Multilayer Perception-Based Hybrid Spectral Band Selection Algorithm for Aflatoxin B1 Detection Using Hyperspectral Imaging

Abstract

Aflatoxin B1 is a toxic substance in almonds, other nuts, and grains that poses potential serious health risks to humans and animals, particularly in warm, humid climates. Therefore, it is necessary to remove aflatoxin B1 before almonds enter the supply chain to ensure food safety. Hyperspectral imaging (HSI) is a rapid, non-destructive method for detecting aflatoxin B1 by analyzing specific spectral data. However, HSI increases data dimensionality and often includes irrelevant information, complicating the analysis process. These challenges make classification models for detecting aflatoxin B1 complex and less reliable, especially for real-time, in-line applications. This study proposed a novel hybrid spectral band selection algorithm to detect aflatoxin B1 in almonds based on multilayer perceptron (MLP) network weights and spectral refinement (W-SR). In the proposed process, the hyperspectral imaging (HSI) spectral rank was firstly generated based on MLP network weights. The rank was further updated using a spectral confidence matrix. Then, a spectral refinement process identified more important spectra from the lower-ranked ones through iterative processes. An exhaustive search was performed to select an optimal spectral subset, consisting of only the most significant spectral bands, to make the entire process suitable for real-time, in-line aflatoxin B1 detection in industrial environments. The experimental results using the artificially contaminated almonds dataset achieved a cross-validation accuracy of 98.67% with an F1-score of 0.982 for the standard normal variate (SNV) processed data with only four spectral bands. Comparative experiment results showed that the proposed MLPW-SR spectral band selection algorithm outperforms baseline methods.

1. Introduction

Grains and nuts are vulnerable to fungi infection such as Aspergillus flavus and A. parasiticus, which can create aflatoxin B1 in the secondary metabolism process [1]. Moreover, aflatoxin B1 contamination can occur in any step from harvesting to packaging. Also, aflatoxin B1 is highly carcinogenic and the continuous consumption of its contaminated food causes serious health problems including liver disease, kidney failures, etc., for both humans and animals [2]. With the changing global climate, the likelihood of aflatoxin contamination has significantly increased over the years. In response to this threat, various countries have implemented distinct regulatory measures aimed at mitigating potential health and economic losses [3].

Hyperspectral imaging stands out as a promising method for food quality estimation applications. Short wave near infrared (SWIR) hyperspectral imaging (HSI) covers spectral wavelengths from 900 to 2500 nm. This allows for the integration of spectral and spatial features [4,5]. Additionally, HSI captures detailed spectral information across hundreds of narrow absorption bands, allowing for the identification of subtle spectral differences indicative of contamination, even at very low concentrations [6]. This detailed information increases the dimensionality of HSI data. Moreover, these data contain both relevant and irrelevant information. Analyzing and classifying high-dimensional data requires substantial computational resources and developed models struggle to perform effectively due to the abundance of spectral data [7]. Furthermore, the large number of spectral data increases the model’s complexity.

Spectral band selection techniques play a crucial role in addressing the challenge of high dimensionality and enhancing the effectiveness of HSI-based detection of aflatoxin B1-contaminated almonds. These techniques aim to identify a subset of the most relevant and informative spectra from the original high-dimensional dataset [8]. Selecting only the most discriminative feature spectra helps to process faster by analyzing fewer spectral data, improving accuracy by focusing on key information, and enhancing interpretability through a simpler model. Techniques like filter-based methods (e.g., Relief, Information Gain) leverage data statistics to rank and select relevant spectra [9,10]. Wrapper-based methods employ learning algorithms to test different feature combinations and select the one leading to the best classification accuracy [7]. Embedded methods, like L1-regularized Support Vector Machines (SVMs), integrate feature selection within the learning process, favoring models with fewer features [11,12]. Additionally, deep learning-based methods, especially Convolutional Neural Networks (CNNs), show potential by implicit learning feature representations through their convolutional layers, potentially surpassing traditional methods in classification performance [13].

To classify hyperspectral images using relevant feature spectra, Li developed a hybrid spectral band selection algorithm named Genetic Algorithm (GA)–SVM [11]. This method combines a GA for spectral subset selection with an SVM classifier. The author used the GA to efficiently search important feature spectra in high-dimensional space, while the SVM leveraged these selected spectra for accurate image classification. Medjahed developed an Improved Grey Wolf Optimizer (IGWO) spectral band selection algorithm for HSI [14]. Unlike traditional feature selection methods that solely focus on maximizing classification accuracy, the developed algorithm incorporates a multi-objective function that additionally prioritizes spectral diversity.

An autoencoder is a special type of unsupervised neural network that encodes by finding key aspects of the provided data. Sun developed a Stochastic Gate-based Autoencoder (SGAE) spectral band selection algorithm for HSI [15]. The SGAE incorporated a stochastic gate layer that allowed the model to learn a subset of informative bands during training. Liu developed an information measure-based feature band selection algorithm using fuzzy rough set theory for hyperspectral image classification of soybean varieties [16]. The author used the Gaussian membership function to find the important feature spectra sets and the approach achieved an average accuracy of 99.11%. Jain developed a two-phase approach for hyperspectral image classification [17]. In the first phase, the author utilized a combination of SVM and self-organizing maps for spectral optimization. The optimized spectral set generated from this process was then used in the second phase for classification. This two-step approach enhances the precision of spectral classification by first refining the feature set and then applying a probabilistic method to differentiate between the regions of interest in the hyperspectral image. Zhao developed a reinforcement learning-based framework for hyperspectral spectral selection in soil organic matter prediction [18]. This framework addresses the limitations of existing methods by utilizing the Markov decision process and reinforcement learning agents to simulate the selection process and improve model performance.

Pal and Foody developed an SVM accuracy-based spectral band selection algorithm [19]. The SVM classification accuracy varied with spectra count, especially for small datasets. Islam developed a neighborhood mutual information-based minimum Redundancy Maximum Relevance (mRMR) spectral band selection algorithm [20]. Xie developed a Salp Swarm Algorithm for Feature Selection (SSAFS) for image-based plant disease detection [21]. SSAFS aimed to identify the optimal subset of features that maximizes classification accuracy while minimizing the number of features used. Islam further developed a hybrid feature selection algorithm, to improve HSI classification [22]. This method combined feature extraction and selection techniques. Zhang described a novel deep-learning based approach for HSI-based seed variety identification [23]. Wang developed a Modified Ant Lion Optimizer (MALO) spectral selection algorithm [24]. MALO combines the Ant Lion optimizer swarm intelligence algorithm with a wavelet support vector machine to address the combinatorial optimization problem of feature selection.

