Enhancing milled rice qualitative classification with machine learning techniques using morphological features of binary images

ABSTRACT

Rice is a globally important agricultural crop, with extensive cultivation and consumption in Asia. In Thailand, it is a primary food crop and a crucial export commodity. However, ensuring the quality standards of Thai rice is challenging due to variations in grain mixtures, making conventional inspection methods laborious and time-consuming. Human judgment in visual inspection introduces the risk of discrepancies. To address this, a swift and accurate solution is needed for quality analysis and differentiation of rice grain categories. Image processing techniques and machine learning offer a promising approach for accurate rice grain classification and reducing human grading errors. In a recent study focused on jasmine rice (KDML 105) samples, images of rice grains were captured using a developed device. Morphological features related to shape and size were extracted through image processing. The Boruta algorithm was employed to select significant features, which were then used to train various machine learning classifiers. After training and validation, the random forest classifier demonstrated the highest performance and was chosen as the main classification model. It was then tested with a new dataset to evaluate its identification accuracy. The selected model successfully classified four categories of rice grains with an accuracy exceeding 99.00%. While research efforts have primarily focused on classifying rice varieties and detecting grain abnormalities, incorporating a combination of morphology, color, and texture features is essential for highly accurate predictions. However, when it comes to predicting rice grain types with distinct shapes and sizes, considering relevant morphological characteristics during the model development process is sufficient to achieve highly precise and accurate results.

KEYWORDS:

• Rice
• rice quality
• rice image processing
• rice classification
• machine learning model

Introduction

Rice is an important global agricultural commodity, widely cultivated and consumed in Asia.^[1] In Thailand, besides being the main food crop, rice also plays an important role in the economy of the country.^[2] Rice has long been Thailand’s most important crop; rice cultivation accounts for 43.7% of all Thai farmland, and over the last 10 years, annual outputs of rice paddy have averaged 31–33 million metric tons, which after processing produces 20–22 million metric tons of milled rice. The domestic market absorbs around 10–12 million metric tons of this, with the balance either being exported or used to build stocks. In the 2021–2022 growing season, Thai-grown rice accounted for 4.0% of global outputs, coming after China (29.0% of total output), India (25.2%), Bangladesh (7.0%), Indonesia (6.7%), and Vietnam (5.2%), and exports ranked 2nd globally with a 13.5% share of the global market after India, which accounted for 38.8% of the total. Thai rice is still recognized for its quality and is in demand on the world market. Important export markets include Iraq, South Africa, China, the United States, Benin, and Japan.^[3] There are many types of Thai rice products, among which high quality and high value rice is Jasmine Rice or Thai Hom Mali Rice.^[4] It is processed rice from paddy varieties Khao Dawk Mali 105 (KDML 105) and RD 15, which are aromatic rice varieties that are sensitive to photoperiod,^[5] so they can be produced only once a year, therefore their yield is limited, and their value is high.^[6] Each year’s paddy yield is processed into milled rice causing some of the grain to be broken, which is less desirable to consumers and of low value. Rice consumers may lose benefits if broken rice is not controlled because they had to buy rice at a high price but got rice that had a lot of broken grains. Hence, the government has set product standards for processed Thai Hom Mali rice of white rice type, which shall be divided into six grades as white rice 100%, white rice 5%, white rice 10%, white rice 15%, white broken rice A1 extra super, and white broken rice A1 super.^[7] The standards specifications of rice component in each grade showed in Tables 1 and 2.^[8] Each grade has five different rice grain categories (Figure 1). In the production of each rice grade, the operator will bring each category of rice that is separated from modern processing to mix according to customer requirements or according to specified standards,^[9] causing the price of rice to be different depending on the mixing ratio. The above information shows that grain categories is therefore an important quality indicator that determines the grade of rice. However, inspecting the grades of rice containing slightly different rice grain size mixtures by conventional measurement is laborious and time-consuming. Inspection of grain size and classification by image processing techniques through machine learning algorithm is another way to be able to solve problems that arise.

Figure 1. Parts of rice kernels mean each part of the whole kernel that is divided lengthwise into 10 equal parts. Categorized into five subgroups, namely whole grain rice, head rice, big broken rice, small broken rice, and small broken C1.^[9]

Table 1. The standards specifications of Thai Hom Mali white rice.

