A Study of Forest Phenology Prediction Based on GRU Models

Abstract

Investigating forest phenology prediction is a key parameter for assessing the relationship between climate and environmental changes. Traditional machine learning models are not good at capturing long-term dependencies due to the problem of vanishing gradients. In contrast, the Gated Recurrent Unit (GRU) can effectively address the problem of vanishing gradients and allow the neural network to capture longer-range dependencies. In this study, an optical camera was used as experimental equipment to obtain forest images. The absolute greenness index (GEI) data of the region of interest (ROI) in the images were calculated to fit the seasonal variation curve of forest phenology. The GRU neural network model was introduced to train and analyze the GEI data, and the performance of the GRU neural network was evaluated using the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) methods. Finally, the model was used to predict the trend of GEI data for the next 60 days. The results showed that: (1) In terms of training and predicting forest phenology, the GRU model was validated using histograms and autocorrelation graphs, which indicated that the distribution of predicted data was consistent with the trend of actual data, the GRU model data was feasible, and the model was stable. (2) The MSE values of the GRU model at rain-fed-CK (preset point 1), sufficient drip irrigation-DIFI (preset point 3), and sufficient furrow irrigation-BIFI (preset point 5) were 9.055 × 10⁻⁵, 12.91 × 10⁻⁵, and 8.241 × 10⁻⁵, respectively. The RMSE values were 9.516 × 10⁻³, 11.36 × 10⁻³, and 7.313 × 10⁻³, respectively. The MAE values were 7.174 × 10⁻³, 8.241 × 10⁻³, and 5.351 × 10⁻³, respectively. These results indicate that the overall performance of the GRU model was good. (3) The predicted GEI data for the next 60 days showed a trend consistent with actual changes in GEI data, as demonstrated by the GRU model. The GRU model has become the preferred method for phenological prediction due to its simple internal structure and relatively short training time. Results show that the GRU model can achieve forest phenological change prediction and can reveal in-depth insights into future forest growth and climate change, providing a theoretical basis for the application of forest phenological prediction.

1. Introduction

Plants exhibit high sensitivity to interannual climate change, and their growth and development are significantly influenced by climate change. As climate change continues to be influenced by human activities such as urbanization, coupled with the exacerbating effects of urban environmental issues such as the urban heat island effect and air pollution, the impact on urban vegetation growth is further intensified. Climate change has had irreversible effects on the life rhythms of many plants, mainly manifested in changes in the phenology of plant growth stages such as budburst, flowering, leaf expansion, peak leaf, autumn coloration, and leaf fall, as well as the shift in phenological synchronicity [1]. By altering their phenology, plants adapt to seasonal changes in their surroundings, providing the most direct evidence for climate change and indicating that species and ecosystems are being affected by global environmental changes, thereby limiting the stable provision of their vegetation ecosystem services [2]. Phenological observations will continue to provide data for global change research, and long-term monitoring of the main phenological phases of urban forests will provide a basis for studying the response of natural ecosystems to future climate change. Accurate phenological models will improve the accuracy of predicting ecosystem productivity and gas exchange with the atmosphere, thus enabling accurate predictions of future climate change trends. Therefore, predicting phenological phases is crucial for understanding how plants respond to climate change. Achieving accurate predictions of forest phenology and enhancing the ecological adaptability of vegetation to climate change can reveal the laws of ecosystem carbon cycling and is closely related to assessing the ecological environment and climate forecasting.

