Crop Disease Detection in Uncontrolled Lighting Conditions

Chapter 1: Introduction
1.1 Overview and Background
Food shortage is expanding at an exponential rate due to inadequate agricultural output. To
address this, there is need for farmers to increase their productivity and one of these ways is to
reduce the yield lost due to crop diseases. The agriculture business has greatly improved thanks to
technological advancements. Using tЇme, cost, and accuracy of crop qualЇty evaluatЇon by eye
inspection is a tough task [cite]. To overcome this problem, researchers created several
approaches for developing new technologies such as object recognition and image processing for
quality assessment . Image processing technology is used in this study to identify and classify plant
illnesses [1-3]. This image processing method requires the use of high-resolution images to detect
and classify diseases that were previously hard to capture [cite]. As a result, accurately and
efficiently anticipating illness is a challenging task. By analyzing [4] and taking the required steps
to protect against such leaf illnesses, our research aims to build a model with an improved
accuracy that correctly diagnoses and distinguishes leaf illness at the onset of disease. To improve
the accuracy of the model this study examines a variety of parameters that are mainly used
together with RF models.

1.2 Motivation
According to [1] researchers use images captured in a controlled environment, on a single
background. Previous studies have collected images using indoor boxes avoiding the influence of
external lighting [2][2]This helps the accuracy of the models they develop however it greatly
reduced the practicability of the model in a natural environment. This leaves the models wanting
in their practicability. In their paper [3] noted that the greatest repository of plant leaf dataset
Kaggle has images which are taken on a single background resulting in a dataset lacking the
complexity of the real natural environment.

2
1.3 Problem Statement
The problem to be addressed through this study is the is the improvement of the practicability of
the crop disease detection models being developed. The study will attempt to improve the
practicability of the models by introducing complexity to the dataset that already exist helping to
mimic the natural environment, an improvement in the practicability of the models in crop
disease detection will greatly improve how these algorithms gets used.

Comment:

The problem statement is not coming out clearly. I know what you want to say, but you will not
get a chance to explain to the marker. I propose that you change your motivation and make the
following statement your problem statement.

The use of Machine learning in the identification of plant diseases for some time. In order to
identify plant disease researchers use images captured in a controlled environment, on a single
background [cite]. Previous studies have collected images using indoor boxes avoiding the
influence of external lighting [2]. Such models have high accuracy in traing and testing but have
proved to be poor in actual production when used in the natural environment since the lighting in
such scenarios is not controlled. This study seeks to improve the practicability of the models by
introducing lighting complexity to pre-existing plant image datasets which will help to mimic the
natural environment and resultantly improve on the usability of the models.

1.4 Research Aims and Objectives

1.4.1 Aims
This paper aims to develop a model with improved practicability in classifying a healthy from a
non-healthy leaf.

1.4.2 Objectives

 To improve the practicability of classifier algorithms in disease detection in cops.

 To maintain the accuracy of crop disease detection algorithms

1.4.3 Research Questions

 Can the practicability be improved?
 Is it feasible to use different background on the dataset?
 How will the use of complex lighting and background affect the accuracy?
 How will the training and detection time be affected?

3
1.5 Impact of the Research
The increase in the practicability of crop disease detection model will result in a model that will
work more effectively in a natural environment since the majority of crops are not grown in
indoors boxes.

Insert Page breaks between chapters

Chapter 2: Literature Review

2.1 Introduction
Crop disease can manifest itself in several places, including but not limited to the leaf, and stem.
We will just focus on the disease that the leaves in this part have detected. [4] successfully used
the Random Forest(RF) algorithm to which had and accuracy score of 77%. Phyllanthus Elegant
Wall disease implant neural network (ANN) implantation was divided into two types by Ishaket et
al. healthy and unhealthy. They altered the color scheme of the vegetable photos using photo
editing tools. The photographs are categorized using the leaf's color and positioning. Dandawate
et al. employed SVM to detect pathogens in soybean leaves. This tool has automatically diagnosed
plant ailments based on the nature of the mutations. This benefits the farmer by requiring less
online effort. Patil et al. extracted characteristics of the tomato leaf. The leaf image is made up of
red, green, and blue components. These characteristics are considered for categorizing diseases.

