Satellite-Derived Spatiotemporal Dynamics of Vegetation Cover and Its Driving Factors in the Three-River Headwaters Region from 2001 to 2022

Abstract

To preserve ecological integrity and promote sustainable progress in the Three-River Headwaters Region (TRHR), it is vital to understand the vegetation alteration patterns and the sensitivity of these patterns to climatic and anthropogenic influences. In this study, we retrieved the fractional vegetation cover (FVC) through the dimidiate pixel model, driven by MODIS reflectance data from 2001 to 2022, and analyzed its spatiotemporal variations and responses to climate variation and human activities via partial correlation and residual analyses. The results indicated that the FVC retrieval accuracy reached 84.2%. From 2001 to 2022, the growing season FVC displayed a fluctuating yet overall increasing trend, with an average growth rate of 0.23% per year (p < 0.01). The vegetation significantly improved in 50.72% of the TRHR, with the Yellow River source area exhibiting the most notable improvement. However, 67.42% of the TRHR experienced a transition from improvement to degradation in vegetation, indicating a pessimistic outlook for future changes. Partial correlation analysis revealed that temperature had a pronounced influence on the southwestern Yellow River Basin and the southern Yangtze River Basin, whereas precipitation had a substantial effect on the southwestern and northeastern sections of the Yellow River Basin. Additionally, residual analysis revealed that climate change served as the predominant factor behind the changes in the FVC, whereas the anthropogenic intervention contributed substantially to vegetation improvements in the northeastern and western portions of the Yellow River Basin. Our study provides scientific support for the construction of ecological security barriers and the harmonious development of humans and nature in the TRHR.

1. Introduction

The Three-River Headwaters Region (TRHR) acts as a vital ecological shield for China and is one of the most delicate and susceptible areas; the environmental fluctuations within the region exert substantial influences on the climate and environment at the national and even continental scales [1,2,3,4]. As an integral component of the ecosystem of the TRHR, vegetation is a crucial metric for gauging the condition of terrestrial vegetation and characterizing environmental changes within the ecosystem [5,6,7]. The fractional vegetation cover (FVC) serves as a fundamental metric for characterizing vegetation coverage and is typically designated as the percentage of the area of the entire study region occupied by the orthogonal projection of plants (encompassing branches, stalks, and leaves) on the ground [8,9]. Over the past few decades, the FVC in the TRHR has experienced degradation, and the region is facing environmental issues such as sharp declines in biodiversity, land desertification, and a reduction in water conservation capacity [6,10,11]. Fortunately, in recent years, the FVC within the TRHR has migrated toward a positive development trajectory under the influence of global environmental changes (e.g., global warming and glacier melting) and a set of ecological measures undertaken by the state [12,13]. Therefore, an in-depth exploration of vegetation variations and their attribution in the TRHR is urgently required [14].

Remote sensing has been extensively applied to large-area and long-term vegetation estimations in the TRHR. The majority of studies directly employ the NDVI as a proxy for vegetation cover [15,16,17,18]. For example, Zhang et al. [6] mapped the spatiotemporal variations in vegetation cover using MOD13Q1 NDVI data and reported that the mean NDVI exhibited a fluctuating yet ascending trajectory, growing at an annual rate of 0.0013. However, the NDVI exhibits minimal variation in high-vegetation-cover regions and is highly responsive to the reflectance of the soil background. In contrast, the FVC, as a proportional value, can more accurately represent the actual state of the vegetation cover in a region. In particular, the dimidiate pixel model has been extensively adopted because of its simple formula with two parameters (NDVI for pixels representing pure bare soil and pure vegetation) and demonstrated physical significance. Although some studies have attempted to retrieve FVC, they often lack extensive ground-observed FVC data for validation [14,19,20].

Many studies have focused on the spatiotemporal variation in the FVC in the TRHR [14,21,22]. These studies have drawn similar conclusions that the vegetation cover within the TRHR has grown markedly in recent years, but there are large differences in the magnitudes of this trend. For example, Liu et al. [21] determined vegetation cover by means of the dimidiate pixel model and reported that the FVC exhibited a very pronounced upward trend during the growing season between 1998 and 2012, experiencing an average annual increase of 0.004. Employing the dimidiate pixel model, Zhang et al. [22] noted an annual increase rate of 0.23% in grassland coverage across the TRHR from 1982 to 2012, with 63.96% of the area exhibiting an increasing trend. Xie et al. [14] employed trend analysis and reported that the FVC notably increased from 2001 to 2020, growing at a rate of 2.1%/10a, of which the northeastern and northwestern portions of the TRHR experienced significant increases. Differences in data sources, methodologies, time spans, and spatial resolutions, as well as the complex terrain and ecology of the TRHR, result in inconsistent conclusions and uncertainty regarding vegetation variations in the TRHR [14].

