Characterizing and Predicting the Quality of Milled Rice Grains Using Machine Learning Models

Abstract

Physical classification is the procedure adopted by the rice unloading, delivery, storage, and processing units for the commercial characterization of the quality of the grains. This step occurs mostly by the conventional method, which demands more time and specialized labor, and the results are subjective since the evaluation is visual. In order to make the operation faster, more accurate, and less dependent, non-destructive technologies and computational intelligence can be applied to characterize grain quality. Therefore, this study aimed to characterize and predict the quality of whole, processed rice grains, as well as classify any defects present. This was achieved by sampling from the upper and lower points of four silo dryers with capacities of up to 40,000 sacks. The grain samples had moisture contents of 16%, 17%, 18%, and 19% and were subjected to drying-aeration until reaching 12% moisture content (w.b.). Near-infrared spectroscopy technology and Machine Learning algorithm models (Artificial Neural Networks, decision tree algorithms Quinlan’s algorithm, Random Tree, REPTree, and Random Forest) were employed for this purpose. By analyzing Pearson’s correlation statistics, a strong negative correlation (R² = 0.98) was found between moisture content and the yield of whole grains. Conversely, a strong positive correlation (R² = 0.97) was observed between moisture content and classified physical defects across the various characterized physicochemical constituents. These findings indicate the effectiveness of near-infrared spectroscopy technology. The Random Tree model (RandT) successfully predicted the grain quality outcomes and is therefore recommended as the model of choice, obtained Pearson’s correlation coefficient (r = 0.96), mean absolute error (MAE = 0.017), and coefficient of determination (R² = 0.92). The results obtained here reveal that the combination of near-infrared spectroscopy technology and Machine Learning algorithm models is an excellent non-destructive alternative to manual physical classification for characterizing the physicochemical quality of whole and defective rice grains.

1. Introduction

The classification step is responsible for characterizing the physical quality of the grains by manually separating the physical defects. Physical defects can come from the crop, for example, due to weather conditions (fermented, burnt, and moldy grains), cultivar type (chalky grain), or due to harvesting processes (broken rice), as well as physical defects can appear or worsen in the post-harvest stages. Regardless of the stage, for quality standardization, there are specific regulations, which attribute maximum levels of defects, as well as specifications for product marketing [1]. During the classification process, the grains with defects are visually identified, classified according to the specific standard, and then removed from the sample [2,3]. The conventional classification process demands more processing time and specialized labor, which can directly interfere with the logistics and flow of the grain mass at the pre-processing and storage unit. Moreover, physical evaluations are often subjective and can lead to errors, impacting the quality of the commercialized product.

To meet this demand, the need for new studies to evaluate technologies for indirect measurement of grain quality has emerged, so that the process becomes faster and more assertive. Non-destructive technologies and computational intelligence algorithms have been applied for characterizing the qualitative parameters of agricultural products [4,5]. Among the currently available technologies, near-infrared spectroscopy (NIR) is one of the most addressed and applied for agricultural product evaluation. Near-infrared spectroscopy is a highly flexible form of analysis that can be used in a wide range of research and industrial process applications. NIR spectroscopy is a method that uses the near-infrared region of the electromagnetic spectrum (from about 700 to 2500 nanometers). By measuring scattered light from and through a sample, NIR reflectance spectra can be used to quickly determine the properties of a material without altering the sample [4]. Scientists applied this technology in the evaluation of rice grains and they are achieving 93% accuracy [4]. Furthermore, the NIR technology obtained satisfactory results in the evaluation of rice quality for different cultivars and fertilizer levels [1].

The use of Machine Learning (ML) algorithms has also presented expressive outcomes when applied to predict the quality of agricultural products. Machine learning focuses on the principle that all complex data points can be mathematically linked by computer systems, provided they have enough data and computing power to process those data. In this context, the use of ML algorithms has offered greater capacity for processing, analyzing, and interpreting data [6]. When properly modeled, ML techniques can offer responses in less time when compared to statistical regression models. Overall, the main algorithms that have been applied in agricultural studies are: Artificial Neural Networks, Decision Trees, Random Forest, and Support Vector Machines [7,8]. Random Forest (RF) is an ML technique successfully used in yield forecasting and quality assessment [9]. This model proved to be an effective and easier-to-use method for predicting corn and wheat quality when compared to multiple linear regression models. Artificial Neural Networks (ANN) is another model that can be trained from data related to corresponding inputs and outputs [10]. ANNs are useful tools for the analysis and interpretation of complex food security data, and predictions of physical and chemical seed quality. During the last few years, research has investigated the results of using ML methods for classification within the context of agricultural problems, such as the prediction of nitrogen content [11], soil correction, seed classification [12], phosphorus reduction in wastewater [13], protein prediction in stored grains [14].

