Study on the Spatiotemporal Variation of Vegetation Characteristics in the Three River Source Region Based on the CatBoost Model

Highlights

What are the main findings?

• The CatBoost model outperformed RF and LightGBM in NDVI data fusion, reconstructing a 1 km, long-term (1982–2014) GIMMS-MODIS NDVI dataset for the Three River Source Region.
• Vegetation in the Three River Source Region exhibited a significant greening trend from 1982 to 2014 (0.0020/10a), with clear spatial heterogeneity—strongest in the Yellow River Source Region and weakest in the Lancang River Source Region.

What are the implications of the main findings?

• The fused NDVI dataset improved the accuracy of long-term vegetation monitoring in the Three River Source Region.
• The results offer insights for ecological restoration and sustainable management in alpine ecosystems.

Abstract

Under the ongoing trend of climate warming and increasing humidity on the Qinghai–Tibet Plateau, the Three River Source Region (TRSR) has shown strong sensitivity to global climate change. Its vegetation change is particularly worthy of attention and research. The Normalized Difference Vegetation Index (NDVI) is a key indicator for assessing the growth status of vegetation. However, the insufficiency of existing NDVI datasets in terms of spatiotemporal continuity has limited the accuracy of long-term vegetation change studies. This study proposed a machine learning-based downscaling framework that integrates the Moderate-resolution Imaging Spectroradiometer (MODIS) NDVI and the Global Inventory Monitoring and Modeling System (GIMMS) NDVI data to reconstruct a long-term, high-resolution NDVI dataset. Unlike conventional statistical fusion approaches, the proposed framework employs machine learning-based nonlinear relationships to generate long-term, high-resolution NDVI data. Three machine learning algorithms—Random Forest (RF), LightGBM, and CatBoost—were evaluated. Their performance was validated using the MODIS NDVI as reference, with the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and Pearson’s correlation coefficient (R) as evaluation metrics. Based on model comparison, the CatBoost model was identified as the optimal algorithm for spatiotemporal data fusion (R² = 0.9014, RMSE = 0.0674, MAE = 0.0445), significantly outperforming RF and LightGBM models and demonstrating stronger capability for NDVI spatiotemporal reconstruction. Using this model, a long-term, 1 km monthly GIMMS-MODIS NDVI dataset from 1982 to 2014 was successfully reconstructed. On the basis of this dataset, the spatiotemporal variation characteristics of vegetation in the TRSR from 1982 to 2014 were systematically analyzed. The research results show that: (1) The constructed long-series high-resolution NDVI dataset has a high consistency with MODIS NDVI data; (2) From 1982 to 2014, the NDVI in the TRSR showed an increasing trend, with an average growth rate of 0.0020/10a (p < 0.05). NDVI showed obvious spatial heterogeneity, characterized by a decreasing gradient from southeast to northwest. (3) The Yellow River source exhibited the most evident vegetation recovery, the Yangtze River Source area showed a moderate improvement, whereas the Lancang River Source area displayed little noticeable change. (4) Broad-leaved forests experienced the most significant growth, while cultivated vegetation displayed a marked tendency toward degradation. This study provides both a high-accuracy long-term NDVI product for the TRSR and a methodological foundation for advancing vegetation dynamics research in other high-altitude regions.

1. Introduction

The Three River Source Region (TRSR), located at the headwaters of the Yellow River, Yangtze River, and Lancang River, has experienced a pronounced warming and humidifying trend in recent decades as a result of global climate change. As one of the most climate-sensitive regions in the world, its rate of temperature increase is approximately 1.2 times higher than the average for the Tibetan Plateau during the same period, while precipitation has also shown an overall upward trend, accounting for about 71% of the total increase observed across the Plateau [1]. Ongoing global climate change has further intensified its impact on local ecosystems. In the TRSR, the ecosystem is highly fragile and particularly susceptible to human and climate factors [2,3]. Vegetation serves as an essential indicator of ecological and environmental change. At the end of the 20th century, the region experienced severe vegetation degradation, manifested in grassland deterioration, soil erosion, and land desertification [4]. However, since the early 2000s, with the establishment of national nature reserves and the implementation of ecological restoration policies such as “grazing prohibition and grassland recovery” [5,6], vegetation conditions have significantly improved, and the ecological environment has gradually recovered. Thus, exploring the long-term changes in vegetation patterns within the TRSR holds important scientific and practical implications for revealing regional ecological evolution and formulating effective strategies for ecological protection and restoration [7].

Satellite remote sensing, with its broad spatial coverage and high spatiotemporal resolution, has become an essential tool for monitoring vegetation dynamics. Various satellite sensors, such as the Moderate-resolution Imaging Spectroradiometer (MODIS), SPOT, and Landsat, provide multiple vegetation indices—including Fractional Vegetation Cover (FVC), Gross Primary Productivity (GPP), and NDVI—that are widely used to analyze vegetation changes across the Tibetan Plateau. Among these indices, the Normalized Difference Vegetation Index (NDVI) is one of the most widely used indicators, it can effectively capture vegetation growth conditions and the temporal–spatial variations in vegetation cover. Consequently, NDVI has been extensively applied in studies of vegetation activity and ecosystem dynamics [8]. Currently, the major long-term NDVI datasets include the Global Inventory Monitoring and Modeling System (GIMMS) NDVI and MODIS NDVI products. The GIMMS NDVI dataset covers the period from 1982 to 2015, offering observations every 15 days at a spatial resolution of approximately 8 km, whereas the MODIS NDVI dataset offers higher resolution data (250 m and 1 km) since 2000. In comparison, although the Landsat series provides long-term, high-resolution observations dating back to 1972, its application in long-term vegetation change monitoring is relatively limited due to its long revisit cycle and susceptibility to cloud contamination [9].

