Research on Temperature Prediction of Passion Fruit Planting Bases in Southwest Fujian Province

Abstract

This article utilized hourly temperature, humidity, pressure and wind speed data from passion fruit meteorological observation stations in three southwestern cities of Fujian Province (Longyan, Sanming, Zhangzhou) from 2020 to 2022, as well as national ground conventional meteorological observation stations. BP neural network and stepwise regression method were applied to construct temperature prediction models for the passion fruit planting bases. The results showed that: (1) The simulation effect of the passion fruit station temperature prediction model based on BP neural network (referred to as BP model) was better than that of the model based on stepwise regression method (referred to as regression model). The average absolute error (MSE) of BP model (2.75–3.42 °C) was smaller than that of regression model (3.32–3.94 °C). (2) For the simulation results of daily temperature changes in the passion fruit station, the difference in hourly average temperature between the BP model predictions (regression model predictions) and observed temperatures at passion fruit station was −4.1–4.4 °C (−6.0–10.2 °C). The BP model showed a daily temperature trend that was closer to the measured values; (3) For the simulation results of high and low temperatures in the passion fruit station, the BP neural network model (regression model) showed a prediction error range of −5.6 °C to 5.2 °C compared to observed temperatures, while the stepwise regression model’s error range was −4.1 °C to 8.8 °C. The BP model’s predicted temperature trend was closer to the measured values. (4) Both models have significant shortcomings in the prediction of high-temperature individual cases and hourly averages, with relatively large errors (generally exceeding 3 °C), especially during the period from 10 to 16 o’clock. The future version needs to be optimized.

1. Introduction

Passion fruit is a perennial evergreen vine plant that thrives in tropical and subtropical regions. It is rich in various nutrients necessary for the human body and has extremely high nutritional value. Market demand for passion fruit has consistently exceeded supply [1,2,3].

Research showed that temperature has a significant impact on the growth, development, yield, and quality of passion fruit. Liang et al. [4] discovered temperature-related genomic regions in the passion fruit genome and determined that temperature is one of the main influencing factors [5,6,7,8]. The most suitable temperature for the growth of passion fruit is 25–28 °C [9], and high or low temperatures seriously affect the normal growth and development of passion fruit [10,11].

Longyan, Sanming and Zhangzhou are the main passion fruit-growing in Fujian. The favorable climate conditions there are highly conducive to nutrient accumulation and sugar conversion in passion fruit, leading to high quality and yield. However, abnormal temperatures remain a challenge for local growers, especially amid global climate change [12,13,14]. Extreme temperatures (excessive or insufficient) frequently occur in Fujian Province, severely affecting passion fruit production.

Typically, temperature prediction is based on data from national conventional ground meteorological observation stations, but there is a significant difference between the temperature at these stations and that at the passion fruit planting bases. Since the growth and development of passion fruit are primarily influenced by the temperature in the planting bases, the key technical challenge is how to accurately predict the temperature at the passion fruit plantation sites based on data from national meteorological stations, thereby providing precise temperature forecasts for passion fruit cultivation.

In research on microclimate temperature prediction models, previous studies have adopted multiple methods, including the stepwise regression method [15,16,17,18,19,20], long short-term memory (LSTM) neural networks [21,22,23], and BP neural networks [24,25,26,27], among other machine learning approaches [28,29]. Many studies focus on modeling greenhouse microclimates, but there is relatively little research on temperature simulation in open-field planting areas. Since many crops grow in natural outdoor environments, temperature simulation for such areas is particularly important. Additionally, most studies rely on small datasets, with observation periods limited to a few months.

This study utilizes hourly meteorological data (temperature, instantaneous wind speed, relative humidity, and air pressure) from 2020 to 2022, collected at both the meteorological stations of the passion fruit planting bases and nearby national conventional ground observation stations in three southwestern Fujian cities (Longyan, Sanming, and Zhangzhou). Using BP neural networks and stepwise regression, we constructed a temperature prediction model for the passion fruit planting bases. The results can help optimize temperature regulation within passion fruit planting bases, thereby ensuring an optimal growth environment and enhancing production and profitability of passion fruit in Fujian.

