Next Article in Journal
Impacts of Plant Configuration on the Outdoor Wind Comfort of Subtropical Coastal Campuses: Evidence from a Study of Quanzhou
Previous Article in Journal
Nitrogen Enrichment Alters Plant Root, Soil Microbial Structure, Diversity, and Function in Mountain Forests of North China
Previous Article in Special Issue
Forest Disturbance and Restoration in China's North-South Transition Zone: A Case from the Funiu Mountains
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Forests
Volume 16
Issue 3
10.3390/f16030460
forests-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Prediction and Spatiotemporal Dynamics of Vegetation Index Based on Deep Learning and Environmental Factors in the Yangtze River Basin

by
Yin Wang
1,
Nan Zhang
1,
Mingjie Chen
1,
Yabing Zhao
1,
Famiao Guo
1,
Jingxian Huang
1,
Daoli Peng
1,* and
Xiaohui Wang
2,*
1
State Key Laboratory of Efficient Production of Forest Resources, Beijing Forestry University, Beijing 100083, China
2
Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, China
*
Authors to whom correspondence should be addressed.
Forests 2025, 16(3), 460; https://doi.org/10.3390/f16030460
Submission received: 29 November 2024 / Revised: 22 February 2025 / Accepted: 27 February 2025 / Published: 5 March 2025
(This article belongs to the Special Issue Mapping and Modeling Forests Using Geospatial Technologies)
Browse Figures
Versions Notes

Abstract

:
Accurately predicting the vegetation index (VI) of the Yangtze River Basin and analyzing its spatiotemporal trends are essential for assessing vegetation dynamics and providing recommendations for environmental resource management in the region. This study selected the key climate factors most strongly correlated with three vegetation indexes (VI): the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and kernel Normalized Difference Vegetation Index (kNDVI). Historical VI and climate data (2001–2020) were used to train, validate, and test a CNN-BiLSTM-AM deep learning model, which integrates the strengths of Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Attention Mechanism (AM). The performance of this model was compared with CNN-BiLSTM, LSTM, and BiLSTM-AM models to validate its superiority in predicting the VI. Finally, climate simulation data under three Shared Socioeconomic Pathway (SSP) scenarios (SSP1-1.9, SSP2-4.5, and SSP5-8.5) were used as inputs to the CNN-BiLSTM-AM model to predict the VI for the next 20 years (2021–2040), aiming to analyze spatiotemporal trends. The results showed the following: (1) Temperature, precipitation, and evapotranspiration had the highest correlation with VI data and were used as inputs to the time series VI model. (2) The CNN-BiLSTM-AM model combined with the EVI achieved the best performance (R2 = 0.981, RMSE = 0.022, MAE = 0.019). (3) Under all three scenarios, the EVI over the next 20 years showed an upward trend compared to the previous 20 years, with the most significant growth observed under SSP5-8.5. Vegetation in the source region and the western part of the upper reaches increased slowly, while significant increases were observed in the eastern part of the upper reaches, middle reaches, lower reaches, and estuary. The analysis of the predicted EVI time series indicates that the vegetation growth conditions in the Yangtze River Basin will continue to improve over the next 20 years.

1. Introduction

An essential component of the terrestrial ecosystem is vegetation. It is a sensitive indicator that may identify changes in the environment, and it also has a significant impact on the global carbon cycle and energy flux [1,2,3,4]. Climate change causes complex and diverse changes in vegetation dynamics, including variations in temperature, changed patterns of precipitation, and extreme weather. Understanding and monitoring these changes are essential for understanding ecological resilience and biodiversity and managing natural resources. To efficiently monitor vegetation changes, vegetation indexes (VIs) such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and the recently proposed kernel Normalized Difference Vegetation Index (kNDVI) are commonly used to represent vegetation growth dynamics and overall health. These indexes play a key role in the observation and formulation of global environmental change response plans [5,6,7]. They provide an intuitive reflection of vegetation growth, degradation, or recovery trends; capture the spatiotemporal variation characteristics of vegetation cover; and reveal the impacts of climatic factors and human activities on the ecological environment. Additionally, the time series changes in VIs offer scientific evidence for evaluating the effectiveness of ecological restoration projects, formulating ecological protection policies, and optimizing resource management.
At the same time, many studies have made significant progress in VI time series prediction, which has attracted the interest of various disciplines [8]. From traditional statistical methods, such as Least Absolute Shrinkage and Selection Operator (LASSO) and Autoregressive Moving Average (ARMA), to machine learning techniques like Support Vector Machine Regression (SVR), Extreme Learning Machine (ELM), and Artificial Neural Networks (ANN), significant progress has been made in VI time series prediction [9,10,11,12,13]. For example, Huang et al. [14] combined ANNs and SVR to improve the prediction accuracy of the NDVI in the Yellow River Basin, while Wu et al. [15] applied time-delayed neural networks for NDVI prediction in arid regions. These models address the complex nonlinear relationships in data that traditional methods struggle with and have been recognized as more robust alternatives. Nevertheless, machine learning methods often fail to fully capture the intrinsic features of time series data [16]. In contrast, deep learning methods, particularly Long Short-Term Memory (LSTM) networks, offer significant improvements by capturing complex temporal dependencies in data [17,18,19]. For example, Cao et al. [20] demonstrated that LSTM outperforms LASSO and Random Forest (RF) in predicting rice yield in China. However, due to structural limitations, most LSTM-based models encounter challenges when processing multi-dimensional, grid-based data. The CNN-LSTM hybrid model, which combines Convolutional Neural Networks (CNN) for spatial feature extraction with LSTM for temporal learning, effectively addresses these limitations [21,22]. Bounoua et al. [23] found that this model captures both spatial and temporal dependencies more effectively, highlighting its advantages in crop water stress prediction. Additionally, incorporating the AM enhances the model’s performance by focusing on the most critical features, thereby reducing the computational load associated with multi-dimensional inputs.
The association between climate change and vegetation dynamics has been a major area of scientific investigation because of the increasingly severe effects of climate change [24,25,26]. Since the ecological restoration project of shelterbelt construction that started in 1990, due to the implementation of ecological projects, along with climate change and population migration, the ecological environment has undergone significant changes. Predicting vegetation dynamics to develop adaptive strategies has become particularly important to ensure ecological balance and sustainable development. Previous deep learning prediction models of the VI only considered the time-varying characteristics of the VI and often ignored the influence of climate factors, which led to poor prediction results for VI time series [27,28,29]. Therefore, studying the spatiotemporal correlation between VI changes and climate factors has become the focus of researchers [5,30,31,32]. As research continued, complex models that integrated climate factors to improve prediction accuracy were developed. For example, Iwasaki [33] used multiple regression models to predict NDVI changes in Mongolian grasslands by combining Global Satellite Mapping of Precipitation (GSMap precipitation) and Japanese Reanalysis 25-year/Japanese Meteorological Agency Climate Data Assimilation System (JRA-25/JCDAS) of temperature. However, despite the fact that an enormous quantity of research has examined how future climate change would affect vegetation, the majority of these studies rely on incremental scenarios that are likely to occur (e.g., equidistant increases in mean temperature and precipitation from baseline to the future) [34] and general circulation models (GCMs) [35] as the main input data. In addition, prediction uncertainty is increased due to the fact that the current research only employs one climate change scenario and has a rather coarse spatial and temporal resolution of data. To overcome these limitations, the Coupled Model Intercomparison Project Phase 6 (CMIP6) data set, which offers the most accurate spatiotemporal resolution for simulating future climate change, is now being utilized. Scenarios like SSP1-2.6, SSP2-4.5, and SSP5-8.5, which stand for low-, medium-, and high-emission scenarios, respectively, are included in this data set [36]. This data set has been extensively utilized to forecast future land use patterns and can represent the combined impact of socioeconomic and climatic elements [37].
This study aims to predict the VI in the Yangtze River Basin from 2021 to 2040 and analyze spatiotemporal dynamic variation trends using the CNN-BiLSTM-AM VI prediction model. The innovation of this study lies in the application of the CNN-BiLSTM-AM deep learning model in the field of VI prediction, which demonstrates high predictive accuracy. Based on different SSP scenarios, this study generates multiple VI prediction results, providing valuable references for responding to various climate change scenarios and supporting ecological protection and sustainable development planning in the context of climate change. Additionally, the prediction results can be used to assess the effectiveness and long-term impact of existing environmental protection policies, offering scientific evidence for policy adjustments and optimization. Land managers and ecological policymakers can adjust land use planning, agricultural crop planting structures, and forest management measures based on predicted VI changes; identify potential high-risk areas; and take proactive prevention and response measures.