A growing number of researchers have investigated vegetation dynamics in the TRSR using various remote sensing datasets. However, most of these studies have relied on a single satellite product. For instance, Shen et al. analyzed vegetation changes from 2000 to 2015 using MODIS data combined with the DBEST method, revealing that most areas experienced ecological restoration, while a few regions showed varying degrees of degradation [10]. Liu et al. used MODIS NDVI and climatic data to examine the spatiotemporal variations and driving factors of vegetation cover between 2000 and 2011, finding that areas of vegetation improvement exceeded those of degradation, with a decreasing NDVI trend observed only in the Lancang River source [11]. In Bei’s study, GIMMS NDVI data (1982–2015), together with meteorological and evapotranspiration data, were used to analyze spatiotemporal variations and vegetation–climate interactions. The results showed that meadows in the northern and western parts of the TRSR exhibited significant growth, with temperature and precipitation identified as the primary influencing factors [12]. However, single-source remote sensing products cannot simultaneously provide both long-term temporal continuity and high spatial resolution, which limits the ability to conduct detailed, high-precision analyses of vegetation dynamics at the regional scale and to accurately capture their spatiotemporal patterns.

To solve the limitations of single remote sensing products, data fusion techniques combine the strengths of different sensors to achieve time-series reconstruction and improve spatial resolution [13]. In recent years, many scholars have applied various fusion methods to remote sensing data for vegetation monitoring. For example, Liu et al. used a linear regression approach to interpolate MODIS NDVI with GIMMS NDVI data, thereby obtaining high-resolution NDVI products [14]. With recent advances in machine learning, higher accuracy in spatiotemporal modeling of remote sensing data has been achieved. Li et al. employed the InENVI model to fuse MODIS and Landsat data, producing a 30 m-resolution long-term NDVI dataset for China covering 2001–2020 [15]. Zhang et al. developed a multi-scale convolutional neural network to fuse AVHRR and Landsat imagery, generating MODIS-like data and addressing the limitations of temporal coverage [16]. Similarly, Yang et al. applied a Random Forest (RF) model to fuse GIMMS and MODIS data, constructing a 250 m spatial resolution monthly NDVI dataset for the Qinghai–Tibet Plateau from 1982 to 2020 [13].

However, most existing fusion studies still have methodological limitations. Many studies construct simple univariate linear relationships between GIMMS and MODIS NDVI, which oversimplify temporal dynamics and fail to capture vegetation variations and interannual changes [14]. Others use weighted linear mixing approaches that slightly enhance fusion accuracy but remain highly sensitive to landcover heterogeneity and uncertainty, often resulting in substantial errors and loss of spatial detail [17]. Although Convolutional Neural Networks can exploit spatial neighborhood information for feature extraction, their performance tends to deteriorate in fragmented and heterogeneous landscapes [16].

Although previous studies have explored NDVI data fusion, high spatiotemporal resolution fusion research for the TRSR remains limited. Current studies on vegetation in this region generally suffer from short temporal coverage and low spatial resolution, which restrict the ability to simultaneously achieve long-term continuity and high spatial detail, and often fail to fully account for growth differences among vegetation types. To address these limitations, this study developed a machine learning-based downscaling framework that integrates multi-source remote sensing data. Specifically, MODIS NDVI data (high spatial resolution) were used to downscale and calibrate the coarse-resolution GIMMS NDVI dataset. The machine learning models were trained using coincident MODIS and GIMMS data, incorporating topographic factors derived from the Digital Elevation Model (DEM) to account for spatial heterogeneity in vegetation distribution. Through this process, a continuous 1 km NDVI dataset for the TRSR covering the period 1982–2014 was constructed. Based on this dataset, spatiotemporal variations of vegetation were systematically analyzed using correlation analysis and the Mann–Kendall (MK) trend test. Furthermore, the study examined the spatiotemporal patterns and differences of vegetation changes across different vegetation types.

2. Materials and Methods

2.1. Study Area

Three Rivers Source Region (TRSR) is located in the central part of the Qinghai–Tibet Plateau (Figure 1), covering about 15% of the plateau’s total area and extending from 89.24°E to 103.30°E and from 31°N to 37.12°N [18]. The terrain of this area is complex, mainly plateau and mountainous, with an average altitude of more than 4000 m. It is characterized by low temperature and strong radiation, and little precipitation [12].

The vegetation types are diverse in the TRSR, including alpine meadow, alpine grassland, shrubland, swamp, and coniferous forests. Among them, alpine grassland and alpine meadow are dominant. They are not only an essential part of the national ecological security barrier but also serve as a vital basis for the livelihoods of local herders [19]. In recent years, affected by rising temperatures, precipitation changes, and permafrost degradation, the regional ecological environment has changed to a certain extent. Therefore, it is meaningful to study its long-term vegetation changes.

2.2. Data

2.2.1. NDVI

The GIMMS NDVI3g product selected in this study is from Global GIMMS NDVI3g v1 dataset (1981–2015)-National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/en/data/9775f2b4-7370-4e5e-a537-3482c9a83d88 (accessed on 1 April 2025). The time span of this product is from 1982 to 2014. The time resolution is once every half month and the spatial resolution is 1/12° (approximately 8 km). Due to its advantage of a long time series, it is widely implemented in vegetation growth and phenological changes. However, its spatial resolution is coarse, making it difficult to capture details and differences.

The MOD13A2 data was obtained from the Google Earth Engine (GEE) platform (https://earthengine.google.com (accessed on 4 April 2025)) and provides 1 km spatial resolution with a 16-day temporal interval. However, the time is relatively short and it performs poorly in the study of vegetation changes over a long period of time.

