Related works

Nowadays, machine learning (ML) is used in developing adaptive intelligent systems that can perform complex tasks that are beyond human abilities. Some of the areas of applications of ML algorithms include pattern recognition, image processing, natural language processing, and medical diagnostics, to mention just a few [8]. Different types of ML algorithms exist, in the context of how they are applied to the field of adulteration detection. There are some technologies available to classify honey, but most of these methods are destructive, and too costly in terms of manpower, money, and time [9]. Here, destruction means that the conventional technologies would damage or compromise the honey samples. This study focuses on improving the detection of honey adulteration using machine learning by comparing the popular algorithms. In this section, the review of related research works is discussed as follows.

Oroian et al. [10] studied honey adulteration with fructose, glucose, and hydrolyzed inulin syrup and reported that it influenced some physicochemical properties such as pH, electrical conductivity, and water activity. Guler et al. [11] investigated changes in viscosity for adulterated honey and reported an increase in viscosity with a sugar syrup concentration increase. Several methods were used to evaluate direct adulteration in honey. Kelly et al. [12] reported the use of near-infrared transfectant spectroscopy to detect Irish honey adulteration by high fructose corn syrup and beet inverted syrup. Gallardo-Velazquez et al. [13] investigated the use of mid-infrared Fourier transform spectroscopy to quantify the content of honey adulterants including HFCS, corn syrup, and inverted sugar. RuizMatute et al. [14] reported the use of GC-MS for the detection of honey adulteration with high fructose Inulin syrups. Liquid chromatography (LC) and gas chromatography (GC) have been used simultaneously to detect exogenous sugars in honey by appropriate fingerprints of adulteration. Kumaravelu and Gopal [15] reported the use of near-infrared spectroscopy and partial least square regression for the detection and quantification of four honey types adulteration by jaggary.

Siddiqui et al. [16] present A thorough analysis of honey adulteration methods for the years 2000–2016. According to the authors, NMR spectroscopy is a potent technique for identifying genuine honey and detecting adulteration by different types of sugar.

To confirm the botanical provenance and authenticity of honey, many studies present recent developments in data mining, machine learning, and multivariate data analysis techniques. Numerous analytical techniques are covered, such as those by Wu et al. [17], Cuevas-Glory et al. [18], and Trifković et al. [19]. In addition, the study determines which parameters are most commonly used in conjunction with these methods to determine the location and floral origin of honey.

The research stated in [20] aims to identify various types of milk adulterants, detection methods, and health hazards associated with milk product adulteration. The proposed project investigates the fractional-order element-based impedance sensor for its potential use in detecting milk adulteration. The low-cost, user-friendly instrumentation system is expected to be commercialized soon. An automated sensing system for synthetic milk detection, based on a microcontroller, has been created to reduce reliance on specialized labor and improve efficiency. The dipole layer capacitance at the interface of the impedance sensor and contaminated milk is considered during the modeling process. An electrical equivalent circuit is built, and the detection of milk adulteration is classified using Recurrent Neural Networks. The proposed work is estimated to have an accuracy rate of 92.31 percent, a sensitivity rate of 75.23 percent, and a specificity rate of 90.12 percent, all higher than the current rate [20].

The recent work stated in [21] proposed an efficient and nondestructive method for detecting fruit freshness using the machine learning algorithm convolutional neural network (CNN). Traditional methods, such as instrumentation, testing reagents, or manual labor, are time-consuming and labor-intensive. Fruit, a high-value food, is susceptible to spoilage during packaging, transportation, and sales. The paper shows that CNNs have good performance in identifying fruit freshness and discusses the overfitting of machine learning based on experimental results.

The majority of the studies aim for the determination of the botanical origin of honey, with only a few aiming for geographical origin. A possible reason for the extra effort by researchers to ascertain the botanical origin is that it causes a greater impact on the composition and properties of honey.