2. Materials and Methods

2.1. Study Area

The Yangtze River Basin is a vast and geographically diverse core region of China, spanning from 90°33′ E to 122°19′ E and 24°28′ N to 35°54′ N. With a total length of approximately 6300 km and a drainage area of about 1.8 million square kilometers, the Yangtze River originates from the Tanggula Mountains on the Qinghai–Tibet Plateau, flows through 11 provinces and municipalities, and ultimately empties into the East China Sea (Figure 1). Extending from west to east, the basin traverses the Qinghai–Tibet Plateau and temperate, subtropical, and tropical climatic zones, showcasing diverse climatic conditions. The average annual temperature is 13.3 °C, with a spatial pattern of higher temperatures in the southeast and lower temperatures in the northwest. Rainfall is unevenly distributed, with an average annual precipitation of 1036 mm, decreasing from southeast to northwest. Topographically, the upper reaches are dominated by mountains and plateaus, the middle reaches by plains and hills, and the lower reaches by alluvial plains. The Yangtze River Basin is one of the most vegetation-rich regions in China, encompassing alpine vegetation, forests, grasslands, and other ecosystems. It plays a vital role in soil and water conservation and regulating river runoff and serves as a critical ecological barrier for maintaining the overall environmental balance of the basin.

2.2. Methodological Framework

In this study, by determining the Pearson correlation coefficient, we first eliminated the meteorological variables that were not strongly connected with the VI. Subsequently, the deep learning model of the VI time series prediction model combining multi-dimensional inputs was constructed by using historical VI and climate data (2001–2015). Next, the model was tested using the data from 2016 to 2020. Thirdly, the above model and the climate simulation data under three different scenarios were used to predict the VI of the Yangtze River Basin in the next 20 years (2021–2040). Finally, the future vegetation change trend in the Yangtze River Basin was analyzed. The workflow is shown in Figure 2.

2.3. Data Sets

2.3.1. NDVI, EVI, and kNDVI

The NDVI, EVI, and kNDVI are all key indicators of vegetation health and vigor. The NDVI is calculated using the reflectance difference between the near-infrared (NIR) and red light bands, which range from −1 to 1: negative values generally indicate water, 0 indicates bare land, and positive values indicate vegetation cover. Although the NDVI is concise and effective, it tends to saturate in densely vegetated areas and may be affected by ground background in sparsely vegetated areas [38]. In contrast, the EVI optimizes vegetation signals by reducing the interference of background and atmosphere, which is especially suitable for high-biomass areas. In addition, the kNDVI introduces a length-scale parameter ( σ ), which can be adjusted to capture the nonlinear sensitivity of the NDVI to vegetation density [39]. The kNDVI is more robust against noise and better correlated with ecosystem productivity [40]. The generalized formula for the kNDVI is as follows:
k N D V I = tanh ( ( NIR − Red 2 σ ) 2 )
With the generalization σ = 0.5 (NIR + Red), the formula for the kNDVI is as follows:
k N D V I = tanh ( N D V I 2 )
Google Earth Engine (GEE) is a cloud computing-based geographical information processing platform that integrates a large amount of remote sensing data and geospatial data to provide users with high-performance data processing and analysis capabilities. In this study, the NDVI and EVI datawere derived from the MOD13A1 data set of MODIS sensors loaded on Terra and Aqua satellites, which is provided by the GEE platform with a temporal resolution of 16 days and a spatial resolution of 1 km. We first imported the MOD13A1 data set into the GEE platform. After conducting preprocessing steps such as selecting bands, filtering regions, filtering periods, quality control, etc., we used the maximum value synthesis method to calculate the monthly maximum values of the NDVI and EVI from 2001 to 2020. Then, the kNDVI was calculated using Formula (2).

2.3.2. Historical Environmental and Future Climate Simulation Data

The National Tibetan Plateau Science Data Center (https://data.tpdc.ac.cn/home, accessed on 23 October 2023) provided all the necessary environmental and climate simulation data (Table 1). These future climate simulation data are generated through the Delta spatial downscaling scheme according to the global spatial resolution greater than 100 km climate model data set released by IPCC CMIP6 and the global high-resolution climate data set released by WorldClim. The data adopt the most recent SSP scenarios published by IPCC, and each scenario includes the climate data of one Global Climate Model (MRI-ESM2-0). Among them, the lowest current radiative forcing scenario, SSP1-1.9, takes into account a global warming level control of 1.5 °C, with radiative forcing reaching approximately 1.9 W/m2 in 2100. The medium-radiative-forcing scenario, SSP2-4.5, takes into account radiative forcing stabilizing at approximately 4.5 W/m2 in 2100, while the high-radiative-forcing scenario, SSP5-8.5, takes into account radiative forcing reaching approximately 8.5 W/m2 in 2100. In the next 20 years, the annual average temperature in the Yangtze River Basin is expected to rise across all scenarios (Table 2). Under the SSP5-8.5 scenario, both the annual average temperature and evapotranspiration are the highest, with the greatest variability, followed by SSP2-4.5, while SSP1-1.9 has the lowest values. In contrast, the annual average precipitation is projected to be the highest under SSP1-1.9, also with the greatest variability, followed by SSP5-8.5, and the lowest under SSP2-4.5 (Figure 3).
All the data were first transformed from NetCDF to GeoTIFF format and then reprojected from the sinusoidal to the GCS_WGS_1984 projection. Due to the significant differences in the spatial and temporal resolutions of the remote sensing data used in this study, achieving uniformity was challenging, and extracting data over large areas required considerable time. To address this, we divided the Yangtze River Basin into 10 × 10 km grids using ArcGIS 10.8, converted the grids into points, and extracted the corresponding data values for subsequent analysis.

2.4. Research Methods

2.4.1. Pearson Correlation Analysis to Select Environmental Factors Correlated with VIs

In deep learning modeling, selecting appropriate input variables can effectively identify and eliminate variables with low correlation with the research target, significantly improve the model’s ability to predict, and reduce unnecessary computational complexity. It has been confirmed that temperature, precipitation, evapotranspiration, surface windspeed, relative humidity, and surface soil moisture are related to vegetation changes [41,42,43,44]. We performed a Kolmogorov–Smirnov normality test on six environmental factors and VIs from 2001 to 2020, followed by Pearson correlation analysis. Strongly correlated factors with a correlation coefficient greater than 0.5 were selected to construct the prediction model [45].
The degree and direction of the trend between two VIs and environmental factors are reflected in the correlation coefficient. Its value falls between −1 and 1, with 0 reflecting no correlation and more correlation indicated by bigger absolute values between the two variables.