Some authors utilizing computational intelligence obtained positive results for soybean seed quality prediction, highlighting the speed of analysis compared to conventional methods [7]. Similarly, Lutz and Coradi [8] verified that the use of ML techniques predicts the deterioration of stored grains, assisting in decision-making. Moreover, Kiratiratanapruk et al. [15] used and developed computational intelligence techniques to classify rice grain varieties, obtaining accuracies above 90% for different models. Therefore, the NIR and ML technologies have a wide and successful application in the characterization and qualitative prediction of different agricultural products, and are of paramount relevance particularly for rice grains, due to the rigorous standardization requirements, justified by the way of commercializing the product and the level of market demand.

In order to reduce errors and the time for decision-making on the quality of rice batches received or shipped from processing and storage units, due to the subjectivity of visual and manual physical classification, the application of the technique of measurement by NIR and prediction by ML models. Therefore, understanding the physical-chemical parameters of rice grains through non-destructive and prediction technologies enables the replacement of the conventional method of physical classification. As a hypothesis, characterizing the quality of whole and defective rice grains by being analyzed through non-destructive technology and with the aid of ML algorithms makes the operation more assertive, fast, and independent of visual evaluations. Thus, the objective of this study was to evaluate the application of near-infrared spectroscopy and Machine Learning models for characterizing and predicting the quality of whole and defective rice grains to replace the conventional method of physical classification. Specifically, we aimed to: (i) physically characterize the quality of rice through manual physical classification; (ii) evaluate the physicochemical quality of whole and defective rice grains for different water contents using near-infrared spectroscopy; (iii) predict the physicochemical quality of whole and defective rice grains for different water contents by applying ML algorithms; and (iv) evaluate the performance of near-infrared spectroscopy combined with ML as an alternative to conventional rice grain classification methods.

2. Materials and Methods

2.1. Description and Experimental Design

The paddy of the IRGA 424 variety was produced in the Cachoeira do Sul municipally, Rio Grande do Sul, Brazil, in the year 2022 in Planossolo Háplico soil. The rice was harvested with different initial moisture contents (

Table 1. Characterization of rice sample collection storage silos.

Silos	Total Stored (Sc of 50 kg)	Moisture Content (% d.b.)
Silo 1	42,218.60	19
Silo 2	36,871.40	18
Silo 3	28,660.20	17
Silo 4	46,212.20	16

), then the grains were subjected to drying in silo dryers up to 12% (w.b.) in four full-scale silo dryer units, model SFP-18314 (Pagé industry, Araranguá, Santa Catarina, Brazil). Sampling was performed at 11 different points for each of the four silo dryers. The first six points were located at the top of the silo dryer, following the alignment of the thermometry cables allocated.

The remaining five points were collected at the bottom of the silo dryers, near the discharge points, and evenly distributed at the base (Figure 1). Subsequently, the rice grain samples were processed and subjected to separation into whole and broken grains, followed by classification according to defects.

Figure 2 illustrates the operations, including: sample collection during storage, processing, manual physical classification, physical-chemical analysis using near-infrared spectroscopy, and quality prediction using Machine Learning models.

2.2. Rice Processing and Physical Classification

For the processing of rice grains, a rice polisher, Paz-1/DTA model (Zaccaria company, Limeira, São Paulo, Brazil), was used. It was calibrated and operated according to the manufacturer’s technical recommendations. The paddy rice grains were gradually added to the input hopper of the polisher to obtain the dehusked and polished rice.

The polishing process involved the passage of grains between two abrasive stones present in the equipment’s huller, which removed the outer layer of the grains. To separate the whole grains from the broken ones, a cylinder separator with 5.5 mm cells, attached to the rice polisher, was used. As the cylinder rotated, the broken grains entered the cells and were discharged by gravity into a horizontal hopper, while the whole grains remained retained in the cylinder for subsequent separation.

After processing, the samples underwent manual physical classification of the rice, following the Normative Instruction No. 02, dated 7 February 2012, which establishes the physical classification standards for grains and commercial information, considering the following defects: red, yellow, scorched, immature, chalky, moldy, cut or stained, broken, streaked, immature, and discolored, as well as impurities and foreign materials adhered to the mass of grains [16]. After the physical classification, the grains with defects, along with the broken grains, were combined into a single sample according to the evaluated moisture content, resulting in samples of whole grains and samples of defective grains for the four moisture contents analyzed.

2.3. Near-Infrared Spectroscopy (NIRS)

For the physicochemical evaluation of the rice grains, near-infrared spectroscopy (NIRS) was used. A Metrohm DS2500 spectrometer (Metrohm company, Herisau, Switzerland) was employed. The samples were homogenized and placed in a sample capsule. They were then illuminated with radiation of a specific wavelength in the near-infrared region. The instrument measured the difference between the amounts of energy emitted by the spectrometer and reflected by the sample to the detector at various bands, creating a spectrum for each sample. The spectral data were recorded in reflectance mode in the spectral range from 400 to 2500 nm, determining the content of starch (ST), crude protein (CP), fat (Fat), ash (AS), and crude fiber (CF) in the whole and defective rice grains for different moisture contents. Additionally, for the whole grains, the apparent specific mass (ASM) was also determined, following the methodology described by Mohsenin [17]. Five replicates were performed for each sample.

