Image Acquisition, Preprocessing and Classification of Citrus Fruit Diseases: A Systematic Literature Review

Abstract

Different kinds of techniques are evaluated and analyzed for various classification models for the detection of diseases of citrus fruits. This paper aims to systematically review the papers that focus on the prediction, detection, and classification of citrus fruit diseases that have employed machine learning, deep learning, and statistical techniques. Additionally, this paper explores the present state of the art of the concept of image acquisition, digital image processing, feature extraction, and classification approaches, and each one is discussed separately. A total of 78 papers are selected after applying primary selection criteria, inclusion/exclusion criteria, and quality assessment criteria. We observe that the following are widely used in the selected studies: hyperspectral imaging systems for the image acquisition process, thresholding for image processing, support vector machine (SVM) models as machine learning (ML) models, convolutional neural network (CNN) architectures as deep learning models, principal component analysis (PCA) as a statistical model, and classification accuracy as evaluation parameters. Moreover, the color feature is the most popularly used feature for the RGB color space. From the review studies that performed comparative analyses, we find that the best techniques that outperformed other techniques in their respective categories are as follows: SVM among the ML methods, ANN among the neural network networks, CNN among the deep learning methods, and linear discriminant analysis (LDA) among the statistical techniques.This study concludes with meta-analysis, limitations, and future research directions.

1. Introduction

In the field of agribusiness, diseases of fruit products initiate the degradation of the economy, just as large-scale manufacturing affects the economy around the world. Some researchers in the last decade demonstrated the criticality of the quality of fruit products, as it impacts human wellbeing [1]. Fruit products ought to be the basis of a sound eating regimen. Citrus fruits are a significant product in agriculture, and nearly everybody consumes them consistently [2]. Citrus fruits include lemons, oranges, grapes, and tangerines. Various diseases affect citrus fruits, including black spot, greasy spot, canker, and greening, as well as many more. Diseases of citrus fruits are a critical subject that significantly influences the quality and number of yields around the world. The utilization of pesticides by farmers to control various diseases and enhance the production of crops is taking place on a vast scale [3]. Diseases of fruit crops cause significant issues, such as low levels of production and monetary misfortunes, for farmers. Therefore, the detection of diseases and the identification of their severity is a primary need in the agricultural world. Generally, symptoms of disease in citrus fruits are identified with regular monitoring using just the naked eye. This procedure is costly in enormous manors and is less precise. In some countries, farmers hire specialists to identify citrus fruit diseases, and again, this is a costly and tedious task. There is a need for high returns in horticultural enterprises, as well as a better-quality yield of fruit products, if automatic systems are developed to help in the early discovery of infection or diseases in citrus fruit [4]. Many systems have been examined and proposed by analysts in the landscape of artificial intelligence, machine learning, digital image processing, and deep learning for the prediction and classification of citrus infections.

Machine vision platforms are indeed a commercial tool for the evaluation of food standards. All such systems are used to assess production throughout the domain and are used for robotic post-harvest or the early diagnosis of possibly lethal diseases [1]. They are often used in post-harvest processing for the computer-controlled investigation of the fruits’ external quality, including the breakneck speed filtering of them together in commercial sections.

2. Systematic Literature Review

The arrangement, organization, and formation of the systematic literature review (SLR) were implemented by the methodology chosen by the authors [5]. The framework of this SLR was also based on the protocols followed by the authors [6,7,8]. After constructing the review methodology, a sequence of steps was carried out for this study. The procedure is shown in Figure 1.

• The first step is identifying the need for the SLR.
• In the second step, research questions are formulated to answer the issues being addressed in the review.
• The third step designed is a search strategy comprised of two further steps, i.e., primary and secondary steps.
• The fourth step is to find relevant studies from the different resources relied upon to answer the research questions. The inclusion and exclusion criteria for screening the related studies are also included in this phase.

To improve consistency, we then developed the quality evaluation criteria. We measured and evaluated the collected data regarding the review questions in the sixth stage of the SLR. Implementing the review process for any SLR is significant enough to decrease the chances of biased studies. In the following section, we delineate the steps we followed in this SLR.

3. Need for SLR

This SLR aims to provide a complete picture of computer vision systems in diagnosing disease, particularly in citrus fruits. The goal for this SLR is to explore the research articles in a more structured way that adopts the concept of a computer vision system for the identification and classification of diseases present in citrus fruits.

3.1. Research Questions

This SLR aims to explain and analyze the scientific evidence from further research using ML, DL, and methodological techniques to evaluate the classification of citrus fruit diseases. For this purpose, the following research questions were designed and evaluated:

• RQ1: Which kind of diseases affect citrus fruits?
• RQ2: Which techniques have been used to capture citrus fruits’ disease-related patterns?
• RQ3: Which techniques outperform other techniques in terms of classification accuracy?
• RQ4: Which hybrid techniques were used to detect citrus fruit diseases?
• RQ5: What evaluation metrics are commonly seen in studies for assessing techniques?
• RQ6: What evaluation metrics are commonly seen in studies for assessing techniques?

3.2. Search Strategy and Study Selection

There were two steps of our core method for finding and downloading relevant studies:

a.

Primary searching;

b.

Secondary searching.

For primary study selection, relevant keywords were selected by including all possible synonyms and using alternative phrases and Boolean operators, such as “OR” and “AND”. The search string designed was:

Citrus AND (fruit) AND (disease OR diseases OR infection OR infections OR decay OR decays OR defect OR defects) AND (detection OR prediction OR classification) AND (“neural networks” OR “neural network” OR “deep learning” OR “machine learning” OR “statistical technique”).