2.4.2. VI Predictive Modeling Based on Deep Learning Method

The CNN-BiLSTM-AM model (Figure 4) is specifically tailored for time series data analysis within the scope of feature engineering. It combines the characteristics of CNN, BiLSTM, and AM to analyze time series data. CNN effectively reduce the number of parameters and extract key features through convolutional layers and pooling layers, while BiLSTM not only solves the problem of vanishing gradient and explosion but also captures the bidirectional dependencies in the sequence by combining the output of the forward and backward LSTM networks and processes long sequence data through memory blocks and gating mechanisms. AM focus on important information in the input and improve model performance by calculating and normalizing weights. The proposed model integrates the feature recognition ability of CNN, the sequence processing efficiency of BiLSTM, and the information screening ability of AM. After performing Pearson correlation analysis to select environmental factors, collinearity tests were conducted to ensure the stability of the model, improve prediction performance, and reduce the risk of overfitting. The model was trained using the NDVI, EVI, kNDVI, and the selected environmental factors from 2001 to 2015. A total of 180 monthly data sets, derived from these 15 years, were split into 70% training and 30% validation sets. Parameters such as loss function, optimizer, and training period were adjusted through multiple training sessions. To assess the impact, three other models of LSTM, CNN-BiLSTM, and BiLSTM-AM were trained under the same data conditions, and the Coefficient of determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) were also selected as accuracy evaluation indicators. Finally, using the data from 2016 to 2020, these models were tested to verify their extrapolation ability, and two curves of the real and predicted values were drawn to examine whether the time series model could accurately capture monthly fluctuations. All models were developed using Python 3.8, Keras 2.9.0, and TensorFlow 2.9.1. The trained model and its preprocessing parameters were retained for future forecasts of the VIs from 2021 to 2040.

2.4.3. VI Time Series Prediction in the Yangtze River Basin from 2021 to 2040

Based on a pre-trained VI time series prediction model, climate simulation data from 2021 to 2040 were used as inputs to forecast VI changes under three scenarios for the next 20 years. Considering the significant geographical span of the Yangtze River Basin and its pronounced variations in climate, environment, and human activities, treating the entire basin as a single prediction unit lacks regional representativeness and fails to provide effective guidance for ecological and environmental policymaking. To address this issue, the basin was divided into 67 sub-regions based on China’s Level-3 river classification system (Figure 5), and predictions were conducted for each region individually. The predicted VI values from 2021 to 2040 under three scenarios were averaged every five years and visualized in vector maps. Moreover, the spatial changes in the VI in the Yangtze River Basin were exhibited based on the rate of VI change. The calculation method is shown in Formula (3).
C R = V ¯ I ¯ n − V ¯ I ¯ p V ¯ I ¯ p × 100 %
where C R represents the rate of VI change, V ¯ I ¯ n represents the mean value of the VI in the next five years, and V ¯ I ¯ p represents the mean value of the VI in the past five years.

3. Results

3.1. Impact of Environmental Factors on VIs

The Kolmogorov–Smirnov normality test showed that the p-values for the significance tests of VIs and environmental factors were both greater than 0.05, indicating that the data follow a normal distribution (Figure 6). According to Table 3, temperature, precipitation, and evapotranspiration showed strong correlations with VIs, with a mean correlation coefficient of 0.662. On the contrary, the remaining three factors had low correlations with VIs from −0.444 to 0.497. And the Variance Inflation Factor (VIF) values between temperature and precipitation, temperature and evapotranspiration, and precipitation and evapotranspiration were 1.843, 1.843, and 1.863, respectively, all below 10, indicating no multicollinearity. Therefore, these three factors were selected for model construction. This reduces the error brought on by erroneous or duplicate input data, which enhances the model’s processing effectiveness and forecast accuracy.

3.2. Optimal Model Configuration and Performance Comparison

After several parameter adjustments, we arrived at an optimal model configuration that achieved good training efficiency and accuracy. The Mae loss function and Adam optimization algorithm were selected for their excellent performance in various training scenarios. The model was initialized with a learning rate of 0.01 and a timestep of 5 and was trained for 200 epochs. To fine-tune the weights in the later stages of training to avoid overfitting, we set the learning rate tuning strategy to a decay factor of 0.1, which, in this study, meant that the learning rate would decrease by 90% after 180 epochs. The experimental results indicate that all four models exhibit strong predictive performance. As shown in Table 4, the CNN-BiLSTM-AM model demonstrates extremely high accuracy in predicting the EVI, with the lowest RMSE (0.023), the lowest MAE (0.015), and the highest R2 (0.951). Additionally, the CNN-BiLSTM-AM model achieves outstanding performance in testing data from 2016 to 2020, particularly for EVI predictions, with accuracy metrics of R2 = 0.981, RMSE = 0.022, and MAE = 0.019. Moreover, the EVI prediction curve based on the CNN-BiLSTM-AM model closely aligns with the trends in the actual curve, effectively capturing abrupt changes at peaks and turning points (Figure 7). These data show that the model combining a CNN and BiLSTM with the AM can more effectively capture and analyze the spatiotemporal characteristics of the data when dealing with VI prediction containing complex temporal dependencies, thus providing more accurate prediction results. Therefore, we finally selected the CNN-BiLSTM-AM model and the EVI to predict the VI in the Yangtze River Basin from 2021 to 2040.

3.3. Scenario Analysis of EVI Change

From 2001 to 2020, the average value of the EVI was 0.269, showing an increasing trend (Figure 8), growing at an average yearly rate of 0.82%. The predicted data show that the EVI will continue to increase from 2021 to 2040 under the three SSP scenarios, and the annual average value under SSP5-8.5 is the highest (0.308), and there is a small difference between SSP1-1.9 (0.291) and SSP2-4.5 (0.290). Compared with the average EVI value from 2001 to 2020, it increases by 8.18%, 7.81%, and 14.50%, respectively. Under the SSP5-8.5 scenario, the EVI is significantly improved, reaching the highest value of 0.324 among all scenarios, and the EVI in 2040 increases by 26.01% compared with that in 2021. Due to the highest growth rate of temperature and evapotranspiration, vegetation will grow and decline periodically, which also leads to the largest fluctuation in the EVI (standard deviation is 0.030). Under the SSP2-4.5 scenario, the EVI has a small fluctuation (standard deviation is 0.024), with a mean value of 0.290, a minimum value of 0.277, and a maximum value of 0.301, and the EVI in 2040 increases by 9.25% compared with that in 2021. Under the SSP1-1.9 scenario, climate changes may be consistent with vegetation growth characteristics in the Yangtze River Basin, and vegetation also shows a relatively stable growth trend (standard deviation is 0.018, minimum value is 0.270, and maximum value is 0.304). The EVI in 2040 was 20.38% higher than that in 2021.
Similarly, from 2021 to 2040, the climate change trends in the Yangtze River Basin under the three SSP scenarios will have a significant impact on EVI dynamics. Compared to the historical conditions from 2001 to 2020, the annual average EVI, temperature, precipitation, and evapotranspiration all increased across all scenarios. In the SSP1-1.9 mild scenario, the changes in precipitation, temperature, and evapotranspiration are relatively small, leading to stable fluctuations in the EVI, with an average value of 0.291. In the SSP2-4.5 intermediate scenario, the magnitude of climate variability increases slightly compared to SSP1-1.9 but not significantly, and the trend in the EVI remains similar to that under SSP1-1.9. However, under the SSP5-8.5 extreme scenario, the impact of climate change on vegetation becomes particularly pronounced. The annual average EVI reaches a maximum of 0.308, and the fluctuation range is also aggravated. High temperatures, increased evapotranspiration, and dramatic changes in precipitation place greater environmental stress on vegetation, resulting in greater instability and more significant fluctuations in the EVI.