Phenological prediction models are important tools for quantitatively studying the relationship between climate factors and plant phenology and can be used to simulate trends in vegetation distribution, reconstruct past climate changes, predict future climate changes, and assess risks in agriculture and forestry production. Many predictive models have been developed to explain the historical trends of plant phenology and predict future phenological trends [3,4]. Currently, vegetation phenological prediction models are often established using different meteorological factors, satellite remote sensing, and ground-based remote sensing parameters. For example, machine learning is used to predict vegetation phenology models using satellite remote sensing and gridded data [5]. Jolly et al. proposed the Growth Status Index (GSI) model, which is a classic bioclimatological phenology model index composed of minimum temperature, evaporation demand, and photoperiod [6]. Lang et al. proposed a predictive model for autumn phenology, which revealed the coupling effect of shortened photoperiod and decreased minimum daily temperature on leaf senescence [7]. Czernecki used machine learning and satellite remote sensing and gridded data to predict vegetation phenology models [5]. Wingate installed phenology cameras in the carbon and water flux network to obtain leaf-out and leaf-fall phenology periods and established a phenology model related to canopy leaf area, chlorophyll, and anthocyanin [8]. Li Ning et al. used rubber tree phenology quantification research and rubber tree observation test data as the basis, combined with crop growth clock model, to establish a rubber tree spring phenology prediction model and developed it into computer software [9]. Xu Gexi used phenological observation data of tropical evergreen broad-leaved trees combined with monthly average temperature and precipitation data of the same period and used the integral regression analysis method to establish a phenology prediction model [10]. Liu Yuxia used Moderate Resolution Imaging Spectroradiometer (MODIS) remote sensing data and carbon flux data, supplemented by phenology camera observation data and meteorological data, to propose a remote sensing-based phenology inversion model [11]. Zhou Lei focused on the broad-leaved Korean pine forest of the Changbai Mountain and explored the role of digital camera images in species-scale phenological simulation and community-scale phenology model improvement [12]. Liu Jian et al. incorporated a water stress response function into the wheat growth model, improving the simulation accuracy of wheat phenology periods [13].

Physics-based models can predict change processes but require a large amount of relevant information such as meteorological data, terrain data, soil characteristics, and vegetation cover, as well as complex mathematical tools and users with a lot of specialized knowledge [14]. Empirical models, on the other hand, are mainly represented by data-driven models, establishing a direct mapping between input features and output targets without the need to understand the physical processes [15]. Machine learning models belong to the latter category [16]. Deep learning is a new research direction in the field of machine learning, which can acquire the internal rules and representation hierarchy of sample data through training. Currently, deep learning has been applied in the field of phenology analysis, no longer limited to extracting vegetation indices but can directly classify phenological stages by learning the features of time series images. At the same time, deep learning has the ability to handle long time series data, such as recurrent neural networks (RNNs), which are one of the most popular neural networks in the field of deep learning and have been proven to be effective tools for time series prediction [17,18]. However, the inherent problem of RNN is the vanishing or exploding gradient, which has not significantly improved its performance in sequence prediction [19]. Long short-term memory neural networks (LSTMs) optimize and solve this problem based on RNN, and many studies have shown that LSTMs perform better than traditional RNNs. In recent years, gate recurrent units (GRUs) [20], which have been optimized based on LSTMs, have appeared and shown good performance in natural language processing, with simpler structure and higher operation speed [21,22]. For example, H. Yalcin used a pre-trained convolutional neural network to analyze the phenological stages of six crops such as wheat, cotton, and corn [23] and compared them with the phenological stages obtained from the extracted image texture features and the Naïve-Bayes classifier. GRU prediction models have been successfully applied in the prediction of industrial, medical, and agricultural yields [24,25]. Bruno Barbosa developed a model using Sentinel-2 images time series and the Welch t-test to identify vegetation removal in the urban-rural interface for fuel management actions to aid firefighting [26]. Peng Fang compared the classification performance and sensitivity of SVM, RF, and CART algorithms for winter wheat identification using Sentinel-2 images, finding SVM and RF to have optimal performance and low sensitivity to parameter changes, while CART was sensitive and had inferior performance [27]. Vittorio Mazzia presented a novel deep learning model for pixel-based land cover and crop classification using multi-temporal Sentinel-2 imagery, achieving high accuracy and reducing the need for manual feature engineering [28].

Carbon neutrality is a key path and regulatory target for adaptation and mitigation of climate change, and forests are the mainstay of urban ecological environment and carbon neutrality. There is currently no relevant research directly using deep learning to predict forest phenology. In this study, the GRU model was employed to train GEI data from three locations in the forest area, effectively predicting the vegetation phenology changes. Based on different irrigation conditions, the GRU model was separately trained for prediction, and the accuracy of the results under different irrigation conditions was discussed. These results can assist forestry and water resource management personnel in better understanding the growth and irrigation needs of forest areas, thus enabling more precise development of management and protection strategies, thereby improving forest productivity and water resource utilization efficiency.