For both NDVI data, the monthly values were derived through the maximum value composite (MVC) technique.

Previous studies have demonstrated that the dynamic trends and spatial patterns of GIMMS NDVI and MODIS NDVI over the Qinghai–Tibet Plateau are generally consistent [20]. As shown in Figure 2, we compared the monthly NDVI data from the overlapping period of the two datasets and found that their variations on both monthly and annual scales are largely consistent, with a high correlation coefficient of 0.97. However, there are slight differences in the timing of the peak values.

Due to the lack of available drone-based or ground truth data, and considering the widespread application and high reliability of MODIS products, the MODIS NDVI was adopted as the reference dataset for the validation of the downscaled NDVI results.

2.2.2. DEM

The digital elevation model (DEM) data used in this study has a spatial resolution of 0.008° × 0.008° and is from the National Qinghai–Tibet Plateau Scientific Data Center (https://www.tpdc.ac.cn/zh-hans/data/ddf4108a-d940-47ad-b25c-03666275c83a (accessed on 5 April 2025)).

2.2.3. Vegetation Type Data

In this study, vegetation type data were integrated with NDVI data to enable a more detailed examination of long-term vegetation dynamics. The spatial distribution map of vegetation is shown in Figure 1.

Vegetation type data for the TRSR were obtained from the1:1,000,000 vegetation type spatial distribution data that were provided by the Resource and Environmental Science Data Platform (https://www.resdc.cn (accessed on 5 April 2025)). The dataset has a spatial resolution of 1 km and provides a detailed refinement of the distribution patterns, horizontal zonation, and vertical stratification characteristics of 11 vegetation types [21]. The vegetation types in the TRSR are highly diverse, comprising a total of nine categories. The TRSR includes nine types of vegetation. The area proportions of each vegetation type in the TRSR are as follows: meadow (64.38%), grassland (14.10%), alpine vegetation (8.56%), shrubland (7.66%), others (2.24%), coniferous forests (2.15%), swamp (0.84%), cultivated vegetation (0.04%), and broad-leaved forests (0.03%).

We selected the period 1982–2014 for analysis in this study. The GIMMS NDVI and DEM data were resampled and interpolated using the bilinear interpolation method, while the vegetation type data were interpolated using the nearest-neighbor method. All datasets were resampled to a spatial resolution of 1 km × 1 km, ensuring consistency with the spatial resolution of the MODIS NDVI data.

2.3. Methods

2.3.1. Machine Learning Algorithms

Considering that the relationship between GIMMS and MODIS NDVI is not a simple linear one, this study selected three different regression models in the machine learning model—Random Forest (RF), CatBoost, and LightGBM. Among them, RF is based on the bagging algorithm, while CatBoost and LightGBM are based on the boosting algorithm. These models are capable of effectively capturing nonlinear relationships, complex feature interactions, and handling high-dimensional data, and mitigating the effects of missing values. By combining one bagging model (RF) with two boosting models (CatBoost and LightGBM), the effects of the three downscaling models are compared to determine the optimal model. This design enables a fair evaluation of ensemble methods while ensuring stable and generalizable results. The overall plan and technology roadmap are illustrated in Figure 3.

Random Forest is a decision-tree-based model built upon the principle of ensemble learning. It constructs multiple decision trees using random subsets of samples and features and averages the predictions from all trees to produce the final output [22]. This approach enhances model stability and predictive accuracy. Random forest effectively captures nonlinear relationships among variables and performs well with high-dimensional data, though its computational efficiency may decrease when handling very large datasets.

CatBoost addresses gradient bias and prediction shift through ordered boosting and categorical feature encoding techniques [23], thereby reducing the risk of overfitting. It demonstrates fast training speed, strong generalization ability, and robustness in handling both numerical and categorical variables.

LightGBM is a gradient boosting–based algorithm that constructs decision trees efficiently [24]. It integrates regularization methods such as L1, L2, and pruning, together with early stopping, to reduce overfitting and enhance the model’s generalization performance [25]. By reducing the number of data samples and feature dimensions, LightGBM achieves rapid and efficient computation [26].

To ensure optimal model performance, parameter selection and model optimization were conducted for all three machine learning algorithms. We selected high-resolution MODIS NDVI as the target variable. The input features included GIMMS NDVI, latitude, longitude, DEM, and monthly time indicators. The dataset was segmented from 2001 to 2014, with monthly data from 2001 to 2011 used for training, and data from 2012 to 2014 reserved for validation. The performance of the model was evaluated by using the five-fold cross-validation. For the RF model, 200 decision trees were used with a maximum depth of 6 and a minimum of 3 samples per leaf. For the LightGBM model, the learning rate was set to 0.05, the maximum depth to 7, and the number of leaves to 31, with both feature and bagging fractions set to 0.8 and a bagging frequency of 5. The CatBoost model was trained with 7500 iterations, a learning rate of 0.05, and a tree depth of 7. RMSE was used as the evaluation metric, and other parameters were kept at their default values to maintain model reproducibility.

2.3.2. Evaluation Metrics

To evaluate the accuracy of the downscaled NDVI data against the MODIS NDVI data, three statistical metrics were employed: the coefficient of determination (R²), root mean square error (RMSE) and Mean Absolute Error (MAE). All calculations were conducted at the pixel level.

