Regional vegetation dynamics and its response to climate change—a case study in the Tao River Basin in Northwestern China

The 30-year normalized-difference vegetation index (NDVI) time series from AVHRR/MODIS satellite sensors was used in this study to assess the regional vegetation dynamic changes in the Tao River Basin, which cuts across the Eastern Tibetan Plateau (ETP) and the Southwestern Loess Plateau (SLP). First, principal component and correlation analyses were carried out to determine the key climatic variables driving ecological change in the region. Then, regression models were tested to correlate NDVI with the selected climatic variables to determine their predictive power. Finally, Sen’s slope method was used to determine how terrestrial vegetation has responded to regional climate change in the region. The results indicated an average winter season NDVI value of 0.14 in the ETP but only 0.04 in the SLP. Primarily driven by increasing temperature, vegetation growth has generally been enhanced since 1981; spring NDVI increased by 0.03 every 10 years in the ETP and 0.02 in the SLP. Further, results from trend analyses suggest vegetation growth in the ETP shifted to earlier-start and earlier-end dates, however in the SLP, the growing season has been extended with an earlier-start and later-end date. The precipitation threshold for vegetation germination, measured by the cumulative spring rainfall, was found to be 44 mm for both the ETP and SLP.


Introduction
Climate and vegetation dynamics are tightly coupled: regional climate affects land surface processes over a range of scales with unprecedented speed (IPCC 2007, Zhao et al 2011, while vegetation, in turn, affects climate through feedbacks via photosynthesis and evapotranspiration, changes in albedo and biogenic volatile organic compound emissions (Henderson-Sellers 1993, Fang et al 2003, Meng et al 2011, Faubert et al 2012, Wang and Dickinson 2012, Henden et al 2013. For example, studies reported that vegetation growth at high latitudes in some Northern hemisphere regions has increased from 1981 to the 1990s due to climate change (e.g., Nemani et al 2005), and changes in vegetation leaf area index have lead to shifts in temperature and precipitation patterns (Ge Environmental Research Letters Environ. Res. Lett. 9 (2014) 125003 (12pp) doi: 10.1088/1748-9326/9/12/125003 Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. et al 2007, Pu and Dickinson 2013). However, these feedback mechanisms are complex, varying greatly from location to location and over time (Turner et al 2003, Pettorelli et al 2005, Nie et al 2012 because the response of terrestrial vegetation to climatic factors such as air temperature and precipitation is spatially and temporally heterogeneous (Curran 1980, Lambin et al 2001. To account for this heterogeneity and to fully understand the response of ecosystems to climate change (Ni 2011, Guo and Zhang 2013, Li et al 2013, it is necessary to conduct location specific case studies for different geographic regions (Fan and Zhang 2010, Horion et al 2013 so that spatially explicit conclusions can be drawn.
Remote sensing provides a vital tool to capture the temporal dynamics of vegetation change in response to climate shifts, at spatial resolutions fine enough to capture the spatial heterogeneity. Frequent satellite data products, for example, can provide the basis for studying time-series of ecological parameters related to vegetation dynamics (Bradley et al 2007, Gu et al 2009, Jacquin et al 2010, Beck et al 2011. Among the many available remote-sensing data products, the normalized-difference vegetation index (NDVI) has been frequently used in vegetation dynamics studies, as this index is highly correlated with the leaf area index, photosynthetic capacity, biomass, dry matter accumulation, and net primary productivity (Wang et al 2010, Cartus et al 2011, Raynolds et al 2012. Therefore, NDVI data are frequently used to assess spatio-temporal changes in regional vegetation dynamics (Kang et al 2011, Zhang et al 2011 in response to changes in regional climate. In this study, we used advanced very high resolution radiometer (AVHRR) and moderate resolution imaging spectroradiometer (MODIS) NDVI data, along with yearly and monthly net radiation, air temperature, and precipitation data to examine the feedback mechanisms between climate and vegetation. To capture the spatial heterogeneity of vegetation responses to a changing climate, we selected the Tao River Basin (TRB) in Northwest China, with a gradient in both climate and vegetation, to study the feedback mechanisms between spatially heterogeneous vegetation and climate regimes. The objective of this study is to better understand the spatial variability of vegetation responses to regional climate change.