Several studies have used Fourier transform infrared spectroscopy to detect adulteration in honey with cane sugar. The accuracy of predicting sugar concentration was 93.75% for one type of honey but below 80% for three different honey types. This shows that spectroscopy can be used with machine learning techniques to predict adulteration in honey, but the ability to predict sugar concentration across a range of honey must be improved. Hyperspectral imaging is a promising tool for food quality assurance, allowing spatial information to be used alongside spectral information. It has been used for various food quality applications, including determining different properties of grapes and predicting the properties of chocolate. The problem needs further investigation by evaluating the effect of feeding bees at different sugar concentrations and evaluating the resulting physiochemical properties of the honey. It is also necessary to develop a new reliable and cost-effective method for detecting adulteration in honey. Therefore, the objective of this study is to use the outperformed algorithm from the comparative analysis among K-nearest neighbors, Support vector machine, and random forest. Decision tree and ANNs to classify and predict. Compared to existing approaches, our proposed model relies on hyper-spectrum images. Moreover, our approach shows the discovery to overcome generalization to unknown honey types problems and improve the accuracy of existing models using hyperparameter tuning with enhanced performance.

Dataset

The dataset used in this study consists of segmented and pre-processed hyperspectral images of adulterated honey samples. Hyperspectral imaging is a non-destructive analytical technique that combines conventional imaging with spectroscopy to capture both spatial and spectral information from an object. In the context of honey adulteration detection, hyperspectral imaging allows for the acquisition of detailed spectral data across a range of wavelengths, enabling the identification of adulterants based on their unique spectral signatures.

Overall, the dataset comprises 12 different honey products from seven different brands with 12 different botanical origins labels as shown in Fig 2. Half of the samples are types of Manuka honey which is a premium NZ honey type, and the other half are various types of other NZ honey. ManukaUMF5 refers to a specific type or grade of Manuka honey with a UMF (Unique Manuka Factor) rating of 5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Class Distribution.

https://doi.org/10.1371/journal.pdig.0000536.g002

UMF is a grading system used to measure the potency and quality of Manuka honey, with higher UMF ratings indicating higher levels of beneficial compounds.

Table 1 shows the makeup of the dataset from these different kinds of honey. In creating the dataset, we sampled and captured images of all the honey at each sugar concentration; however, some mixtures were of low quality, and we could not include the images in the final dataset. The full dataset of honey samples contains 8675 total instances.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The top 10 Classes with the most entries in the dataset.

https://doi.org/10.1371/journal.pdig.0000536.t001

Each instance is the result of spatial segmentation of the hyperspectral imagery. It contains 128 features ‐ representing the spectral wavelengths of the hyperspectral camera.

Table 1 represents the number of training and testing examples available. Each sample contains 25 examples following segmentation.

Fig 3 shows the brand distribution along counts of the dataset. As the inspection graph indicates, the brand distribution has variety. Some brand has more count scores while some other brands have less.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of Brand.

https://doi.org/10.1371/journal.pdig.0000536.g003

Data pre-processing

Before applying machine learning algorithms, the dataset undergoes pre-processing steps to enhance its quality and suitability for analysis. This includes normalization of spectral data to ensure consistent scale across features, segmentation of hyperspectral images into training examples, and removal of low-quality samples. The pre-processing aims to standardize the data and improve the performance of subsequent machine learning models.

The pre-processing step follows the segmentation and uses a simple normalization approach, where the mean of the spectra is forced to be at 0, and the standard deviation to be 1. Other normalization approaches were also considered; however, this was experimentally found to be the best approach for this dataset.

The hyperspectral imaging system first calibrates using a dynamic white reference technique, and then it performs preprocessing segmentation. Using this technique, each hyperspectral image is divided into a five-by-five grid, yielding twenty-five training examples. The segments are averaged to represent the sample of honey as a whole, and the final data set is normalized before machine learning techniques are applied.