Grades	Size of head rice (parts)	Size of brokens (parts)	Grain composition (%)
			Whole kernels	Head rice	Brokens and Small broken C1
					Total	Small brokens	Small broken C1
100%	≥8.0	≥5.0 to < 8.0	≥60.0	residue	≤4.0	0	0
5%	≥7.5	≥3.5 to < 7.5	≥60.0	residue	≤7.0	≤0.5	≤0.1
10%	≥7.0	≥3.5 to < 7.0	≥55.0	residue	≤12.0	≤0.7	≤0.3
15%	≥6.5	≥3.0 to < 6.5	≥55.0	residue	≤17.0	≤2.0	≤0.5

Table 2. The standards specifications of Thai Hom Mali white broken rice.

Grades	Grain Composition (%) ***
	A	B	C	D	E	F
A1 extra super	<15.0	-	≤74.0	-	<10.0	≤1.0
A1 super	-	-	<15.0	<80.0	-	≤5.0

***A: Whole kernels, B: Whole kernels and Brokens having the length as from 6.5 parts onward and combined, C: Brokens having the length as from 5 parts onward, D: Brokens rice not reaching 6.5 parts, E: Brokens rice not reaching 5.0 parts, and F: Small brokens C1.

The image processing and machine learning are being developed for use in a variety of agricultural applications. For example, Qadri et al.^[10] developed an automated seed classification system to classify canola varieties. Eight distinct canola varieties were photographed using a digital camera. Binary, texture, RST-invariant, histogram, and spectral features were extracted from these images. Subsequently, a correlation-based feature selection (CFS) algorithm was employed to reduce the dimensionality of the multi-dataset. The selected features were then utilized to train and test an Artificial Neural Network (ANN) classifier, employing a 10-fold stratified cross-validation approach. The results demonstrated the effectiveness of these techniques, yielding classification accuracy rates ranging from 95% to 98%. Aslam et al.^[11] introduced an automated system for mango variety discrimination based on leaf properties. Their approach involved capturing leaf images, extracting 57 features, reducing dimensionality using the CFS method, and employing both Logistic Model Tree (LMT) and K-Nearest Neighbors (KNN) classifiers for classification. The KNN classifier demonstrated superior accuracy compared to the LMT classifier, achieving accuracy rates ranging from 88.33% to 97%. In term of rice, image processing techniques have been widely studied in the context of rice seed sorting and characterization, offering benefits such as speed, low cost, and ease of operation. Various studies have utilized image processing algorithms to analyze the morphological features of rice grains, achieving high classification accuracies. For instance, Hanibah et al.^[9] achieved an average accuracy of 98% by focusing on morphological features like area, perimeter, major axis length, and minor axis length. Other studies employed important features and machine learning algorithms, such as Fuzzy Logic, Artificial Neural Network, and Support Vector Machine to classify rice quality with accuracies ranging from 80.00% to 97.33%.^[12–14] Qadri et al.^[15] assessed the machine vision (MV) techniques to classify six Asian rice varieties. Each selected rice variety contained 1800 grains were captured using cell phone camera. All the images were enhanced and converted into the standard 8-bit gray-scale format, then Six radius-based non- overlapping ROI form each image. Binary (B), Histogram (H), and Texture (T) features from each image were extracted. The correlation-based feature selection (WEKA) method with Best First Search algorithm was deployed reducing the dimensional of extracted features. Fan and Yang^[16] developed an identification model for milled rice varieties based on multiple image features and explored the contributions of each image feature to the accuracy of the identification. The extracted image features were 13 morphology features, 18 color features, and four texture features. Experiment results showed that, by selecting 20 features from the original 35 features, the average milled rice variety identification accuracy was 92.50%. Recently, Kurade et al.^[17] presented a low-cost rice quality assessment system using image processing and machine learning, achieving 77% accuracy with the random forest classifier. These studies highlight the potential of image processing and machine learning for accurate and cost-effective rice quality analysis.