Study area
The TRB covers a total area of 25 527 km 2 between 101°36′E and 104°20′E, and 34°06′N and 36°01′N (figure 1) and includes portions of the Eastern Tibetan Plateau (ETP) and Southwestern Loess Plateau (SLP). In the ETP, forests dominate mountains and grasslands mainly cover open valleys, whereas in the SLP, the vegetation are primarily grasslands at low coverage. The region averaged annual mean air temperature increases from 1°C in the ETP at elevations ranging from 4560 to 2800 m asl (above sea level), to 9°C in the SLP at elevations ranging from 2800 to 1730 m asl. The annual mean precipitation decreases from 600 mm in the West to 300 mm in the East . In addition, the annual net radiation is higher in the ETP because of its higher elevation. In short, from upstream in the West to downstream in the East, the dominant climate varies from an alpine cold humid and sub-humid climate to a temperate semi-arid climate, while the terrestrial vegetation ranges from alpine grasslands in the upstream regions to forest and arid grasslands in the middle and downstream regions.
From 1980s to 2000s, the dominant land cover and land use (grasslands, forests, rainfed cultivated lands and little urban areas) changed insignificantly with less than 1% in total area of the watershed. Further, based on limited statistical data from the 'grain for green' project survey, it was concluded that total land use conversion during this period of time was less than 0.5%, suggesting human-related land use change is insignificant . Land use change during our study period (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000), including grazing intensity and expansion of irrigated agriculture is insignificant in the TRB. The grazing ban policy was implemented in early 1970s with traditional wire fences, and since then there has been no change. Irrigated agriculture, based on our survey and an existing Land Use Map of China (produced by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, 2000), does not appear to exist in the region. Therefore, at the TRB-basin scale, changes in land use have had little impact on regional scale vegetation vigor, during our study period from 1981 to 2010.
Preliminary remote sensing analyses in this study indicated an increasing trend in NDVI in the ETP and SLP regions, suggesting that total vegetative cover increased significantly, especially in the spring seasons, NDVI has increased by 34% and 29% for the ETP and SLP regions, respectively, since the 1980s beyond a reasonable level expected due to human impact alone. Therefore, we hypothesize that climate is the likely dominant driver of vegetation change in the study area. The objectives of this study are to determine whether or not climate is the dominant driver of vegetation change, and if so, what are the specific climate variables that are most effective for predicting vegetation change.

NDVI data cross-calibration
In this study, the NDVI products derived from AVHRR (8 km resolution) and MODIS (1 km resolution) were acquired for the period from 1981 to 2006 and from 2001 to 2010, respectively. We obtained the largest annual NDVI value for each pixel in the region using the maximum-value composite (MVC) method, and resampled all the MODIS data to an 8 km resolution to correspond to the AVHRR data resolution (Holben 1986, Yang et al 1988. We recognized there exist some discrepancies between the AVHRR and MODIS datasets, and there is a need for cross calibration (Gitelson and Kaufman 1998, Fensholt et al 2009. Gitelson and Kaufman (1998) used spectral characteristics to fuse the two datasets, while Brown et al (2008) attempted to use neural network as data fusion tool. These analyses on the data consistency fusion techniques focus on high spatial resolution imagery such as MODIS and Landsat/MISR/ASTER ( Stefanov and Netzband 2005, Naud et al 2007, Vermote et al 2007, Roy et al 2008, Hilker et al 2009, Li et al 2013, Walker et al 2014. For much coarser spatial resolution datasets such as AVHRR (GIMMS) and MODIS or AVHRR and SPOT, given that they are often acquired during different time periods (e.g. AVHRR/GIMMS was acquired during 1981-2006 while MODIS is 2001-present), consistency assessment is often carried out by correlating corresponding pixels of overlapping time period (e.g. Song et al 2010). This method would avoid any issues associated with spectral and spatial calibrations of each sensor and, therefore, was adopted in this study. We used the subset of two datasets acquired during the overlapping time period from 2001 to 2006 and a linear regression method to assimilate the two time series so that they have the same mean values during the overlapping dates. The regression equation for the assimilation was found to be:

MODIS AVHRR
We further checked the consistency of the two NDVI series after assimilation (figure 2). The correlation coefficient between the two series was 0.77 (P < 0.01). The absolute error of NDVI was less than 0.1 and the relative error was less than 20% for 87.1% of the pixels. Thus, there was high consistency between the AVHRR and MODIS NDVI values after the cross calibration, and it was therefore reasonable to use the assimilation approach to analyze the regional vegetation dynamics for the entire 30 year-period from 1981 to 2010.