The multilayer perception (MLP) network can be used to find important feature spectra in a high-dimensional dataset [25,26]. While MLP can achieve higher accuracy, it often lacks interpretability [27]. Zhang used single-layer MLP to find the important features band [28]. The authors used network weight to find the feature’s importance score from the trained network. Romero and Sopena used an MLP network to train and classify various data and sequentially remove irrelevant features based on certain criteria [29]. The developed method removed feature spectra one by one and retrained the network, which increased computational cost, but the method led to a notable boost in performance. Kosarirad developed a hybrid feature band selection using the Grasshopper Optimization Algorithm (GOA) and MLP [30]. The authors found an optimal spectral set using GOA and evaluated the performance by classifying sonar data using the MLP network. Despite numerous feature spectra selection algorithms for HSI proposed in various research studies, none have consistently delivered optimal results across all applications. Moreover, many algorithms rely on a random search approach, which does not ensure the identification of the optimal feature spectra set. Classical feature selection algorithms evaluate the importance of individual features, which often fail to produce the optimal feature set for classification. Also, these methods often struggle to capture the high correlations and complex relationships inherent in spectral data, which can hinder optimal classification performance [31]. In contrast, we proposed an MLP-based feature spectra selection method that addresses these challenges by utilizing the multilayer perceptron’s ability to learn intricate interactions, thereby enhancing classification performance and making it suitable for real-time food safety applications.

This research aimed to develop an efficient and reliable HSI-based method for detecting aflatoxin B1-contaminated almonds to ensure food safety in the almond industry. Also, various feature selection techniques were investigated for hyperspectral image classification of aflatoxin B1-contaminated almonds for comparison.

2. Materials and Methods

2.1. Research Sample Collection and Preparation

The developed spectral band selection algorithm was extensively tested on three hyperspectral datasets. The datasets were made to develop aflatoxin b1 quantification and classification in almonds for industrial in-line application. All datasets contain both aflatoxin B1 contamination and contamination-free almond’s hyperspectral response. The contamination-free samples were considered zero aflatoxin B1 concentration. Almonds were artificially contaminated in four different aflatoxin B1 contamination levels in each dataset. Since the aflatoxin B1 contamination rate is low, the presence of highly contaminated almonds significantly impacts the test results [32]. Therefore, in this research, almonds were contaminated in 250, 500, 750, and 1000 ppb. In dataset 1, almonds of similar thickness and shape were used to obtain a consistent spectral response. The almond’s thickness was measured using a vernier caliper. Then all contaminated and contamination-free almonds were used to acquire hyperspectral images using an SWIR camera. The camera has an operating wavelength range between 900 and 1700 nm, divided into 224 spectral bands. The length of the acquired image was set to 200 pixels, while the width was set to 640 pixels. Consequently, the dimension of the acquired hyperspectral image cube was $640\times 200\times 224$. To reduce the data volume, a mean value was calculated for each spectral image and converted each hyperspectral image data cube into 224 spectral data points. The artificial aflatoxin B1-contaminated concentration was used as the ground truth. In this way, dataset 1 was created using 471 almond samples. Similarly, dataset 2 was made of 465 contaminated and non-contaminated almonds. In dataset 2, almonds were chosen by visual inspection with similar shape, thickness, and texture to create data diversity.

However, almonds naturally vary in shape, thickness, and texture. Therefore, dataset 3 was created, which consisted of almonds with different thicknesses, shapes, and sizes. These three datasets were utilized to assess the classification performance of the developed spectral band selection algorithm for detecting aflatoxin B1 contamination. The feature spectra were selected for both raw and standard normal variate (SNV) processed data. The mean spectral response is called raw data, which could be affected by the light scattering phenomenon. In the SNV, each spectrum is normalized by subtracting its mean and dividing by its standard deviation. This process effectively reduces the spectral scattering effect caused by variations in sample size or surface properties, while preserving the original spectral features and response. By minimizing unwanted variations, SNV enhances the signal quality without altering the essential information in the data. In comparison to the developed spectral band selection algorithm, all three datasets were tested on existing classical feature selection algorithms. The developed hybrid spectral band selection algorithm based on MLP network weights and spectral refinement (MLPW-SR) is described in the subsequent subsection.

2.2. Proposed Feature Selection Method

The proposed Algorithm 1 is based on the combination of the modified weight-based feature selection of MLP (MLPW) coupled with a refinement pass for accurate feature rank generation, and an exhaustive search. The proposed algorithm aims to select two sets of feature spectra. These sets can then be utilized to create an aflatoxin B1 classification system for industrial quality control and in-line applications. The algorithm can be divided into three main parts, MLP weight-based spectral rank generation, spectral refinement, and exhaustive search-based spectral set selection. At first, MLPW was used, focusing on the spectral confidence matrix and voting-based spectral rank position to select the topmost important spectra among the generated rankings. Therefore, a new confidence matrix ($C_{\mu }$) and confidence position matrix ($C_p$) were established, which were used to rank spectra and update the rank. The proposed algorithm iteratively removed the least important spectra and generated an important ranked matrix. In weight-based spectral band selection, randomly initializing weights may result in a misleading spectral rank matrix [33]. Therefore, we developed a spectral refinement approach that identifies relatively important spectral bands among the removed ones and adds them to the important spectra list. In the final phase, the most significant spectra T_e were selected to create small spectra sets for in-line aflatoxin B1 detection. An exhaustive feature selection approach was implemented in this phase, combined with an SVM performance evaluation algorithm to assess and refine the selected features.

The dataset ɸ_n is the collection of the hyperspectral responses of aflatoxin B1-contaminated and non-contaminated almonds of 224 spectra. The first and last several spectra of the SWIR hyperspectral image contain some noise to ramp up and down the imaging sensor [34]. Therefore, these feature spectra were not considered in the band selection process. The spectral data were normalized and partitioned for training, testing, and validation before being fed into the MLP classifier. The dataset was then trained, tested, and validated for ‘n’ times of the same MLP network with two hidden layers. To avoid the possible biases of feature position among the spectral dataset, feature spectra were randomized for the MLP model. This process increases the diversity of spectral data and increases the reliability of the output model’s weight. In the MLP classifier, a feature spectra selection layer was added between the input layer and the hidden layer, which is identical to the input layer. Multiple MLP classifiers were trained, tested, and validated using different learning parameters (η = 0.0001 to 0.1) to find an optimal learning rate, which was incorporated into the final models. In this experiment, a learning rate of 0.001 was used. The LeakyReLU activation function was applied in the feature spectra selection and hidden layers, while the SoftMax function was used in the output layer. In the first hidden layer, 256 nodes were utilized, while in the second layer, 128 nodes were employed, resulting in improved classification accuracy. To overcome the overfitting problem in the MLP models, L2 regularization was used to achieve optimal classification results [35]. The network was trained separately with a small number of randomized initial weights (±0.5) and bias value and backpropagation algorithms were used to update the weight in each node. The trained network provided an optimal weight matrix $\psi$ for higher cross-validation accuracy, such that $\psi ϵ\mathbb{R}$. This matrix is normalized and creates a normalized vector ${\mu }_i$ for each MLP model.