3.4. Characteristics of EVI Spatial Distribution Variation

The prediction results indicate that over the next 20 years, under all three scenarios, the EVI in the Yangtze River Basin exhibits a distinct longitudinal zonal distribution, gradually increasing from west to east (Figure 9). Additionally, the EVI in 67 sub-regions shows increasing or decreasing fluctuations over time (Figure 10).
Under the SSP1-1.9 scenario, the EVI in 2021 is generally low, with most areas having EVI values around 0.270. The average change rate of the four time periods is 3.25%. The lowest value in 20 years (0.063) also occurs in 2021 in the hinterland of the Qinghai–Tibet Plateau at the source of the Yangtze River (Cheerchen River). From 2021 to 2025, the main trend is an increase in the upper reaches and middle reaches, with the growth area accounting for 86.12% of the basin’s total area. Among the sub-regions, a sub-region in the southwestern-most part of the upper reaches shows a significant increase, accounting for 0.49% of the total basin area. From 2026 to 2030, except for the source, 60 sub-regions will continue to see slight increases in the EVI, accounting for 89.65% of the total area. Throughout this time, the EVI in the upper-middle part of the Baoshixia in the upper reaches of the Yangtze River arrives at the highest value in 20 years, at 0.462. From 2031 to 2035, the main trend is a slight decrease in the EVI across the upper reaches, middle reaches, lower reaches, and estuary areas, covering 86.43% of the total basin area. From 2036 to 2040, there is another shift to slight increases accounting for 89.66% of the total area.
Under the SSP2-4.5 scenario, the fluctuations in the EVI are more pronounced compared with the SSP1-1.9 scenario. The average change rate across the four time periods is 1.88%. From 2021 to 2025, the EVI shows a decreasing trend compared with the year 2021, mainly in the upper reaches and middle reaches regions, covering 59.37% of the total basin area. From 2026 to 2030, the EVI recovers to slight increases in all sub-regions except for the upper reaches, accounting for 88.57% of the total area. Around this time, the EVI in the upper reaches (Cheerchen River) arrives at the lowest value of 0.060 in 20 years, while the EVI in the lower reaches region (Wangjiaba) arrives at the highest value of 0.429 over 20 years. From 2031 to 2035, there is another large-scale slight decrease in the EVI across the basin, excluding the upper reaches and the southeastern part of the middle reaches and lower reaches regions, covering 77.66% of the total area. From 2036 to 2040, the EVI shows a recovery, with slight increases covering 83.58% of the area.
Under the SSP5-8.5 scenario, the EVI exhibits the most notable fluctuations, with the largest increases and the smallest decreases occurring under this scenario. The average change rate across the four time periods averages 7.07%. In 2021, the EVI in the lower reaches region (Wangjiaba) arrives at the highest value of 0.481 over 20 years. From 2021 to 2025, the EVI shows mixed trends, with a slight increase in an area covering 63.33% of the total area, excluding the upper reaches and northern parts of the middle reaches and lower reaches. From 2026 to 2030, significant increases, increases, and slight increases are observed in the upper reaches, middle reaches, and lower reaches regions, accounting for 0.49%, 11.07%, and 72.95% of the total area, respectively. Throughout this time, the EVI in the upper reaches (Cheerchen River) drops to the lowest value of 0.074 over 20 years. From 2031 to 2035, significant and slight decreases appear in the previously growing areas, accounting for 0.49% and 84.01% of the total area, respectively. From 2036 to 2040, the EVI shows a recovery, with decreases in the upper reaches and estuary regions covering 13.88% of the total area and increases in the upper reaches, middle reaches, and lower reaches regions covering 86.12% of the total area.
In summary, the EVI in the entire Yangtze River Basin will gradually rise to a peak during the period from 2021 to 2030, exhibit a declining trend from 2031 to 2035, and then begin to recover from 2036 to 2040. Nevertheless, the overall trend remains as an increase compared to the period from 2001 to 2020.

4. Discussion

4.1. Directions for Improving the CNN-BiLSTM-AM Model

The CNN-BiLSTM-AM model has a significant effect on predicting the VI, but this model usually requires a large number of iterations to converge, especially when dealing with multi-dimensional spatiotemporal data, which require a long training time. Moreover, the model has multiple hyperparameters to be adjusted, including the number of filters in the CNN layer, the number of units in the BiLSTM layer, and the configuration of the AM layer. The effectiveness of the model is greatly affected by these various hyperparameter setups, and finding the best configuration often requires a lot of experiments and experience, which further increases the time required for our experiment. In addition, all existing deep learning models are usually regarded as “black box” models whose internal mechanisms are difficult to explain. For VI prediction, understanding the decision-making process of the model is particularly important for ecological environmental management and policymaking. In future research, more efficient network architectures, such as lightweight CNNand BiLSTM variants, can also be explored to reduce the computational resource requirements and shorten the training time. In addition, introducing visualization technology and Explainable Artificial Intelligence (XAI) methods is expected to further improve the transparency and interpretability of the model, providing more intuitive and understandable support for research.

4.2. Effects of Different Climate Scenarios on Vegetation Dynamics

The trend in the EVI under different SSP scenarios reveals important information about vegetation health and vegetation cover in the Yangtze River Basin. Under the three high-, medium-, and low-emission scenarios, the EVI increases compared with it in the past 20 years, which could be caused by global warming and ecological restoration policies. The observed extreme changes under the SSP1-1.9 and SSP2-4.5 scenarios are smaller, with standard deviations of 0.018 and 0.024 for the two scenarios, respectively, which are identical to previous findings and suggest that vegetation health may remain relatively stable under the low-emission scenario [46,47]. The SSP5-8.5 high-emission scenario has a standard deviation of 0.030, showing more pronounced EVI oscillations. This finding indicates that the higher the emission scenario, the greater the variability in and extreme values of each climate factor, and vegetation health is stressed by more extreme climate conditions, leading to more pronounced EVI fluctuations. The process of growing vegetation is intricate, and the relative stability under the low-emission scenario and the significant fluctuations in the high-emission scenario reflect the various effects of climate change on vegetation health. Therefore, among all scenarios, SSP1-1.9 is particularly significant for sustainable ecological management and achieving long-term vegetation stability. The SSP5-8.5 scenario provides critical insights into potential risks under severe climate change, which are essential for developing targeted mitigation and adaptation strategies. Depending on the evaluation objectives, each scenario offers unique insights, whether aimed at maintaining current vegetation conditions, preparing for moderate changes, or mitigating risks under worst-case scenarios. This multi-scenario approach facilitates comprehensive planning to address both current and future challenges in vegetation management in the Yangtze River Basin.