2. Data Sources and Research Methods

2.1. Research Area

Longyan, Sanming and Zhangzhou are all located in the southwest of Fujian Province, with a mix of mountainous terrain and coastal plains. The terrain is quite rugged, featuring many peaks and canyons (Figure 1). The region has a subtropical monsoon climate, which is highly suitable for the cultivation of passion fruit. The black boxes represent the national stations, and the red triangles represent the passion fruit stations.

The meteorological data used in this study are all from the Fujian Provincial Meteorological Bureau, including Sanmingren Village Station (Passion Fruit Planting Bases Meteorological Station, abbreviated as Passion Fruit Station), Sanming Station (National Ground Conventional Meteorological Station, abbreviated as National Station), hourly meteorological data from September 2020 to November 2022. Additionally, hourly meteorological data for Longyan Zhuoyang Village Station (Passion Fruit Station) and Shanghang Station (National Station) from February 2021 to November 2022. Additionally, hourly meteorological data from Puli Village Station in Zhangzhou (Passion Fruit Station) and Zhangzhou Space Station (National Station) from September 2020 to December, 2022. The data measurements conducted by the meteorological station all represent the readings taken at a height of 1.5 m above the ground.

The DEM data come from the Spatial Geography Cloud Data, and the China’s provincial administrative boundary data comes from the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences.

2.2. Phenological Period Division of Passion Fruit

Through literature research [30,31], the time range of different phenological stages of Fujian passion fruit is determined (

Table 1. Time Range of Different Phenological Stages of Passion Fruit.

Phenological Period	Budding	Flowering and Fruiting	Overwintering
Data	3.21~4.10	4.11~11.20	11.21~3.20

).

2.3. BP Model

The BP neural network is a common artificial neural network model used to solve various tasks such as classification and regression. It consists of an input layer, a hidden layer and an output layer, which is trained using forward propagation and back-propagation algorithms. Forward propagation calculates the network’s output, while back-propagation adjusts weights and biases to minimize errors. This process iterates on the training data until the desired performance is achieved.

Calculate the correlation between hourly temperature at the Passion Fruit Meteorological station and various meteorological elements (pressure, instantaneous wind speed, national station temperature, relative humidity, etc.) at the National Station. Input meteorological elements that have a good correlation with the temperature of passion fruit station into a BP neural network for training, and construct the BP model. The structure of the BP neural network is shown in Figure 2.

To solve the problem of inconsistent units of input variables, the Min-Max Normalization method is used to normalize meteorological elements:

$${\mathrm{x}}_{\mathrm{i}}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{ij}}-{\mathrm{X}}_{\mathrm{imin}}}{{\mathrm{X}}_{\mathrm{imax}}-{\mathrm{X}}_{\mathrm{imin}}}}$$

(1)

where, x_i is the normalized data of the ith input variable (i = 1, 2, 3, 4), X_ij is the jth number of the ith input variable, and X_imin and X_imax are the maximum and minimum values of the input variable.

The normalized data are input into the input layer, passed through a rectified linear unit activation function (ReLU) to the intermediate layer, and finally to the output layer. ReLU (Rectified Linear Unit) is a simple activation function that outputs the maximum value between the input and zero. It is computationally efficient and helps alleviate the problem of vanishing gradients. Then, the outputs are denormalized to obtain final results, and there is no activation function is applied to the output layer. This article uses the Conjugate Gradient method for back propagation training, continuously adjusting weights and biases to minimize the loss function.

The parameters of the BP model are set through the following methods, and the model with the smallest average absolute error between the predicted value and the actual value is selected: (1) The number of hidden layers is set to 4, 8, 12, 16, and 20. (2) The training times are set to 200, 500, 800, 1000, and 2000 respectively. (3) The learning rates are set to 0.01, 0.001, and 0.0001 respectively. (4) The minimum errors are set to 0.01, 0.001, and 0.0001 respectively.

The final settings for the BP neural network parameters are as follows: training epochs = 1000, learning rate = 0.001, minimum error threshold = 0.001. The dataset is randomly divided into 70% for training and 30% for validation.