Outlier detection is a crucial step in identifying parameters affected by outlier tools from thousands of parameters. It involves separating log files based on recipe and tool numbers, processing them to calculate means and standard deviations using MapReduce, and performing outlier detection. An outlier detection technique (ODT) is used to detect anomalous observations that do not fit the typical statistical distribution of a dataset. In this study, simple methods that use statistical tools like boxplots have been employed to represent the distributions of samples corresponding to each feature. Fig 4 likely presents boxplots for each feature in the dataset, providing a visual representation of the distribution of values and identifying potential outliers. Fig 4 provides valuable insights into the distribution of features in the dataset and highlights potential outliers that may require further investigation. By addressing outliers and understanding the characteristics of each feature, researchers can enhance the effectiveness of their machine-learning algorithms for detecting honey adulteration. Understanding the distribution of features is essential for developing accurate machine-learning models for honey adulteration detection. Identifying outliers helps ensure that the models are trained on high-quality data, leading to more reliable predictions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Outlier detection.

https://doi.org/10.1371/journal.pdig.0000536.g004

Data pruning

We classify the botanical origins of honey using the most promising methods currently available, and we also take into account the theory of adulteration from related applications. By lowering the problem’s complexity, feature reduction can lessen overfitting and enhance performance. Principal component analysis (PCA), feature ranking-based feature selection of the k best features, and autoencoder techniques are the feature reduction techniques employed. One common technique for extracting and reducing features is PCA. A set of orthogonal vectors is formed when the features are combined linearly. This approach retains only the features with high variance.

Feature extraction

Feature extraction involves identifying and selecting relevant information from the dataset to represent each honey sample. In this study, spectral features extracted from hyperspectral images serve as input variables for the machine learning models. Additionally, attributes such as brand, acquisition, and concentration class are considered for classification and prediction tasks.

The primary characteristics of hyperspectral images are the camera’s operating wavelengths, which are utilized to analyze honey. To test the algorithms’ generalization ability and divide the dataset into training and testing sets, additional attributes are taken into account. All honey types within a brand can be categorized using the ’Brand’ attribute, which represents the honey’s manufacturer. The ’Acquisition’ attribute denotes the various image samplings for identical honey types and brands. Six samples are taken, numbered one through six, for every distinct honey jar. As a result, instances from a segment that is part of the training set are avoided and a balanced distribution of all honey types is obtained. The botanical class of honey is indicated by the ’Class’ attribute.

Hyper parameters tuning

The model parameters are enhanced or tuned by the training process. We run data through the operations of the model, compare the resulting prediction with the actual value for each data instance, evaluate the accuracy, and adjust until to get the best values. Hyperparameters are tuned by running the whole training data to look at the aggregate accuracy and adjust. Dataset pruning is the process of removing sub-optimal tuples from a dataset to improve the learning of a machine learning model. The idea of pruning is to consider a subset of hyperparameter configuration space to avoid unnecessary functions or attributes of data.

Prediction model

Various machine learning algorithms are employed for classification and prediction tasks, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), K-nearest neighbors (KNN), Random Forests (RF), and Decision Trees (DT). Each algorithm is trained on the pre-processed dataset using appropriate hyperparameters to optimize performance. The modeling process consists of selecting models that are based on various machine learning techniques used in the experimentation. The model selection process for detecting honey adulteration using hyperspectral imaging data involves choosing machine learning algorithms based on their suitability for handling high-dimensional data, learning complex patterns, and achieving accurate classification. The selected algorithms include Artificial Neural Networks (ANN), Support Vector Machines (SVM), K-nearest neighbors (KNN), Random Forests (RF), and Decision Trees (DT).

Artificial Neural Networks (ANN): Chosen for its ability to learn complex patterns from large datasets, making it suitable for image classification tasks. It learns to distinguish between genuine and adulterated honey samples based on their spectral signatures extracted from hyperspectral images.

Support Vector Machines (SVM): Selected for its capability to separate classes by finding the optimal hyperplane with maximum margin. It’s effective for binary classification tasks and classifies honey samples as pure or adulterated based on their spectral features.

K-Nearest Neighbors (KNN): Chosen for its simplicity and effectiveness in handling non-linear relationships in data. It assigns the majority class among the nearest neighbors of each sample in the feature space, allowing for accurate classification of honey samples.

Random Forests (RF): Selected for its robustness and ability to handle noisy and high-dimensional data. It combines multiple decision trees to improve classification accuracy by learning from different subsets of features.