2.2 Limitations of ML in crop disease detection

Never have 2 headings follow each other like that. Always put some text in between

2.2.1 Controlled environment

Researcher [3] noted the impact of a controlled environment to the overall practical nature of the
models produced from such datasets. The datasets generated under such conditions will only
serve to produce highly accurate algorithms to that dataset. Faced with a dataset from the natural
environment these models performance drops significantly. Most researchers briefly discuss the
limitations of the experiments done under controlled environments but they are not considered
at all in their experiments [5]

4
2.2.2 Dataset
In a review of imaging techniques for plant disease detection by [6] they noted the limited
availability of public images to train a CNN model resulting in researchers using transfer learning
instead. The lack of variety leads to the same dataset being used which lacks complexity and the
reduction of how practical these trained models are in a natural environment.

2.2.3 Related work

Diverse research methods have been carried out to correctly detect and classify plant disease of
many plants. These comprise of the conventional on field inspection by experts, the use of image
processing techniques and recently the use of machine and deep learning algorithms. However
the advantages of these techniques are not fully utilised due to high cost in case of some methods
and lack of robustness in others. Transfer learning and deep learning techniques stack up well
against these odds and the current project is implemented using these methods to accurately
detect and classify maize disease. Before diving in detail into the specifics of the project, let
understand some of the methodologies used previously and understand their benefits and
shortcomings.

2.2.4 Spectroscopy and its evolution in agriculture domain

Initial researches in the fЇeld made use of spectroscopy to ЇdentЇfy and classЇfy dЇseases. Spectral
frequencЇes of plants were analyzed as Їnfected plants exhЇbЇted dЇfferent characteristics in
comparison to healthy plants due to difference in amount of light absorbed in the near-
infrared(NIR) range which helped in detecting anomalies.

This method was used by [7] in his research to detect powdery mildew in wheat plants.
Correlation, regression and independent sample t tests were used to analyze the 32 spectral
features extracted. As the whole analysis was based on spectral data, it was not effective in early
detection of diseases that affected visual characteristics of plant before altering its internal
structure.

New technologies named multispectral and hyper spectral imaging aimed at reducing the
shortcomings of the previous method by combining spectroscopy and imaging evolved in due
course. In research on strawberry leaves [8] hyperspectral data was used and algorithms like k-
Nearest neЇghbor(kNN) and FЇsher dЇscrЇminant analysis(FDA) were utilized to detect if the plants
were infected by anthracnose crown. Even though they were able to achieve an accuracy of 70

5
percent, the research had several drawbacks which included the tedious process of data collection
using spectroradiometer that necessitated the need for experts that operate it. Further, the data
collected can also be flawed due to shadows cast by sunlight leading to reduced performance as
stated by the authors.

In another research on wheat by Zheng et al. (2019), hyperspectral data was used for analysis as it
was found that spectral data and indices like PRЇ (PhotochemЇcal Reflectance Їndex) and ARЇ
(Anthocyanin Reflectance index) changed with plant development. The results obtaЇned from
theЇr study were better in comparison to previous researches in the field.

Various parameters were tuned and tested in the quest for improving accuracy by other
researchers but the inherent problem of data collection and expert knowledge to understand the
results remained unaddressed , stressing the need for alternative technologies.

2.2.5 Use of machine learning in disease identification and classification

Machine learning is a technical advancement that can be used to tackle various problems related
to predictions or regressions. This has paved way for its use in different domains and has
produced good results. This section shall discuss about its use in disease classification and identify
its benefits and limitations.

Some of the algorithms in machine learning that were popular for classification were KNN and
SVM and have been used widely for disease classification in plants. As previous researches in the
field had placed too much stress on data collection, researchers Dhaware and Wanjale (2017)
have tried to mitigate this by making use of images obtained from handheld devices for analysis
and incorporating image processing in their methodology which involved segmentation. SVM was
used for classification. However, results obtained indicated scope for improvement with usage of
techniques like feature extraction.

An accuracy of 90 percent was obtained by Francis et al. (2016) when composed segmentation
was used to identify diseases like quick wilt and berry spot in pepper plants. It involved green pixel
masking and similarity-based segmentation which was followed by feature extraction.
Propagation neural network was used for classification. As the analysis was carried out on small
dataset, the results are not totally reliable.