Many scientists have highlighted how vegetation responds to climatic variability and anthropogenic activities in the TRHR. Some scientists have investigated the consequences of climate variability for vegetation and have posited that it substantially influences vegetation growth [14,19,23,24]. For example, Zhang et al. [23] and Chen et al. [24] found that both precipitation and temperature support vegetation recovery, with temperature exerting a greater influence than precipitation. Xie et al. [14] and Zhao et al. [19] proposed that precipitation was instrumental in driving FVC. Furthermore, other scientists have investigated the impact of anthropogenic activities on vegetation cover, and most have revealed that both factors contribute to vegetation recovery, with the effects of climatic variability being stronger than those of anthropogenic actions [6,15,25,26]. For example, Li et al. [25] recognized climate variability as the key determinant of vegetation variation, yet human activities can intensify the rate of change over a short period. Sun et al. [26] concluded that vegetation alterations in the TRHR were primarily controlled by climatic factors, with ecological engineering having a limited impact. Conversely, He et al. [27] posited that anthropogenic activities had a more significant positive impact than climatic variability, contributing 59.07% and 40.93%, respectively, to the overall effect. Scientists have drawn diverse and sometimes starkly opposing conclusions because of the complexity of driving factors and the diversity of methodologies, which has led to uncertainty regarding the influence of climatic and anthropogenic parameters on vegetation. Hence, a quantitative appraisal of the impacts of climatic parameters and anthropogenic interventions on vegetation cover is vitally necessary.

In this study, we applied the MODIS NDVI dataset to examine the spatial-temporal dynamics of the FVC and its determinants in the TRHR. We had three main objectives: (1) to retrieve and validate the monthly FVC; (2) to examine the spatiotemporal variations in FVC over a 22-year period; and (3) to explore the impacts of climatic parameters and anthropogenic interventions on the FVC.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area

The TRHR, situated within the coordinates 89°45′–102°23′E and 31°39′–36°12′N, lies in the interior of the Qinghai-Tibet Plateau and southern Qinghai Province, covering an area of 302,500 km² (Figure 1) [6,17,18,28]. The TRHR operates as the headwater region for the Yangtze River, the Yellow River, and the Lancang River and is widely recognized as the “Chinese water tower” [2,29]. The mean altitude of the region exceeds 4500 m, and the Bayan Har Mountains, Tanggula Mountains, Anyemaqen Mountains, and Hoh Xil Mountains traverse it, forming the core framework of the TRHR [17,18,28,29]. The area features a characteristic plateau continental climate, with a mean yearly temperature ranging from −5.6 to 3.8 °C and annual rainfall ranging from 262.2 to 772.8 mm, of which the precipitation from June to September constitutes approximately 75% of the yearly total [18,29,30,31]. The TRHR has a climate characterized by distinct cold and hot seasons, pronounced wet and dry periods, strong solar radiation, and extended sunshine duration. The study area, which is densely covered with rivers, swamps, and lakes, is recognized as the largest wetland ecosystem on the planet [17,32]. The total wetland area is 73,300 km², accounting for 24% of the study area. Alpine grasslands, which include both alpine meadows and alpine steppes, constitute the predominant land cover type, accounting for 79% of the vegetation-covered area in the TRHR, and the growing season extends from May to September [6,33]. The TRHR administrative region encompasses 16 counties spread across the Tibetan Autonomous Prefectures of Huangnan, Hainan, Guoluo, and Yushu, in addition to Tanggula Township in Golmud city, covering approximately 43% of Qinghai Province [1,34].

2.1.2. Data

MODIS NDVI data were downloaded from the Land Processes Distributed Active Archive Center in the US (https://lpdaac.usgs.gov). We selected the MOD13A3 NDVI data covering the growing season months of May through September between 2001 and 2022. This dataset has a spatial resolution of 1 km and is provided on a monthly basis. The data were preprocessed, which included calibration, geometric correction, and atmospheric correction. The MODIS Reprojection Tool (MRT 4.1) software was used for converting data formats and performing projection transformations. The FVC ground observations that were used to verify the accuracy of the FVC estimations were obtained from the Qinghai Eco-Environment Monitoring Center (http://www.qheemc.com/), and a total of 122 effective FVC observations from 2009 were selected after screening.

Monthly meteorological data, including mean temperature and total precipitation throughout the growing season, spanning the years 2001 to 2022, at a spatial resolution of 1 km, were retrieved from the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn/). The elevation data were acquired from the Geospatial Data Cloud (http://www.gscloud.cn/) at a 90 m spatial resolution. The land cover data were derived from the GlobeLand30 dataset (https://www.ngcc.cn/, which has a 30 m spatial resolution. The data encompassed 10 land cover types: forest, shrub, grassland, artificial surface, cropland, water, bare land, wetland, tundra, and snow/ice, and there was no tundra in the TRHR. All the data were processed using the ArcMap 10.8 software, which involved steps such as clipping, conversion to the WGS 1984 coordinate system, and rescaling to a 1 km resolution.