2.4. Pearson Correlation Network

Pearson correlation network analysis was performed using the free R software “ggfortify”, following the methodology by Naldi et al. [18]. In the correlation network, the proximity between the nodes was determined by proportionality to the absolute value of the correlation between the nodes. Additionally, the thickness of the edges was controlled by applying a cutoff value of 0.60, indicating that |r_xy| ≥ 0.60 had their edges highlighted. Positive correlations were highlighted in green, while negative correlations were represented in red.

2.5. Machine Learning Algorithms

Data analysis using Machine Learning algorithms involved the application of the following models: Artificial Neural Networks (ANNs), decision tree algorithms Quinlan’s algorithm (M5P), Random Tree (RandT) and REPTree (ReepT), and Random Forest (RF). Multiple Linear Regression (MLR) was used as a control technique. Based on these models, the following variables were predicted: crude protein (CP), ash (AS), fat (Fat), crude fiber (CF), and starch (ST) for whole rice grains, and for defective grains with different moisture contents. Additionally, the variable apparent specific mass (ASM) was included only for the analysis of whole grains. The following variables were considered as input for each prediction model of the physicochemical properties of rice grains: whole grain yield (YIE), defects (GD), and moisture content (MC).

The ML analyses were performed using stratified cross-validation with k-fold = 10 and ten repetitions (100 runs for each model) and adopting the default configuration for all model parameters [19]. All prediction analyses were performed on the Weka software version 3.9.5 on an Intel^® CoreTM i5-3317U CPU with 4 GB of RAM. Weka aims to aggregate algorithms from different approaches in artificial intelligence dedicated to the study of machine learning. This sub-area intends to develop algorithms that allow a computer to “learn” either inductively or deductively. Weka performs computational and statistical analysis of the data provided, resorting to data mining techniques, inductively trying to generate hypotheses for solutions from the patterns found and, at the extremes, even theories about the data in question. The ANN algorithm used consists of a single hidden layer formed by a number of neurons equal to the number of attributes plus the number of classes divided by 2 [20] REPTree model is an adaptation of the C4.5 classifier and can be used in regression problems with an additional pruning step based on an error reduction strategy [21]. RandomTree model is a class for constructing a tree that considers K randomly chosen attributes at each node. It does not perform pruning and also has the option to allow the estimation of class probabilities based on a waiting set. The M5P model is a reconstruction of Quinlan’s M5 algorithm based on the conventional decision tree with the addition of a linear regression function at the leaf nodes [22]. The RF (Random Forest) model can generate multiple prediction trees for the same dataset and use a voting scheme among all the learned trees to predict new values [23]. The MLR (Multiple Linear Regression) model was used as a control model as it is suitable for predicting relationships between variables.

The statistics used to verify the quality of fit of the prediction models were the mean absolute error (MAE) and Pearson correlation coefficient (r) between observed and predicted values by each model. For comparison of the models, MAE and r means for each model were grouped by the Scott–Knott test at 5% probability and shown through boxplot graphs. These analyses were performed on the R software using the ExpDes.pt and ggplot2 packages.

3. Results and Discussion

3.1. Whole Rice Grains

Table 2. Physical and physicochemical quality of whole rice grains in function of moisture content.