2.4. Stepwise Regression Model

Stepwise regression is a method used in multiple linear regression analysis to select optimal independent variable. Its main purpose is to determine the model that best explains the changes in the dependent variable by gradually adding or removing independent variables. The regression equation is:

$$\mathrm{y}=\mathrm{b}+{\mathrm{b}}_1\times {\mathrm{x}}_1+{\mathrm{b}}_2\times {\mathrm{x}}_2+{\mathrm{b}}_3\times {\mathrm{x}}_4+{\mathrm{b}}_4\times {\mathrm{x}}_4$$

(2)

where, x₁ is the national station atmospheric pressure, x₂ is the national station wind speed, x₃ is the national station temperature, x₄ is the national station relative humidity, y is the temperature of the passion fruit station, b is a constant, b₁, b₂, b₃, b₄ are coefficients.

We randomly allocated 70% of the data for training and 30% for validation.

2.5. Error Analysis

To clarify error of the BP model and regression model, the average absolute error MSE and root mean square error RMSE between the predicted temperature y₁, passion fruit station temperature y₂, and national station temperature x₂ in the statistical test data are calculated.

$${\mathrm{MSE}}_0={\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{1\mathrm{i}}-{\mathrm{y}}_{2\mathrm{i}}\right|}{\mathrm{n}}}$$

(3)

$${\mathrm{MSE}}_1={\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{2\mathrm{i}}-{\mathrm{x}}_{2\mathrm{i}}\right|}{\mathrm{n}}}$$

(4)

$${\mathrm{RMSE}}_0=\sqrt{{\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{1\mathrm{i}}-{\mathrm{y}}_{2\mathrm{i}}\right)}^2}{\mathrm{n}}}}$$

(5)

$${\mathrm{RMSE}}_1=\sqrt{{\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{2\mathrm{i}}-{\mathrm{x}}_{2\mathrm{i}}\right)}^2}{\mathrm{n}}}}$$

(6)

where, n is the number of samples, i is the ith observation data, y₁ is the predicted temperature, y₂ is the temperature of passion fruit station, x₂ is the national station temperature, MSE₀ and RMSE₀ are the average absolute error and root mean square error of the predicted temperature (y₁) and passion fruit station temperature (y₂), MSE₁ and RMSE₁ are the average absolute error and root mean square error of the passion fruit station temperature (y₂) and national station temperature (x₂), respectively.

2.6. Model Effect Verification

To evaluate the performance of the temperature prediction model for passion fruit station in southwestern Fujian, the following validation methods are employed: (1) Statistical analysis: The model’s hourly predicted temperatures are compared with observed values from passion fruit station. The differences are calculated and visualized through line charts. (2) Extreme event validation: We selected recorded extreme temperature events (high/low) in southwestern Fujian (2020–2022). Discrepancies are statistically analyzed and presented as anomaly time series.

3. Results and Analysis

3.1. Correlation of Temperatures Between Meteorological Elements at National Stations and Passion Fruit Station

Table 2. Correlation statistics between temperature in passion fruit station and meteorological elements at national stations.

Regional	Meteorological Elements of National Stations
	Air Pressure	Instantaneous Wind	Atmospheric Temperature	Relative Humidity
Long yan	−0.68 **	−0.15 **	0.65 **	−0.07 **
San ming	−0.73 **	−0.18 **	0.71 **	−0.04 **
Zhang zhou	−0.68 **	0.04 **	0.64 **	0.31 **

Note: ** 0.01 significance, * 0.05 significance.

displays the statistical analysis of the correlation between temperature in passion fruit station and meteorological elements at national stations. It can be seen that: (1) The correlation coefficients between the meteorological elements are statistically significant at p < 0.01, indicating that air pressure, instantaneous wind speed, national station temperature, and relative humidity were all related to the temperature of the passion fruit station. (2) The correlation between temperature, air pressure in the passion fruit station and temperature in the national station is the best, ranging from −0.68 to −0.71.