Traditional methods for classifying different grades of milled rice rely on expert knowledge, high levels of skill, and experience. The approach requires a considerable amount of effort and patience. Moreover, it faces challenges when dealing with slight variations in grain size, making it prone to errors and causing confusion regarding rice quality and pricing. This can lead to unfair trading practices as rice with different grade amounts can have significantly different prices, affecting both rice customers and those involved in rice trading. To address these issues, the above studies show that image processing techniques through machine learning algorithms can be used to solve these problems, and the success of tasks varies according to the selected features and classification models used for processing. In this study, an automated system utilizing image processing and machine learning has been developed. This system offers advantages such as increased accuracy, speed, simplicity, and transparency in rice grain size classification. The primary objective of this study was to identify and select the optimal machine learning model for classifying rice grains based on their size using binary image morphological features as the criteria. The study’s process encompasses several key stages, including image acquisition, image pre-processing, feature extraction, feature selection using the Boruta algorithm, selection of a suitable classifier model, and evaluation of the chosen model’s performance.

Materials and methods

The steps in the development of the rice grain classification model are shown in Figure 2. After the preparation of samples, images of rice kernels were captured. The images were pre-processed to extract the morphological features of each rice kernel. The Boruta algorithm is a feature selection technique used to identify the most important variables in a dataset. The selected features were applied as inputs to 11 machine learning classifiers for separating various classes of rice kernels. The results of each classifier were evaluated and compared for classification efficiency, and then the best model was selected. Finally, the selected classifier was tested on a new image dataset for performance evaluation.

Figure 2. Steps of development the rice grain classification model.

Milled rice samples

The study utilized a 5 kg sample of Khao Dawk Mali 105 (KDML 105) variety obtained from the Rice Research Centers, Ministry of Agriculture and Cooperatives, and Rice Science Center in Khon Kaen Province, Thailand. The sample consisted of five distinct categories of white rice, namely whole rice, head rice, big broken rice, small broken rice, and small broken C1, with each class weighing 1 kg. The initial moisture content of the sample was determined using the hot air oven method^[18] and was between 11 and 12% (w.b.). Randomly a hundred rice grains from each rice category, as illustrated in Figure 3, measurements of both length and width using a micrometer (Mitutoyo Series 422 Blade Micrometer, Mitutoyo America Corporation, Texas, USA). The results disclosed specific size ranges for each category: Small broken C1 rice exhibited lengths ranging from 1.14 to 2.19 mm and widths of 1.01 to 2.10 mm. Small broken rice ranged from 2.03–2.67 mm in length and 1.67–2.47 mm in width. Big broken rice displayed lengths from 3.93–6.03 mm and widths between 1.96–2.79 mm. Head rice measured 6.19–6.89 mm in length and 1.98–2.70 mm in width, while whole rice grains had lengths spanning from 6.90–7.63 mm and widths from 1.98–2.85 mm. The distribution characteristics of rice grain sizes according to their category are illustrated in Figure 4. Statistical analysis with one-way analysis of variance (ANOVA) demonstrated significant differences (P < .01) in both length and width among the different rice groups. Subsequent comparison of average values using Duncan’s method at a 95% confidence level revealed specific size averages for each category, both in terms of length and width, as detailed in Table 3.

Figure 3. Samples of five categories of milled rice grains for measuring length and width.

Figure 4. The distribution of rice grain size samples is classified by rice grain category.

Table 3. Comparing the average rice grain size across five rice categories.

Rice categories	N	Length (mm)	Width (mm)
Small broken C1	100	1.82 ± 0.31e	1.58 ± 0.37c
Small broken	100	2.33 ± 0.15d	2.14 ± 0.18b
Big broken	100	5.05 ± 0.72c	2.35 ± 0.17ab
Head rice	100	6.56 ± 0.19b	2.37 ± 0.15ab
Whole rice	100	7.20 ± 0.18a	2.47 ± 0.16a

Note: Different letters in Length and Width columns differ significantly (p≤.05).

Image-capturing system

The rice samples used in the study were imaged using the device shown in Figure 5. The image was captured using an Olympus color digital camera (OM-D, E-M5 Mark II), and lighting was provided by an LED lamp composed of four 15-watt bulbs. This type of lamp produces less heat and allows for adjustable brightness. The mean illuminance level at the object stage was set to 2500 lux using the light control box and lux meter (Testo 540). The rice samples were set on an object stage 600 mm away from the camera lens. The images were captured at a resolution of 3456 × 3456 pixels and saved in JPEG format. These images were then transferred and stored on a computer’s hard disk using a USB interface to perform further analysis.

