Multi-Layer Perceptron Neural Network Model Development for Chili Pepper Disease Diagnosis Using Filter and Wrapper Feature Selection Methods

The goal of the current study is to develop a diagnosis model for chili pepper disease diagnosis by applying filter and wrapper feature selection methods as well as a Multi-Layer Perceptron Neural Network (MLPNN). The data used for developing the model include 1) types, 2) causative agents, 3) areas of infection, 4) growth stages of infection, 5) conditions, 6) symptoms, and 7) 14 types of chili pepper diseases. These datasets were applied to the 3 feature selection techniques, including information gain, gain ratio, and wrapper. After selecting the key features, the selected datasets were utilized to develop the diagnosis model towards the application of MLPNN. According to the model’s effectiveness evaluation results, estimated by 10fold cross-validation, it can be seen that the diagnosis model developed by applying the wrapper method along with MLPNN provided the highest level of effectiveness, with an accuracy of 98.91%, precision of 98.92%, and recall of 98.89%. The findings showed that the developed model is applicable. Keywords-chili pepper diseases; feature selection; multi-layer perceptron neural network; wrapper


INTRODUCTION
Chili peppers are plants used and consumed by Thai people in many forms. They are fundamental spices that can enhance the flavor, odor, and color of food. Chili peppers are considered unique due to their nutritional and medical benefits, flavors, and colors, so they cannot be replaced by other plants, hence, they are highly economically important in Thailand [1]. Chili peppers must be appropriately stored and maintained before being sold. Many farmers usually face problems in chili pepper plantations. These include inevitable natural disasters, weeds, and pests. Therefore, the farmers need to be educated about chili pepper diseases in order to protect their plants appropriately. Multi-Layer Perceptron Neural Networks (MLPNNs) have been broadly applied to diagnose diseases. For example, such models were developed in [2] and in [3] for predicting lung cancers and heart diseases respectively. These two studies developed disease diagnosis models which were more than 90% effective.
The current research aims to apply MLPNNs to diagnose chili pepper diseases. It also employs the filter and wrapper feature selection methods to develop the diagnosis model. The feature selection methods will select only significant features for synthesizing the model rapidly and classifying data more effectively [4]. The most effective model in terms of data classification will be further applied to the development of a chili pepper disease diagnosis system that can assist the farmers in diagnosing the diseases and treat their plants timely and adequately.

A. Filter Feature Selection Methods
Filter feature selection methods are the processes that evaluate the level of effectiveness of each feature regarding its compatibility with data analysis. These processes do not rely on any learning approach. These filter methods select the data by ranking each feature in a priority queue. The priority list is based on the number specified by the user or the threshold of the selected feature. The advantages of these methods are rapid data processing and independence from learning methods [4]. This research applies two types of filter feature selection methods, namely the Information Gain (IG) and Gain Ratio (GR), as follows.

1) Information Gain (IG)
IG is a measure used for classifying data by calculating the gain value of each attribute. If any attribute has the highest gain value, it will be selected as a subset with classification power, as shown in (1) [5]. Equation (1) demonstrates the calculation of entropy, while (2) explains the calculation of the gain value [5].
where c refers to the number of classes, ( ) i p t refers to the frequency value of class i for node t, Entropy(parent) refers to the entropy of the parent node, k refers to the total number of feature values, N refers to the total amount of data of the parent node, and N(v i ) refers to the total amount of instances of child node i.

2) Gain Ratio (GR)
GR is a measure that uses the gain ratio as the indicator for dividing datasets into sub-datasets based on the IG value. However, when the IG value is used for classifying datasets, there would be the occurrence of bias in favor of attributes with large numbers of values. Thus, the GR can be derived from the gain value divided by the SplitInfo value that is calculated in (3) [6], where k refers to the total number of split data. Therefore, the GR is calculated in (4) [5]:

B. Wrapper Feature Selection
Wrapper feature selection methods are processes that select subsets from all features. They explore the feature subsets which specifically match a learning approach. Therefore, these methods can increase the effectiveness of the learning process to the greatest extent. This research applies an evolutionary wrapper for selecting features. The features, which are predictor variables, are randomized into the equation each by each. Then, their effectiveness in prediction is tested. If the level of effectiveness in prediction increases, the effective features will be maintained. In contrast, if the level of effectiveness in prediction decreases, the ineffective features will be removed [7].

C. Multi-Layer Perceptron Neural Network
MLPNN is a form of perceptron neural network with multiple layers. It is suitable for complex computational tasks. It consists of 3 layer lavels, namely the Input layer, the Hidden layer, and the Output layer. The Hidden layer could be composed of many layers, but there must be at least one layer in it [6,8]. The data are the input of the Input layer. The output is sent out from the Output layer. The summation function of MLP is calculated in (5): where n is the total sum gained from the summation function, p i is the input of neuron i, w i is the weight of neuron I, k is the number of Input layer neurons, and b is the bias value. The MLPNN model is illustrated in Figure 1.

D. Model's Effectiveness Evaluation
The model's effectiveness was evaluated by the confusion matrix, which can calculate the precision, recall, and accuracy of the model [9][10][11][12].