Researchers, Padol and Yadav (2016) made use of k- means algorithm for segmentation. Noises
were removed from images using Gaussian filter and around 54 features were extracted from it

6
for analysis. This was however time consuming and tedious, and yielded results close to 90
percent making us wonder if such complex procedures had any significant impact.

Global-singular value decomposition (SVD) was used by researchers Zhang and Wang (2016) in
their quest to identify cucumber diseases. It proved to be effective in feature extraction. SVM,
coupled with an improved recognition method based on watershed algorithm helped them
achieve good and accurate results. The whole analysis was conducted on a small dataset due to
computational complexity of the techniques involved.

Feature extraction was done using SURF by Aravind et al. (2018) and then k-means was used for
clustering them. Grayscale image and its occurrence matrix with histogram served as classifier’s
input in the study. A variation of this study by Maniyath et al. (2018) involved the use of
Histogram of an Oriented Gradient (HOG) as feature extractor. This made use of Hu moments,
texture and color histogram for analysis. Both studies produced good results and supported the
use of random forest when the amount of data available for analysis is small.

In general, one can comprehend that all the above researches made use of relatively small
datasets for analysis and are heavily dependent on pre-processing for accurate classification. This
is computationally expensive and requires high domain knowledge [cite]. These seem to be the
main reasons why we are observing a paradigm shift in recent years as more and more work in
this field are making use of deep learning techniques instead of traditional machine learning
techniques. These are capable of addressing the bottlenecks and provide better results.

2.2.6 Importance of deep learning and transfer learning techniques

Deep learning is a new breed of machine learning algorithm that do not need much pre-
processing for analysis. They are capable of extracting features on their own making the process
of disease identification simpler.

Studies using transfer learning include the research conducted by Shrivastava et al. (2019) on rice
plant. Here CNN architectures were used as feature extractors and output of these were used by
SVM for classification. Models produced accuracies of 92 which is commendable. However, the
dataset used was of size 619 which is too small for analysis. The limitations of these were
addressed making use of data augmentation by authors Coulibaly et al. (2019) which improved
the results to a certain extent . Using CNN as feature extractors can however be computationally
intensive. Further, the training time of models are not discussed by authors which could have
helped us understand their utilities better.
7
In research on rice plants by Lu et al. (2017), CNN with different number of layers were utilised to
identify and classify rice diseases. Although, the techniques achieved good accuracy, the tests
were performed on a small dataset of 500 images. Hence, the results obtained must taken with a
pinch of salt and further analysis must be done on the same.

An extension of the research was done by Ferentinos (2018) by making use of dataset of different
species of plants. Architectures like Alexnet, VGG, Googlenet which are based on CNN were
trained from scratch and resultant models obtained accuracies of about 99 .. In similar analysis by
Mohanty et al. (2016), transfer learning versions of Alexnet and Googlenet were used for
classification of various plant diseases resulting in accuracies similar to previous study. However,
both these models failed miserably when tested on datasets other than the ones under study. The
authors suggest that these can be mitigated using data augmentation and training some layers of
the architectures with problem specific data.

2.2.7 Boosted Trees for image classification

A relatively new breed of algorithms that makes use of boosted trees, their potential is yet to be
fully explored [cite]. In initial research conducted using these techniques by Gao et al. (2017) for
object classification, these techniques performed good and produced accuracies close to 92
percent. Xu and Wang (2019) in their risk analysis of Type 2 diabetes made use of random forest
and XGboost for analysis. The results obtained were highly promising and were close to those
obtained by Gao et al. (2017) supporting their use in current research for classifying and
identifying plant disease.

The current research aims to address the shortcomings of previous works in this area by making
use of data augmentation and transfer learning using trainable layers. The performance of new
algorithms like Gradient and XGboost shall be compared against these to find an optimal model
that is useful in identifying and classifying maize disease. The methodology and implementation of
these techniques is elaborated in further sections.

2.2.8 Solving shortcomings of ML in crop disease detection

Several studies have been done to try to solve the shortcomings of ML in crop diseases detection,
from attempting to solve the lighting weakness. Marko Arsenovic et.al 2019 attempted to solve
the complexity problems by gathering images from different sources.

