Establishing a Prediction Model for Tea Leaf Moisture Content Using the Free-Space Method’s Measured Scattering Coefficient

Abstract

The rapid and nondestructive detection of tea leaf moisture content (MC) is of great significance to processing tea with an automatic assembly line. This study proposes an MC detection method based on microwave scattering parameters (SPs). Through the established free-space electromagnetic measurement device, 901 different frequency points are taken between 2.45 and 6 GHz using a vector network analyzer (VNA). The SPs of tea leaves with different moisture contents (5.72–55.26%) at different bulk density and different sample thicknesses were measured. The relationship between frequency, S₂₁ amplitude and moisture content, thickness, and bulk density of tea was analyzed using correlation coefficients, significance analysis, and model construction. Back propagation (BP) neural network, decision tree (DT), and random forest (RF) MC prediction models were established with the frequency, amplitude, and phase of the SPs, thickness, and bulk density of the samples as inputs. The results showed that the RF-based model had the best performance, with determination coefficient (R²) = 0.998, mean absolute error (MAE) = 0.242, and root mean square error (RMSE) = 0.614. Compared to other nondestructive testing processes for tea, this method is simpler and more accurate. This study provides a new method for the detection of tea MC, which may have potential applications in tea processing.

1. Introduction

Moisture content (MC) is an important index of tea quality [1]. In most tea processing methods, the last step of drying plays a decisive role in the tea’s MC. At present, tea processing mainly depends on the subjective experience of the producer or the fixed parameters of the processing machine. The former consumes manpower and cannot be quantified, while the latter ignores the different initial conditions of each batch of tea. In this case, dried tea has a high rate of tea damage caused by excessive drying, or the drying is insufficient and the MC is too high, which necessitates repeated drying and affects the taste or causes mold to breed. Therefore, rapid and nondestructive MC detection before tea drying may help the dryer automatically adjust the specific parameters during operation to improve the production efficiency and quality.

At present, the MC detection technology used for processing tea is mainly divided into the direct method and the indirect method [2]. The direct method determines the MC of the sample by weighing the moisture loss of the sample at a certain temperature [3,4]. The direct method usually includes the constant-temperature oven method and the rapid moisture determination method. The advantage of this method is that the MC is measured accurately. The disadvantage is that the determination period is long, and the determination process is destructive to the sample, so it is only widely used in a laboratory setting. There has been much research into the indirect method, including capacitance [5], machine vision [6,7], near-infrared spectroscopy [8,9], hyperspectral [10], microwave, and so on. The capacitance method needs to be calibrated before each measurement, and the measured material needs to be in full contact with the electrode plate to ensure accurate measurement. Thus, it is only suitable for measuring lower water content. Machine vision has high requirements of fixed light. When collecting information, it can only focus on the information on the tea’s surface, ignoring the internal state of the tea. Due to its weak penetration, the reflected light received by the near-infrared spectrum can only provide the moisture distribution on the surface of the tea, losing the internal distribution. Although hyperspectral imaging technology can extract rich information, it is not convenient for rapid tea water content measurements because it needs to be in a specific experimental space over a long period of time during each process.