4.3. Impacts of Human Activities on Vegetation Dynamics and Limitations of This Study

Apart from how the dynamics of vegetation are affected by climate change, human actions also have an important effect. In the next 20 years, the vegetation in the Yangtze River Basin will show some spatial heterogeneity [29], such as the continuous upward trend in the EVI in the central and eastern parts of the region, partly due to the high level of human activities in the region and the initiatives of effective regional conservation measures and ecological restoration programs [47,48,49]. Around the 21st century, China launched large-scale ecological projects such as the Yangtze River Shelter Forest, the conversion of farmland to forest, and the protection of natural forests. These programs have significantly contributed to the increase in vegetation cover, improving the ecological environment, curbing soil erosion, regulating regional climate, and promoting biodiversity recovery. Due to the relatively backward economic development, complex terrain, and other reasons, human activities have insufficient vegetation management in the Qinghai–Tibet Plateau area at the source of the Yangtze River, and therefore, EVI growth in the western part of the basin is slow. These spatial differences highlight the importance of vegetation management and environmental conservation approaches in specific regions.
This study achieved certain results but still has some limitations. Firstly, this study only considered three climate factors (temperature, precipitation, and evapotranspiration) affecting the VI, while other factors, such as sunshine duration, different soil types, topography, and human activities, have significant impacts on vegetation changes but were not included as variables in the model. Secondly, differences in ecosystem responses to climate change may result in spatial distribution variations. Incorporating these factors into the model in future studies may further improve the prediction accuracy of the VI [50,51,52], which was not discussed in this study. Thirdly, some research has demonstrated that the interaction between vegetation and climate parameters is not uniform in terms of time delay [53,54]. This study did not consider the time delay effects of temperature, precipitation, and evapotranspiration on vegetation dynamics, which may have influenced the results. Fourth, vegetation dynamics may exhibit nonlinear responses under extreme climate conditions. For instance, abrupt changes may occur when temperature or precipitation reaches critical thresholds. However, such critical points may not have been captured in historical data, and thus, models trained on historical data have certain limitations in predicting the future.
In addition, the variations in the VI are analyzed based on the rate of change. The limitation of this method is that the trend is assumed to be linear, which cannot capture nonlinear dynamics, and it is difficult for it to accurately reflect complex fluctuations. Future research should consider the use of more advanced trend analysis methods to effectively deal with nonlinear and complex trend changes. For instance, the Mann–Kendall test does not rely on the data distribution assumption and can effectively detect monotonic trends in time series data. Therefore, when dealing with vegetation changes affected by extreme climate, this method can more accurately identify trend alterations [55]. The RESTREND method, which reconstructs and evaluates trend stability, effectively removes noise and seasonal fluctuations in the data, thereby providing a more accurate depiction of long-term vegetation trends, especially in the context of climate change [56]. Additionally, logarithmic regression can effectively simulate vegetation dynamics that change rapidly in the early stages and gradually tend to be stable in later stages [57].

5. Conclusions

In this study, the CNN-BiLSTM-AM deep learning model was employed to predict the VI in the Yangtze River Basin from 2021 to 2040. This model combines CNNs’ ability to extract spatial features, the time series data processing ability of BiLSTM, and the key information highlighting function of AMs. It effectively handles the nonlinear changes in VI data and achieves good accuracy, with an R2 of 0.981, RMSE of 0.022, and MAE of 0.019.
The results indicate that by using this model and three different SSP scenarios to predict the VI in the Yangtze River Basin, we accurately captured and revealed possible changes in vegetation health and dynamics under different climate and socioeconomic development scenarios over the next 20 years. The eastern part of the upper reaches, the middle reaches, the lower reaches, and the estuary have a warm, humid climate and multiple human activities like urban greening, water resource management, and ecological restoration that promote vegetation growth, and therefore, the EVI will significantly increase in these areas. In contrast, the vegetation in the source and western part of the upper reaches of the Yangtze River Basin is expected to grow slowly due to more complex terrain and harsher climate conditions. These scenarios provide important scientific support for the formulation of regional ecological protection and resource management strategies to adapt to potential future climate changes. Particularly under the high-emission scenario (SSP5-8.5), vegetation health may be under greater stress due to more extreme climate conditions, requiring more positive responses from policymakers and managers.
Although this study achieved some positive results, the data sources and influencing factors used are relatively limited. Future research should expand data sources and integrate more types of remote sensing data, ground observation data, and environmental factors to improve the model’s accuracy and robustness. Since vegetation conditions are also affected by land use changes and human activities, these anthropogenic factors should be included in prediction models. In order to provide continuous data support for demonstrating the model’s adaptability and long-term effects, we also advise implementing ongoing vegetation monitoring projects. This will give a strong scientific foundation for ecological preservation and agricultural planning in the Yangtze River Basin, as well as critical support for the creation of international strategies to combat climate change and safeguard the environment.

Author Contributions

Conceptualization, Y.W. and D.P.; methodology, Y.W.; software, Y.W.; validation, Y.W. and J.H.; formal analysis, Y.W.; investigation, Y.W. and Y.Z.; resources, D.P.; data curation, Y.W. and F.G.; writing—original draft preparation, Y.W.; writing—review and editing, X.W.; visualization, M.C. and N.Z.; supervision, D.P.; project administration, D.P.; funding acquisition, X.W. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Key R&D Program of China (Grant No. 2022YFF0711602) and National Key R&D Program of China (Grant No. 2023YFD2200403).

