An improved moth flame optimization algorithm based on rough sets for tomato diseases detection

An improved


Introduction
Plants are very crucial source of food and energy for humankind.Plant diseases can cause major economical, and ecological losses as well as reduction in both quantity and quality of agricultural products.Therefore, diagnosing and detecting plant diseases in a timely an accurate way is very important.Usually, the observation of experts using their naked eyes is the traditional approach followed in practice for the diagnosing and detection of plant diseases.Moreover, in some developing countries, small farmers could nd diculties to get experts making consulting these experts very expensive and time consuming.This could lead to the spreading of the disease into all crops.Thus, automatic/computerbased plant diseased detection approaches are of high importance.The automatic detection system usually consists of two main phases.Firstly, the plant leaf image is captured using a digital camera.Secondly, the detection and classication of leaf diseases can be achieved through dierent steps: extracting the infected region, computing some features representing each disease and they classify these features to identify the diseases.The importance of automatic diagnosing and detection of plant diseases emerges as it could support benets in monitoring big elds of crops, hence provide automatic detection of diseases based on the symptoms which appear on the plant leaves (24).
In last years, automatic detection of plant diseases attracts many researchers in dierent elds.Bauer et.al., (8), proposed an approach for the automatic classication of leaf (i.e.,sugar beet) diseases using high resolution multi-spectral and stereo images.In (36), Weizheng et al., introduced a new fast and accurate approach for grading plant diseases using computer image processing technique.
They rst used Otsu method to extract the leaf region, and then used Sobel operator to detect edges of the diseased spot.Finally, plant diseases are graded through the information of the quotient of disease spot and leaf areas as indicator.In another study (25), Naidu et al. suggested a method to identify virus infected grapevine using the discriminant analysis and they obtained a maximum accuracy of 81% of the classication results.Also, cotton diseases (10) were automatically identied using preprocessing operation and the use of SVM classier to identify visual symptoms of cotton diseases.Moreover, in (20) a new method for wheat disease identication using image recognition was proposed.In this method, after computing features of diseased region of leaf images, samples are trained and recognized using the RBF-SVM classier.In (29) to classify the leaf brown spot and the leaf blast diseases of rice plant, an automated system has been developed.This system is based on the morphological changes of the plants caused by the diseases and used the Bayes and SVM classiers in the disease identication.Also an approach to detect the symptoms of nutrient diseases (4) was suggested and it is based on the vision system and pattern recognition.
The feature selection process is one of the most important tasks for pattern recognition and classication systems, e.g.plant disease detection system.The main goal of this process is to nd a minimal feature subset from a problem domain such that to give a high accuracy in representing the original features (12).It improves the predictive accuracy of algorithms by reducing the number of features, removing irrelevant, noisy and redundant features.It is also helps in the improvement of the classication performance.The feature selection mechanism has been successfully employed to eectively solve classication problem in various areas, such as bioinformatics (32), image processing (31), data mining (22), pattern recognition (34), medical diagnosis (2; 33).Dierent techniques were used to achieve feature selection.This includes the rough set theory (28) and bio-inspired techniques.The basic idea of using rough set-based for feature selection is to generate all possible feature reductions and then choose the one with minimal cardinality (19).The rough set has already used to accomplish a features selection task in dierent area such as: (13; 38; 6).Also, many bio-inspired methods have been used for feature selection process and thes include genetic algorithm (GA) (21; 27), ant colony optimization (ACO) (7; 1), Bat Algorithm (BA) (26; 30) and Grey Wolf Optimizer (GWO) (14).
Eorts have been targeted to combine the RS approach with bio-inspired algorithms to improve the performance.Bello et al. (9) proposed an feature selection approach which integrates Ant Colony System with rough set.The approach rstly generates a number of ants which are placed randomly on the graph and then they traverse edges probabilistically until a traversal stopping criterion is satised to output the best rough set reduct.This method achieved a high ratio in features reduction but the classication accuracy and execution time are not good enough.Similar to the Bello's approach (9), Wang et al., (35) introduced an approach integrating between rough set and the particle swarm optimization (PSO)to achieve the feature selection task.They followed the same idea but only applied PSO instead of ACS.Wang's approach was able to nd the optimal reducts on most of the used datasets and minimizing the execution time.
In another eort, Guo et al., (18) proposed an approach combining between Genetic Algorithm, GA, and rough set for the feature selection.Firstly, rough set was used to carry out the feature selection, then to nd the optimal subset in the remaining feature subset, they used the GA improved with Population Clustering.The SVM (Support Vector Machines) was then applied to evaluate the eectiveness of the selected feature subset.
In this paper we proposed a Moth-Flame Optimization (MFO) and rough set (MFORSFS) approach for automatically detecting some kinds of tomato disease.The tomato was chosen to be the application of the automatic disease detection in this study because of its importance.It is ranked number one among 40 vegetables/fruits in terms of "relative contribution to human nutrition" and contains a high nutrition value.To achieve tomato disease detection, feature selection is a important phase.Thus, we rst have introduced a new feature selection technique based on MFO and Rough Set called MFORSFS.This MFORSFS was evaluated to prove its robustness and then we have used in the detection of the tomato diseases.The proposed MFORSFS algorithm was compared against using (1) Particle Swarm Optimization (PSO) and (2) Genetic Algorithm (GA) with the rough sets.The results showed that the MFORSFS gave a higher accuracy of classication results while preserve low number of features compared to the other two optimization algorithms.
The rest of this paper is organized as follows: Section 2 gives an overview of the moth ame optimization and rough sets.Section 3 presents the details of the proposed system.In Section 4, experimental results and discussion are given.Finally in Section 5, conclusions and future work are presented.

