Plant and Animal Ecology

Research on Forest Phenology Prediction based on LSTM and GRU Model

  • GUAN Peng , 1, 2, 3 ,
  • ZHENG Yili , 1, 2, 3, *
  • 1. School of Technology, Beijing Forestry University, Beijing 100083, China
  • 2. Beijing Laboratory of Urban and Rural Ecological Environment, Beijing Municipal Education Commission, Beijing 100083, China
  • 3. Key Lab of State Forestry Administration for Forestry Equipment and Automation, Beijing 100083, China
*ZHENG Yili, E-mail:

GUAN Peng, E-mail:

Received date: 2021-07-15

  Accepted date: 2022-01-20

  Online published: 2023-01-31

Supported by

The Fundamental Research Funds for the Central Universities(2021ZY74)


Research on forest phenology is an important parameter related to climate and environmental changes. An optical camera was used as a near-earth remote sensing satellite device to obtain forest images, and the data of Green excess index (GEI) in the images were calculated, which was fitted with the seasonal variation curve of GEI data by double Logistic method and normalization method. LSTM and GRU deep learning models were introduced to train and test the GEI data. Moreover, the rationality and performance evaluation of the deep learning model were verified, and finally the model predicted the trend of GEI data in the next 60 days. Results showed: In the aspects of forest phenology training and prediction, GRU and LSTM models were verified by histograms and autocorrelation graphs, indicating that the distribution of predicted data was consistent with the trend of real data, LSTM and GRU model data were feasible and the model was stable. The differences of MSE, RMSE, MAE and MAPE between LSTM model and GRU model were 0.0014, 0.013, 0.008 and 5.26%, respectively. GRU had higher performance than LSTM. The prediction of LSTM and GRU models about GEI data for the next 60 days both showed a trend chart consistent with the change trend of GEI data in the first half of the year. GRU and LSTM were used to predict GEI data by deep learning model, and the response of LSTM and GRU deep learning models in forest phenology prediction was realized, and the performance of GRU was better than that of LSTM model. It could further reveal the growth and climate change of forest phenology in the future, and provide a theoretical basis for the application of forest phenology prediction.

Cite this article

GUAN Peng , ZHENG Yili . Research on Forest Phenology Prediction based on LSTM and GRU Model[J]. Journal of Resources and Ecology, 2023 , 14(1) : 25 -34 . DOI: 10.5814/j.issn.1674-764x.2023.01.003

1 Introduction

Long-term monitoring of major forest phenological periods can accurately predict and explain the carbon cycle of ecosystem, which is closely related to the future forest environment and climate change. Many predictive models have been developed to explain historical trends in plant phenology and to predict future trends (Yun et al., 2017; Asse et al., 2020). For example, satellite remote sensing and gridded data can predict vegetation phenology models through machine learning (Czernecki et al., 2018). A phenological camera was installed in the carbon-water flux network to obtain the phenological periods of leaf emergence and deciduous leaves, and establish phenological models related to canopy leaf area, chlorophyll and phylerubin (Wingate et al., 2015). The phenological observation test data of rubber tree were used as the basis to combine the crop growth bell model and establish the prediction model of spring phenological periods of rubber tree (Li et al., 2020). Broad-leaved Korean pine forest in Changbai Mountain was used as the research object to explore the role of digital camera images in specie-scale phenology simulation and community-scale phenology model improvement (Zhou et al., 2013). The prediction models of climatic factors such as illumination, temperature, rainfall and phenological period of Korean pine were established to provide accurate prediction of phenological period (Sun and Yu, 2006). Machine learning models was used to predict daily pixel-level flowering rates, and the resulting models could accurately predict flowering seasons (Dixon et al., 2021).
Machine learning models (Gao et al., 2020) mainly include artificial neural network models (ANN) (Chua et al., 2008; Kan et al., 2015), neuro-fuzzy models (Nayak et al., 2005; Mukerji et al., 2009) and support vector machine models (SVM) (Chen and Yu, 2007; Lin et al., 2009; Zhang et al., 2013). Machine learning models does not have any prior knowledge and it is difficult to determine the optimal time step. When machine learning algorithm prediction is used, it needs to screen the best effect of time step in prediction so as to obtain the best prediction accuracy. The development of deep learning technology presents an opportunity to simplify this problem. At present, deep learning has achieved great success in many fields, especially in computer vision and natural language processing (Kim, 2014). Deep learning methods have the ability to process long time series data. For example, recurrent neural networks (RNN) is one of the most popular neural networks in the field of deep learning. RNN has been proved to be an effective tool for processing time series prediction. However, the performance of traditional RNN in sequence prediction is not significantly improved due to the inherent problems of gradient disappearance or explosion. Long Short-Term Memory networks (LSTM) is optimized based on RNN to solve this problem. Many studies have shown that LSTM performs better than traditional RNN. In recent years, it has appeared in Gated Recurrent Unit Networks (GRU) based on LSTM optimization (Cho et al., 2014; Salinas et al., 2020), and performed well in the field of natural language processing. GRU has a simpler structure and higher operating speed (Greff et al., 2017). LSTM and GRU network models have been successfully applied to the prediction of industrial, medical and agricultural yields (Jia et al., 2020; Feng, 2021). However, the deep learning model prediction of forest phenology has not been reported.
This study proposed the following two hypotheses. 1) LSTM and GRU models can be applied to forest phenology prediction. 2) GRU is better than LSTM. Therefore, the 2018-2019 forest growth status of Louisiana, USA was selected for the study. The first item is that the absolute Green excess index (GEI) data extracted from forest images by near-earth remote sensing satellite is used to verify whether the LSTM and GRU models can be used for short-term phenological prediction. The second purpose is that histograms and kernel density maps, autocorrelation and performance indicators are used to evaluate the rationality of the model. The third objective is to extend the prediction period to 60 days and explore the advantages of LSTM or GRU models in short-term phenological prediction.