The negative correlation between temperature and air pressure at Passion Fruit Station is because the high-pressure system usually accompanies clear weather and night-time radiational cooling, resulting in a decrease in temperature; the negative correlation with wind speed is because strong winds promote air mixing and evaporation for heat dissipation, enhancing the cooling effect; the negative correlation with relative humidity is because in high-humidity conditions, an increase in cloud cover weakens solar radiation, and the influence of water vapor condensation releasing latent heat may be masked by other factors. The positive correlation between the temperature at Passion Fruit Station and that at the national station mainly stems from both being controlled by large-scale weather systems, and their temperature change trends are consistent. However, due to the influence of local microclimate conditions such as terrain, vegetation, or altitude at Baixiangguo station, the specific temperature variation amplitude differs from that of the surrounding national stations.

3.2. Simulation Results of BP and Regression Model

The dataset was randomly split into training (70%) and validation (30%) subsets. Following the methodology described in Section 2.5, the methodology described in Section 2.5 (

Table 3. Various indicators of BP model.

Regional	Phenological Phase	MSE0 (°C)	RMSE0 (°C)	MSE1 (°C)	RMSE1 (°C)
Longyan	Budding	3.0	3.9	5.0	7.2
	Flower and fruit	3.1	3.9	5.0	6.7
	overwintering	3.0	3.9	4.7	6.7
Sanming	Budding	2.6	3.5	4.4	6.1
	Flower and fruit	3.3	4.1	4.7	6.3
	overwintering	2.8	3.7	4.0	5.8
Zhangzhou	Budding	3.2	4.1	4.7	6.5
	Flower and fruit	3.4	4.3	5.3	6.7
	overwintering	3.0	3.9	5.8	7.3

and

Table 4. Various indicators of Regression model.

Regional	Phenological Phase	MSE0 (°C)	RMSE0 (°C)	MSE1 (°C)	RMSE1 (°C)
Longyan	Budding	3.7	4.8	4.5	6.4
	Flower and fruit	3.5	4.3	5.0	6.7
	overwintering	3.5	4.5	4.4	6.3
Sanming	Budding	3.5	4.6	4.4	6.2
	Flower and fruit	3.8	4.6	4.7	6.2
	overwintering	3.3	4.3	4.0	5.8
Zhangzhou	Budding	3.6	4.6	4.8	6.3
	Flower and fruit	3.8	4.7	5.4	6.7
	overwintering	3.4	4.4	5.9	7.5

). The stepwise regression equation is shown in

Table 5. Stepwise regression equation.

Regional	Phenological Phase	Equations	R
Longyan	Sprout	y = 338.04 − 0.34 × x1 − 0.20 × x2 + 0.48 × x3 + 0.11 × x4	0.4
	Flower and fruit	y = 193.23 − 0.20 × x1 − 0.23 × x2 + 0.56 × x3 + 0.12 × x4	0.3
	overwintering	y = 113.59 − 0.13 × x1 − 0.05 × x2 + 0.71 × x3 + 0.17 × x4	0.4
Sanming	Sprout	y = 318.56 − 0.32 × x1 − 0.36 × x2 + 0.32 × x3 + 0.09 × x4	0.4
	Flower and fruit	y = 237.48 − 0.24 × x1 − 0.51 × x2 + 0.56 × x3 + 0.10 × x4	0.5
	overwintering	y = 42.17 − 0.05 × x1 − 0.19 × x2 + 0.47 × x3 + 0.15 × x4	0.5
Zhangzhou	Sprout	y = 360.28 − 0.35 × x1 − 0.19 × x2 + 0.32 × x3 + 0.09 × x4	0.3
	Flower and fruit	y = 127.42 − 0.14 × x1 + 0.21 × x2 + 0.72 × x3 + 0.19 × x4	0.5
	overwintering	y = 80.56 − 0.09 × x1 + 0.24 × x2 + 0.72 × x3 + 0.18 × x4	0.4

Note: R: correlation coefficient, x₁: the national station atmospheric pressure (hPa), x₂: the national station wind speed (m/s), x₃: the national station temperature (°C), x₄: the national station relative humidity (%), y: predicted temperature (°C).