Gabor Features
Gabor lter-base method is an eective method for extracting texture feature.It has been used in many applications such as biometrics and segmentation.Gabor lters are known as convolution kernel, the product of a cosine and Gaussian functions.It enjoys the characteristic of specied orientation and spatial frequency.The 2-D Gabor lter is like a local band-pass lter with some localization properties in the spatial and frequency domain.Gabor lter is proved his eciency in characterizing texture features (17), like in our case: extracting texture features from tomato's leaves.
A 2D Gabor function g(x, y) is dened as follows: where σ x and σ y characterize the spatial extent and frequency bandwidth of the Gabor lter, and W represents the frequency of the lter.Let g(x, y) be the mother generating function for the Gabor lter family.A set of dierent Gabor functions g m,n (x, y) can be generated by rotating and scaling g(x, y) to form an almost complete and non-orthogonal basis set, that is, Where x = a −m (x cos θ n + y sin θ n ),ý = a −m (−x sin θ n + y cos θ n ) , a > 1, θ n = nπ/K, m = 0, 1, . . ., S − 1, and n = 0, 1, . . ., K − 1. Parameter S is the total number of scales, and parameter K is the total number of orientations.
So, S and K represents the total number of generated functions.
Given an image I(x, y), its Gabor-ltered images are

Feature Selection Overview
In the past few decades, classication problems resolved using machine learning techniques usually contains high dimensional of data.Such high dimensionality lead to challenges such as the curse of dimensionality or a large number of features.These challenges tends to overt problem which results in performance degeneration.To address this problem, feature selection has been introduced.
The main purpose of feature selection is to determine a minimal feature subset of a problem domain such that retaining a suitably high accuracy in representing the original features (12).
According to using labeled or unlabeled training set, feature selection techniques can be classied into unsupervised (10)

Rough set basics
Rough set theory (37) is a mathematical approach to imprecision, vagueness and uncertainty.Rough Set Attribute Reduction (RSAR) (11) provides a lter-based tool for extracting feature from a domain in a concise way whilst reducing the amount of knowledge involved.To formalize the rough set, consider I = (U, A) is an information system, where U is a non-empty set of nite objects (the universe) and A is a non-empty nite set of attributes such that for ∀a ∈ A determines a function f a : U → V a .With any P ⊆ A, there is an associated equivalence relation IND(P): The Let X ⊆ U , the P-lower approximation PX and P-upper approximation PX of set Xcan be dened as: Let P,Q ⊆ A be equivalence relations over U, then the positive, negative and boundary regions can be dened as: The positive region of the partition U/Q with respect to P(POS P (Q)), is the set of all objects of U that can be certainly classied into blocks of the partition.
An important issue in attribute reduction is discovering dependencies between attributes.U/Q by means of P For P,Q ⊆ A, we say that Q depends on P in a degree k (0 ≤ K ≤ 1) If k = 1, Q depends totally on P, if 0 < k < 1, Q depends partially (in a degree k) on P, and if k = 0 then Q does not depend on P.
In a decision system, an attribute set includes two sets: decision attribute set D and condition attribute set C, i.e.A = C ⊂ D. The degree of dependency between these two sets, γ C (D) , which is known as the quality of approximation of classication, is induced by the decision attributes set (37) .
When P is a set of condition attributes and Q is the decision, γ p (Q) is the quality of classication (37).The goal of attribute reduction is to remove redundant attributes so that the reduced set provides the same quality of classication as the original.A reduct is dened as a subset R of the conditional The set of all reducts is dened as: In rough set attribute reduction, a reduct with minimal cardinality is the one being searched for.To locate a single element of the minimal reduct set Red min ⊆ Red, the following equation is used : The intersection of all reducts is called the core, the elements of which are those attributes that cannot be eliminated.The core is dened as: The mathematical model for the MFO is based on two components, moth and ame.The moths are actual search for agents that move around the search space, whereas ames are the best position of moths that obtains so far.As mentioned above the inspiration of this algorithm is the transverse orientation.