For the secondary phase, references for the selected paper were also included to obtain more relevant papers for our SLR. The research references that were focused on were the following for the selection and compilation of our primary studies:

• IEEE Xplore;
• Scopus;
• Springer;
• Science Direct;
• Wiley;
• Google Scholar;
• ACM Digital Library.

After selecting the databases to search, the primary step was to screen the relevant studies of interest. The full-text papers identified the relevant studies that fulfilled the inclusion and exclusion criteria. We also included papers using the references of the selected studies. The total number of 1357 papers was identified by searching from the databases; the screening of papers was based upon the search motive using the selection criteria.

3.3. Quality Assessment Criteria

The quality evaluation was treated as an additive step to select the relevant studies for the SLR. We constructed a quality assessment questionnaire to score the selected studies. The studies with lower scores were further excluded from the SLR.

• Q1. Are the goals of the study explicitly stated?
• Q2. Are the techniques of analysis well-established and reasoned?
• Q3. Are the experiments implemented with adequate datasets?
• Q4. Is the experiment extended to dataset(s) processed with images?
• Q5. Are the findings and outcomes explicitly stated?

Based upon these five questions, the papers were evaluated and scored from 0 to 4. However, most researchers applied a binary scoring system (i.e., 0 or 1) for each question. Binary scores, i.e., 0 or 1, are not the best indicators to evaluate studies. However, we somewhat followed the authors’ method to use fuzzy linear variables [9,10,11,12,13,14,15]. However, instead of using a crisp set, we chose a numeric set ranging from 0 to 4 for each question. Mainly, we set the score for each question as follows:

• 0 (No);
• 1 (Rarely);
• 2 (Partly);
• 3 (Mostly);
• 4 (Yes).

Since we used the question scoring method mentioned above, the overall score for each study can fall in the following ranges:

• 0–1.5 (no);
• 1.6–3.5 (moderate);
• 3.6–5.0 (yes).

As a result of the quality assessment criteria, a further 9 papers were discarded, and 78 papers were finalized. Figure 2 represents the distribution of the selected studies from various research libraries.

All studies having “average” and “yes” indicators were included in the SLR. Studies rated as “yes” were considered the highest quality studies according to the quality assessment criteria.

3.4. Data Extraction and Data Synthesis

We next assessed the details of the selected papers that answered the research questions. The primary aim of data synthesis is to collect and collate the selected studies’ information to formulate responses with regard to the research questions. We included the authors’ names, the article’s title, publishing details, dataset details, image acquisition and processing details, feature extraction details, and the technique used. The retrieved data were saved into an Excel file for further analysis and synthesis. The accuracy measures used in different studies were also evaluated to validate the approaches used. The

Table 1. Selected studies addressing our research questions.

PAPER	RQ1	RQ2	RQ3	RQ4	RQ5	PAPER	RQ1	RQ2	RQ3	RQ4	RQ5
[16]	Yes	Yes	No	No	Yes	[17]	Yes	Yes	No	No	Yes
[18]	Yes	Yes	No	No	Yes	[19]	Yes	Yes	No	No	Yes
[20]	Yes	Yes	No	No	Yes	[21]	Yes	Yes	Yes	No	Yes
[22]	Yes	Yes	Yes	No	Yes	[23]	Yes	Yes	Yes	No	Yes
[24]	Yes	Yes	No	No	Yes	[25]	Yes	Yes	No	No	Yes
[26]	Yes	Yes	No	No	Yes	[27]	Yes	Yes	No	No	Yes
[28]	Yes	Yes	No	No	Yes	[29]	Yes	Yes	No	No	Yes
[30]	Yes	Yes	No	No	Yes	[31]	Yes	Yes	No	No	Yes
[32]	Yes	Yes	Yes	No	Yes	[33]	Yes	Yes	Yes	No	Yes
[34]	Yes	Yes	No	No	Yes	[35]	Yes	Yes	No	No	Yes
[36]	Yes	Yes	No	Yes	Yes	[37]	No	Yes	No	No	Yes
[38]	Yes	Yes	No	No	Yes	[39]	Yes	Yes	No	No	No
[40]	Yes	Yes	Yes	No	Yes	[41]	Yes	Yes	Yes	No	Yes
[40]	Yes	Yes	No	No	Yes	[42]	Yes	Yes	Yes	No	Yes
[43]	Yes	Yes	No	No	Yes	[44]	Yes	Yes	No	No	Yes
[45]	Yes	Yes	No	No	Yes	[46]	Yes	Yes	No	Yes	Yes
[47]	Yes	Yes	No	No	Yes	[48]	Yes	Yes	Yes	Yes	Yes
[49]	Yes	Yes	Yes	No	Yes	[50]	Yes	Yes	No	No	Yes
[51]	Yes	Yes	No	No	Yes	[21]	Yes	Yes	No	No	Yes
[52]	Yes	Yes	No	No	Yes	[53]	Yes	Yes	No	No	Yes
[17]	Yes	Yes	No	No	Yes	[54]	Yes	Yes	No	No	Yes
[55]	Yes	Yes	Yes	No	Yes	[56]	Yes	Yes	No	No	Yes
[57]	Yes	Yes	No	No	No	[58]	Yes	Yes	No	No	Yes
[59]	Yes	Yes	No	No	Yes	[60]	Yes	Yes	Yes	No	Yes
[61]	Yes	Yes	No	No	Yes	[62]	Yes	Yes	No	No	Yes
[63]	Yes	Yes	No	No	Yes	[64]	Yes	Yes	No	No	Yes
[65]	Yes	Yes	No	No	Yes	[66]	Yes	Yes	No	Yes	Yes
[67]	Yes	Yes	No	No	No	[68]	Yes	Yes	No	No	Yes
[69]	Yes	Yes	Yes	No	Yes	[59]	Yes	Yes	No	No	Yes
[70]	Yes	Yes	No	No	Yes	[71]	Yes	Yes	No	No	Yes
[72]	Yes	Yes	No	Yes	Yes	[73]	Yes	Yes	No	No	Yes
[74]	Yes	Yes	Yes	No	Yes	[75]	Yes	Yes	No	No	Yes
[76]	Yes	Yes	No	No	Yes	[77]	Yes	Yes	No	No	Yes
[78]	Yes	Yes	Yes	No	Yes	[79]	Yes	Yes	No	No	Yes
[80]	Yes	Yes	No	No	Yes	[81]	No	Yes	Yes	No	Yes
[82]	Yes	Yes	No	No	Yes	[83]	Yes	Yes	No	No	Yes
[84]	Yes	Yes	No	Yes	Yes	[85]	Yes	Yes	No	No	No
[86]	Yes	Yes	Yes	Yes	Yes	[87]	Yes	Yes	No	Yes	Yes
[88]	Yes	Yes	No	No	Yes	[89]	Yes	Yes	No	No	Yes
						[90]	Yes	Yes	No	No	Yes