The ${\mu }_i$ matrix was used to create a spectral importance position matrix. The highest normalized weight archiving spectrum was the most important and obtained position 1, the second highest weight-containing spectra were position 2, and so on. Therefore, any spectra obtained a position from 1 to g, and g is the number of spectra in each MLP model. In this way, we obtained an $n\times \mathrm{g}$ normalized weight matrix, which was used to generate a weight confidence matrix $C_{\mu }$, defined as follows:

(1)$C_{\mu }={\sum }_{i=1}^n{(\psi }_i/max\left(\psi \right))$

The value of the weight confidence matrix varied between n and δ, and δ was the minimum weighted confidence value. The weight confidence value will be n when a feature spectrum maintains the highest weight for all MLP networks. So, the weight confidence value closer to n is a more important spectrum than other spectra. The primary spectral rank F is generated based on the weight confidence matrix.

The generated spectral rank might not be optimal, given that the MLP network utilized only two hidden layers and random initial weights. Therefore, the same MLP model selects a different spectral rank matrix [36]. Increasing the number of hidden layers improves classification performance by distributing the weights across more nodes. However, the weights assigned to the spectral band selection node may not accurately reflect the true importance of each spectral band. The confidence of the position matrix was used to update and obtain a stable optimal spectral rank. The confidence of the position matrix $C_p$ was generated based on the $n\times g$ important position contingency table. The contingency table was generated using a sorting weight matrix ${\mu }_i$. The contingency table ($C_t$) stored the position voting information of each feature spectra. The $C_p$ matrix measured the level of importance that each feature’s position holds within the spectral data and defined the confidence level associated with that position. If $\delta$ is a Kronecker delta function and l represents spectral rank position, then the equation can be written as follows:

(2)$C_p(l,j)={\sum }_{j=1}^g\delta (C_t\left(l,j\right),j)/n$

The position confidence value measures the stability of spectrums in the spectral rank. A confidence value of 1 indicates that a particular feature holds the same rank across all networks, reflecting complete stability and consistency in its importance. The final spectral rank $F$ was generated to update the primary spectral rank using the position confidence value.

Algorithm 1: selection of significant spectral bands
Input: dataset ɸn
Output: Two spectra set $F_{T_c}$ and $F_{f_x}$
1:	Initialize: $T_c$, $T_e$, $F_n$ spectra set, $T_f$ length of $F_n$, n number of MLP networks
2:	while $T_f$ less than $T_c$ do
3:		for i = 1 to n do
4:		Initialize: epoch, max_epoch, network weight $\psi$
5:			while epoch < max_epoch do
6:				Train, test, and validate network performances
7:				Find optimal weight ${\psi }_i$ for which validation accuracy maximum
8:				Compute normalized weight ${\mu }_i={\psi }_i/\mathrm{m}\mathrm{a}\mathrm{x}\left(\psi \right)$
9:			end
10:			Compute spectral position contingency table such that $C_t\left(i\right)$ = sorted ${\mu }_i$
11:		end
12:		Compute weight confidence matrix $C_{\mu }$ by summing spectral weight $\sum _{i=1}^n{\mu }_i$
13:		Compute spectral rank $F$ by sorting $C_{\mu }$
14:		Compute $C_p$ using Equation (2)
15:		for j = 1 to $T_f$ do
16:			if Equation (3) is true do
17:				Exchange spectral rank position $F\left(j\right)$ and $F(j-1)$
18:			end
19:		end
20:		Set $F_{top}=F(\mathrm{t}\mathrm{o}\mathrm{p} 80\% \mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{r}\mathrm{a})$ and $F_{bot}=F(\mathrm{b}\mathrm{o}\mathrm{t}\mathrm{t}\mathrm{o}\mathrm{m} 20\% \mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{r}\mathrm{a})$
21:		for k = 1 to the length of $F_{bot}$ do
22:			spectra_set = addition of spectra $F_{top}$ and $F_{bot}\left(k\right)$
23:			fit(k) = classification accuracy using SVM of spectra_set
24:			$ig\left(k\right)$ = calculate information gain of spectra $F_{bot}\left(k\right)$
25:		end
26:		Compute ${fit}_N=\mathrm{s}\mathrm{o}\mathrm{r}\mathrm{t}\left(fit/\mathrm{m}\mathrm{a}\mathrm{x}\left(fit\right)\right)$
27:		Compute ${ig}_N=\mathrm{s}\mathrm{o}\mathrm{r}\mathrm{t}\left(ig/\mathrm{m}\mathrm{a}\mathrm{x}\left(ig\right)\right)$
28:		for l =1 to the length of $F_{bot}$ do
29:			if ${fit}_N\left(l\right)$ greater than 0.5 and ${ig}_N\left(l\right)$ greater than 0.05 do
30:				Set spectra $F_{bot}\left(l\right)$ as significant and assign to $F_r$
31:			end
32:		end
33:		Updated spectra $F_n$ by concatenating $F_{top}$ and $F_r$
34:		if $T_f$ equals to $T_e$ do
35:			Generate all subsets of length $f_x$ such as $F_{f_x}\left(s\right)\in F_n$
36:			Compute classification accuracy for all sub spectra sets $F_{f_x}$ using SVM classifier
37:			Find best spectra set $F_{f_x}$ for online application
38:		end
39:		if $T_f$ equals to $T_c$ do
40:			Set $F_n$ as $F_{T_c}$ for quality control application
41:		end
42:	end

The spectral rank was updated based on the value of confidence weight and confidence position of two different spectra. At first, the confidence weight was checked for two spectra ith and (i + 1)th. If the confidence weight difference decreased by 5%, the confidence position value was then reassessed. So, the condition can be written as follows:

(3)${(\psi }_{F\left(j\right)}-{\psi }_{F(j-1)}) \ge 0.05 \ast {\psi }_{F\left(j\right)} \& C_p(F(j))<C_{p}F(j-1)$

The decreasing value was randomly chosen for the best fit. If the confidence position value of the ith feature spectra is less than that of the (i + 1)th spectra, then exchange the spectral position within the spectral rank matrix F. Otherwise, the spectral rank matrix remains the same. The spectrum set was selected in an iterative process. Then, from the spectral rank matrix, the lowest-ranked 20% of spectra (F_bot) were removed in each iteration and the top 80% of spectra were considered significant spectra (F_top). However, the removed feature spectra set F_bot may contain important spectra due to the random weight assignment for each MLP network.

