Synergistic Drivers of Vegetation Dynamics in a Fragile High-Altitude Basin of the Tibetan Plateau Using General Regression Neural Network and Geographical Detector

Abstract

The internal response mechanism of vegetation change in fragile high-altitude ecosystems is pivotal for ecological stability. This study focuses on the Lhasa River Basin (LRB) on the Tibetan Plateau (TP), a typical high-altitude fragile ecosystem where vegetation dynamics are highly sensitive to climate change and human activities. Utilizing MODIS surface reflectance data (MOD09Q1), a general regression neural network (GRNN) was applied to create a 250 m resolution fractional vegetation cover (FVC) dataset from 2001 to 2022, whose accuracy was verified with field survey data. Through methods like the Theil–Sen Median trend analysis, Mann–Kendall significance test, Hurst exponent, and geographical detector, the collaborative mechanism of 14 driving factors was systematically explored. Key conclusions are as follows: (1) The FVC in the LRB evolved in stages, first decreasing and then increasing, with 46.71% of the basin area expected to show an improvement trend in the future. (2) Among natural factors, elevation (q = 0.480), annual mean potential evapotranspiration (q = 0.362), and annual mean temperature (q = 0.361) are the main determinants of FVC spatiotemporal variation. (3) In terms of human activities, land use type has the highest explanatory power (q = 0.365) for FVC. (4) The interaction of two factors on FVC is stronger than that of a single factor, with the elevation–land use interaction being the most significant (q = 0.558). These results deepen our understanding of the interactions among vegetation, climate, and humans in fragile high-altitude ecosystems and provide a scientific basis for formulating zoned restoration strategies on the TP.

1. Introduction

Vegetation, as a core component of terrestrial ecosystems, maintains the multi-interface coupling processes among the atmosphere, hydrosphere, pedosphere, and biosphere by regulating carbon–water cycles and energy balance [1,2]. It not only fulfills critical functions in sustaining ecological stability and driving material cycling and energy transfer but also serves as a core indicator of ecosystem structure and function [3,4]. Under the backdrop of global change, the response mechanisms of vegetation dynamics in ecologically fragile regions to climatic fluctuations and human activities have emerged as a key research focus—these regions face high risks of ecosystem degradation due to their climatic sensitivity and weak resistance to disturbances [5,6,7,8]. Systematically decoding the evolution patterns of vegetation dynamics and their driving mechanisms in fragile ecosystems, particularly clarifying their nonlinear response relationships to climatic variations and human activities [5,9], has become a fundamental scientific basis for predicting ecosystem evolution trends and formulating adaptive management strategies [10,11].

In the research progress exploring vegetation dynamics and influencing factors in ecologically fragile regions, the rapid advancement of remote sensing technology has provided strong support for acquiring large-scale, long-term, and high-resolution vegetation information [12,13,14]. Among these, FVC, a key parameter characterizing the spatial heterogeneity of surface vegetation, has been widely applied in assessing vegetation growth conditions and analyzing ecosystem functions [12,15]. Traditional FVC inversion has mostly relied on Normalized Difference Vegetation Index (NDVI) empirical models, yet their regional applicability and physical interpretability remain limited [16,17,18]. In recent years, machine learning methods represented by the GRNN have overcome the aforementioned limitations thanks to their nonlinear fitting ability and computational efficiency, making it possible to generate large-scale, high-precision, and spatio-temporally continuous FVC datasets [18,19,20,21,22,23]. For example, Jia et al. (2015) used a GRNN to fuse MODIS and Landsat data, producing a global FVC product with a validation accuracy significantly superior to traditional empirical models [18]. Xiao et al. (2016) derived FVC from GLASS (Global Land Surface Satellite) leaf area index products using GRNN assimilation, achieving a root mean square error (RMSE) with in situ measurements that were over 30% lower than the GEOV1 product [20]. Regarding the analysis of driving mechanisms, correlation analysis, multiple regression, and residual analysis are suitable when linear relationships exist between factors and drivers [24]. The Geodetector model, by quantifying the nonlinear effects of multi-factor interactions, offers a methodological innovation for revealing the coupling mechanisms of complex terrestrial systems [6,25,26].

As the world’s highest and largest plateau ecosystem, the TP’s alpine vegetation serves not only as a key carrier for regional hydrological regulation and carbon storage but also as a typical climate-sensitive fragile ecosystem [9,27]. Factors such as permafrost degradation, altered precipitation patterns, and intensified human activities are exerting complex impacts on vegetation growth by modifying soil moisture and nutrient cycles in this region [28,29]. Previous studies have demonstrated that climatic factors dominate the spatiotemporal differentiation of vegetation by regulating water and heat conditions [2,3,28]. For example, Shi et al. (2023) found that vegetation in northern TP exhibited a greening trend between 2000 and 2020, primarily attributed to rising temperatures and increased precipitation [27]. Temperature increases have prolonged the vegetation growing season to some extent, providing more adequate thermal conditions for plant growth [30]. However, in high-altitude areas of the TP, excessively high temperatures may trigger issues such as excessive water evaporation, which can instead inhibit vegetation growth [31]. Increased precipitation generally improves soil moisture, alleviating water limitations for vegetation and promoting growth and development [3], yet extreme precipitation events like floods or droughts can have adverse effects on plant growth [32]. Topographic factors reshape local water and heat regimes through altitude, slope, and aspect, driving the vertical zonation distribution of vegetation [33,34]. However, Wei et al.’s (2022) study on the NDVI in the TP revealed that the intensity of anthropogenic pressures far exceeds the impacts of climate change [29]. Additionally, human disturbances such as road network expansion and intensified urbanization are profoundly altering the distribution and growth of plateau vegetation [35,36,37].