2. Method

2.1. Profile of the Study Area

The study area is located in Guotang Forestry Farm, Gaotang County, Shandong Province, China (116°5′25″ E, 36°48′47″ N), with an altitude of 30 m (as shown in Figure 1). The area has a temperate, semi-humid continental monsoon climate with significant seasonal variations and is one of the main cultivation areas for Chinese white poplar. The annual average precipitation in this area is 544.7 mm, mainly concentrated in July and August, and the annual evaporation is 1880 mm. The average groundwater level is approximately 6.2 m [29].

2.2. Study Object

This study used the triploid clone B301 of Populus tomentosa, which was planted in the spring of 2015 for afforestation [30]. The trees were planted in a uniform pattern with a spacing of 2 m × 3 m, resulting in a forest density of 1667 plants/hm⁻², covering an area of 10,296 m². Five irrigation points were set up in a completely randomized design, with six experimental replicates.

2.3. Experimental Design

At the experimental site of the forest farm, three groups of different irrigation methods and amounts were designed, including rain-fed (no irrigation), sufficient drip irrigation, and sufficient furrow irrigation, to cultivate the poplar B301. The irrigation was conducted from 25 March when the poplar buds began to swell until the arrival of the rainy season on 26 June. Time-series images were acquired using near-ground remote sensing equipment, and the three irrigation methods were calibrated as preset points 1, 3, and 5.

The fully drip-irrigated (DIFI) system employed drip irrigation tubes produced by Netafim, Israel, with a drip head flow rate of 1.6 L·h⁻¹ and a spacing of 50 cm between drip heads, placed on the ground surface in a row-and-two-belt layout along the tree row direction (with drip irrigation pipes placed 30 cm away from the tree on each side). During the growing season, based on the quantitative relationship between the growth of the white poplar and the soil moisture effectiveness [31], the fully drip-irrigated system was set to begin irrigation when the soil water potential at a depth of 20 cm directly below the drip head reached around −18 kPa (i.e., 79% of the field water holding capacity θf and 73% of the soil moisture effectiveness rθ) and to stop when the soil moisture content within the soil wetting zone was raised to the field water holding capacity [29,32].

The system of full furrow irrigation (BIFI) uses PVC pipes to ensure synchronous water supply to each canal and installs tension sensors directly below the canals to record soil water potential. Groundwater is used for irrigation, and the water flow is automatically controlled by a water meter installed at the top of the valve. Irrigation is stopped when the average soil moisture content in the shallow soil layer (50 cm) reaches the field capacity θf, and the water meter reading is recorded. In contrast, the control group (CK) receives no irrigation treatment and is left unwatered throughout the year.

2.4. Experimental Monitoring Equipment

The high-definition network digital camera DS-2DC4223IW-D/GLT(C) manufactured by China Hangzhou Hikvision Digital Technology was utilized as the remote sensing equipment for capturing the growth status of Populus tomentosa B301 in this experiment. It is equipped with a 1/2.8″ progressive scan CMOS sensor, with a maximum resolution of 1920 × 1080 pixels, and a lens with a horizontal viewing angle of 57.6–2.7° (wide angle-telephoto), 360° horizontal rotation, and −15° to 90° vertical direction (automatic flipping). The camera can operate normally in a temperature range of −30 °C to +65 °C and supports various functions, including 23× optical zoom, 16× digital zoom, 3D digital noise reduction, and strong light suppression.

The camera was installed at a height of 20 m above ground level on the upper end of a state-owned forest meteorological station, facing due north (as shown in Figure 2). Due to its height, the camera is susceptible to external environmental factors such as wind and rain, which may cause inevitable movements and changes in its field of view, necessitating azimuthal and image registration calibration. Additionally, the camera was set to “automatic mode” for exposure and white balance adjustments. It was programmed to automatically scan three fixed preset points in the forest (Preset point 1, 3, and 5), corresponding to three different irrigation areas: CK (rain-fed), fully drip-irrigated, and fully furrow-irrigated. The images were transmitted via the China Mobile 4G signal network, using the FTP network protocol (IP address 39.102.35.120) to transfer them to an Alibaba Cloud server. To minimize the effects of different solar angles, images were taken every hour between 7:03 AM and 7:00 PM local time, resulting in a time interval of 1 h. The images were named based on the place and time of shooting and were automatically saved in JPEG format without compression, forming a time series. A total of 10,296 images were selected for extracting vegetation indices, covering the period from 24 March 2021 to 3 January 2022 [29].