2 Material and methods

2.1 Overview of the study area

This study was carried out in Louisiana (91°58′05″W, 32°58′34″N), which is located along the Gulf of Mexico, bordering Arkansas in the north, Texas in the west, Mississippi in the east, and the Gulf of Mexico in the south. It is one of the main producing areas of wood, paper products, and wooden boards in America. The forest area is 590 ha, accounting for 47% of the land area of the state. This state is in the subtropical moist monsoon climate zone, and the average growth period is 220-320 days.

2.1.1 Data acquisition

In this study, commercial webcams were selected as near-earth remote sensing satellite equipment (Deng et al., 2019) to obtain time series images. Shooting will take place from 7:05 to 17:05 every day. The image was saved in uncompressed JPEG format. The resolution of the image was 2560×1440 pixels. Raw images were transmitted wirelessly and stored on servers named after when and where they were taken. In this study, a total of 4260 images were selected from November 1, 2017 to December 31, 2018.
No archive images have been selectively edited or artificially enhanced. It tries to maintain this level of objectivity. The Region of Interest (ROI) in each image was defined as the study area (Fig. 1). The PyCharm compiler was used by the Python language’s own algorithms to process many hours of images of an “image processing tool”. The R, G and B brightness values of the region of interest of each frame were extracted from the image. In order to reduce camera colors to balance changes due to fog and shadows, an average method of 1D was used for a period of time and an average of daily RGB brightness values for all photos. Equation 1 is calculated to obtain GEI data time series diagram and to quantify vegetation canopy dynamics. Second, the data is read to carry out the bilogistic model and normalization (Eq.2-3). The normalized data is in the range [01].
In the formula, GEI is Green excess index, R, G and B represent the brightness values of red, green and blue channels respectively. g (t) is the curve that characterizes the seasonal changes of vegetation, m is the minimum value of the year, W is the maximum value of the year, t is time, S and A are the inflection points representing the start and end of season (SOS, EOS), respectively; mSx and mAx are the rates at points S and A, indicating the greenup rate in spring (RSP) and the senescence rate in autumn (RAU), respectively. Xnorm, Xi, Xmin, and Xmax are the normalized, observed, minimum, and maximum values of the absolute greenness index, respectively.
Fig.1 Region of interest in different season
A GEI dataset with complete and no missing values was selected, which was calculated directly from the original data. In Python 3.6 software, Numpy (van der Walt et al., 2011), Panases (McKinney, 2010) and Scijit-Learn (Pedregosa et al., 2012) LSTM and GRU models were run on the basis of software package. Data sets were defined and made. It divided the building model data sets into two categories. In the first category, data of 10 months from November 2017 to August 2018 were used as a training set for the training model. The second category were the August to December months data during 2018 as the test set. The cross-validation method was adopted (Basler, 2016; Sirsat et al., 2019) to adjust neural network hyperparameters.