Figure 5. Image-capturing device used to record milled rice images.

Image acquisition

The images of the rice grain samples were taken by separately placing the samples on the object stage with a matte black background. In this work, the obtained image data is divided into two parts. The first part of the image data is for training and validation of the models. Random samples of rice grains were selected from five distinct rice categories, comprising small broken rice C1, small broken rice, big broken rice, head rice, and whole rice, with each category containing 400 grains. The rice grain samples from each type were placed on the image stage one category at a time, arranged so that the rice grains were not adjacent. Subsequently, images of individual categories were captured until all five categories were obtained. This process resulted in obtaining five images, which were utilized for training and validation of the model. The example images of rice grains from all five categories are illustrated in Figure 6.

Figure 6. The images of five milled rice categories for the training and validation process.

The second part is the image as a dataset for testing the efficiency of the selected model. Four captured images with different numbers of rice grades according to Table 1, namely white rice grade 100%, white rice grade 5%, white rice grade 10%, and white rice grade 15%, were performed as a test set. Figures 7 show a captured rice images with different rice grades used in the testing process.

Figure 7. The images of various rice grades in terms of grain quantity were used for testing the selected model.

Images pre-processing

The image pre-processing was performed using Python software. Initially, the RGB images were converted to grayscale. This step reduces the image to a single channel. The grayscale image was then transformed into a binary image. This conversion was achieved using the Otsu method, which automatically calculates an optimal threshold level to separate the foreground from the background. After thresholding, the resulting binary image consists of a background image and the rice grain images. By following these preprocessing steps, all the images were prepared for feature extraction of the individual kernels present within the image (the details are presented in: https://github.com/Sokudlo/Milled_Rice_Classification_Using_ML, section “01_Preparation of training dataset” and “02_Preparation of testing dataset”).

Feature extraction

In this study, the scikit-image library was utilized to extract 11 morphological features related to size and shape for each object. The resulting feature values correspond to the number of pixels attributed to each rice grain. For a detailed list of the morphological features given in Table 4.^[19,20]

Table 4. Morphological features extracted from each object.

No.	Feature	Explanation
1	Area (A)	The measure of the total number of pixels within the boundary of an object, representing its.
2	Major Axis Length (Ma)	The longest distance between two points on the boundary of an object, indicating its length.
3	Minor Axis Length (Mi)	The longest line on a grain of rice that can be drawn perpendicular to the major axis.
4	Eccentricity (Ec)	A measure of an object deviates from a perfect circle, indicating its elongation or roundness.
5	Equivalent Diameter (Eq)	The diameter of a circle with the same area as the object, providing a representative measure of its size. The calculation formula is: $\mathrm{E}\mathrm{q}=\sqrt{\frac{4\mathrm{x}\mathrm{A}}{\pi }}$
6	Solidity (S)	The ratio of the area of an object to the area of its convex hull, indicating the compactness or solidity of the object. The calculation formula is: $\mathrm{S}\mathrm{o}=\frac{\mathrm{A}}{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\mathrm{e}\mathrm{x}\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}}$
7	Extent (E)	The ratio of the number of pixels in an object to the total number of pixels in its bounding box, representing the spatial distribution. The calculation formula is: $\mathrm{E}=\frac{\mathrm{A}}{\mathrm{B}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{b}\mathrm{o}\mathrm{x}\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}}$
8	Perimeter (P)	The total length of the boundary of an object, indicating its overall shape complexity or outline.
9	Aspect Ratio (As)	The ratio of the length of the major axis to the length of the minor axis, indicating the elongation of an object. The calculation formula is: $\mathrm{A}\mathrm{s}=\frac{\mathrm{M}\mathrm{a}}{\mathrm{M}\mathrm{i}}$
10	Compactness (C)	The ratio of the object’s perimeter squared to its area. Higher values indicate more compact shapes. The calculation formula is: $\mathrm{C}=\frac{\mathrm{E}\mathrm{q}}{\mathrm{M}\mathrm{a}}$
11	Roundness (R)	A measure of an object resembles a perfect circle. The calculation formula is:$\mathrm{R}=\frac{4\hspace{0.17em}x\pi x\mathrm{A}}{{\left(\mathrm{P}\right)}^2}$