below summarizes and displays the results.

4. Results and Discussion

In this section, the answers to our research questions are given based on the selected studies.

• RQ1: What kinds of diseases affect citrus fruits?

Citrus fruits, such as oranges, grapefruit, limes, and lemons, are affected by different citrus diseases, including chilling/freezing, anthracnose, pitting/splitting/greening, scab, greasy spot, etc. These are presented in Figure 3. Figure 4 shows many studies that have encountered the particular diseases present in citrus fruits.

The most widespread defects according to count are surface defects (17), P. digitatum and other fungus infections (14), canker (12), melanose (10), HLB (9), and blackspot (7).

Table 2. Description of diseases present in citrus fruits observed by studies.

Type of Diseases	Description	Research Papers
Anthracnose	It is an initial colonizer of damage, with brown-colored lesions on fruits damaged by many factors, such as sunburn, chemical exposure, high temperatures, and extended storage duration.	[76,78,88]
Melanose	This disease creates a slight bruise on the fruit that can rarely affect the ultimate fruit production yield but induces visible imperfections that decrease the profitability of fruits destined for the produce market.	[17,22,26,44,50,59,63,70,78]
Citrus scab	These fungal pathogens of foliage are globally present in various citrus-cultivar-producing regions. Generally, in in-plane areas with regular rain, the occurrence of scab is more extreme than in tropical regions. Scab injuries can be found on fruits just before seven days of contamination. These infections also present as small patches with a coarse and distorted appearance.	[27,59,72,76,78]
Black spot	These are small, round, and dangerous spots with a diameter of 0.12 to 0.4 in fruits.	[26,41,59,65,72,76,78]
Canker	The diameter of the fruit scar is about 1–10 mm and is surrounded by water-dipped and yellow disc-like lesions.	[17,22,27,40,41,50,59,63,68,70,78]
Huanglongbing (HLB)	Another name of HLB is greening. It occurs because of liberibacter, phloem-limited fastidious, and candidatus. It is a bacterial and most destructive disease that reduces fruit production globally. With low soluble solids and high acidity, fruit tainted by HLB juice becomes abnormally sour. It causes the fruit to be malformed and break. It causes malformations and cracks on fruit.	[16,23,41,44,59,64,72,78]
Wind scar	Windy weather can cause tree branches to rub or hit growing fruit. These scars are likely to only appear on the fruit’s surface and do not, in general, compromise its consistency. Tissue damage contributes to the entry of bacteria and fungi and colonize the tissue, resulting in further damage.	[22,26,34,44,58,70,75]
P. digitatum fungus	Infections because of Penicillium spp. fungi are a primary concern impacting citrus development. Actual losses due to this fungus are complex and rely on the processing location, citrus type, age, environmental conditions of the developing and harvesting stage, the degree of apparent damage during handling, and the climate after harvest.	[21,25,31,48,53,56,59,60,62,65,79,86,88]
Greasy spot	It is more intense where concentrations of rust mites are high. Rind blotch considerably lowers the extrinsic properties of fruit for sale at the market. Grapefruit is of significant concern, but this disease can also be present in oranges and other citrus fruits. Rind blotch decreases the amount of grapefruit that is suitable for the fresh market.	[26,44,50,70]
Surface defects	Mechanical injury (bruises) and indications of illnesses entail surface defects. Surface injuries include fruit blemishes, burns, and wrong structures.	[18,30,32,38,40,43,49,55,61,71,77,79,82,83,85,87]
Chilling/freezing disorders	This post-harvest disorder occurs due to the chilling effect, and it is a catastrophic disorder that evolves with the storage of fruits at low temperatures that notably demotes the quality of citrus fruits in the market or disqualifies them from the market. The impact of the chilling disorder depends upon the temperature at which fruits are to be stored or the duration of the time spent by the fruit in cold storage. Its severity increases with lower storage temperatures and long durations of exposure.	[29,46]
Other color defects	Other miscellaneous defects are stubborn, brown rot, black mould, green spot, color defects, copper burn, blemishes, morphological disorders, stem-end rot, etc.	[17,19,20,22,24,26,27,30,34,63,68,70]

provides a brief introduction to each disease, along with relevant references. Other miscellaneous defects with their count shown in brackets are stubborn (3), brown rot (1), black mould (3), green spot (1), color defects (3), copper burn (2), blemishes (3), morphological disorders (2), and stem end rot (1).