. The results shows that: (1) The BP model exhibited MSE₀, RMSE₀, MSE₁, and RMSE₁ ranges of 2.8–3.4 °C, 3.9–4.3 °C, 4.0–5.8 °C, and 6.1–7.3 °C, respectively, while the regression model showed corresponding ranges of 3.3–3.8 °C, 4.3–4.9 °C, 4.0–5.9 °C, and 5.8–7.6 °C. (2) Both models exhibited lower MSE₀ values compared to MSE₁ and lower RMSE₀ than RMSE₁. The BP model outperforms the regression model with lower MSE₀ (2.8–3.4 °C vs 3.3–3.8 °C) and RMSE₀ (3.9–4.3 °C vs 4.3–4.9 °C) values, demonstrating superior prediction accuracy.

3.3. BP and Regression Model Predict Daily Temperature Changes

In order to clarify the daily variation characteristics of temperature predicted by the model, following the methodology in Section 2.6, the predicted values and measured daily variations are plotted (Figure 3, Figure 4 and Figure 5). It can be seen that: (1) The mean hourly temperature is 7.5–30.0 °C (observed at passion fruit station), 10.0–25.0 °C (BP model predictions), and 5.0–20.0 °C (regression model predictions). (2) The hourly average temperature differences between BP model predictions and observed values at the passion fruit station ranged from −4.1 to 4.1 °C during the budding stage, −3.8 to 4.4 °C during the flowering and fruiting stage, and −3.6 to 4.4 °C during the overwintering stage. In comparison, the regression model exhibited larger deviations, with prediction errors ranging from −6.0 to 10.2 °C in the budding stage, −1.9 to 8.9 °C in the flowering and fruiting stage, and −2.1 to 9.4 °C in the overwintering stage. (3) Between 14:00 and 16:00, both models exhibited relatively large deviations from the observed temperatures at the passion fruit station, typically exceeding 3 °C, with the BP model consistently showing smaller prediction errors compared to the regression model. (4) The daily variation trend of average temperature in the BP model matches the observed diurnal pattern at the passion fruit station, while the regression model fails to capture significant diurnal variations.

In summary, during 08:00–18:00, compared with the regression model, the BP model’s predicted temperatures showed relatively small deviations from the observed hourly averages at passion fruit station, and its predicted diurnal temperature trends closely matched the measurements. Therefore, the simulation effects of the BP model are better than that of the regression model. However, in reality, the BP model’s prediction of hourly average temperatures at passion fruit station is not very accurate. Further optimization is needed in the future.

3.4. High and Low Temperature Simulation by BP and Regression Model

To clarify the modeling effects of the two models on high and low temperatures, following the methodology in Section 2.6, the BP and regression models are used to predict the high temperature event on 25 July 2021 and the low temperature event on 11 January 2022 in Longyan, Sanming and Zhangzhou. The hourly X-axis and temperature Y-axis are used to plot the daily temperature changes of the high and low temperature events (Figure 6).

In the high-temperature simulation, (1) The observed temperatures at the passion fruit station ranged from 20.0 to 35.0 °C, while the BP model predictions matched this range (20.0–35.0 °C) and the regression model showed a narrower prediction range (20.0–30.0 °C). (2) The temperature differences between BP models predictions and observed temperatures at passion fruit station are −3.0–4.7 (Longyan), −4.2–2.6 (Sanming), and −5.6–5.2 (Zhangzhou) °C. The temperature differences between regression models predictions and observed temperatures at passion fruit station are −0.5–8.8 (Longyan), −4.3–8.2 (Sanming), and −4.1–8.7 (Zhangzhou) °C. (3) The BP model demonstrates better temperature prediction accuracy than the regression model at passion fruit cultivation sites, with deviations from observed temperatures generally remaining below 5 °C compared to the regression model’s typically larger discrepancies exceeding 5 °C, particularly during daytime hours (10:00–16:00). (4) The BP model’s predicted temperature trends shows closer agreement with observations at passion fruit station, whereas the regression model generated artificially attenuated temperature fluctuations.