Due to higher quality penetration and the advantages of nondestructive and rapid detection, the microwave measurement method is gradually being applied in crop MC detection [11,12,13,14,15]. Shivling et al. [16] developed a patch antenna based on microwave resonance technology. At a frequency of 5 GHz, it was observed that the change in tea MC would cause a change in resonance frequency and S₁₁ parameter regularity, and the linear model parameters were calculated to be suitable for the prediction of tea MC. However, the authors only studied tea with a low MC (2–10%). Based on the microwave free-space method, Wu et al. [17] optimized the microwave attenuation and phase shift characteristic signals at 2–10 GHz frequency using an ant colony optimization algorithm. The three main models of support vector regression (SVR), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP) were used to predict the MC of tea, which ranged from 16.25% to 77.65%. The determination coefficient (R²) result was 0.994 and the mean absolute error (MAE) result was 0.349, representing a significant improvement in the accuracy of the model. However, unlike other crops, tea has different stacking densities in a fixed space due to its own characteristics. Authors have not paid sufficient attention to the influence of tea stacking density on microwave signals. Yigit et al. [18] used a vector network analyzer based on the microwave free-space method to collect the scattering parameters (SPs) of three crops at different frequency points ranging from 1–2.48 GHz. Three machine learning algorithms, K-nearest neighbor (KNN), SVR, and artificial neural network (ANN), were established to predict the MC of grains by using the data set constructed from the frequency, S₁₁, S₂₁, and MC. The method used by these authors skips the conventional steps of calculating the dielectric constant and uses the SPs to directly obtain the MC of the crop. The problem of phase ambiguity is thus avoided and the measurement steps are simplified. However, the authors did not study the feasibility analysis of crops with complex surfaces, measured thicknesses, and a high MC.

To address the limitations of current tea MC detection methods, this study used the free-space method to explore the effects of different green tea MC levels and related factors on the SPs. The SPs were measured with a vector network analyzer (VNA), and the extracted data were modeled using back propagation (BP), random forest (RF), and decision tree (DT) prediction models to establish the levels of tea MC. This method skips the measuring instrument calibration process and exerts the influence of the device on the measurement data into the learning algorithm. Different from the previous method of combining the MC with the dielectric constant, the MC of tea using this method is instead measured by establishing the relationship between the MC and the SPs. At the same time, the influences of thickness and bulk density are added to improve the accuracy of the model to a certain extent. This method can measure the MC of tea faster and more accurately than other current methods, signifying its value in various applications.

2. Materials and Methods

2.1. Principle of Measurement

When the microwave signal passes through a sample, the sample produces a polarization reaction. When measuring tea, this reaction consumes some of the energy; the water molecules consume the most, and the microwave attenuation caused by this will change the SPs [19], depending on the MC of the tea. This principle is shown in Figure 1. The expressions of the parameters are listed in formulas (1)–(6).

$$b_1=S_{11}{a}_1+S_{12}{a}_2$$

(1)

$$b_2=S_{21}{a}_1+S_{22}{a}_2$$

(2)

$$S_{11}=\frac{b_1}{a_1}|{}_{a_2=0}$$

(3)

$$S_{21}=\frac{b_2}{a_1}|{}_{a_2=0}$$

(4)

$$S_{22}=\frac{b_2}{a_2}|{}_{a_1=0}$$

(5)

$$S_{12}=\frac{b_1}{a_2}|{}_{a_1=0}$$

(6)

where $a_1$ and $b_1$ are port incident and reflected waves; $a_2$ and $b_2$ are port 2 incident and reflected waves. $S_{11}$ and $S_{22}$ are reflection coefficients; $S_{12}$ and $S_{21}$ are transmission coefficients.

2.2. Sample Preparation

In this experiment, Rizhao green tea was used as the experimental sample, and the specific variety was Xueqing. The MC of the tea samples was 5.72% according to the drying method; they were divided into several portions of 2 kg each and placed in a sealed bag for use. Since the MC of tea before drying is approximately 55%, the MC range was set to 5.7–55.2% [20,21] and the MC gradient was set to about 2%. The MC formula was used to calculate the amount of distilled water that needed to be sprayed on each sample [22]. After distributing the water, the samples were placed in a refrigerator for 48 h. During this period, the samples were taken out at intervals (8 h) to be shaken to ensure that the water could be evenly absorbed. The MC of the wet basis is shown in formula (7).

$$\mathrm{MC}=\frac{M_{\mathrm{t}}-M_{\mathrm{d}}}{M_{\mathrm{t}}}\times 100\%$$

(7)

where the initial mass of the tea samples is $M_{\mathrm{t}}$, the mass of the dried tea samples is $M_{\mathrm{d}}$, and the set moisture content is MC.