Feature selection

Feature selection is a fundamental concept in machine learning that significantly influences model performance. Its importance in classification stems from its ability to eliminate irrelevant features, thereby improving model performance, enhancing model interpretability, and reducing computational requirements.^[21] The primary objective of feature selection in this study is to identify the most informative and discriminative features that have a significant relationship with the classes of rice grains, thereby improving the predictive performance of machine learning models. Among the various feature selection methods available, the Boruta algorithm has gained substantial recognition and utilization. The Boruta algorithm is a wrapper-based approach that leverages the random forest classifier, widely acknowledged for its effectiveness.^[22] By introducing shadow features as permuted copies of the original features, Boruta assesses feature importance by combining Z-scores and maximum importance comparisons. Through an iterative process, Boruta progressively refines the feature set by eliminating irrelevant features while considering feature interactions and dependencies.^[23] This methodology empowers researchers and practitioners to select prominent features, thereby enhancing the performance, interpretability, and efficiency of the models. Furthermore, the convenient implementations of the Boruta algorithm are available in Python libraries such as Boruta and scikit-learn with the BorutaPy implementation, facilitating the application of Boruta to datasets and streamlining the feature selection process.

Rice grains classification and model selection

The scikit-learn Python library is used for implementing machine learning algorithms. To enable classification of rice grains based on their classes, 11 different machine learning models are performed on the selected feature data, namely, KNeighbors Classifier, Logistic Regression, Decision Tree Classifier, Gradient Boosting Classifier, Random Forest Classifier, Bagging Classifier, AdaBoost Classifier, Gaussian NB, MLP Classifier, SVC Linear, and Gaussian Process Classifier. In this study, 70% of the selected feature datasets from each rice grain category were randomly selected for training the models, while the rest of the datasets were used as validation sets. All models were run three times, and the best-performing model results were selected by comparison of the accuracy, precision, and macro-averaged F1-scores of the models. The accuracy of the model is assessed by employing a confusion matrix, which serves as a predictive tool for evaluating the performance of the model across various rice grain classes. Precision signifies the degree of correctness in predicting the class, whereas recall quantifies the model’s efficacy in accurately identifying samples belonging to their true classes. Mathematically, precision and recall are expressed as equation 1 to 4^[7,17]:

(1) $\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{s}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}\ast 100$(1)

(2) $\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}\ast 100$(2)

(3) $\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e}\mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{s}}\ast 100$(3)

(4) $\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2\hspace{0.17em}\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}\ast 100$(4)

True positives are the count of kernels whose actual class is positive and are correctly classified as positive by the classifier. True negatives represent the number of kernels whose actual class is negative and are accurately classified as negative by the classifier. False positives refer to the count of kernels whose actual class is negative but are incorrectly classified as positive. Lastly, false negatives indicate the number of kernels whose actual class is positive but are incorrectly classified as negative. The performance indicators of each classifier were subjected to an analysis of variance (ANOVA), and subsequently, the means were compared using Duncan’s method at a 95% confidence level to evaluate the optimal classifier.

Selected model evaluation

The objective is to assess the performance of the selected model in accurately categorizing rice grains. The evaluation is conducted using testing datasets derived from four different rice image samples (Figure 7), with the corresponding quantity of rice grain provided in Table 5. The model’s performance is evaluated through the utilization of a confusion matrix,^[21,24] which compares the classification results obtained from the system with the expected classification results. This analysis helps determine the accuracy and effectiveness of the model in classifying the rice grain samples.

Table 5. The quantity of rice grain in four testing datasets.

Samples	Number of rice grains in each sample					Remarks
	Whole rice	Head rice	Big broken	Small broken	Small broken C1
Sample 1	240	140	18	2	0	Grade 100%
Sample 2	240	132	26	2	0	Grade 5%
Sample 3	220	132	44	3	1	Grade 10%
Sample 4	220	112	64	3	1	Grade 15%

Results and discussion

The 11 morphological features were extracted from five binary images, each containing 400 rice grains, for a total of 2000 rice grains. These features exhibited varying ranges across the dataset. The area of the rice grains ranged from 55 to 1345 pixels, while the major axis length spanned from 10.00 to 77.18 pixels. Minor axis length varied between 7.26 and 24.40 pixels, and eccentricity values ranged from 0.18 to 0.97. The equivalent diameter fell between 8.37 and 41.52 pixels, while the solidity values ranged from 0.84 to 0.99. Extent values spanned from 0.36 to 0.91, and perimeters ranged from 24.17 to 166.23 pixels. Aspect ratios varied from 1.02 to 4.27, compactness ranged from 0.48 to 0.97, and roundness values ranged from 0.40 to 1.18. The summary of feature datasets and the distribution of these feature data, categorized by rice grain type, are shown in Table 6, and visualized in Figure 8, respectively.