R² quantifies model fit (0–1), where larger values denote stronger explanatory power and improved model accuracy. The calculation of R² is as follows:

$$R^2=1-{\displaystyle \frac{\sum _{i=i}^n{\left(y_i-\widehat{y_i}\right)}^2}{\sum _{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(1)

RMSE represents the square root of the average squared differences between the predicted and observed values. It measures the overall magnitude of prediction errors and is more sensitive to large deviations. The calculation of RMSE is as follows:

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

(2)

MAE measures the average absolute difference between the predicted and observed values, which intuitively reflects the model’s overall predictive accuracy. The MAE is calculated as follows:

$$MAE={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|y_i-\widehat{y_i}\right|$$

(3)

where $n$ is the number of samples; $y_i$ is the true observed value of NDVI of the $i\text{-}\mathrm{th}$ sample; $\widehat{y_i}$ is the predicted value of the $i\text{-}\mathrm{th}$ sample; and $\overline{y}$ is expected value of $y$.

2.3.3. Correlation Analysis

The Pearson correlation coefficient (r) assesses both the strength and direction of a linear relationship between two variables. In this study, Pearson’s correlation analysis was conducted to evaluate the relationships between GIMMS NDVI and MODIS NDVI, as well as between downscaled NDVI and MODIS NDVI. The Pearson correlation coefficient (r) is calculated as follows:

$$r={\displaystyle \frac{\sum _{i=1}^n\left(x_i-\bar{X}\right)\left(y_i-\bar{Y}\right)}{\sqrt{\sum _{i=1}^n{\left(x_i-\bar{X}\right)}^2\sum _{i=1}^n{\left(y_i-\bar{Y}\right)}^2}}}$$

(4)

where $x_i$ and $y_i$ represent the values of variables $x$ and $y$ in the i-th sample, $\bar{X}$ and $\bar{Y}$ denote the mean values of variables $x$ and $y$, $n$ is the sample size.

2.3.4. Trend Analysis

The Theil–Sen median trend analysis and Mann–Kendall (MK) test are commonly used for long-term trend analysis. In this study, the former was applied to investigate the spatiotemporal variation trends of NDVI in the TRSR at different time scales, while the latter was employed to assess the significance of these trends.

The Theil–Sen median trend analysis is a robust nonparametric statistical method and performs better than traditional regression approaches [18,27]. The calculation formula is as follows:

$$slope=Median\left({\displaystyle \frac{x_i-x_j}{i-j}}\right) for 1\le j\le i\le n$$

(5)

The MK trend test is a non-parametric statistical method used to detect monotonic trends in time series data, with the advantage of being independent of data distribution assumptions [28]. It is particularly suitable for analyzing environmental and climatic datasets [18,29]. In remote sensing studies, the MK trend test is widely applied to analyze long-term variations in vegetation indices such as NDVI and EVI, helping researchers to identify the rate and significance of vegetation cover changes and providing a scientific basis for ecological environment monitoring. Its calculation formulas are as follows:

$$S=\sum _{i=1}^{n-1}\sum _{j=i+1}^n\mathrm{s}\mathrm{g}\mathrm{n}\left(x_{i+1}-x_i\right)=\sum _{i<j}\mathrm{s}\mathrm{g}\mathrm{n}\left(x_j-x_i\right)$$

(6)

$$Var\left(S\right)={\displaystyle \frac{n\left(n-1\right)\left(2n+5\right)-\hspace{0.25em}\sum _{t=1}^k\left(n_t\left(n_t-1\right)\left(2n_t+5\right)\right)}{18}}$$

(7)

$$Z=\left\{\begin{array}{ll}{\displaystyle \frac{S-1}{\sqrt{\mathrm{Var}\left(S\right)}}},& \mathrm{if} S > 0\\ 0,& \mathrm{if} S=0\\ {\displaystyle \frac{S+1}{\sqrt{\mathrm{Var}\left(S\right)}}},& \mathrm{if} S < 0\end{array}\right.$$

(8)

where $S$ is the Mann–Kendall (MK) test statistic; $Var$ is the variance function; $Z$ is the significance statistic; n denotes the total number of observations in the time series; $k$ is the number of tied groups, and $n_t$ represents the number of observations in the $n_t$ tied group. A positive Z value represents an upward trend, whereas a negative Z value denotes a downward trend. The significance level (α) is set to 0.05. When the absolute value of $Z$ ($\left|Z\right|$) is greater than or equal to 1.645, 1.96, and 2.576, the trend is considered statistically significant at the 90%, 95%, and 99% confidence levels, respectively.

In this study, five significance levels were defined: significant increase, slight increase, no change, slight decrease, and significant decrease—to analyze the spatiotemporal variations of vegetation. The specific classification criteria and calculation methods are presented in

Table 1. Criteria for determining trend significance.

Slope	Z	Trend Classifications	Trend Characteristics
β > 0	$\left|Z\right|$ > 1.96	2	Significant increasing trend
	$\left|Z\right|$ ≤ 1.96	1	Slight increasing trend
β = 0		0	No change
β < 0	$\left|Z\right|$ ≤ 1.96	−1	Slight decreasing trend
	$\left|Z\right|$ > 1.96	−2	Significant decreasing trend

.

3. Results

3.1. Model Evaluation

In this study, three regression algorithms—RF, LightGBM, and CatBoost—were employed to construct NDVI downscaling models. As shown in

Table 2. Evaluation indicators of three downscaling models.

Model	Train_RMSE	Test_RMSE	Train_MAE	Test_MAE	Train_R2	Test_R2
RF	0.0772	0.0808	0.0509	0.0539	0.8661	0.8585
LightGBM	0.0705	0.0737	0.0467	0.0494	0.8886	0.8821
CatBoost	0.0639	0.0674	0.0418	0.0445	0.9083	0.9014

, all three downscaling models effectively improved the spatial resolution of GIMMS NDVI. The validation results indicate that the coefficient of determination (R²) values for all models exceed 0.85, while the RMSE values are mostly below 0.08 and the MAE values are below 0.06. Among them, the CatBoost model outperforms both RF and LightGBM in all evaluation metrics for both the training and validation sets, with no apparent overfitting or underfitting. These results show the robustness and suitability of CatBoost for constructing NDVI downscaling models. Therefore, all subsequent regional assessments and analyses of the downscaled NDVI characteristics were based solely on the results obtained from the CatBoost model.