2.3. Experimental Installation

The measuring device is shown in Figure 2. A pair of double-ridge pyramid horn antennas operating at a range of 2–16 GHz are fixed on an aluminum frame with a width of 100 cm. The photosensitive resin structure lens is installed between the antennas to allow the microwaves to propagate as plane waves in free space. The data acquisition equipment is a Keysight P9371A USB VNA; a computer (Lenovo (Beijing) Co., Ltd., Beijing, China); an MB45 halogen moisture analyzer (Shanghai Ohaus Instruments Ltd., Shanghai, China); a YH-type electronic balance (Shanghai Yingheng Weighing Co., Shanghai, China); and an acrylic box with dimensions of 4 cm × 30 cm × 30 cm, 7 cm × 30 cm × 30 cm, 10 cm × 30 cm × 30 cm. Additionally, there is a sealing bag, disposable gloves, a spray kettle, an aluminum plate, etc.

2.4. Experimental Design

In this experiment, the thickness, MC, and bulk density of the tea samples were used as dependent variables, and their influence on the SPs of tea (the independent variables) were explored. Before the measurement and after the samples were taken out of the refrigerator, due to the temperature difference, it was necessary let them stand for 2 h and shake them regularly (30 min) during the period to ensure uniform moisture, which was carried out at room temperature (20 °C). As shown in Figure 3, 3 g samples (±0.005 g) were selected from each sample and placed in the MB45 halogen moisture tester. The working mode was set to 100 °C, and the work was stopped when the water loss was less than 0.001 g within 90 s. Each sample was tested five times, and the average was taken as the MC value of the sample. During the measurement, the measurement frequency of the VNA was set to 2.45–6 GHz, and 901 data points were collected each time for the amplitude and phase of the SPs to construct a data set. A total of 16 MC samples were collected; each MC corresponds to three thicknesses (4 cm, 7 cm, and 10 cm, respectively) and three bulk densities (0.198 g/cm³, 0.218 g/cm³, and 0.238 g/cm³, respectively), 144 experiments, and 129,744 data sets. The moisture content data are shown in

Table 1. Tea MC grouping.

Number	Moisture Content (%)	Number	Moisture Content (%)
1	5.72	9	30.27
2	11.18	10	32.58
3	15.08	11	34.29
4	17.76	12	40.82
5	20.35	13	45.24
6	22.74	14	49.56
7	25.07	15	52.54
8	27.43	16	55.26

. Figure 4 is the research route flowchart of this paper.

3. Prediction Model

3.1. Back Propagation Neural Network Model

A BP neural network is a neural network trained according to the error back propagation algorithm. It has strong nonlinear mapping abilities and can realize a strict mathematical mapping relationship or formula between uncertain input and output data [23,24]. Based on the premise that the network itself learns and trains the data, the nonstrict mapping rules between the input data and the output data are obtained. Given the input data, the prediction of the output data results is realized. This algorithm is most widely used in the field of neural networks and its structure is generally composed of an input layer, a hidden layer, and an output layer. There is no corresponding mathematical formula to determine the number of neurons in the hidden layer. In practical engineering applications, the hidden layer is usually determined according to empirical formula (8) [25]. In this study, the number of hidden layers was between 1 and 13. By comparing the results of different configurations, the final number was determined to be 13.

$$Z=\sqrt{n+m}+a$$

(8)

where $Z$ is the number of neurons in the hidden layer of the neural network, $n$ is the number of nodes in the input layer, $m$ is the number of nodes in the output layer, and $a$ is the adjustment coefficient between 1 and 10.