Summary: One of the most exemplary illustrations of the diverse tactics bacteria use to invade particular parent species is the interaction of citrus species with various bacterial diseases. Among several mechanisms, the most prominent and frequent are discoloration, foul smell, and cracking. Among the proteins potentially used in obtaining transgenic plants resistant to bacterial citrus diseases are planted recognition receptors, master regulators of the SA pathway, cecropin, and thionins.

• RQ2: Which techniques have been used to capture citrus fruits’ disease-related patterns?

It is evident that statistical techniques such as linear regression, MBLR, SLR, HMM, etc., were used during the initial phase. Not only were these approaches overly mathematical and often unable to manage the noise contained in the data, more accurate models based on machine learning techniques were implemented in later phases from 2000 onwards. Three different classification techniques were analyzed in this SLR:

• Machine learning;
• Deep learning;
• Statistical techniques.

This literature study also found that in 1995, only one paper used two different techniques, i.e., a neural network model and a Bayesian approach, to detect different kinds of diseases in grapefruits, tangerine, and oranges. In 1998, one paper was found to detect surface defects using a deep learning technique. After that, one paper was found in the year 2001 that used the deep learning technique, and in the year 2004, one paper was found that used two different statistical techniques for the detection of detects in citrus fruits. Figure 5 shows the contribution of three techniques (machine learning, deep learning, and statistical techniques) from 2006 to 2020.

• Deep learning models used for capturing disease-related patterns of citrus fruits.

This SLR analyzed DL models used by the included studies. As deep models are more complex versions of or extensions of neural networks with a more significant number of hidden layers, we also included all types of neural networks in this section, which are written as follows:

• Artificial neural networks (ANN);
• Convolution neural networks (CNN);
• Probabilistic neural networks (PNN);
• Multilayer perceptrons (MLP);
• Associative neural networks (AANN);
• Radial basis probabilistic neural networks (RBPNN);
• Back propagation neural network (BPNN).

Figure 6 shows the number of studies that used DL techniques for citrus fruit diseases. It can be observed that the most commonly employed technique was ANN, which was employed in about 15 studies. The second most used DL estimation technique is CNN, which was investigated in 11 different studies. BPNN was investigated in about six studies, and RBPNN was investigated in four, while multilayer perceptrons (MLP) were employed in three different studies. Finally, PNN and AANN were employed in one study assessed in this SLR.

• Machine learning models used for capturing disease-related patterns of citrus fruit

The analyzed ML techniques used by all the selected studies in SLR are as follows:

• Support vector machines (SVM);
• Bayesian networks (BN);
• K-nearest neighbors (KNNs);
• Decision trees (DT);
• Genetic programming (GP);
• Classification and regression tree (CART);
• Naïve Bayes;
• Random forest (RF);
• K-means clustering;
• Fuzzy;
• Extreme learning machine (ELM);
• Ensemble learning (Adaboost);
• Ensemble boosted tree (EBT).

Figure 7 shows the count of ML techniques that have been used in the selected studies. The most widely employed ML technique is SVM, which was employed in nearly 17 research papers. Different kinds of SVM were used, such as multi-class SVMs, RBF kernel (RBF-SVM), Mahalanobis kernel (MK-SVM), etc. Further, the second-most-used estimation technique was K-nearest neighbors (kNNs), which was evaluated in about 13 different selected studies. Different types of K-nearest neighbors (kNNs) have been used here, such as the edited multi-seed nearest neighbor technique, the nearest neighbor prototype technique, and weighted K-nearest neighbors (W-KNN). Fuzzy was employed in nearly 4 selected studies; CART was employed in 2 studies, while Bayesian networks were employed in 2 different selected studies. GP, CART, and naïve Bayes were used in two studies, and ELM, Adaboost, and EBT were used in one study for each technique. Finally, DT was employed in four studies assessed in this SLR. RF and DTare were considered by one group, whereas CART was also classified in this category by some researchers. Different studies used the genetic algorithm and K-means clustering for feature selection and image processing (segmentation) purposes, respectively.

• Statistical techniques used for capturing disease-related patterns in citrus fruits.

• PLS-discriminant;
• Discriminant analysis;
• Regression;
• Decision tree (LDA);
• Principal component analysis (PCA).

Figure 8 shows many statistical techniques that have been used in the selected studies. It was found that the most widely used statistical technique is PCA, which was used in about 12 studies, followed by LDA, which was used in 8 studies. The regression technique evaluated in around seven different experiments was the third most-used estimation technique. Different types of regression techniques were used, such as logistic regression (LR), principal component regression, Gaussian process regression, multiple linear regression method (MLR), linear regression, and multivariate regression. PLS-discriminant regression was investigated in about three studies, and other discriminant analysis methods were investigated in three studies.

• RQ3: Which techniques outperform other techniques in terms of classification accuracy?

Table 3. Overall techniques that outperformed other techniques.