8
2.3 Conclusion

Chapter 3: Methodology
3.1 Introduction

3.2 Research Method

In this study the researcher used quasi-experiment method which seeks to establish a relationship
between the independent and dependant variables that is the cause and effect relationship. Best
and Khan in 1993 suggested the pretest-posttest non-equivalent groups design which will be used
in this experiment. Here results are taken before and after the experiment to see the effects of
the experiment on a group. The design is as follows:

Experimental Group S1 X S2

Control Group S3 C S4

Where :

S1 : Pretest of experimental group

S2 : Posttest of experimental group

S3 : Pretest of control group

S4 : Posttest of control group

X : Treatment of experimental group by using non complex datasets

9
C : Treatment of control group by using complex dataset

3.3 Research Variables

The study variables according to Best and Khan 1993, these are characteristics or the conditions
that the researcher will manipulate controls and observe. These are classified into dependent
variables and independent variables.

3.3.1 Independent variables

This variable is a characteristic that has the ability to influence or affect an outcome of a
dependent variable according to Creswell 2012. In this study, the complexity of the dataset.

3.3.2 Dependent variables

This variable is a characteristic that is dependent on or is influenced by independent variable
according to Creswell 2012. In this study we will use the training time, accuracy and consider the
practicability as our dependent variable.

3.4 Research Hypothesis

The introduction of complex background and light to the dataset will improve practicability of
crop disease detection.

3.5 Techniques for Collecting the data

Tests and research Instrument will be used to collect data.

3.5.1 Tests
A test is a method of measuring performance in a given setting Brown 2004. The researcher will
use two kinds of test, the first one is applied as a pretest and the posttest. The test was training a
model without complex dataset and the data was captured. The posttest was done after
considering the complexity of the dataset introduced.

3.5.2 Research Instrument Analysis

The instrument is designed for analysing the effect of the complex and non-complex dataset
pretest and posttest activities. The same environment used for the non-complex dataset pretest is
the same for the complex dataset posttest activities.

3.5.2.1 Validity Test

Defined as the appropriateness, usefulness and meaningfulness of the specific inferences
researchers make based on the data they collect according to Fraenkle, et. Al 2012. This research
10
is going to concentrate on the construct validity defined as the nature of the characteristic being
measured by the instrument in this case the training time, accuracy as well as the practicability.

3.6 Conclusion

 I am not going to approve Research Methodology Chapter with scant information on the
research methodology. Please redo this chapter and provide all the pertinent headings that
include the data source, the correct methodology, population (size of dataset) and steps
taken in cleaning the data (EDA).
 For methodology since this is a purely Data Science problem use CRISP-DM, TDSP, KDD or
any other which was created for solving data science problem

Chapter 4: Results and Discussion of the Results

4.1 Introduction
In order to improve the practicality of the crop disease detection this study will introduce the image
complexity addition in the image pre-processing stage illustrated by the following figure.

11
Figure 4.1.1: Pr0p0sed m0del

4.1 Tools For Experimental

Using Google colab as the environment of choice and python as the pr 0gramming language, this thesis
programmed ML algorithms namely RF, SVM and KNN in order to compare the models made from non
complex dataset and the complex dataset. The app imp0rted data from Kaggle plantVillage.

4.2 Dataset complexity improvement

Never have two headings back to back without any text between them

4.2.1 Practicability measurement

In order to measure practicability the thesis has chosen to assign a 1 for an image that has non-
gray background found on the grape fruit leaves on the plantVillage dataset from Kaggle. A 0.1 for
an image which has 10nits lighting and 1 for an image with 1nit of lighting.
12
4.2.2 The background
The background of half of the grape plant images was altered from the greyish to green, blue, red,
pink, yellow, brown, orange and half of the dataset was left the same. The results are presented
on the following table.