Figure 8. The distribution of feature data categorized by rice grain type.

Table 6. Summary of feature datasets extracted from rice grains, groping by rice category.

	Small Broken C1			Small Broken Rice			Big Broken Rice
Features	Min.	Max.	Mean	Min.	Max.	Mean	Min.	Max.	Mean
Area	55.0	454.0	334.4	405.0	635.0	531.9	516.0	1025.0	760.6
Major Axis Length	10.0	30.5	24.2	28.5	40.9	34.6	37.4	68.6	49.3
Minor Axis Length	7.3	22.7	17.7	16.1	23.0	20.0	14.8	23.5	20.2
Eccentricity	0.2	0.9	0.7	0.6	0.9	0.8	0.8	1.0	0.9
Equivalent Diameter	8.4	24.0	20.4	22.7	28.4	26.0	25.6	36.1	31.1
Solidity	0.8	1.0	1.0	0.9	1.0	1.0	0.9	1.0	1.0
Extent	0.6	0.9	0.7	0.5	0.9	0.7	0.4	0.9	0.6
Perimeter	24.2	84.4	64.9	75.2	104.4	86.0	93.3	166.2	112.4
Aspect ratio	1.0	2.0	1.4	1.3	2.4	1.7	1.7	4.1	2.5
Compactness	0.7	1.0	0.8	0.6	0.9	0.8	0.5	0.8	0.6
Roundness	0.7	1.2	1.0	0.7	1.0	0.9	0.4	0.9	0.8
	Head Rice			Whole Rice
Features	Min.	Max.	Mean	Min.	Max.	Mean
Area	804.0	1149.0	1000.7	952.0	1354.0	1154.6
Major Axis Length	53.0	73.8	63.5	61.4	77.2	69.7
Minor Axis Length	16.6	23.5	20.5	18.2	24.4	21.6
Eccentricity	0.9	1.0	0.9	0.9	1.0	1.0
Equivalent Diameter	32.0	38.2	35.7	34.8	41.5	38.3
Solidity	0.9	1.0	1.0	0.9	1.0	1.0
Extent	0.4	0.9	0.6	0.4	0.9	0.6
Perimeter	121.7	159.9	137.0	137.1	161.9	149.1
Aspect ratio	2.4	4.3	3.1	2.6	4.0	3.2
Compactness	0.5	0.6	0.6	0.5	0.6	0.6
Roundness	0.5	0.8	0.7	0.6	0.7	0.7

The dataset underwent analysis using the Boruta algorithm, and the findings are presented in Table 7. The original features were assigned importance scores ranging from 0.00009 to 0.28957, with the major axis length feature attaining the highest score. Conversely, the shadow features received importance scores ranging from 0.00042 to 0.00075, with the extent feature achieving the highest score. Upon comparing the scores of all features with those obtained from the shadow features, it was determined that the feature set was reduced from 11 to 4 features. The selected features include major axis length, perimeter, area, and equivalent diameter.

Table 7. Result of feature selection by Boruta feature selection methodology.

Features	Important score (Original features)	Important score (Shadow features)	Hits	Decision
Area	0.20482	0.00075	True	Confirmed
Major Axis Length	0.28957	0.00060	True	Confirmed
Minor Axis Length	0.00009	0.00066	False	Rejected
Perimeter	0.28469	0.00053	False	Rejected
Eccentricity	0.00022	0.00057	True	Confirmed
Solidity	0.00036	0.00042	False	Rejected
Extent	0.00031	0.00074	False	Rejected
Equivalent Diameter	0.21267	0.00065	True	Confirmed
Aspect ratio	0.00022	0.00057	False	Rejected
Compactness	0.00021	0.00052	False	Rejected
Roundness	0.00027	0.00059	False	Rejected