Although previous studies on the influencing factors of vegetation dynamics in the TP have provided important references for understanding vegetation changes in ecologically fragile regions and achieved notable progress, most existing research focuses on the linear effects of single driving factors. There remains a lack of systematic understanding of the interactions between natural and anthropogenic factors and the spatiotemporal heterogeneous response mechanisms [2,3] as well as a significant shortage of high-resolution, long-term studies at the watershed scale [12,38]. These limitations collectively hinder the formulation of refined ecosystem management strategies for ecologically fragile areas at the watershed scale in the TP. The LRB, serving as a core area for population and economic activities on the TP, faces dual pressures of escalating land desertification and decline in ecosystem services in its middle and lower reaches [39]. Therefore, this study aims to utilize MODIS surface reflectance data and a GRNN to invert and generate a 250 m resolution FVC dataset for the LRB from 2001 to 2022; analyze the spatiotemporal dynamic characteristics of FVC; and employ the Geodetector model to quantify the independent and interactive effects of natural and anthropogenic factors, thereby elucidating the spatial heterogeneous patterns of multi-scale driving mechanisms. These efforts aim to provide a scientific basis for ecosystem protection, restoration, and sustainable development in the LRB and, more broadly, in ecologically fragile regions of the TP.

2. Materials and Methods

2.1. Research Area Overview

The study area, a typical high-altitude ecologically fragile region within the TP, is located in the LRB of Tibet (Figure 1), spanning coordinates 90°05′–93°20′E and 29°20′–31°15′N. The LRB covers 32,425 km² and exhibits a plateau temperate semi-arid climate influenced by the Southwest Monsoon, characterized by distinct dry and wet seasons. Precipitation is concentrated between May and September, with annual totals ranging from 257 to 699 mm. The region experiences intense solar radiation, a mean annual temperature of 6.3–9.1 °C, a large daily temperature range, an average elevation exceeding 4500 m, and a mean slope gradient of approximately 0.29%. The Lhasa River originates from Luoburula, flows through Mozhugongka and Dazi, and finally converges into the Yarlung Zangbo River in Qushui, Lhasa. Rapid economic growth and population expansion within the basin have posed unprecedented challenges to its ecological stability [40].

2.2. Data Sources

2.2.1. FVC Data

Field survey data: A total of 60 sample plots were surveyed within the study area from July to August 2022 (Figure 1). For each plot, a 250 m × 250 m area was selected, and three 1 m × 1 m subplots were randomly established within it. The FVC of each subplot was obtained via visual estimation, with the average value of the three subplots used to represent the actual FVC of the sample plot [41].

Remote sensing data: The remote sensing data were obtained from the MODIS Surface Reflectance product (MOD09Q1) provided by the National Aeronautics and Space Administration (NASA, https://ladsweb.modaps.eosdis.nasa.gov/search/, accessed on 8 July 2024), which has an 8-day temporal resolution and a 250 m spatial resolution, with 46 tiles annually. Data from the vegetation growing season (May–September) of 2001–2022 were selected to generate a 250 m resolution FVC dataset for the LRB using a GRNN. The Maximum Value Composite (MVC) method was applied to derive annual growing season data.

2.2.2. Driving Factors

The primary explanatory variables for the FVC were determined to be ten natural factors (elevation, slope, aspect, annual mean temperature (AMT), annual mean precipitation (AMP), annual mean potential evapotranspiration (AMPE), geomorphological type, vegetation type, soil type, and distance to the rivers) and four anthropogenic factors (distance to the roads, distance to the residences, annual mean nighttime light (AMNL), and land use type) (

Table 1. Driving factors.

Factor Type	Code	Factors	Unit	Original Resolution
Natural factors	X1	Elevation	m	30 m
	X2	Slope	°	30 m
	X3	Aspect	°	30 m
	X4	AMT	°C	1 km
	X₅	AMP	mm	1 km
	X₆	AMPE	mm	1 km
	X₇	Landform type	/	1 km
	X8	Vegetation type	/	1 km
	X9	Soil type	/	1 km
	X10	Distance to the rivers	km	250 m
Human factors	X11	Distance to the roads	km	250 m
	X12	Distance to the residences	km	250 m
	X13	AMNL	/	1 km
	X14	Land use type	/	30 m

). The elevation, slope, and aspect data were derived from 30 m spatial resolution DEM data. The AMT, AMP, and AMPE data came from the National Tibetan Plateau Science Data Center [42,43,44,45,46]. River, road, and residence datasets were sourced from OpenStreetMap, while nighttime light data, at a 1 km resolution, were obtained from the Harvard Dataverse platform [47]. The landform type data, vegetation type data, soil type data, and land use type were all sourced from the Resource and Environment Science Data Center of the Chinese Academy of Sciences. Distance to the rivers, roads, and residences to each pixel were calculated using the ArcGIS 10.8 Distance Tool.