We further analyzed the spatial distribution of R², RMSE, and MAE for the CatBoost downscaling model. The validation results in Figure 4 show that 76% of the TRSR achieved R² values above 0.7, where vegetation is predominantly meadow and grassland. Areas with RMSE values in the range [0, 0.075) accounted for 73.27% of the region, while areas with MAE between 0 and 0.05 accounted for 64.25%. These results demonstrate that the downscaling algorithm performs well across most parts of the TRSR.

To evaluate the accuracy of the downscaled NDVI across different vegetation types while accounting for both spatial and temporal variations, 60,000 pixels were randomly selected throughout the TRSR. Monthly downscaled NDVI values were extracted and compared with MODIS NDVI values, and correlation analyses were conducted for each vegetation type. As shown in Figure 5, the results indicate that the CatBoost downscaling model can capture the spatial variability of NDVI for most vegetation types, particularly meadow, grassland, broad-leaved forests, coniferous forests, swamp, and shrubland, with R² values exceeding 0.8. However, for the others vegetation type, the model’s performance was relatively poor, with an R² of only 0.56. These findings demonstrate the model’s high accuracy in representing both spatial heterogeneity and temporal dynamics.

3.2. Time Variation of Downscaled NDVI

We analyzed the NDVI time series to assess long-term vegetation changes in the TRSR. As shown in

Table 3. Annual NDVI statistics (max, min, and growth rate) for different vegetation types across regions from 1982 to 2014. Asterisks (*) denote statistical significance at the 95% confidence level.

Region	Vegetation Type	Max	Min	Growth Rate (/10a)
TRSR	Meadow	0.3080 (2010)	0.2778 (1983)	0.0037 *
	Grassland	0.1854 (2010)	0.1635 (1985)	0.0027 *
	Alpine vegetation	0.2045 (2010)	0.1816 (1983)	0.0011
	Shrubland	0.4106 (2010)	0.3870 (2008)	0.0015
	Coniferous forests	0.4380 (2010)	0.4207 (1983)	0.0017 *
	Swamp	0.4903 (2010)	0.4643 (1982)	0.0037 *
	Cultivated vegetation	0.3521 (2010)	0.3293 (1995)	0.0009
	Broad-leaved forests	0.3768 (2010)	0.3502 (1995)	0.0032 *
	Others	0.1570 (2010)	0.1412 (2014)	0.0003
Yellow River	Meadow	0.3582 (2010)	0.3230 (1983)	0.0033 *
	Grassland	0.2815 (2010)	0.2462 (1995)	0.0039 *
	Alpine vegetation	0.2317 (2010)	0.1934 (1983)	0.0027 *
	Shrubland	0.4243 (2010)	0.3953 (2008)	0.0027 *
	Coniferous forests	0.4187 (2010)	0.3950 (1984)	0.0024 *
	Swamp	0.4903 (2010)	0.4643 (1982)	0.0037 *
	Cultivated vegetation	0.3487 (2019)	0.3064 (1995)	0.0037 *
	Broad-leaved forests	0.3541 (2010)	0.3229 (1995)	0.0035 *
	Others	0.2000 (2002)	0.1708 (2014)	−0.0009
Yangtze River	Meadow	0.2505 (2010)	0.2230 (1983)	0.0016
	Grassland	0.1642 (2010)	0.1442 (1985)	0.0025 *
	Alpine vegetation	0.1900 (2010)	0.1689 (1983)	0.0012
	Shrubland	0.3590 (2010)	0.3232 (2008)	0.0006
	Coniferous forests	0.3847 (2010)	0.3525 (2008)	0.0009
	Cultivated vegetation	0.3542 (1998)	0.3176 (2008)	−0.0040 *
	Broad-leaved forests	0.4458 (2004)	0.4162 (1996)	0.0025
	Others	0.1389 (2011)	0.1234 (1995)	0.0013
Lancang River	Meadow	0.3188 (1988)	0.2948 (2008)	0.0003
	Alpine vegetation	0.2327 (1988)	0.2119 (1983)	−0.0002
	Shrubland	0.4081 (1988)	0.3903 (2008)	0.0003
	Coniferous forests	0.4560 (2004)	0.4366 (1983)	0.0016 *
	Cultivated vegetation	0.3757 (1990)	0.3556 (2008)	−0.0016
	Broad-leaved forests	0.3986 (2010)	0.3767 (1982)	0.0023 *
	Others	0.1966 (1994)	0.1781 (1983)	0.0012

, NDVI exhibited a fluctuating yet overall increasing trend during 1982–2014, with annual mean values ranging from 0.2655 to 0.2916. The highest mean NDVI occurred in 2010, while the lowest was in 1983, with an average growth rate of 0.0020/10a (p < 0.05). Among vegetation types, the annual mean NDVI ranked as follows: swamp > coniferous forests > shrubland > broad-leaved forests > cultivated vegetation > meadow > alpine vegetation > grassland > others. For all nine vegetation types, the maximum NDVI values were observed in 2010. In terms of temporal trends, NDVI for all vegetation types showed an increasing tendency, with growth rates from highest to lowest: swamp (0.0037/10a), broad-leaved forests (0.0032/10a), grassland (0.0027/10a), meadow (0.0021/10a), coniferous forests (0.0017/10a), shrubland (0.0015/10a), alpine vegetation (0.0011/10a), cultivated vegetation (0.0009/10a). Among these, swamp, coniferous forests, broad-leaved forests, meadow, and grassland exhibited statistically significant upward trends (p < 0.05). Overall, the regional mean NDVI of the TRSR has exhibited minimal fluctuations over the past 33 years. Consequently, a zonal analysis was undertaken to provide a more detailed assessment of vegetation change.