Climate data
The climatic data used in this study included net radiation − ( ) R , MJ m n 2 , maximum, minimum and mean monthly air temperatures (°C) (T mean , T max , T min ), cumulative monthly air temperatures above 0, 3, 5, and 10°C (CT 0 , CT 3 , CT 5 , and CT 10 , respectively), and cumulative precipitation (mm) in the present year (CP 0 ), and since December (CP 1 ) and November (CP 2 ) of the previous year until the month when the NDVI value was recorded. Combining the climate data with the assimilated monthly average NDVI data, we constructed a monthly climate and vegetation time-series. The climate data were interpolated from weather stations in the TRB (figure 1) using a krigging method, and were subsequently used for zonal statistics analyses. For seasonal analyses we further divided the annual data into four seasons: winter (December-February), spring (March-May), summer (June-August), and autumn (September-November). This allowed us to examine vegetation responses to climate in different seasons.

Trend analysis
Trend analysis is an active subject for variability determination of natural time series such as climate, hydrology and vegetation (Hamed 2008, Zhao et al 2008. In this study, we conducted trend analysis to better understand the interactive nature between climate and vegetation dynamics. Numerous trend analysis methods have been developed (e.g., Henebry and Betancourt 2010, Ma et al 2011, Ye et al 2013) and three frequently used methods tested in this study are the Mann-Kendall test (Mann 1945, Kendall 1955, least-squares linear regression, and Sen's slope trend test (Sen 1968, Nayak et al 2010. Among these methods, the Mann-Kendall test was found to be the best for detecting significant trends, but not their magnitude, whereas least-squares linear regression and Sen's slope trend test were better for determining the magnitude of the slope. The Mann-Kendall test was found to be less sensitive to extreme values in the data series, whereas least-squares linear regression was affected by both extreme values and autocorrelation within the data series. Sen's slope trend test was able to eliminate the impact of missing data or anomalous trends by using the median of the series of slopes as the judgmental foundation (Zhang et al 2002, Stow et al 2004. In this study, we compared all three methods and selected Sen's slope to detect and characterize trends in the vegetation dynamics and climate parameters in the TRB.

Principal-components analysis (PCA)
We investigated 11 climatic factors that potentially affect vegetation dynamics in the study area. PCA revealed the relative contribution of these factors or groups of factors to vegetation dynamics, as measured by changes in NDVI. We used the eigenvalue and cumulative contribution of the principal components to classify the standardized driving factors into different categories. We standardized the climatic factors that affected the vegetation dynamics using the following equation: where X j is the standardized value of the jth driving factor; X jobs is the observed value of the jth driving factor; X j is the arithmetic mean of the jth driving factor; σ j is the standard deviation of the jth driving factor; and j ∈ (1, n), where n represents the number of driving factors.
Based on orthogonal transformation, the n standardized driving factors can be converted into m principal components: where, F m is the standardized equivalent value of principal component m estimated by the driving factors X n ; a mn is the score coefficient for the components and is equal to the principal component load matrix divided by the corresponding eigenvalue. It reflects the impact of a climate factor on the principal component. The principal component load matrix requires mutual independence among the variables as much as possible, and must converge after several iterations of orthogonal transformation to minimize the number of variables for the highest load component. The resulting estimated load matrix reflects the relevance of the variables and principal components. The higher the value of the rotated principal component matrix, the greater the impact of the driving factors on this principal component, and the more easily a change in this principal component will cause changes in the vegetation.

Correlation analysis
We ranked the climatic factors according to their contributions in the PCA. Since some climatic factors may be correlated, we used correlation analysis to test for a linear relationship between any two variables. We divided the absolute value of the correlation coefficients (Pearson's r) into a weak correlation (0 < |r| ⩽ 0.3), a low correlation (0.3 < | r| ⩽ 0.5), a moderate correlation (0.5 < |r| ⩽ 0.8), and a strong correlation (0.8 < |r| ⩽ 1). We then used the t-test to look for significant correlations between driving factors: xy xy 2 where r xy is the correlation coefficient between x and y, and N is the sample size. The critical value (t α ) at different significance levels can be found in standard tables for the distribution of this parameter. If t > t α/2 , the correlation is significant. Combining the PCA and correlation analysis, the climatic factors with high correlations were rejected and the key climatic factors with low correlations were retained to avoid auto correlation. For example, both radiation and temperature contributed strongly to dynamic vegetation change (table 1), but they are highly correlated (table 2) through the PCA analysis and correlation. Therefore, because the radiation parameter had a weaker contribution to the total variance, the temperature parameter was retained as the climate variable for further analyses.