3.2. Decision Tree Model

A DT is an important classification regression method in data mining technology. CART DT is a typical algorithm in DT, which is generally called a classification regression tree. It is an efficient and accurate classification method [26]. The main construction process involves generating and pruning a DT. Generating a CART DT is the process of recursively constructing a binary DT, which can be used for both classification and regression. For a classification tree, CART uses the Gini coefficient minimization criterion; for a regression tree, the square error minimization principle is used to select features and generate a binary tree. For decision tree pruning, eigenvalue attributes can be used multiple times, and continuous and missing values can also be effectively processed. Therefore, due to the characteristics and advantages of the CART decision tree algorithm, this paper applies it to the importance judgment and classification of index variables. The CART regression tree generation algorithm is as follows:

$$f_{(x)}={\displaystyle \sum }_{m=1}^M{c}_{m}I\left(x\in R_m\right)$$

(9)

where the $f$ represents the regression tree value of the model output; $M$ represents the $M$ regions divided by the input space; $R$ represents the unit data set; $m$ is the number of elements that divide the input space, that is, $R_1$, $R_2$,..., $R_m$; $I$ represents a constant; and $c_m$ is the fixed output value of each unit, so that the error range between the output value and the actual value can be obtained.

The optimal output value on each unit is solved by the criterion of minimum square error:

$${\widehat{c}}_m=\mathrm{ave}\left(y_i\right|x_i\in R_m)$$

(10)

where $\overrightarrow{y}$ represents the output variable and x represents the input variables.

For the division of space, a heuristic method is used and two regions are defined:

$$R_1(j,s)=\left\{x\mid x^{(j)}\le s\right\}$$

(11)

$$R_2(j,s)=\left\{x\mid x^{(j)}>s\right\}$$

(12)

where $j$ represents the segmentation variable and $s$ represents the value corresponding to the segmentation variable, that is, the segmentation point. $R_1$ and $R_2$ represent the feature space after partition.

Firstly, the optimal segmentation variable (optimal initial index) and optimal segmentation point (optimal initial index value) are determined in order to find the representative values of the two regions $c_1$ and $c_2$ (according to the optimal initial index and the optimal initial index value) and minimize the square difference on their respective intervals. The solving process is as follows:

$${\min }_{j,s}=\left[{\min }_{c_1}{\displaystyle \sum _{x_1\in R_1(j,s)}{\left(y_i-c_1\right)}^2}+{\min }_{c_2}{\displaystyle \sum _{x_2\in R_2(j,s)}{\left(y_i-c_2\right)}^2}\right]$$

(13)

Secondly, all input variables (initial indicators) are traversed to find the optimal segmentation variable (optimal initial indicator), and the input space is divided into two regions (two nodes) in turn:

$${\widehat{c}}_1=\mathrm{ave}\left(y_i\mid x_i\in R_1(j,s)\right)$$

(14)

$${\widehat{c}}_2=\mathrm{ave}\left(y_i\mid x_i\in R_2(j,s)\right)$$

(15)

Finally, the above partitioning process is repeated for each region until the stopping condition is satisfied to generate a regression tree, which is also called the least squares tree.

3.3. Random Forest Model

RF is an ensemble learning algorithm based on DT [27] and can be roughly divided into a classification algorithm and a regression algorithm; this paper only uses the latter, the principle of which is based on a combination model composed of a set of regression decision subtrees.

$$\left\{h\left(x,{\theta }_t\right),t=1,2,\dots ,T\right\}$$

(16)

where ${\theta }_t$ is an independent and identically distributed random variable, $x$ is an independent variable, and $T$ is the number of DT.

Using the idea of ensemble learning, the mean of each DT $\left\{h\left(x,{\theta }_t\right)\right\}$ is taken as the regression prediction result:

$$\overline{h}(x)=\frac{1}{T}{\displaystyle \sum _{t=1}^T\left\{h\left(x,{\theta }_t\right)\right\}}$$

(17)

where $\left\{h\left(x,{\theta }_t\right)\right\}$ is the output based on $x$ and $\theta$. In order to overcome the phenomenon of low accuracy and overfitting that may occur in the DT model, the RF model introduces the concepts of a bagging algorithm and random subspace. When the random forest model performs regression calculations, a set of hierarchical rules is used to divide the random subspace so that the data set is recursively grouped based on similar instances. A set of covariates is used to recursively split the values of the variables of interest, resulting in multiple parent nodes and child nodes similar to the tree structure. The candidate variables that were randomly selected from the total number of covariates at each node are called mtry parameters. The DT will evaluate each candidate to find the best split for maximizing purity, so that the error range of each child node is minimized. The RF model generates different DTs from different subsets of the training data through the bagging algorithm, which can avoid the influence of the possible multicollinearity between the independent variables on the model results. In this paper, when the RF model performs best, the number of mtry parameters is 11 and the number of decision trees is 100.