2.2. Methods

2.2.1. Dimidiate Pixel Model

The dimidiate pixel model is a hybrid pixel decomposition model commonly employed for vegetation coverage estimation and is known for its simple operation and high accuracy [35]. The principle of the model assumes that an image element is composed of bare soil and vegetation; then, the spectral information of an image element can be obtained from the weighted combination of the two parts, and the weight is the proportion of each part within the image element. Pixels composed entirely of vegetation cover are regarded as pure vegetation pixels, whereas those completely covered by bare soil are identified as pure bare soil pixels, thereby obtaining the contributions of vegetation and bare soil to a mixed pixel [36,37,38,39,40,41]. Following the foundational principle of the dimidiate pixel model, we obtained the vegetation cover estimation model using the NDVI, with the formula presented below:

$$FVC={\displaystyle \frac{NDVI-{NDVI}_{soil}}{{NDVI}_{veg}-{NDVI}_{soil}}},$$

(1)

where $NDVI$ represents the NDVI measurement of the mixed raster cell; ${NDVI}_{soil}$ represents the NDVI measurement of a pure bare land raster cell; and ${NDVI}_{veg}$ represents the NDVI measurement of a pure vegetation raster cell. However, due to interference from the atmosphere, light, soil moisture, and other factors [35,42], ${NDVI}_{soil}$ and ${NDVI}_{veg}$ are not equal to the theoretical values of 0 and 1, and previous studies instead often use the minimum and maximum NDVI measurements of image data within the specified confidence interval [20,43,44]. In this study, based on the NDVI values in the TRHR and reference to relevant studies [21,22], 0.05 and 0.85 were chosen to represent ${NDVI}_{soil}$ and ${NDVI}_{veg}$, respectively. The FVC was categorized into 5 levels according to relevant studies and the specific conditions of the TRHR (

Table 1. FVC categorization standards.

Class Name	Categorization Criteria
Lower	0–15%
Low	15–30%
Medium	30–45%
High	45–60%
Higher	60–100%

).

2.2.2. Statistical Metrics

To quantify the precision of the dimidiate pixel model, we selected four evaluation metrics: coefficient of determination (R²), bias, root mean square error (RMSE), and Kling–Gupta efficiency (KGE). The RMSE was applied to assess the discrepancy between the ground-measured FVC and the estimated FVC. Bias reflects the disparity between the actual observations and the corresponding estimated FVC. R² was applied to indicate the fitting degree of the dimidiate pixel model. The formulas are as follows:

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

(2)

$$Bias={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left(y_i-x_i\right),$$

(3)

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

(4)

KGE can comprehensively reflect the correlation, variability, and deviation between the estimated and ground-measured FVC values, which can be applied to assess the model’s accuracy. The KGE value range is (−∞, 1), with values closer to 1 suggesting greater model precision.

$$KGE=1-\sqrt{{\left(r-1\right)}^2+{\left({\displaystyle \frac{\frac{{\sigma }_y}{\overline{y}}}{\frac{{\sigma }_x}{\overline{x}}}}-1\right)}^2+{\left({\displaystyle \frac{\overline{y}}{\overline{x}}}-1\right)}^2},$$

(5)

where in Formulas (2)–(5), $x_i$ and $y_i$ denote the ground-observed and estimated FVC, respectively, whereas $\overline{x}$ and $\overline{y}$ signify the mean figures of the ground-observed and estimated FVC, respectively. ${\sigma }_x$ and ${\sigma }_y$ correspond to the standard deviations of the ground-measured and estimated FVC, respectively. $r$ refers to the correlation coefficient.

2.2.3. The Theil–Sen Median and the Mann-Kendall Significance Test

To examine the trend of the average growing season FVC, the Theil-Sen median and the M-K significance test [45,46,47,48] were applied, and the trend was classified into 5 types, as shown in

Table 2. Categorization standards for trend analysis.

β	Z	Categorization Standards
$\beta <-0.001$	$\left|Z\right|>1.96$	Significant decrease
$\beta <-0.001$	$\left|Z\right|\le 1.96$	Nonsignificant decrease
$-0.001\le \beta \le 0.001$	$-$	Basically stable
$\beta >0.001$	$\left|Z\right|\le 1.96$	Nonsignificant increase
$\beta >0.001$	$\left|Z\right|>1.96$	Significant increase

[6]. The method is effective at mitigating the impact of outlier data and measurement errors, which can minimize the interference of outliers [49,50]. The expression is presented below:

$$\beta =median\left({\displaystyle \frac{{FVC}_j-{FVC}_i}{j-i}}\right)\left(1<i<j<n\right),$$

(6)

where $\beta$ refers to the slope of the FVC, ${FVC}_j$ and ${FVC}_i$ represent FVC values in different years, and $j$ and $i$ are time series. The FVC increases when $\beta >0$ and decreases when $\beta <0$. The M-K test was performed to assess the significance of the FVC pattern with the following formulas:

$$S=\sum _{i=1}^{n-1}\sum _{j=i+1}^{n}sgn\left(x_j-x_i\right),$$

(7)

$$Z=\left\{\begin{array}{c}{\displaystyle \frac{S-1}{\sqrt{var\left(S\right)}}} (S>0)\\ 0 (S=0)\\ {\displaystyle \frac{S+1}{\sqrt{var\left(S\right)}}} (S<0)\end{array}\right.,$$

(8)

$$Var\left(S\right)={\displaystyle \frac{n\left(n-1\right)\left(2n+5\right)}{18}},$$

(9)

where $n$ denotes the overall count of the data. When $n\ge 8$, $S$ is approximately normally distributed, with $Z$ serving as the standardized test statistic. When $\left|Z\right|>1.96$, the trend can be deemed statistically significant with 95% confidence [51].