Best Technique	Outperformed Techniques
SVM	ANN [76]	W-KNN, EBT, DT, LDA [41]	CNN, KNN [71]	ANN, KNN, SRS [81]	KNN, NaïveBaYes [23]	DT, Fuzzy, ANN [33]	MLP, RBF [49]	KNN [69]
ANN	SRC [81]	SVM, LR, [23,29]	DT [33]	Naïve Bayes [23,32]	KNN [33,81]	Fuzzy [33]	LDA [29]	QDA [29]
DT	Naïve Bayes [32]	MLP, RB [49]	Fuzzy [33]	EBT [41]	LDA [86]	ANN [32]	SMO [49]
LDA	PCA [60]	EBT [41]	CART [48]
BaYesian	NN [22]
CNN	KNN [69]	ANN, DT, Fuzzy [33]	SVM [33,69]
NNRB	KNN [40]
ASNN	SVM, [55]	BPNN [55]
PLS	MLR [74]
FA	PCA [21]	SM [21]
Adaboost	SVM [23]	LR [23]	KNN [23]	Naïve Bayes [23]	NN [23]
RF	SVM [42]
KNN	CNN [71]	FUZZY, DT [33]
PCA	SM [21]
LR	SVM [23]	Naïve baYes [23]	KNN [23]
W-KNN	EBT [41]
ELM	SVM [42]
BPNN	SVM [55]

Table Abbreviations: NNRB—neural network radial basis, AANN—associative neural network, PLS—partial least squares regression, FA—factor analysis, RF—random forest, ELM—extreme learning machine, RB—radial basis, BPNN—backpropagation neural net, SM—Sammon mapping.

shows the techniques that outperformed the other techniques in the comparison studies performed by the different experiments. It was found in the studies that more than one experiment may be included by a single study based on different datasets or methodology. The first technique that performed best is SVM, which outperformed other techniques in 8 different experiments. ANN is the second technique that performed better than the other nine techniques in 5 experiments. The next most-well-performing techniques are DT and CNN, which performed better than eight and five different techniques in eight and four experiments, respectively. The most promising statistical technique we found in the studies is LDA, which was compared with other techniques in different experiments. In three different experiments, LDA performed better than PCA, EBT, and CART.

Summary: It is observed from Table 3 that in six different experiments, SVM was the best technique, outperforming W-KNN, EBT, DT, naïve bayes, fuzzy, and RBF techniques. The second best-performing technique is the decision tree, which was assessed in 5 different experiments and compared to the naïve Bayes, RB, fuzzy, EBT, and SMO techniques. We found that the Adaboost ML technique is a significantly less-explored technique in the classification of citrus fruit diseases. The Adaboost ML technique was used in only one study and outperformed many ML and DL techniques. Lastly, some more techniques, such as the random forest, KNN, ELM, and FA techniques were also well-performing ML techniques.

Similarly, in some experiments, ML techniques performed better than other DL techniques. SVM again performed better than DL techniques such as ANN, CNN, and MLP in 5 different experiments. Additionally, we observed that the DT, Bayesian, Adaboost, KNN, W-KNN, and ELM techniques also performed better than other DL techniques in one or two experiments.

It can be observed that ANN and CNN are the two best techniques that performed better than other ML techniques in 5 and 2 experiments, respectively. We also found that the other DL techniques, such as neural network radial basis, associative neural networks, and backpropagation neural networks, outperformed the other ML techniques in different experiments. It can be observed that the LDA technique is the best technique among all the statistical techniques in comparative studies assessed in our SLR. We also found other well-performing statistical techniques, which include partial least squares regression, PCA, and LR.

The SVM ML algorithm is found to be the most highly performing technique compared to all others. The ability of SVM to handle high-dimensional data comes into play in different comparison papers with unknown distributions. Other algorithms outperform SVM in some general papers, but these algorithms are not generally able to classify and address unknown variables with accuracy and efficiency compared to SVM. SVM claims to provide a significant improvement in classification accuracy over ANN. SVM proved to be a powerful method for automatically classifying the plant diseases studied in this study.

• RQ4: Which hybrid techniques were used to classify citrus fruit diseases?

A hybrid approach combines different ML, DL, and other techniques to improve classification accuracy. This survey found that 7% of studies used hybrid models that were either used for the feature extraction process or classification of the citrus fruit diseases. Figure 9 represents the distribution of the studies using different techniques for citrus fruit disease classification. It is observed that the majority of studies used ML techniques (37%), followed by DL techniques (31%). Statistical techniques provide a total contribution of 25% in this SLR.

Table 4. Hybrid methods used in the selected studies.

