Environmental Simulation Model for Rapid Prediction of Tea Seedling Growth

Abstract

Accurate and effective monitoring of environmental parameters in tea seedling greenhouses is an important basis for regulating the seedling environment, which is crucial for improving the seedling growth quality. This study proposes a tea seedling growth simulation (TSGS) model based on deep learning. The Internet of Things system was used to measure environmental change during the whole seedling process. The correlation between the environmental parameters and the biomass growth of tea seedlings in various varieties was analyzed. A CNN-LSTM network was proposed to build the TSGS model of light, temperature, water, gas, mineral nutrition, and growth biomass. The results showed that: (1) the average correlation coefficients of air temperature, soil temperature, and soil moisture with the biomass growth of tea seedlings were 0.78, 0.84, and −0.63, respectively, which were three important parameters for establishing the TSGS model. (2) For evaluating the TSGS model of a single variety, the accuracy of ZM’s TSGS based on the CNN-LSTM network was the highest (Rp² = 0.98, RMSEP = 0.14). (3) For evaluating the TSGS model of multiple varieties, the accuracy of TSGS based on the CNN-LSTM network was the highest (Rp² = 0.96, RMSEP = 0.17). This study provided effective technical parameters for intelligent control of tea-cutting growth and a new method for rapid breeding.

1. Introduction

Tea plant (Camellia sinensis (L.) O. Kuntze.) is an important cash crop widely planted worldwide. In general, tea plant is a vegetatively propagated crop, and the primary way of propagation is cutting in a seedling greenhouse. However, due to the influence of the natural environment, seedling greenhouses have problems such as difficulty keeping warm in winter and cool in summer, and high humidity leads to diseases. Therefore, the tea seedlings have slow breeding speed, poor ability, and high seedling cost in the northern region of China, which seriously restrict the development of tea seedling industrialization [1].

The growth of tea seedlings is the basis of cultivating strong seedlings and plays a key role in improving the survival rate of tea seedlings. Furthermore, the growth rate of tea seedlings is closely related to the environmental changes in the greenhouse [2]. Light, temperature, humidity, carbon dioxide, and mineral nutrition are the decisive factors for the growth and development of tea seedlings [3,4,5,6]. Accurate and effective monitoring and control of environmental parameters are essential for cultivating strong seedlings. However, the traditional methods of manually monitoring the environment have problems such as low accuracy, low efficiency, and long time span.

With the development of the information perception technology of Internet of Things (IoT), it has gradually become the key technology of the agricultural IoT. The information in the air and soil were acquired through environmental (temperature, humidity, and light) sensors. They were quickly uploaded to the monitoring platform by information transmission technology [7]. At present, IoT information perception technology has been widely applied in tomato [8], millet [9], rice [10], cotton [11], and other crops. For example, the IoT system obtained the environmental parameters in the sugarcane field, and the sugarcane yield model was constructed by machine learning methods [12]. The IoT system obtained the RGB images of weeds in the wild, and deep learning methods constructed an intelligent weed detection model [13]. The IoT system obtained the greenhouse’s environmental parameters then a photosynthetic phenotype analyzed the photosynthesis’s interaction with the fluctuating environment and canopy structure in two seasons [14]. Therefore, the environmental parameters and real-time images of the tea-seeding greenhouse can be real-time monitored by an IoT system.