Moisture Content (% d.b.)	Whole Grain Yield (%)	Crude Protein (%)	Fat (%)	Crude Fiber (%)	Ashes (%)	Starch (%)	Specific Apparent Mass (kg m−3)
19	49.884	8.13	1.85	2.08	0.92	70.85	585.51
19	50.529	9.06	1.82	2.07	0.89	71.82	538.25
19	51.015	8.23	1.86	2.06	0.80	70.75	562.79
19	52.536	8.90	1.68	2.04	0.97	71.42	517.52
19	52.944	7.58	2.02	2.09	0.78	70.32	585.98
19	53.836	9.07	1.64	2.01	0.92	73.21	493.52
19	53.836	7.67	1.94	2.06	0.85	72.91	588.97
19	54.395	8.01	1.77	2.01	0.88	71.59	555.68
19	54.531	8.27	1.86	2.07	0.88	72.53	541.94
19	54.976	8.74	1.65	2.02	0.95	71.94	524.56
19	54.976	7.78	1.92	2.07	0.87	72.91	584.46
19	55.057	7.78	1.92	2.07	0.87	72.91	584.46
Average	53.836 d	8.18 a	1.855 a	2.065 a	0.88 b	71.88 b	559.235 a
Standard deviation	1.760	0.523	0.116	0.026	0.053	0.932	30.826
18	58.237	8.44	1.62	2.09	1.09	71.34	548.90
18	58.668	10.24	1.51	2.07	1.12	70.62	469.23
18	58.903	7.19	1.91	2.13	1.02	72.47	571.40
18	59.030	8.61	1.61	2.06	1.07	72.30	493.12
18	59.298	8.02	1.86	2.11	1.01	71.19	524.85
18	59.537	10.19	1.71	2.05	0.95	68.98	499.48
18	59.564	7.78	1.87	2.10	1.02	71.47	561.26
18	60.075	10.14	1.75	2.00	0.97	71.56	523.58
18	60.115	7.410	1.94	2.11	1.01	72.65	560.69
18	60.702	10.81	1.74	2.06	1.09	67.39	521.58
18	61.143	7.490	1.89	2.14	1.03	72.38	516.77
18	61.223	7.490	1.89	2.14	1.03	72.38	516.77
Average	59.5505 b	8.23 a	1.805 a	2.095 a	1.025 a	71.515 b	522.58 b
Standard deviation	0.923	1.270	0.134	0.040	0.048	1.519	29.205
17	55.089	7.45	1.77	2.15	0.94	72.46	550.98
17	55.752	8.1	1.74	2.11	1.02	72.90	498.30
17	55.848	7.74	1.96	2.13	0.97	72.60	519.91
17	56.002	8.25	1.88	2.14	1.07	70.06	500.51
17	56.398	7.73	1.86	2.10	1.06	72.76	509.76
17	56.431	8.83	1.72	2.15	1.04	70.89	523.38
17	56.586	8.02	1.78	2.11	1.08	72.18	535.72
17	56.586	8.30	1.76	2.15	1.00	72.68	514.19
17	57.058	8.11	1.78	2.12	1.01	71.93	521.70
17	57.122	7.71	1.77	2.12	0.94	73.12	470.51
17	57.352	8.24	1.76	2.14	1.10	71.32	520.98
17	57.657	8.24	1.76	2.14	1.10	71.32	520.98
Average	56.5085 c	8.105 a	1.77 b	2.135 a	1.03 a	72.32 a	520.445 b
Standard deviation	0.708	0.349	0.066	0.017	0.055	0.896	19.207
16	61.534	7.47	1.85	2.05	0.99	72.26	550.14
16	61.759	9.00	1.67	2.10	1.11	70.51	544.36
16	62.547	8.33	1.67	2.10	1.00	70.38	491.48
16	62.547	9.13	1.78	2.07	1.02	72.01	527.28
16	62.797	7.56	1.80	2.08	0.98	71.35	507.31
16	62.818	7.72	1.85	2.10	1.07	72.68	552.41
16	63.235	8.10	1.95	2.07	0.87	71.99	552.61
16	63.941	8.78	1.63	2.07	1.06	71.22	551.15
16	64.083	8.08	1.85	2.04	0.82	72.05	551.16
16	64.724	9.32	1.66	2.05	1.10	69.93	548.89
16	65.784	7.40	1.94	2.09	0.97	72.09	534.21
16	66.456	7.40	1.94	2.09	0.97	72.09	534.21
Average	63.0265 a	8.09 a	1.825 a	2.075 a	0.995 a	72.00 a	546.625 a
Standard deviation	1.464	0.682	0.113	0.020	0.083	0.829	18.960

Means followed by the same letters do not differ by the Scott–Knott test at 5% probability.

shows the results of the physicochemical characterization of rice grains based on the initial moisture content (MC) before drying and the percentage of whole grains (YIE) obtained after drying. We observed that lower initial moisture content in the grains led to higher percentages of beneficiated whole grains, resulting in higher percentages of starch (ST) and fat (Fat). In grains with moisture content (MC) between 18 and 19%, higher values of apparent specific mass (ASM) and crude protein (CP) were observed, along with lower values of ash content (AS).

By Pearson’s correlation (Figure 3 and

Table 3. Coefficients of the associations between the variables (Pearson’s correlation)—whole rice grains.

Variables	MC	YIE	CP	FAT	CF	AS	ST	ASM
MC	1	−0.76864	0.11273	0.06508	−0.30666	−0.43654	0.03568	0.24787
YIE	−0.76864	1	0.00219	−0.03730	0.03704	0.40562	−0.12241	−0.16039
CP	0.11273	0.00219	1	−0.66926	−0.45180	0.26111	−0.63285	−0.36902
FAT	0.06508	−0.03730	−0.66926	1	0.25374	−0.49700	0.30056	0.54024
CF	−0.30666	0.03704	−0.45180	0.25374	1	0.38640	0.12262	−0.12615
AS	−0.43654	0.40562	0.26111	−0.49700	0.38640	1	−0.23959	−0.45760
ST	0.03568	−0.12241	−0.63285	0.30056	0.12262	−0.23959	1	0.10019
ASM	0.24787	−0.16039	−0.36902	0.54024	−0.12615	−0.45760	0.10019	1

), it is possible to verify a negative strong correlation for MC × YIE, indicating an inverse relationship between them. The mass of grains with higher initial moisture content (MC) accumulated a higher amount of heat at the end of drying, increasing thermal damage and decreasing the yield of whole benefited grains [24]. Weak negative correlations were found for MC × CF × AS, occasioned by the physical changes in the morpho-cellular tissues that affected the physicochemical compositions of the grains. Moreover, there is a weak inverse relationship between MC × Fat, since the lipid content was affected by the degradation of the aleurone layer due to the metabolic activity of the grains resulting from the water contents [1,24]. Apparent specific mass (ASM) had a positive and weak correlation with moisture content (MC). Some authors obtained higher ASM values in paddy rice grains stored with increased moisture content [25]. According to the authors, the ASM was altered as the moisture content (MC) of the grains decreased during drying [25].

