of diverge image is required to train CNNs.

The larger the collection the richer the features that CNN learns. Training a CNN

with large collection of diverse images is not an easy task. However, there is an easy

way. We can use pretrained Fast RCNN to leverage the power of CNN. It saves a

huge amount of time and effort when we use pertained CNN as feature extractor. In

this phase, we used „ResNet-50‟ as pertained CNN and Medleys image dataset.

Image classification using convolutional neural network is a very exciting topic.

Though you can apply this concept on different cases here in this research we have

used to identify the diseased or infected coffee leaf hoping to find a better result than

other methods that have been used to implement by other researchers. And also, we

will be taking some of the results of some researchers during their research to

compare with our outcome of this research. Object recognition is another excellent

field where you can use the method shown in this lesson. Image classification using

Fast RCNN in MATLAB r2018a is not a straightforward approach. However, we

have used a good strategic method to use this approach to achieve the goal of this

research. We have also used three algorithms KNN, K-Means, SVM with same

49
dataset and compare with proposed system Fast RCNN to show up Fast RCNN is

better identification approaches

4.2 Experiment Steps

4.2.1 Dataset Preparation: diseased coffee leaves should be collected and

prepared. My first source of dataset is south shoa zone, a publicly available image

database, contains large number of images of diseased and healthy plant leaves of

different species collected under controlled conditions and the ground truths are also

provided.

4.2.2 Dataset Preprocessing

The dataset used for testing and training of all our models is collected from the south

west shoa zone, which comprises 11340 images of infected or dead coffee leaves

directly from coffee plantations, and not in controlled lab environments. These

images were given as input, and enhancement of the input image is done for the

quality of the input image. These Images processing techniques used includes

background avoidance, setting up contrast, noise avoidance, and edge detection,

texture-based clustering and filtering. Experimentation was done to implement these

techniques Artificial Neural Networks, which come up with better output as

compared to other feature engineering algorithms. Image extraction was estimating

the leaf size, Features considered for, yellowing and wilt infected leaves, foliage

covered surroundings and spots on leaves.

K-Means, KNN and SVM results analysis.

50
 The results of the implemented models are presented below. As mentioned in the

methodology, the selected models were: KMeans, Support vector machine, KNN,

ResNet50.In principle, Orange Data Mining (Janez et al., 2013) was used to observe which

models would be the most viable to implement, it was observed that the neural network,

SVM and KNN were the best candidates since they presented higher metrics than the others.

Therefore, the VGG16 network was used as a feature extractor, since in other investigations

it gave positive results. After extracting the characteristics, we proceeded to increase the data

with the window technique and they were implemented in the Google Net platform, since it

provides us with powerful resources to train the models. The results obtained were the

following:

Model Accuracy Recall Precision F1 AUC

K-means

0.601 0.601 0.681 0.603 0.812

SVM RBN 0.524 0.524 0.65 0.554 0.92
Linear SVM 0.472 0.472 0.774 0.5 0.841
KNN 0.59 0.512 0.709 0.608 0.821

results The results obtained were not were as expected, the metrics showed that these models

did not meet expectations and could not predict the images correctly. Their predictions

indicated high dispersion of the data, and the metric recall in each model was the lowest of

all, which indicates that the classifiers were not able to discriminate between positive and

negative cases correctly

4.3 Fast RCNNs RESULTS ANALYSIS

51
 For these models it was decided to implement a network per classifier, in other words, one

model for infected coffee leave. Thus, you would have more control over each model and see

its possible shortcomings. As mentioned above the selected models were ResNet50. These

models needed the input images resized, so for ResNet50, V2 the images were resized to 224

* 224.As mentioned in the methodology, the last layer was removed from these networks and

two more layers were added. The other output layer with two neurons and activation SoftMax

function. After their training, the results were not adequate, although they presented excellent

metrics, the validation was not as expected. The predictions with images were very similar to

the previous models. These were affected by noise and there was no improvement in contrast

to the previous models.

As in the R-CNN detector, the Fast R-CNN detector also uses an algorithm like Edge Boxes

to generate region proposals. Unlike the R-CNN detector, which crops and resizes region

proposals, the Fast R-CNN detector processes the entire image. Whereas an RCNN detector

must classify each region, Fast R-CNN pools CNN features corresponding to each region

proposal. Fast R-CNN is more efficient than R-CNN, because in the Fast R-CNN detector,

the computations for overlapping regions are shared. Use the

trainFastRCNNObjectDetector() function to train a Fast R-CNN object detector. The

function returns a

fastRCNNObjectDetector () that detects objects from an image.