Background Accuracy complexity

RF KNN SVM
Ordinary 0.98125 0.934375 0.934375 0
Green 0.78251 0.754523 0.754523 1
Blue 0.88235 0.875231 0.875231 1
Pink 0.89523 0.878523 0.878523 1
Orange 0.98563 0.958527 0.958527 1
red 0.95865 0.935287 0.935287 1
yellow 0.91258 0.888527 0.888527 1
brown 0.87257 0.85258 0.85258 1
Black 0.965625 0.93125 0.93125 1

4.2.3 The lighting

The brightness of half of the images was also altered from 1cd/mm2 (candela per square meter)
to 10cd/mm2 during the experiment.

lighti ng eff ect to accuracy

1

0.98

0.96

0.94

0.92

0.9

0.88

0.86
1unit 3units 5units 7units 10units

RF KNN SVM

13
 Accuracy Complexity
RF KNN SVM
ordinary 0.98125 0.934375 0. 934375 0
3units 0.9625 0.9343 0.925 1
5units 0.978125 0.934375 0.925 1
7units 0.975 0.940625 0.9125 1
10units 0.96875 0.94375 0.940625 1

Chart Title
1.2

0.8

0.6

0.4

0.2

0
1 3 5 7 10

RF KNN SVM

4.2.4 Lighting and background

We cross referenced the lighting and the background and the results are as follows

Accuracy Complexity
RF KNN SVM
ordinary ordinary 0.98125 0.934375 0. 934375 1
3units 0.9625 0.9343 0.925 1
5units 0.978125 0.934375 0.925 1
7units 0.975 0.940625 0.9125 1
14
 10units 0.96875 0.94375 0.940625 1
black ordinary 0.97125 0.9421 0. 92437 1
3units 0.9525 0.95465 0.925 2
5units 0.928125 0.956375 0.925 2
7units 0.915 0.970625 0.9325 2
10units 0.905 0.98375 0.94036 2
red ordinary 0.961 0.92325 0. 932 1
3units 0.9725 0.9243 0.915 2
5units 0.978125 0.93652 0.912 2
7units 0.945 0.94252 0.9125 2
10units 0.9875 0.9362 0.940625 2
yellow ordinary 0.973 0.934 0. 934375 1
3units 0.9625 0.932 0.925 2
5units 0.972 0.934 0.925 2
7units 0.975 0.95 0.9125 2
10units 0.96875 0.955 0.91625 2

4.3 Performance Evaluation

With the introduction of varying backgrounds and varying light we have seen the fluctuations in
performance meaning the accuracy is affected by the light, RF is negatively affected by the increase in light
as compared to KNN which seemingly increase in accuracy for the increase in lighting.

4.4 Conclusion
Based on the experiments done in an attempt to increase the complexity of the dataset inorder to
improve the practicality of the trained models, we have noticed that accuracy is greatly affected
by the increase of the complexity which was expected some algorithms are more sensitive than
others and some seems to gain accuracy instead of losing such as KNN.

15
 Chapter 5: Conclusion and Future Work

5.1 Conclusion
Based on the results presented by in the previous chapter and the work that was done throughout
this study, we have seen that accuracy is greatly affected by the complexity of the dataset.

5.1.1 Effects of changing the background

The background has an effect on the accuracy as well with different backgrounds showing
different effects on the performance the used ML algorithms.

5.1.2 Effects of different lighting conditions

The lighting had a great impact on the accuracy as compared to the background with signs of
performance improvement to other algorithms such as KNN which was gaining accuracy as
compared to the RF which was losing accuracy the brighter the image.

5.1.3 Recommendations
References

[1] J. Liu and X. Wang, “Plant diseases and pests detection based on deep learning: a review,” Plant
Methods, vol. 17, no. 1. BioMed Central Ltd, Dec. 01, 2021. doi: 10.1186/s13007-021-00722-9.

[2] F. Martinelli et al., “Advanced methods of plant disease detection. A review,” Agronomy for
Sustainable Development 2014 35:1, vol. 35, no. 1, pp. 1–25, Sep. 2014, doi: 10.1007/S13593-014-
0246-1.

[3] M. Arsenovic, M. Karanovic, S. Sladojevic, A. Anderla, and D. Stefanovic, “Solving current

limitations of deep learning based approaches for plant disease detection,” Symmetry (Basel), vol.
11, no. 7, p. 21, Jul. 2019, doi: 10.3390/sym11070939.

[4] M. Robnik-ˇ Sikonja, “Improving Random Forests,” Springer, 2004.