Also, it was noticed that the removed spectra set contains some spectra whose correlated spectra were selected as important spectra. Therefore, spectral refinement was performed to find the potentially important spectra from F_bot. In spectral refinement, the important spectra were calculated by finding the SVM and information gain fitness score. Therefore, each spectrum from F_bot was trained and cross-validated, coupled with F_top, to calculate the fitness score. To make the spectral refinement process reliable, we considered only highly relevant spectra for which normalized fitness scores were greater than 0.5. Moreover, only highly relevant spectra were used to calculate information gain (IG). IG is widely used to find feature selection and the decision tree classifier [37]. In the proposed algorithm, IG was used to find important feature sets. In general, if the IG of any feature is higher than 0.05, then the feature is considered an important feature [38]. Therefore, the IG value of 0.05 was used to find the fitted features spectra set F_r. This way, the final important spectra were selected from F_bot. These potentially important spectra were made with an important spectrum set and considered as important spectra as F_n = F_top + F_r.

The above process iteratively selects a predefined number of significant spectral features. In online applications, the spectra set should be small, with optimal speed and accuracy. However, quality control applications need high accuracy with optimal speed and computational cost. Therefore, in the in-line application, the F_fx spectra set was selected and for the quality control application, the F_Tc spectra set was selected (F_Tc > F_fx). To select the F_Tc spectra set for quality control application, the iterative process continued until the T_c spectral band. However, to select F_fx spectra, F_Te top-ranked spectra were selected as mentioned above, and then an exhaustive search algorithm was performed. The exhaustive search algorithm finds all possible combinations of f_x spectrum features among T_e spectra. The SVM classification algorithm was used as a performance evaluation model and the top performing spectra set F_fx was selected as the most important spectra for in-line aflatoxin B1 detection.

2.3. Classical Feature Selection Algorithms

The developed spectral band selection algorithm was compared with existing classical techniques like Random Forest (RF) [39], IG [37], mRMR [40], Adaboost [41], Chi-Square (Chi2) [42], Analysis of Variance (ANOVA) [43], and ReliefF [44] to detect aflatoxin B1 in almonds for industrial in-line and quality control applications. RF is an ensemble machine learning algorithm that uses bagging techniques. It selects the optimal feature from a randomly chosen subset. RF finds the best splitting tree based on the subset data and features based on majority voting. It enhances overall classification performance. The model randomness can be increased by imposing a threshold for each feature, rather than pursuing the optimal threshold [39]. IG is also used to select relevant features in many applications. Decision tree classifiers use IG to split datasets into different groups for efficient classification [37]. Adaptive Boosting (AdaBoost) is a powerful binary ensemble learning technique that combines the prediction of several weak learners and makes a single strong learner achieve higher classification performance [41]. In AdaBoost feature selection, the importance of each feature is inferred from its frequency of use across multiple weak learners trained during the boosting process. Features that were consistently selected across different weak learners were considered more important, while those rarely chosen were deemed less relevant. The mRMR method prioritizes features that exhibit a high relevance to the target variable while simultaneously minimizing redundancy among the selected features [40].

The Chi2 assesses the statistical significance of the relationship between each feature and the target variable, and it identifies the most discriminative features for the training model [42]. Due to simplicity, computational efficiency, and robustness, Chi2 feature selection has been widely applied in diverse fields such as text classification, image recognition, and bioinformatics. The ANOVA compares means among multiple groups [43]. By partitioning the total variance into between-group and within-group components, ANOVA determines if the differences between group means are statistically significant. Renowned for its versatility and robustness, ANOVA is employed across diverse disciplines, including psychology, biology, economics, and engineering, to analyze experimental data and draw meaningful conclusions. The ReliefF assesses the relevance of features based on their ability to distinguish between instances [44]. It efficiently identifies the most informative features while minimizing the impact of noise.

3. Results and Discussion

The developed MLPW-SR spectral band selection and other existing feature selection algorithms were tested on three different hyperspectral imaging datasets. In this research work, the analysis has shown only 4 and 10 spectra sets. Four spectra were targeted to use for in-line application and 10 spectra were set for quality control application to meet industrial requirements. The performance was evaluated using classification accuracy and F1-score. Classification accuracy reflects the overall proportion of correct predictions by the model, whereas the F1-score balances precision and recall, making it particularly valuable for evaluating performance on imbalanced datasets.

3.1. Performance Evaluation of Binary Classification

The binary aflatoxin B1 classification accuracy and F1-score are tabulated in Table 1, Table 2 and Table 3 for the selected sets of 4 and 10 spectra using different feature selection methods for datasets 1, 2, and 3, respectively. The confusion matrix illustrating the binary classification accuracy of the support vector machine (SVM) using raw data from dataset 1 is shown in Figure 1. In this analysis, 4 spectra were selected for in-line detection of aflatoxin B1, while 10 spectra were utilized for quality control, where precise measurements are essential. The developed MLPW-SR algorithm selected 4 and 10 spectral bands for both raw and SNV-processed data. For raw data, the selected 4 spectra were 73, 132, 137, and 214, while the 10 spectra set included 44, 68, 97, 100, 115, 137, 158, 171, 181, and 214. For SNV-processed data, the 4 selected spectra were 98, 115, 133, and 167, and the 10 spectra set consisted of 15, 33, 44, 49, 50, 64, 98, 115, 133, and 206.

The SVM classification model achieved 4, 10, and full spectrum accuracy of 91.51%, 94.69%, and 93.63% for raw data and 98.67%, 95.74%, and 94.43% for SNV-processed data, respectively. Also, the F1-score was 0.784 and 0.870 for raw data and 0.982 and 0.920 for SNV data, respectively, indicating the classification models’ reliability. In raw data, 10 spectral bands achieved the highest classification accuracy compared to 4 and the full spectrum model. However, the 4 spectra set model outperformed the 10 and full spectra model for SNV-processed data. This is because the SNV removes the light-scattered effect from each spectrum; therefore, the 4 selected feature spectra become a more significant set than the 10 spectra set. Also, the 10 spectra may include some irrelevant bands, which reduces the classification accuracy. The SNV with 4 spectral bands achieved higher classification accuracy than the full spectrum model; this is because the full spectrum model contains both relevant and irrelevant spectra.

In contrast with other classical feature selection algorithms, the developed algorithm achieved better classification accuracy and F1-score. The tabulated classification parameters in Table 1 have shown that the developed MLPW-SR feature selection algorithm selected the best feature spectra set compared to the RF, IG, AdaBoost, mRMR, Chi2, ANOVA, and ReliefF algorithms, achieving a higher classification and F1-score. The classification accuracy of the selected feature spectra for different feature selection algorithms is also illustrated in Figure 2 for dataset 1.

For dataset 2, the developed MLPW-SR spectral band selection algorithm for 4, 10, and full spectra achieved 88.17%, 90.32%, and 93.82% for raw data and 95.16%, 89.25% and 90.59% for SNV-processed data, respectively. The algorithm-selected 4 and 10 spectra sets were 65, 141, 150, 208 and 4, 65, 125, 129, 138, 141, 143, 150, 170, 208 for raw data and 9, 91, 118, 171 and 41, 57, 64, 65, 91, 96, 106, 117, 120, 190 for SNV-processed data. The developed algorithm achieved comparable results for 4 and 10 spectra for both raw and pre-processed data. Compared with the other existing feature spectra selection algorithms, the developed algorithm achieved better results except in a few instances. The performance comparison of different feature spectra selection algorithms for dataset 2 is illustrated in Figure 3.