The rise of deep learning and machine learning provides cutting-edge technologies for agricultural information processing [15,16]. For example, Li et al. [17] proposed a recognition technology of tea plant diseases based on RGB images and deep learning (Mask R-CNN, Four-channel residual network (F-RNet)), which can accurately and quickly identify and warn brown blight, target spot, and tea coal disease. Lu et al. [18] proposed a monitoring technology for the nitrogen nutrition index of wheat based on near-infrared UAV images, and combined three machine learning methods (random forest (RF), support vector machine (SVM), extreme learning machine (ELM)) to evaluate leaves and plant nitrogen content. Li et al. [19] proposed a monitoring technology for important phenotypes of tea plants based on multi-source remote sensing (RGB, multispectral, thermal infrared, LiDAR) and machine learning (SVM, RF, back propagation (BP), partial least squares (PLS)), which could large-scale obtain tea height, leaf area index, leaf water content, leaf chlorophyll, and nitrogen content. In addition, researchers evaluated the antioxidant ability of plants by Fourier transform infrared spectroscopy (FT-IR), X-Ray diffraction (XRD), and energy dispersive X-ray (EDX) studies [20,21]. However, currently, most studies focus on monitoring crops through imaging cameras (high-precision RGB, multispectral, hyperspectral, thermal infrared, and LiDAR) combined with machine learning, while there are relatively few studies using environmental sensors combined with deep learning methods to monitor plant growth. For a long time, environmental phenotype identification and its data analysis was an important data support for plant breeding; however, it was neglected by most breeding programs [22].

In this study, the research focuses on the influence of environmental parameters on the growth of tea seedlings. The IoT system was used to monitor the environmental changes in a greenhouse during the whole seedling process, and the growth biomass of shoots and roots in tea seedling was measured. Main contributions: (1) The tea seedlings’ growth curve was quantified; the correlations between light, temperature, water, gas, mineral nutrition, and the growth of tea seedlings were clarified; then, the optimal environmental parameters for tea seedling growth were screened. (2) A TSGS model was constructed based on an IoT system and deep learning. (3) A convolutional neural network-long short-term memory (CNN-LSTM) network was proposed for constructing the TSGS model; then, CNN-LSTM was compared with CNN, LSTM, and SVM. (4) The effects of different genotypes of tea seedlings on the model were eliminated, and the performance of single-variety and multi-variety TSGS models was explored.

2. Materials and Methods

2.1. Experimental Design

The experimental site was located in Fuyuanchun Ecological Tea Plantation (35°66′ N, 119°47′ E), Rizhao City, Shandong Province, China. The structure of the seedling greenhouse was a semi-slope type greenhouse with an area of 600 m². On 13 November 2021, short-spike cuttings were started, and the seedlings were completed on 2 July 2022. Three varieties of cutting seedlings, namely Yujinxiang (YJX), Zhongbai 1 (ZB), and Zhongming 6 (ZM), were used in this experiment. The cutting seedlings were planted in the way of plug trays, 240 plug trays were planted for each variety, and each plug tray had 36 cutting seedlings. During the whole growth process of the cutting seedlings, samples were taken every 25 days, and one plug was taken for each variety of cutting seedlings, and a total of 10 samples were taken.

2.2. Determination of Growth Biomass

The cutting seedlings’ growth mainly includes the growth of shoots and roots, so the biomass of shoots and roots was determined to represent the growth biomass of cutting seedlings. The shoots and roots were separated from the cutting seedlings and placed in an oven (105 °C, 20 min). Then, the oven temperature was adjusted to 90 °C and dried to a constant weight. Finally, an electronic balance measured the total weight of shoots and mother leaves, and the weights were recorded. In order to increase the amount of data, a single-variable cubic regression was established between the growth biomass and the time series. The cutting seedlings’ growth curves were quantified by establishing a regression, and the daily growth biomass was obtained as the model’s input.

2.3. Collection of Environmental Data

The climate collection system was installed in the seedling greenhouse, including air temperature and humidity transmitter (KR-BYH-M, Kerun, Weihai, China), light intensity transmitter, carbon dioxide transmitter, trinity (soil temperature, moisture, and conductivity) transmitter (KR-ECTH-1-TR-1, Kerun, Weihai, China), and soil fertility (nitrogen, phosphorus, potassium) transmitter (KR-NPK-N01-TR, Kerun, Weihai, China), as shown in Figure 1. The climate collection system was used to obtain the environmental changes’ data during the whole seedling process, 24 times a day, 225 days in total. All the collected data were uploaded to the remote monitoring platform through the transmission module (KR-XZJ-100-Y-4G, Kerun, Weihai, China). The average of the daily data was used as the input to the model.