In low-temperature simulation, (1) the temperature of passion fruit station, BP model predicted temperature, and regression model predicted temperature is 2.5–12.5, 2.5–20.0, and 0.0–12.5 °C, respectively. (2) The temperature differences between BP models predictions and passion fruit station temperatures are −0.6–2.1 (Longyan) °C, −3.9–2.6 (Sanming) °C, and −4.8–1.0 (Zhangzhou) °C. The temperature differences between regression models predictions and observed temperatures at passion fruit station are 3.0–5.8 (Longyan), −4.1–2.7 (Sanming), and 2.5–4.7 (Zhangzhou) °C. (3) The temperature differences between BP models predictions and passion fruit station temperatures are generally less than 3 °C, while the temperature differences between regression models predictions and observed temperatures at passion fruit station are generally greater than 3 °C, especially. (4) The BP model’s predicted temperature trends shows closer alignment with observations from passion fruit station. In contrast, the stepwise regression model produces artificially attenuated daily temperature fluctuations with statistically insignificant variation characteristics.

In summary, for high/low temperature predictions, the BP model demonstrates closer alignment with observed values at passion fruit station (mean absolute error: 2.8–3.4 °C), and significantly smaller deviations from field measurements compared to the regression model. These results confirm the simulation effects of the BP model are better than that of the regression model. However, in reality, the BP model’s predictions for high temperatures at passion fruit station are not very accurate. Further optimization is needed in the future.

4. Conclusions and Discussion

Through the conclusion, it can be analyzed that: (1) For temperature simulation in passion fruit station, the BP model’s MSE (2.8–3.4) and RMSE (3.9–4.3 °C) were lower than the regression model’s (MSE: 3.3–4.0; RMSE: 4.3–4.9 °C). (2) The simulation of daily temperature changed in passion fruit station showed that the differences in hourly average temperature between BP model predictions (regression model predictions) and observed temperatures at passion fruit station was −4.1–4.4 °C (−6.0–10.2 °C),and the BP model was relatively small. (3) High temperature simulation of passion fruit station, the temperature difference between BP models predictions (regression model predictions) and observed temperatures at passion fruit station was −5.6–5.2 °C (−4.1–8.8 °C). The difference between the predicted values of the BP model and the measured values of passion fruit station was smaller than that of the regression model. (4) Low temperature simulation of passion fruit station, the temperature difference between BP model (regression model) predictions and observed temperatures at passion fruit station was −4.8–2.6 °C (−4.1–5.8 °C). The difference between the predicted values of the BP model and the measured values of the passion fruit station was smaller than that of the regression model. (5) The BP model predicted the change trend of temperature was closer to observations than the regression model. (6) During 10–16 o’clock, the prediction values of the BP model and the regression model for the high temperature and hourly average temperature deviated significantly from the actual measured values. In the future, improvements are needed.

Consistent with the studies of Huang Longfei, Liu Ruolan, Xu Yuchun, etc. [18,19,25], this research has found that the simulation effect of the BP neural network model is superior to that of the regression model. However, unlike other studies on temperature prediction using the BP model, in the actual temperature simulation of this paper, the predicted values of the high-temperature during the 10–16 o’clock by the BP model showed a significant deviation from the actual measured values. This discrepancy suggests that while BP networks maintain their relative advantage over regression approaches, their performance in peak temperature prediction under field conditions may require specific optimization. The research conducted by Xu Yu et al. [24]. indicates that the LSTM model outperforms the BP model. Meanwhile, the support vector machine model constructed by Pandey Kamal, Fan Linan, and others [28,29] demonstrates excellent predictive performance. Therefore, in the future, we will adopt the LSTM and SVR methods to build an air temperature prediction model and test the prediction results.

Currently in China, meteorological data from specialized agricultural areas such as passion fruit plantations are not publicly available, creating a significant knowledge gap for practitioners regarding the microclimatic conditions within these cultivation systems. To address this critical information barrier, this study develops a novel temperature prediction model specifically for passion fruit plantations using BP neural networks and stepwise regression methods. Our model enables growers to estimate plantation-specific temperature conditions based on publicly available weather forecast data, thereby providing valuable decision-support for cultivation management. This approach represents an important step toward bridging the current data accessibility gap in precision agriculture for tropical fruit production.