2.5. Phenological Evaluation Indicators

2.5.1. Region of Interest Selection

The study selected the diagonal center points of different irrigation zones in preset points 1, 3, and 5 as regions of interest (ROI) as shown in Figure 3, with irrigation equipment located below. Images were processed using a custom algorithm developed in Python and executed in the PyCharm compiler (Version: PyCharm 2020.2.2). Due to varying lighting conditions, rain, snow, fog, and condensation on camera housing windows, the quality of the images occasionally suffered. To ensure objectivity of the results, no selective editing or manual enhancement was performed on any archived images, and no smoothing or filtering was applied to the resulting time series analysis. The mean brightness values (digital number, DN) of the R, G, and B bands in the image ROI were extracted using Equation (1). To mitigate the effects of variations in illumination on color balance, the mean brightness values of the R, G, and B bands were averaged on a daily basis using a 1-day time interval.

$${\mathrm{DN}}_{\mathrm{channel}}={\left(\mathrm{sumDN}\right)}_{\mathrm{channel}}/{\mathrm{N}}_{\mathrm{channel}}$$

(1)

DN represents the value of each wave band, ${\left(\mathrm{sumDN}\right)}_{\mathrm{channel}}$ indicates the sum of all pixel brightness values of a certain band in the image, and ${ \mathrm{N}}_{\mathrm{channel}}$ indicates the number of pixels in the ROI.

2.5.2. Vegetation Index Calculation

The vegetation index is used to describe the greenness of each pixel by the spectral reflection and absorption characteristics of the vegetation and accurately reflects the vegetation growth status. In this study, we selected and calculated the time series of the Greenness Index (GEI) to quantify the dynamics of the vegetation canopy.

$$\mathrm{GEI}=2\mathrm{G}-\left(\mathrm{R}+\mathrm{B}\right)$$

(2)

The selection of the GEI dataset was based on the fact that it was complete and did not contain any missing values, allowing for direct calculation using the raw data. After obtaining the complete raw data, the calculation and analysis were performed using the Numpy [33], Panases [34], and Scijit-Learn [35] software packages in Python3.6.

2.6. Pre-Built Model Selection for Vegetation Index

The GRU model consists of two control gates (Figure 4): the update gate (${\mathrm{Z}}_{\mathrm{t}}$) and the reset gate (${\mathrm{r}}_{\mathrm{t}}$) [36], which utilize an update gate to perform the functions of both the forget gate and the input gate and utilize the reset gate to directly act on the previous hidden state. The reset gate determines how to combine the new input information with the previous memory, while the update gate defines the amount of previous memory to be preserved up to the current time step. Essentially, these two gate control vectors determine which information can ultimately be output from the gated recurrent unit. The unique feature of the GRU model is its ability to maintain information in long-term sequences without clearing or removing it over time, even if it is not relevant to the prediction. The update equation is computed as follows:

$${\mathrm{Z}}_{\mathrm{t}}=\mathsf{\sigma }\left({\mathrm{W}}_{\mathrm{zx}}{\mathsf{\chi }}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{zh}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{z}}\right)$$

(3)

$${\mathrm{r}}_{\mathrm{t}}=\mathsf{\sigma }\left({\mathrm{W}}_{\mathrm{rx}}{\mathsf{\chi }}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{rh}}{\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{b}}_{\mathrm{r}}\right)$$

(4)

$${\tilde{\mathrm{c}}}_{\mathrm{t}}=\tanh \left({\mathrm{W}}_{\mathrm{cx}}{\mathsf{\chi }}_{\mathrm{t}}+{\mathrm{W}}_{\mathrm{ch}}\left({\mathrm{r}}_{\mathrm{t}}{\ast \mathrm{h}}_{\mathrm{t}-1}\right)+{\mathrm{b}}_{\mathrm{c}}\right)$$

(5)

$${\mathrm{c}}_{\mathrm{t}}=\left(1-{\mathrm{Z}}_{\mathrm{t}}\right){\mathrm{c}}_{\mathrm{t}-1}+{\mathrm{Z}}_{\mathrm{t}}{\tilde{\mathrm{c}}}_{\mathrm{t}}$$

(6)

In the formula, ${\mathrm{Z}}_{\mathrm{t}}$ is update gate, ${\mathrm{r}}_{\mathrm{t}}$ is reset gate. ${\mathrm{c}}_{\mathrm{t}}$ is the output vector of the hidden layer at time t. ${\tilde{\mathrm{c}}}_{\mathrm{t}}$ is the updated candidate vector. The sigmoid function maps the result between 0–1. The closer it gets to 1, the easier it is to retain information.