2.4. Establishment of CNN-LSTM Model

The CNN-LSTM consisted of two parts, the CNN spatial feature extraction module and the LSTM module (Figure 2). The CNN spatial feature extraction module was mainly used to extract feature vectors related to the growth of cutting seedlings in meteorological data. The LSTM module was used to obtain the temporal correlation of the growth of cutting seedlings [23,24]. The Fold represents sequence folding, the BN represents the Batch Normalization, the RELU is the activation function, AVG POOL is the average pool layer operation, the Unfold represents sequence unfolding, and the Flatten represents the flattening operation. Before feeding the environmental data and growth biomass data into CNN-LSTM, the cell functions were used to convert the data into a 2D matrix. Then, the data were input into a 3 × 3 filter for convolution, convolution five times in a row, after average pooling, sequence expansion, and flattening, and then input to the LSTM network. After three recurrent gatings, the predicted data were input to the fully connected layer and output by the regressor.

In order to further verify the performance of the CNN-LSTM network, machine learning method SVM [25] and deep-learning methods CNN [26] and LSTM [27] were selected to compare with the CNN-LSTM network. In order to compare the performance of deep-learning methods more reliably, the number of layers in both CNN and LSTM networks was 20. The hyperparameters were determined during the training process; then, the final deep-learning model adopted an Epoch of 120 and a learning-rate of 0.01 to prevent the model’s overfitting. The kernel function of SVM is a polynomial. In this study, the TSGS model was established by these four methods.

2.5. Test Environment and Model Evaluation

The hardware environment for experimental data processing is a dual-processor [Inter (R) Core (TM) i7-6700HQ CPU @2.60 GHz 2.60 GHz, DELL, Xiamen, China], and the on-board RAM is 8 GB. TSGS model established by MATLAB 2020 (MathWorks, Natick, MA, USA).

A 5-fold cross validation method was employed to divide the training set and the test set to verify the model’s accuracy. In order to evaluate the performance of the model more accurately, the coefficient of determination (R²) and the root mean square error (RMSE) were used to evaluate the effect of the regression model [28]. Generally speaking, the higher the R² value is, the closer it is to 1, the higher the accuracy of the built model. In contrast, the lower the RMSE value, the closer to 0, the higher the accuracy of the built model. The formulas for R² and RMSE are as follows:

$${\mathrm{Rc}}^2,{\mathrm{Rp}}^2=\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\overline{\mathrm{y}}}_{\mathrm{i}}\right)}^2-{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\hat{\mathrm{y}}}_{\mathrm{i}}\right)}^2}{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\overline{\mathrm{y}}}_{\mathrm{i}}\right)}^2}$$

(1)

$$\mathrm{RMSEC}, \mathrm{RMSEP}=\sqrt{\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\hat{\mathrm{y}}}_{\mathrm{i}}\right)}^2}{\mathrm{n}} }$$

(2)

where Rc² and RMSEC represent training set, Rp² and RMSEP represent test set, n is the number of samples in the corresponding data set, ${\hat{\mathrm{y}}}_{\mathrm{i}}$ and ${\mathrm{y}}_{\mathrm{i}}$ are the predicted and measured values, respectively, and ${\overline{\mathrm{y}}}_{\mathrm{i}}$ is the average measured value of the sample.