To explore and select a machine learning classifier that enables accurate, rapid classification of rice grains based on their length categories. The 11 machine learning classifiers were trained and validated with selected features. About 70% of the features selected in the dataset were randomly selected for training the models, while the rest were used as validation sets. The performance evaluation of each classifier was based on accuracy, precision, recall, and F1-score. To assess the statistical significance of these results, a One-way ANOVA analysis was conducted. The analysis revealed that the accuracy, precision, and F1-score of each classifier exhibited statistically significant differences (P < .01), whereas the average recall values did not vary significantly. Upon comparing the average values using Duncan’s method at a 95% confidence level, it was observed that the Random Forest Classifier achieved the highest average accuracy at 96.67%. This result was significantly superior to the other classifiers. The Multilayer Perceptron Classifier achieved an average accuracy of 72.33%, which was the lowest among the classifiers. Regarding average precision values, the Random Forest Classifier also outperformed with an average precision of 96.60%. This result was significantly different from the Ada Boost Classifier, which achieved an average precision of 67.35%, but not significantly different from the other classifiers. In terms of the F1-score, the Random Forest Classifier maintained a high average of 96.52%. However, it did not exhibit significant differences from the other classifiers, except for the Ada Boost Classifier (70.63%) and the Multilayer Perceptron Classifier (66.85%). Detailed information regarding the mean comparison is shown in Table 8.

Table 8. The performance output of machine learning classifiers for rice grain classification.

Models	Accuracy (%)	Precision (%)	Recall (%)	F1_score (%)
K-Neighbors Classifier	96.50b	96.48a	96.33	96.38a
Logistic Regression Classifier	85.83h	85.30ab	85.26	85.00ab
Decision Tree Classifier	95.17f	94.99a	94.95	94.96a
Gradient Boosting Classifier	96.17c	96.13a	96.00	96.04a
Random Forest Classifier	96.67a	96.60a	96.49	96.52a
Bagging Classifier	96.00d	95.89a	96.81	95.84a
Ada Boost Classifier	75.33i	67.35b	77.57	70.63b
Gaussian NB Classifier	94.83g	94.84a	94.62	94.66a
Multilayer Perceptron	72.33j	79.87ab	72.10	66.85b
Support Vector Classifier	95.50e	95.52a	95.39	95.40a
Gaussian Process Classifier	95.50e	95.41a	95.33	95.36a

Note: Different letters in the Accuracy, Precision, and F1_score columns differ significantly (p≤.05), while the Recall column is not differentiated.

The results in Table 8 exhibit that all classifiers manifest accuracy in effectively classifying the various rice grain categories present in the rice samples. The primary goal of the classification model revolves around accurately discerning these rice grades. When considering the results of mean comparison for accuracy, precision, and F1-score, it was found that the Random Forest Classifier emerges as the optimal choice for the segregation of rice grain categories based on their length. This signifies the classifier’s ability to effectively capture the subtle nuances and variances in grain length, leading to a more accurate classification outcome. The Random Forest Classifier demonstrates robust performance on both the training and validation data, rendering it a suitable choice for selection. The Random Forest Classifier was chosen and subsequently retrained using the entire training dataset, utilizing only the features comprising area, major axis length, equivalent diameter, and perimeter. Following the training phase, the model was tested by employing the new four distinct rice image datasets. The prediction results are depicted in Figure 9, where the confusion matrix showcases the classification outcomes of rice grains in each sample. Additional details about this process are accessible through a link in the Supplementary Materials section.

Figure 9. Result of selected model testing with new datasets.

The result of testing sets in Figure 9 shows that the highest percentage of accuracy is 100%, given by sample 4 (white rice grade 15%), and the lowest is 99.00%, gathered from sample 3 (white rice grade 10%). The accuracy rate of the sample 1 classification was 99.75%, with a computation time of 0.1897 s. Only one grain of whole rice has been identified as head rice (in panel (a)). In panel (b), the accuracy rate of the sample 2 classification was 99.25%, with a total computation time of 0.1931 s. The output showed that two grains of head rice have been identified as whole rice and one grain of whole rice has been identified as head rice. The lowest accuracy of this classification occurred with sample 3 (in panel (c)), where two grains of head rice have been identified as whole rice and two grains of whole rice have been identified as head rice. The error in identification reduces the accuracy to 99.00%. And finally, in panel (d), the grains were identified collectively at 100%, with a total computation time of 0.1877 s.