Regarding YIE, weak negative correlations were observed with MEA and ST. Starch (ST) is composed of amylose chains that form molecular structures, which directly influence the hardness of the grain. Thus, rice grains with higher amylose contents are more resistant to abrasion in processing, achieving a higher yield (YIE) of whole grains [1,26]. Medium and weak negative correlations were observed between CP × Fat and CP × ST, respectively. The highest concentrations of crude protein (CP) were located in the endosperm of the grain, along with the starch content (ST), where the increase in one implied the reduction of the other [1,26]. Furthermore, according to Nunes et al. [24] the drying operation interferes with the decrease in the crude protein (CP) extraction, especially in the protein-starch ratio. The inverse relationship between fat (Fat) and crude protein (CP) content in whole rice grains was verified by Müller et al. [1]. According to Denardin and Silva [26], lipid bodies called triacylglycerols are stored in the endosperm of grains, where they are stabilized by hydrophobic proteins, which mobilize fatty acid catalysis.

A negative correlation was observed between crude fiber content (CF) and moisture content (MC). Although weak, the correlation indicated an inverse relationship between the variables. Thus, as the grain mass dried, lower water contents resulted in higher crude fiber levels. According to Nunes et al. [24], the higher CF content may be related to the increase in compounds in the cell wall, into structures such as cellulose and hemicellulose, providing greater stiffness to the grain. Thus, rice grains with higher CF in their composition were less physically affected by mechanical processing operations. Ash contents (AS) showed positive correlations with CF × CP × YIE. Ash contents (AS) were considerably reduced in the polishing process of rice grains with higher whole grain yield (YIE). According to Cecchi [27], the AS corresponds to the inorganic residue that remains after the burning of organic matter, consisting mainly of large amounts of K, Ca, Na, and Mg.

Table 4. Machine Learning models applied to physicochemical quality of whole rice grains with different initial moisture contents.

Models	r	MAE	R2	r	MAE	R2
	Starch (ST)			Ashes (AS)
MLR	0.8169	0.6235	0.6674	0.2596	0.0700	0.0673
ANNs	0.8251	0.7657	0.6808	0.5125	0.0621	0.2626
M5P	0.9613	0.4299	0.9241	0.4609	0.0636	0.2124
RF	0.9758	0.6594	0.9522	0.4609	0.0636	0.2124
REPTree	0.9570	12.591	0.9160	0.5160	0.0620	0.2663
RandTree	0.9456	13.227	0.8942	0.5160	0.0620	0.2663
	Crude Fiber (FB)			Crude Protein (CP)
MRL	0.3488	0.0324	0.1217	0.0404	0.4557	0.0016
RNAs	0.8118	0.0204	0.6590	0.3461	0.4108	0.1198
M5P	0.7805	0.0192	0.6091	0.0651	0.4547	0.0042
RF	0.7913	0.0175	0.6261	0.7246	0.3185	0.5250
REPTree	0.8391	0.0178	0.7041	0.2029	0.4889	0.0411
RandTree	0.8228	0.0178	0.6770	0.8614	0.1520	0.7421
	Fat (Fat)			Apparent Specific Mass (ASM)
MRL	0.2065	0.1017	0.0426	0.1830	21.4990	0.0335
RNAs	0.2266	0.1081	0.0513	0.4278	20.1854	0.1830
M5P	0.0500	0.0981	0.0345	0.1830	21.4990	0.0335
RF	0.6490	0.0614	0.4212	0.5183	15.7880	0.2687
REPTree	0.4683	0.0846	0.2193	0.3920	18.7946	0.1540
RandTree	0.7325	0.0415	0.5366	0.4886	15.5575	0.2387

Pearson’s correlation coefficient (r), mean absolute error (MAE) and coefficient of determination (R²) for Machine Learning models: Artificial Neural Network (ANN), Decision Tree (REPTree), Random Tree (RandTree), Quinlan’s M5 algorithm (M5P), Random Forest (RF), and Multiple Linear Regression (MLR).

shows the results of the observed and estimated grain quality values for the different Machine Learning models, while Figure 4 illustrates the potential results of the models for predicting moisture (MC) and starch (ST), ash (AS), and crude fiber (CF) contents in whole grain rice (YIE). In the prediction of starch (ST) as a function of MC and YIE, similar correlation coefficients were observed for all models.

RF model showed the highest correlation (r > 0.97), followed by the M5P and REPTree models. However, the lowest MAE was archived by the M5P model (MAE < 0.5), followed by the RF model. Given this, RF and M5P were the most suitable for predicting the starch content (ST) in whole rice grains. The RF model has wide applicability in the agricultural industry. The efficiency and versatility of RF were evidenced by Zeymer et al. [28], who satisfactorily predicted dry matter loss in soybeans as a function of water content and storage time. Furthermore, Ramos et al. [9] verified the great ability of the RF model to predict soybean plant height through spectral bands.