Exploration of regressive modules
Based on the results of the correlation analysis, we explored possible regressive modules for the remaining key parameters to determine the relationships of vegetation dynamics with regional climate change. NDVI was the dependent variable and the key climate factors were the independent variables: where b i is the regression coefficient for key climatic factor (C i ) of factor set i, and ε is an error term. The regression equation can be then used to examine how vegetation responds to changes in climate.

Trend analyses
The decadal trends and the magnitudes of the effects of the key climatic factors estimated by means of least-squares linear regression, Sen's slope trend test, and the Mann-Kendall test score are presented in figure 3. To facilitate visual comparison, the scales of the climate parameters were adjusted by multiplying the precipitation values by 0.1, the mean annual air temperatures by 10, NDVI values by 200, and net radiation by 0.1 as shown in figure 3. The least-squares linear regression and Sen's slope trend test showed similar magnitudes of trends for the climatic factors and NDVI. The three statistical results showed considerable consistency in the trends, indicating a high level of confidence in the trends for vegetation and climate change. Because of its ability to account for extreme values, we chose Sen's slope trend test to estimate the yearly and monthly trends and the magnitudes of vegetation and climate variation.

Correlations between vegetation dynamics and climatic factors
4.2.1. PCA results. The PCA analysis revealed two principal components whose eigenvalues were greater than 1 in the ETP region and the SLP region, with cumulative contributions that explained 96.0% and 97.3% of the total variance, respectively, in the two regions. Table 1 shows the rotated principal-components matrix for the climatic factors in each region. The higher the value in this matrix, the greater the impact of the driving factor on this principal component  Table 2 summarizes the correlations among the monthly climatic factors and NDVI in both regions. There were strong correlations among most of the climatic factors, especially among the air temperatures, with correlation coefficients close to or greater than 0.9. There were also strong correlations among the cumulative precipitation factors, but these factors showed weak correlations with the other climatic factors in both regions.
Based on the results of the correlation analysis, the air temperature factors were all significantly correlated with NDVI and the cumulative precipitation was weakly or nonsignificantly correlated with NDVI. In section 4.3.3, we will discuss the key factors revealed by this analysis and their effect on NDVI.