2.1.2 Long Short-Term Memory (LSTM)

LSTM model is similar to the chain structure of traditional RNN. But the internal operations of LSTM are more complex. Compared with traditional RNN, LSTM solves the gradient vanishing and gradient explosion problems in time series training. The LSTM neural network (Xiong et al., 2020) maintained and regulated the unit state (Ct) and hidden state (ht), forgetting gate (ft), an input gate (it), and an output gate (Ot) (Fig. 2). The forgetting gate (ft) determines what information will be moved out of the cell state (Ct). The input gate (it) determines what new information will be stored in the unit state (Ct). The output gate (Ot) specifies what information from the cell state is used as the output (Ot). Neurons can capture complex correlations in short and long term time series through three gate functions. This is a significant improvement. The calculation of unit state (Ct) and hidden state (ht) of LSTM neural network A is shown in formula 4-9 below.
$C_{t}=f_{t}\odot C_{t-1}+i_{t} \odot \bar{C}_{t}$
$h_{t}=O_{t} \odot tanh(C_{t})$
In the formula, ft is the control function of the forgetting gate, Wf is the weight matrix of the true value of the canopy color index of the forget gate phenology, xt represents the input at time t, ht-1 is the predicted value of the phenological canopy color index at the previous moment, bf is the forgetting gate deviation vector; it is the control function of the input gate, Wi is the weight matrix of the true value of the color index of the input gate phenology observation, ht-1 is the state of the hidden layer at the previous moment, [ht-1, xt] is the long vector connected by two vectors, bi is the input gate deviation vector; $\bar{C}_{t}$ is the conversion value of the tanh function of the actual value of the color index at the current time, as the color index vector at the current time, Wc is the weight matrix of the stored value of the phenological canopy color index at the current moment; bc is current time deviation vector. Ct of the output gate is used to update the state of the neuron. ⊙ is the vector element multiplication operation conforms to; tanh uses the tangent hyperbolic function to scale the value into the range of [–1,1]; The control functions Ot, Wo is the output weight matrix, bo is the bias of the output gate. σmeans the activation function sigmoid, which is taking the value [01].
Fig.2 LSTM neural network structure

2.1.3 Gated Recurrent Unit (GRU)

The GRU model (Ayzel and Heistermann, 2021) is structured with two control gates (Fig. 3), including a renewal gate (Zt) and a reseted gate (rt). An update gate is used to implement the functions of the forget gate and the input gate in the LSTM. Reset doors are utilized to act directly on the front hidden state. The reset gate determines how the new input information is combined with the previous memory. The update gate is defined as the amount of previous memory saved by the current time step. These two gated vectors determine what information is ultimately available as the output of a gated loop unit. The GRU model is characterized by the ability to preserve information in long-term sequences and not to be removed over time or irrelevant to the prediction, calculated as following Eq. (10)-(13).
In the formula, Zt is update gate, rt is reset gate. t is the updated candidate vector. Ct is the output vector of the hidden layer at time t. xt is the input vector at times t; r is the reset gate vector at time t; Zt is the update gate vector at time t; ht is the output vector of the hidden layer at time t, $\widetilde{c}_{t}$is the updated candidate vector; Wch, Wcx, Wzx, Wzh, Wrx, Wrh are the weight matrix between each connection vector; σ is the sigmoidfunction. The sigmoid function maps the result between 0-1. The closer it gets to 1, the easier it is to retain information.
Fig. 3 GRU neural network structure

2.1.4 Test environment configuration

The hardware platform for training and testing is configured as Intel (R) Core (TM) i5-4460 CPU @3.20 GHz processor, 8GB memory, 11GB Ge force GTX1080 Ti graphics card, 500G solid state hard disk and 2 TB mechanical hard disk. Python programming is implemented under the Ubuntu16.04 system and Tensorflow framework, and CUDA, Cudnn and Open CV third party libraries are invoked in the procedure.