2.2.4. Hurst Index

The Hurst index [52], originating from R/S analysis, constitutes a reliable approach to quantitatively assess the enduring dependency in temporal sequence data [49]. The Hurst index has been successfully applied to vegetation studies to reflect the sustainability of vegetation development trends [53,54,55]. Consequently, we computed the Hurst index to examine the FVC variation, as detailed below:

• The average FVC time series was determined as follows:

$${\overline{FVC}}_{\left(\tau \right)}={\displaystyle \frac{1}{\tau }}\sum _{t=1}^{\tau }{FVC}_{\left(\tau \right)} \tau =1, 2, 3, \cdots , n.$$

(10)

2.

The cumulative deviation was computed in the following manner:

$$X_{\left(t,\tau \right)}=\sum _{j=1}^t\left({FVC}_{\left(t\right)}-{\overline{FVC}}_{\left(\tau \right)}\right) 1\le t\le \tau .$$

(11)

3.

The range of deviation was determined in the following way:

$$R_{\left(\tau \right)}=\underset{1\le t\le \tau }{\max }{X}_{\left(t,\tau \right)}-\underset{1\le t\le \tau }{\min }{X}_{\left(t,\tau \right)}.$$

(12)

4.

The standard deviation was determined in the following manner:

$$S_{\left(\tau \right)}=\sqrt{{\displaystyle \frac{1}{\tau }}\sum _{t=1}^{\tau }{\left({FVC}_{\left(t\right)}-{\overline{FVC}}_{\left(\tau \right)}\right)}^2}.$$

(13)

5.

The Hurst index was computed in the following manner:

$${\displaystyle \frac{R_{\left(\tau \right)}}{S_{\left(\tau \right)}}}={\left(c\tau \right)}^H,$$

(14)

where $H$ refers to the Hurst index, which varies between 0 and 1 and is classified as detailed in

Table 3. Standards for categorizing the Hurst index.

Nature of Change	Categorization Standards
Persistence	0.5 < Hurst < 1
Random	Hurst = 0.5
Anti-persistence	0 < Hurst < 0.5

.

2.2.5. Partial Correlation Analysis

To investigate how climatic elements affect FVC, partial correlation analysis was carried out to compute the partial correlation coefficients between different climatic elements and the FVC. This method can calculate the correlation coefficients between various independent elements and the dependent variable. When calculating the correlation coefficient for a specific element relative to the dependent variable, the contributions of other independent elements are temporarily disregarded, allowing for a more accurate reflection of the relationship between the two variables [56,57,58]. The formula for the calculation is presented below:

$$R_{abc}={\displaystyle \frac{r_{ab}-r_{ac}\times r_{bc}}{\sqrt{\left(1-{r_{ac}}^2\right)+\left(1-{r_{bc}}^2\right)}}},$$

(15)

where $R_{abc}$ denotes the partial correlation coefficient between element $a$ and element $b$, with the effect of element $c$ removed. $r_{ab}$, $r_{ac}$, and $r_{bc}$ refer to the correlation coefficients between elements $a$ and $b$, $a$ and $c$, and $b$ and $c$, respectively.

2.2.6. Residual Trend Analysis

To differentiate the influences of climatic variation and anthropogenic interventions on the FVC trend, we employed residual trend analysis and calculated the respective contribution rates of these two factors [59]. First, a multiple linear regression analysis involving FVC and various climatic factors was conducted, with the residual calculated as follows [60,61]:

$${FVC}_{pre}=\alpha \times T+\beta \times P+\phi ,$$

(16)

$${FVC}_{res}={FVC}_{obs}-{FVC}_{pre},$$

(17)

where ${FVC}_{pre}$ represents the predicted FVC value, which represents the FVC value influenced by climate. $T$ denotes the average temperature. $P$ represents precipitation, $\alpha$ and $\beta$ are the regression coefficients of FVC and climate factors, respectively, and $\phi$ is a constant term. ${FVC}_{res}$ refers to the residual value, which represents the FVC value influenced by human activities [62]. ${FVC}_{obs}$ refers to the FVC value estimated via remote sensing.

Through the trend analysis of ${FVC}_{pre}$ and ${FVC}_{res}$, the FVC trends modulated by climatic factors and human-induced influences were obtained. $S>0$ suggests that climatic factors or anthropogenic influences enhance vegetation growth, while $S<0$ indicates inhibition.

Table 4. Evaluation methods for impact factors.

$\mathit{S}\left({\mathit{F}\mathit{V}\mathit{C}}_{\mathit{o}\mathit{b}\mathit{s}}\right)$	$\mathit{S}\left({\mathit{F}\mathit{V}\mathit{C}}_{\mathit{p}\mathit{r}\mathit{e}}\right)$	$\mathit{S}\left({\mathit{F}\mathit{V}\mathit{C}}_{\mathit{r}\mathit{e}\mathit{s}}\right)$	Impact Factors
>0	>0	>0	CC 1 and HA 2
	>0	<0	CC
	<0	>0	HA
<0	<0	<0	CC and HA
	<0	>0	CC
	>0	<0	HA

¹ CC stands for climatic factors. ² HA stands for anthropogenic influences.

presents the methods used to evaluate the impact factors [63]. The contribution rates of climatic elements and anthropogenic influences on FVC variations are presented below:

$$r_1={\displaystyle \frac{S\left({FVC}_{pre}\right)}{S\left({FVC}_{obs}\right)}}\times 100\%,$$

(18)

$$r_2={\displaystyle \frac{S\left({FVC}_{res}\right)}{S\left({FVC}_{obs}\right)}}\times 100\%,$$

(19)

where $r_1$ is the proportionate contribution of climatic elements to the FVC and $r_2$ denotes the proportionate contribution of anthropogenic influences to the FVC.