4. Results and Discussion

4.1. Effects of Variables on Tea SPs

This section analyzes the influence of tea MC, thickness, and bulk density on the measurement of scattering parameters. The correlation between each factor and MC was determined using the Pearson correlation coefficient and significance analysis (the significance level was 0.01).

Table 2. The correlation and significance between SPs and MC.

Feature	Correlation Coefficient	Significance	Feature	Correlation Coefficient	Significance
S11 (dB)	0.009	0.263	S12 (dB)	0.819	0.000
S11 (DEG)	0.010	0.195	S12 (DEG)	0.006	0.028
S21 (dB)	0.819	0.000	S22 (dB)	0.009	0.001
S21 (DEG)	0.018	0.022	S22 (DEG)	0.000	0.871

shows that among the SPs, the correlation coefficient with MC was the largest, and the most significant were the amplitudes of S₂₁ (S_21A) and S₂₂ (S₂₁ = S₁₂, so only S_21A is analyzed). Furthermore, Table 2 shows the correlation coefficient and significance analysis of each parameter and the MC.

Table 3. The correlation and significance between various factors and MC.

Feature	Correlation Coefficient	Significance
Frequency	0.000	0.986
S21 (dB)	0.773	0.000
Thickness	0.456	0.000
Bulk density	0.161	0.000

shows that there was a definite correlation between thickness, bulk density, and moisture content, and it was more significant.

4.1.1. Effect of Tea MC on SPs

The prepared tea leaves with 16 MC gradients were placed in a test box with a thickness of 7 cm, and the bulk density was controlled at approximately 0.218 g/cm³. The amplitude and phase values of SPs were measured at a room temperature of 20 °C. Taking the frequencies of 2.5 GHz, 3.6 GHz, 4.7 GHz, and 5.8 GHz as examples, Figure 5 shows each measurement result.

Figure 5 shows that the relationship between S_21A and the MC is the strongest. Table 2 also confirms this view. The essential reason for this is that when the microwaves penetrate the tea samples, they cause an energy loss and phase shift, and the most relevant transmission coefficient changes accordingly. Water molecules have the greatest influence on microwaves in tea, so the higher the MC, the greater the microwave loss, and the smaller the S_21A.

4.1.2. Effect of Thickness on the Measurement of Tea S_21A

At room temperature (20 °C), the bulk density of tea was controlled at 0.218 g/cm³, and the tea was placed in test boxes with a thickness of 4 cm, 7 cm, and 10 cm, respectively, to measure the SPs. Taking the frequencies of 2.5 GHz, 3.6 GHz, 4.7 GHz, and 5.8 GHz as examples, Figure 6 shows the measurement results. The results show that the three gradients of tea MC and the thickness of the sample had a significant effect on the measurement of S_21A. With the increase in sample thickness, the S_21A decreased, and this effect continued to expand with the increase in MC. Therefore, when establishing an MC prediction model, the tea sample thickness should be regarded as a fixed quantity or variable.

4.1.3. Effect of Bulk Density on the Measurement of Tea S_21A

At room temperature (20 °C), three gradients of tea (0.198 g/cm³, 0.218 g/cm³, and 0.238 g/cm³) were placed in the 10 cm thick test box to measure the SPs. Taking the frequencies of 2.5 GHz, 3.6 GHz, 4.7 GHz and 5.8 GHz as examples, Figure 7 shows the measurement results. The results show that the three gradients of tea MC and the bulk density of the sample had a significant effect on the measurement of S_21A. As the bulk density of the sample increased, the S_21A value decreased, and this effect continued to expand as the MC increased. Therefore, the bulk density of tea samples should be regarded as a fixed quantity or variable when establishing an MC prediction model.