2.2 Performance evaluation index

The mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean percentage error (MAPE) was 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 (Kırbaş et al., 2020; Saiful et al., 2020). The smaller the matrix value, the better the model and the higher the prediction accuracy. The GRU model is characterized by the ability to preserve information in long-term sequences and not to be removed over time or irrelevant to the prediction, calculated as follows Eq.14-17.
where, N is the number of predicted points. yt is the true value of the predicted point.$\bar{y}_{t}$is the predicted value of the predicted point.

3 Results

3.1 Extract data from GEI

ROI uses Eq.1-3 to extract GEI data. It was obtained that the beginning of the forest growing season (SOS) and the time of reaching exuberance, the beginning of the litters (COS) and the end of the growing season (EOS) were the 67, 108, 288 and 337 days of the annual sequence days, respectively. The GEI data accurately reflect the forest phenological growth and litter processes. The whole process correspond to three phases of vegetation phenology. The first stage is the rapid growth period. GEI data shows a rapid upward trend. The second stage are the maturity stage. All GEI data remain relatively stable only in the period of small fluctuations. The third stage are the litters. A period during which GEI data dropped significantly to a minimum (Fig. 4).
The results showed that the GEI data were influenced by environmental factors and diurnal differences, but there were obvious seasonal variations. The forest is dormant until the 67th day of the year. Except for the influence of snowfall, the overall change trend of GEI data is relatively gentle. Under the influence of rising temperature and increasing precipitation, the forest gradually began to germinate. GEI data rose and vegetation SOS began. The time point at which the forest reached its peak and the GEI value reached a high value until the 108th day of the year, which lasted about 57 days. The forest fall period began (COS) until 15 October (288 days). Thereafter, the forest activity gradually weakened with the decrease of temperature and the GEI data decreased. The forest growing season (EOS) endson day 337 of the year. The forest was all withered and the GEI data returned to the low value again.

3.2 Contrastive analysis

In order to verify whether the two models are suitable for forest phenology prediction, LSTM and GRU models were used in the same training set for quantitative analysis. In order to observe the training prediction results more intuitively, the actual and predicted values of the model are processed visually (Fig. 5).
Fig. 5 Test results of (a) LSTM model and (b) GRU model

Note: The values 16 and 2 in the frame in the figure represent the difference between the true value and the predicted value.

Figure 5 shows the fitting effect of the model on the training set in the part left of the red dotted line. The right of the dashed blue line is the model’s predictions on the test set. Compared with Fig. 4, LSTM and GRU models have good fitting effects on both the training set and the test set. This is in line with the trend of temporal data, which presents an annual cycle component and a long-term trend component. Secondly, the prediction chart shows that the maximum difference between the predicted LSTM data and the real value data is 16. The maximum difference between the predicted GRU data and the real data is 2, indicating that the predicted GRU data is closer to the actual data than the LSTM data.

3.3 Forecast data analysis

In order to further verify the prediction data distribution results, the prediction data of LSTM and GRU models were analyzed by histogram and kernel density curve (KDE). They were substituted into the Python package editor. The kernel density of the histogram can be obtained by transforming the frequency of the histogram into frequency (Fig. 6).
Fig. 6 Data distribution graph based on (a) LSTM model and (b) GRU model
The results of Fig. 6a show that the histogram structure distribution is not symmetrical and the kernel density curve is not uniform. The data distribution is consistent with the data prediction of LSTM model in Fig. 5a. The difference between the value of the test set and the actual value gradually increases with time and then decreases. The histogram structure and the kernel density curve of the results of Fig. 6b show a symmetrical distribution and a uniform and gentle state. The data distribution is consistent with the prediction of GRU model data in Fig. 5b. The difference between the value of the test set and the actual value remains relatively stable. This indicates that both the test data and the real data can be applied to the phenological prediction model.