NDVI responses to climate factors.
From the 11 factors that we initially selected, the PCA and correlation analysis showed that net radiation and air temperature had a stronger contribution to vegetation dynamics than precipitation factors in the TRB. Based on these analyses, we divided the factors that affected vegetation dynamics into two parts: (1) the air temperature and net radiation factors and (2) the precipitation factors. We explored possible regressive modules to determine the response of NDVI to these two groups of factors.
The regression results showed that the air temperature and net radiation factors predicted NDVI very well, whereas the cumulative precipitation did not. However, the sine of the cumulative precipitation predicted NDVI well in both regions, particularly for CP 1 . Based on the degree of independence of the parameters (correlation analysis) and the results of the regressive module exploration, we chose T mean and CP 1 as the independent parameters to predict NDVI and obtained the following: For the ETP: To calibrate these equations, we used data from the assimilated NDVI data series from 1981 to 2000; to validate the equations, we used data from 2001 to 2010. Figure 4 shows the scatterplots for the calibration and validation in the two regions.
The correlation coefficients for the regressive module were greater than 0.85 for both the calibration and validation periods in the two regions ( figure 4). This indicates that the accuracy of the regression equations was high. We used the Nash coefficient to estimate the stability of the two models: for the ETP, the coefficient was 0.77 for calibration and 0.88 for validation; for the SLP, these values were 0.86 and 0.87, respectively. This means that both regression equations were stable and that they can represent the relationship between vegetation dynamics and climatic factors in both regions of the TRB. 4.2.4. Response of vegetation dynamics to regional climate change. Based on the regression equations, we examined Table 2. Analysis of the correlations (r) between climatic factors in the Eastern Tibetan Plateau (shaded area below the diagonal) and the Southwestern Loess Plateau (unshaded area above the diagonal). vegetation dynamics as a function of air temperature and precipitation. The results (figure 5) suggest that air temperature was more important to vegetation dynamics in the ETP than in the SLP. This is understandable because the vegetation in cold regions is typically more sensitive to air temperature than the vegetation in warmer regions. Precipitation was more important in the SLP than in the ETP, which is also reasonable because the vegetation in dry regions is more sensitive to precipitation than in humid regions. During the cold season, the low cumulative precipitation in the beginning of the year and the high cumulative precipitation at the end of the year led to a positive value for sin (CP 1 0.3 ), but a negative value for the corresponding coefficient. The negative value reflects the low NDVI during the early growing season and defoliation of the vegetation late in the year. During the warm season, the moisture and thermal conditions are favorable for vegetation growth, so NDVI in both regions was higher than it was during the cold season. Because evergreen needle-leaved forest grows in the ETP, the lowest NDVI during the cold season is about 0.14, versus 0.04 in the SLP, where there are few evergreen species. These values represent the lowest NDVI levels in the two regions. Due to the combined effects of high air temperatures and cumulative precipitation, NDVI in both regions was highest in July or August. The red solid circles in figure 5 show that under the same moisture and thermal conditions, NDVI was higher in the ETP because of the high background NDVI value due to the presence of evergreen forest. The white arrows represent the approximate direction of the NDVI change between seasons, and the length of the arrows represents the magnitude of the change. In general, the vegetation cover was higher in the ETP, and the annual NDVI change was gentler. The changes in NDVI and the two key climatic parameters were well synchronized during the past 30 years. The air temperature and NDVI both increased while the precipitation decreased in the two regions. However, at time scales of seasons or months, the characteristics of the variation were different.
The seasonal trends for NDVI, T mean , and CP 1 from 1981 to 2010 are summarized in table 3. As indicated by the slopes, there was no vegetation change in winter or summer in both regions. In the ETP, NDVI increased during the spring and decreased during the autumn. This indicates that the growing season may be starting earlier and ending earlier. In the SLP,  NDVI increased in both the spring and the autumn, suggesting that the growing season is starting earlier and ending later. These trends are not necessarily directly related to the length of the growing season, but may instead be based on the background of increasing temperatures in all seasons. However, the trend of decreasing precipitation in all seasons may lead to long-term problems with moisture availability that reduce the effects of rising temperatures on vegetation growth. Determining the start and end of the growing season is beyond the scope of this study, but would be an interesting topic for future research. 4.3.2. Spatial patterns of vegetation dynamics. Using the period from 2001 to 2010 as an example, we can see that NDVI increased in some areas of the ETP and decreased in others. Overall, NDVI increased slightly during this period in this region. In the SLP, NDVI generally increased during the study period ( figure 6). This increase may be related to the 'Grain for Green ' program, which was implemented in 1999(Li et al 2010, Lü et al 2012 to return slope agricultural lands to natural grasslands. Under this program, farmers and livestock herders were encouraged to stop farming or herding their animals on ecologically fragile lands in exchange for government compensation, leading to vegetation recovery in some degraded ecosystems. The decrease in the area of farmland in the TRB under this program has been small (less than 0.5% of the basin's area), and has been concentrated in dry arable regions mainly with rain-fed agriculture. The effect of land cover changes and human activities is therefore expected to be small, and the vegetation dynamics at a decadal scale are likely to be mainly affected by climate change. Based on the relationships between the decadal vegetation change (NDVI) and the two climatic parameters (T mean and CP 1 ), there appears to be a fairly loose correspondence between them. In addition, the trends in the changes of air temperature, precipitation, and NDVI, and the magnitudes of the changes, differ from decade to decade.    (5) and (6)), NDVI appears to be positively linearly related to air temperature in both regions, but negatively related to the sine of cumulative precipitation. The coefficients of air temperature and precipitation have the same sign in the equations for both regions. Numbers and arrow lines in the top of figure 7 describe the average ranges of climatic values for the growing season (autumn included in the summer ranges) in the two regions in the TRB. In the ETP, the coefficient for air temperature is higher than that in the SLP, indicating that vegetation dynamics are more sensitive to air temperature in this region. The higher precipitation coefficient in the SLP indicates that the vegetation in this arid region is significantly affected by precipitation. Because precipitation has decreased in all seasons in both regions since 1981, the increase in NDVI during this period can be attributed entirely to the increasing air temperature. Particularly during the spring, NDVI increased rapidly (by 0.03 and 0.02 per 10 years in the ETP and the SLP, respectively). The annual increase in NDVI has been 0.01 per 10 years in both regions since 1981 under different backgrounds of regional climate change (table 3).
It is noted that using vegetation indicators to assess subtle change of vegetation over large area is challenging as it relies on, e.g., NDVI's sensitivity to vegetation signals and insensitivity to external environmental noises such as atmospheric and soil variations. Given the data availability and the time period of interest in this area, NDVI remains the only choice and its subtle increase (decadal mean value of 0.03) appears to be questionable. However, if we examine its relative change, then the decadal 0.03 is substantial. Annual variation of NDVI, along with corresponding temperature and precipitation is shown in figure 8 for both ETP and SLP sub regions, to allow a visual examination of the annual NDVI changes.

