Prediction modeling using deep learning for the classification of grape-type dried fruits

The antioxidants, potassium, fiber, and iron included in raisins make them a healthy snack and a concentrated source of carbs for energy [1]. Globally, raisin output is comparable to that of rice, wheat, and corn in terms of importance to the agricultural industry [2]. Carbohydrates and starch abound in raisins. Because of its low cost and high satiety value, raisins play a crucial role in human diets around the world. Additionally, they have found extensive application in manufacturing. As one of the world’s leading producers of grapes, Türkiye is among the world’s top producers [3]. About 30% of Türkiye’s grape harvest is used for table consumption, 37% for drying, 3% for wine production, and the remaining 30% for other purposes [4].

Machine learning is a cutting-edge technology that can be used to identify and categorize a wide range of conditions, such as fire forest prediction, rice grain classifications, seed classifications and raisin varieties [5, 6]. This technique is used in a wide variety of fields, from food products to renewable energy [7,8,9]. The application of machine learning (ML), which makes use of computer algorithms to discover patterns in vast datasets with many complex features, has been substantially accelerated in the field of food industry by recent breakthroughs in big data [10]. ML makes use of computer algorithms to identify patterns in enormous datasets with numerous complex characteristics.

The intelligent application of ML prediction algorithms can aid in an early, more affordable approach to a variety of raisin grains. Our work here contributes to the field by allowing us to forecast raising grain classification using ML methods, which first requires the identification of important characteristics in the raw dataset through preprocessing. The results of this research could potentially make it easier to maintain quick and accurate classification of raisin grains within the framework of any realistic and reliable identification of this type of raisins. The results provided in this research are compared to those that have already been published [11].

The present paper is organized in the following manner: The literature review concerning the utilization of machine learning techniques in the context of raisin grains is exposed in part 2. The dataset utilized for this research endeavor is offered in section 3. The methodology employed in this study is delineated in section 4. Lastly, the findings derived from this investigation are deliberated about in section 5. In sections 6, 7, and 8, we provide our findings experimental results and discuss prospective avenues for future research and conclusion.

According to a study conducted in 1995 by Valiente et al. in [12], raisins are widely considered to be one of the most important sources of dietary fiber. Martin-Carron et al. in [13] have found that when there was a higher degree of tannin polymerization, there was also a higher inhibitory effect on digestive enzymes. This was the case regardless of the type of tannin. This contributes to the diet being more easily digested than it would be otherwise. Yeung et al. in [14] have demonstrated that Golden Thompson Seedless raisins have significantly higher antioxidant activity in addition to a higher total phenolics content than dipped and sun-dried Thompson Seedless raisins. Additionally, they found that the total phenolics content of these raisins was higher. Because of this discovery, the hypothesis that enzymatic activity plays a role in the reduction of phenolic content and antioxidant activity receives additional support.

When working in the food industry, it is not always simple to differentiate between the various kinds of food and levels of quality that are on the market. This manual process requires a lot of labor, which adds to the expense and extends the amount of time it takes. On the other hand, due to an inefficient method of classifying and sorting the products, there has been a reduction in the total number of products that have been shipped. It is possible that the utilization of an automatic raisin sorting system that is based on machine vision will have the potential to improve product quality, eliminate inconsistencies that are caused by manual evaluation, and lessen the reliance on readily available manpower.

Researchers have already started to utilize ML algorithms in order to carry out a wide variety of computations and conduct data analysis thus leading to relevant findings that will help improve people’s lives. The findings of studies that employ machine learning algorithms are beneficial to a wide variety of fields, including business, medicine, science, and mathematics, amongst others.

When it comes to evaluating and figuring out the quality of food, there are a lot of different ways in which traditional methods can be used. Nevertheless, these activities can be both labor-intensive and financially burdensome. Because of these drawbacks, researchers have been working hard on new techniques to swiftly and accurately assess most fundamental qualities of products, like the sweetness of their raisins. One such option is a machine learning system. Machine learning allows us to extract features from images that can be used to quantitatively assess the state of a product’s quality [15, 16].