3.4 Stationary test

The stationarity of series is a prerequisite for time series analysis. The large number theorem and the central theorem require the sample to satisfy the stationarity requirement, that is, the stationarity on time series. If it is not satisfied, many of the conclusions obtained are unreliable. Observation method and unit root method were used to test the stationarity. The observation method is to observe whether the trend graph and correlation graph of a series shows the periodicity factor of a certain time series over time. The observational test was used. It edited a statistical test module called test_stationarity in Python to construct autocorrelation graphs. The delay value of the X-axis was limited to 20 to make the statistical test result more intuitive (Fig. 7).
Fig. 7 Autocorrelation graph of (a) LSTM model and (b) GRU model
As shown in Fig. 7a, the autocorrelation coefficient gradually and steadily decreases. The data of the LSTM model on the training set and the test set run stably with the times series. In Fig. 7b, the autocorrelation coefficient decreases gradually. GRU model data on training set and test set also progressed steadily with time series. The time series of LSTM and GRU model tests are stable. However, the autocorrelation coefficient in Fig. 7a decreases more than that in Fig. 7b, which indicates that GRU is more stable than LSTM model.

3.5 Performance evaluation

In order to better verify the accuracy of the proposed model and evaluate the fitting degree of the model in sample, LSTM neural network and GRU neural network were compared under the same experimental platform and environment. Adaptive matrix was used for estimation and optimization. The loss function was evaluated by MSE, RMSE, MAE and MAPE. The comparison results under the same indicators are shown in Table 1.
Table 1 Performance evaluation
Moudle Inspection criteria
LSTM 0.003 0.054 0.041 17.81
GRU 0.0016 0.041 0.033 12.55
Model validation is an important step in verifying system performance by comparing actual data with predicted data. Mean square error (MSE) refers to the expected value of the square of the difference between the estimated value of the parameter and the true value of the parameter, which can evaluate the change degree of the data. The smaller the MSE value, the better the accuracy of the prediction model in describing the experimental data. Root mean square error (RMSE) is the distance between the expected value and the actual value. The mean absolute error (MAE) can avoid the problem of error cancellation, so it can accurately reflect the actual prediction error and describe accuracy. Mean absolute percentage error (MAPE) is a statistical index commonly used to measure the accuracy of prediction, such as the prediction of time series. The results in Table 1 show that the MSE, RMSE, MAE and MAPE of the LSTM model were 0.003, 0.054, 0.041 and 17.81%, respectively. The MSE, RMSE, MAE and MAPE of GRU model were 0.016, 0.041, 0.033 and 12.55%, respectively. The fitting degree between the predicted value and the actual value was high. MAPE was obviously larger than other values. Because MAPE is a relative error, its size is related to the GEI value range. MAE and RMSE are absolute errors, and their data are not affected by GEI size. LSTM model and GRU model have a small gap in performance indicators. In the evaluation comparison, it is shown that both models can achieve accurate prediction. However, according to the comparison between MAPE value and MAE value, the prediction accuracy of GRU model was higher.

3.6 GEI data forecasting

In the model, the time series data set was divided into training set and test set by 80% and 20%. The moving average method was used for prediction. The next value in the sequence was predicted from the average of a previously fixed finite number, and all data was iterated through the model 100 times. For GEI data prediction, the moving average method was adopted, and the size of the sliding window was set as 30 days, which means using 30 days of data as input and 31 days of values as predictions. Then, gradually, the training speculation.The corresponding weights were quantified by minimizing the error between the actual value and the predicted value of the training. Jupiter Notebook Environment took forecast data for the next 60 days and produced forecast results (Fig. 8).
Fig. 8 Model prediction results of (a) LSTM model and (b) GRU model

Note: The orange region is the prediction curve of LSTM; The blue area is the prediction curve of the GRU.

The results indicated that LSTM model showed that the upward trend gradually stabilized at the beginning and increased significantly on the 70th day. The GRU model showed an increase first, then a gradual decline, then a slow rise, and an increase within 50 days. According to the growth of the forest in 2018, the time for the forest to be dormant is 60 days. Except for snowfall, GEI data showed a gentle trend of change. Then, affected by rising temperatures and increased rainfall, the forest began to sprout. GEI data rose with the start of the growing season. Comparing the forest growth trend presented by LSTM and GRU models with the growth trend in 2018, it was found that the predicted GEI data trend was in line with the actual growth change development trend. The results showed that the LSTM and GRU models predicted little difference in the 60 days growth trend in 2019. GRU model had few structural parameters in terms of convergence time. Predicting trends in advance can reduce overfitting and improve training efficiency.