2.7. Model Performance Evaluation Metrics

The mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) were selected to accurately quantify the prediction performance of the model, which is used as the basis to judge the prediction effect of the model and to measure the performance of the system [37,38]. The smaller the matrix value, the better the model and the higher the prediction accuracy

$$\mathrm{MSE}=\frac{1}{\mathrm{N}}{{\displaystyle \sum }}_{\mathrm{t}-1}^{\mathrm{N}}{\left({\mathrm{y}}_{\mathrm{t}}-\overline{{\mathrm{y}}_{\mathrm{t}}}\right)}^2$$

(7)

$$\mathrm{RMSE}=\sqrt{\frac{1}{\mathrm{N}}\sum _{\mathrm{t}-1}^{\mathrm{N}}{({\mathrm{y}}_{\mathrm{t}}-\overline{{\mathrm{y}}_{\mathrm{t}}})}^2}$$

(8)

$$\mathrm{MAE}=\frac{1}{\mathrm{N}}{\displaystyle \sum }_{\mathrm{t}=1}^{\mathrm{N}}\left|\left({\mathrm{y}}_{\mathrm{t}}-\overline{{\mathrm{y}}_{\mathrm{t}}}\right)\right|$$

(9)

where N is the number of predicted points, ${\mathrm{y}}_{\mathrm{t}}$ is the true value of the predicted point, and $\overline{{\mathrm{y}}_{\mathrm{t}}}$ is the predicted value of the predicted point.

3. Results and Analysis

3.1. Extracting GEI

GEI data was extracted from the ROI using Equation (2), and the start of the forest growing season (SOS), the peak growing season and onset of senescence (COS), and the end of the growing season (EOS) were obtained as the 98th, 262nd, and 332nd day of the year, respectively. The GEI data accurately reflects the forest phenology growth and senescence process, which corresponds to three stages of vegetation phenology: the first stage is the rapid growth period, where GEI data shows a rapid upward trend; the second stage is the mature period, where GEI data remains relatively stable with only small fluctuations; and the third stage is the senescence period, where GEI data decreases sharply to its lowest value (Figure 5).

The results showed that the GEI data were influenced by environmental factors and diurnal differences, but there were obvious seasonal variations. The forest was in a dormant period, and the GEI data, except for the impact of snowfall, had a relatively flat overall change trend. Under the influence of rising temperature and increasing precipitation, the forest gradually began to germinate. GEI data rose and vegetation began to enter the forest growing season, the time point at which the forest reached its peak and the GEI value reached a high value until the 98th day of the year, which lasted about 62 days. The peak growing season and onset of senescence lasted until 19 September (262 days). Thereafter, the forest activity gradually weakened with the decrease of temperature, and the GEI data also decreased. On the 332nd day of the year, the end of the growing season arrived. The forest was all withered and the GEI data returned to the low value again.

3.2. Model Comparison Analysis

To evaluate whether the GRU model can be applied to forest phenology prediction, we conducted a quantitative analysis training using the GRU model. To visualize the training and prediction results more intuitively, the actual values and predicted values of the model were processed visually (Figure 6).

The results in Figure 6 show that the red part represents the fitting effect of the model on the training set, while the blue part represents the prediction results of the model on the test set. Comparing with Figure 5, it is found that the GRU model has a good fitting effect on both the training and test sets, which conforms to the trend of time series data and presents annual periodic and long-term trend components. Furthermore, from the perspective of the three preset points, the GRU predicted values are in good agreement with the real values, indicating that the GRU model performs well.

3.3. Model Prediction Data Analysis

To further verify the distribution of the predicted data, the GRU model’s predicted data results were analyzed using histograms and kernel density estimation (KDE) plots. The frequency of the histogram was converted to a frequency by incorporating the results into a Python package editing program, resulting in a histogram KDE plot (Figure 7).