As a result of this, many different aspects of food goods, including raisins, have been studied in recent years utilizing machine learning and image processing techniques. These aspects include color, texture, quality, and size. Several sensory aspects determine the quality of trading mass products. Consumers’ decisions and preferences are heavily influenced by aesthetic qualities, making them more crucial [17]. The raisins can be blended with other raisin varieties (such as seeded raisins) and low-quality raisins due to the high profit margins and the inconvenience of standard drying processes.

Okamura et al. in [18] were able to extract characteristics of raisins by using images of the fruit. The naive Bayes method was utilized by them in order to classify these characteristics. These were more effective than the traditional method of manually classifying raisins, according to Okamura et al. in [18].

Using this approach, Cinar et al. in [4] have photographed the raisins inside the box. The gathered statistical data was then classified using three separate machine approaches. Classification outcomes were obtained with the employment of the LR, MLP, and SVM algorithms. One of these methods, the support vector machine (SVM), has the highest success rate of 86.44 % [4].

Dirik et al. in [19] have proposed the hybrid PSO-ANN approach, which is compared to k-Nearest Neighbor (KNN) and Random Tree (RT) on a dataset of 900 raisin grains from two different categories. The hybrid PSO-NN strategy outperformed the other approaches in the evaluation, achieving a fantastic classification performance with an accuracy rate of 100%. KNN and RT had respective accuracies of 87.39% and %94.91.

Karimi et al. in [15] have developed an expert system to evaluate the consistency and freshness of raisins. The images of 1400 individual raisins were analyzed for their textural characteristics. As a result, the combinations were categorized using an Artificial Neural Network (ANN) and a Support Vector Machine (SVM). Using the top 50 features, the SVM classifier outperformed the ANN in terms of efficiency and accuracy of classification (an average of 92.71%).

For four distinct types of raisins, Mollazade et al. in [16] collected color pictures and extracted 36 colors and 8 shape attributes. They started with 44 different features and narrowed it down to 7 utilizing a correlation-based feature selection process. They have attempted to categorize raisins using artificial neural networks (ANN), support vector machines (SVM), decision trees (DT), and Bayes networks (BN). With the ANN method, they were able to achieve a classification accuracy of 96.33% [20].

Omid et al. in [21] have created a system for imaging and categorizing raisins to extract their size and color characteristics using image processing. They obtained approximately 96% classification accuracy.

Tarakci et al. in [22] have conducted a classification algorithm comparison of KNN and WKNN. The K closest neighbor (kNN) algorithm, the most widely used classification algorithm in machine learning, and the weighted kNN (WKNN) algorithm, which accounts for the importance of each feature’s index, are both employed. Using the inverse squared distance (w=1d2) (w=\frac{1}{{{d}^{2}}}) as the weighting factor produces a weighting that is proportional to the square of the distance. The results showed that the WKNN algorithm was more effective at classification than the kNN.

On the basis of the raisin grain dataset, a system for predicting the classification of raisin grains was developed by Yavuj et al. [23]. Using Machine Learning Methods, the 7-attribute, 2-class data set was acquired as-it was classified using Decision Trees and Random Forest. Analysis results showed that 85.44% and 85.22% accuracy were achieved with 020 Random Forest and Decision Trees, respectively.

Angadi et al. [24] have used the application of MATLAB for the classification of raisins through the use of image processing techniques. Through the use of color and size features, they were able to achieve an average accuracy of 95% in their study.

Khojastehnazhand and other researchers [25] have conducted research on the overall quality of raisins in bulk, using both high-quality and lower-quality raisins in addition to wood. The data model that was utilized contains a division into two modes, the first of which contains six classes and the second of which contains fifteen classes. The support vector machine (SVM) model and the linear discriminant analysis (LDA) algorithm are both examples of the proposed classification methods for the supervised learning category. Based on the results, it can be deduced that the SVM model that incorporates the grey level run length matrix (GLRM) performs better and more accurately.