[5] J. G. A. Barbedo, “Factors influencing the use of deep learning for plant disease recognition,”
Biosyst Eng, vol. 172, pp. 84–91, Aug. 2018, doi: 10.1016/j.biosystemseng.2018.05.013.

[6] V. Singh, N. Sharma, and S. Singh, “A review of imaging techniques for plant disease detection,”
Artificial Intelligence in Agriculture, vol. 4, pp. 229–242, Jan. 2020, doi:
10.1016/J.AIIA.2020.10.002.

16
[7] M. Z. Y. C. B Zhang, “Crop pest identification based on spatial pyramid pooling and deep
convolution neural network,” Trans Chin Soc Agric Eng, vol. 35, no. 19, pp. 209–215, 2019.

[8] S. P. S. G. S Kaur, “Plants disease identification and classification through leaf images: a survey,”
Arch Comput Methods Eng, vol. 26, no. 4, pp. 1–24, 2018.

[9] A. A. Bernardes et al., “Identification of foliar diseases in cotton crop,” Lecture Notes in
Computational Vision and Biomechanics, vol. 8, pp. 67–85, 2013, doi: 10.1007/978-94-007-0726-
9_4.

[10] V. B. Devi, R. Prabavathi, P. Subha, and M. Meenaloshini, “An Efficient and Robust Random Forest
Algorithm for Crop Disease Detection,” 2022 International Conference on Communication,
Computing and Internet of Things, IC3IoT 2022 - Proceedings, 2022, doi:
10.1109/IC3IOT53935.2022.9767937.

[11] M. Prabhakar, R. Purushothaman, and D. P. Awasthi, “Deep learning based assessment of disease
severity for early blight in tomato crop,” Multimed Tools Appl, vol. 79, no. 39–40, pp. 28773–
28784, Oct. 2020, doi: 10.1007/S11042-020-09461-W.

[12] A. Picon, M. Seitz, A. Alvarez-Gila, P. Mohnke, A. Ortiz-Barredo, and J. Echazarra, “Crop conditional
Convolutional Neural Networks for massive multi-crop plant disease classification over cell phone
acquired images taken on real field conditions,” Comput Electron Agric, vol. 167, Dec. 2019, doi:
10.1016/J.COMPAG.2019.105093.

[13] C. Yang, “Remote Sensing and Precision Agriculture Technologies for Crop Disease Detection and
Management with a Practical Application Example,” Engineering, vol. 6, no. 5, pp. 528–532, May
2020, doi: 10.1016/J.ENG.2019.10.015.

[14] M. Ouhami, A. Hafiane, Y. Es-Saady, M. el Hajji, and R. Canals, “Computer Vision, IoT and Data
Fusion for Crop Disease Detection Using Machine Learning: A Survey and Ongoing Research,”
Remote Sensing 2021, Vol. 13, Page 2486, vol. 13, no. 13, p. 2486, Jun. 2021, doi:
10.3390/RS13132486.

[15] Y. Zhao et al., “An effective automatic system deployed in agricultural Internet of Things using
Multi-Context Fusion Network towards crop disease recognition in the wild,” Applied Soft
Computing Journal, vol. 89, Apr. 2020, doi: 10.1016/J.ASOC.2020.106128.

17
[16] “Remote Sensing | Free Full-Text | Computer Vision, IoT and Data Fusion for Crop Disease
Detection Using Machine Learning: A Survey and Ongoing Research.”
https://www.mdpi.com/2072-4292/13/13/2486 (accessed Nov. 24, 2022).

[17] A. dos Santos Ferreira, D. Matte Freitas, G. Gonçalves da Silva, H. Pistori, and M. Theophilo Folhes,
“Weed detection in soybean crops using ConvNets,” Comput Electron Agric, vol. 143, pp. 314–324,
Dec. 2017, doi: 10.1016/J.COMPAG.2017.10.027.

[18] S. A. Fennimore, D. C. Slaughter, M. C. Siemens, R. G. Leon, and M. N. Saber, “Technology for

Automation of Weed Control in Specialty Crops,” Weed Technology, vol. 30, no. 4, pp. 823–837,
Dec. 2016, doi: 10.1614/WT-D-16-00070.1.