Data Availability Statement

The data presented in this study are available on request from the corresponding authors.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Jiang, L.; Bao, A.; Guo, H.; Ndayisaba, F. Vegetation dynamics and responses to climate change and human activities in Central Asia. Sci. Total Environ. 2017, 599, 967–980. [Google Scholar] [CrossRef] [PubMed]
  2. Peng, J.; Li, Y.; Tian, L.; Liu, Y.; Wang, Y. Vegetation dynamics and associated driving forces in Eastern China during 1999–2008. Remote Sens. 2015, 7, 13641–13663. [Google Scholar] [CrossRef]
  3. Liu, H.; Jiao, F.; Yin, J.; Li, T.; Gong, H.; Wang, Z.; Lin, Z. Nonlinear relationship of vegetation greening with nature and human factors and its forecast—A case study of Southwest China. Ecol. Indic. 2020, 111, 106009. [Google Scholar] [CrossRef]
  4. Piao, S.; Wang, X.; Ciais, P.; Zhu, B.; Wang, T.; Liu, J. Changes in satellite-derived vegetation growth trend in temperate and boreal Eurasia from 1982 to 2006. Glob. Change Biol. 2011, 17, 3228–3239. [Google Scholar] [CrossRef]
  5. Chu, H.; Venevsky, S.; Wu, C.; Wang, M. NDVI-based vegetation dynamics and its response to climate changes at Amur-Heilongjiang River Basin from 1982 to 2015. Sci. Total Environ. 2019, 650, 2051–2062. [Google Scholar] [CrossRef]
  6. Xu, H.-J.; Wang, X.-P.; Yang, T.-B. Trend shifts in satellite-derived vegetation growth in Central Eurasia, 1982–2013. Sci. Total Environ. 2017, 579, 1658–1674. [Google Scholar] [CrossRef] [PubMed]
  7. Rojo-Álvarez, J.L.; Martínez-Ramón, M.; Munoz-Mari, J.; Camps-Valls, G. Digital Signal Processing with Kernel Methods; John Wiley & Sons: Hoboken, NJ, USA, 2018. [Google Scholar]
  8. Ferchichi, A.; Abbes, A.B.; Barra, V.; Farah, I.R. Forecasting vegetation indices from spatio-temporal remotely sensed data using deep learning-based approaches: A systematic literature review. Ecol. Inform. 2022, 68, 101552. [Google Scholar] [CrossRef]
  9. Aghelpour, P.; Mohammadi, B.; Biazar, S.M. Long-term monthly average temperature forecasting in some climate types of Iran, using the models SARIMA, SVR, and SVR-FA. Theor. Appl. Climatol. 2019, 138, 1471–1480. [Google Scholar] [CrossRef]
  10. Hu, H.; Zhang, J.; Li, T. A novel hybrid decompose-ensemble strategy with a VMD-BPNN approach for daily streamflow estimating. Water Resour. Manag. 2021, 35, 5119–5138. [Google Scholar] [CrossRef]
  11. Kuila, S.; Dhanda, N.; Joardar, S. ECG signal classification and arrhythmia detection using ELM-RNN. Multimed. Tools Appl. 2022, 81, 25233–25249. [Google Scholar] [CrossRef]
  12. Panwar, V.S.; Pandey, A.; Hasan, M.E. Generalised regression neural network (GRNN) architecture-based motion planning and control of an e-puck robot in V-Rep software platform. Acta Mech. Et Autom. 2021, 15, 209–214. [Google Scholar] [CrossRef]
  13. Zarei, A.R.; Mahmoudi, M.R. Ability assessment of the stationary and cyclostationary time series models to predict drought indices. Water Resour. Manag. 2020, 34, 5009–5029. [Google Scholar] [CrossRef]
  14. Huang, S.; Ming, B.; Huang, Q.; Leng, G.; Hou, B. A case study on a combination NDVI forecasting model based on the entropy weight method. Water Resour. Manag. 2017, 31, 3667–3681. [Google Scholar] [CrossRef]
  15. Wu, T.; Fu, H.; Feng, F.; Bai, H. A new approach to predict normalized difference vegetation index using time-delay neural network in the arid and semi-arid grassland. Int. J. Remote Sens. 2019, 40, 9050–9063. [Google Scholar] [CrossRef]
  16. Wang, S.; Huang, G.; Baetz, B.; Huang, W. A polynomial chaos ensemble hydrologic prediction system for efficient parameter inference and robust uncertainty assessment. J. Hydrol. 2015, 530, 716–733. [Google Scholar] [CrossRef]
  17. Di, G.; Qian, L.; Yong-xing, D.; Bao-shan, L. Research on NDVl Prediction Based on BiLSTM-Attention Model. Comput. Simul. 2023, 40, 278–283+439. [Google Scholar]
  18. Cao, J.; Zhang, Z.; Luo, Y.; Zhang, L.; Zhang, J.; Li, Z.; Tao, F. Wheat yield predictions at a county and field scale with deep learning, machine learning, and google earth engine. Eur. J. Agron. 2021, 123, 126204. [Google Scholar] [CrossRef]
  19. Dang, C.; Liu, Y.; Yue, H.; Qian, J.; Zhu, R. Autumn crop yield prediction using data-driven approaches: Support vector machines, random forest, and deep neural network methods. Can. J. Remote Sens. 2021, 47, 162–181. [Google Scholar] [CrossRef]
  20. Cao, J.; Zhang, Z.; Tao, F.; Zhang, L.; Luo, Y.; Zhang, J.; Han, J.; Xie, J. Integrating multi-source data for rice yield prediction across China using machine learning and deep learning approaches. Agric. For. Meteorol. 2021, 297, 108275. [Google Scholar] [CrossRef]
  21. Kim, N.; Na, S.-I.; Park, C.-W.; Huh, M.; Oh, J.; Ha, K.-J.; Cho, J.; Lee, Y.-W. An artificial intelligence approach to prediction of corn yields under extreme weather conditions using satellite and meteorological data. Appl. Sci. 2020, 10, 3785. [Google Scholar] [CrossRef]
  22. Hu, F.; Yang, Q.; Yang, J.; Luo, Z.; Shao, J.; Wang, G. Incorporating multiple grid-based data in CNN-LSTM hybrid model for daily runoff prediction in the source region of the Yellow River Basin. J. Hydrol. Reg. Stud. 2024, 51, 101652. [Google Scholar] [CrossRef]
  23. Bounoua, I.; Saidi, Y.; Yaagoubi, R.; Bouziani, M. Deep Learning Approaches for Water Stress Forecasting in Arboriculture Using Time Series of Remote Sensing Images: Comparative Study between ConvLSTM and CNN-LSTM Models. Technologies 2024, 12, 77. [Google Scholar] [CrossRef]
  24. Fu, B.; Li, S.; Yu, X.; Yang, P.; Yu, G.; Feng, R.; Zhuang, X. Chinese ecosystem research network: Progress and perspectives. Ecol. Complex. 2010, 7, 225–233. [Google Scholar] [CrossRef]
  25. Liang, S.; Yi, Q.; Liu, J. Vegetation dynamics and responses to recent climate change in Xinjiang using leaf area index as an indicator. Ecol. Indic. 2015, 58, 64–76. [Google Scholar]
  26. Ren, S.; Li, Y.; Peichl, M. Diverse effects of climate at different times on grassland phenology in mid-latitude of the Northern Hemisphere. Ecol. Indic. 2020, 113, 106260. [Google Scholar] [CrossRef]
  27. Gong, X.; Du, S.; Li, F.; Ding, Y. Study of mesoscale NDVI prediction models in arid and semiarid regions of China under changing environments. Ecol. Indic. 2021, 131, 108198. [Google Scholar] [CrossRef]
  28. Lehner, F.; Hawkins, E.; Sutton, R.; Pendergrass, A.G.; Moore, F.C. New potential to reduce uncertainty in regional climate projections by combining physical and socio-economic constraints. AGU Adv. 2023, 4, e2023AV000887. [Google Scholar] [CrossRef]
  29. Reddy, D.S.; Prasad, P.R.C. Prediction of vegetation dynamics using NDVI time series data and LSTM. Model. Earth Syst. Environ. 2018, 4, 409–419. [Google Scholar] [CrossRef]
  30. Li, M.; Hou, X.; Ying, L.; Lu, X.; Zhu, M. Spatial-temporal dynamics of NDVI and its response to temperature and precipitation in the Yellow River Delta during the period 1998–2008. Resour. Sci. 2011, 33, 322–327. [Google Scholar]