3. Results and Discussion

3.1. Quantify the Growth Curve of Cutting Seedings

In order to quantify the growth curve and increase the amount of data, a one-variable cubic regression was established between the growth biomass data and growth time. Table 1 shows the accuracy of the evaluated one-variable cubic regression model. It can be seen from Table 1 that the regression curve of YJX has the highest precision, but its growth rate is the lowest of 0.0151 g/d. The growth rate of ZM is 0.0384 g/d, and the precision of regression curve is higher than that of ZB. The regression curve of ZB has the lowest precision, and the growth rate is 0.0250 g/d. Figure 3 shows the growth curves of cutting seedings in the three varieties. It can be seen from Figure 3 that the growth rate of cutting seedlings starts to accelerate at 100 days and reaches the maximum at 200 days. From the established growth curve, it can be seen that the fitting effect of YJX and ZM is better than that of ZB. The regression formulas of the three varieties are as follows:

$${ \mathrm{Y}}_1=-0.07631+0.21415\times \mathrm{X}-0.00043\times {\mathrm{X}}^2+2.18\times {10}^{-6}\times {\mathrm{X}}^3$$

(3)

$${\mathrm{Y}}_2=-0.26819+0.41848\times \mathrm{X}-0.00082\times {\mathrm{X}}^2+3.80\times {10}^{-6}\times {\mathrm{X}}^3$$

(4)

$${\mathrm{Y}}_3=-0.16971+0.036648\times \mathrm{X}-0.0008\times {\mathrm{X}}^2+4.62\times {10}^{-6}\times {\mathrm{X}}^3$$

(5)

Y₁, Y₂, and Y₃ represent the growth biomass of YJX, ZB, and ZM respectively, X represents the days.

Table 1. Evaluating the accuracy of regression models.

Varieties	Multiple R	R Square	Adjusted R Square	Average Growth Rate (g/d)
YJX	0.996	0.993	0.988	0.0151
ZB	0.933	0.871	0.794	0.0250
ZM	0.994	0.989	0.982	0.0384

Table 1. Evaluating the accuracy of regression models.

Varieties	Multiple R	R Square	Adjusted R Square	Average Growth Rate (g/d)
YJX	0.996	0.993	0.988	0.0151
ZB	0.933	0.871	0.794	0.0250
ZM	0.994	0.989	0.982	0.0384

Figure 3. Growth curve of cutting seedings. The solid line is the true value of the growth biomass, and the dashed line is the predicted value of growth biomass after regression. Varieties of cutting seedlings: Yujinxiang (YJX), Zhongbai 1 (ZB), and Zhongming 6 (ZM).

Figure 3. Growth curve of cutting seedings. The solid line is the true value of the growth biomass, and the dashed line is the predicted value of growth biomass after regression. Varieties of cutting seedlings: Yujinxiang (YJX), Zhongbai 1 (ZB), and Zhongming 6 (ZM).

3.2. Changes of Environmental Parameters

Meteorological and soil sensors were employed to collect the air temperature, soil temperature, light intensity, and air humidity data during the whole seedling process, to monitor the environmental changes in the greenhouse, as shown in Figure 4. The results showed that the air temperature and soil temperature showed an upward trend, mainly because the gradual increase of the outdoor temperature affected the indoor temperature. The light intensity showed a downward trend because the outdoor light intensity increases with the passage of time. In the Northern Hemisphere, the light intensity is weak in winter and strong in summer. However, tea plants prefer light, tolerates shade, and avoids direct sunlight. Therefore, the seedling greenhouse was shaded to reduce light intensity. The air humidity was maintained at 70–95%, which was the most suitable air humidity range for the growth of tea seedlings. Soil moisture was maintained at 15–25%. The changes of carbon dioxide concentration and soil NPK content were relatively gentle.

3.3. Screening of the Optimal Environment Parameters

In order to explore the influence of environmental parameters on the growth biomass of tea seedlings in various varieties, Pearson correlation was used to analyze the environmental parameters and the growth biomass of tea seedlings, as shown in Figure 5. The results showed that the air temperature and soil temperature had the highest correlation with the growth biomass of tea seedlings, and the average correlation coefficients were 0.78 and 0.84, respectively. This indicated that air temperature and soil temperature were the two key parameters that had the most significant effects on the growth of tea seedlings. From another perspective, the growth rate of tea seedlings began to accelerate at 100 days in Figure 3. At this time, the air temperature and soil temperature changed significantly. Previous studies showed that temperature was one of the most important environmental parameters affecting plant growth. Plant growth was directly and indirectly affected by temperature, and the optimal temperature was crucial to ensure quality and improve material levels [29]. For example, temperature changes affect the growth of potatoes by affecting the root system development, respiration, transpiration, flowering, and dormancy in potato seedlings [30].