[19] J. M. Duarte-Carvajalino, D. F. Alzate, A. A. Ramirez, J. D. Santa-Sepulveda, A. E. Fajardo-Rojas, and

M. Soto-Suárez, “Evaluating late blight severity in potato crops using unmanned aerial vehicles
and machine learning algorithms,” Remote Sens (Basel), vol. 10, no. 10, Oct. 2018, doi:
10.3390/RS10101513.

[20] M. Dyrmann, R. N. Jørgensen, and H. S. Midtiby, “RoboWeedSupport - Detection of weed

locations in leaf occluded cereal crops using a fully convolutional neural network,” Advances in
Animal Biosciences, vol. 8, no. 2, pp. 842–847, 2017, doi: 10.1017/S2040470017000206.

[21] T. van Klompenburg, A. Kassahun, and C. Catal, “Crop yield prediction using machine learning: A
systematic literature review,” Comput Electron Agric, vol. 177, Oct. 2020, doi:
10.1016/J.COMPAG.2020.105709.

[22] T. Domingues, T. Brandão, and J. C. Ferreira, “Machine Learning for Detection and Prediction of
Crop Diseases and Pests: A Comprehensive Survey,” Agriculture, vol. 12, no. 9, p. 1350, Sep. 2022,
doi: 10.3390/agriculture12091350.

[23] J. Liu, I. Abbas, and R. S. Noor, “Development of Deep Learning-Based Variable Rate Agrochemical
Spraying System for Targeted Weeds Control in Strawberry Crop,” Agronomy 2021, Vol. 11, Page
1480, vol. 11, no. 8, p. 1480, Jul. 2021, doi: 10.3390/AGRONOMY11081480.

[23] J. Liu and X. Wang, “Plant diseases and pests detection based on deep learning: a review,” Plant
Methods, vol. 17, no. 1. BioMed Central Ltd, Dec. 01, 2021. doi: 10.1186/s13007-021-00722-9.

18
[24] F. Martinelli et al., “Advanced methods of plant disease detection. A review,” Agronomy for
Sustainable Development 2014 35:1, vol. 35, no. 1, pp. 1–25, Sep. 2014, doi: 10.1007/S13593-014-
0246-1.

[25] M. Arsenovic, M. Karanovic, S. Sladojevic, A. Anderla, and D. Stefanovic, “Solving current

limitations of deep learning based approaches for plant disease detection,” Symmetry (Basel), vol.
11, no. 7, p. 21, Jul. 2019, doi: 10.3390/sym11070939.

[26] M. Robnik-ˇ Sikonja, “Improving Random Forests,” Springer, 2004.

[27] J. G. A. Barbedo, “Factors influencing the use of deep learning for plant disease recognition,”
Biosyst Eng, vol. 172, pp. 84–91, Aug. 2018, doi: 10.1016/j.biosystemseng.2018.05.013.

[28] V. Singh, N. Sharma, and S. Singh, “A review of imaging techniques for plant disease detection,”
Artificial Intelligence in Agriculture, vol. 4, pp. 229–242, Jan. 2020, doi:
10.1016/J.AIIA.2020.10.002.

[29] M. Z. Y. C. B Zhang, “Crop pest identification based on spatial pyramid pooling and deep
convolution neural network,” Trans Chin Soc Agric Eng, vol. 35, no. 19, pp. 209–215, 2019.

[30] S. P. S. G. S Kaur, “Plants disease identification and classification through leaf images: a survey,”
Arch Comput Methods Eng, vol. 26, no. 4, pp. 1–24, 2018.

[31] A. A. Bernardes et al., “Identification of foliar diseases in cotton crop,” Lecture Notes in
Computational Vision and Biomechanics, vol. 8, pp. 67–85, 2013, doi: 10.1007/978-94-007-0726-
9_4.

[32] V. B. Devi, R. Prabavathi, P. Subha, and M. Meenaloshini, “An Efficient and Robust Random Forest
Algorithm for Crop Disease Detection,” 2022 International Conference on Communication,
Computing and Internet of Things, IC3IoT 2022 - Proceedings, 2022, doi:
10.1109/IC3IOT53935.2022.9767937.