1. As with the earlier techniques, we take an image as an input.

img = imread('Test.jpg');

2. This image is passed to a ConvNet which in turns generates the Regions of Interest.

options = trainingOptions('GoogleNet', ...

52
 'MiniBatchSize', 10, ...

'InitialLearnRate', 1e-3, ...

'MaxEpochs', 10, ...

'CheckpointPath', tempdir);

3. A RoI pooling layer is applied on all of these regions to reshape them as per the input of

the ConvNet. Then, each region is passed on to a fully connected network.

rng (0); shuffledIdx =

randperm(height(stopSigns)); stopSigns = stopSigns (shuffledIdx,:);

4. A SoftMax layer is used on top of the fully connected network to output classes. Along with

the SoftMax layer, a linear regression layer is also used parallelly to output bounding box

coordinates

for predicted classes.

ds = transform(ds,@(data)preprocessData (data,[920 968 3])); frcnn =

trainFastRCNNObjectDetector(ds, fastRCNNLayers, options, ... 'NegativeOverlapRange', [0

0.1], ... 'PositiveOverlapRange', [0.7 1]);

The investigation yielded important results to determine the presence of the disease in coffee leaf and

identifies the type of disease. Where, it was possible to identify and delimit the areas of the plague,
53
allowing better visualization of the locations of the infected areas for farmers. To predict the disease,

the Fast RCNN network was chosen for its good metrics and accurate predictions, in addition, using

the program labelimg, the use of this network for artificial vision problems in agriculture is facilitated

since it facilitates the location of the pest for humans.

54
 CHAPTER FIVE
CONCLUSION AND FUTURE WORK

5.1 Conclusion

This paper presents the survey on different diseases classification techniques used for

Coffee Arabica leaf disease detection and an algorithm for image segmentation technique that

can be used for automatic detection as well as classification of plant leaf diseases later.

Therefore, these diseases with very less computational efforts the optimum results were

obtained, which also shows the efficiency of proposed algorithm in recognition and

classification of the leaf diseases. Another advantage of using this method is that the plant

diseases can be identified at early stage or the initial stage.

All the experiments are performed in MATLAB. For input data disease, samples of Coffee

wilt disease. The co-occurrence features are calculated after mapping the R, G, B components

of the input image to the thresholded images. The co-occurrence features for the leaves are

extracted and compared with the corresponding feature values that are stored in the feature

library. The classification is first done using the Minimum Distance Criterion with K-Mean

Clustering and shows its efficiency with accuracy of 80%. The detection accuracy is improved

to 93.63% by proposed algorithm. In the second phase classification is done using SVM

classifier and shows its efficiency with accuracy of 92%. Now the detection accuracy is

improved to 92% by SVM with proposed algorithm. The training and the testing sets for each

type of leaf along with their detection accuracy is shown in Table 4.1 use other methods from

reference articles experiment result and search result Fast RCNN from Google. From the

results it can be seen that the detection accuracy is enhanced by SVM with proposed

55
algorithm compared to other approaches reported. Now the efficiency of detection of infected

leaf improved to 97% by implementing Fast RCNN.

As results are showing in this research, we have got a good approach to detect coffee disease

which is very help full to diagnose disease easily.so as we have discussed earlier as studies are

showing Ethiopia has encountered huge amount of annual yield of coffee so by implementing

this approach it’s possible to reduce loss and increase production. We hope this research could

help some to increase the productivity the country wants.

56
 5.2 Future work

For future removing the drawbacks of Fast RCNN better to implement Faster

RCNN method and its very nice to make the implementation training and testing to

undertake on fast servers with high specifications like a computer with a good

graphics memory for better and accurate result with a minimum time.

57
 REFERENCES

al, T. e. (2016). An approach based on CNN to recognize apple pathological images. High

Technol, 64-74.

CalvinHung*, Z. (2014). Feature Learning Based Approach for Weed Classification Using

High Resolution Aerial Images from a Digital Camera Mounted on a UAV. ISSN, 1-18.

Debasu, A. (2018). An Automatic Coffee Plant Diseases Identification Using Hybrid Approaches of

Image Processing and Decision Tree . Indonesian Journal of Electrical Engineering and

Computer Scienc, 806-811.

Deep Neural Network Based Recognition of Plant Diseases by Leaf Image Classification.
(2016).

Computational Intelligence and Neuroscience, 1-8.