Hybrid Techniques	Description	Result
ALEX-NET and Random Forest	Feature extraction takes place using the Alex-Net model. Lastly, multiclass classification is implemented using the random forest technique to classify different diseases.	The described AGC-A algorithm reliably identifies various forms of citrus fruit diseases with an overall identification rate of 97.29% [72].
PCA and NN	The radial neurons are used as input and linear output neurons are used as the hidden layer in the RBF neural network model. The analysis was formulated by the newrb function in the Matlab programme, but this function is unstable. PCA was used for feature selection to enhance relative stability. A total of 10 variables were employed as the input of the neural network.	The shelf lives of 40 samples were estimated, and the prediction accuracy of the model developed in this paper was 80 [84].
SVMs with the Mahalanobis kernel	The paper adopted a methodology determined by the collective use of hyperspectral images and a classifier of the Mahalanobis kernel. More detailed and accurate findings were obtained relative to other approaches in multiple scenarios and acquired images. The different intrinsic importance of spectral channels was assessed more effectively by this kernel, as their relative value was learned from the results.	The total accuracy on the MK-SVM classifier test set through various levels of training samples for the clemenules and clemenvilla mandarin variants was 95% [86].
Genetic algorithms based on LDA	LDA-based genetic algorithms were used to determine one of the most suitable bands. The fitness function was chosen to re-examine the response of every execution while undertaking any GA method. The more extraordinary fitness feature could be chosen to produce a new generation. The study utilized the overall precision of LDA as its fitness function, and hence, the technique was called GALDA.	The overall result for clemenules using the LDA classifier with the GALDA system was 90%, and the results for clemenules using the CART classifier was 95% [48].
PCA-ANN	PCA was first used to decrease the dimensionality of the spectral data, which were then considered the input for the ANN modeling technique.	Soft independent class analogy modeling (SIMCA), a form of PCA combined with ANN and SVM techniques, was performed to recognize freezing damage in sweet lemons, leading to various experimental simulated freezing environments on the entire spectral content. Using PCA-ANN, the accuracy of classification was 100% [46].
CNN regression	The Brix/acid ratio of flesh juice was measured in this research by executing a regression analysis with a CNN (CNN regression). The most common CNN styles, comprising a convolution layer, a pooling layer, and a fully connected layer, were introduced. The regression layer was positioned in the final layer to implement CNN regression and measure the Brix/acid ratio.	The absolute error in the Brix/acid ratio was 2.48 employing CNN regression [36].
Principal component analysis followed by PLS-discriminant analysis	PLS-discriminant analysis (PLS-DA) with principal component analysis (PCA) was used to distinguish fruit per the canopy position. PCA and PLS-DA methods are focused on spectra obtained before harvesting, being unable to distinguish fruit depending on their canopy position.	The precision of the two regression approaches revealed that for screening between internal and outside fruit and sorting fruit depending on susceptibility to RBD, both strategies could be used solely or in collaboration [66].
Convolutional neural networks and a fuzzy motor	Centered on artificial intelligence algorithms using CNN and a fuzzy motor, a fruit-oriented automated quality evaluation framework was introduced that examined a set of external attributes of critical value to assess the Persian lemon quality in addition to mitigating the loss variables of objectivity that could lead to an operator to diverge from the selection criterion.	Invalidating characteristics, the CNN trained by transfer learning gained 97.5% performance, and the methodology developed for the characterization procedure was performed correctly. In the three classes suggested, each lemon was correctly categorized by the fuzzy method according to the characteristics of each lemon and the classification score provided by CNN [87].
PCA and multi-class support-vector machine (M-SVM)	The most relevant features were selected by implementing a hybrid feature selection technique that comprised entropy, PCA score, and a covariance vector premised on skewness. For the final classification of citrus disease, the identified features were fed to the multi-class support vector machine (M-SVM).	The proposed method significantly outperformed the other techniques and attains 97% accuracy for the image gallery dataset, 89% accuracy for the consolidated dataset, and 90.4% accuracy for their local dataset [69].
Mask R-CNN and merging algorithm	The ability to recognise fruits in various occluded situations is a crucial area of expertise for a fruit-grading robot system. An unified approach was created to concurrently recognise and quantify citrus fruits and branches using a branch segment merging method and a mask regional convolutional neural network. To increase the accuracy of the mask R-CNN, a segmental labelling technique was presented for randomized and asymmetrical branches. The minimal enclosing rectangle of the segmental mask portions predicted by the model was calculated to produce a much more exact bounding box. After that, the branches and the trunk were rebuilt using a branch segment merging technique. Casting the colour image out onto the contour allows for the measurement of the fruit and branch diameters.	Fruit and branch detection have typical accuracies of 88.15 and 96.27%, respectively. The average measurement errors for the diagonal dimension of fruit, the longitudinal radius of fruit, and the radius of branches are 2.52, 2.29, and 1.17 mm, respectively. The outcomes of the tests demonstrate that the detecting system performs well for all sorts of fruits and geometric distortions. The robot can effectively design a proper picking path and avoid collisions with the aid of this imaging system.

shows the hybrid techniques used in this SLR, along with their brief introduction and results.

Summary: Hybrid methods of classifying datasets have not been used popularly, but they produce results with significant accuracy. These methods produce greater accuracy and efficiency by combining classification methods instead of using them separately [23,32,46]. The above table notes the improved results using hybrid methods in different papers compared to the accuracy obtained by applying these methods separately to the previous question. Combining ML, DL, and statistical methods proved beneficial for the classification process.

• RQ5: Which features are to be extracted to classify citrus fruit diseases?

Numerous features can be utilized to depict an item and can be further contrasted with the details collected from non-objects for classification into different classes. Usually, the most sustainable features that are simple to measure and significantly contribute towards classification are the best [91]. The number of studies using different extracted features is shown in Figure 10. Our review found that the color features are the most widely used feature, followed by textural features. The results of our study show that 45% of studies extracted color features, 34% extracted texture features, and 21% extracted shape features.

Figure 11 shows the distribution of color spaces being studied. It can be observed from Figure 11 that the most frequently used color space is RGB, which was used in about 17 (35%) studies. The second most widely utilized color space was used in around 10 (23%) different studies. LAB color space was utilized in about nine (18%) studies; HSI was used in seven studies (14%), and YIQ was investigated in 2 (4%) different studies. Finally, NIR, YIQ, CMY, and YC${}_b$C${}_r$ were used in one (2%) study for each color space.

Several visual characteristics associated with fruit and vegetables are called features. Initially, fruit images are taken by a camera, and then pre-processing and segmentation techniques are applied to the images to filter, smoothen, and remove the noise of the images. After these steps, feature extraction takes place, which further helps to classify the diseases. Color is the most persuasive aspect and substantial descriptor that frequently improves feature extraction for the image analysis of fruits and vegetables. Color features play a crucial role in detecting and classifying disease in fruits. Different color spaces, such as HSV, RGB, HIS, and YCbCr, can be employed for classification purposes. Two important size features are area and perimeter, and these can also be evaluated by obtaining the pixel count of the images and adding up the distance of each adjacent pixel at the boundary, respectively. For food and vegetable quality analysis, the most common size features are the area, perimeter, length, and width. Apart from these features, major axis and minor axis features can also be determined for classification purposes. The major axis is the largest line through the fruit or vegetable product, which is determined by the measurement of the distance between the two boundary pixels of each mixture and the selection of the longest distance. The textural feature computed from the pixel group reflects the distribution of components and the morphology of the surface and is useful for computer vision, which determines the surface in the context of entropy, roughness orientation, contrast, etc. Numerous features can be utilized to depict an item, which can be further contrasted with the details collected from a non-object for classification into different classes.