4 Discussion

Limited data can be quite challenging on modeling and forecasting. In this study, two different models of GEI data modeling predictions were explored to study the future growth of forest phenology. Results LSTM and GRU could realize phenological prediction. Compared with LSTM, GRU had a higher prediction rate and conformed to our initial hypothesis. In order to clarify this phenomenon, we extracted the GEI data from the forest image and predicted it with LSTM and GRU models. The model performance was evaluated and predicted by histogram, kernel density, and autocorrelation analysis. The results are as follows. Firstly, both LSTM and GRU had good similarity with the actual value and meet the prediction requirements after data set training and testing. Secondly, through histogram, kernel density map and autocorrelation diagram verification, the test data was reasonable and the prediction model was stable. Thirdly, through the evaluation of model performance, the accuracy of the model met the prediction requirements. Fourthly, all models can predict the trend of data over the next 60 days. However, GRU showed a phenological trend in advance.
LSTM and GRU models have become phenological prediction models. Many scholars have made predictions on other aspects of forestry (Yang and Ma, 2005; Mei et al., 2009). The deep learning model was not used to predict the phenology. Phenology can reflect changes in the ecological environment (Pei et al., 2011), which is crucial to the growth of forestry. In order to try LSTM and GRU can achieve prediction, it can be found that these two models can be well trained to predict the changing trend of training and testing. In Fig. 5a and 5b, it is found that the numerical difference between the real value and the predicted value is different. The reason for this is that the model controls memory. The LSTM model was transferred to the next unit with output gate control. GRU models, on the other hand, were passed directly to the next unit without any control. Second, the LSTM calculated new memory (t) without any control over the information about the previous moment. Instead, it is implemented independently with the forget gate. The GRU model calculated the new memory (t) by using the information of the previous moment controlled by the reset gate.
The histogram and the kernel density curve correctly corresponded to the data distribution. For continuous variable data, the histogram was constructed to understand the data distribution and the continuous curve of kernel density wase drawn to visualize the potential probability distribution of data. The results showd that the data distribution of GRU and LSTM model test sets was relatively consistent with the histogram distribution. But GRU model data distribution was more constant and uniform. In order to prove the probability density function of the random variable of the data population, the kernel density continuum plot was used for analysis. It is found that the data of LSTM model were symmetrically distributed. The symmetrical distribution of GRU model data was more constant and uniform. This indicates that the two sets of values are basically consistent, and the sample reproducibility is good, which further confirms that LSTM and GRU can be used as phenological prediction models.
The prediction model of time series data was stable. Autocorrelation was used to measure the correlation between current sequence values and past sequence values. In order to verify the cross-correlation between predicted values and true values in different time periods (Liu et al., 2017), autocorrelation analysis was applied. The results showed that the value of LSTM self-phase relation decreased greatly. However, GRU self-correlation values declined gently, indicating that GRU has stronger correlation and better stability between current and past sequence values than LSTM. Because GRU is less structured and more stable than LSTM.
The performance of the proposed model has good accuracy. The actual data and the predicted data were compared and tested, and the performance matrix method (Sun et al., 2020) was chosen to check the error rate of the model to measure the system performance. Because the degree of fit of any model can be explained by its error rate. The MSE, RMSE and MAE of LSTM and GRU models were between 0 and 1, showing that the LSTM and GRU models have high prediction accuracy. However, MAPE values of GRU model was 5.26% lower than that of LSTM model. From the perspective of evaluation value, GRU model performed better in phenology prediction than LSTM model.
Both LSTM and GRU models predicted 60 day data and were consistent with the forest growth trend. The results showed that LSTM model had an upward trend in the first 60 days. GRU showed a trend of rising first and then declining slowly. This is because both LSTM and GRU models introduce addition when updating from T to T-1. This prevents gradient dispersion and alleviates the problem of gradient disappearance. Therefore, it can predict phenological growth changes. However, there are three gate structures of LSTM, which have large number of parameters, slow convergence and slow calculation time. Compared with LSTM, GRU has fewer parameters and is relatively fast in calculation, which reduces the risk of overfitting and the tendency of change in advance.