In addition, soil moisture was also an important environmental parameter for the growth of tea seedlings, with an average correlation coefficient of −0.63. This indicated that within the suitable soil moisture range for tea seedlings, the lower the soil moisture, the faster the tea seedlings grow. Carbon dioxide and light had the highest correlation with the growth of ZM, which was −0.51 and −0.47, respectively. This indicated that the influence of carbon dioxide and light on the growth of ZM was greater than that of YJX and ZB. This may be because the average growth rate of ZM was much higher than that of YJX and ZB, and more organic matter was produced. However, carbon dioxide and light were the prerequisites for plants to photosynthesize to produce organic matter, so ZM was more responsive to carbon dioxide and light. The average correlation coefficient of air humidity and soil nitrogen, phosphorus, and potassium were about −0.3, which had little effect on the growth of tea seedlings.

Due to the influence of economic conditions, controlling all environmental parameters is expensive. Therefore, in future research, under the condition of ensuring the suitability of other environmental parameters, the research will focus on regulating the three environmental parameters: air temperature, soil temperature, and soil moisture. This will help to find the best node more efficiently and shorten the seedling cycle faster.

3.4. Evaluation and Comparison of Models

In order to eliminate the influence of different varieties of tea seedlings on the model and improve the robustness and generalization of the model, the growth rates of tea seedlings in three varieties were used as input parameters.

Table 2. Evaluation of model accuracy.

Target	Model	Training Sets		Test Sets
		Rc2	RMSEC	Rp2	RMSEP
YJX growth amount	CNN-LSTM	0.99	0.01	0.96	0.17
	SVM	0.98	0.14	0.94	0.22
	CNN	0.95	0.20	0.92	0.64
	LSTM	0.99	0.01	0.94	0.57
ZB growth amount	CNN-LSTM	0.97	0.18	0.94	0.25
	SVM	0.98	0.15	0.93	0.31
	CNN	0.93	0.31	0.88	0.85
	LSTM	0.99	0.01	0.92	0.74
ZM growth amount	CNN-LSTM	0.99	0.05	0.98	0.14
	SVM	0.98	0.36	0.95	0.61
	CNN	0.95	0.63	0.92	1.23
	LSTM	0.99	0.01	0.96	0.45
YJX + ZB + ZM growth amount	CNN-LSTM	0.98	0.18	0.96	0.17
	SVM	0.94	0.48	0.92	0.52
	CNN	0.82	1.45	0.76	1.92
	LSTM	0.99	0.01	0.89	1.12

shows the evaluation results of single-variety and multi-variety models. For evaluating GSTS of a single variety, the GSTS model based on CNN-LSTM had the highest accuracy, and the GSTS accuracy of ZM was the highest (Rp² = 0.98, RMSEP = 0.14). The GSTS model based on CNN had the lowest accuracy, and the GSTS accuracy of ZB was the lowest (Rp² = 0.88, RMSEP = 0.85); For evaluating GSTS of multi-variety, the GSTS model based on CNN-LSTM had the highest accuracy (Rp² = 0.96, RMSEP = 0.17). The GSTS model based on CNN had the lowest accuracy (Rp² = 0.76, RMSEP = 1.92). Figure 6 shows the scatter plots of the validation set for evaluating single- and multi-variety GSTS based on CNN-LSTM.