  31. Liu, X.; Zhang, J.; Zhu, X.; Pan, Y.; Liu, Y.; Zhang, D.; Lin, Z. Spatiotemporal changes in vegetation coverage and its driving factors in the Three-River Headwaters Region during 2000–2011. J. Geogr. Sci. 2014, 24, 288–302. [Google Scholar] [CrossRef]
  32. Piao, S.; Yin, G.; Tan, J.; Cheng, L.; Huang, M.; Li, Y.; Liu, R.; Mao, J.; Myneni, R.B.; Peng, S. Detection and attribution of vegetation greening trend in China over the last 30 years. Glob. Change Biol. 2015, 21, 1601–1609. [Google Scholar] [CrossRef] [PubMed]
  33. Iwasaki, H. NDVI prediction over Mongolian grassland using GSMaP precipitation data and JRA-25/JCDAS temperature data. J. Arid Environ. 2009, 73, 557–562. [Google Scholar] [CrossRef]
  34. Nunes, J.P.; Seixas, J.; Pacheco, N.R. Vulnerability of water resources, vegetation productivity and soil erosion to climate change in Mediterranean watersheds. Hydrol. Process. 2007, 22, 3115–3134. [Google Scholar] [CrossRef]
  35. Sitch, S.; Huntingford, C.; Gedney, N.; Levy, P.E.; Lomas, M.; Piao, S.L.; Betts, R.; Ciais, P.; Cox, P.; Friedlingstein, P.; et al. Evaluation of the terrestrial carbon cycle, future plant geography and climate-carbon cycle feedbacks using five Dynamic Global Vegetation Models (DGVMs). Glob. Change Biol. 2008, 14, 2015–2039. [Google Scholar] [CrossRef]
  36. Popp, A.; Calvin, K.; Fujimori, S.; Havlik, P.; Humpenöder, F.; Stehfest, E.; Bodirsky, B.L.; Dietrich, J.P.; Doelmann, J.C.; Gusti, M. Land-use futures in the shared socio-economic pathways. Glob. Environ. Change 2017, 42, 331–345. [Google Scholar] [CrossRef]
  37. Sanderson, B.M.; Booth, B.B.; Dunne, J.; Eyring, V.; Fisher, R.A.; Friedlingstein, P.; Gidden, M.; Hajima, T.; Jones, C.D.; Jones, C.; et al. The need for carbon emissions-driven climate projections in CMIP7. Geosci. Model Dev. 2024, 17, 8141–8172. [Google Scholar] [CrossRef]
  38. Pettorelli, N.; Vik, J.O.; Mysterud, A.; Gaillard, J.-M.; Tucker, C.J.; Stenseth, N.C. Using the satellite-derived NDVI to assess ecological responses to environmental change. Trends Ecol. Evol. 2005, 20, 503–510. [Google Scholar] [CrossRef]
  39. Feng, X.; Tian, J.; Wang, Y.; Wu, J.; Liu, J.; Ya, Q.; Li, Z. Spatio-temporal variation and climatic driving factors of vegetation coverage in the Yellow River Basin from 2001 to 2020 based on kNDVI. Forests 2023, 14, 620. [Google Scholar] [CrossRef]
  40. Camps-Valls, G.; Campos-Taberner, M.; Moreno-Martínez, Á.; Walther, S.; Duveiller, G.; Cescatti, A.; Mahecha, M.D.; Muñoz-Marí, J.; García-Haro, F.J.; Guanter, L. A unified vegetation index for quantifying the terrestrial biosphere. Sci. Adv. 2021, 7, eabc7447. [Google Scholar] [CrossRef]
  41. Zhong, L.; Ma, Y.; Salama, M.S.; Su, Z. Assessment of vegetation dynamics and their response to variations in precipitation and temperature in the Tibetan Plateau. Clim. Change 2010, 103, 519–535. [Google Scholar] [CrossRef]
  42. Wang, S.; Fu, B.; Gao, G.; Yao, X.; Zhou, J. Soil moisture and evapotranspiration of different land cover types in the Loess Plateau, China. Hydrol. Earth Syst. Sci. 2012, 16, 2883–2892. [Google Scholar] [CrossRef]
  43. Tchebakova, N.M.; Parfenova, E.; Soja, A. The effects of climate, permafrost and fire on vegetation change in Siberia in a changing climate. Environ. Res. Lett. 2009, 4, 045013. [Google Scholar] [CrossRef]
  44. Munson, S.M.; Belnap, J.; Okin, G.S. Responses of wind erosion to climate-induced vegetation changes on the Colorado Plateau. Proc. Natl. Acad. Sci. USA 2011, 108, 3854–3859. [Google Scholar] [CrossRef]
  45. Cohen, J. Statistical Power Analysis for the Behavioral Sciences; Routledge: New York, NY, USA, 1988. [Google Scholar]
  46. Kong, R.; Zhang, Z.; Huang, R.; Tian, J.; Feng, R.; Chen, X. Projected global warming-induced terrestrial ecosystem carbon across China under SSP scenarios. Ecol. Indic. 2022, 139, 108963. [Google Scholar] [CrossRef]
  47. Miao, Q.; Huang, M.; Li, R.-Q. The impacts of climate change on vegetation net primary productivity of the Yangtze River Basin. J. Nat. Resour. 2010, 25, 1296–1305. [Google Scholar]
  48. Chen, Q.; Zhou, Q.; Zhang, H.; Liu, F. Spatial disparity of NDVI response in vegetation growing season to climate change in the Three-River Headwaters Region. Ecol. Environ. 2010, 19, 1284. [Google Scholar]
  49. Tong, X.; Wang, K.; Yue, Y.; Brandt, M.; Liu, B.; Zhang, C.; Liao, C.; Fensholt, R. Quantifying the effectiveness of ecological restoration projects on long-term vegetation dynamics in the karst regions of Southwest China. Int. J. Appl. Earth Obs. Geoinf. 2017, 54, 105–113. [Google Scholar] [CrossRef]
  50. Wu, X.; Liu, H.; Li, X.; Ciais, P.; Babst, F.; Guo, W.; Zhang, C.; Magliulo, V.; Pavelka, M.; Liu, S. Differentiating drought legacy effects on vegetation growth over the temperate Northern Hemisphere. Glob. Change Biol. 2018, 24, 504–516. [Google Scholar] [CrossRef]
  51. Xu, H.-J.; Wang, X.-P.; Zhao, C.-Y.; Yang, X.-M. Diverse responses of vegetation growth to meteorological drought across climate zones and land biomes in northern China from 1981 to 2014. Agric. For. Meteorol. 2018, 262, 1–13. [Google Scholar] [CrossRef]
  52. Zhang, Q.; Kong, D.; Singh, V.P.; Shi, P. Response of vegetation to different time-scales drought across China: Spatiotemporal patterns, causes and implications. Glob. Planet. Change 2017, 152, 1–11. [Google Scholar] [CrossRef]
  53. Wen, Y.; Liu, X.; Yang, J.; Lin, K.; Du, G. NDVI indicated inter-seasonal non-uniform time-lag responses of terrestrial vegetation growth to daily maximum and minimum temperature. Glob. Planet. Change 2019, 177, 27–38. [Google Scholar] [CrossRef]
  54. Zhao, J.; Huang, S.; Huang, Q.; Wang, H.; Leng, G.; Fang, W. Time-lagged response of vegetation dynamics to climatic and teleconnection factors. Catena 2020, 189, 104474. [Google Scholar] [CrossRef]
  55. Liu, J.; Liu, S.; Tang, X.; Ding, Z.; Ma, M.; Yu, P. The response of land surface temperature changes to the vegetation dynamics in the Yangtze River Basin. Remote Sens. 2022, 14, 5093. [Google Scholar] [CrossRef]
  56. Xu, L.; Yu, G.; Tu, Z.; Zhang, Y.; Tsendbazar, N.-E. Monitoring vegetation change and their potential drivers in Yangtze River Basin of China from 1982 to 2015. Environ. Monit. Assess. 2020, 192, 642. [Google Scholar] [CrossRef]
  57. Li, J.; Zhang, X.; Rousta, I.; Olafsson, H.; Krzyszczak, J. Evaluating vegetation dynamics in the Yangtze river basin in relation to climatological parameters using remote sensing data from 2001 to 2022. Int. Agrophys. 2024, 38, 407–422. [Google Scholar] [CrossRef]
Forests 16 00460 g001
Figure 1. Geographical location map of Yangtze River Basin, China (1: Source; 2: Upper reaches; 3: Middle reaches; 4: Lower reaches; 5: Estuary).
Figure 1. Geographical location map of Yangtze River Basin, China (1: Source; 2: Upper reaches; 3: Middle reaches; 4: Lower reaches; 5: Estuary).
Forests 16 00460 g001
Forests 16 00460 g002
Figure 2. The workflow (TMP: temperature; PRE: precipitation; ET: evapotranspiration; WD: windspeed; RH: relative humidity; SM: soil moisture).