The annual average NDVI of the TRSR and its three subregions showed a moderate increase. The Yellow River Source had the fastest NDVI growth rate, at 0.0033/10a (p < 0.05). The Yangtze River Source showed a slower growth rate, at 0.0016/10a. Vegetation in the Lancang River Source remained largely unchanged, with a growth rate of only 0.0004/10a.

Across all vegetation types, the NDVI of all vegetation types in the Yellow River source region increased significantly. In the Yangtze River Source, grassland experienced a significant increase, whereas cultivated vegetation decreased, with change rates of increase of 0.0025/10a and −0.0040/10a, respectively. In the Lancang River Source Region, broad-leaved and coniferous forests exhibited the largest NDVI increases, at 0.0023/10a and 0.0016/10a, respectively. Overall, vegetation recovery progressed at the slowest rate.

However, due to the influence of regional differences in climate conditions, vegetation growth amplitude, and growth rate, different vegetation types exhibit significant seasonal variations in NDVI [20]. As shown in

Table 4. Growth rates of NDVI for different vegetation types across various regions and seasons from 1982 to 2014. Asterisks (*) denote statistical significance at the 95% confidence level.

Region	Vegetation Type	Growth Rate Spring (/10a)	Growth Rate Summer (/10a)	Growth Rate Autumn (/10a)	Growth Rate Winter (/10a)
TRSR	Meadow	0.0026 *	0.0023	0.0014	0.0030
	Grassland	0.0017 *	0.0040 *	0.0023 *	0.0028 *
	Alpine vegetation	0.0011	0.0019	0.0004	0.0012
	Shrubland	0.0035 *	−0.0006	0.0011	0.0027
	Coniferous forests	0.0036 *	−0.0003	0.0004	0.0031
	Swamp	0.0092 *	−0.0004	0.0032	0.0034
	Cultivated vegetation	0.0010	0.0039	0.0009	−0.0018
	Broad-leaved forests	0.0028	0.0042	0.0031	0.0027
	Others	0.0007	−0.0000	−0.0012	0.0013
Yellow River	Meadow	0.0043 *	0.0016	0.0033	0.0050 *
	Grassland	0.0022	0.0060 *	0.0050 *	0.0029 *
	Alpine vegetation	0.0030 *	0.0033	0.0022	0.0033
	Shrubland	0.0043 *	−0.0005	0.0034	0.0045
	Coniferous forests	0.0031	0.0016	0.0037 *	0.0019
	Swamp	0.0092 *	0.0092 *	0.0032	0.0034
	Cultivated vegetation	0.0030	0.0090 *	0.0041	−0.0007
	Broad-leaved forests	0.0024	0.0066 *	0.0045 *	0.0010
	Others	0.0003	0.0023	−0.0016	0.0002
Yangtze River	Meadow	0.0009	0.0041	0.0001	0.0015
	Grassland	0.0016 *	0.0037 *	0.0017	0.0028 *
	Alpine vegetation	0.0006	0.0026	−0.0002	0.0010
	Shrubland	0.0024	0.0020	−0.0008	−0.0013
	Coniferous forests	0.0029 *	0.0003	−0.0010	0.0017
	Cultivated vegetation	−0.0019	−0.0021	−0.0046 *	−0.0073 *
	Broad-leaved forests	0.0043 *	−0.0022	−0.0012	0.0077 *
	Others	0.0002	0.0005	0.0002	0.0014
Lancang River	Meadow	0.0018	−0.0003	−0.0007	0.0003
	Alpine vegetation	0.0012	−0.0006	−0.0006	0.0014
	Shrubland	0.0028 *	−0.0012	−0.0012	0.0000
	Coniferous forests	0.0039 *	−0.0009	−0.0003	0.0022
	Cultivated vegetation	−0.0012	−0.0023	−0.0019	0.0037
	Broad-leaved forests	0.0022 *	0.0010	0.0020	−0.0005
	Others	0.0043 *	0.0013	−0.0014	0.0014

, NDVI exhibited an overall increasing trend across all four seasons from 1982 to 2014, with the most significant increase in spring and the weakest change in autumn. This finding is consistent with the conclusion of Yang et al. [30] that spring exhibits the greatest rate and magnitude of NDVI increase, while autumn shows no significant variation. The overall NDVI variation trend for each vegetation type in spring was highly consistent with that of the annual mean NDVI. In the TRSR, the NDVI of four vegetation types—swamp, coniferous forests, grassland, and meadow—exhibited trends consistent with the annual mean, all showing significant increases during spring. Among them, swamp experienced the most significant increase (p < 0.05), with a growth rate of 0.0092/10a, while the other three vegetation types also showed noticeable increases, with growth rates exceeding 0.0020/10a. During summer, autumn, and winter, only grassland NDVI exhibited a significant upward trend, with growth rates greater than 0.0023/10a, whereas other vegetation types showed no evident changes.

From a regional perspective, vegetation growth was the most pronounced in the Yellow River source region across all four seasons. The NDVI of swamp, cultivated vegetation, broad-leaved forests, and grassland all showed evident increases, with average growth rates exceeding 0.0060/10a. In the Yangtze River source region, cultivated vegetation exhibited significant decrease during autumn and winter (−0.0046/10a or lower), while grassland and broad-leaved forests showed distinct increasing trends. In the Lancang River source region, NDVI increased significantly only in spring, with no evident changes for any vegetation types in summer, autumn, or winter.