Decision Trees (DT): Chosen for their simplicity and interpretability, making them suitable for understanding the decision-making process behind classification tasks. They partition the feature space based on spectral characteristics to accurately classify honey samples.

Overall, these machine learning algorithms are applied to develop accurate and robust classification models for detecting honey adulteration, ensuring food safety and quality. The goal is to identify the best classifier for the analyzed problem. Each classifier must therefore be trained on the featured set and the classifier with the best classification results is used for prediction. The details of the algorithms used in this study are discussed as follows.

Artificial Neural Network (ANN)

Artificial neural networks (ANNs) have gained popularity for classification, clustering, pattern recognition, and prediction in various disciplines, including artificial intelligence, information security, big data, cloud computing, internet, and forensic science [22]. ANNs’ full applications can be evaluated based on factors such as accuracy, processing speed, latency, performance, fault tolerance, volume, scalability, and convergence. ANNs are particularly useful for image recognition and natural language processing due to their self-learning, adaptivity, fault tolerance, and advancement in input-to-output mapping. ANNs are effective in handling complex and non-complex problems in various spheres of life, including agriculture, science, medical science, education, finance, management, security, engineering, trading commodities, and art [23]. However, there is a growing need to adopt a systematic approach in the development phase to improve performance, such as addressing factors such as data set choice, data accuracy, data instrument, data standardization, data inputs, data division, preprocessing, validations, processing, and output techniques [24].

In this study, Artificial Neural Networks (ANNs) were employed as a machine learning technique to classify and predict adulteration in honey samples using hyperspectral imaging data. ANNs are a type of deep learning algorithm inspired by the structure and function of the human brain’s neural networks. They consist of interconnected nodes organized in layers: an input layer, one or more hidden layers, and an output layer.

For this study, ANNs were trained on segmented and pre-processed hyperspectral images of adulterated honey samples. Each pixel in the hyperspectral images represents a specific wavelength of light reflected by the honey sample, providing detailed spectral information. The ANN model was designed to learn patterns and relationships within these spectral data to accurately classify honey samples as either pure or adulterated, as well as to predict the degree of adulteration.

During training, the ANN adjusts its internal parameters through a process called backpropagation, where errors between predicted and actual outcomes are used to update the connections between nodes. This iterative process continues until the model achieves satisfactory performance on a validation dataset.

The performance of the ANN was evaluated based on metrics such as precision, recall, F1-score, and accuracy. Precision represents the proportion of correctly classified adulterated honey samples among all samples classified as adulterated, while recall measures the proportion of correctly classified adulterated honey samples among all actual adulterated samples. The F1-score is the harmonic mean of precision and recall, providing a balance between the two metrics. Accuracy, on the other hand, represents the overall correctness of the classification results.

ANNs were utilized in the study to leverage the spectral information captured by hyperspectral imaging for the accurate detection and quantification of honey adulteration. Through the training and evaluation of ANN models, the study aimed to develop a reliable and efficient tool for ensuring the authenticity and safety of honey products.

Support Vector Machine (SVM)

The support vector machine approach is based on the notion of estimating (maximum) margins. The algorithm aims to discover a hyperplane (decision boundary) that is placed between the data points of the different classes and is as far away from the data points as possible. The support vectors are the data points nearest to the hyperplane [25]. Because the support vectors influence the position and orientation of the hyperplane, they are utilized to maximize the margin of the classifier. The hyperplanes (H) are defined by giving weights (w) to each feature as well as some bias (b), the combination of which predicts the target variable (y).

The bias term ensures that the separating hyperplane does not have to go through the origin. The weights are proportional to the feature’s importance. The features most important for splitting the data do have higher weights. When different classes are not linearly separable, the support vector machine uses a technique called ‘the kernel trick’. The basic idea of the support vector machine kernel is that the function transforms a low-dimensional input space into a higher-dimensional space, to be able to separate the target classes with a hyperplane. Different kernels can be specified, such as the linear kernel, the polynomial kernel, the radial basis function kernel, and the sigmoid kernel. Since a thorough explanation of the different kernels is beyond the scope of this thesis, this is not discussed here.