4.2. Modeling and Analysis

In order to further determine the relationship between thickness, bulk density, and SPs and the influence of tea moisture content on the prediction model, the 129,744 sets of data were divided into a training set and test set according to a ratio of 8:2. With frequency and SPs as input, the thickness and bulk density were retrieved, respectively. The mean absolute percentage error (MAPE) and R² were used as evaluation indexes. In order to establish the best prediction model of tea MC, the thickness and bulk density were modeled as fixed quantities and variables. In modeling them as fixed quantities, taking the thickness of 7 cm and the bulk density of 0.218 g/cm³ as an example, 14,416 sets of data were divided into a training set and a test set according to the ratio of 8:2 [28]. SPs and frequency were used as inputs for modeling, and water content was used as the output. In the modeling of thickness and bulk density as variables, 12,9744 sets of data were also divided into a training set and test set according to the ratio of 8:2. SPs, frequency (F), thickness (T), and bulk density (BD) were used as inputs for modeling, and the MC was used as the output. The MAE, root mean square error (RMSE), and R² were used as the evaluation indexes of the model performance [29]. A flowchart of the predictive modeling is presented in Figure 8.

4.2.1. Model Analysis

With frequency and S_21A as inputs, three models of BP, DT, and RF were used to retrieve moisture content, thickness, and bulk density, respectively. The results are shown in

Table 4. Retrieval of thickness and bulk density.

Object	Model	Feature Combination	Evaluation Metrics
			MAPE (%)	R2
Moisture content	RF	F, S21A	15.602	0.848
	DT		16.919	0.889
	BP		23.528	0.717
Thickness	RF	F, S21A	17.477	0.478
	DT		17.959	0.601
	BP		24.684	0.274
Bulk density	RF	F, S21A	6.879	0.464
	DT		7.557	0.585
	BP		9.186	0.009

. The results show that the correlation coefficient between moisture content and frequency and S_21A was the highest, followed by thickness and bulk density.

In the case of modeling thickness and bulk density as fixed quantities, the frequency and scattering parameters were divided into different combinations and input into the BP, DT, and RF models, respectively. The results are shown in

Table 5. The influence of different parameter combinations on the model when the thickness and bulk density are constant.

Model	Feature Combination	Evaluation Metrics
		MAE (%)	RMSE (%)	R2
RF	F, SPs	0.439	0.981	0.995
	F, S21A	0.616	1.441	0.989
	S21A	5.279	6.955	0.763
DT	F, SPs	0.435	1.566	0.994
	F, S21A	0.676	1.895	0.991
	S21A	5.304	7.096	0.869
BP	F, SPs	1.893	2.594	0.967
	F, S21A	2.101	2.848	0.959
	S21A	5.535	7.048	0.694

. Although previous analysis has shown that S_21A is closely related to MC, it can be seen from Table 5 that when S_21A was used as an input alone, the performance of the model was poor. When combining the frequency with water content, the performance of the model was significantly improved. After the remaining SPs were matched as inputs, the performance of the model was slightly improved. The three models showed the best performance under the combination of frequency and SPs. The test set prediction results R² diagram is shown in Figure 9.

In the case of modeling the thickness and bulk density as variables, the frequency, SPs, thickness, and bulk density were divided into different combinations and input into BP, DT and RF models, respectively. Compared to the modeling of bulk density and thickness as fixed quantities, the combination of frequency and S_21A in variable modeling significantly reduced performance. After inputting the data of thickness and bulk density, the performance of the model was significantly improved. At the same time, it can also be seen in

Table 6. The influence of different parameter combinations on the model when the thickness and bulk density are variables.