5 Conclusions

By collecting forest images of near-earth remote sensing satellite, the short-term prediction of ROI phenological events based on LSTM and GRU models was concluded. GRU and LSTM used GEI time series data from input forest images to accurately simulate the process of forest phenology. GRU and LSTM have the ability to filter out redundant information when simulating temporal phenological relationships. Both LSTM and GRU show high precision prediction in phenological prediction. But LSTM is controlled by three gates, namely input gate, forgetting gate and output gate. The GRU combined the forgotten gate and the input gate into the update gate to control the information update and retention. The reset gate was also set to control the contents of the old hidden state and participate in the calculations of the new hidden state. Therefore, GRU is structurally simpler than LSTM. It showed a trend of variation in forest phenology prediction in advance. Phenological prediction model can be used to simulate the change trend of tree species distribution, reconstruct past climate change, predict future climate change, agricultural and forestry production and risk assessment. It is an important way to quantitatively study climate factors and plant phenology. In practice, the faster the prediction results of forest phenology, the more effective the corresponding forest protection measures will be. This has important implications for climate change research.


The authors would like to thank Special Fund for Beijing Common Construction Project. Thank the camera phenology observation network (PhenoCam) for providing the support of digital photo observation data of the site.
Asse D, Randin C F, Bonhomme M, et al. 2020. Process-based models out compete correlative models in projecting spring phenology of trees in a future warmer climate. Agricultural and Forest Meteorology, 285-286: 107931. DOI: 10.1016/j.agrformet.2020.107931.


Ayzel G, Heistermann M. 2021. The effect of calibration data length on the performance of a conceptual hydrological model versus LSTM and GRU: A case study for six basins from the CAMELS dataset. Computers & Geosciences, 149: 104708. DOI: 10.1016/j.cageo.2021.104708.


Basler D. 2016. Evaluating phenological models for the prediction of leaf-out dates in six temperate tree species across central Europe. Agricultural and Forest Meteorology, 217: 10-21.

Chen S T, Yu P S. 2007. Pruning of support vector networks on flood forecasting. Journal of Hydrology, 347(1-2): 67-78.


Cho K, van Merrienboer B, Gulcehre C, et al. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg,

Chua L H C, Wong T S W, Sriramula L K. 2008. Comparison between kinematic wave and artificial neural network models in event-based runoff simulation for an overland plane. Journal of Hydrology, 357(3-4): 337-348.


Czernecki B, Nowosad J, Jabłońska K. 2018. Machine learning modeling of plant phenology based on coupling satellite and gridded meteorological dataset. International Journal of Biometeorology, 62(7): 1297-1309.


Deng L J, Lin Y, Yan L, et al. 2019. Urban plant phenology monitoring: Expanding the functions of widespread surveillance cameras to nature rhythm understanding. Remote Sensing Applications: Society and Environment, 15: 100232. DOI: 10.1016/j.rsase.2019.05.001.


Dixon D J, Callow J N, Duncan J M A, et al. 2021. Satellite prediction of forest flowering phenology. Remote Sensing of Environment, 255: 112197. DOI: 10.1016/j.rse.2020.112197.


Feng Z S. 2021. An image detection method based on parameter optimization of support vector machine. Circuits Systems and Signal Processing, 15: 306-314.

Gao S, Huang Y F, Zhang S, et al. 2020. Short-term runoff prediction with GRU and LSTM networks without requiring time step optimization during sample generation. Journal of Hydrology, 589: 125188. DOI: 10.1016/j.jhydrol.2020.125188.


Greff K, Srivastava R K, Koutník J, et al. 2017. LSTM: A search space odyssey. IEEE Transactions on Neural Networks and Learning Systems, 28(10): 2222-2232.


Jia C Y, Kang K X, Gao W, et al. 2020. Fault prediction of electro-hydraulic servo valve based on CNN+LSTM neural network. Chinese Hydraulics Pneumatics, 44(12): 173-181. (in Chinese)

Kan G Y, Yao C, Li Q L, et al. 2015. Improving event-based rainfall-runoff simulation using an ensemble artificial neural network based hybrid data-driven model. Stochastic Environmental Research and Risk Assessment, 29(5): 1345-1370.


Kim Y. 2014. Convolutional neural networks for sentence classification. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg, USA: Association for Computational Linguistics.