The primary reason for using the support vector machine is the ability of the algorithm to capture complex relationships without applying transformations to the data. By projecting the data into a higher-dimensional space, the support vector machine can model non-linear patterns in the data [26].

For this study, SVMs were trained on segmented and pre-processed hyperspectral images of adulterated honey samples, similar to the approach used for Artificial Neural Networks (ANNs). Each pixel in the hyperspectral images represents a specific wavelength of light reflected by the honey sample, providing detailed spectral information. The SVM model was designed to learn patterns and relationships within these spectral data to accurately classify honey samples as either pure or adulterated.

The key concept behind SVMs is to find the optimal hyperplane that best separates the different classes in the feature space. In the case of binary classification (pure vs. adulterated honey), this hyperplane acts as a decision boundary that maximizes the margin between the classes, thereby enhancing the model’s generalization ability [27].

During training, the SVM algorithm determines the optimal hyperplane by identifying a subset of training samples called support vectors, which lie closest to the decision boundary. The model aims to maximize the margin between these support vectors while minimizing classification errors.

K-Nearest Neighbors (KNN)

K-nearest-neighbor (KNN) classification is a fundamental and simple method used in classification studies when data distribution is unknown or difficult to determine. It is based on the idea that the nearest patterns to a target pattern provide useful label information. KNN assigns the class label to the majority of K-nearest patterns in the data space [28].

The KNN algorithms classify new data based on the class of the k nearest neighbors. The distance from neighbors can be computed using many distance metrics, such as Euclidean distance, Manhattan distance, Minkowski distance, and so on. In this study, Manhattan distance was used to compute the distance. The new data class may be determined by a majority vote or inverse proportion to the calculated distance. KNN is a non-generalizing approach since the algorithm stores all of its training data in memory, which may be translated into a fast-indexing structure such as a ball tree or a KD tree. For this research, KNN was trained on pre-processed hyperspectral images of adulterated honey samples. Each pixel in these images represents a specific wavelength of light reflected by the honey, providing detailed spectral information.

During training, the KNN algorithm stores all available cases and their class labels. When presented with a new data point, it calculates the distances to all other points in the training set and identifies the k nearest neighbors.

Classification is then determined by the majority class label among the k neighbors, with the most frequent label assigned to the new data point. The choice of the parameter k is critical, influencing the model’s performance and generalization ability.

Random forest

L. Breiman [29] presented the random forest algorithm, and it has proven to be an effective technique for regression and classification. It combines several randomized decision trees and averages the results to get an overall prediction. This approach is flexible, appropriate for solving complex issues, and appropriate for a range of learning assignments. With an emphasis on mathematical forces, parameter selection, resampling mechanisms, and variable importance measures, this article examines recent theoretical and methodological advancements for random forests [29].

Random Forest is a mast er learning technique used for grading and regression. It’s a kind of ensemble technique that combines a bunch of weak models to create a powerful model. The random forest generates several tresses. Voting for that class should be used to categorize each tree. This is a categorization. The forest chooses the classification with the most votes. Take the test to evaluate the features and choose trees to predict and save the findings [30].

In the study focused on detecting honey adulteration using hyperspectral imaging, the Random Forest algorithm played a key role. Random Forest is an ensemble learning method utilized for classification tasks. In this context, it was applied to analyze hyperspectral images of honey samples, where each pixel represents a specific wavelength of reflected light. These images provide detailed spectral information crucial for identifying adulteration. During training, Random Forest constructs numerous decision trees based on random subsets of the training data and features. This randomness helps mitigate overfitting and enhances the model’s ability to generalize to new data. In the context of honey adulteration detection, Random Forest aggregates the predictions of individual decision trees to classify samples.

By considering spectral patterns across multiple wavelengths, it accurately discerns between pure and adulterated honey.