Table 5. Mathematical expressions of important feature matrices.

Metrics	Equation
Co-occurrence metric $Z_{▵p▵q}(x,y)$	${\sum }_{P=1}^{\alpha }{\sum }_{q=1}^{\beta }\{1,if{I}_m\left(p,q\right)=xand{I}_m\left(p+▵p,q+▵q\right)=y0,otherwise\}$
Inertia moment	${\sum }_x^{}\frac{Z(x,y)}{Normalization}{|x-y|}^2$
The absolute value of the difference	${\sum }_x^{}{\sum }_y^{}\frac{Z(x,y)}{Normalization}|x-y|$
Regular value of the difference	${\sum }_x^{}{\sum }_y^{}\frac{Z(x,y)}{Normalization}|y-x|$
The modified absolute value of the difference	${\sum }_{x,y}^{}\frac{Z(x,y)}{{\sum }_{Im}^{}Z(I,m)}\frac{|x-y|}{|x+y|}$
Mean $\left(\underset{\_}{\mu }\right)$	${\sum }_{x=y=0}^{M-1}\left(p-q\right)Z_{▵p,▵q}\left(x,y\right)$
Standard deviation	${\sum }_{x=y=0}^{M-1}{(x-y)}^2{Z}_{▵p,▵q}\left(x,y\right)$
Entropy	${\sum }_{M=0}^{M-1}{m}^2\left({\sum }_{x=y=0}^{M-1}{Z}_{▵p,▵q}\left(x,y\right)\right)$
Variance	${\sum }_{x=y=0}^{M-1}{\left(i-\underset{\_}{\mu }\right)}^2{Z}_{▵p,▵q}\left(x,y\right)$
Kurtosis	${\sum }_{x=y=0}^{M-1}{(x-y)}^4{Z}_{▵p,▵q}\left(x,y\right)$
Skewness	${\sum }_{x=y=0}^{N-1}{(x-y)}^3{Z}_{▵p,▵q}\left(x,y\right)$
Range	max (Z (x,y)) – min (Z (x,y))
Homogeneity	${\sum }_{x=y=0}^{M-1}{Z}_{▵p,▵q}{\left(x,y\right)}^2$
Correlation	${\sum }_{x=y=0}^{N-1}{Z}_{▵p,▵q}\left(x,y\right)(x-\underset{-x}{\mu })$$(y-\underset{-y}{\mu })$
Contrast	${\sum }_{x=y=0}^{M-1}{Z}_{▵p,▵q}\left(x,y\right)\left(Z_{▵p▵q}\left(x,y\right)\right)$
Energy	${\sum }_{x=y=0}^{M-1}{Z}_{▵p,▵q}{\left(x,y\right)}^2$
Gradient module	${\sum }_{x=1}^M{\sum }_{y=1}^{M}Z(x,y)$ g${\left(x\right)}^2$, g(x) = x − M + 1
Intensity Symmetry	1 − ${\sum }_{x,y}^M\left(Z\left(x,y\right)-Z(M-1-x,M-1-j)\right|$
HSV Space
Hue component	$\{\theta b_c\le g_{c}360-\theta b_c\ge g_c\}$$\theta$ = $\{\frac{1}{2}(\frac{\lfloor \left(r_c-g_c\right)+\left(r_c-b_c\right)\rfloor }{\lfloor {\left(r_c-g_c\right)}^2+{(r}_c-b_c)\sqrt{g_c-b_c}\rfloor })\}$
Saturation component	1 − ( $\frac{3}{r_c+g_c+b_c}$) $\left(min(r_c{,g}_c,b_c)\right)$
Intensity component	$\frac{1}{3}$ ($r_c+g_c+b_c)$
Smoothness S	1 − $\frac{1}{1+{\sigma }_x^2\left(Z_L\right)}$
Consistency C	${\sum }_{x=0}^{M-1}{H}^2\left(Z_i\right)$

shows the mathematical expressions of the important feature metrics, such as the co-occurrence metric, the entropy, standard deviation, HIS components, etc. Table 5 shows the various extracted shape/size features used in the selected studies.

• RQ6: What evaluation metrics are commonly seen in studies for assessing techniques?

A range of metrics is used to measure the performance of different classifiers used for citrus fruit disease classifications. These evaluation parameters are often used to assess models developed using different DL, ML, and statistical methods. The evaluation parameters, their mathematical formulas and descriptions, and the count of the studies in which they are used are shown in

Table 6. Performance Measures used in Studies.

Evaluation Parameter	Formula Used	Description	Count
Accuracy	$\frac{correctlyclassifiedimages}{Totalnumberofimages}$	The proportion of accurate assessments among all the assessments.	47
Sensitivity	$\frac{TruePositives\left(TP\right)}{TruePositives\left(TP\right)+FalseNegatives\left(FN\right)}$	The percentage of infected classes correctly predicted among all existing infected classes. It is also called recall and the true-positive rate (TPR).	17
Specificity	$\frac{TrueNegatives\left(TN\right)}{TrueNegative\left(TN\right)+FalsePositives\left(FP\right)}$	Healthy fruit classes that are correctly predicted among all authentic, healthy classes.	8
Precision	$\frac{TruePositives\left(TP\right)}{TruePositives\left(TP\right)+FalsePositives\left(FP\right)}$	The percentage of infected fruit classes that are appropriately categorized among the total number of cases classified.	7
FPR	$\frac{FalsePositives\left(FP\right)}{FalsePositives\left(FP\right)+TrueNegatives\left(TN\right)}$	The percentage of healthy classes identified as infected classes.	3
FNR	$\frac{FalseNegatives\left(FN\right)}{FalseNegatives\left(FN\right)+TruePositives\left(TP\right)}$	The percentage of infectious categories classified as the regular class.	2
F-measure	$\frac{2\ast Precision\ast Recall}{Precision+Recall}$	The harmonic means of accuracy and sensitivity. It expresses the equilibrium between recall and precision.	7
Percentage error	$\frac{|estimatedno.-actualno.|}{actualno.}\times 100$	The percentage error is the difference between the expected number and the actual number relative to the actual number measured in the percentage format.	5

.