For raw data, among the different feature spectra selection algorithms, only the developed MLPW-SR algorithm achieved a higher than 88% classification accuracy for 4 spectra. Similarly, for 10 spectra, the algorithm achieved the highest classification accuracy of 90.32%. In comparison to SNV data, the MLPW-SR achieved the highest classification accuracy for 4 spectra compared to all different feature spectra selection algorithms, including the full spectrum accuracy. This is because MLPW-SR successfully finds the best 4 spectra set, which is also the most relevant set to classify aflatoxin B1. On the other hand, SNV with 10 spectra model achieved lower classification accuracy than the 4 spectra model. This could be because the addition of more spectra with the less relevant spectra made the spectra set less relevant. Therefore, the MLPW-SR algorithm achieved less classification accuracy than RF (90.32%), AdaBoost (91.94%), mRMR (94.62%), and ReliefF (91.94%). However, it achieved higher classification accuracy than IG (83.26%), Chi2 (87.63%), and ANOVA (88.98%).

For dataset 3, the classification accuracy for 4, 10, and full spectra achieved 85.24%, 87.50%, and 90.06% for raw data and 86.75%, 89.55%, and 87.05% for SNV-processed data, respectively. The developed MLPW-SR demonstrates superior aflatoxin B1 classification accuracy for 4 spectra for raw and SNV data, respectively, compared to other feature spectra selection algorithms. Notably, AdaBoost achieved an accuracy of 87.05% for SNV 4 spectra. Furthermore, MLPW-SR attained a higher classification accuracy for raw and SNV data of 10 spectra, surpassing other feature selection algorithms and aligning closely with the accuracy obtained using the full spectrum. The classification accuracy of selected spectra for raw and SNV-processed data for dataset 3 is visualized in Figure 4. Dataset 3 represents the actual variation in naturally found almond samples; therefore, the overall classification accuracy is less than other datasets. Classical feature selection methods like Chi2, IG, RF, AdaBoost, MRMR, ANOVA, and ReliefF have limitations in capturing complex, non-linear interactions between features. These methods often focus on individual feature ranking or subsets, potentially missing interactions between features that could enhance model performance [31]. For example, the Chi2 and ANOVA rely on statistical relevance, while IG struggles with redundancy [45]; in contrast, the proposed MLP-based feature spectra selection effectively captures complex, non-linear interactions, enhancing predictive accuracy. Based on the overall analysis, the proposed MLPW-SR algorithm can be used effectively to select spectral bands from hyperspectral image datasets to detect aflatoxin B1 in industrial real-time applications.

3.2. Performance Evaluation of Different Steps of the Proposed Method

The primary goal of spectral updates was to generate stable and reliable spectral ranks. The experimental results have shown that the inclusion of spectral updates helped to stabilize the feature spectra and increased the classification accuracy significantly. This is because a relatively significant spectrum in spectral rank may not be ranked as significant; however, the update process brings back this spectrum into a significant list. The spectral refinement finds the strong aflatoxin B1-correlated spectra using the SVM fitness score and IG score and finds those spectra significant that are found strong in both fitness scores. The experiment results have shown that spectral refinement improves the spectral rank and corresponding classification accuracy significantly. The comparison of aflatoxin B1 classification of different combinations of the developed model is illustrated in Figure 5. It is observed that the refinement pass with rank update requires higher iteration than other combinations to reach 4 spectra. However, at the initial stage, the spectral refinement does not improve the classification accuracy. This is due to a higher number of irrelevant spectra in the acquired hyperspectral images. In each iteration, the proposed algorithm finds the least significant spectra and removes them from the spectra set. Therefore, when the amount of aflatoxin B1-correlated spectra becomes dominant, the classification accuracy increases. So, the classification accuracy reached its highest peak when the spectra set contained the highest aflatoxin B1-sensitive spectral bands. Then, the reduction in the spectral band accounted for losses of relevant aflatoxin B1-sensitive information. The removal of the most significant spectra results in a reduction in information, making it more difficult to distinguish aflatoxin B1-contaminated almonds from uncontaminated ones. Conversely, the inclusion of relevant spectral bands improves model performance by capturing subtle variations in the data that are essential for accurate classification. Therefore, the refinement pass becomes crucial at a later stage when all selected spectra show some sensitivity to aflatoxin B1, and it helps retain more relevant spectra over less significant ones. For this phenomenon, the aflatoxin B1 classification accuracy decreases with a decrease in the number of significant spectral bands. A lower number of spectra increases the speed of aflatoxin B1 detection. Selecting the most significant spectra helps achieve optimal classification accuracy for industrial applications.

4. Conclusions

In this work, the hybrid spectral band selection algorithm MLPW-SR has been developed, which integrates multilayer perceptron network weights with feature refinement to classify aflatoxin B1 in almonds using hyperspectral images. The MLPW-SR algorithm effectively selects important spectral bands that can be utilized for detecting aflatoxin B1 contamination in almonds, making it suitable for industrial in-line and quality control applications. By leveraging advanced techniques for spectral selection, this method enhances detection accuracy and operational efficiency in food safety processes. The selected spectral bands of various algorithms were tested using a support vector machine (SVM) classifier. In comparison with other feature selection algorithms like RF, information grain, AdaBoost, mRMR, ANOVA, Chi-2, and ReliefF, the MLPW-SR achieved better classification performances for 4 spectra and 10 spectra models for three different artificially aflatoxin B1-contaminated almond’s hyperspectral images dataset. Datasets 1 and 2 consisted of similar thickness almonds and the MLPW-SR algorithm achieved higher accuracy for 4 and 10 spectra. However, almonds should have different thicknesses and sizes as in dataset 3. The developed spectral band selection algorithm selected significant spectra in dataset 3, which helped to achieve comparable accuracy with the full spectra model using 4 and 10 spectra. Therefore, the developed spectral band selection is well suited to finding optimal feature sets from the hyperspectral image dataset, and the feature spectra could be used effectively in the in-line and quality control aflatoxin B1 detection system.

Author Contributions

Conceptualization, M.A.K. and S.-H.L.; methodology, M.A.K. and S.-H.L.; software, M.A.K., I.L., and S.-H.L.; validation, M.A.K., I.L., and S.-H.L.; formal analysis, M.A.K., I.L., and S.-H.L.; investigation, M.A.K., I.L., C.B.S., and S.-H.L.; resources, M.A.K. and S.-H.L.; data curation, M.A.K. and S.-H.L.; writing—original draft preparation, M.A.K.; writing—review and editing, I.L., C.B.S., G.M., B.K.P., and S.-H.L.; visualization, M.A.K.; supervision, I.L., C.B.S., and S.-H.L. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

Further inquiries can be communicated to the corresponding author.