Overall, from 1982 to 2014, the NDVI in the TRSR showed an upward trend across all time scales, with the most significant increase occurring in spring and the least noticeable change in autumn. At the regional scale, the Yellow River source area exhibited a generally higher rate of NDVI growth compared to the source areas of the Yangtze and the Lancang River.

3.3. Spatial Variation of Downscaled NDVI

We analyzed the seasonal variations and spatial distribution of vegetation. Figure 6 shows that NDVI values decrease gradually from the southeast toward the northwest in all four seasons, consistent with the distribution of precipitation. Vegetation growth exhibits distinct seasonal variations. NDVI values are higher in the Yellow River and Lancang River Source Regions, while they are generally lower in the Yangtze River source. The southeastern and northeastern sections of the Yellow River Source Region, along with the southeastern Lancang River area, exhibit relatively high NDVI values, while lower values occur in the western and northern portions of the Yangtze River Source Region and the northwestern Lancang River area.

As shown in Figure 7a, vegetation change characteristics in the TRSR show some spatial heterogeneity on an annual scale from 1982 to 2014, but the trend of vegetation degradation has reversed in most areas. Regarding the rate of change, the average NDVI increase was 0.0020/10a passing the 95% significance test. Regionally, vegetation recovered in 78.23% of areas, with 34.63% showing significant increases, also passing the 95% significance test. Among the three regions, the Yellow River Source Region showed the best vegetation recovery, with 41.53% of areas experiencing a significant increase in NDVI (0.0033/10a), primarily in the southeastern and northwestern regions. The Yangtze River Source Region was second, with 35.26% experiencing a significant increase in NDVI (0.0016/10a), primarily in the western and northern regions. In the Lancang River Source, only 16.81% of areas showed significant improvements in vegetation, with a more dispersed distribution.

We further analyzed the seasonal characteristics of vegetation change. As shown in Figure 7b–e, vegetation in all seasons across the study area exhibited varying degrees of recovery, with the proportion of areas showing NDVI increases accounting for 77.42%, 65.93%, 64.72%, and 67.30%, respectively. Areas with significant NDVI increases were substantially larger than those with decreasing trends, indicating that most regions passing the significance test experienced vegetation improvement. Spring was the season with the best vegetation recovery, with an NDVI growth rate of 0.0025/10a (p < 0.05) and the largest proportion of significantly increasing areas (32.14%), which largely coincided with the spatial distribution of annual NDVI increases. In contrast, the proportion of significantly increasing areas was below 20% for the other three seasons. During summer and autumn, vegetation growth slowed relative to spring, with growth rates of 0.0025/10a and 0.0016/10a, respectively, neither of which passed the significance test. Although winter showed the highest NDVI growth rate (0.0028/10a), it was not statistically significant, and the proportion of significantly increasing areas was relatively small (17.11%).

At the sub-regional scale, vegetation recovery in the Yellow River Source area was greater than the other two regions during spring, autumn, and winter, with notable NDVI increases observed in its northwestern part. In the Yangtze River Source area, vegetation showed the most pronounced recovery in summer, with significantly increasing NDVI areas accounting for 31.75% of the sub-region, while the increases in spring, autumn, and winter were weaker. Significant vegetation growth was mainly concentrated in the northwestern Yangtze River Source region. By contrast, the Lancang River Source exhibited the weakest vegetation recovery.

In summary, the growth of NDVI reflects that the ecological environment of the area has improved to some extent.

3.4. Statistical Analysis of Vegetation Trend Changes

To better examine the NDVI changes of various vegetation types from 1982 to 2014, we combined the NDVI trends obtained with the classification by vegetation types. The statistical results for NDVI trends of different vegetation types in the TRSR are shown in Figure 8. We found significant differences in seasonal changes among the three subregions within each vegetation type. Overall, most vegetation types in the TRSR showed an overall improvement, with the improved area far exceeding the degraded area. Among them, the NDVI of marshes increased most significantly in spring, accounting for 86.12% of the total area, but experienced some degree of deterioration in summer. Cultivated vegetation experienced relatively severe degradation. broad-leaved forests, grassland, and meadow all showed varying degrees of improvement across all four seasons, with improved areas accounting for over 60% of the total area.

From a regional perspective, the overall vegetation in the Yellow River Source area has improved significantly, and the improved area of all vegetation types is much larger than the degraded area. Among them, the growth of swamp is most significant in spring, with the significantly increased area accounting for 80%. Except for others vegetation types of this type, the remaining vegetation types show an improvement trend in all four seasons; the cultivated vegetation in the Yangtze River Source area has degraded in all four seasons, especially in winter, but more than 80% of the grassland in the region has improved in the four seasons, the improvement trend of meadow is relatively smaller than that of grassland, and the degradation trend of shrubland and coniferous forests in autumn and winter is greater than the improvement trend; the improvement area of broad-leaved forests in the Lancang River Source area has reached 100% in the four seasons, and the improvement of coniferous forests in spring is also greater than that in the Yellow River and Yangtze River Source, but the improvement speed of most vegetation in the region is relatively slow.

In summary, the overall vegetation condition in the TRSR showed some improvement between 1982 and 2014, but significant differences persisted between different vegetation types and regions.

