Separating the Impact of Climate Changes and Human Activities on Vegetation Growth Based on the NDVI in China

Vegetation growth is aﬀected by both climate changes and human activities. In this study, we investigated the vegetation growth response to climate change (precipitation and temperature) and human activities in nine subregions and for nine vegetation types in China from 1982 to 2015. The normalized diﬀerence vegetation index (NDVI) and the RESTREND method based on a multiple linear regression model were employed to this end. An overall increasing trend in the NDVI was observed in recent decades, and the fastest increases were identiﬁed in southern China (Trend NDVI � +0.0190) and evergreen broad-leaved forests (Trend NDVI � +0.0152). For > 66% of China, vegetation is more sensitive to temperature and less sensitive to precipitation based on the regression coeﬃcients. The water demand for vegetative growth increased signiﬁcantly from 1999 to 2015 with global warming, especially in parts of the temperate zone. We deﬁned a relative change in the residual trend to quantify the impact of human activities on vegetation. RESTREND NDVI /NDVI in two periods (P1, 1982–1998 and P2, 1999–2015) markedly increased, indicating that human activities play a key role in the reversal of


Introduction
Vegetation impacts the water cycle via various processes, including transpiration [1,2], canopy interception [3,4], and the conservation of water and soil [5][6][7]. In the context of climate change, precipitation and temperature are recognized as the two major climatic factors affecting the biophysical processes of vegetation [8,9]. A warmer climate can increase the length of the growing season [10] and wetter conditions mitigate water demand stress and improve plant productivity in arid and semiarid regions [11]. Under the influence of both climate changes and human activities, the changes in vegetation have become more complicated in various climatic regions [9,12]. With rapid economic development, ecological systems have suffered severe damage through overgrazing and land reclamation in China, for example, in the Qinghai-Tibet Plateau [13] and Inner Mongolian grasslands [14]. Moreover, rapid urbanization and population growth limit the potential growth areas for vegetation.
In 1999, the Grain for Green Project was initiated to reduce the soil erosion and improve local ecological conditions, including for the Loess Plateau, in the northern areas of Shaanxi Province [15], for the upper and middle Yangtze River in Sichuan Province [16], and the typical arid/semiarid rainfed agricultural areas in Gansu Province [17]. e Grain for Green Project has led to the conversion of ∼16,000 km 2 of rainfed agriculture land in semiarid regions to grassland and forests, resulting in an ∼25% increase in vegetation cover of the Loess Plateau from 1999 to 2010. e benefit of both water resource and land-use management (e.g., conservation of water and soil [5,18], water-saving agriculture [19,20], and returning farmland to forests [21]) and climate change (e.g., more precipitation in arid/semiarid regions and temperature increasing in high-latitude regions [22,23]) affects local ecological conditions. Especially entering the 21st century, human activities have played an important role in vegetation change [24,25]. It is necessary to identify effects due to climate changes and human activities at regional scales. Two periods of 1982-1998 (before the Grain for Green Project) and 1999-2015 (after) were analyzed in this study.
With the development of remote sensing, massive amounts of data are widely employed to investigate changes in vegetation [26] and human activities [27,28] at large scales. Since the 1980s, a series of vegetation indices have been developed using remote sensing datasets, for example, the normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI), derived from infrared and near-infrared spectral bands to quantify long-term changes in vegetation growth [29,30]. Several studies have focused on attributing the changes in vegetation using remote sensing means [9,10]. A significant increase in the NDVI is observed in some arid and semiarid regions of China, for example, in Northwest China, due to the vegetation's enhanced sensitivity to wet conditions [31]. A decreasing trend in the EVI was detected in the Yangtze River Basin due to urbanization [9]. Based on previous studies, there is a positive relationship with warming in most parts of China; however, the response sensitivity of vegetation differed with seasons and vegetation types [32,33].
In recent decades, a series of approaches have been developed to quantify the impact of climate changes and human activities on vegetation. e most widely used effective methods can be divided into three categories [15]: (i) regression model-based methods, (ii) biophysical modelbased methods, and (iii) residual trend-based methods. e regression model-based method uses regression statistics to examine the relationship between vegetation change and climate or human-related factors. However, it is unsuitable to distinguish the complex interactions between vegetation and influencing factors (climatic and anthropogenic factors) [34]. Second, the biophysical model-based method is common; however, several physiological and ecological parameters need to be measured or calibrated and the effect of equifinality for different parameters is unavoidable in a complex system [34]. ird, as one of the most widely used methods for separating the impact of climate changes and human activities on vegetation, the residual trend-based method (RESTREND) is a robust approach. e trend in the residuals of multiple regression between climatic factors and vegetation indicators (e.g., NDVI, GPP, and NPP) can quantify the impact of human activities on vegetation.
In this study, we quantify the sensitivity of climate variables in response to vegetation growth in a warming climate across various climatic regions and different vegetation types in China. To this end, MLR and the residual trend (RESTREND) analysis method were employed to separate the impact of climate changes and human activities. e rest of the article is organized as follows: in Section 2, we describe the long-term meteorological and remote sensing datasets and statistical methods. In Section 3, we report the assessment and attribution results of long-term changes in vegetation from 1982 to 2015. We present our discussion and conclusions in Section 4 and Section 5, respectively.