The RandT, REPTree, and ANNs models showed the highest correlations and lowest errors for observed and predicted starch contents (ST) (0.51 and 0.06), respectively. Furthermore, the M5P and RF models also showed similar fits to the other models. Despite the low mean absolute error, the correlation was considered low (less than 0.7), and for this reason, the models studied are not the most suitable for predicting the influence of MC and YIE on ash contents (AS). Similar fit patterns were found for all ML models used to predict CF, except for the conventional MLR model. Random Tree, REPTree and RF models showed correlation coefficients around 0.8 and MAE around 0.016. Thus, the three models were suitable for predicting crude fiber (CF) in whole rice grains, with the REPTree model standing out.

Figure 5 shows the performance of the MLR model to predict CP, Fat, and ASM. The Random Tree model demonstrated a better fit for predicting the interference of moisture content (MC) on crude protein (CP) levels. However, the RF model showed greater potential to predict the same variable, with a correlation coefficient higher than 0.72 and MAE around 0.32.

Among the models studied, the Random Tree archived the highest correlation coefficient and the lowest mean absolute error and hence is suitable for predicting the effect of MC and YIE on fat levels (Fat) in whole rice grains. These findings are supported by Walter et al. [29], who reported a decreased lipid concentration in the grain milling process as they were present in different layers of the grain, even associated with the starch granules. Likewise, Müller et al. [1] noted a progressive decrease in the lipid content on the surface of whole-milled rice grains.

When analyzing apparent specific mass, all ML models achieved low Pearson’s correlation coefficients (r < 0.52), in which the RF model outperformed the others. Similarly, all models presented high MAE between observed and predicted ASM values, showing poor fits. Therefore, none of the applied models was suitable to predict the direct relationship between water content (WC) and apparent specific mass (ASM) in whole rice grains. Finally, we verified that the Random Tree model presented the highest consistency among the models for predicting the variables studied, even with correlations lower than 0.7. In the evaluation of ash (AS) and ASM), the Random Tree model remained among the best models. Therefore, this Random Tree, which is based on random choices in the attribute tree, has high potential to predict the physicochemical variables of whole grain rice grains under different initial water content (MC) and whole grain yield (YIE).

3.2. Defective Rice Grains

The higher the initial water contents (WC) of the grains, the higher the percentages of defects obtained in the processing after drying (

Table 5. Physical and physicochemical quality of rice grains with defects in function of moisture content.

Moisture Content (% d.b.)	Grain Defects (%)	Crude Protein (%)	Fat (%)	Crude Fiber (%)	Aches (%)	Starch (%)
16	0.768	10.77	2.09	2.48	1.65	65.46
16	0.798	11.49	3.34	2.35	1.75	62.64
16	0.816	11.52	3.24	2.57	2.01	61.51
16	0.816	11.52	3.24	2.57	2.01	61.51
16	0.858	11.59	3.81	3.07	2.22	60.9
16	0.861	10.48	2.22	2.70	1.75	63.9
16	0.871	11.17	3.38	2.59	1.76	64.01
16	0.880	11.07	3.68	2.73	1.63	63.40
16	0.960	11.72	3.46	3.00	2.08	62.17
16	0.969	11.37	2.93	2.74	1.82	62.95
16	1.009	11.19	1.97	2.63	1.78	63.03
16	1.024	11.60	2.67	2.97	2.07	64.39
Average	0.866 d	11.43 a	3.24 a	2.665 c	1.80 b	62.99 a
Standard deviation	0.0815	0.3562	0.5998	0.2090	0.1832	1.2795
17	1.026	11.65	3.34	2.60	1.87	62.62
17	1.269	11.00	3.78	2.96	1.74	61.5
17	1.295	11.73	3.46	2.67	1.97	61.73
17	1.307	11.89	2.72	2.51	1.75	62.13
17	1.332	12.41	3.25	2.87	2.41	60.08
17	1.385	10.84	2.25	2.73	1.94	65.38
17	1.442	11.06	1.87	2.51	1.89	64.64
17	1.528	10.76	2.38	2.60	2.00	63.89
17	1.528	10.76	2.38	2.60	2.00	63.89
17	1.549	11.35	3.51	2.88	2.08	62.28
17	1.555	11.02	2.06	2.57	1.81	65.79
17	1.663	11.26	3.09	2.63	1.91	62.29
Average	1.4135 c	11.16 a	2.905 b	2.615 c	1.925 a	62.455 a
Standard deviation	0.1661	0.4925	0.6146	0.1439	0.1706	1.6338
18	1.713	11.18	3.46	2.53	1.69	62.71
18	1.713	11.18	3.46	2.53	1.69	62.71
18	1.953	10.81	2.27	2.59	1.81	64.01
18	1.966	12.07	3.73	3.05	2.24	59.86
18	1.966	12.07	3.73	3.05	2.24	59.86
18	2.094	12.20	3.32	2.87	2.01	60.03
18	2.195	10.98	3.30	2.68	1.87	63.43
18	2.380	11.79	3.02	2.86	1.99	61.85
18	2.408	11.37	2.97	2.91	1.94	62.39
18	2.420	11.66	3.76	2.88	2.04	59.41
18	2.433	10.70	2.94	2.60	1.86	64.36
18	2.444	12.15	3.16	2.87	2.03	60.99
Average	2.1445 b	11.515 a	3.31 a	2.865 a	1.965 a	62.12 a
Standard deviation	0.2664	0.5227	0.4094	0.1822	0.1726	1.6616
19	2.799	12.05	3.84	3.10	2.33	59.11
19	2.895	11.23	2.97	2.77	2.05	61.79
19	3.063	11.99	3.45	2.46	2.06	59.49
19	3.167	11.41	3.35	3.11	1.99	61.02
19	3.211	12.25	2.84	2.67	1.90	62.26
19	3.293	12.66	2.98	2.63	1.76	60.88
19	3.617	11.25	2.77	2.76	1.79	63.38
19	3.645	12.43	3.25	2.57	1.95	61.99
19	4.079	10.88	2.73	2.96	1.94	63.00
19	4.213	10.32	4.31	4.85	2.04	58.36
19	5.692	12.13	2.70	2.38	2.13	60.09
19	5.704	11.66	2.51	2.61	1.89	62.62
Average	3.455 a	11.825 a	2.975 b	2.715 b	1.97 a	61.405 c
Standard deviation	0.9520	0.6604	0.5042	0.6264	0.1475	1.5496