E. Related Works
Authors in [2] developed a model to predict lung cancer using the MLPNN, comparing the efficacy of disease classification with the K-Nearest Neighbor (KNN) technique. The results showed that the disease classification model using the MLPNN was more effective. Authors in [3] applied the MLPNN to predict heart diseases. The efficacy of the heart disease classification model was compared with 9 classification methods and the results showed that the MLPNN was the most effective.
Authors in [13] applied the Wrapper method to select feature data to classify cardiac arrhythmias. The model's effectiveness was tested using 10-fold cross-validation, comparing 5 modeling techniques. The results showed that the model used to classify cardiac arrhythmias using Wrapper combined with MLP had the best performance. Authors in [14] developed an MLP model by applying the Correlation-based Feature Selection (CFS) and IG to analyze Thai water buffalo diseases. The experimental results showed that the developed model by CFS and MLP was efficient with accuracy, precision, and recall greater than 99.0%. III. METHODS In this work, a MLPNN model for chili pepper disease diagnosis was developed using filter and wrapper feature selection methods. Six 6 processes were involved: data collection, data preparation, feature selection, modeling, evaluation, and deployment.

A. Data Collection
Primary data were collected from 5 agricultural professionals and 33 chili pepper farmers in Ubon Ratchathani Province of Thailand with the use of questionnaires. This is a significant chili pepper cultivation area in 3 districts, namely Muang Sam Sip District, Mueang District, and Khueang Nai. The secondary data were collected from the review in [15]. An analysis of the data revealed 14 diseases in chili peppers in those plantations, as shown in Figure 2. The details of the data are illustrated in Table I.

B. Data Preparation
At the data preparation stage, the selected data were prepared before being applied to the model. The data were prepared in the following steps: 1) data selection, 2) data cleaning, and 3) data transformation. Ultimately, 863 questionnaires were processed, and were divided into 80% for training and 20% for testing. The data used for developing the chili pepper disease diagnosis model are illustrated in Table II   Once the data in Table II had been collected, the research team verified their accuracy. Then, they were converted into a csv file to be computed in Weka 3.9, as illustrated in Figure 3. After converting the data, the authors input the data into Weka to select features by applying IG, GR, and Wrapper.

C. Modeling
The data were divided into 4 datasets, namely 3 datasets undergoing the feature selection process and 1 original dataset used for MLPNN modeling. The research team specified the parameters for the MLPNN modeling as follows: Hidden layer = 2, Training time = 500, Learning rate = 0.3, and Momentum = 0.2. These parameters provided the best results.

D. Evaluation
The effectiveness of the chili pepper disease diagnosis model was evaluated by 10-fold cross-validation with the test dataset (20% of the total data). In addition, it was measured by examining the precision, recall, and accuracy of the developed models [6].

E. Deployment
After constructing the model, the most effective sample was applied to develop the chili pepper disease diagnosis prototype system. This system can diagnose chili pepper diseases timely and suggest proper treatments for each disease rapidly. The research model framework is illustrated in Figure 4. IV. RESULTS The chili pepper disease diagnosis model was developed by applying filter and wrapper methods along with MLPNN. The number of features remaining after the feature selection process is shown in Table III. The effectiveness of the models was evaluated by 10-fold cross-validation. The research results showed that the model developed by Wrapper and MLPNN (Wrapper+MLP) provided the highest results with accuracy of 98.91%, precision of 98.92%, and recall of 98.89%. Next in the rankings were IG+MLP, GR+MLP, and MLP (original data), as shown in Table IV and Figure 5. Furthermore, compared to other studies that detected or classified chili pepper diseases (Table V)   Once the feature selection process was completed, the selected data were applied to develop the chili pepper disease diagnosis model with the use of the MLPNN The experimental results indicated that the diagnosis model developed by the Wrapper method and the MLPNN provided the highest level of effectiveness, with an accuracy of 98.91%. This result means that the developed model can be used in a chili pepper disease diagnosis system, suggesting chili pepper disease prevention and treatment. The results of this research conform to the study conducted in [3], in which the authors applied the MLP method to develop a disease diagnosis model which proved to perform better than other known techniques. The research findings also conform to the findings in [13]. The authors employed the wrapper method and the MLP to classify the symptoms of cardiac arrhythmia. Their research results showed that the wrapper feature selection method could increase the effectiveness of the classification process.
In general, a disease on chili peppers can be identified by its symptoms. Most research is focused on image processing [16,17,20] and deep learning [18,19] disease analysis from chili pepper leaves that require large image datasets to train the model. However, some symptoms that appear on chili pepper leaves may be similar, resulting in discrepancies in the efficiency and accuracy of classification. In addition, the disease identification of chili peppers from leaves is limited to some diseases with distinctive features appearing on the leaves. It is often necessary to analyze symptoms in other parts of the chili pepper plant, such as roots, stems, leaves, and fruits. Therefore, this study used the information presented on plants throughout the chili pepper life cycle and considered the environmental conditions to help identify 14 occurring diseases. Besides, the Wrapper feature selection process proved to be more efficient than the IG, and GR, and the classification without feature selection.