The histogram and kernel density plot in Figure 7 exhibit a symmetric and uniform distribution, indicating a stable and consistent pattern in the data. The distribution of the data aligns with the GRU model data prediction in Figure 6, with the difference between test set values and real values being relatively stable. From the three preset points, the data distribution graphs of the GRU model demonstrate a good fit, indicating that the test data can be applied to phenology prediction models.

3.4. Model Stability Test

The stationarity of a time series is a prerequisite for time series analysis, which is required in the law of large numbers and the central limit theorem. Many modeling processes are established under the premise of these theorems, and many conclusions obtained are unreliable if these conditions are not met. Stationarity testing can be performed using two methods: the observation method and the unit root test method. The observation method examines whether the trend and correlation graph of the sequence exhibit certain periodic factors with time. The test for stationarity was conducted using the observation method, and a statistical test module named teststationarity was edited in Python to construct an autocorrelation graph, limiting the lag values on the x-axis to 20 to more intuitively display the statistical test results (Figure 8).

The results showed that the autocorrelation coefficients gradually approached zero in Figure 8. Among the three preset points, preset point 1 exhibited an overall tendency to approach zero, preset point 3 underwent oscillations before approaching zero, and preset point 5 exhibited small disturbances after approaching zero. All three preset points demonstrated that the data of the GRU model on the training and testing sets stabilized as the time series progressed. The time series tested by the GRU model exhibited stationarity.

3.5. Model Performance Evaluation

In order to better validate the accuracy of the proposed model and evaluate the goodness of fit of the model within the sample, performance evaluation was conducted on the GRU neural network model. The adaptive moment estimation method was used for optimization, and the loss function was evaluated using the MSE, RMSE, and MAE methods. The results of the GRU model for the three preset points are compared in

Table 1. Performance evaluation.

Evaluation Indicators	Preset Point 1	Preset Point 3	Preset Point 5
MSE/10−5	9.055	12.91	5.349
RMSE/10−3	9.516	11.36	7.313
MAE/10−3	7.174	8.241	5.351

.

Model validation is an important step to evaluate the performance of a system by comparing actual data with predicted data. As shown in Table 1, the MSE of preset point 1 is 9.055 × 10⁻⁵, RMSE is 9.516 × 10⁻³, and MAE is 7.174 × 10⁻³; the MSE of preset point 3 is 12.91 × 10⁻⁵, RMSE is 11.36 × 10⁻³, and MAE is 8.241 × 10⁻³; the MSE of preset point 5 is 8.241 × 10⁻⁵, RMSE is 7.313 × 10⁻³, and MAE is 5.351 × 10⁻³. The high goodness-of-fit between the predicted values and the real values indicates that the GRU model can achieve accurate predictions. The MAE and RMSE are absolute errors and are not affected by the size of GEI. The small differences in performance indicators between the GRU model and the comparative evaluation suggest that the GRU model can achieve precise predictions.

3.6. GEI Data Prediction

The time series dataset was split into training and testing sets using an 80–20 split in the model. The moving average method was used for prediction, which involves selecting a sliding window size. A simple moving average model was used to predict the next value in the time series based on a fixed number of past values. The model was trained by iterating over all data 100 times and minimizing the error between the actual and predicted values during the training period. For the prediction of future 60-day data, the predicted values were obtained in a Jupyter notebook environment, and the results were plotted (Figure 9).

The model predicts the future results for the next 60 days, and the GRU model performs well as the predicted values are close to the real values. Upon analysis of three preset points, it can be observed that the error between the predicted and real values of GEI for the next 60 days is minimal. Furthermore, the predictive performance of the GRU model demonstrates remarkable alignment with the ground truth data. Based on the forest growth conditions in 2021, the forest was in a dormant period for the next 60 days. The GEI data showed a relatively flat trend with minor fluctuations due to factors such as snowfall. The forest growth trend predicted by the GRU model is consistent with the actual growth patterns, indicating its efficacy in predictive modeling. The minor differences observed in the predicted growth trend for the 60-day period in 2022 suggest that the GRU model has achieved a high degree of accuracy in its predictions.