Means followed by the same letters do not differ by the Scott–Knott test at 5% probability.

).

In the Figure 6 and

Table 6. Coefficients of the associations between the variables (Pearson’s correlation)—whole rice grains—defective rice grains.

Variables	MC	GD	CP	FAT	CF	AS	ST
MC	1	0.87568	0.28664	0.16684	0.22435	0.21161	−0.43199
GD	0.87568	1	0.23186	0.03914	0.21896	0.18746	−0.36516
CP	0.28664	0.23186	1	0.29798	−0.15680	0.45045	−0.57519
FAT	0.16684	0.03914	0.29798	1	0.49972	0.35167	−0.72743
CF	0.22435	0.21896	−0.15680	0.49972	1	0.36886	−0.47669
AS	0.21161	0.18746	0.45045	0.35167	0.36886	1	−0.61148
ST	−0.43199	−0.36516	−0.57519	−0.72743	−0.47669	−0.61148	1

, it is possible to see a positive and strong correlation between the input variables physical defects (GD) and moisture content (MC). Nunes et al. [24] reported that high moisture contents negatively affect the quality of stored rice grains due to increased metabolic activity and increased percentages of physical defects at the end of storage time. Exposing the grain mass to longer drying time left the grains more susceptible to breakage during mechanical processing operations.

Starch content (ST) showed a negative correlation with MC which, although weak, indicated an inverse relationship. Walter et al. [29] reported that drying and storage interfere with the ST of rice. Moreover, ST also showed a very weak negative correlation with physical defects (GD), indicating an inverse relationship between the variables. According to Scariot et al. [30], high MC can influence the formation of chalky grains, which are considered defects by the industry due to the opaque appearance and interference in the cooking of the product, caused by the non-compaction of the starch and protein granules arrangement in the grains that form air spaces between them, resulting in diffraction of the incident light. Chalky conditions reduce the hardness of the grain, making it more fragile to the polishing operation and leading to grain breakage, reducing the physical and chemical quality of the product, which justifies the negative correlation between ST and the percentage of grains with defects (GD) observed in the correlation network.

Starch contents (ST) and the other variables had a medium negative correlation with Fat. There is also a weak negative correlation between ST and ash (AS) contents, corresponding to an inverse relationship between them for rice grains with physical defects (GD). Under the presence of moisture, starch granules expand due to diffusion and absorption, and this procedure is reversible through the drying process of the grain. However, besides altering the starch granules, there may be changes in macronutrients such as lipids and proteins, generating impacts on the physicochemical properties of the grain [26]. Overall, increasing starch content (ST) also interfered inversely with grain mass yield due to susceptibility to the occurrence of physical defects (GD) [1,27]. Furthermore, a positive correlation was observed between Fat × CP and Fat × CF contents, indicating a direct relationship between the variables. This influence is justified by the direct relationship between lipid concentration and grain hardness, which is indirectly associated with the protein and fiber contents present in the physicochemical constitution of the grains. The crude protein content (CP) directly contributed to maintaining the lipid layers in the grains, and its decrease implies the reduction of the fat content [31,32].

Moisture contents (MC) and percentages of physical defects (GD) also showed a weak positive correlation with CP, indicating a direct relationship between them. Nunes et al. [24] observed that rice grains less exposed to high drying temperatures had lower percentages of broken grains and consequently higher crude protein contents. According to Lima et al. [33], high moisture content increases the respiration rate of the grain mass causing oxidation and, as a result, the loss of total carbohydrates, starch, proteins, and other physicochemical components of the grains.

Table 7. Machine Learning models applied to physicochemical quality of rice grains with defects for different initial moisture contents.