Threshold of spring precipitation for vegetation growth.
The black solid dot on the left diagram in figure 7 is the cumulative spring precipitation threshold (44 mm in this case) required for NDVI to increase (changes of NDVI ⩾ 0). This value represents a reasonable precipitation requirement to support spring vegetation growth. If the amount of precipitation is below this threshold, vegetation growth will be poor due to a lack of sufficient soil moisture. There appears to be little difference in the thresholds for ETP and SLP regions. The background NDVI in winter is especially low in the SLP and thus this threshold is particularly meaningful assessing vegetation dynamics in the region.
The concept of precipitation threshold (44 mm in this study) is important in climate change research, particularly in ecosystem assessment and water resource management in these semi-humid and semi-arid regions. Note that the 44 mm threshold in figure 7 is the threshold beyond which NDVI starts to change (increase in this case), an indicator of onset of vegetation growth. It should also be noted that vegetation growth (onset) is determined by many environmental variables, including climate, snow melting, ground water, soil and topography. The combination of these key factors influences vegetation germination and growth rate in general, but in the study region snowfall in the winter seems to play a dominant role in the following spring growth. In the study area, previous work  indicated frequent droughts in the winter, therefore spring rainfall becomes critical to vegetation growth in the region, in addition to temperature.
It is further noted that ETP and SLP regions have unique climates and are expected to have different precipitation thresholds to trigger spring growth (ΔNDVI > 0). However, the reality is that the two regions were found to have approximately the same (44 mm) precipitation thresholds.

Conclusions
In this paper, we assimilated the AVHRR and MODIS NDVI time-series to produce a comprehensive NDVI time-series for the TRB. We identified the key climatic factors that affected the vegetation dynamics (measured by changes in NDVI) Figure 7. Changes in NDVI as a function of (left) mean air temperature (T mean ) and (right) cumulative precipitation (CP 1 ) in the Tao River Basin. The red lines represent the Eastern Tibetan Plateau and the blue lines represent the Southwestern Loess Plateau. Arrows and numbers show the change in T mean and CP 1 during the indicated season. The solid black dot represents the cumulative spring precipitation threshold required for vegetation to grow. throughout the region by means of PCA and correlation analysis, and conducted the exploration of regressive modules to determine the relationships between NDVI and these factors. Several conclusions can be drawn from this study: First, we were able to successfully assimilate the AVHRR and MODIS NDVI series, as indicated by the small error for pixels during the same period and the high stability of the regression for the two NDVI series, allowing us to create continuous data from 1981 to 2010 for three decades of trend analysis of vegetation dynamics.
Second, the vegetation and climate trend analyses revealed by least-squares linear regression, Sen's slope trend test, and the Mann-Kendall test were generally consistent, suggesting that all three methods can be used to analyze hydrometeorological and NDVI trends in our study area.
Third, based on the PCA and correlation analysis, we identified the mean air temperature and cumulative monthly precipitation since December of the previous year as the key climatic parameters influencing vegetation dynamics in the study area. Regression of NDVI against these climate parameters provided useful clues indicating that climatic variables can predict NDVI in the ETP and the SLP with acceptable accuracy in the TRB.
Fourth, the dominant climatic factor responsible for the vegetation dynamics differed between the two parts of the TRB. In the colder ETP, air temperature had the strongest effect on vegetation dynamics, whereas precipitation was most important in the drier SLP.
Fifth, Sen's slope trend test showed that NDVI has generally been increasing throughout the TRB since 1981, largely as a result of increasing temperatures, despite a declining precipitation trend. The rate of increase in spring NDVI was higher in the ETP sub-region than in the SLP, with values of 0.03 and 0.02 per 10 years, respectively. The results suggest that the vegetation growing season in the ETP is starting earlier and ending earlier, while that in the SLP, the growing season is starting earlier and ending later.
Finally, the NDVI regression equations that we constructed suggest that the threshold of cumulative spring precipitation for vegetation germination is 44 mm. If spring precipitation is below this threshold, there will be a significant negative impact on vegetation growth.