2.
1. Data Sources. GPCC datasets are recommended for hydro-meteorological model verification and water cycle studies owing to the high data quality and long-term time series. In this study, we selected the monthly precipitation GPCC datasets (0.5°× 0.5°longitude by latitude) covering the mainland of China from 1982 to 2015. is analysis is based on meteorological stations established before 1900, and GPCC datasets span the period from 1891 to 2016. In this study, the period from 1982 to 2015 is covered. e coincident monthly spatial resolution (0.5°× 0.5°longitude by latitude) and the period (1982 to 2015) air temperature datasets, and the Hadley Centre and Climatic Research Unit HadCRUT4 datasets including 3839 grid boxes were used in this study.
Semimonthly, the long-term NDVI of GIMMS NDVI3g datasets were used. GIMMS NDVI3g datasets are global vegetation index change data provided by NASA (https://ecocast.arc.nasa.gov/data/pub/gimms/3g.v1/). e dataset is preprocessed by data fusion, radiometric correction, geometric correction, bias correction, and image enhancement. It is commonly used in the study of regional and global vegetation change owing to the long time frame and homogeneous series [10,35]. In this study, we obtained GIMMS NDVI data from January 1982 to December 2015.
To minimize the influence of clouds, atmosphere, and monthly phenology, which is considered a possible source of noise, the maximum value composite method was adopted to merge the datasets from the semimonthly to monthly scale. To ensure consistent spatial resolution for precipitation, air temperature, and NDVI, we uniformly aggregated to 0.5 using the nearest-neighbor resampling method. e land cover map for 2015 from the European Space Agency Climate Change Initiative (https://maps.elie.ucl.ac. be/CCI/) was also re-gridded to the 0.5 grid using the nearest-neighbor method and used to classify vegetation types. e nine subregions of China are shown in Figure 1(a). We reclassified the 38 land cover types into nine categories, namely croplands-CRO, evergreen broadleaved forests-EBF, deciduous broad-leaved forests-DBF, evergreen needleleaf forests-ENF, deciduous needleleaf forests-DNF, mixed forests-MF, shrublands-SHR, grasslands-GRA, and non-vegetation regions-NOV (Figure 1(b)).

Methods
e linear regression equation is expressed as follows: where Y t is the dependent variable in the year t (Y is the annual series of NDVI for a given grid box), X is the independent variable (X is the annual series from 1982 to 2015 2 Advances in Meteorology for a given grid box), a and b are the slope and the intercept, respectively, and ε is the residual for the regression equation. e slope of NDVI is given as follows: e statistic value S of the Mann-Kendall test is defined as follows: where e variance of S is as follows: e statistic Z is defined as follows: Based on a confidence level of 0.05 using the Mann-Kendall test, if |Z| ≥ 1.96, the statistical value of the trend is significant [36].
MLR was used to quantify the contribution of meteorological variables. In this sensitivity analysis, sunshine duration (solar radiation) was eliminated because of the multicollinearity with precipitation [9]. To compare two independent variables, we first normalized the annual series of all variables (both dependent and independent) as a dimensionless series with a large sample size (μ � 0; σ � 1): where x is the original annual series, that is, P, T, and NDVI, and X is the dimensionless series. e regression coefficient (hereafter as RC) obtained from the MLR is an appropriate measure for identifying a variable as being independent. e RESTREND method examines the change in residual differences between the observed and predicted annual NDVI based on the MLR model using climate variables (e.g., temperature and precipitation) as independent parameters [37]. To distinguish this influence before or after the Chinese Grain for Green program, the long-term climatic change was divided into two periods (Period 1, 1982-1998 and Period 2, 1999Period 2, -2015. e annual NDVI residual trend was calculated based on MLR with the annual air temperature and precipitation. e calculation of the RESTREND method is as follows: where P is the annual precipitation, T is the annual temperature, and ε is the residual based on the MLR model. en the RESTREND can be quantified by linear regression: where years is a given period (1982-2015, 1982-1998, or 1999-2015) corresponding to the residual of ε.