3. Results

3.1. Validation of FVC Retrieval in the TRHR

In this study, we constructed a dimidiate pixel model to estimate the vegetation cover throughout the growing season within the TRHR from 2001 to 2022. The mean FVC for each month was subsequently considered the average FVC. The FVC obtained via NDVI remote sensing retrieval was compared with the FVC measured on the ground, and four evaluation metrics, R², RMSE, bias, and KGE, were calculated to assess the model accuracy (Figure 2). The RMSE (15.8%) showed that the precision of FVC estimation based on the NDVI was 84.2%, which suggests that it is feasible to estimate FVC using the model adopted in this study. The bias (7.3%) showed a slight discrepancy between the observations and the estimated values. An R² of 0.60 (p < 0.01) and a KGE of 0.68 suggest that the model was well fitted and highly accurate.

3.2. Spatial Distribution of the FVC in the TRHR

The geographical distribution of the mean growing season FVC over the period from 2001 to 2022 is depicted in Figure 3. The FVC exhibited a progressive decrease from southeast to northwest, featuring an average vegetation cover of 37.69%. Regions of high FVC values were predominantly positioned in the southern and eastern TRHR, which are marked by relatively low altitudes and appropriate hydrothermal conditions. The FVC was low in the northwestern part of the TRHR, with values ranging between 0 and 30%, and this area featured a high altitude and a frigid and arid climate characterized by barren land, numerous glaciers, and a multitude of lakes and swamps.

3.3. Spatiotemporal Dynamics of the FVC from 2001 to 2022

3.3.1. Temporal Variation in the FVC

The 22-year temporal trend of the FVC within the TRHR was assessed using the annual mean FVC during the growing season (Figure 4). The FVC exhibited a fluctuating yet overall upward trajectory, with a mean annual increase of 0.23% (p < 0.01). The minimum FVC was registered in 2001 at 33.65%, whereas the maximum occurred in 2018 at 42.2%. From 2001 to 2002, the FVC increased greatly, and it decreased in 2003; over the next four years (2003–2006), the FVC continued to grow before decreasing again in 2007; there was a surge in the FVC in 2009 and 2010, after which the value plummeted in 2011 and regained its peak in 2012; the FVC continued to decrease from 2013 to 2016 and rebounded sharply in the following two years; the value remained relatively high from 2018 to 2021 and decreased in 2022.

The FVC in the TRHR from 2001 to 2022 was categorized into different classes based on Table 1, and Figure 5 presents the variations in the proportions of various FVC categories relative to the total area over the 22 years. There was a notable decrease in the proportion of areas with lower vegetation cover, which decreased from 24.15% in 2001 to 15.97% in 2022 (k = −0.303%/a, p < 0.001) and was lowest in 2018, at 15.42%. Conversely, the area with higher vegetation cover increased significantly from 13.22% in 2001 to 22.29% in 2022 (k = 0.331%/a, p < 0.01), reaching its peak in 2021 (28.01%). Regions with high, medium, and low vegetation cover showed no significant changes and remained relatively stable. In conclusion, the growth of the average FVC in the TRHR was caused mainly by a decrease in the lower vegetation cover area and an increase in the higher vegetation cover area.

3.3.2. Spatial Variation in the FVC

To characterize the changes in the FVC within the TRHR in greater detail, the Theil–Sen slope method and the M-K test for significance were employed to examine the spatial variations in the FVC at the pixel scale (Figure 6). The regions that met the significance test criteria (95% significance level) comprised 51.91% of the TRHR and were largely found in the western and central sections of the YaRB and the western regions of the YeRB (Figure 6a). The FVC predominantly displayed an increasing trend, encompassing 87.11% of the total area, with 50.72% demonstrating a significant increase and 36.39% experiencing a nonsignificant increase (Figure 6b). Regions with notable increases in FVC were situated primarily across the west, southwest, and northeast of the YeRB and throughout the western and central sections of the YaRB. The areas with stable vegetation constituted 8.39% of the TRHR and were predominantly dispersed throughout the western and central sections of the YaRB. A mere 4.5% of the TRHR exhibited a declining trend, which was primarily situated in the central section of the eastern YaRB. In summary, the origin area of the Yellow River was recognized as the most important region of vegetation improvement within the TRHR.