To sum up, the GSTS model based on CNN-LSTM has very high accuracy, which is better than SVM, CNN, and LSTM networks. The CNN-LSTM’s high accuracy is mainly attributed to the CNN’s strong feature extraction ability; meanwhile, the LSTM can discover interdependences in time series data and automatically detect the optimal combination suitable for related data [31]. Therefore, the combination of CNN and LSTM can maximize both advantages. In previous studies, Hassan et al. [32] used the CNN network to extract depth features from time series images, and used image depth time series features as the input of LSTM to build a rapeseed nutrient diagnosis model. Finally, the CNN-LSTM model was compared with another standard multi-class support vector machine (MCSVM) model. The results of the research were consistent with our findings that the prediction performance of the model established by CNN-LSTM was better than that of traditional machine learning methods. Qiao et al. [33] proposed a spatial-spectral temporal neural network (SSTNN) to obtain the joint spatial-spectral features of multispectral images of wheat and corn by using CNN, and performed temporal dynamic analysis by using RNN composed of multiple bidirectional LSTM units. Compared with random forest (RF), SVM, LSTM, and CNN algorithms, the results showed that the accuracy of SSTNN (R² = 0.74, RMSE = 0.82) is higher than that of RF, SVM, LSTM, and CNN. The results of the research were consistent with our findings that the accuracy of hybrid network is better than that of single network.

The effect of the GSTS model based on CNN was the worst. This was because the change in tea seedlings’ growth biomass was positively correlated with the change in time. However, CNN was not completely suitable for learning time series, requires various auxiliary processing, and the effect was not necessarily good [34]. Therefore, the model established by only using the CNN network is less effective. The LSTM network could obtain the temporal correlation of the growth of cutting seedlings. In our study, the precision of LSTM model is higher than that of CNN and SVM model, with Rp² of 0.89–0.96. This shows that LSTM network is more effective in monitoring plant growth. In previous studies, researchers used spatio-temporal short-term memory (ST-LSTM) and memory in memory (MIM) to predict wheat future growth and development image sequences. The RMSE is 77.78–118. The results also prove that LSTM network is more effective in monitoring plant growth [24]. However, different from our study, it uses images to monitor wheat growth, while we use environmental parameters to monitor tea seedling growth. In future research, images and environmental data of tea seedling growth will be collected at the same time for joint analysis to explore a better model.

3.5. Outlook

Plant breeding mainly depends on the big data of genotype, phenotype, and environmental type. The joint analysis of genotype-phenotype-environmental type is the direction of breeding in the future, which is of great value for realizing the true sense of directional breeding, breeding to adapt to climate change, and breeding to adapt to specific environments. At present, genotype identification has developed to the stage of multi-omics, and the nondestructive detection of the phenotype has also made significant progress with the development of optical sensors. However, the identification of environmental type is just beginning. In this study, the effects of various environmental parameters on the growth of different genotypes of tea seedlings were analyzed, and a TSGS model was constructed with the method of deep learning. By inputting the growth rate into the TSGS model, the influence of tea seedlings of different genotypes on the model is eliminated, which provides a way for future genome environment group-integrated prediction-driven intelligent breeding. In addition, the data collected in different space-time conditions (including multiple years, multiple periods, and multiple locations) plays an important role in improving the generalization and robustness of the model. This section may be divided into subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

4. Conclusions

The IOT system was used to measure environmental change during the whole seedling process. CNN-LSTM, CNN, LSTM, and SVM were used to build the TSGS model of environmental parameters and growth biomass. The conclusions were as follows:

(1)

The average correlation coefficients of air temperature, soil temperature, and soil moisture with the biomass growth of tea seedlings were 0.78, 0.84, and −0.63, respectively, which were three important parameters for establishing the TSGS model.

(2)

For evaluating the TSGS model of single variety, the accuracy of ZM’s TSGS based on CNN-LSTM network was the highest (Rp² = 0.98, RMSEP = 0.14).

(3)

For evaluating the TSGS model of multiple varieties, the accuracy of TSGS based on CNN-LSTM network was the highest (Rp² = 0.96, RMSEP = 0.17).