According to Hu et al. [26], the implementation of classification methods that are founded on data mining (DM) techniques can serve as an efficient instrument for the production of accurate classifier models. There are powerful algorithms for selecting and extracting useful features, as well as training various DM models to adapt to difficult input output mappings [16,21,27]. These algorithms were developed by Kirkos et al. [27,28]. Image processing techniques have been successfully used in recent years to differentiate between the various types and sizes of food [29].

In a different study, Koklu et al. [30] have used the morphological traits that were found by looking at images of pumpkin seeds to carry out classification procedures. They employed five different machine learning techniques LR, MLP, SVM, Random Forest (RF), and k-Nearest Neighbor (k-NN) to carry out classification tasks. According to the findings of the study, the SVM approach, which produced an accuracy percentage of 88.64%, was the most accurate. Table 1 below illustrates the performance of previously stated results.

The “Raisin” data set was obtained from the website https://www.muratkoklu.com/datasets/. The two types of grapes used to construct the 900-piece data set were cultivated in Türkiye in equal quantities. On the basis of seven distinguishing features [4], raisins were divided into two broad categories: Kecimen and Besni. In Table 2, we see that 7 morphological characteristics were determined for each raisin.

Figure 1 shows the data for both types of raisin samples, including the minimum, average, and maximum values, as well as the standard deviation, skewness, and Kurtosis.

Fig. 1

The primary goal of our investigation is to develop a robust and precise algorithm for the identification of raisin grains. The work entails extensive data gathering, cleaning, processing, training, and testing. The algorithm 1 depicts the procedures involved in machine learning classifiers.

Predictions can be made after finishing the three phases of gathering datasets, training machine learning classifiers, and evaluating those classifiers. In this step, the data is divided into training and testing data in a 70–30 ratio. As a result, 70% of the data is used for training, with the remaining 30% used to test the performance of the previously developed model. The more training data there is, the easier it is to learn from. As shown in the flow chart, the methodology of decision making system is based on supervised machine learning classifiers. As shown in the flow chat, Figure 2 shows that the methodology is observed.

Fig. 2

4.2

Data pre-processing

The variables in the dataset each have each their own unique distribution of values, which may make it more challenging for the models to produce accurate results. Data reduction to a single order improves precision when there is a substantial disparity between the data. To do this, min-max normalization was employed to uniformly distribute the attributes present in the data sets utilized in the analysis between 0 and 1. Equation (1) provides the formula for min-max normalization [32].

(1) x′=x−min(x)max(x)−min(x). x'=\frac{x-\min (x)}{\max (x)-\min (x)}.

Figure 3 illustrates the correlation between each of the morphological features.

Fig. 3

4.4

Flowchart of proposed methodology

Reading a variety of papers and becoming familiar with the nature of the work involved is the first step in gaining domain expertise [35]. The proposed approach is applied to the 8-column dataset. Eleven different machine learning methods, including five deep learning-based neural networks, are used. Using the R programming language, the code is written in a Jupyter Notebook. In this study, the LightGBM Classifier achieves a 98.40% accuracy and a 92.57% ROC AUC with adequate parameter tuning utilizing cross-validation in the training set. The flow chart depicts the steps in our methodology workflow, Figure 4.

Fig. 4

4.5

Performance evaluation

In order to fortify machine learning classifications, cross validation has become a regular method of evaluation. In cross-validation, a dataset is split into k groups for separate training and testing purposes. The system is trained using the remaining sets while one set is used as a test set. This is done until k is reached, and then the system is put through its paces [36]. The following Figure 5 depicts the results of a cross-validation [4].

Fig. 5

Classification algorithms can be evaluated using standard measures including accuracy, specificity, sensitivity, recall, and F1 score. The confusion matrix is used to determine these values. Confusion matrices are tables used to describe the performance of a classification model on a set of test data that is already well-known. There are 4 variables that make up the confusion matrix. True positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) are the four possible outcomes. The format of the confusion matrix is displayed in Table 3 [37].