Moth Flame Optimization
In order to mathematically model this behaviour, the position of each moth is updated with respect to a ame using the following equation: where M i indicates the i-th moth, F j refers to the j-th ame, and S is the spiral function.The logarithmic spiral for the MFO algorithm is defended as follows: Where D i indicates the distance of the i-th moth for the j-th ame and is as dened in 16, b is a constant for dening the shape of the logarithmic spiral, and t is a random number in [ -1, 1].D is calculated as follows: Where M i indicates the i-th moth, F i denotes the j-th ame and D i refer to the distance between M i and F i .
The t parameter in the spiral equation 15 controls the direction of moth navigation around the ame.(t = -1 is the closest position to the ame, while t = 1 shows the farthest) The spiral equation allows a moth to y around a ame and not necessarily in the space between them.Therefore, the exploration and exploitation of the search space can be guaranteed.
In order to further emphasize exploitation, t is dened as random number in [r , 1] where r is linearly decreased from -1 to -2 over the course of iteration.
According to equation 15,each moth is restricted to move towards a ame that may lead to local optimum stagnation.In order to prevent this, at each iteration, a list of ames must be updated and sorted based on their tness values.The moths then update their positions with respect to their corresponding ames.
Since the position updating of moths with respect to n dierent locations in the search space may degrade the exploitation of the best promising solutions, an adaptive mechanism for the number of ames has been proposed as in the following formula: where l is the current number of iteration, N is the maximum number of ames, and T indicates the maximum number of iterations.

The proposed MFO-based rough set tomato diseases detection approach
The proposed MFO-based rough set tomato diseases detection approach is comprised of ve fundamental phases: image acquisition, pre-processing, feature extraction, feature selection and nally classication.These phases are described in details below.The overall architecture of the proposed system is illustrated in Figure 1.

Image acquisition phase
The rst phase of the proposed MFO-based rough tomato diseases detection approach is the image acquisition phase.This phase plays an important role in any image classication system.These images must select carefully to achieve

Pre-processing Phase
In this phase, after collecting the dataset, the images were enhanced by removing noise that caused by defects of camera ash or hight lights to increase the eciency of classication and prediction process.Firstly, every leaf was isolate and extract in single image.Secondly, captured images were resized to 512 x 512 resolution, thus minimizing the storage capacity and reduce the computational time in the post-processing.Finally, the background of each image was removed using background subtraction technique with some morphological operations.Gaussian Mixture-based Background/Foreground Segmentation Algorithm (39) was used to subtract the background and morphological techniques Figure 2: Samples of infected tomato using in this work (dilation followed by erosion) to remove noise.

Feature extraction phase
In this phase, Gabor transform was used to describe the textural pattern of diseased tomato leaves.The total number of extracted features are 402.For more details of this phase reader can refer to (24).Each of used Gabor lters was implemented as a 8 x 8 convolution mask for each of its real and imaginary components.The acquired images were converted to HSV color space and 6 components of the image (R,G,B,H,S,V) have been extracted.To construct feature vector of each image components; a vector of 64 length was obtained from the average output for every i th lter.Vector of 3 length consisted of: cost function J(i), maximum average output D i max and minimum average output At the end of this step feature vector of (64+3) x 6 = 402 length that describe the image has been obtained.