Sensitivity Analysis for Vegetation Using MLR.
To explore the sensitivity of vegetation to precipitation (indicating water stress) and temperature, a multiple regression model was designed using standardized NDVI (dependent variable) and standardized climatic variables (precipitation and temperature as independent variables) merged into an annual time series to remove the lagged response of vegetation within monthly periods.
To eliminate the interference factors in a non-vegetated area, we extracted the regression coefficient of precipitation (RC P ) and temperature (RC T ) to annual NDVI without counting the grid boxes of the non-vegetated land cover type at grid scales (Figure 3(a)). On the whole of China, 70.2% of grid boxes indicated that the vegetation is more sensitive to temperature (|RC P | < |RC T |). e contribution (regression coefficient) of precipitation to vegetation trends has a large spatial difference at grid scales, ranging from −0.60 in the north of Northeast China to +0.69 in Eastern Inner Mongolia (Figure 3(a)). e regression coefficient of temperature showed a different spatial pattern, ranging from −0.37 in the south of the Tibetan Plateau to +0.79 in the north of Central China over the same period (Figure 3(a) and 3(b)).
For the nine subregions of China, we note a significantly positive response of vegetation to precipitation (an average RC P over +0.1) in nonhumid regions of China, for example, Inner Mongolia (RC P � +0.302), North China (RC P � +0.270), and Northwest China (RC P � +0.140). e response of vegetation to precipitation is relatively insensitive in the other subregions (the average RC P around zero, greater than −0.1 and less than +0.1). By contrast, the significantly positive response of vegetation to temperature (an average RCP over +0.1) was found in all subregions of China, especially in Central China (RC T � +0.547), North China (RC T � +0.426), the Tibetan Plateau (RC T � +0.438), and South China (RC T � +0.399).
Using multiple linear regression (MLR) analysis, we obtained the regression coefficient of precipitation with different land covers (vegetation types). On average, the regression coefficient of precipitation ranged from −0.205 in deciduous needleleaf forests to 0.144 in grasslands. e deciduous needleleaf forests, with a negative response to precipitation, are the most sensitive land-use type (RC P � −0.205). As shown in Figure 3(b), increased precipitation also favors vegetation increases in other individual land-use types (e.g., evergreen broad-leaved forests, RC P � −0.017; deciduous broad-leaved forests, RC P � −0.094; evergreen needleleaf forests, RC P � −0.059; shrublands, RC P � −0.143). On average, the regression coefficient of temperature ranged from 0.149 in deciduous needleleaf forests to 0.405 in evergreen broad-leaved forests. Compared with the regression coefficient of precipitation, we found that the absolute value of the regression coefficient for temperature is significantly higher in most land-use types, except for deciduous needleleaf forests (|RC P |>|RC T | in the deciduous needleleaf forests in China). is means that the response of vegetation to precipitation is significantly weaker than that of temperature, except for deciduous needleleaf forests (Figure 3(b)).