This study employed the Jenks natural breaks method to objectively classify the 14 driving factors, with their geographical distributions illustrated in Figure 2. To ensure data consistency, all coordinate systems were projected to WGS_1984_UTM_Zone_48N, and the spatial resolution was resampled to 250 m × 250 m. A total of 8104 2 km × 2 km grid sampling points were generated using the ArcGIS Fishnet Tool, from which attribute tables containing FVC and driving factor values were extracted and constructed.

2.3. Methods

2.3.1. The Inversion Method of FVC

First, a GRNN was employed for inversion, and FVC was calculated based on vegetation canopy transmittance following the specific workflow described below.

The GRNN can be divided into 4 layers: the input layer, the hidden layer, the summation layer, and the output layer. The input layer distributes the input vector to each unit of the hidden layer. The hidden layer includes all training samples (I = 1, 2, …, n). After determining the input vector, the distance between the input vector and the training samples is computed and substituted into the probability density function. The summation layer is then split into two separate computational units, labeled A and B. The output layer calculates A divided by B to obtain the final prediction result.

The GRNN’s probability density function (kernel function) is calculated via the Gaussian function, with the output expression:

$$Y^{\mathrm{\prime }}\left(X\right)={\displaystyle \frac{{\sum }_{i=1}^n{Y}^{i}exp\left(-\frac{D_i^2}{{2\sigma }^2}\right)}{\sum _{i=1}^{n}exp\left(-\frac{D_i^2}{{2\sigma }^2}\right)}}$$

(1)

where $D_i^2$ = (X$-X^i$)^T(X$-X^i$), Xⁱ and Yⁱ (i = 1, 2, …, n) are the input and output of i-th sample, respectively; n is number of samples; X is the input vector; $Y^{\prime }$(X) is the output when the predicted input is X; and $\sigma$ is the corresponding parameter that controls the smoothness of the fitting result. Preprocessing operations such as atmospheric correction, identification of contaminated data, removal of contaminated data, and filling of missing data were carried out on the MODIS surface reflectance data [19]. Then, these data were used as input data, and the trained neural network was employed to invert and generate a long-term time series LAI data product with a resolution of 250 m.

Based on the LAI dataset, the FVC was calculated using the transmittance of the vegetation canopy, where the light passing through the canopy at its top was measured [20]:

$$P_{tr}\left(\phi \right)=e^{-\sqrt{a}\times k_c\left(\phi \right)\times \Omega \times LAI}$$

(2)

where $\phi$ is the solar zenith angle, a is the leaf absorptivity, Ω is the clumping index, LAI is the leaf area index, and $k_c\left(\phi \right)$ is the canopy extinction coefficient, which is calculated by the following formula:

$$k_c\left(\phi \right)={\displaystyle \frac{\sqrt{x^2+{tan}^2\left(\phi \right)}}{{x+1.774\times \left(x+1.182\right)}^{-0.733}}}$$

(3)

where $x$ is the ratio of the average projected area of the horizontal surface of the leaves to that of the vertical surface. According to the definition of FVC, when the vegetation cover is 1, subtract the canopy transmittance when $\phi$ is at 0 degrees, that is:

$$FVC=1-P_{tr}\left(0\right)$$

(4)

This method has been directly validated at over 28 sites worldwide [20]. Results indicate that this method outperforms mainstream products (such as GEOV1 and MOD15) in terms of accuracy, spatiotemporal continuity, and physical consistency, thus validating its reliability and applicability [18,19,20,22].

2.3.2. FVC Accuracy Verification

In this study, the confusion matrix and Kappa coefficient were used to validate the accuracy of remotely sensed inverted FVC data. Remotely sensed inverted FVC data close in time to field surveys were selected, and the corresponding remotely sensed inverted FVC values were extracted and compared with the measured values. Both remotely sensed inverted FVC values and measured values were classified into two categories according to a unified classification standard: FVC ≤ 0.5 and FVC > 0.5. A 2 × 2 confusion matrix (

Table 2. The confusion matrix.

		Remote Sensing
		FVC ≤ 0.5	FVC > 0.5
Observation	FVC ≤ 0.5	a	b
	FVC > 0.5	c	d

) was constructed, and the Kappa coefficient was calculated to evaluate the consistency of classification results. A paired-samples t-test was conducted using G*Power 3.1.9.7 software [48], with an a priori effect size (Cohen’s d) set at 0.5 and a significance level (α) of 0.05. The power analysis revealed that a minimum of 54 samples were required to achieve a statistical power of 0.95. In this study, 60 samples were used for validation, exceeding the calculated sample size requirement.

Here, a denotes the number of plots where both measured and remotely sensed inverted FVC values are ≤ 0.5; b is the number of plots with measured FVC ≤ 0.5 but remotely sensed inverted FVC > 0.5; c is the number of plots with measured FVC > 0.5 but remotely sensed inverted FVC ≤ 0.5; and d represents the number of plots where both measured and remotely sensed inverted FVC > 0.5 [49].

The Kappa coefficient’s calculation formula is:

$$K={\displaystyle \frac{N\sum _{i=1}^n{x}_{ii}-\sum _{i=1}^n\left(\sum _{j=1}^n{x}_{ij}·\sum _{j=1}^n{x}_{ji} \right)}{N^2-\sum _{i=1}^n\left(\sum _{j=1}^n{x}_{ij}·\sum _{j=1}^n{x}_{ji} \right)}}$$

(5)

where N is the total sample size; x_ii is the value on the diagonal; and x_ij and x_ji are the marginal sums of rows and columns, respectively. The range of the Kappa index is between −1 and 1. Among them, K = 1 indicates a high level of agreement, K = 0 means that the agreement is equivalent to random classification, and K < 0 indicates that the agreement is lower than random classification. Generally, K values exceeding 0.75 suggest strong consistency, those ranging from 0.40 to 0.75 denote moderate consistency, and values at or below 0.40 indicate low consistency.

2.3.3. TSM and M-K Trend Analysis

Theil–Sen median (TSM) trend analysis is a popular technique for analyzing the FVC evolution trend [50]. The Mann–Kendall (M-K) test is used to verify the statistical significance of the trend [51]. Figure 3 shows the process of TSM and M-K trend analysis.

2.3.4. Hurst Exponent

The Hurst exponent (H), a crucial parameter in time series analysis [52], can effectively capture the long-term memory effect and the trend persistence characteristics of a data series [53]. In the context of FVC time series, when H lies within the interval (0, 0.5), the FVC time series exhibits anti-persistence. When H equals 0.5, it adheres to the characteristics of Brownian motion. When H is in the range (0.5, 1), the series shows persistence, and notably, the greater the value of H, the stronger the persistence. This is how the computation method is shown:

(1)

Create a pixel time series (V_t), where t ranging from 1 to n.

(2)

Mean value calculation:

$${\mu }_V={\displaystyle \frac{1}{n}}{\sum }_{t=1}^n{V}_t$$

(6)

where $V_t$ is the mean value of the series.

(3)

Cumulative fluctuation series:

$$C(t)={\sum }_{i=1}^t\left(V_i-{\mu }_V\right), \mathit{t}=1, 2, \dots , \mathit{n}$$

(7)

(4)

Range calculation:

$$R_V={\mathrm{m}\mathrm{a}\mathrm{x}}_{1\le t\le n}C\left(t\right)-{\mathrm{m}\mathrm{i}\mathrm{n}}_{1\le t\le n}C\left(t\right)$$

(8)

(5)

Standard deviation calculation:

$$S_V=\sqrt{{\displaystyle \frac{1}{n-1}}{\sum }_{t=1}^n{\left(V_i-{\mu }_V\right)}^2}$$

(9)

(6)

Hurst exponent calculation:

$$\mathrm{l}\mathrm{n}\left({\displaystyle \frac{R_V}{S_V}}\right)=H·\mathrm{l}\mathrm{n}n+C$$

(10)

2.3.5. Geographical Detector

The Geodetector is a set of statistical methods used to detect spatial heterogeneity and dissect its driving factors.

(1)

Factor Detection

$$q=1-{\displaystyle \frac{{\sum }_{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}}$$

(11)

where q is a factor’s explanatory power for the spatial heterogeneity of FVC, with its value falling within the interval (0,1). The magnitude of the q value indicates the strength of the factor’s influence on the spatial differentiation of FVC. L represents the number of FVC or independent variable X classifications. Category h sample sizes are denoted by $N_h$ and the overall sample size by N. The total variance of FVC is denoted by ${\sigma }^2$, while the variance of the dependent variable FVC in category h is denoted by ${\sigma }_h^2$.

(2)

Interaction Detection

The interaction detector measures the explanatory power of the two driving factors’ interaction on FVC. The intersection of the spatial regions corresponding to the two factors is taken. By comparing the q-values of the two factors’ individual effects on the FVC with the q-value of their combined effect, the type and effect of the interaction can be determined (Figure 4).

(3)

Risk Detection

The risk detection module is used to determine whether there are significant differences in the mean values of FVC between two sub-regions, and a hypothesis verification system of the t-test is constructed for verification [54].

$$t={\displaystyle \frac{{\overline{X}}_1-{\overline{X}}_2}{\sqrt{\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}}}}$$

(12)

where ${\overline{X}}_1$ and ${\overline{X}}_2$ represent the mean values of FVC within the two sub-regions, respectively; $s_1^2$ and $s_2^2$ are the sample variances; and $n_1$ and $n_2$ are the sample sizes.

3. Results

3.1. Verification of FVC in the LRB

The validation results of the confusion matrix indicate that the Kappa coefficient is 0.78, suggesting a high degree of consistency between the remotely sensed FVC and the measured FVC. According to Figure 5, among the samples with measured FVC ≤ 0.5, nine samples were correctly classified by remote sensing, and there were no misclassified samples. This demonstrates that the remote-sensing classification has a high level of accuracy in identifying cases where FVC ≤ 0.5. Among the samples with measured FVC > 0.5, 47 samples were correctly classified by remote sensing, but 4 samples with measured FVC > 0.5 were misclassified as FVC ≤ 0.5. This indicates that there is some misjudgment in identifying cases where FVC > 0.5, yet the overall number of correct classifications is relatively large.

3.2. Spatiotemporal Variation of FVC in the LRB

Based on the analysis of the FVC time series data in the LRB from 2001 to 2022 (Figure 6), the FVC values during the study period ranged from 0.369 to 0.447, with an average of 0.416, exhibiting pronounced interannual fluctuation characteristics. Specifically, the lowest FVC value (0.369) occurred in 2007, while the peak value (0.447) was observed in 2020. Linear trend analysis revealed a slight upward trend in interannual FVC changes, with a growth rate of 0.5%/10a, though this trend did not pass the significance test (p > 0.05), suggesting that climatic fluctuations or human activities may have obscured the long-term trend signal. To uncover the underlying patterns of change, the original data were further smoothed using a 5-year moving average method. The results indicated that FVC evolution showed a trend of first decreasing and then increasing.

The change trends of the FVC in the LRB showed significant differences in space (see Figure 7a). Among them, 15.72% of the area had stable FVC, mostly found in the basin’s northwest high-elevation regions; 4.34% of the area had significant FVC degradation, concentrated around towns with frequent human activities; 8.83% of the area had significant FVC improvement, which was in the valley region of the lower Lhasa River reaches (Figure 7b). In addition, 29.87% of the area had nonsignificant FVC degradation, and 41.24% of the area had nonsignificant FVC improvement. Overall, about 65.79% of the LRB was in a stable or improving state, with a favorable FVC trend.

3.3. Sustainability of FVC in the LRB

This study utilized the Hurst exponent to determine the sustainability of future vegetation changes in the LRB. Figure 8a shows that Hurst exponents ranged from 0.20 to 0.77, with an average of 0.54, and approximately 70.29% of the region exhibited Hurst exponents greater than 0.5. This indicates that vegetation change trends in the LRB are likely to persist. Areas with Hurst exponents less than 0.5, accounting for 29.71% of the basin, were discretely distributed across the entire LRB. To predict the trend of FVC changes in the study area, the Hurst exponent was coupled with the slope β from the Theil–Sen Median trend analysis (Figure 8b), classifying future changes into six types [55,56,57] (

Table 3. Classification of future sustainability changes of FVC in the LRB.

Hurst	β	Future Change	Percentage
>0.5	≥0.0005	Continued improvement	36.91%
<0.5	≥0.0005	Future degradation	13.26%
<0.5	β ≤ −0.0005	Future improvement	9.80%
>0.5	β ≤ −0.0005	Continued degradation	24.51%
>0.5	−0.0005 < β < 0.0005	Continued stability	8.88%
<0.5	−0.0005 < β < 0.0005	Random change	6.64%

). Approximately 46.71% of the LRB is projected to exhibit a future improvement trend, concentrated in Linzhou, Maizhokunggar, and the upper reaches of the Lhasa River valley (Figure 8b). Regions with a projected vegetation degradation trend accounted for 37.78%, primarily distributed in the low-altitude areas of the northwestern basin. Additionally, areas with continued stability and random changes accounted for 8.88% and 6.64%, respectively, mainly concentrated in the high-altitude regions of the northwestern basin. Sensitivity analysis revealed that the length of the input data sequence influenced the proportion of areas with vegetation improvement trends (

Table 4. Hurst exponents and fluctuations in improvement proportion under different data sequence lengths.

Time Period	Data Length	Areas with Future Improvement Trend	Mean Hurst Exponent
2001–2022	22 years	46.71%	0.54
2007–2022	16 years	50.76%	0.52
2013–2022	10 years	47.49%	0.49

). Longer data sequences can more comprehensively reflect changes in the research object over extended periods, reducing interference from short-term fluctuations and thus making the results more representative of long-term stable trends.

3.4. Driving Factors of FVC in the LRB

3.4.1. Factor Detection Analysis

Geodetector analysis was conducted to evaluate the impacts of 14 driving factors on the FVC in the LRB (Figure 9). The results revealed that all factors, except for the AMNL, exhibited extremely significant effects on the FVC in the LRB (p < 0.001). Among them, elevation had the highest q value, with an explanatory power of 0.480, which is the most crucial factor affecting FVC in the LRB. The explanatory powers of land use type, AMPE, and AMT were close, at 0.365, 0.362, and 0.361, respectively. Soil type, vegetation type, distance to roads, and AMP were also key factors for vegetation change in the LRB, with descriptive powers of 0.338, 0.249, 0.115, and 0.087, respectively. Although their effects were minimal, the distance to rivers, landform type, slope, aspect, distance to residences, and AMNL all affected FVC.

3.4.2. Interaction Detection Analysis

Interaction detection revealed the types of interactions between two driving factors of FVC. The impact of the two-factor interaction on FVC was significantly greater than that of any single factor (Figure 10). Among the 91 sets of interactions between factors, 37 sets showed two-factor enhancement, and 54 sets exhibited nonlinear increases. Specifically, the interaction between elevation and all other factors had the most significant impact on FVC, with q-values ranging from 0.484 to 0.558. When land use type, AMPE, AMT, and soil type interacted with other factors, the ranges of their q-values were 0.373–0.558, 0.364–0.526, 0.366–0.506, and 0.343–0.520, respectively, indicating the necessity of these four components for vegetation growth (Figure 10). Among them, the interaction between elevation and land use type had the highest interaction effect among all pairwise factor interactions (q = 0.558).

3.4.3. Risk Detection

The risk detector identified the optimal ranges for factors influencing vegetation growth (Figure 11), with significant differences in mean FVC values across factors. The FVC exhibited a significant gradient variation with elevation, first increasing and then decreasing as elevation changed. Within the 4540–4871 m elevation range, the mean FVC reached a peak of 0.60, representing the most ideal elevation for vegetation growth. When the elevation exceeded 4871 m, vegetation growth was inhibited, reflecting the limiting effect of environmental conditions on vegetation growth. Forest land use type had the highest FVC at 0.62, while water bodies showed the lowest FVC at only 0.17. As AMPE increased, mean FVC first rose and then declined, peaking at 0.61 when AMPE ranged from 673 to 816 mm. The response of FVC to AMT also followed a unimodal trend (Figure 11). When AMT was between −2.54 °C and 0.57 °C, FVC reached 0.58, indicating this temperature range is optimal for vegetation growth. Soil types significantly differed in their impacts on LRB’s FVC: vegetation thrived best in meadow soil and frigid calcic soil (Figure 11), with an annual mean FVC as high as 0.59, while frigid frozen soil showed the lowest FVC. Shrub-dominated areas had the highest vegetation coverage (FVC = 0.57) among vegetation types. As the distance to the roads increased, the average FVC value decreased. Within the region of 0–1.54 km from the road, the average FVC value was the highest at 0.51. The vegetation was not significantly affected by fluctuations in precipitation. AMP had a weaker effect on vegetation growth: mean FVC reached 0.48 when AMP was 345.3–393.7 mm, dropping to the lowest 0.26 when AMP fell below 345.3 mm. Mean FVC declined with increasing distance from rivers. Plains exhibited the best vegetation growth (FVC = 0.47), while hilly areas had the lowest (0.35). As slope gradient increased, mean FVC decreased: slopes of 0–7.74° had the highest mean FVC (0.47). North-facing aspects showed the highest mean FVC (0.46), while southeast-facing aspects had the lowest (0.37), indicating shaded slopes are more favorable for vegetation than sunlit slopes. Distance to residences had a minimal impact, with the mean FVC consistently above 0.35. Vegetation grew best in areas with the lowest AMNL.

Table 5. Appropriate ranges or types of factors.

Factor Type	Factors	FVC Suitable Range or Types	FVC
Natural factors	Elevation	4540–4871 m	0.60
	Slope	0–7.74°	0.47
	Aspect	North	0.46
	AMT	−2.54–0.57 °C	0.58
	AMP	345.3–393.7 mm	0.48
	AMPE	673–816 mm	0.61
	Landform type	Plain	0.47
	Vegetation type	Scrub	0.57
	Soil type	Meadow soil, cold calcic soil	0.59
	Distance to the rivers	0–2.03 km	0.48
Human factors	Distance to the roads	0–1.54 km	0.51
	Distance to the residences	7.98–15.1 km	0.43
	AMNL	0–3	0.42
	Land use type	Forest	0.62

lists the factors associated with optimal vegetation growth conditions.

4. Discussion

4.1. Spatiotemporal Variation of FVC

As a core indicator of ecosystem structure and function, FVC plays a critical role in maintaining ecological stability, preventing soil erosion, and mitigating climate change by regulating processes such as surface albedo, evapotranspiration, and carbon sink capacity [1,2]. Previous studies have shown that, as a sensitive and vulnerable region to global warming, the TP exhibits a significant increasing trend in its vegetation cover [31,58]. For example, the NDVI across the TP increased at a rate of 0.5%/10a from 1982 to 2020 [31], with a 2.7%/10a increase from 2000 to 2020 [57]. The Enhanced Vegetation Index (EVI) during the growing season across the TP showed an average increase of 1%/10a from 2000 to 2020 [27], and the Tibet section of the Lancang River Basin experienced a significant 0.4%/10a NDVI increase from 1998–2020 [59]. Similarly, FVC in the Minjiang River Basin of eastern TP trended upward from 2000–2020 [60]. Consistent with these findings, our study revealed a weak positive correlation in overall LRB FVC growth during 2001–2022 (β = 0.5%/10a, p > 0.05), with 50.07% of the basin area showing vegetation improvement higher than the proportion of degraded areas (34.21%) (Figure 7b). It is worth noting that the FVC in the basin increased significantly after 2009, which is highly consistent with the implementation period of the “Tibet Ecological Security Barrier Protection and Construction Plan (2008–2030)” [61], suggesting that the systematic promotion of ecological projects has a positive driving influence on vegetation restoration.

Spatial differentiation characteristics (Figure 7a) revealed the highest FVC stability in the northwestern Nyainqêntanglha Mountains of the basin (elevation > 5500 m) where vegetation communities dominated by alpine meadows and tundra exhibited resilience comparable to the Yangtze River Source region due to low-temperature constraints and minimal human disturbance [62]. A “degradation-improvement” spatial game feature can be seen in the lower reaches of the river valley: the FVC in the built-up areas of Lhasa City and around traffic trunk lines has significantly degraded due to the expansion of construction land, while a series of ecological protection and ecological projects have increased the FVC in the LRB [40]. This spatial heterogeneity confirms the regional characteristics of the coexistence of “stress–restoration” dual mechanisms, indicating that scientifically planned ecological projects can effectively offset the negative impacts of urbanization and provide a practical paradigm for the coordination of human–land relationships in alpine river valleys [63,64,65]. Through the combined analysis of Hurst exponents and TSM trend analysis, this study revealed that approximately 46.71% of the LRB is projected to exhibit a future vegetation improvement trend, while 37.78% is expected to undergo degradation. Although the Hurst exponent is widely used to predict vegetation change trends, it cannot clarify the specific duration of the trend [26,57,66].

4.2. Driving Forces of Vegetation Change

Vegetation dynamics in the TP region are jointly regulated by climate change and human activities [29,57]. Moreover, its spatial heterogeneity mainly originates from natural factors rather than social and economic factors [60,67,68,69]. Findings from this study align with these conclusions: elevation, land use type, AMPE, and AMT collectively dominate variations in FVC in LRB, with the natural factor of elevation exhibiting the highest explanatory power.

4.2.1. Natural Factors

This study reveals that among the 10 natural factors, elevation, AMPE, and AMT are the most influential factors affecting FVC in the LRB, with the q values being 0.480, 0.362, and 0.361, respectively. Elevation determines the vegetation type and its distribution pattern by causing vertical differentiation of hydrothermal conditions [25,70,71]. To further explore the mechanisms by which elevation influences vegetation change in the LRB, this study categorized the elevation of the LRB into three gradients and analyzed the driving factors at each gradient (Figure 12). Changes in LRB vegetation are predominantly affected by the AMPE in the low-elevation area (<4500 m). In this area, the heat conditions are sufficient, and vegetation growth is mainly restricted by the availability of soil moisture [33]. An increase in temperature intensifies evapotranspiration, which may accelerate the depletion of soil moisture, resulting in a water deficit in vegetation and thus inhibiting photosynthesis efficiency and biomass accumulation [71]. The primary determinant of vegetation change in the LRB in the medium-elevation (4500–5500 m) and high-elevation (>5500 m) regions is land use type, underscoring the significance of ecological projects and land management models (converting grazing land back to grassland) [72,73]. Overall, the elevation gradient modifies the action hierarchy of every driving factor by controlling the intensity of human activity and the hydrothermal combination.

The study also found that among natural factors, AMPE (q = 0.362) and AMT (q = 0.361) exhibited similar explanatory power for vegetation spatial differentiation, closely related to the high sensitivity of plant phenology to temperature changes in high-elevation regions [74]. Under the backdrop of climate warming, vegetation response mechanisms exhibit threshold effects: moderate temperature increases can prolong the vegetation growing season and enhance productivity [25]. However, when temperature rises exceed the physiological tolerance of vegetation, soil water stress induced by increased evapotranspiration may offset the photosynthetic benefits of warming [71]. The influence of soil type is primarily reflected in meadow soil and frigid calcic soil, which exhibited the highest FVC values closely linked to their high organic matter content, cation exchange capacity, and water-holding capacity [75]. The analysis of vegetation types showed that the shrub-covered area has the highest FVC value, mainly distributed at an elevation of 4540–4871 m. This area has the optimal hydrothermal combination conditions [57], and the adaptability of shrubs to the alpine environment has been verified in multiple dimensions [57,76]. Notably, although precipitation is widely regarded as a dominant driver of vegetation change on the TP [57,69], this study showed that its explanatory power for vegetation dynamics was only 0.087. This may be attributed to the hydrological regulation mechanism of deep groundwater recharge to surface ecosystems in the LRB, which mitigates the impact of precipitation on vegetation changes [77]. For example, vegetation in arid regions directly utilizes groundwater through deep roots, with its productivity strongly correlated with groundwater depth rather than precipitation [78,79]. Deep groundwater forms a stable water source by slowly recharging surface ecosystems, thereby mitigating the impact of short-term precipitation fluctuations on vegetation [80]. Vegetation actively acquires groundwater by adjusting root depth, reducing its dependence on precipitation variability [81].

4.2.2. Human Factors

Based on the results of the geographical detector analysis, land use type exhibited the most significant effect among anthropogenic factors on vegetation changes, with a q-value of 0.365, suggesting the ecological effects of human intervention [82]. This study quantitatively analyzed the process of land use transition to deeply explore its driving mechanisms for FVC changes (Figure 13). From 2005 to 2020, the grassland area decreased by 11.20%, while the areas of cultivated land, forest, water area, construction land, and unused land increased by 0.64%, 6.67%, 1.24%, 0.36%, and 2.22%, respectively (Figure 13a). The reduction of grassland area was mainly converted into unused land (9.65%) and forest (8.56%), which led to a decrease of 0.346 and an increase of 0.126 in FVC, respectively (Figure 13b). Grassland degradation and irrational utilization may have contributed to vegetation degradation to some extent [83]. In contrast, ecological measures such as afforestation have significantly promoted vegetation restoration [84]. Urbanization in the LRB has converted cultivated land, forest, and grassland into construction land, resulting in a decrease of 0.397, 0.438, and 0.309, respectively, in the annual average FVC of the LRB (Figure 13b). The direct impacts of urbanization on vegetation are primarily reflected in the reduction of vegetation cover caused by urban expansion [65]. Previous studies have noted that a series of ecological projects were implemented in the LRB to promote vegetation restoration (e.g., afforestation, mine restoration, riparian ecological restoration, and grassland ecological restoration), which offset the negative impacts of urbanization on basin ecology [40,85]. This study also confirmed this (Figure 13b). Furthermore, this study revealed that the FVC value reaches a peak (0.51) within the 0–1.5 km buffer zone from the road, suggesting that moderate human activities may have a positive ecological effect. This phenomenon may be due to the improvement of the micro-habitat brought by the road corridor [86] and the associated effect of the artificial irrigation system.

4.2.3. Impact of Factor Interactions on FVC

The interaction detector results showed that the interactions between the two driving factors in the LRB significantly enhanced the explanatory power for vegetation changes, indicating that interactions among multiple factors impose complex impacts on vegetation dynamics [87,88]. This also confirms the nonlinear influence mechanism of multi-factor synergies on vegetation changes [57,69,87]. The LRB’s vegetation change is most significantly influenced by the interaction between elevation and land use type, with an explanatory power of 0.558, further demonstrating that both natural and human influences affect vegetation change. This conclusion aligns with Mukhtar et al.’s research findings on the Himalayas’ southern slope [34].

4.3. Limitations and Prospects

This study systematically revealed the driving mechanism of vegetation change in alpine regions by integrating the vegetation cover data inverted by the GRNN and the geographical detector method. Compared with traditional linear statistical models, this method has unique advantages in measuring spatial differentiation and analyzing the interactions of explanatory variables [89]. However, this study still has the following limitations: First, during the FVC validation stage, due to transportation constraints and environmental limitations on the TP, validation points were primarily distributed in mid-lower reaches below 4500 m altitude, which may introduce biases in validation accuracy for high-altitude regions. Second, while the Hurst exponent can predict the persistence of vegetation change trends, it cannot determine the specific duration of these trends [26,57,66]. Third, the q-value of the geographical detector is sensitive to data aggregation scales, potentially affecting the absolute values of factor influence while maintaining stable relative rankings [90]. Additionally, the driving factor system does not quantify the impacts of soil moisture, CO₂ fertilization, and ecological projects on vegetation, which may weaken the completeness of causal analysis. Furthermore, the insufficient integration of high-resolution urban vegetation mapping data (e.g., Sentinel-2, WorldView-3) may limit the in-depth analysis of interactive mechanisms between small-scale driving factors and FVC [91,92]. Finally, existing driving factor analysis methods still have room for improvement in capturing nonlinear relationships, necessitating the integration of techniques such as machine learning and geographically weighted regression to construct hybrid models [87,93,94], thereby enhancing the comprehensiveness and practical utility of the research.

Future work is recommended to be carried out in the following aspects: Obtain higher-resolution and multi-source integrated surface reflectance data to more accurately invert vegetation cover and improve the ability to capture the details of vegetation change. At the same time, increase the sample plot survey data to improve the accuracy of the verified FVC dataset. Consider driving factors more comprehensively, and introduce advanced models to deeply analyze the interaction mechanism. Expand the research scope to provide universal experience for vegetation protection in ecologically vulnerable areas, and support the construction of the ecological barrier function of the TP.

5. Conclusions

This study employed the GRNN and Geodetector to comprehensively analyze vegetation dynamics in the ecologically fragile high-altitude LRB. The findings offer novel perspectives on the complex interactions between natural and human driving forces in this vulnerable region. The principal conclusions are as follows: (1) The LRB’s FVC first declined and then rose between 2001 and 2022. In the study area, the majority of vegetation changes are sustainable, and 46.71% of the area is expected to exhibit an improving trend in the future. (2) Among natural factors, elevation (q = 0.480), AMPE (q = 0.362), and AMT (q = 0.361) form a dominant driving framework. The three factors shape the vegetation distribution pattern by regulating hydrothermal redistribution and nutrient supply. (3) Among human factors, land use type (0.365) is the most important, confirming that ecological improvement projects have offset the negative impacts of urbanization. (4) The two-factor interaction generally shows a non-linear enhancement feature, and the interaction between elevation and land use type is the strongest (q = 0.558), revealing that vegetation changes are influenced by both natural and human factors.

This research advances understanding of vegetation–climate–human dynamics in data-scarce alpine regions and provides a methodological framework for quantifying synergistic drivers using GRNN and Geodetector. The results highlight the necessity of zonal policies: optimizing water management in low-altitude areas, enhancing land use planning in mid-elevation zones, and prioritizing cryosphere vegetation protection in high-altitude regions. These insights are critical for maintaining ecological security on the TP and inform adaptive governance in fragile high-altitude ecosystems globally.