Moth ame based features selection phase
As it was mentioned above, the output of the feature extraction phase is 402 features.Such large number of features usually contains irrelevant and redundant features.To achieve the feature selection phase, the MFO algorithm was employed through using both of rough set and SVM classier as a tness function for the MFO to evaluate the best set of features helping achieving the highest accuracy.The MFO algorithm was adopted in this paper for the following reasons.Firstly, in the original paper introducing the MFO (23), it is reported that the MFO algorithm has advantages on other related algorithms such as PSO, GA, and GSA in the context of optimization problems.Secondly, it is proved that MFO has the ability to solve real problem such as marine propeller design (so it could be useful algorithm in our case too (the detection of tomatoâs diseases).Thirdly, the MFO convergence is guaranteed since the moths always have the habit of updating their positions according to ames which are the most promising solutions.
The overall proposed MFO based rough set feature selection algorithm is described in Algorithm 1.
In the MOF-rough-set feature selection approach, the solution space represents all possible selections of features selection.Each moth position represents binary selection of feature sets of length N , where N is the total number of attributes.Every bit represents an attribute where the value `1' means that the corresponding attribute is selected while `0' means it is not selected.Each position is an attribute subset.The frequency of a position updating for each moth is represented as a positive integer, varying between 1 and max-update.It implies how many of the moth's bits (features) should be changed, at a particular moment in time.
The maximum range of position updating serves is a constraint to control the global exploration ability of a moth.After many tests, it was found that an appropriate maximum of position updating of each moth value is (1/3)*N .
Also, this maximum range was proven to achieve good results as reported in (35).Figure 1 illustrates the Layout structure of the proposed MFO-based rough set approach.
It is important to highlight the used parameters in the feature selection approach, as given in For the population initialization: The population initialization mechanism was used in the proposed algorithm and in all PSO and GA based ones using in the experimental evaluation, see Section 4. When population is randomly initialized, a feature subset (solution) should be produced randomly by Where i ∈ {1, 2, ....P N } and j ∈ {1, 2, ...F N }, where P N is population size and F N is number of feature.
For the tness function: it was a measure to determine the goodness or quality of a single solution in a population.At the end of each iteration, tness value is calculated of each agent for evaluating quality search.In this paper, classication accuracy was adopted as tness function and the Support vector machine SVM classier was used to evaluate the performance of each solution.
The classication accuracy obtained was based on the average of the 10-fold cross validation method.Since we must take into account two important issues, the classication quality and feature subset length.So, the tness function is calculated according to the following equation: Where γ R is the classication quality of condition attribute set R relative to decision D, |R| refer to the length of elected attribute subset.|C| is the total number of features.α and β are two parameters corresponding to the importance of classication quality and subset length, α ∈ [0, 1] and β = 1 − α.We adopt this approach based on the work done in (35), they states that classication quality is more signicance than the size of subset, as a result both parameters have been set as follow: α = 0.9, β = 0.1.

SVM-based classication phase
In the classication phase, the SVM was employed to assess whether features selected using MFORSFS method can help in detecting infecting tomato leaves.
The inputs of this phase are trained feature vectors, whereas the outputs are the decision of whether the tomatoâs leaf is infected or not and if it is infected, it determines the type of disease (Powdery mildew and early blight).It is worth to mention that the SVM was used in two dierent phases.In the feature selection, it was used as a tness function to evaluate which set of features is best to represent the leaf (infected or healthy).In the classication phase, the SVM was also used to classify between the infected and healthy leaves.
To evaluate the performance of a classication system, the k-cross-validation, a common method to deal with small training sets in machine learning (3), was used.Cross-validation is a method to evaluate classier or predictive models.
In this method, the original sample is partitioned into two sets

Experimental Results and Discussion
To evaluate the proposed approach, two main scenarios were designed and tested.The rst scenario was for the evaluation of the MFO-Rough-Set based feature selection approach using benchmark datasets.Also, in this scenario, to make the MOF+rough set feature selection approach comparable with related work, PSO and GA were also combined with the rough sets to achieve the feature selection.The three proposed features selection algorithms (MOF+rough set, PSO+rough sets, and GA+rough sets) were compared with each other to select the best one to choose a suitable combination of features in wrapper mode for maximizing classication performance and minimizing the data dimensionality.To make the results of the three algorithms are comparable, it was important to unify bases for all adopted bio-inspired algorithms.Thus, Population Initialization, Fitness Function are setup as described in Section 3.4 and the other parameters given in 1.All adopted bio-inspired algorithms were initialized identically and the used tness function was the same.
In the second scenario, the performance of the overall MFO-rough-set based tomato diseases detection approach was investigated.Three sub-scenarios were also designed here.Firstly, a simple classier, KNN, was used a tness function of MFO and its results were compared to the SVM-based ones.Secondly, a traditional feature selection, i.e., mRMR, was used to select the best features and the classication results were reported and compared with our proposed method.
Thirdly, three features selection algorithms (MOF+rough set, PSO+rough sets, and GA+rough sets) were applied in the feature selection phase to choose the best one.All algorithms were implemented using MatLab R2014b and all experiments were run under a computer with Intel(R) Core (TM) i7 CPU Q820@1.73GHZ and 8 GB memory and the system is Windows 8 Professional.
To evaluate the results in both the mentioned scenarios, several measurements were used.These measurements are Accuracy, specicity, Recall and F-Score.They are dened mathematically at Equations ( 20), ( 21),( 22) and ( 23) respectively (16).Using multi-level confusion matrix, each measure were calculated for each class, then the overall value were calculated on average of all classes.