Erhan, D. (June 2015). Going deeper with convolutions. (pp. 1-13). Dimitru Erhan.

Faridah*1, G. O. (2011). Coffee Bean Grade Determination Based on Image Parameter.

TELKOMNIKA, 547-554.

Feng Qin1, D. L. (2016). Identification of Alfalfa Leaf Diseases Using Image Recognition

Technology. PLOS, 1-26.

GuilingSun, X. ,. (2018). Plant Diseases Recognition Based on Image Processing Technology.

Journal of Electrical and Computer Engineering , 1-8.

Kemin Du, F. x. (2018). A recognition method for cucumber diseases using leaf symptom
images based on deep convolutional neural network. ResearchGate, 18-24.

58
Lucas ximenes Boa Sorte, C. T. (2019). Coffee Leaf Disease Recognition Based on Deep

Learning and Texture Attributes. 23rd International Conference on Knowledge-Based

and Intelligent Information & Engineering Systems (pp. 135-144). Elsevier B.V.

M Dian Bah, A. H. (n.d.). „Deep Learning with Unsupervised Data Labeling for Weed Detection in

Line Crops in UAV Images‟.

Monzurul Islam, A. D. (2017). Detection of potato diseases using image segmentation and multiclass

support vector machine. 2017 IEEE 30th Canadian Conference on Electrical and Computer

Engineering (CCECE) (pp. 30-36). Monzurul Islam.

Sanat Bhandarkar, R. P. (July,2019). Deep Learning and statstical models for detection of white stem.
(pp.

1-9). Sanat Bhandarkar.

Sharada Prasanna Mohanty1, 2. D. (2016). Using Deep Learning for Image-Based Plant

Disease Detection . EPFL, 1-7.

Vijai Singh, A. M. (2017). „Detection of plant leaf diseases using image segmentation and soft
computing

techniques‟. elsevier, 1-9.

Zhicheng Yan†, H. Z. (2015). HD-CNN: Hierarchical Deep Convolutional Neural Networks for

Large Scale Visual Recognition. ISSV (pp. 1-10). Hong Kong: Yizhou Yu
.
(n.d.). Retrieved from towardai.net:

https://www.google.com/url?sa=We&url=https%3A%2F%2Ftowardsai.net%2Fp%2F deep-

learning%2Fplant-disease-detection-

usingfasterrcnn&psig=AOvVaw13ss9pFEL0Vr9tqEyjAyQv&ust=1613678973958000&so

urc

e=images&cd=vfe&ved=2ahUKEwj4sLvh3PHuAhVHw4UKHTINCI4Qr4kDegU

59
 IARCsA

Degaga, J. (2020). Review on Coffee Production and Marketing in Ethiopia.

Researchgate.

Han, X. (2017). Pre-Trained AlexNet Architecture with Pyramid. MDPI.

Hossain, E. (2019). A Color and Texture Based Approach for the Detection and

Classification of Plant Leaf Disease Using KNN Classifier. 2019 International Conference

on Electrical, Computer and Communication Engineering (ECCE), 79

February, 2019. Eftekhar Hossain.

Isotani, S. (2009). RECOGNITION OF FACIAL EXPRESSION BY IMAGE PROCESSING

AND. Anais do IX Congresso Brasileiro de Redes Neurais

/Inteligência Computacional (IX CBR. Sadao Isotani.

Jambo, A. M. (2015). Importance and Characterization of Coffee Berry Disease

(Colletotrichum. Journal of, 6:9.

Maity, S. (n.d.). Fault Area Detection in Leaf Diseases using k-means.

PEDRO TALHINHAS 1, 2. ,. (2016). The coffee leaf rust pathogen Hemileia vastatrix:

one and a half. ResearchGate.

Rupali Patil1, S. U. (Apr-2016). Grape Leaf Disease Detection Using K-means Clustering

Algorithm. International Research Journal of Engineering and Technology

(IRJET), 1-4.

Sultana, F. (n.d.). Advancements in Image Classification using. IEEE.

Talaviya, T. (2020). Implementation of artificial intelligence in agriculture for

optimisation of irrigation and application of pesticides and herbicides. Science

direct, 58-79.

Teferi, K. B. (2015 ). Climatic Variables and Impact of Coffee Berry Diseases

(Colletotrichum Kahawae) in Ethiopian Coffee Productio. Journal of Biology, Agriculture and

Healthcare.

60
Tylor, F. O. (2020). Coffee production constraints and opportunities at major growing districts of

southern Ethiopia.

APPENDIX

61
62
63