Acknowledgments

This research was supported by the Research Training Program International (RTPi) scholarship from Commonwealth Australia, which provided crucial funding and support throughout this study. Additionally, we are grateful for the top-up scholarship provided by SureNut Ltd., whose contribution significantly enhanced the resources available for this research. We would like to sincerely thank SureNut Ltd. for supplying all the almonds used in this study. Their generous provision of high-quality materials was essential for completing our experiments.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Binary SVM classification of aflatoxin B1-contaminated almonds using 4 spectra: (a) training, (b) testing, (c) cross-validation.

Figure 2. Illustration of binary aflatoxin B1 classification accuracy comparison for dataset 1. (a) Raw data, (b) SNV pre-processed data.

Figure 3. Illustration of binary aflatoxin B1 classification accuracy comparison for dataset 2. (a) Raw data, (b) SNV pre-processed data.

Figure 4. Illustration of binary aflatoxin B1 classification accuracy comparison for dataset 3. (a) Raw data, (b) pre-processed data.

Figure 5. Illustration of achieved binary classification accuracy of different phases of developed feature selection algorithm for three different datasets.

Table 1

Binary classification performance comparison of proposed and existing feature selection algorithm for dataset 1 (equal thickness, measured by vernier caliper).

Data Type	Band Selection Algorithm	Full Spectra	Classification Accuracy (%)
			4 Spectra				10 Spectra
		CV	Training	Testing	CV	F1-Score	Training	Testing	CV	F1-Score
Raw	RF	93.63	89.12	81.91	87.80	0.680	90.45	81.91	86.74	0.658
	IG		77.45	73.40	77.45	0	86.74	81.66	83.02	0.500
	AdaBoost		88.59	82.98	86.74	0.632	92.57	85.11	91.78	0.797
	mRMR		77.45	73.40	77.72	0.045	89.12	82.98	86.74	0.597
	Chi2		79.31	77.45	77.45	0	84.08	82.98	81.70	0.389
	ANOVA		77.45	77.45	77.45	0	77.45	77.45	77.45	0
	ReliefF		89.12	82.98	87.53	0.671	91.78	86.17	90.19	0.755
	Proposed		92.84	90.17	91.51	0.784	95.49	92.30	94.69	0.870
SNV	RF	94.43	87.80	80.85	87.53	0.684	96.55	86.17	94.96	0.886
	IG		79.58	74.34	77.45	0.023	81.17	75.53	79.58	0.206
	AdaBoost		86.47	80.60	87.30	0.652	97.61	90.43	93.90	0.861
	mRMR		96.02	88.30	94.43	0.874	97.61	92.55	94.69	0.884
	Chi2		85.68	77.66	85.15	0.588	96.82	88.30	93.90	0.866
	ANOVA		95.76	88.30	93.37	0.840	97.08	90.43	96.02	0.906
	ReliefF		90.98	84.98	90.98	0.764	94.96	90.43	93.90	0.848
	Proposed		99.20	96.81	98.67	0.982	99.73	95.74	96.55	0.920

Table 2

Binary classification performance comparison of proposed and existing feature selection algorithm for dataset 2 (equal thickness, measured by visual inspection).

Data Type	Band Selection Algorithm	Full Spectra	Classification Accuracy (%)
			4 Spectra				10 Spectra
		CV	Training	Testing	CV	F1-Score	Training	Testing	CV	F1-Score
Raw	RF	93.82	81.72	78.49	81.72	0.752	90.05	89.25	86.56	0.824
	IG		80.91	77.27	79.84	0.730	80.38	75.27	80.11	0.724
	AdaBoost		85.48	88.17	84.14	0.790	91.40	89.25	89.63	0.860
	mRMR		84.41	83.32	82.80	0.776	84.14	87.10	83.06	0.779
	Chi2		81.18	78.49	81.45	0.749	82.80	82.80	83.60	0.778
	ANOVA		81.45	77.42	81.45	0.749	84.14	83.87	83.60	0.778
	ReliefF		65.86	64.52	66.94	0.400	70.70	66.67	70.16	0.484
	Proposed		90.05	92.47	88.17	0.846	91.40	92.47	90.32	0.878
SNV	RF	90.59	90.86	92.47	90.86	0.883	92.47	93.55	90.32	0.875
	IG		83.60	86.02	82.80	0.770	87.10	86.02	83.26	0.773
	AdaBoost		92.74	93.55	90.86	0.880	95.97	89.25	91.94	0.899
	mRMR		92.47	91.40	90.32	0.872	96.77	89.25	94.62	0.932
	Chi2		88.44	91.40	87.37	0.832	88.71	91.40	87.63	0.832
	ANOVA		88.44	91.40	87.37	0.832	91.40	91.40	88.98	0.857
	ReliefF		81.72	80.65	77.96	0.687	93.55	89.25	91.94	0.897
	Proposed		95.43	94.62	95.16	0.938	91.67	89.25	89.25	0.862

Table 3

Binary classification performance comparison of proposed and existing feature selection algorithm for dataset 3 (different thickness).

Data Type	Band Selection Algorithm	Full Spectra	Classification Accuracy (%)
			4 Spectra				10 Spectra
		CV	Training	Testing	CV	F1-Score	Training	Testing	CV	F1-Score
Raw	RF	90.06	84.79	78.92	84.49	0.398	86.30	82.53	85.09	0.476
	IG		84.04	78.31	83.28	0.335	84.19	77.71	83.73	0.365
	AdaBoost		85.24	80.72	84.64	0.446	85.69	80.72	85.24	0.500
	mRMR		84.79	77.71	84.34	0.435	86.30	81.33	85.84	0.478
	Chi2		83.28	78.92	82.08	0.261	84.19	78.92	83.73	0.333
	ANOVA		83.28	79.52	82.38	0.273	84.64	80.72	84.04	0.377
	ReliefF		83.28	80.12	82.68	0.311	8509	79.31	83.89	0.382
	Proposed		85.84	83.33	85.24	0.467	88.10	84.72	87.50	0.574
SNV	RF	87.05	83.89	76.51	83.13	0.364	90.81	84.34	88.86	0.661
	IG		82.68	77.71	81.33	0.101	84.49	78.31	83.28	0.343
	AdaBoost		87.80	80.72	87.05	0.596	90.51	84.34	87.35	0.611
	mRMR		85.09	79.51	84.19	0.432	88.10	83.31	84.64	0.519
	Chi2		83.89	76.51	83.13	0.364	87.20	82.12	86.14	0.545
	ANOVA		83.89	76.51	83.13	0.364	87.05	77.11	84.94	0.485
	ReliefF		83.89	76.51	83.13	0.364	88.86	82.53	86.45	0.563
	Proposed		87.95	85.13	86.75	0.581	91.87	88.34	89.55	0.685

References

1. Kabir, M.A.; Lee, I.; Mishra, G.; Singh, C.B.; Panda, B.K.; Lee, S.H. Quantifying Aflatoxin B1 Contamination Levels in Almonds Using Hyperspectral Imaging Utilizing Gaussian Process and Support Vector Regression. Proceedings of the 2024 IEEE Symposium on Wireless Technology & Applications (ISWTA); Kuala Lumpur, Malaysia, 20–21 July 2024.