Evaluating the proposed MFO-Rough-Set feature selection approach
To test our proposed feature selection approach, dataset from the UCI data repository ( 5) was used, Table 2 summarizes the 6 used data set for further experiments.To evaluate the proposed MFO-Rough-Set selection algorithm, the average classication accuracy of the selected feature subsets was used and it was measured using the 10-fold cross-validation method was used.This means that all values were veried ten times to ensure the reliability of the experiment.
The dataset was randomly separated into 10 segments.In each iteration, one segment was selected as test data (nonrepetitively) and the others were used as training data.To obtain a value of classication accuracy, the average of the results in each iteration was calculated.All of the experimental results are averaged over the 10 runs of 10-fold Cross-Validation.
In this experiment, all of MFO-Rough-Set, PSO-Rough-set, and GA-Rough-Set were tested on the 6 datasets mention above for selecting the best subset of features that eectively describe the dataset.As we mention before, several measurements are used to evaluate the performance of the proposed features selection algorithms.Table 3 shows the number of features selected in the best solution obtained for each optimization technique.As it can be observed from this table, the best obtained results produced from the new MFO feature selection algorithm that for most of the used dataset.Also the number of features resulted after using the new MFO feature selection algorithm always smaller than (or equal in some cases) other algorithms.

Evaluating MFO-based tomato Diseases Detection Approach
To assess the performance of the proposed MFO-based tomato diseases detection approach, rstly a real dataset of diseased tomato leaves were collected.
Then, a set of features describing the diseased tomato leaves were extracted.
These features were in a m × n matrix, where m = 200 is the number of used leaves and n = 402 is the number of features that describe each leaf.Three sub-scenarios were also designed here.Firstly, a simple classier, KNN, was used a tness function of MFO and its results were compared to the SVM-based ones.Secondly, a traditional feature selection, i.e., mRMR, was used to select the best features and the classication results were reported and compared with our proposed method.Thirdly, three features selection algorithms (MOF+rough set, PSO+rough sets, and GA+rough sets) were applied in the feature selection phase to choose the best one.

SVM-based vs KNN-based Fitness Function
Both of the SVM and KNN classiers were used in the evaluation of the quality of the MFORSFS methods.Two kernel functions (RBF, and Polynomial) of the SVM were used and KNN with k=1,3,5, and 7 were also used.A comparison were also conducted between the two classiers and the results are summarized in Table (5), and (6).
From Table (5), it can be noticed that when using the KNN as a classier with k=5, the highest results 87%, in terms of accuracy, precision and recall, was obtained from features were selected using with MFORS when its parameters are KNN with k= 5 Table (5), it could be seen that the highest results, 91.5%, in terms of accuracy, precision and recall, was obtained using: SVM-Polynomial as a classier from the feature selected by MFORSFS method with KNN is a tness function and k = 5.
From Table ( 6) and ( 5), it can be noticed that SVM-based classication, applied to the MFORSFS-based features with KNN as tness function, gave better results than that of the KNN-based ones.Where latter gave accuracy at 90.5% while the latter gave accuracy at 87