3.3.3. Future Trends in FVC

To forecast the future trends in the growing season FVC, the Hurst index was computed at the pixel scale (Figure 7a). The Hurst index in the TRHR varied between 0.128 and 0.875, with a mean value of 0.45, indicating relatively strong anti-persistence. A Hurst value > 0.5 suggests that the future change in the FVC will align with the current state, and the regions where Hurst > 0.5 shows persistent characteristics encompass 27.55% of the total area. The regions where Hurst < 0.5 makes up 72.45% of the TRHR show anti-persistent trends, and the future changes in FVC are contrary to the present conditions.

The outcomes of the Hurst index and Sen slope estimation were combined to further explore the prospective development of the FVC, and four types of changes were obtained, as shown in Figure 7b. The regions transitioning from improvement to degradation accounted for the largest proportion (67.42%), and these regions were extensively dispersed in the southwest and east of the TRHR, and the future changes were not optimistic. The second area was continuously improving, accounting for 25.23% of the TRHR; this area was predominantly located in Qumalai County and Maduo County and was scattered in other areas. The areas that transitioned from degradation to improvement and those that experienced continuous degradation constituted a small proportion of the total area (5.03% and 2.32%, respectively). Spatially, both types of areas were similarly observed in the eastern section of Zhiduo County, the southern section of Qumalai County, and along the border between Gande County and Dari County. This distribution indicates progress in vegetation restoration work, but some degraded areas have yet to show signs of improvement.

3.4. Contribution of Driving Factors to FVC

3.4.1. Climate Factors Affecting FVC

Climate variation is a critical factor influencing FVC, where precipitation and temperature are the key climatic drivers [21,64]. Figure 8a,b present the geographical differences in the average precipitation and temperature within the vegetative period. The mean temperature from 2001 to 2022 was −10.91~14.55 °C, and the large regional temperature differences were attributed primarily to differences in altitude. The temperatures in the southern and eastern valleys were greater than those in the western plateau with glaciers and snow cover. The total precipitation across the growing season varied from 158.44 to 679.80 mm, with great regional differences, and the spatial pattern was “relatively low amounts in the northwestern area and relative high amounts in the southeastern area”, which was the same as that of the FVC.

Figure 8c shows the time series analysis results of the climatic factors and FVC, which largely exhibited similar variations, especially during 2008–2012 and 2017–2021. In 2016 and 2022, the average temperature was relatively high, and the FVC decreased, which may have been due to low precipitation and was restricted by water conditions. To delve deeper into how the FVC responds to climate change, we carried out partial correlation analysis to assess the connection between the FVC and various climate parameters. The analysis indicated that the partial correlation of the FVC with temperature was 0.63 (p < 0.01), whereas the partial correlation with precipitation was slightly greater at 0.65 (p < 0.01). Therefore, on an annual scale, precipitation exerted a marginally stronger influence than temperature on FVC. At the monthly level, temperature had a more pronounced effect than precipitation, as illustrated by a partial correlation value of 0.79 (p < 0.001) for FVC and temperature, which was much greater than the value for FVC and precipitation (0.32, p < 0.001).

To more effectively reveal the role of meteorological elements in the FVC, partial correlation coefficients were computed at both the site-specific scale and the pixel-specific scale (Figure 9). At the site scale, the FVC was positively associated with both climatic elements, with the impact of precipitation being notably stronger than that of temperature. Figure 9a shows that a direct relationship between FVC and temperature was detected at 88.19% of the sites, with 21.26% of these sites passing the significance test; a total of 11.81% of the sites demonstrated a negative correlation, predominantly positioned in the eastern portion of the TRHR, and no sites passed the significance test. As depicted in Figure 9b, the sites with a significant positive association and a nonsignificant positive relationship between precipitation and FVC constituted 59.84% and 37.01%, respectively, whereas only 3.15% of the sites exhibited a negative relationship. Moreover, the spatial variability of the partial association between precipitation and FVC was not pronounced.

At the pixel scale, the FVC was directly associated with precipitation and temperature across most of the study region. As illustrated in Figure 9c, the zone with a markedly positive relationship between temperature and FVC comprised 36.17% of the TRHR and was located mainly in the southern YaRB and the southwestern YeRB. A negative correlation was evident in 12.82% of the TRHR, which was scattered in the northwest zone of the YaRB and the northeast zone of the YeRB. As depicted in Figure 9d, the area with a markedly direct relationship between precipitation and FVC comprised 36.33% of the TRHR and was largely found in the northeast and southwest sections of the YeRB, along with the southeastern zone of the YaRB. The nonsignificant positive correlation regions occupied the largest area (57.67%) and were largely found in the western and central sections of the TRHR. To summarize, the overall influences of precipitation and temperature on FVC were roughly the same; the FVC in the southern zone of the YaRB and the southwestern portion of the YeRB was most susceptible to temperature effects, whereas the FVC in the northeastern zone of the YeRB and the southeastern zone of the YaRB was most affected by precipitation, and the FVC in Dari County was greatly impacted by both of these factors.