Contribution of Climate Changes and Human Activities to Interannual Variation of the NDVI in the Two Periods.
With dramatic global warming occurring at global and regional scales, quantitatively determining the contribution of climate changes and human activities to the NDVI is essential for understanding the trends in vegetation change. Previous studies have highlighted that vegetation growth is relatively sensitive to climate changes (e.g., temperature and precipitation), and the influence of human activities on vegetation growth must not be ignored [9,38]. e correlation coefficient of the MLR model (CC) mentioned in Section 3.2 can indicate the impact of climate changes on vegetation with global warming. Here, we prepared a longterm annual temperature series spanning 1982 to 2015 (Figure 4). We noted that the annual temperature slowly Advances in Meteorology increased in Period 1 (P1), from 1982 to 1998, with a mean annual temperature of 6.66, followed by a dramatic increase after 1998 and decadal warming hiatus early in the 21st century. Period 2 (P2) from 1999 to 2015 was characterized by a higher annual temperature (7.20°C), an intensification of global warming. e MLR model for quantifying the sensitivity of vegetation to climate was also applied to P1 and P2, and then the RESTREND method was used to quantify the contribution of climate changes and human activities in the periods. e widely used "binning method" was applied to reveal the pattern and characteristics of the vegetation sensitivity to climate changes in different precipitation or temperature conditions. We stratified the annual NDVI data based on the annual precipitation (or air temperature) in bins of 100 mm, from 100 mm for arid region cases to >1700 mm for humid regions (also 2°C bins, from -6°C for sub-frigid regions to 24°C for tropical regions). Here, we focus on each bin greater than 10 grid boxes.
First, we integrated the regression coefficient of temperature (RC T ) and precipitation (RC P ) for each temperature and precipitation bin in P1 (Figure 5(a)-5(c)). In this period, the vegetation growth was more than twice as sensitive to temperature as to precipitation, especially in the bins above 12°C or below 4°C. e negative response to precipitation (RC P ) was found in the bins below 0°C or above   16°C, and a significant peak occurred between 4°C and 8°C. In this period, RC P decreases with an increase in precipitation, and the RC P is larger/less than zero in nonhumid/ humid regions (nonhumid/humid regions are the areas where the mean annual precipitation is larger/less than 800 mm). e results indicate that more precipitation could boost vegetation growth in nonhumid regions, whereas excessive precipitation has a negative effect on vegetation growth in humid regions in this period. We also investigated both RC T and RC P for each bin in P2 ( Figure 5(b)-5(d)). During this period, there is more sensitivity to precipitation than the temperature in arid/ semiarid regions (mean annual precipitation is below 400 mm) or temperate regions (mean annual temperature is more than 8°C and less than 12°C) ( Figure 5(b)-5(d)). With global warming, the positive response to precipitation was found in all temperature bins (except for the -4 to -2°C bin). Precipitation is more sensitive than temperature to vegetation growth in some regions, that is, where the mean annual temperature is between 8 and 12°C.
Compared to the probability density function of the correlation coefficient with the mean of 0.408 and standard deviation of 0.168 from 1999 to 2015 (a period of heightened global warming), we found a significantly higher correlation coefficient (CC) with the mean of 0.494 and standard deviation of 0.185 from 1982 to 1998 at the grid box level (Figure 6). e results support that the influence of human activities on vegetation growth became more obvious in the later period. e MLR model, which cannot be explained by climate changes and is usually interpreted as the effects of human activities, increased from 1999 to 2015. erefore, it is necessary to explore and separate the impact of human activities on vegetation growth using the RESTREND method.
To quantify the influence of human activities on vegetation growth, we define a relative change in the residual trend, which is the ratio of the change in the residual trend in the NDVI to the Long-time average annual NDVI value (defined as RESTREND NDVI /NDVI, unit: (%)·10a −1 ). e value of relative change in the residual trend (RESTREND NDVI /NDVI � 1.84%·10a −1 ) indicates that human activities played a key role in the reversal of land degradation during P1. e spatial distribution of the residual trend showed large regional differences in P1 (Figure 7 For P2, the results indicate that the influence of human activities on vegetation is more evident (Figure 7(b)). ere was a significant increase in the relative change in the residual trend in the NDVI of all of nine subregions (with the highest positive values in North China, RESTREND NDVI /NDVI � 6.120%·10a −1 ), except for Southeast China which had a slight negative effect from human activities on vegetation growth (RESTREND NDVI /NDVI � -0.299%·10a −1 ). Interestingly, the brown points (degradation in vegetation variation) surrounded by green are the megacities of the Pearl River Delta and Yangtze River Delta in P1 (Figure 7(a)) and the megacities of Chengdu, Xi'an, Zhengzhou, Beijing, Tianjin, Nanjing, Qingdao, and Shanghai in P2 (Figure 7(b)).
is suggested that rapid urbanization is one of the causes of vegetation degradation.