Figure 12 presents a study count that assesses performance metrics. The most widely used performance metric is the accuracy, which is followed by recall and precision. Specificity, p-value, F-measures, and percentage error are other widely used metrics of assessment. Some metrics that are not counted in the graph with only one number, namely the G-mean, coefficient of correlation, and MCC, are less general.

Findings: The accuracy of an experiment, object, or value is measured by how closely its results agree with the actual or accepted value. This is the most reliable and most commonly used performance metric. We found different research papers utilizing classification accuracy to compare the different methods in the conducted studies. The second-most-used metric was found to be the recall, as the research papers in question discuss its findings using the metric to measure sensitivity to changes. The table highlights the importance of accuracy over other performance metrics.

The data were analyzed from the selected studies published from 1995 to 2020. In 1995, one study was found, which was published in the USA; in 1998, the one selected study was published in Belgium. Between duration 1999 and 2000, no papers were found that worked mainly on diseases of citrus fruits (excluding citrus leaves, stems, etc.). For the years 2001, 2004, 2006, and 2007, only a single study was found for each year; these studies were published in the USA, Switzerland, India, and United Kingdom. It can be observed that in recent year, an enormous amount of work has been done in this area, specifically in 2019; i.e., 15 studies were published in this year. Additionally, Figure 13 depicts that a maximum contribution is provided by the Netherlands (23), India (21), and the USA (12) in the field of the detection of diseases of citrus fruits.

5. Summary and Findings

In terms of valuation, citrus fruits are a vital fruit crop in the global market and have also contributed to a considerable effort to simplify multiple assessment practices along the supply chain. From automated fruit inspection inside and outside the environment to yield analysis with verified efficiencies and accuracies, machine vision has been demonstrated to have outstanding potential and pragmatism.

Table 7. Meta-synthesis.

S.No	Point	Facts
1	Which technique was popularly used for the classification of diseases of citrus fruits in the studies assessed in our SLR?	Machine Learning
2	Popularly used deep learning technique for citrus fruit disease classification	CNN
3	Popularly used machine learning technique for citrus fruit disease classification	SVM
4	A popularly used statistical technique for citrus fruit disease classification	PCA
5	Popularly used image processing technique	Thresholding
6	Technique that outperformed other techniques (ML/Non-ML) in terms of accuracy	Machine Learning (SVM)
7	Defect type on which most of the state-of-the-art work has been performed	Surface defects
8	The most frequently studied citrus fruit	Orange
9	The most commonly used evaluation parameter	Accuracy
10	The number of studies using ML/DL/statistical methods
11	Name of the disease on which most of the work has been performed	Canker
12	Name of the fungi on which most of the work has been performed	P. digitatum

includes the meta-analysis conducted in this SLR.

6. Limitations

There were many challenges that the researchers faced during the execution of this work. The predominant issue was the affordability of a regular database because of the disease, pathogens, and infections present in fruits. This shortage of accessibility for researchers and scholars decreases the ability of a database to facilitate work to be carried out in this area. Standard publicly accessible databases are also required to enhance the overall efficiency of such initiatives and make wide-ranging computer-aided prognostic models suitable for identifying and classifying various diseases with more precision. The implementation scenarios are constrained in some situations since the development of fruit trees is dynamic; the collection of image datasets at various durations of growing time reflects different characteristics that significantly contribute to complex differences in the output of the system. It is not easy to obtain real-time datasets from the orchards of citrus fruits because of environmental variability. The choice of the disease type and the signs that are individually described or classified for samples from another set of citrus fruits is another important consideration for authors and researchers. There is a significant need to implement an automated system for image analysis and classification that largely depends on the chosen ideal wavelengths to improve citrus disease detection performance. It is also inferred from the literature review that the fruit sample should be collected from different areas or regions with different characteristics to achieve a fair outcome.

7. Conclusions

This paper described an SLR focused on disease identification and classification in citrus fruits using machine learning, deep learning, and statistical techniques covering almost two decades. A total of 78 studies were selected from 1995 to 2020 (March) for further analysis and evaluation to obtain important information for the users. The latest review of the outcomes associated with citrus fruit disease classification and integral methodologies is introduced in this SLR. In the era of smart agriculture, image acquisition, image processing, feature extraction, and classification techniques are essential components for recognizing and predicting various diseases present in citrus fruits. This paper presented different conceptualized theories related to all the essential components of the recognition and classification of citrus fruit diseases. This SLR has addressed nearly all the state-of-the-art frameworks applicable to the detection of diseases in citrus fruits. Our goal is to make researchers and scholars more interested in developing and applying new technologies in this area. The paper also addressed stepwise measures to build a necessary automatic framework to protect fruits from apparent disease by answering nine research questions. As for the results and comparisons, a meta-analysis section has been included in this SLR. In future work, more importance can be given to the technical aspects or methodology used in the most significant papers for the better promotion of the research work.