3.4.2. Contribution Rates of Climate Change and Human Activities

Vegetation variations are chiefly influenced by climate variations and anthropogenic activity, and we used residual analysis to distinguish their contributions. As shown in Figure 10a, both climate variations and anthropogenic activity contribute positively to vegetation restoration. At the regional scale, climate variations explained 47.5% of the change in the FVC, whereas anthropogenic interventions contributed to 52.5% of the FVC change, indicating that the influence of anthropogenic activities was slightly more impactful. Figure 10b illustrates the spatial pattern of how human activities affect the FVC. The area where human activities play an active role accounted for 81.03% of the TRHR and was largely present in the western and northeastern zones of the YeRB, which indicates that the ecological conservation and development initiatives have begun to yield positive outcomes. The area where anthropogenic activities suppressed the FVC was 18.97%, which was largely found in the eastern YaRB and the southeastern YeRB.

Overlaying the residual analysis results with the Theil–Sen trend analysis results allowed us to ascertain the individual contributions of climatic elements and anthropogenic interventions to the impact on FVC, and we classified the impact types into six categories (Figure 10c). Among them, the FVC growth resulting from both climate variations and human interventions represented the largest proportion, comprising 47.65% of the TRHR, and was predominantly located in the southeastern YaRB and LRB, along with the northeastern YeRB. Second, the FVC growth area was dominated solely by climate variations, accounting for 44.13% of the region, and was positioned northwest of the TRHR, a region characterized by high altitude and a sparse population. The relative contribution rate was calculated for the regions with combined impacts of the two factors (Figure 10d). We found that the contribution of climate variations throughout the study area was high, above 50%, and was above 80% in most areas. In conclusion, the fluctuations in the FVC within the TRHR were chiefly driven by climatic conditions, with some areas with restricted impacts from human activities, which were primarily positioned in the eastern zone of the YaRB and the northern zone of the YeRB.

4. Discussion

4.1. Spatiotemporal Variation in FVC

In this study, we integrated 122 ground observations and employed four evaluation metrics—RMSE, bias, R², and KGE—to validate FVC retrieval. This approach addresses the shortcomings of previous research, which often lacked a thorough validation of inversion accuracy or relied on a limited number of validation samples [14,19,20]. Subsequently, we employed the Theil–Sen median and the Mann-Kendall significance test to investigate FVC trends over the past 22 years, which, compared to simple linear regression, reduces the influence of outliers and errors, providing a more robust analysis. We noted that the growing season FVC displayed a trend of gradual increase, with fluctuations from 2001 to 2022, which matched the results of related studies [6,14,21,65]. A continuous decline in FVC was observed from 2012 to 2016, potentially attributed to the abrupt decrease in both temperature and precipitation during this period and the influence of the El Niño event around 2015, which led to anomalies in precipitation and temperature [17]. Furthermore, we incorporated the calculation of trends in the FVC across different classes and discovered that the growth in the FVC was due primarily to the reduction in zones with lower vegetation cover and the expansion in zones with higher vegetation cover.

At the pixel level, our findings indicated that regions with notable increases in the FVC were mostly found in the western, southwestern, and northeastern zones of the YeRB, as well as in the western and central sections of the YaRB. Conversely, areas with a decline in FVC were predominantly situated in the Bayan Har Mountains of the eastern YaRB, along with the southeastern section of the YeRB. These findings corresponded closely with the results reported in previous studies [6,14,23,27]. Additionally, our study revealed that the future trends of FVC were dominated by anti-persistence, with substantial continuous growth observed solely in Qumalai County and Maduo County, which aligns with the findings of prior studies [6,14]. The proportion of areas experiencing a transition from vegetation improvement to degradation was the highest, constituting 67.42% of the total area, which suggests a pessimistic outlook for future changes. As predicted by Liu et al. [65], with further increases in temperature, potential evapotranspiration is also expected to rise, leading to rising temperatures and declining humidity in the TRHR, thereby inhibiting the growth of vegetation.

4.2. Effects of Climate Change and Human Activities

We explored the impact of climate variations at different scales (regional, site, and pixel scales) and reported that the FVC was positively related to both precipitation and temperature, which aligns with prior study findings [14,15,21,23]. Nevertheless, the impacts of these two climatic elements on FVC differ among various studies, with precipitation exhibiting the most notable difference when compared to the other factors. This discrepancy may arise from the substantial spatial variability in precipitation, coupled with its high intra-annual variability and interannual fluctuations. Consequently, across various temporal and spatial dimensions, the trends and patterns of precipitation can exhibit substantial variations [65]. On the other hand, atmospheric precipitation is not the only source of soil moisture in the TRHR. For example, Zhang et al. [6] demonstrated that artificial irrigation and artificial rainfall diminished the restrictive influence of precipitation on the growth of grassland vegetation, especially after human-led ecological conservation efforts were implemented.

Additionally, we performed a quantitative analysis of the contributions of climate change and human activities as two driving factors. The results revealed that both climatic and human-driven factors positively influenced FVC, with climate variation having the greatest impact, which resembles the conclusions by Xie et al. [14] and Zhao et al. [19]. Human activities have substantially improved vegetation in the western and northeastern zones of the YeRB, which is strongly tied to the range of conservation initiatives undertaken by the state (

Table 5. Summary of conservation initiatives in the TRHR.