The occurrence of errors in rice grain classification can often be attributed to the misclassification of whole rice, head rice, and big broken rice. After a thorough analysis of the range of pixel values for selected features within each rice category from each sample, it became evident that there was significant overlap in the pixel ranges associated with each rice grain category. This overlap, particularly in the cases of whole rice, head rice, and big broken rice, can lead to prediction errors. However, it is important to note that these misclassifications only account for a small percentage of the overall error, and they do not significantly impact the total computation time. Despite these challenges, the proposed method demonstrates an impressive accuracy rate of 99.50%. Moreover, the process proves to be suitable for assessing rice grain characteristics, specifically in terms of identifying the number of rice grains. A comparison of our proposed technique with the other results reported in the literature is presented in Table 9. The comparison results show that the method presented in this study has a better accuracy rate than other methods.

Table 9. Comparison between Existing and Proposed Approach.

References	Methodology/Features	Propose Classification	Accuracy (%)
[13]	The morphology features of rice grains, including parameters like area, perimeter, equivalent diameter, and roundness, used a fuzzy logic algorithm.	Fuzzy Logic	80.00
[19]	Applied preprocessing, segmentation, feature extraction, and feature selection techniques. And then training with various ANN structures.	Artificial Neural Network	97.33
[15]	Extracted for binary, histogram, and texture features. These features were used to create a feature vector for each image. Then employed the feature to classifiers	LMT Tree	97.39
[14]	HSV color and GLCM texture were extracted and trained in the classifiers	Support Vector Machine	96.83
Proposed Approach	RGB images were preprocessed to binary, feature extraction, and feature selection using the Boruta algorithm. The classifier was trained with selected features.	Random Forest Classifier	99.50

Conclusion

In this study, the Boruta algorithm was used for morphology feature selection with a high importance score. Four features were selected, including area, major axis length, equivalent diameter, and perimeter. The selected morphology features were trained in 11 machine learning algorithms, and it was found that the random forest classifier gave the best results, achieving high percentages of accuracy, precision, recall, and F1-score. The random forest classification model was chosen to evaluate its performance using four new datasets. The results of the classification demonstrated exceptional performance, with an average accuracy of 99.50% and a computation time of 0.1905 s. The study achieved high accuracy rates in rice grain classification through a combination of feature extraction techniques, feature reduction using the Boruta algorithm, and the Random Forest classifier. There are some limitations that need to be controlled when working with this technique, including the camera’s resolution, the distance between the camera lens, the intensity and evenness of the lighting, and the color and surface properties of the image stage. Another crucial challenge related to the correctness of this model is the preparation of rice samples and their placement for photography. Expert sorting of samples for each category of rice grain is a time-consuming and precision-demanding task. Moreover, when placing the rice grains on the image stage, meticulous placement is required to prevent the grains from sticking together because this technique cannot operate with rice grains that are stuck together. However, these methodologies demonstrate potential for application in the quality classification of milled rice. Further development is planned to develop some method for addressing the overlapping rice grains, which will be extended to encompass additional rice varieties, facilitating the creation of a comprehensive database. This will enable the development of an efficient model for milled rice grain classification across various rice varieties. Finally, plan to create dedicated devices for inspecting the quality of milled rice grains. That will benefit rice consumers and those involved in rice trading by enabling them to assess rice quality and purchase prices more effectively.

Acknowledgments

This research was supported by the Fundamental Fund of Khon Kean University, under the National Science, Research and Innovation Fund (NSRF), and the authors gratefully acknowledge the Postharvest Technology Innovation Center, Science Research and Innovation Promotion and Utilization Division, Office of the Ministry of Higher Education, Thailand. The Agricultural Machinery and Postharvest Technology Center and Department of Agricultural Engineering, Khon Kaen University, Thailand, provided support for all the facilitation in this study.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/10942912.2023.2264533.

Additional information

Funding

This research was supported by the Fundamental Fund of Khon Kaen University, under the National Science, Research and Innovation Fund (NSRF).