Models	r	MAE	R2	R	MAE	R2
	Ashes (AS)			Crude Fiber (CF)
MLR	0.0555	0.1511	0.0030	0.2546	0.2276	0.0648
ANNs	0.0309	0.1575	0.0009	0.3639	0.2218	0.1324
M5P	0.0555	0.1511	0.0030	0.7904	0.2033	0.6247
RF	0.8790	0.0586	0.7726	0.9267	0.1053	0.8588
REPTree	0.5787	0.1153	0.3348	0.9128	0.1437	0.8333
RandTree	0.8449	0.0443	0.7138	0.9184	0.0842	0.8434
	Fat (Fat)			Crude Protein (CP)
MLR	0.1785	0.4664	0.0318	0.3574	0.4531	0.1278
ANNs	0.3430	0.2847	0.1177	0.6615	0.5165	0.4376
M5P	0.3548	0.5056	0.1258	0.6678	0.4007	0.4459
RF	0.9221	0.1731	0.8504	0.7317	0.2793	0.5355
REPTree	0.6133	0.3434	0.3762	0.7462	0.3060	0.5568
RandTree	0.9640	0.0757	0.9292	0.5577	0.2814	0.3110
	Starch (ST)
MLR	0.2063	1.4960	0.0425
ANNs	0.2589	1.4880	0.0670
M5P	0.2063	1.4960	0.0425
RF	0.7096	0.7586	0.5036
REPTree	0.5300	1.0770	0.2809
RandTree	0.7540	0.5515	0.5686

Pearson’s correlation coefficient (r), mean absolute error (MAE), and coefficient of determination (R²) for Machine Learning models: Artificial Neural Network (ANN), Decision Tree (REPTree), Random Tree (RandTree), Quinlan’s M5 algorithm (M5P), Random Forest (RF), and Multiple Linear Regression (MLR).

shows the correlation coefficient (r), mean absolute error (MAE), and coefficient of determination (R²) between the observed and estimated values of rice grain quality with defects for the different ML models.

Fits obtained by the ML models are shown in Figure 7. The decision tree (REPTree) and Random Forest (RF) models presented the highest correlation coefficients between the observed and predicted variables for CP, and the lowest MAE was observed for the RF model. Thus, both models are suitable for predicting crude protein levels in rice grains with physical defects. The Artificial Neural Networks (ANNs) model obtained the highest MAE for predicting the CP content in rice grains with defects, not being recommended for the prediction of this variable.

For the crude fiber (CF) variable, the RF model showed the highest correlation coefficient (r > 0.92), followed by the Random Tree (RandT) and REPTree models, with r above 0.90. Conversely, among the highlighted models, the lowest MAE was observed for the Random Tree model (RandT), which was lower than 0.085. Given the observed variations, the three models can be indicated to predict the influence of MC on CF. ANN model presented the lowest r and the highest MAE.

Regarding ST prediction (ST), the highest r and the lowest MAE were observed for the Random Tree (RandTree) model (around 0.75 and 0.55, respectively). Additionally, the RF model showed a similar fit to the Random Tree, with r around 0.7 and MAE of 0.75. REPTree, Artificial Neural Networks (ANNs), and M5P models did not provide good prediction fits, with correlation coefficients lower than 0.53.

The fit parameters obtained by ML models for the variables fat (Fat) and ash (AS) are shown in Figure 8. The Random Tree model showed the best fit (r > 0.96 and MAE < 0.076), being indicated to predict the fat contents in rice grains with defects. Likewise, the Random Forest also achieved a high correlation (r > 0.92) [34], and a low MAE between the observed and predicted fat content values, being indicated for predicting the fat content in rice grains with physical defects.

For ash content (AS), the RF model showed the highest correlation coefficient, followed by the Random Tree (RandTree), with correlation coefficients of 0.87 and 0.84, respectively. However, the lowest MAE was found for the Random Tree model (MAE < 0.045), followed by the Random Forest. The increased ash content (AS) is a result of the organic fraction degradation of the grains due to the metabolic activity arising from the presence of water [35,36,37].

Random Tree (RandT) decision tree model presented the best fit to predict the physicochemical variables of rice grains with physical defects as a function of different initial moisture content (MC). Thus, Random Tree is the most suitable among the models studied [38]. The random choice among the attributes present in the tree, which is the property of this model, allowed its constancy in relation to the others studied [39,40].

4. Conclusions

The combination of the non-destructive technology Near-Infrared Spectroscopy and the Machine Learning models characterized successfully the physicochemical composition of whole and defective rice grains, being an alternative to the conventional method of physical classification. The Random Tree model (RandT) was the indicated model to predict the physicochemical quality in whole and defective rice grains for different moisture contents, obtained Pearson’s correlation coefficient (r = 0.96), mean absolute error (MAE = 0.017), and coefficient of determination (R² = 0.92). The use of near-infrared (NIR) spectroscopy evaluation methods and machine learning models can ensure greater precision, robustness, and agility in evaluating the quality of rice samples in the processing and storage units to reduce subjective errors in manual and visual physical classification.