4. Discussion

Modeling and predicting with limited data can be quite challenging. In this study, we utilized the GRU model to model GEI data and predict future forest phenology growth patterns. The results indicate that the GRU model is capable of predicting phenology with remarkable accuracy, aligning with our initial hypotheses. To elucidate this phenomenon, we extracted GEI data from forest images and attempted to predict it using the GRU model. We then validated our analysis using histograms, kernel density plots, and autocorrelation plots. Furthermore, we evaluated the model performance and predictive capabilities. The results indicate that: (1) the GRU model, after being trained and tested on the dataset, exhibited good similarity with the actual values, fulfilling the predictive requirements. (2) Validation, conducted through histograms, kernel density plots, and autocorrelation plots, demonstrated the reasonability of the testing data and confirmed the stability of the predictive model. (3) Model performance evaluation confirmed that the model accuracy meets the predictive requirements. (4) The GRU model is capable of predicting the forest’s trends for the next 60 days.

The GRU model has become a phenological prediction model. While many scholars have made predictions regarding other aspects of forestry [39,40], phenology has not been given sufficient attention, even though it plays a crucial role in reflecting the impact of environmental changes on forest growth [41]. To demonstrate the predictive capabilities of GRU, we trained and tested the mode and found that it was capable of effectively predicting trend changes. The discrepancy between the real and predicted values observed in Figure 6 can be attributed to the memory control mechanism of the model. Specifically, the GRU model directly passes the information to the next unit without any control, and, in addition, it uses the reset gate to regulate the information from the previous time step when calculating the new memory h^(t).

The histograms and kernel density plots correspond correctly to the distribution of the data. For continuous variable data, histograms are constructed to understand the distribution, and kernel density plots are drawn to visualize the potential probability distribution of the data. The results showed that the distribution of the testing dataset of the GRU model was relatively consistent with the distribution of the histogram, and the data distribution of the GRU model remained constant and uniform. To demonstrate the probability density function of the random variable of the entire dataset, kernel density plots were used for analysis. It was found that the GRU model data showed a uniformly distributed and symmetrical distribution, indicating that the values were consistent and that the sample repeatability was good. This further confirms that the GRU model can be used as a phenology prediction model.

Under the time series data prediction model, the correlation between predicted values and real values at different time periods was analyzed using autocorrelation analysis, which is used to measure the correlation between current sequence values and past sequence values [42]. The results showed that the autocorrelation coefficient values of the GRU model decreased smoothly, indicating that the model is very stable. This indicates that the GRU model measures a strong correlation and has good stability between the current and past sequence values. This is due to the simple structure of the GRU model, which leads to relatively greater stability.

To evaluate the performance of the proposed model, real data was compared with predicted data using performance metrics such as MSE, RMSE, and MAE [43]. These metrics were chosen to measure the error rate of the model and provide a measure of the accuracy of the system’s performance. The goodness of fit of any model can be explained by its error rate, and, therefore, MSE, RMSE, and MAE were used to assess the accuracy of the proposed model. The results showed that the values of MSE, RMSE, and MAE for the GRU model range from 0 to 1, indicating high prediction accuracy. The performance of the GRU model in phenology prediction is good.

The GRU model predicted 60 days of data and was consistent with the trends of forest growth. It was observed that the overall trend of the GRU model showed a relatively flat change The reason for this is that the GRU model introduces addition when updating from t to t−1, which prevents gradient dispersion and mitigates the problem of gradient vanishing. Therefore, it can predict the changes in phenological growth. The GRU model has fewer parameters, faster computation, and reduced overfitting risk, which allows it to anticipate changes ahead of time.

5. Conclusions

Based on the study of short-term prediction of phenological events using the GRU model on ROI using forest images collected by low Earth-orbit remote sensing satellites, it can be concluded that the GRU model accurately simulates the process of forest phenological changes by using the GEI time series data in the input forest images. The GRU model has the ability to filter out redundant information when simulating the temporal phenological relationship and shows high accuracy in phenological prediction. The GRU model merges the forget gate and input gate into the update gate, which controls information updating and retention and sets the reset gate to control the contribution of old hidden state to the calculation of new hidden state. Therefore, the GRU structure is simple and can show the trend of changes in forest phenology prediction in advance. In practical applications, the faster the prediction of forest phenology results, the more effective it is to take corresponding forest maintenance measures in advance, and forest phenology prediction is of great significance for studying climate change.