4. Discussion

4.1. Value of the Downscaling Model

In this study, three regression algorithms—RF, LightGBM, and CatBoost—were employed to construct NDVI downscaling models. The spatiotemporal validation results demonstrate that the CatBoost model achieved the best overall performance. Across the TRSR, approximately 76% of the area achieved R² values above 0.7, while RMSE and MAE were generally below 0.075 and 0.05, respectively. In particular, strong correlations were observed for major vegetation types such as meadow, grassland, broad-leaved forests, coniferous forests, swamp, and shrubland, highlighting the model’s reliability in representing complex ecological environments. However, for the others vegetation type, the model’s performance was relatively poor, with an R² of only 0.56. This lower accuracy may be attributed to the smaller NDVI values of this vegetation type compared with grassland and meadow, making the data more susceptible to noise from clouds and other factors. Additionally, its sparse spatial distribution limits the number of available training samples, further affecting the model’s learning and fitting performance.

Compared with previous downscaling approaches [13,14], the proposed algorithm and modeling framework exhibit higher accuracy in describing the spatiotemporal variations of NDVI across different vegetation types, demonstrating strong potential for regional-scale vegetation monitoring.

4.2. The Spatiotemporal Dynamics of NDVI in the TRSR

This study reveals the spatiotemporal variation of vegetation in the TRSR from 1982 to 2014. Overall, vegetation in the region exhibited a significant improvement over the 33-year period, consistent with previous studies by Zhang et al. [31], particularly for swamp, grassland, meadow, and broad-leaved forests. It may be closely related to the relatively rapid warming, improved hydrothermal conditions, and relatively stable vegetation in the Yellow River Source [32,33]. NDVI across the region showed a fluctuating upward trend, with the maximum value recorded in 2010 and the minimum in 1983, aligning with the background of warming and increased precipitation and broadly consistent with the findings of Ma et al. [34]. The trend of climatic warming and wetting may serve as a driver of vegetation recovery in the TRSR, particularly in spring, when rising temperatures result in the most pronounced vegetation growth in swamp, meadow, and grassland. However, local degradation observed during summer may be related to changes in evapotranspiration, precipitation, or human activities such as overgrazing [35]. Spatially, NDVI increases were more pronounced in the southeastern part of the Yellow River Source area and the northwestern part of the Yangtze River Source area, whereas the Lancang River Source area showed relatively slower growth, likely due to the relatively weak trend of climate warming and humidification [36]. These spatiotemporal patterns are largely consistent with previous research, further confirming the overall trend of vegetation recovery in the TRSR.

4.3. Vegetation Trend Variations Among Regions

The NDVI trends from 1982 to 2014 indicate that vegetation generally improved in the TRSR, though the magnitude of change differed by vegetation type and region. Most vegetation types showed a significant greening trend, with the area of improvement far exceeding that of degradation. Swamp showed the most pronounced increase, particularly in spring, likely reflecting enhanced soil moisture [29]. In contrast, cultivated vegetation showed clear signs of degradation in all seasons, suggesting the negative influence of intensive human activities, land use, and potential soil degradation [37]. Broad-leaved forests, grassland, and meadow all showed sustained improvement throughout the year, consistent with the effects of ecological restoration initiatives such as grazing exclusion and afforestation implemented [6].

Regionally, vegetation recovery was most pronounced in the Yellow River Source area. In the Lancang River source area, broad-leaved forests improved noticeably, but the overall recovery rate was low, possibly because temperature and precipitation changed only slightly in this region.

4.4. Limitations and Uncertainties of the Downscaled NDVI

While our study provides a detailed assessment of vegetation changes in the TRSR, several limitations and uncertainties remain. First, due to the lack of in situ NDVI observations, the validation relied solely on MODIS NDVI data, which may introduce certain biases and uncertainties. Second, the downscaling model only considered topographic factors such as DEM while excluding key environmental variables such as climate and soil conditions. This simplification may overemphasize the role of elevation in the model training process.

Moreover, NDVI data cannot fully capture changes in vegetation biomass and ecological functions. Therefore, future research could incorporate other remote sensing data, such as the Leaf Area Index (LAI) and Fractional Vegetation Cover (FVC), with field survey data to provide a more comprehensive analysis. While this study reveals temporal and spatial trends, it does not delve into the driving mechanisms of NDVI changes, such as climate and human activities. Therefore, future research could combine climate models with land use data to further explore the vegetation dynamics.

5. Conclusions

This study used machine learning algorithms to construct a high-resolution, long-term, monthly NDVI dataset from 1982 to 2014. Compared with previous studies, the time span was extended. Based on different time scales, combined with vegetation types and statistical methods such as the MK test, a systematic analysis was conducted on the spatiotemporal variation characteristics of NDVI in the TRSR and the differences between different vegetation types. The main conclusions are as follows:

(1)

The vegetation changes in the TRSR showed an increasing trend, with the most significant increases occurring in swamp, coniferous and broad-leaved forests, meadow, and grassland. NDVI generally decreased from southeast to northwest.

(2)

Most vegetation types showed improvement. Most vegetation types exhibited improvement, with the area of greening far exceeding degradation. Swamp improved most notably in spring, while cultivated vegetation experienced more degradation.

(3)

Regional differences were evident: the Yellow River Source Region showed the strongest vegetation recovery; the Yangtze River Source Region exhibited grassland improvement but cultivated vegetation decline; and the Lancang River Source Region showed notable broad-leaved forests recovery, though overall change was moderate.

These findings are important for understanding the dynamics of vegetation ecosystems and for developing targeted ecological protection and restoration strategies. Strengths of this study include the construction of a long-term, high-resolution monthly NDVI dataset using advanced machine learning algorithms and its integration with statistical trend detection to comprehensively assess spatiotemporal vegetation dynamics. Future research could integrate climate and soil data to better identify the drivers of vegetation change, incorporate field observations for model validation, and further improve the downscaling framework to enable more detailed analyses of vegetation dynamics in the TRSR.