In order to perform an accurate evaluation of the machine learning classifiers, important measures were collected from the confusion matrix. In addition to the correct classification rate or accuracy, other metrics such as True Positive Rate (TPR), False Positive Rate (FPR), Precision, Recall, F1 score, and ROC area were used to evaluate the machine learning classifiers. These metrics were used to evaluate the classifiers’ performance (See Table 4).

The data were computed and reviewed to determine the optimum technique for predicting varieties of raisin grains concerns. The performance criteria such as accuracy, precision, recall, F1-score, and area under the curve (AUC) were used to evaluate eleven machine learning and five deep learning methods.

It is insufficient to rely just on accuracy as a performance measure for a model. The AUC value is an important indicator for assessing a model’s performance and ability to discriminate across classes. Machine learning algorithms such as SVM, DT, LR, NB, RF, AdaBoost, XgBoost, LightGBM, and Deep learning were used to achieve performance values. In all models used, the cross-validation k value was set to 10 in Table 5 and 5 in Table 6. Table 5 and 6 show the confusion matrix for all of the algorithms employed in the study.

At various points along the probability curve, the True Positive Rate and False Positive Rate are calculated. The performance of GaussianNB, DecisionTree, K-Nearest Neighbor, RandomForest, SVM, XGBoost, LightGBM, AdaBoost, Logistic Regression, Conviluation Neural Network, Radiaal Basis Function Network, Recurrence Neural Network, Artificial Neural Network, and Deep Neural Network ML algorithms on the raisin grains dataset is represented in Table 5.

According to the model performance with a 10-fold cv compared to 5-fold cv, LightGBM has the maximum accuracy score of 99.00%, while Artificial Neural Networks has the lowest accuracy score of 64.81%, as shown in Figure 6. Furthermore, AdaBoost performed well, scoring 90.30%. The SVM, Logistic regression, and Knn values are 87.78%, 87.41%, and 87.04%, respectively.

Fig. 6

How well the models can tell the difference between the positive and negative classes is quantified by the AUC. The AUC can take on values between 0 and 1, with higher values indicating more favorable outcomes. Incorrect results are indicated by an AUC of 0.5, while a perfect score of 1 shows that the test was absolutely accurate. AUCs between 0.7 and 0.8 are generally accepted as being below the raisin classification or variety detection threshold. A performance rating of 0.8 to 0.9 is considered outstanding, and a rating of 0.9 or above is exceptional [38].

It can be difficult to obtain a large and diverse dataset of identified raisin images for training. Manual work is required to gather, classify, and annotate images appropriately, which can be time-consuming and expensive. Identifying meaningful indicators that effectively differentiate between different quality aspects of raisin grains might be difficult. It is critical for accurate classification to develop adequate feature extraction methods that capture the key properties of raisins.

The dataset for this study was gathered from Turkish raisin grains. Using a real-world dataset of raisin types, this study investigates how to forecast raisin kinds grown in Türkiye or any other country. Machine learning and deep learning are used in this study to create a model for predicting the seed of a raisin dataset. Several studies were examined to assess objectives, analyze the methods used, and identify any problems in the research.

Moreover, during the process of our research, we made use of fourteen different machine learning algorithms and evaluated the level of accuracy achieved by each. When we looked at all of the different methods for predicting raisin grain based on the information that we utilized, LightGBM had the greatest accuracy of 98.40%.

The Supervised machine learning technique has been utilized successfully in a variety of applications, including women career and work happiness in constructing appropriate algorithms for predicting employees’ job satisfaction with a variety of individual characteristics [39].

The work can be expanded in the future by gathering more data from different regions of the world to analyze. The accuracy of the machine learning model predictions would improve with the addition of more data. Some additional deep learning approaches may also be used.