Climate Changes Sensitivity during the Two Periods.
Before the Chinese Grain for Green program was implemented, the vegetation in nonarid areas had a negative response to precipitation as shown in Figures 3 and 5. A more plausible explanation is that temperature and radiation were the limiting factors of vegetation. Precipitation in drought years could satisfy the basic needs for vegetation growth in P1 in semiarid and humid regions of China, because of lower transpiration [39,40]. With increased vegetation growth and the associated increase in transpiration, the water demand will consequently increase. According to the statistical results from the regression coefficients of multiple linear regression, the sensitivity of vegetation to precipitation increased significantly with vegetation increases in P2. Jiao et al. pointed out that vegetation water constraints are associated with greening trends, which leads to the vegetation becoming increasingly sensitive to precipitation [41]. is finding may explain that the negative regression coefficients of precipitation in P1 shifted to positive coefficients with greening trends in the most subhumid and humid regions in P2.

e Impacts of Human Activities on Vegetation Change.
Previous studies have highlighted that vegetation on the Qinghai-Tibet Plateau has declined significantly due to excessive human interference. Using the RESTREND  Advances in Meteorology method, as shown in Figure 7, we identified the vegetation degradation in P2 [42]. More than 75% of the new cropland has been cultivated at the expense of forest from 1980 to 2000 across the tropical regions. e Grain for Green program induced a considerable reduction in the cultivation of low productivity crops, and any area taken out of cultivation was offset by putting at least an equal additional area into cultivation [43]. e rapid expansion of urbanization is also an important factor leading to vegetation degradation even with the GFG project. One study showed that urban areas increased by 24.6 × 10 3 km 2 (0.26% total area of China), cropland decreased by 4.9 × 10 3 km 2 (0.05% total area of China), and the total area of forest and grassland decreased by 16.4 × 10 3 km 2 (0.17% total area of China) with urbanization from 2000 to 2015 [44]. In this study, the negative effects of urbanization are shown in Figure 7(b), which is in agreement with previous studies [44,45].

Limitations and Future Research Directions.
Due to the lack of meteorological observation stations in the Qinghai-Tibet Plateau and Northwest China, the global grid meteorological datasets (GPCC and CRU) were used in this study. However, Liang et al. highlighted that a large difference between observation data and global grid data could bias results [46]. is means that the accuracy of the results needs to be further verified using observation datasets. e sensitivities of temperature and precipitation to vegetation growth were quantified at an annual scale. Solar radiation as the energy source of vegetation photosynthesis is also one of the important meteorological elements. Besides, the lag of vegetation response to climate changes should be considered at monthly/seasonal scales [47,48].

Conclusions
In this study, we focus on vegetation responses to climate changes and human activities in various climate areas or land types from two periods (P1, 1982-1998 and P2, 1999-2015) at the grid box level (3839 grid boxes) in China. Nine subregions and nine categorized land covers (vegetation types) were defined. We analyzed the vegetation trends using the NDVI from the whole period of 1982 to 2015. Significant greening trends were detected in most regions, especially in Central China (Trend NDVI � +0.0190·10a −1 ), Southwest China (Trend NDVI � +0.0165·10a −1 ), and South China (Trend NDVI � +0.0147·10a −1 ). en, we explored the sensitivity of vegetation to climate variables (precipitation and air temperature) using the MLR model. We found that the vegetation is more sensitive to temperature in 70.2% of the areas, and a positive response of vegetation to precipitation was found in nonhumid regions. A binning method was used to reveal the characteristics of the sensitivity of vegetation to climate changes in various climate conditions during the two periods. According to the regression coefficient of the MLR model, the impact of precipitation on vegetation increased significantly in P2 with the intensification of global warming.
is indicates that the water demand for vegetation increased significantly in P2.
e RESTREND method demonstrates its effectiveness in the quantification of vegetation response to climate changes and human activities.

Data Availability
All data and models used during the study appear in the submitted article.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this article.