2. Daou, R.; Hoteit, M.; Bookari, K.; Joubrane, K.; Khabbaz, L.R.; Ismail, A.; Maroun, R.G.; Khoury, A.E. Public health risk associated with the co-occurrence of aflatoxin B1 and ochratoxin A in spices, herbs, and nuts in Lebanon. Front. Public Health; 2023; 10, 1072727. [DOI: https://dx.doi.org/10.3389/fpubh.2022.1072727] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36699892]

3. Mishra, G.; Panda, B.K.; Ramirez, W.A.; Jung, H.; Singh, C.B.; Lee, S.H.; Lee, I. Application of SWIR hyperspectral imaging coupled with chemometrics for rapid and non-destructive prediction of Aflatoxin B1 in single kernel almonds. LWT; 2022; 155, 112954. [DOI: https://dx.doi.org/10.1016/j.lwt.2021.112954]

4. Feng, Y.; Zhu, J.; Song, R.; Wang, X. S2EFT: Spectral-Spatial-Elevation Fusion Transformer for hyperspectral image and LiDAR classification. Knowl.-Based Syst.; 2024; 283, 111190. [DOI: https://dx.doi.org/10.1016/j.knosys.2023.111190]

5. Yang, J.; Wu, C.; You, T.; Wang, D.; Li, Y.; Shang, C.; Shen, Q. Hierarchical spatio-spectral fusion for hyperspectral image super resolution via sparse representation and pre-trained deep model. Knowl.-Based Syst.; 2023; 260, 110170. [DOI: https://dx.doi.org/10.1016/j.knosys.2022.110170]

6. Wang, X.; Tan, K.; Du, P.; Han, B.; Ding, J. A capsule-vectored neural network for hyperspectral image classification. Knowl.-Based Syst.; 2023; 268, 110482. [DOI: https://dx.doi.org/10.1016/j.knosys.2023.110482]

7. Chaudhuri, A. Search space division method for wrapper feature selection on high-dimensional data classification. Knowl.-Based Syst.; 2024; 291, 111578. [DOI: https://dx.doi.org/10.1016/j.knosys.2024.111578]

8. Kabir, M.A.; Mishra, G.; Panda, B.K.; Singh, C.B.; Lee, I.; Lee, S.-H. Classification of Aflatoxin B1 Contamination Level Using Hyperspectral Images with Random Forest and QDA. Proceedings of the 2024 IEEE International Conference on Big Data, Cloud Computing, and Data Science (BCD); Kitakyushu, Japan, 16–18 July 2024.

9. Gao, J.; Wang, Z.; Jin, T.; Cheng, J.; Lei, Z.; Gao, S. Information gain ratio-based subfeature grouping empowers particle swarm optimization for feature selection. Knowl.-Based Syst.; 2024; 286, 111380. [DOI: https://dx.doi.org/10.1016/j.knosys.2024.111380]

10. Zhang, B.; Li, Y.; Chai, Z. A novel random multi-subspace based ReliefF for feature selection. Knowl.-Based Syst.; 2022; 252, 109400. [DOI: https://dx.doi.org/10.1016/j.knosys.2022.109400]

11. Li, S.; Wu, H.; Wan, D.; Zhu, J. An effective feature selection method for hyperspectral image classification based on genetic algorithm and support vector machine. Knowl.-Based Syst.; 2011; 24, pp. 40-48. [DOI: https://dx.doi.org/10.1016/j.knosys.2010.07.003]

12. Wang, C.; Gu, Z.; Wei, J.-M. Spectral clustering and embedding with inter-class topology-preserving. Knowl.-Based Syst.; 2024; 284, 111278. [DOI: https://dx.doi.org/10.1016/j.knosys.2023.111278]

13. Paul, D.; Bardhan, S.; Saha, S.; Mathew, J. ML-KnockoffGAN: Deep online feature selection for multi-label learning. Knowl.-Based Syst.; 2023; 271, 110548. [DOI: https://dx.doi.org/10.1016/j.knosys.2023.110548]

14. Medjahed, S.; Saadi, T.A.; Benyettou, A.; Ouali, M. Gray wolf optimizer for hyperspectral band selection. Appl. Soft Comput.; 2016; 40, pp. 178-186. [DOI: https://dx.doi.org/10.1016/j.asoc.2015.09.045]

15. Sun, H.; Zhang, L.; Wang, L.; Huang, H. Stochastic gate-based autoencoder for unsupervised hyperspectral band selection. Pattern Recognit.; 2022; 132, 108969. [DOI: https://dx.doi.org/10.1016/j.patcog.2022.108969]

16. Liu, Y.; Wu, T.; Yang, J.; Tan, K.; Wang, S. Hyperspectral band selection for soybean classification based on information measure in FRS theory. Biosyst. Eng.; 2019; 178, pp. 219-232. [DOI: https://dx.doi.org/10.1016/j.biosystemseng.2018.12.002]

17. Jain, D.K.; Dubey, S.B.; Choubey, R.K.; Sinhal, A.; Arjaria, S.K.; Jain, A.; Wang, H. An approach for hyperspectral image classification by optimizing SVM using self organizing map. J. Comput. Sci.; 2018; 25, pp. 252-259. [DOI: https://dx.doi.org/10.1016/j.jocs.2017.07.016]

18. Zhao, L.; Tan, K.; Wang, X.; Ding, J.; Liu, Z.; Ma, H.; Han, B. Hyperspectral feature selection for SOM prediction using deep reinforcement learning and multiple subset evaluation strategies. Remote Sens.; 2023; 15, 127. [DOI: https://dx.doi.org/10.3390/rs15010127]

19. Pal, M.; Foody, G.M. Feature selection for classification of hyperspectral data by SVM. IEEE Trans. Geosci. Remote Sens.; 2010; 48, pp. 2297-2307. [DOI: https://dx.doi.org/10.1109/TGRS.2009.2039484]

20. Islam, M.R.; Ahmed, B.; Hossain, M.A.; Uddin, M.P. Mutual information-driven feature reduction for hyperspectral image classification. Sensors; 2023; 23, 657. [DOI: https://dx.doi.org/10.3390/s23020657]

21. Xie, X.; Xia, F.; Wu, Y.; Liu, S.; Yan, K.; Xu, H.; Ji, Z. A novel feature selection strategy based on salp swarm algorithm for plant disease detection. Plant Phenomics; 2023; 5, 0039. [DOI: https://dx.doi.org/10.34133/plantphenomics.0039]