Kırbaş İ, Sözen A, Tuncer A D, et al. 2020. Comparative analysis and forecasting of COVID-19 cases in various European countries with ARIMA, NARNN and LSTM approaches. Chaos Solitons & Fractals, 138: 110015. DOI: 10.1016/j.chaos.2020.110015.


Li N, Bai R, Wu L, et al. 2020. Impacts of future climate change on spring phenology stages of rubber tree in Hainan, China. Chinese Journal of Applied Ecology, 31(4): 1241-1249. (in Chinese)


Lin G F, Chen G R, Huang P Y, et al. 2009. Support vector machine-based models for hourly reservoir inflow forecasting during typhoon-warning periods. Journal of Hydrology, 372(1-4): 17-29.


Liu L L, Zhang X Y, Yu Y Y, et al. 2017. Real-time and short-term predictions of spring phenology in North America from VIIRS data. Remote Sensing of Environment, 194: 89-99.


Mei Z X, Xu S J, Wang J Q. 2009. Spatio-temporal integrated forecast method of forest fire area based on DRNN and ARIMA model. Scientia Silvae Sinicae, 45(8): 101-107.

McKinney W. 2010. Data structures for statistical computing in Python. Proceedings of the 9th Python in Science Conference. Austin, USA: SciPy.

Mukerji A, Chatterjee C, Raghuwanshi N S. 2009. Flood forecasting using ANN, Neuro-Fuzzy, and Neuro-GA models. Journal of Hydrologic Engineering, 14(6): 647-652.


Nayak P C, Sudheer K P, Rangan D M, et al. 2005. Short-term flood forecasting with a Neurofuzzy model. Water Resources Research, 41(4): 91-119.

Pedregosa F, Varoquaux G, Gramfort A, et al. 2012. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12: 2825-2830.

Pei S X, Guo Q S, Xin X B, et al. 2011. Response of 4 common broad-leaved arbors phenology to climate change in the northeast China. Scientia Silvae Sinicae, 47(11): 181-187.

Salinas D, Flunkert V, Gasthaus J, et al. 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3): 1181-1191.


Saiful I M, Hossain E. 2020. Foreign exchange currency rate prediction using a GRU-LSTM hybrid network. Soft Computing Letters, 3: 100009. DOI: 10.1016/j.socl.2020.100009.


Sirsat M S, Mendes-Moreira J, Ferreira C, et al. 2019. Machine learning predictive model of grapevine yield based on agroclimatic patterns. Engineering in Agriculture, Environment and Food, 12(4): 443-450.

Sun G Z, Yu L J. 2006. Climate factors and establishment of prediction model of Korean pine phenology. Forestry Science and Technology, 31(5): 20-21, 61.

Sun M, Wei S K, Wang Y J, et al. 2020. Prediction model of water quality based on wavelet decomposition and LSTM. Computet Systems Applications, 29(12): 55-63.

Wingate L, Ogée J, Cremonese E, et al. 2015. Interpreting canopy development and physiology using a European phenology camera network at flux sites. Biogeosciences, 12(20): 5995-6015.


van der Walt S, Colbert S C, Varoquaux G. 2011. The NumPy array: A structure for efficient numerical computation. Computing in Science & Engineering, 13(2): 22-30.

Xiong H L, Ji H, Fan C J, et al. 2020. Forecast of travel index of air passengers based on LSTM-SVR model. Journal of Systems Management, 29(6): 1169-1176.

Yang J B, Ma X Q. 2005. On the basis of artificial neural network to forecast the forest fire in Guangdong Province. Scientia Silvae Sinicae, 41(4): 127-132.

Yun K, Hsiao J, Jung M P, et al. 2017. Can a multi-model ensemble improve phenology predictions for climate change studies? Ecological Modelling, 362: 54-64.


Zhang Y, Song K Y, Tong D. 2013. Prediction of mature wood anatomical properties of Pinus banksiana plantation based on Support Vector Machines (SVM). Scientia Silvae Sinicae, 49(9): 119-125.

Zhou L, He H L, Sun X M, et al. 2013. Species- and community-scale simulation of the phenology of a temperate forest in Changbai Mountain based on digital camera images. Journal of Resources and Ecology, 4(4): 317-326.