Decision tree

Decision tree classifiers are widely used in many domains, such as text classification, user smartphone classification, images, and medical disease analysis. A decision tree is a graph that resembles a tree, with internal nodes standing in for attribute tests, branches for test results, and leaf nodes for class labels. The process of choosing a path from the root node to the leaf forms the classification rules. Each intermediate node’s input data is analyzed by identifying attributes and the values that correspond with them to construct the tree [31]. Depending on the test conditions at each node along the path, it can prefigure new data by going through all of the internal nodes. By evaluating vast amounts of intrusion detection data, decision trees can enhance real-time security systems and detect noteworthy network features that point to malicious activity. They offer a comprehensive set of guidelines that are simple to comprehend and work well with real-time technologies. It uses real data-mining algorithms to help with classification. A decision-tree process will generate the rules followed in a process [32]. Decision trees are useful for helping you choose among several courses of action and enable you to explore the possible outcomes for various options to assess the risks and rewards for each potential course of action. These decisions generate rules, which then are used to classify data [33]. Decision trees are the favored technique for building understandable models.

In the realm of honey adulteration detection using hyperspectral imaging, the Decision Tree algorithm serves as a fundamental tool. This method operates by recursively partitioning the dataset based on different spectral features at each node of the tree. The decision-making process involves evaluating specific wavelengths captured by the hyperspectral imaging system, expressed in nanometers (nm).

In the context of our study, the Decision Tree algorithm analyzes the hyperspectral data, focusing on the wavelengths denoted by labels such as Anm, Bnm, Cnm, Dnm, Enm, Fnm, and Gnm. These labels represent the specific nanometer values corresponding to the wavelengths at which spectral data were collected. By assessing the spectral characteristics associated with different honey samples, the Decision Tree algorithm builds a hierarchical structure that facilitates the classification of samples into distinct categories.

The decision nodes in the tree represent specific wavelength bands, and the branches emanating from each node signify the outcomes of the decision based on the observed spectral information. As the algorithm traverses the tree, it ultimately leads to leaf nodes, each corresponding to a classification decision regarding the presence or absence of honey adulteration.

By leveraging the Decision Tree algorithm, our research aims to determine the crucial spectral features indicative of adulteration, contributing to the development of an effective and interpretable model for honey quality assurance.

Performance measures

To accurately detect adulterated honey using various algorithms, proper evaluation needs to be conducted. This paragraph will discuss the various evaluation criteria used to compare the different algorithms. First, the data will be partitioned into a training set and a testing set. A resampling procedure known as 5-fold cross-validation will then be used on the training set to prevent potential biases. All classifiers were validated using 5-fold cross-validation. Performance is measured based on precision, recall, and accuracy. Precision is the ratio of correctly predicted positive observations to the total predicted positive observations [34]. Recall is the ratio of correctly predicted positive observations to all observations in actual class [35]. The easiest performance metric to understand is accuracy, which is just the proportion of properly predicted observations to all observations [36]. In this study, performance is measured based on the following parameters as shown below in Eqs 1, 2, and 3. (1) (2) (3) (4) where TP  =  True Positive, TN  =  True Negative, FP  =  False Positive, and FN  =  False Negative.

Results and discussion

This section describes the accuracy of the adopted models. The results of the decisions made in the prediction phase were collected, for each algorithm, in the relative confusion matrix. This is a matrix where the values predicted by the classifier are shown in the columns and the real values of each instance of the test-set are shown in rows. To proceed with the performance evaluation, we used the confusion matrix to derive a series of fundamental metrics to quantitatively express the efficiency of each algorithm: recording F1 Score, precision, and recall.

Environmental setup

Python was selected for code development with tools like Notepad and Anaconda because of its simple syntax. The mean values were computed using NumPy and Pandas, and the model was assessed using Scikit-learn’s random forest and tree module. Machine learning tools such as classification, regression, clustering, dimensionality reduction, model selection, and pre-processing are included in Sclera, also referred to as sclera. In the study, well-known algorithms like ANN, SVM, KNN, random forests, and decision trees were employed. Packages for collections and surprises were also employed. The algorithms’ performance was assessed and compared using cross-validation techniques.