[33] M. Prabhakar, R. Purushothaman, and D. P. Awasthi, “Deep learning based assessment of disease
severity for early blight in tomato crop,” Multimed Tools Appl, vol. 79, no. 39–40, pp. 28773–
28784, Oct. 2020, doi: 10.1007/S11042-020-09461-W.

[34] A. Picon, M. Seitz, A. Alvarez-Gila, P. Mohnke, A. Ortiz-Barredo, and J. Echazarra, “Crop conditional
Convolutional Neural Networks for massive multi-crop plant disease classification over cell phone

19
 acquired images taken on real field conditions,” Comput Electron Agric, vol. 167, Dec. 2019, doi:
10.1016/J.COMPAG.2019.105093.

[35] C. Yang, “Remote Sensing and Precision Agriculture Technologies for Crop Disease Detection and
Management with a Practical Application Example,” Engineering, vol. 6, no. 5, pp. 528–532, May
2020, doi: 10.1016/J.ENG.2019.10.015.

[36] M. Ouhami, A. Hafiane, Y. Es-Saady, M. el Hajji, and R. Canals, “Computer Vision, IoT and Data
Fusion for Crop Disease Detection Using Machine Learning: A Survey and Ongoing Research,”
Remote Sensing 2021, Vol. 13, Page 2486, vol. 13, no. 13, p. 2486, Jun. 2021, doi:
10.3390/RS13132486.

[37] Y. Zhao et al., “An effective automatic system deployed in agricultural Internet of Things using
Multi-Context Fusion Network towards crop disease recognition in the wild,” Applied Soft
Computing Journal, vol. 89, Apr. 2020, doi: 10.1016/J.ASOC.2020.106128.

[38] “Remote Sensing | Free Full-Text | Computer Vision, IoT and Data Fusion for Crop Disease
Detection Using Machine Learning: A Survey and Ongoing Research.”
https://www.mdpi.com/2072-4292/13/13/2486 (accessed Nov. 24, 2022).

[39] A. dos Santos Ferreira, D. Matte Freitas, G. Gonçalves da Silva, H. Pistori, and M. Theophilo Folhes,
“Weed detection in soybean crops using ConvNets,” Comput Electron Agric, vol. 143, pp. 314–324,
Dec. 2017, doi: 10.1016/J.COMPAG.2017.10.027.

[40] S. A. Fennimore, D. C. Slaughter, M. C. Siemens, R. G. Leon, and M. N. Saber, “Technology for

Automation of Weed Control in Specialty Crops,” Weed Technology, vol. 30, no. 4, pp. 823–837,
Dec. 2016, doi: 10.1614/WT-D-16-00070.1.

[41] J. M. Duarte-Carvajalino, D. F. Alzate, A. A. Ramirez, J. D. Santa-Sepulveda, A. E. Fajardo-Rojas, and

M. Soto-Suárez, “Evaluating late blight severity in potato crops using unmanned aerial vehicles
and machine learning algorithms,” Remote Sens (Basel), vol. 10, no. 10, Oct. 2018, doi:
10.3390/RS10101513.

[42] M. Dyrmann, R. N. Jørgensen, and H. S. Midtiby, “RoboWeedSupport - Detection of weed

locations in leaf occluded cereal crops using a fully convolutional neural network,” Advances in
Animal Biosciences, vol. 8, no. 2, pp. 842–847, 2017, doi: 10.1017/S2040470017000206.

20
[43] T. van Klompenburg, A. Kassahun, and C. Catal, “Crop yield prediction using machine learning: A
systematic literature review,” Comput Electron Agric, vol. 177, Oct. 2020, doi:
10.1016/J.COMPAG.2020.105709.

[44] T. Domingues, T. Brandão, and J. C. Ferreira, “Machine Learning for Detection and Prediction of
Crop Diseases and Pests: A Comprehensive Survey,” Agriculture, vol. 12, no. 9, p. 1350, Sep. 2022,
doi: 10.3390/agriculture12091350.

[45] J. Liu, I. Abbas, and R. S. Noor, “Development of Deep Learning-Based Variable Rate Agrochemical
Spraying System for Targeted Weeds Control in Strawberry Crop,” Agronomy 2021, Vol. 11, Page
1480, vol. 11, no. 8, p. 1480, Jul. 2021, doi: 10.3390/AGRONOMY11081480.

21