Figure 2. The workflow (TMP: temperature; PRE: precipitation; ET: evapotranspiration; WD: windspeed; RH: relative humidity; SM: soil moisture).
Forests 16 00460 g002
Forests 16 00460 g003
Figure 3. Trends in three climatic factors from 2021 to 2040 ((A): Temperature; (B): Precipitation; (C): Evapotranspiration).
Figure 3. Trends in three climatic factors from 2021 to 2040 ((A): Temperature; (B): Precipitation; (C): Evapotranspiration).
Forests 16 00460 g003
Forests 16 00460 g004
Figure 4. CNN-BiLSTM-AM model structure diagram.
Figure 4. CNN-BiLSTM-AM model structure diagram.
Forests 16 00460 g004
Forests 16 00460 g005
Figure 5. The spatial distribution of third-level river basins in the Yangtze River Basin.
Figure 5. The spatial distribution of third-level river basins in the Yangtze River Basin.
Forests 16 00460 g005
Forests 16 00460 g006
Figure 6. Results of Kolmogorov–Smirnov normality test for VIs and environmental factors.
Figure 6. Results of Kolmogorov–Smirnov normality test for VIs and environmental factors.
Forests 16 00460 g006
Forests 16 00460 g007
Figure 7. EVI prediction effects of CNN-BiLSTM-AM model.
Figure 7. EVI prediction effects of CNN-BiLSTM-AM model.
Forests 16 00460 g007
Forests 16 00460 g008
Figure 8. Temporal variation in EVI in Yangtze River Basin from 2001 to 2040.
Figure 8. Temporal variation in EVI in Yangtze River Basin from 2001 to 2040.
Forests 16 00460 g008
Forests 16 00460 g009
Figure 9. Spatial trends in the EVI from 2021 to 2040 under three SSP scenarios in the Yangtze River Basin. (A–E) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, and 2036–2040, respectively, under the SSPl-1.9 scenario; (F) represents the locations and periods with the maximum and minimum EVIs under the SSPl-1.9 scenario; (G–K) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, 2031–2035 and 2036–2040, respectively, under the SSP2-4.5 scenario; (L) represents the locations and periods with the maximum and minimum EVIs under the SSP2-4.5 scenario; (M–Q) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, and 2036–2040, respectively, under the SSP5-8.5 scenario; (R) represents the locations and periods with the maximum and minimum EVIs under the SSP5-8.5 scenario.
Figure 9. Spatial trends in the EVI from 2021 to 2040 under three SSP scenarios in the Yangtze River Basin. (A–E) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, and 2036–2040, respectively, under the SSPl-1.9 scenario; (F) represents the locations and periods with the maximum and minimum EVIs under the SSPl-1.9 scenario; (G–K) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, 2031–2035 and 2036–2040, respectively, under the SSP2-4.5 scenario; (L) represents the locations and periods with the maximum and minimum EVIs under the SSP2-4.5 scenario; (M–Q) represent EVI distribution in the years of 2021, 2021–2025, 2026–2030, and 2036–2040, respectively, under the SSP5-8.5 scenario; (R) represents the locations and periods with the maximum and minimum EVIs under the SSP5-8.5 scenario.
Forests 16 00460 g009
Forests 16 00460 g010
Figure 10. Change rates in 67 sub-regions at five-year intervals relative to the previous five-year period ((A–D) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP1-1.9 scenario; (E–H) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP2-4.5 scenario; (I–L) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP5-8.5 scenario).
Figure 10. Change rates in 67 sub-regions at five-year intervals relative to the previous five-year period ((A–D) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP1-1.9 scenario; (E–H) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP2-4.5 scenario; (I–L) represent the trends in the EVI for the periods of 2021–2025, 2026–2030, 2031–2035, and 2036–2040, respectively, under the SSP5-8.5 scenario).
Forests 16 00460 g010
Table 1. Historical environmental and future climate simulation data.
Table 1. Historical environmental and future climate simulation data.
TimeNameSpatial Resolution (km)Model Name
Historical environmental data
2001–2020		TMP (°C)1
Pre (mm)1
Et (mm)1
SM (g/m3)1
WD (m/s)25
RH (%RH)25
Future climate simulation data
2021–2040		Tmp (°C)1MRI-ESM2-0
Pre (mm)1MRI-ESM2-0
Et (mm)1MRI-ESM2-0
Table 2. The average values of climate factors under historical conditions from 2001 to 2020 and three climate scenarios from 2021 to 2040.
Table 2. The average values of climate factors under historical conditions from 2001 to 2020 and three climate scenarios from 2021 to 2040.
ScenarioClimate Factors
Pre (mm)Tmp (°C)Et (mm)
Historical669.17.2747.3
SSP1-1.9735.18.4775.1
SSP2-4.5688.38.5775.7
SSP5-8.5729.08.6776.3
Table 3. Correlation coefficients between VIs and climate factors.
Table 3. Correlation coefficients between VIs and climate factors.
VariablesNDVIEVIkNDVI
Tmp (°C)0.7180.7520.682
Pre (mm)0.5500.6230.599
Et (mm)0.6430.7490.640
WD (m/s)−0.121−0.384−0.444
RH (%RH)0.4970.4920.492
SM (kg/m3)0.4740.4470.437
Table 4. Accuracy comparison of various modeling methods.
Table 4. Accuracy comparison of various modeling methods.
NameMethodValidation DataTest Data
RMSER2MAETraining Time (s)RMSER2MAE
EVILSTM0.1080.8760.06835.50.1060.9400.076
CNN-BiLSTM0.1020.8960.05742.00.0980.9450.068
BiLSTM-AM0.0840.9060.04637.10.0630.9580.025
CNN-BiLSTM-AM0.0230.9510.01538.00.0220.9810.019
NDVILSTM0.1600.6920.08121.90.1310.9150.105
CNN-BiLSTM0.1040.8710.07544.80.1040.9270.075
BiLSTM-AM0.1020.8770.07533.40.1090.9380.078
CNN-BiLSTM-AM0.0340.9130.02947.60.0370.9600.032
kNDVILSTM0.1240.8360.07220.90.1070.9340.084
CNN-BiLSTM0.1030.8870.07634.80.1030.9410.074
BiLSTM-AM0.0970.9010.06628.60.0970.9170.066
CNN-BiLSTM-AM0.0250.9410.02134.00.0410.9450.036
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Wang, Y.; Zhang, N.; Chen, M.; Zhao, Y.; Guo, F.; Huang, J.; Peng, D.; Wang, X. Prediction and Spatiotemporal Dynamics of Vegetation Index Based on Deep Learning and Environmental Factors in the Yangtze River Basin. Forests 2025, 16, 460. https://doi.org/10.3390/f16030460

AMA Style

Wang Y, Zhang N, Chen M, Zhao Y, Guo F, Huang J, Peng D, Wang X. Prediction and Spatiotemporal Dynamics of Vegetation Index Based on Deep Learning and Environmental Factors in the Yangtze River Basin. Forests. 2025; 16(3):460. https://doi.org/10.3390/f16030460

Chicago/Turabian Style

Wang, Yin, Nan Zhang, Mingjie Chen, Yabing Zhao, Famiao Guo, Jingxian Huang, Daoli Peng, and Xiaohui Wang. 2025. "Prediction and Spatiotemporal Dynamics of Vegetation Index Based on Deep Learning and Environmental Factors in the Yangtze River Basin" Forests 16, no. 3: 460. https://doi.org/10.3390/f16030460

APA Style

Wang, Y., Zhang, N., Chen, M., Zhao, Y., Guo, F., Huang, J., Peng, D., & Wang, X. (2025). Prediction and Spatiotemporal Dynamics of Vegetation Index Based on Deep Learning and Environmental Factors in the Yangtze River Basin. Forests, 16(3), 460. https://doi.org/10.3390/f16030460

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop