The Interaction Between Vegetation Change and Land–Atmosphere Heat Exchange on the Tibetan Plateau

Abstract

Vegetation–heat flux feedbacks have a great influence on ecosystems, but the interaction between them is still unclear. This is particularly critical in ecologically fragile areas, where plant growth is especially sensitive to land–atmosphere interactions that help plants withstand environmental pressures. To the causal relationship between vegetation and heat flux under different topographies on the Tibetan Plateau, we improved the Granger causality model to handle nonstationary scenarios, enabling us to uncover previously unknown interaction patterns between unstable vegetation change and heat fluxes. Further sensitivity analysis was performed to assess the strength of causal influences. The results showed that the sensible heat (SH) and latent heat (LH) fluxes were increasing at rates of 0.28 W·m⁻²·decade⁻¹ and 0.105 W·m⁻²·decade⁻¹, respectively. The interaction between them on vegetation change depends on terrains, at low elevations below 3000 m and high elevations of 5000–6000 m, SH and LH jointly regulate vegetation growth of shady and gentle to moderate slopes, predominantly involving dense grasslands, but the influence of SH is stronger. While at middle elevations of 3000–5000 m and on steep slopes, LH and vegetation of all types interact to form an intensive local energy cycle. Conversely, vegetation change also influences heat flux. Below 6000 m (excluding the 2000–3000 m), vegetation only regulates LH, and this influence appears largely independent of terrain, contributing to energy redistribution and water cycle maintenance in these regions. These interactions suggest that vegetation plays a central role in shaping energy distribution on the plateau, maintaining the water cycle, and regulating climate in alpine regions by regulating heat flux.

1. Introduction

Vegetation, as a vital component of terrestrial ecosystems [1], is highly sensitive to climate change [2]. It plays a critical role in the exchange of matter and energy among the pedosphere, hydrosphere, and atmosphere, serving as a key element in surface energy balance, the hydrological cycle, and biogeochemical processes [3,4]. The current trend of global warming exerts considerable impacts on vegetation [5,6] and alters its role in matter and energy exchanges with the atmosphere.

The Tibetan Plateau (TP), where annual mean temperatures remain below 0 °C in most regions [7], is often referred to as the “Third Pole”. Due to its extensive glacial and snow cover and its function as the headwaters of many major Asian rivers, it is also known as the “Asian Water Tower”. Since the 1950s, the TP has exhibited a pronounced warming and wetting trend. The rate of warming ranges between 0.16 and 0.67 °C per decade, approximately double the global average [8], with projections indicating a continuation of this trend throughout the 21st century. Precipitation has also increased across the region, though with marked spatial variability (3.99–16.84 mm/a) [9].

Due to its fragile ecosystems, the TP is one of the most climate-sensitive regions [10]. Previous studies have identified precipitation, temperature, and human activities as the dominant drivers of vegetation dynamics in this area [9,11,12,13,14,15]. Responses to these drivers vary across different subregions. Increases in precipitation can reduce temperature and solar radiation [16], promote soil erosion [17], and negatively affect vegetation in certain zones [5,18]. However, most studies suggest a positive correlation between precipitation and vegetation growth, especially where water is not a limiting factor [19,20,21]. Rising temperatures have also been associated with higher vegetation coverage [5], longer growing seasons [22], and increased net primary productivity (NPP), thereby enhancing the region’s carbon sequestration potential [23]. Nevertheless, excessive warming may suppress vegetation growth by intensifying respiratory losses [24]. Human activities also exert both negative and positive influences on vegetation dynamics [23,25,26].

Vegetation influences surface energy balance by adjusting net radiation [27,28,29]. Its coverage, type, and structure alter surface albedo and aerodynamic roughness, affecting sensible (SH) and latent heat (LH) fluxes [30,31]. For example, higher vegetation density reduces albedo and increases net radiation, enhancing warming, while greater roughness promotes turbulent heat dissipation and can lower daytime temperatures [32]. Through evapotranspiration, vegetation increases LH and decreases SH, contributing to evaporative cooling [32,33]. A unit increase in Leaf Area Index [34] has been shown to raise LH by approximately 3.7 W m⁻² and reduce SH by about 3.3 W m⁻² globally [35]. This shift lowers the Bowen ratio, meaning that a greater fraction of available energy is allocated to latent (moisture-related) processes as vegetation increases [35,36]. Moreover, accounting for the spatial clumping of vegetation reveals an increase in ground SH but a decrease in vegetation SH, resulting in a net reduction in total SH and a slight rise in LH, especially in tropical regions [37]. Seasonal variations in vegetation structure, particularly in forests, can also cause substantial changes in both SH and LH [36]. In urban areas, higher vegetation cover is associated with increased LH and reduced SH, helping to alleviate urban heat through evapotranspiration [38]. In agricultural landscapes, vegetation growth during the growing season typically elevates LH and lowers SH, whereas bare or dormant fields show the opposite trend [39,40]. Wetlands and forests often exhibit lower SHF and higher LHF than shrublands or grasslands, especially during the growing season [41]. Under extreme conditions such as heatwaves, forests tend to maintain LH while increasing SH, whereas grasslands and croplands may shift more energy toward SH under water stress [41]. The magnitude and direction of these effects vary with vegetation type, density, and environmental setting.

Time-series analysis is widely used to study vegetation–climate interactions. A core assumption is stationarity, i.e., constant statistical properties over time [42]. However, under ongoing climate change and anthropogenic disturbances, this assumption often fails. Nonstationary data with trends and cycles can lead to biased and unreliable conclusions if untreated [43].

Correlation analysis reflects overall associations but fails to determine causality, especially under nonstationary conditions. To address this, [44] proposed the Granger causality test, which extended to nonlinear and multivariate systems [45,46]. However, most methods lack robustness for nonstationary data.

As a hotspot of land–atmosphere interactions [47,48,49], the TP presents challenges in identifying nonlinear, nonstationary feedback between vegetation and heat fluxes. While meteorological factors such as temperature and radiation have been widely studied, the role of vegetation in regulating surface heat fluxes remains unclear. Building on prior work [47,50], this study explores three key questions: (1) Do the heat fluxes (sensible heat flux and latent heat flux) have different effects on vegetation across terrains? (2) Can vegetation, in turn, regulate sensible heat flux and latent heat flux? (3) How can these interactions be quantitatively assessed under nonlinear and nonstationary conditions? To answer these, we propose an improved Granger causality framework capable of handling nonlinear, nonstationary data. The model is applied to examine vegetation–heat flux interactions across various topographic zones on the TP. A sensitivity analysis is further conducted to quantify the strength of causal relationships. The results aim to elucidate the feedback between vegetation dynamics and energy exchange processes, providing scientific insights into ecological conservation, thermal regulation, and climate adaptation strategies in the region.

2. Materials and Methods

2.1. Study Area

The Tibetan Plateau (TP) is located in the western Sichuan Basin, between 23°N–44°N and 66°E–106°E (Figure 1a). Most of the TP is distributed in China, including many alpine canyons, rivers, and lakes. The elevation of TP at the highest point is more than 8800 m (Figure 1b), while the average elevation is about 4000 m [51]. The TP, commonly known as the ‘Third Pole’, is the highest and most extensive plateau in the world (Figure 1d). Due to its high terrain and sensitivity to climate forcing, the TP is widely considered to be the ‘amplifier’ and ‘starter’ of global climate change [52]. The natural vegetation of the TP is dominated by alpine grasslands and meadows, accounting for about 60% of the total area of the TP [1].

Due to the influence of the monsoon, the southeast of the plateau is mainly covered by mountain forests, and the vegetation type is dominated by mesophytic forests (Figure 1c). The northwestern hinterland of the TP is dominated by alpine grassland, alpine desert, mountain grassland, and mountain desert, which belong to the drought-tolerant type. On the southern side of the TP, vegetation transitions with increasing altitude, starting from tropical rainforests or evergreen broad-leaved forests at lower altitudes, followed by coniferous mixed forests, dark coniferous forests, shrubs, alpine meadows, and eventually reaching permanent snow and ice fields above the snow line [53]. See Figure 1 for the geographical background of the TP.

2.2. Data Sources

2.2.1. Enhanced Vegetation Index Data

The Enhanced Vegetation Index (EVI) is an advanced vegetation index used in remote sensing analysis, based on the traditional Normalized Difference Vegetation Index (NDVI). EVI integrates data from red, blue, and near-infrared bands, offering a more accurate reflection of surface vegetation coverage and growth status. It outperforms NDVI in vegetation monitoring and ecosystem analysis, in studies demanding precise and sensitive vegetation information. This study utilized EVI data from the Terra MOD13A3 version 061 monthly-scale product, provided by NASA’s Land Processes Distributed Data Archive (LPDDA). The dataset has a spatial resolution of 1 km and was processed using Python 3.8 to generate annual data from 2000 to 2022 (https://ladsweb.modaps.eosdis.nasa.gov, accessed on 20 August 2025). EVI values range from −1 to 1. This study excluded values below 0, as they indicate no vegetation coverage, retaining only those with EVI ≥ 0. Higher EVI values refer to more extensive, healthier, and denser vegetation coverage. Figure 2a illustrates the distribution of the multi-year average EVI on the TP from 2000 to 2022. Figure 2b shows the overall trend of vegetation greening on the TP.

2.2.2. Heat Fluxes Data

In this study, ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF) were used, focusing on sensible heat (SH) and latent heat flux (LH) at a spatial resolution of 0.1° × 0.1° (11.1 km). These data were employed to examine vegetation change sensitivity to heat fluxes on the TP. The EVI data for the period 2000 to 2022 were resampled using bilinear interpolation to align with the heat fluxes’ spatial resolution. Heat flux measures the amount of heat transferred between the Earth’s surface and the atmosphere due to turbulent air motion. According to the ERA5 data specifications, positive values represent downward fluxes and negative values represent upward fluxes, measured in joules per square meter (J/m²). For the convenience of later analysis, this study designates upward fluxes as positive and downward fluxes as negative, converting the units to watts per square meter (W/m²). Figure 3 illustrates the spatial distribution of mean SH and LH on the TP from 2000 to 2022.

2.2.3. Digital Elevation Model Data

The Digital Elevation Model (DEM) data were obtained from SRTMDEM provided by Geospatial Data Cloud (https://www.gscloud.cn) with a spatial resolution of 90 m. The data were resampled to match the heat flux resolution, and the slope and aspect of the TP were calculated using ArcGIS 10.2 software. In this study, the elevation, slope, and aspect were divided into 9 intervals for analysis and discussion. This division results in multiple potential combinations of these categories, but only the combinations present in the study area were analyzed. The classification results are shown in

Table 1. The classification intervals for elevation, slope, and aspect.

Elevation (m)	Slope (°)	Aspect
0–1000	0–10	Flat
1000–2000	10–20	Northeast
2000–3000	20–30	East
3000–4000	30–40	Southeast
4000–5000	40–50	South
5000–6000	50–60	Southwest
6000–7000	60–70	West
7000–8000	70–80	Northwest
8000–9000	80–90	North

.

2.3. Research Methodologies

To improve the accuracy of causality detection, this study enhanced the traditional Granger causality test by incorporating the principle of Ensemble Empirical Mode Decomposition (EEMD). The improved method was then employed to evaluate whether SH and LH function as causal drivers of variations in the EVI across different topographic regions of the TP. Upon identifying statistically significant causal relationships between SH, LH, and EVI, a multiple linear regression model was constructed to further quantify the relative sensitivity of EVI to SH and LH.

2.3.1. EEMD–Granger Causality Test

To determine whether two variables have a lead-lag relationship over time and whether this relationship is unidirectional or bidirectional, [44] proposed a method to test the causal relationship between variables, known as the Granger causality test. It is crucial to acknowledge that this methodology requires the time series of each variable to be stationary, as nonstationary data can yield spurious or fake regression, leading to a pseudo-causal relationship. Therefore, to accurately identify the causal relationships among SH, LH, and EVI, it is necessary to find a method that transforms nonstationary time series into stationary ones.

EEMD is a technique capable of decomposing nonlinear and nonstationary time series, and can effectively filter out noise to stabilize the original nonstationary data [54,55,56]. In light of the aforementioned considerations, this study proposed a coupling method for nonstationary time series, namely, an EEMD–Granger causality test, for application in the field of hydrometeorology, to handle or analyze nonlinear and nonstationary time-series data. By repeatedly adding Gaussian white noise, the original nonstationary time-series data are filtered out and converted into a stationary time series by this method, enabling more accurate identification of whether unidirectional or bidirectional causal relationships exist between SH and EVI or LH and EVI. The steps of this method are as follows:

• Gaussian white noise with a standard deviation equal to 20% of the original data’s standard deviation is added to the original time-series data of SH, LH, and EVI. The length of the white noise matches the original time series.
• The noise-augmented time series is treated as a whole, and all local extrema (both maxima and minima) are identified. A cubic spline interpolation is then applied to connect all local maxima to form an upper envelope, and all local minima to form a lower envelope.
• The mean of the upper and lower envelopes is calculated at each corresponding time point to obtain a local mean line. This local mean is subtracted from the noise-augmented time series to derive the first Intrinsic Mode Function (IMF) component and the residual.
• Steps 1 through 3 are repeated 100 times in this study, with a new Gaussian white noise sequence of the same standard deviation used for each iteration. The IMFs obtained from all iterations are averaged to reduce uncertainties caused by the added noise, yielding stable IMF components for SH, LH, and EVI.
• The lowest-frequency IMF component from the stable IMFs obtained in step 4 is selected. The selected IMF time series is subjected to a Granger causality test to evaluate the Granger causal relationships among SH, LH, and EVI.

EEMD is an improved method of Empirical Mode Decomposition (EMD) [55]. EEMD introduces white noise into the original time series before performing EMD decomposition, effectively addressing the mode mixing problem encountered in practical applications of EMD. By predefining the standard deviation of the white noise and the number of ensemble iterations, I, EEMD can decompose the SH, LH, and EVI time series into multiple IMF components and a residual term. The workflow of the EEMD process is illustrated in Figure 4.

In this study, a filtering mechanism was implemented during the extraction of IMF components. A threshold value (0.2) was set, and the filtering process was considered complete when the difference in standard deviation between two consecutive iterations of selected IMF components was less than the threshold. Establishing a reasonable filtering mechanism ensures the extraction of meaningful IMF components. If too many IMF components are obtained, the results may lack physical significance. The typical range for the threshold value is 0.2 to 0.3. The formula for calculating the standard deviation of IMF components is provided in Equation (1).

$$D_S={\displaystyle \sum _{j=0}^J\frac{{\left|\left[h_{\left(w-1\right)}\left(j\right)-h_w\left(j\right)\right]\right|}^2}{{h_w}^2\left(j\right)}}$$

(1)

where D_s is the standard deviation of the IMF. J is the length of the time series, $h_{\left(w-1\right)}\left(j\right)$ and $h_w\left(j\right)$ are the time series of two consecutive processing results during the IMF filtering process.

Next, the Granger causality test was performed. Based on the results of the EEMD decomposition, the low-frequency IMF components of each variable under different topographic factors were selected. These low-frequency components are used to reduce the influence of short-term fluctuations and random noise, thereby highlighting the dominant long-term variations that are more relevant to interannual to decadal feedback mechanisms. The low-frequency stationary IMF sequences of the two variables, SH and EVI, or LH and EVI, are defined as I_X and I_Y, respectively. The following regression equations were then established to conduct the causality test:

$$I_{Y,t}={\delta }_0+{\displaystyle \sum _{i=1}^p{\delta }_i{I}_{Y,t-i}}+{\displaystyle \sum _{i=1}^p{\phi }_i{I}_{X,t-i}+{\mu }_t}$$

(2)

$$I_{X,t}={\sigma }_0+{\displaystyle \sum _{i=1}^p{\sigma }_i{I}_{Y,t-i}}+{\displaystyle \sum _{i=1}^p{\lambda }_i{I}_{Y,t-i}+{\nu }_t}$$

(3)

where ${\delta }_0$ and ${\sigma }_0$ are constant terms; ${\delta }_i$, ${\phi }_i$, ${\sigma }_i$, and ${\lambda }_i$ are regression coefficients; p is the maximum lag order; ${\mu }_i$ and ${\nu }_i$ are residual terms.

For this study, the value of p is set to 2 (i.e., 2 years) to ensure that the sample size T satisfies T > 2p + 1. This criterion ensures that the residual degrees of freedom T − 2p are sufficient for conducting an effective test.

If the null hypothesis (H₀) states that I_Y does not cause changes in I_X, or I_X does not cause changes in I_Y, i.e., H₀: ${\phi }_1={\phi }_2=\cdot \cdot \cdot ={\phi }_p=0$ (regression coefficients in Equation (2)) or H₀: ${\lambda }_1={\lambda }_2=\cdot \cdot \cdot ={\lambda }_p=0$ (regression coefficients in Equation (2)), the alternative hypothesis (H₁) is that at least one ${\phi }_i\ne 0$ or ${\lambda }_i\ne 0$. Under the null hypothesis, the F-statistic is calculated, which follows an F_{(p, T−2p)} distribution:

$$F={\displaystyle \frac{\left(SS{E}_q-SS{E}_u\right)/p}{SS{E}_u/\left(T-2p\right)}}~F_{\left(p,T-2p\right)}$$

(4)

The sum of squared errors (SSE) is calculated as follows:

$$SSE={\displaystyle \sum _{t=1}^T{\left(y_t-{\widehat{y}}_t\right)}^2}$$

(5)

where F is the test statistic; SSE_q is the sum of squared errors for the restricted model, where the lagged terms of I_X (I_Y) are assumed to not affect the dependent variable, corresponding to the null hypothesis (H₀); SSE_u is the sum of squared errors for the unrestricted model, where the lagged terms of I_X (I_Y) are assumed to have an effect, corresponding to the alternative hypothesis (H₁); T is the sample size, which spans 22 years in this study; y_t is the observed value for year t; ${\widehat{y}}_t$ is the predicted value for year t; $\left(y_t-{\widehat{y}}_t\right)$ is the residual for year t.

The testing procedure is as follows: if SSE_u is significantly smaller than SSE_q, and the computed F-value is significantly less than the critical value from the F_{(p, T−2p)} distribution, the null hypothesis is rejected. This indicates that the lagged values of I_Y (I_X) have a significant predictive effect on I_X (I_Y). Alternatively, if the p-value is less than 0.01, 0.05, or 0.1, the null hypothesis is rejected, suggesting the existence of a Granger causal relationship between the two variables.

This EEMD preprocessing step decomposes nonlinear and nonstationary time series into stationary components while retaining important nonlinear characteristics of the original data. This approach is suitable for analyzing hydrometeorological variables where such characteristics are essential for identifying potential causal links. Although the subsequent Granger causality test is based on a linear regression framework, which may limit its ability to fully capture highly complex nonlinear dynamics, the preserved nonlinear features in the IMF components allow the method to detect causal relationships that might not be identified when applying traditional Granger tests directly to raw data.

2.3.2. Sensitivity Analysis Method of Vegetation and Heat Fluxes

The multiple regression model is a widely utilized statistical methodology for analyzing the relationship between multiple independent variables and a dependent variable [57]. The regression coefficients derived from the model indicate the extent to which each independent variable contributes to the dependent variable. Larger regression coefficients indicate greater influences of the independent variable on the dependent variable. A positive regression coefficient indicates that as the independent variable increases, the dependent variable also increases. A negative coefficient indicates that as the independent variable increases, the dependent variable decreases. The most similar studies employed only the regression coefficients derived from this model as sensitivity coefficients. Inconsistencies in the unit of measurement of the independent variables may introduce bias into the results of sensitivity analyses, particularly if the dependent variable is also measured in a non-standard manner. To eliminate the interference of the different units of the independent variables and the dependent variables in the process of sensitivity analyses, this study employs a methodology in which the regression coefficients of the model are multiplied by the ratio between the variance of the independent variable and the variance of the dependent variable. This yields the sensitivity coefficients of the dependent variable to the independent variable β, whose absolute values directly reflect the degree of sensitivity of the dependent variable to the independent variable. As EVI is continuous data, the Variance Inflation Factor (VIF) for SH and LH is below 5, indicating low multicollinearity [58]. Therefore, to ensure consistency in the input results, a multiple linear regression model is established in this study, in which the low-frequency series of EVI, previously selected, is set as the dependent variable (Y), and the low-frequency series of SH and LH are set as the independent variables (X₁ and X₂), as shown in Equation (6).

$$X_3=a{X}_1+b{X}_2+c$$

(6)

where X₁ and X₂ represent the independent variables SH and LH, respectively, while X₃ represents the dependent variable EVI; the parameters a and b correspond to the regression coefficients in the multiple regression model; c is the constant term.

Where the standard deviation (s) of EVI, SH, and LH are calculated as:

$$s=\sqrt{{\displaystyle \frac{1}{f-1}}{{\displaystyle \sum _{k=1}^f\left(x_k-\overline{x}\right)}}^2}$$

(7)

where s is the standard deviation; f is the sample size of the independent variable SH or LH. $x_k$ is the observed value of SH or LH in the k-th year; $\overline{x}$ is the mean of the independent variable SH or LH.

The sensitivity coefficients for SH and LH in this study are calculated as follows:

$${\beta }_1=\left|a\times {\displaystyle \frac{s_s}{s_e}}\right|; {\beta }_2=\left|b\times {\displaystyle \frac{s_l}{s_e}}\right|$$

(8)

where β₁ is the sensitivity coefficient corresponding to SH (X₁); β₂ is the sensitivity coefficient corresponding to LH (X₂); a and b are the regression coefficients in Equation (6); $s_s$, $s_l$, $s_e$ are the standard deviations of SH, LH, and EVI, respectively.

3. Results

3.1. The Temporal Variation Characteristics of SH and LH on the TP

Figure 5 illustrates the annual area-averaged SH (Figure 5a), LH (Figure 5b), and Bowen ratio (Figure 5c) over the TP from 2000 to 2022. Reanalysis data indicate that over the past 23 years, the SH and LH have increased at rates of 0.28 W·m⁻²·decade⁻¹ and 0.105 W·m⁻²·decade⁻¹, respectively. The Bowen ratio decreased in 2004, increased in 2007, declined again in 2013, and rose once more in 2020. These variations contrast with the overall decreasing trend of the Bowen ratio observed during 1979–2021 [59].

During the study period, in some years, the decrease in SH transport stabilized the atmosphere and suppressed convection and precipitation. At the same time, the increase in LH enhances the water vapor cycle, increasing atmospheric water content. On the contrary, in other years, the increase in SH destroys the stability of the atmosphere and promotes convection and precipitation, while the decrease in LH weakens the water vapor cycle and leads to a decrease in atmospheric water content. It is important to note that these processes involve substantial uncertainties. For example, atmospheric warming caused by increased SH can reduce relative humidity by raising air temperature [60,61]; however, the net effect on precipitation depends on factors such as regional moisture availability, vertical mixing, and atmospheric circulation [62,63]. LH, as a key source of moisture flux, generally supports convection and precipitation, but its effect may be modulated by local climate and topography [64,65,66]. In general, the atmospheric stability of the TP in the 21st century is relatively low, which may lead to more frequent extreme weather events. As the plateau continues to experience greening and humidification [67,68], the unstable atmosphere is likely to influence vegetation growth and the ecosystem in the plateau [69]. It may also exacerbate the rapid fluctuations of SH and LH throughout the region.

3.2. The Causality Relationships Between Vegetation and Heat Fluxes

To study the potential association between vegetation and SH and LH on the TP, the EEMD–Granger causality test was conducted on SH, LH, and EVI.

Table 2. EEMD–Granger causality test results for SH, LH, and EVI across different elevations.

Elevations/m	SH to EVI		EVI to SH		LH to EVI		EVI to LH
	F	p-Value	F	p-Value	F	p-Value	F	p-Value
0–1000	4.029	0.041 **	1.037	0.38	11.509	0.001 ***	7.762	0.005 ***
1000–2000	6.067	0.023 **	0.945	0.343	0.92	0.021 **	4.13	0.039 **
2000–3000	5.108	0.019 **	0.374	0.694	4.345	0.030 **	1.971	0.177
3000–4000	3.797	0.066 *	0.115	0.738	8.416	0.006 ***	0.314	0.026 **
4000–5000	1.695	0.219	1.293	0.305	1.856	0.016 **	0.749	0.045 **
5000–6000	0.664	0.027 **	0.614	0.444	2.07	0.033 **	0.432	0.658
6000–7000	0.406	0.674	0.568	0.579	0.197	0.223	0.336	0.119
7000–8000	0.043	0.958	1.399	0.275	0.366	0.699	1.132	0.347
8000–9000	1.759	0.208	0.929	0.418	0.354	0.708	0.072	0.931

Note: * indicates significance at the 90% level, ** at the 95% level, and *** at the 99% level in EEMD–Granger Causality Testing.

,

Table 3. EEMD–Granger causality test results for SH, LH, and EVI across different slopes.

Slopes/°	SH to EVI		EVI to SH		LH to EVI		EVI to LH
	F	p-Value	F	p-Value	F	p-Value	F	p-Value
0–10	1.599	0.233	0.571	0.576	0.59	0.566	0.664	0.528
10–20	6.067	0.006 **	0.114	0.893	1.334	0.017 **	1.507	0.012 **
20–30	6.98	0.001 ***	0.199	1.046	5.209	0.024 **	3.602	0.073 *
30–40	4.283	0.032 **	1.096	0.358	5.143	0.035 **	1.766	0.2
40–50	0.268	0.61	0.297	2.235	1.496	0.254	1.114	0.049 **
50–60	6.788	0.018 **	0.134	1.017	4.321	0.021 **	4.506	0.028 **
60–70	3.143	0.069 *	0.738	0.587	4.197	0.023 **	3.336	0.048 **
70–80	1.569	0.226	0.516	0.482	0.741	0.495	1.331	0.296
80–90	0.3	0.591	0.022	0.884	0.949	0.408	0.038	0.963

Note: * indicates significance at the 90% level, ** at the 95% level, and *** at the 99% level in EEMD–Granger Causality Testing.

and

Table 4. EEMD–Granger causality test results for SH, LH, and EVI across different aspects.

Aspects	SH to EVI		EVI to SH		LH to EVI		EVI to LH
	F	p-Value	F	p-Value	F	p-Value	F	p-Value
Flat	0.448	0.723	0.376	0.772	3.814	0.066 *	0.044	0.837
North	5.731	0.017 **	2.159	0.148	4.334	0.027 **	3.804	0.045 **
Northeast	5.798	0.015 **	0.899	0.668	4.783	0.037 **	2.247	0.15
East	3.443	0.042 **	0.396	0.121	2.274	0.135	0.021	0.979
Southeast	4.467	0.050 **	0.443	0.514	0.405	0.674	0.131	0.878
South	3.97	0.057 *	1.27	0.274	0.913	0.421	0.325	0.727
Southwest	0.61	0.557	0.062	0.94	4.401	0.676	0.386	0.686
West	4.871	0.047 **	0.458	0.765	7.741	0.007 ***	6.57	0.014 **
Northwest	5.613	0.022 **	0.606	0.623	1.222	0.343	0.088	0.965

Note: * indicates significance at the 90% level, ** at the 95% level, and *** at the 99% level in EEMD–Granger Causality Testing.

present the Granger causality test results for SH, LH, and EVI at varying elevations, slopes, and aspects. Figure 6 intuitively shows the causal relationship between SH, LH, and EVI.

Combined with the data analysis of Table 2, Table 3 and Table 4, Figure 6a reveals that SH and EVI show a unidirectional causal relationship, while LH and EVI show a bidirectional causal relationship in the south and north aspects, altitude of 0–2000 m, slope of 10–20°, and 50–60°. This means that SH influences vegetation change, but vegetation does not influence SH. In contrast, LH can influence vegetation change, and LH can be influenced by vegetation change.

Figure 6b reveals that, in the northeast aspect at elevations of 2000–3000 m and 5000–6000 m, with slopes of 20–40°, both SH and LH demonstrate unidirectional causality with EVI. SH and LH both show unidirectional causality with EVI. This suggests that SH and LH can drive vegetation change, but vegetation change does not influence SH or LH change.

Figure 6c reveals that at elevations of 3000–5000 m, with slopes of 60–70°, only LH and EVI demonstrate bidirectional causality. Therefore, it can be concluded that vegetation change can be caused by LH and that vegetation can also exert an influence on LH.

Figure 6d reveals that in regions at elevations of 6000–9000 m with slopes of 0–10° and 70–90° on the south and southwest aspects, as well as in flat regions, there are no causal relationships between SH, LH, and EVI. This suggests that neither SH nor LH can cause vegetation change, and that vegetation cannot cause a change in SH and LH.

Figure 6e reveals that in regions with east, southeast, and northwest aspects, the unidirectional causal relationship between SH and EVI is the only one observed, while there is no causal relationship between LH and EVI. This implies that in this region, only SH can cause vegetation change, while vegetation cannot cause change in SH and LH. In regions with east, southeast, and northwest aspects, only SH has a unidirectional causal relationship with EVI, and there is no causal relationship between LH and EVI, implying that in this region, only SH can cause vegetation change, while vegetation cannot cause change in SH and LH. The results of Figure 6a show that in the west aspect region, SH has a unidirectional causal relationship with EVI, and LH has a bidirectional causal relationship with EVI. This suggests that SH can cause vegetation change, but that vegetation cannot cause SH change. Conversely, LH can both cause and be caused by vegetation change.

Figure 6f reveals that in regions with slopes of 40–50°, only LH and EVI show unidirectional causality. This means that, in these regions, vegetation influences LH, but neither SH nor LH influences vegetation change.

The analysis of existing data and figures shows that vegetation change on the TP has not significantly influenced SH change. This indicates that the direct correlation between SH and vegetation on the TP is not particularly robust. In regions of high elevation, due to the scarcity of vegetation coverage, the causal relationship between SH, LH, and EVI is meaningless. The primary drivers of vegetation change at these elevations remain factors such as temperature, moisture distribution, wind direction, and wind speed, all of which are influenced by high-elevation conditions.

3.3. Analysis of the Strength of Causality Between Vegetation and Heat Fluxes

The EEMD–Granger causality test results of SH, LH, and EVI above indicate that in some regions, e.g., the northeast aspect with an elevation of 2000–3000 m and a slope of 20–40°, SH and LH have caused changes in plateau vegetation. This raises the problem of which variable in SH or LH has a greater influence on vegetation. Therefore, the sensitivity coefficients β of vegetation to SH and LH were calculated based on the coefficients of SH and LH in the multiple linear regression model. The results are shown in Figure 7.

As illustrated in Figure 7a, vegetation in regions below 1000 m elevation shows higher sensitivity to SH (0.52) than to LH (0.37), both significant at the 95% level, indicating SH as the dominant influence. At elevations of 1000–2000 m, only the SH-EVI sensitivity passed the 95% significance test, suggesting a continued dominance of SH. In the 2000–5000 m elevation band, more LH shows a more pronounced influence on EVI (${\beta }_1(SH)<{\beta }_2(LH)$), with significant sensitivity at the 95% level for 2000–3000 m and 4000–5000 m, and at the 99% level for 3000–4000 m. At 5000–6000 m, EVI is more sensitive to SH (0.25) than to LH (0.01), both statistically significant. At elevations of 6000–8000 m, previous EEMD–Granger tests indicated no causal relationship between EVI and either SH or LH; the sensitivity analysis supports this, as neither flux exhibits significant influence. Above elevations of 8000 m, especially in the western part of the plateau, where vegetation is nearly absent and the surface is dominated by glaciers and snow, both SH and LH have low but comparable sensitivity coefficients (>0.01). This may reflect interference from snow’s red-light absorption, leading to deviations in EVI measurements, and suggests that glacier snow may still respond weakly to LH variations.

As illustrated in Figure 7b, the sensitivity of EVI to SH and LH varies notably across regions with different slope gradients. In regions with slopes of 0–10°, neither SH nor LH significantly influences vegetation change, as their sensitivity coefficients fail to reach the 90% significance level. This suggests that vegetation in these flat regions is primarily controlled by other factors, such as precipitation and temperature. In regions with slopes of 10–40°, both SH and LH can influence vegetation, but EVI responds more strongly to SH, especially on slopes of 30–40°, where the sensitivity to SH reaches 0.95 and is significant at the 99% level, indicating SH as the dominant driver of vegetation change. In regions with slopes of 40–70°, EVI shows increased sensitivity to LH. Particularly in regions with slopes of 50–60°, both SH (1.0) and LH (1.1) exert significant influence, with coefficients significant at the 95% level. A coefficient above 1.0 reflects that a one-standard-deviation change in SH or LH corresponds to an EVI change exceeding one standard deviation, suggesting a strong and amplified vegetation response under these topographies. In regions with slopes of 60–70°, LH becomes the primary influence, with a sensitivity coefficient of 0.27, significant at the 99% level. In regions with slopes of 70–90°, neither SH nor LH shows significant effects on EVI, likely due to sparse vegetation and limited transpiration and respiration on steep terrain, which weakens feedback between vegetation and surface heat fluxes.

As illustrated in Figure 7c, the aspect is divided into sunny slopes (south, southwest, west, northwest) and shady slopes (north, northeast, east, southeast). Compared with SH, LH has generally higher sensitivity coefficients in the north, northeast, east, south, and west aspects, but only the sensitivity coefficients in the north, northeast, and west aspects are statistically significant at the 95% confidence level, indicating that LH has a stronger impact on vegetation in these regions. It is worth noting that, among them, only the west faces the sun. The vegetation in the northeast direction is most sensitive to LH, which may be due to the balance between the intensity and duration of sunlight, which enhances the growth of vegetation and the variability of LH. In contrast, SH has a greater impact on the east, south, and northwest aspects. The sensitivity of the east aspect to SH is the highest (0.35), which is significant at the 90% level, while the other two aspects exceed the 95% level. However, the vegetation on the south aspect is not significantly influenced by SH or LH, making the sensitivity coefficient statistically meaningless. In general, vegetation changes on the east and southwest aspects are more influenced by SH, while LH has a greater influence on some shady slopes.

4. Discussion

Topography significantly influences the vegetation responses to SH and LH on the TP. In some regions, vegetation change is simultaneously influenced by both SH and LH; the degree of influence varies depending on the topography. Numerous studies have demonstrated that vegetation dynamics and precipitation, temperature, solar radiation, and relative humidity on the TP [70,71,72,73], and other climate factors are closely related. On this basis, the results of this study indicate that SH and LH also play an important role in shaping the growth and spatial distribution of vegetation under different topographies on the TP (Figure 8).

SH refers to the heat exchange between the surface and the atmosphere due to the temperature difference, while LH represents the heat related to the water vapor phase transition during the interaction between the earth and the atmosphere. This shows that the heat exchange between the surface and the atmosphere not only affects the growth of vegetation on the plateau but also plays a significant role in the heat exchange driven by the water vapor phase transition. The different heat transfer mechanisms associated with SH and LH lead to a weaker positive feedback effect of vegetation change on the minimum temperature relative to the maximum temperature in all seasons on the TP [74]. In addition to the above new findings, this study also revealed that vegetation in different terrains had an effect on LH but had no effect on SH (Figure 9). The mechanisms of heat transfer between the surface and the atmosphere within the TP are complex. The heat transfer mechanism between the surface and the atmosphere in the TP is complex. The fundamental driving force of heat exchange is the temperature difference between the surface and the atmosphere, and the evaporation and condensation of water vapor are important supplements to heat transfer. This shows that vegetation does not change the temperature gradient caused by the highest and lowest temperatures on the TP. The interaction between vegetation growth and LH indicates that the transpiration of large-scale vegetation plays an indispensable role in the interaction between the plateau surface and the atmosphere.

In low-elevation regions (below 2000 m, Figure 8d), vegetation change and distributions remain relatively stable [75]. Both SH and LH can influence vegetation in these regions (Figure 8a and Figure 9a), but SH has a greater influence on vegetation change. This indicates that to preserve such stability, the SH change in this region may also be relatively stable, showing fluctuations that deviate from a broader oscillation trend. Below 3000 m on the TP, snow cover is minimal (<4%), while above 6000 m, snow cover can reach 77% [76]. Low-elevation areas generally have a higher surface–air temperature difference, especially under strong solar radiation, which enhances SH [77]. In addition, these regions have little snow cover, limiting the pathway for LH; however, active vegetation further strengthens SH through surface–atmosphere turbulent heat transfer [78].