Start Date	Project
19 August 2000	Three-River Source Provincial Nature Reserve
24 January 2003	Three-River Source National Nature Reserve
30 August 2005	The first phase of Ecological Protection and Construction Project of the Three-River Headwaters (2005–2013)
18 December 2013	The second phase of Ecological Protection and Construction Project of the Three-River Headwaters (2013–2020)
10 January 2014	National Ecological Conservation Comprehensive Experimental Zone of Three-River Headwaters in Qinghai Province
5 March 2016	The pilot program of Three-River-Source National Park
12 October 2021	Three-River-Source Natural Park

). The implementation of measures such as cropland-to-grassland conversion, ecological migration, livestock reduction, rodent pest control, wildlife conservation, and black soil degradation management has contributed to enhancing the ecological conditions of the TRHR [12,13,14,17,30,66,67].

4.3. Limitations and Prospects

While we quantified the role of climate variability and anthropogenic interventions on FVC, several limitations exist that warrant further investigation, as detailed below.

The dimidiate pixel model was employed to estimate FVC, with RMSE (15.8%), bias (7.3%), R² (0.6), and KGE (0.68) indicating room for improvement. It is important to note that vegetation cover estimation inherently involves uncertainties arising from remote sensing data, model assumptions, and ground validation [55,68,69]. Firstly, systematic errors in remote sensing data, such as sensor noise, radiometric calibration errors, and atmospheric correction residuals, may introduce biases in pixel reflectance values. Although rigorous preprocessing workflows were implemented to mitigate these errors, complete elimination remains unachievable. Secondly, the model assumption may become problematic in heterogeneous landscapes with multiple vegetation types or transitional land covers, leading to potential overestimation/underestimation. While the model parameters were optimized through iterative trials, their selection remains subjective and inevitably affects FVC accuracy. Finally, ground-measured FVC has its own limitations, such as discrepancies between small sample plots and pixel scales, as well as phenological mismatches. And an uneven distribution of ground measurement points makes it difficult to cover high-altitude areas and complex terrains, limiting the representativeness of validation results. Future work could integrate multi-source data to address these challenges.

The Hurst index provides insights into long-term persistence but has inherent limitations: (1) It assumes historical trends will persist, which may not hold under abrupt environmental or anthropogenic changes. (2) There are no established metrics to quantify prediction uncertainties. (3) It does not account for external drivers (e.g., policy shifts). These constraints highlight the need for complementary approaches to enhance prediction robustness.

Numerous factors affect vegetation cover, and in this study, only climatic factors and human activities represented by residuals were selected. The TRHR has a rich topography with large elevation differences, and the influences of topographic elements, including altitude, gradient, and orientation, on the FVC merit further investigation. Additionally, climate change is the predominant driver influencing the FVC variation in the TRHR, but this paper considers the roles of only precipitation and temperature when applying partial correlation analysis. Other factors can also impact vegetation cover, such as wind speed, relative humidity, and solar radiation. Therefore, additional meteorological factors should be selected to explore their mechanisms of influence in future studies. It is worth noting that the meteorological dataset itself contains uncertainties, including errors introduced during actual observations at weather stations and the interpolation process to generate gridded data, both of which contribute to uncertainties in the results [68].

When applying residual analysis for driving factor attribution, error propagation may lead to error accumulation and resultant uncertainties. Concurrently, temperature, precipitation, and human interventions have varying degrees of lag effects on the enhancement of vegetation [70]. These lag effects inevitably influence the results and are worthy of further investigation. Moreover, residual analysis cannot fully disentangle the impacts of climate change and human activities, as there exist intricate interactions between the two [57,71,72]. Finally, this study adopted a 22-year scale from 2001 to 2022 to study the spatiotemporal dynamics in the growing season FVC in the TRHR and its driving factors without undertaking segmentation research. Multiscale temporal studies should also be a focus of future research.

5. Conclusions

In this study, by leveraging MOD13A3 NDVI data, we estimated the FVC by applying the dimidiate pixel model. We used the Theil–Sen median method and the M-K test to detect the FVC variations over space and time, and the Hurst index was used to assess the long-term consistency of the FVC. Subsequently, partial correlation and residual analyses were carried out to measure the impacts of climatic variability and anthropogenic interventions on the FVC. The key results are outlined below:

• The dimidiate pixel model employed in this paper yielded an FVC estimation accuracy of 84.2%, indicating that the model is viable and exhibits strong correlation and high precision.
• The FVC in the TRHR exhibited a volatile yet growing trend from 2001 to 2022, with a mean annual increase of 0.23% (p < 0.01). Spatially, an increasing trend in FVC was detected in 87.11% of the area, of which 50.72% demonstrated significant increases in FVC. Notably, the origin area of the Yellow River experienced the greatest increase in vegetation coverage.
• The FVC displayed a direct connection with both rainfall and temperature, with the influences of these two climatic elements being approximately equivalent. The southwestern zone of the YeRB, as well as the southern zone of the YaRB, is mostly influenced by temperature, whereas the northeastern zone of the YeRB and the southeastern portion of the YaRB are mostly influenced by precipitation. Notably, Dari County exhibited a significant response to both temperature and precipitation.
• Residual analysis revealed that climate variability was the predominant influence on the FVC variations within the study area. Despite the vastness and low population density of the region, which results in minimal human impact, vegetation restoration efforts in the origin of the Yellow River have achieved success.