MFORSFS-based features vs mRMR-based features
A traditional feature selection, i.e., mRMR, was used to select the best features and the classication results were reported and compared with our proposed method.The mRMR experiments, four sets of features (rst 50,100,150, 200) were evaluated and the results are summarized in Table (7).From this table, it can be noticed that the highest accuracy results 90.5%, was obtained from using the rst 200 features ranked by mRMR when classied by the SVM-Polynomial.
Based on the obtained results and the results of our method in ), it can be noticed that our method is better than mRMR-based results.
From Table ( results (90.5%, the highest results in (Table (7).Both these results are obtained using the same kernel functions (polynomial) of the SVM classier.So, it could be claimed that our proposed method is better than the mRMR, the traditional   strates a comparison between the three methods in terms of the nal reduct size and execution time, respectively.
From, Figure 4, it can be noticed that the MFO-based selection algorithm gave the best results for the classication evaluation, and in the execution time.
Although, MFO-based method came the second in the reduct size (after the GAbased one), it gave the best in the classication performance and this is the most important in our case.The good performance of the MFO-based approach could be explained by the exploration power of the MFO and the the high performance of rough sets for the feature selection.Where the MFO algorithm uses the t parameter of the spiral equation 15.This parameter controls the direction of moth navigation around the ame, thus allowing each moth to y around ame sand not necessarily in the space between them.Consequently, the exploration and exploitation of the search space can be guaranteed.
Although the database was manually built in this study, an automatic process could be achieved as in the following scenario.A mobile app could be de- In this paper, a new approach for tomato diseases detection called MFO-512 based rough set tomato diseases detection approach was introduced.In this 513 approach, a now algorithm for feature selection (i.e.MFORSFS) was proposed, 514 implementedm, and evaluated.This approach is a combination of the MFO 515 and the rough set and used in the dimension reduction phase of the tomato diseases detection approach.Firstly, the MFORSFS was tested on well dened 6 datasets obtained from the UCI machine learning data repository and it was found that MFO-based approach outperformed PSO and GA-based ones.The MFORSFS was then employed the tomato disease detection approach to reduce the number of features to the ones that can eectively describe each leaf of the diseased tomatoes.The MFORSFS algorithm was compared against feature selection based on PSO and GA.It was found that MFORSFS gave much better performance, robustness and faster convergence.In the future, our approach could be improved by applying other parameters selection algorithms for best parameter values selection.
partition of U, generated by IND(P), is denoted U/P.The equivalence classes of the P-indiscernibility relation are denoted [x] p .The indiscernibility relation is the mathematical basis of rough set theory.
Moth-Flame Optimization (MFO) is a new optimization algorithm which simulate the moths navigation manner in nature.The main inspiration of this optimizer is the navigation method of moths in nature called transverse orientation(23).It is a population-based evolutionary computation search technique which mimics the behavior of moths in their special navigation methods at night.The idea of the MFO is based on a mechanism called transverse orientation for navigation in night throw the moon light.Using this mechanism, moth ies with a xed angle with respect to the moon.When moths see a human-made articial light, they try to maintain a similar angle with the light to y in straight line.Since such a light is extremely close compared to the moon, maintaining a similar angle to the light source causes a useless or deadly spiral y path for moths(15).

Figure 1 :
Figure 1: Layout structure of the proposed MFO-based rough set approach : a training set to train a given model, and another test set to evaluate this model.The general type of this method is k-fold cross-validation in which the original sample is divided randomly into k subsamples of equal size.From all these k subsamples, one subsample is used as the validation data to test the model while the remaining k − 1 subsamples are used as training data.The process of the k-fold cross-validation is repeated k folds (times) where each k subsamples is used as the validation data only one time.The main advantage of this validation method is that all samples are used for both training and validation, and each samples is used for validation exactly once.

Monks 3 3 3
Also in terms of the classication accuracy, Figure (3:a) the accuracy results before applying any feature selection (i.e. using all features) for all datasets.While gures (3:b-f) demonstrates the comparison, in terms of Accuracy, Recall, Precision and F-Score, results of classication evaluation after using the three feature selection algorithms.From these results, it can be seen that the classication evaluation results of the Monks dataset are the same as the Adult a: Whole dataset before feature selection b: Adults dataset c: Zoo dataset d: Iris dataset e: Soybean dataset f: Lung cancer dataset

Figure 3 :
Figure 3: Comparison between the results before and after employing the MFO, PSO, and GA based features selection algorithms using dierent datasets in terms of Accuracy, Precision, Recall and F-Score a) and (4: b) summarize the comparison results before and after employing the three features selection algorithms to original tomato's features (i.e., the 402 Gabor features).Also, gures (4: c) and (4: d) demon-

Figure 4 : 5 .
Figure 4: Visualization for the results of MFORSFS-based tomato diseases detection approach

Table 1 :
Parameters values used in experiments

Table 2 :
Description of the data sets used in experiments

Table 3 :
Number of features selected for each optimization technique

Table 5 :
(7)le(5)and (Table(7), it can be noticed that the MFORSbased classication results (91.5%) is better than that of the mRMR-based Classication results using KNN classier when the KNN (with dierent k values) and SVM-linear-Kernel were used as tness function in the features selection phase

Table 6 :
Classication results using SVM classier when KNN (with dierent k values) and SVM-linear-Kernel were used as tness function in the features selection phase.