22. Islam, M.R.; Islam, M.T.; Uddin, M.P. Improving hyperspectral image classification through spectral-spatial feature reduction with a hybrid approach and deep learning. J. Spat. Sci.; 2024; 69, pp. 349-366. [DOI: https://dx.doi.org/10.1080/14498596.2023.2227948]

23. Zhang, L.; Wei, Y.; Liu, J.; Wu, J.; An, D. A hyperspectral band selection method based on sparse band attention network for maize seed variety identification. Expert Syst. Appl.; 2024; 238, 122273. [DOI: https://dx.doi.org/10.1016/j.eswa.2023.122273]

24. Wang, M.; Wu, C.; Wang, L.; Xiang, D.; Huang, X. A feature selection approach for hyperspectral image based on modified ant lion optimizer. Knowl.-Based Syst.; 2019; 168, pp. 39-48. [DOI: https://dx.doi.org/10.1016/j.knosys.2018.12.031]

25. Garrido, C.; de Oña, R.; de Oña, J. Neural networks for analyzing service quality in public transportation. Expert Syst. Appl.; 2014; 41, pp. 6830-6838. [DOI: https://dx.doi.org/10.1016/j.eswa.2014.04.045]

26. Paliwal, M.; Kumar, U.A. Assessing the contribution of variables in feed forward neural network. Appl. Soft Comput.; 2011; 11, pp. 3690-3696. [DOI: https://dx.doi.org/10.1016/j.asoc.2011.01.040]

27. Luíza da Costa, N.; de Lima, M.D.; Barbosa, R. Evaluation of feature selection methods based on artificial neural network weights. Expert Syst. Appl.; 2021; 168, 114312. [DOI: https://dx.doi.org/10.1016/j.eswa.2020.114312]

28. Zhang, W.; Shen, X.; Zhang, H.; Yin, Z.; Sun, J.; Zhang, X.; Zou, L. Feature importance measure of a multilayer perceptron based on the presingle-connection layer. Knowl. Inf. Syst.; 2024; 66, pp. 511-533. [DOI: https://dx.doi.org/10.1007/s10115-023-01959-7]

29. Romero, E.; Sopena, J.M. Performing feature selection with multilayer perceptrons. IEEE Trans. Neural Netw.; 2008; 19, pp. 431-441. [DOI: https://dx.doi.org/10.1109/TNN.2007.909535]

30. Kosarirad, H.; Nejati, M.G.; Saffari, A.; Khishe, M.; Mohammadi, M. Feature selection and training multilayer perceptron neural networks using Grasshopper optimization algorithm for design optimal classifier of big data sonar. J. Sens.; 2022; 2022, 9620555. [DOI: https://dx.doi.org/10.1155/2022/9620555]

31. Li, J.; Cheng, K.; Wang, S.; Morstatter, F.; Trevino, R.P.; Tang, J.; Liu, H. Feature selection: A data perspective. ACM Comput. Surv. (CSUR); 2017; 50, pp. 1-45. [DOI: https://dx.doi.org/10.1145/3136625]

32. Whitaker, T.B.; Slate, A.; Birmingham, T.; Adams, J.; Jacobs, M.; Gray, G. Correlation Between Aflatoxin Contamination and Various USDA Grade Categories of Shelled Almonds. J. AOAC Int.; 2010; 93, pp. 943-947. [DOI: https://dx.doi.org/10.1093/jaoac/93.3.943]

33. Pentoś, K. The methods of extracting the contribution of variables in artificial neural network models–Comparison of inherent instability. Comput. Electron. Agric.; 2016; 127, pp. 141-146. [DOI: https://dx.doi.org/10.1016/j.compag.2016.06.010]

34. Mishra, P.; Karami, A.; Nordon, A.; Rutledge, D.N.; Roger, J.-M. Automatic de-noising of close-range hyperspectral images with a wavelength-specific shearlet-based image noise reduction method. Sens. Actuators B Chem.; 2019; 281, pp. 1034-1044. [DOI: https://dx.doi.org/10.1016/j.snb.2018.11.034]

35. Phaisangittisagul, E. An analysis of the regularization between L2 and dropout in single hidden layer neural network. Proceedings of the 2016 7th International Conference on Intelligent Systems, Modelling and Simulation (ISMS); Bangkok, Thailand, 25–27 January 2016.

36. Cao, M.; Qiao, P. Neural network committee-based sensitivity analysis strategy for geotechnical engineering problems. Neural Comput. Appl.; 2008; 17, pp. 509-519. [DOI: https://dx.doi.org/10.1007/s00521-007-0143-5]

37. Qu, K.; Xu, J.; Hou, Q.; Qu, K.; Sun, Y. Feature selection using Information Gain and decision information in neighborhood decision system. Appl. Soft Comput.; 2023; 136, 110100. [DOI: https://dx.doi.org/10.1016/j.asoc.2023.110100]

38. Prasetiyowati, M.I.; Maulidevi, N.U.; Surendro, K. Determining threshold value on information gain feature selection to increase speed and prediction accuracy of random forest. J. Big Data; 2021; 8, 84. [DOI: https://dx.doi.org/10.1186/s40537-021-00472-4]

39. Breiman, L. Random Forests. Mach. Learn.; 2001; 45, pp. 5-32. [DOI: https://dx.doi.org/10.1023/A:1010933404324]

40. Zhao, Z.; Anand, R.; Wang, M. Maximum relevance and minimum redundancy feature selection methods for a marketing machine learning platform. Proceedings of the 2019 IEEE International Conference on Data Science and Advanced Analytics (DSAA); Washington, DC, USA, 5–8 October 2019.

41. Wang, R. AdaBoost for Feature Selection, Classification and Its Relation with SVM, A Review. Phys. Procedia; 2012; 25, pp. 800-807. [DOI: https://dx.doi.org/10.1016/j.phpro.2012.03.160]

42. McHugh, M.L. The chi-square test of independence. Biochem. Medica; 2013; 23, pp. 143-149. [DOI: https://dx.doi.org/10.11613/BM.2013.018]

43. Fraiman, D.; Fraiman, R. An ANOVA approach for statistical comparisons of brain networks. Sci. Rep.; 2018; 8, 4746. [DOI: https://dx.doi.org/10.1038/s41598-018-23152-5]

44. Urbanowicz, R.J.; Meeker, M.; La Cava, W.; Olson, R.S.; Moore, J.H. Relief-based feature selection: Introduction and review. J. Biomed. Inform.; 2018; 85, pp. 189-203. [DOI: https://dx.doi.org/10.1016/j.jbi.2018.07.014]

45. Guyon, I.; Elisseeff, A. An introduction to variable and feature selection. J. Mach. Learn. Res.; 2003; 3, pp. 1157-1182.