In the mid-elevation regions (2000–5000 m, Figure 8d), there is a negative correlation between altitude and vegetation coverage, and the increase in altitude corresponds to the decrease in vegetation. On the one hand, the simultaneous increase in vegetation’s ability to adapt to environmental changes has led to a slight increase in species richness due to increased viability [79]. On the other hand, the overall temperature vertical lapse rate in this region shows a temperature drop of about 0.41–0.6 °C·(100 m)⁻¹ [80]. This means that the vegetation ecosystem in the middle altitude area is fragile and vulnerable to climate change. It is worth noting that LH has a more significant effect on vegetation change in this area (Figure 9a), which emphasizes that the middle altitude region of the TP is particularly sensitive to climate change. These mechanisms indicate that the results are not randomly distributed but have clear physical foundations. This also indirectly shows that the energy released by the phase transition of water vapor and plant transpiration exceeds the energy released by thermal convection, making the region crucial for vegetation ecological protection and maintenance of ecological security barriers. Therefore, it is necessary to closely monitor the dynamic changes of vegetation and environment in this region, especially to strengthen the ecological protection of vegetation and prevent the decline of species richness.

In high-elevation regions (above 6000 m), extensive ice and snow cover and the presence of permafrost significantly limit the habitable space of vegetation, resulting in [81,82] the vegetation coverage becoming increasingly sparse, and the meteorological conditions worsening. With the decrease in atmospheric pressure and density, the air is more directly influenced by the change in surface temperature, so the SH increases. At the same time, the temperature at high altitudes decreased significantly, resulting in a decrease in saturated water vapor pressure, which was directly attributed to the decrease in temperature, which in turn reduced LH. However, as long as vegetation exists, transpiration can still release LH, especially in regions where the snow or frozen layer is thin, making LH the dominant flux [78]. As mentioned above, SH and LH are no longer the main factors influencing the growth and distribution of vegetation on the TP in the extreme environment of high-elevation regions. The influence of sparse surface vegetation on these states is minimal. It can be reasonably inferred that vegetation changes in high-altitude areas are mainly affected by direct environmental factors, including water availability, temperature fluctuations, and wind speed, rather than changes in SH and LH.

The analysis of this study found that there is no consistent relationship between the causal relationship and sensitivity of EVI to SH and LH in the topographic factor of slope direction. Vegetation on the sunny slopes (e.g., west, south, and northwest aspects) and shady slopes (e.g., north, northeast, and east aspects) showed a single sensitivity to SH or LH (Figure 9c,d). This indicates that sunlight duration and intensity may not be the dominant factors regulating the interaction between vegetation and SH or LH on the TP.

The results of this study show that the heat exchange process between the surface and the atmosphere of the TP has different influences on vegetation in different topographies. It is currently unclear whether this effect is beneficial or detrimental to vegetation growth. It is not clear whether this effect is beneficial or detrimental to vegetation growth. However, the TP continued to experience greening as a whole [67,83] and concluded that the effects of SH and LH on vegetation are more beneficial than harmful to the ecosystem. This does not rule out that factors such as precipitation, temperature, and sunshine hours driven by climate change may contribute more to vegetation growth than heat flux. This hypothesis needs to be further verified by extensive empirical data collection. Considering the different vertical climate characteristics of the plateau, the influence of slope and aspect on the sensitivity of vegetation to SH and LH is not as significant as that of elevation. Although elevation is the main topographic factor influencing the sensitivity of vegetation to SH and LH on the TP, the effects of slope and aspect cannot be ignored. Due to the limitation of spatial resolution and accuracy of remote sensing data, some detailed information may be missed in the analysis process, which may underestimate the observation effect of slope and aspect. The specific physical mechanisms by which SH influences vegetation involve complex land–atmosphere interactions. Due to space limitations, these mechanisms have not been explored in detail in this study, but will be the focus of future research. In addition, due to the limited time length of high-resolution satellite remote sensing vegetation data, the long-term trend of the data will also be ignored in the study. In short, for vegetation protection, ecological construction, and research TP [84,85], researchers must consider the causal relationship between vegetation changes and SH and LH under different terrains, and the different sensitivities of vegetation to these factors.

5. Conclusions

The topographic control of the interaction between vegetation change and heat fluxes in the TP was studied. In this study, we improved the Granger causality model to consider nonstationary conditions and applied it to explore the causal relationship between vegetation and surface heat fluxes (SH and LH) under different terrains. Further sensitivity analysis was performed to assess the strength of causal influences. The results show that the SH and LH are increasing at the rates of 0.28 W·m⁻²·decade⁻¹ and 0.105 W·m⁻²·decade⁻¹, respectively. SH dominates in low elevations (<3000 m) and gentle-to-moderate slopes (10–40°), and LH dominates in high elevations (5000–6000 m) and steep slopes (50–70°). SH and LH jointly regulated the vegetation growth of shady and gentle slopes, mainly tending dense grassland, but the influence of SH was stronger. In contrast, LH interacts with all types of vegetation at middle elevations and steep slopes of 3000–5000 m, forming an intensive local energy cycle. Below 6000 m, except for the range of 2000–3000 m, vegetation change has a significant regulatory influence on LH, and this feedback is largely independent of topography and contributes to energy redistribution and water cycle maintenance. It is worth noting that although vegetation is influenced by both SH and LH, only LH is influenced by vegetation, indicating that unidirectional feedback highlights the regulatory role of vegetation in the LH process. Among the topographic variables, elevation has the strongest control effect on vegetation–heat flux interaction, and slope and aspect jointly enhanced this effect. The results of this study reveal the performance of the interaction between vegetation and heat flux in different terrains and provide support for maintaining the stability of the ecological environment and land–atmosphere heat exchange in the TP.