Model	Feature Combination	Evaluation Metrics
		MAE (%)	RMSE (%)	R2
RF	F, SPs, BD, T	0.242	0.614	0.998
	F, S21A, BD, T	0.461	1.409	0.991
	F, S21A, D	1.595	3.634	0.937
	F, S21A, T	1.311	2.526	0.969
	F, S21A	3.623	5.638	0.848
	S21A	6.858	8.985	0.615
DT	F, SPs, BD, T	0.249	1.160	0.997
	F, S21A, BD, T	0.515	3.419	0.992
	F, S21A, BD	1.625	4.475	0.952
	F, S21A, T	1.410	3.062	0.977
	F, S21A	3.939	6.710	0.889
	S21A	7.218	9.619	0.767
BP	F, SPs, BD, T	2.443	3.260	0.946
	F, S21A, BD, T	2.826	3.774	0.927
	F, S21A, BD	4.812	6.303	0.765
	F, S21A, T	3.051	4.030	0.916
	F, S21A	5.141	6.830	0.717
	S21A	6.132	7.789	0.592

that the thickness generated a greater improvement in the model than the bulk density. This also verifies that the correlation coefficient between thickness and MC was greater than that between the bulk density and MC in Table 3. The three models showed the best performance under the combination of frequency, SPs, bulk density, and thickness. The test set prediction results R² diagram is shown in Figure 10.

4.2.2. Model Evaluation

In the established model, the performance of the RF model is the best regardless of the combination of the input data sets. Compared with BP and DT models, the MAE and RMSE values of the RF model are smaller, and the R² is higher. Therefore, the moisture prediction of the RF model is more accurate and more suitable for the prediction of tea MC.

Table 7. Performance comparison of recently reported methods for nondestructive testing of tea.

Method	MC Range	Model	R2	Reference
Capacitance	3.43–7.49%	VISSA-GWO-SVR	0.970	[5]
Machine vision	53.14-–78.48%	PCA-GA-BP and PCA-PSO-BP	0.941	[7]
Near-infrared	6.36–78.62%	SNV-PCA-GWO-SVR	0.989	[9]
Microwave	16.25–77.65%	MRACO-Stacking	0.994	[17]
Microwave	5.72–55.26%	RF	0.998	This study

lists some nondestructive testing methods reported in recent studies and the performance of microwave-based methods introduced in our study of tea moisture detection. Compared with most nondestructive testing methods, microwave moisture measurements have higher accuracy. By contrast, this method only needs a simple model structure to establish a highly accurate prediction model.

5. Conclusions

In this study, the changes in tea SPs were measured under different MCs, thicknesses, and bulk densities and analyzed using the controlling variable method. A prediction model for measuring the MC of tea from the SPs based on the free-space method was established. The following conclusions were reached:

Under the same thickness and bulk density, the relationship between S_21A and MC in SPs is the strongest (however, the introduction of other SPs will also improve the accuracy of the model, so other parameters cannot be ignored in modeling). The increase in MC will cause a decrease in S_21A, showing a negative correlation; under the same MC and bulk density, the increase in thickness will cause a decrease in S_21A, showing a negative correlation; and under the same MC and thickness, the increase in bulk density will cause a decrease in S_21A, showing a negative correlation. This study addressed the flaw in Wu et al. [17], namely that the authors did not consider the influence of tea density on microwave signals, and also addressed the limitations of Yigit et al. [18], which accounted for neither complex surface crops nor the possible related effects of thickness and bulk density.

In order to accurately predict the moisture content of tea, the moisture content, thickness, and bulk density were retrieved using frequency and S_21A as inputs when constructing the moisture content prediction model. The results show that the R² between moisture content and frequency and S_21A was the highest, followed by thickness and bulk density. Subsequently, the frequency, SPs, thickness, and bulk density were divided into different combinations as inputs, and three models (BP, DT, and RF) were used to establish the MC prediction model. At the same time, the influence of thickness and bulk density as variables and fixed amounts on the performance of the model was explored. The results show that the model was more accurate when the thickness and bulk density were used as variables. Among them, the RF model had the best performance, the MAE value could reach 0.242, the RMSE value was 0.614, and the R² value was 0.998. Compared to the traditional method of converting SPs through complex permittivity, this process is simpler and more effective.