Human activities modulate greening patterns: a case study for southern Xinjiang in China based on long time series analysis

The study area is southern Xinjiang, Xinjiang Province, China (longitude 73.5° E–93.8° E, latitude 34.3° N–43.6° N; figure 1). Southern Xinjiang is dominated (>60%) by sand and bare land/rock (the Taklamakan Desert) and the mean annual potential evapotranspiration (>1000 mm) is much higher than precipitation (about 100 mm or less). The main natural vegetation is grassland (approximately 26% of the total area), whereas oases (approximately 3% of the total area) are mostly covered by irrigated crops such as cotton.

Zoom In Zoom Out Reset image size

The global inventory modeling and mapping studies (GIMMS) datasets of third-generation NDVI_3g (Pinzon and Tucker 2014) the USA National Oceanic and Atmospheric Administration (NOAA) climate data record (CDR) of advanced very high resolution radiometer (AVHRR) NDVI dataset, version 5 (Vermote et al 2019), and LAI_3g (Zhu et al 2013) from 1982 to 2015 were used in this study (table 1). These data provide the longest available vegetation index data on a global scale (the first year 1981 with half-year data availability was excluded). The NDVI_3g time series is an improved 8 km NDVI dataset produced by AVHRR instruments (Pinzon and Tucker 2014). The CDR NDVI dataset contains gridded daily NDVI with a longer period to present, derived from the NOAA AVHRR surface reflectance product at a resolution of 0.05° and computed globally over land surfaces (Vermote et al 2019). The LAI_3g dataset was generated by combining GIMMS NDVI_3g and moderate resolution imaging spectroradiometer (MODIS) LAI products with the feed-forward neural network algorithm (Zhu et al 2013).

In order to describe greening patterns, i.e. greening rates in different vegetated lands, the China's multi-period LULC remote sensing monitoring data set (Xu et al 2018) was used (table 1). These data are available for the period 1990–2020 at a 5 years interval and a spatial resolution of 1 km and have already been used in comparative regional studies of greening and browning (Zhang et al 2020).

Monthly NDVI and LAI were composed by the maximum value composite method (Du et al 2015) to minimize cloud contamination and were then averaged over the growing season for each year as annual mean NDVI and LAI. As the growing season likely increases due to warmer climate, for example, earlier leaf emergence or shifts in sowing and harvest dates (Zeng et al 2011, Fridley 2012, Park et al 2016), we fixed the growing season from April to November to cover most crop types in southern Xinjiang, a region experiencing a significant temperature increase since 1982. The annual mean NDVI and LAI data were resampled to the 1 km spatial resolution of the LULC data. As a result of paired comparison of two NDVI data and LAI data, pixels with annual NDVI lower than 0.1 for NDVI_3g, annual NDVI lower than 0.13 for CDR NDVI, or LAI equal to 0 were regarded as non-vegetated land in respective data processing and thus not considered in the analysis.

The original LULC dataset was reclassified by merging similar classes of cropland, grassland, waterbody, built-up area, and other lowly vegetated lands (called 'others', including sand, Gobi, saline-alkali land, and bare land/rock) and omitting classes with an area less than 1% of the total pixel area. Finally, four major LULC types were used, i.e. cropland, forest/shrubs, grassland, and others. Because of the LULC dynamics, the majority composite method (i.e. the most frequent LULC type of each pixel) was used to assign a representative LULC type of each pixel and compose a representative map for the 1990–2020 period (figure 1). The representative LULC was used for identifying the greening patterns of different LULC types because the differences between each year's LULC and representative LULC were small (data not shown).

In medium- to long-term analyses of satellite data, linear regression modeling can help to reveal the directions and the rates of land surface greenness change (de Jong et al 2012, Liu et al 2018). Positive and negative slopes indicate greening and browning trends, respectively, and their absolute values depict the rates during the study period:

$$\begin{equation}{\text{Slope}} = {\text{ }}\frac{{n \times \sum\nolimits_{i = 1}^n {{x_i}} - \sum\nolimits_{i = 1}^n i \sum\nolimits_{i = 1}^n {{x_i}} }}{{n \times \sum\nolimits_{i = 1}^n {{i^2}} - {{\left( {\sum\nolimits_{i = 1}^n i } \right)}^2}}}\end{equation} \tag{ 1 }$$

where ${\text{Slope}}$ is the changing rate of $x$ (either NDVI or LAI); $n$ is the length of time series, i.e. number of years and $i$ is year.

In order to evaluate how significant the NDVI and the LAI trends are based on the 1982–2015 data, we used F-test (equations (2) and (3)) and deemed a significant increase (greening trend; P < 0.05), no significant change (stabilization; P > 0.05) and a significant decrease (browning trend; P < 0.05):

$$\begin{equation}F = \frac{{\sum\nolimits_{i = 1}^n {{{\left( {{x_i} - \bar x} \right)}^2}} }}{{\sum\nolimits_{i = 1}^n {{{\left( {i - \bar i} \right)}^2}} }}\end{equation} \tag{ 2 }$$

$$\begin{equation}p = P({\text{Fv}}\left( {{\text{df}}1,{\text{df}}2} \right) > F)\end{equation} \tag{ 3 }$$

where p is the possibility of the item change trend according to the F-test statistic, F value; Fv(df1, df2) is the random variable with an F distribution and degree of freedom: df1, df2.

For more specific trend detection in NDVI and LAI time series, the locally weighted scatterplot smoothing (LOWESS) regression method was used, with model parameters determined empirically (Chu et al 2018). It is a regression tool used for analyzing time series by fitting a smooth line through the time scatter plot, assessing the relationship between variables, and foreseeing trends. The procedure involves setting a smoothing parameter, i.e. a fraction of the total number n of data points used in each local fit (set to 0.49), and iterative reweighting (set to 3) to provide robust fitting when there are outliers in the data.

In order to distinguish between greening and browning and their trends in different vegetated lands, two area indexes were compared between vegetated classes, i.e. relative area (RA) and contribution to greening (C; equations (3) and (4)), the former showing the area percentage of greening and browning in each LULC type and the latter depicting the area percentage of each LULC type in increased NDVI or LAI:

$$\begin{equation}{A_i} = {\text{s}}{{\text{p}}_i} \times n \times {\text{res}}\end{equation} \tag{ 4 }$$

$$\begin{equation}{C_i} = \frac{{{A_i}}}{{\mathop \sum \nolimits_i {A_i}}} \times 100\% \end{equation} \tag{ 5 }$$

where $\,{A_i}$ is the sum of increased NDVI or LAI in the LULC type; ${\text{s}}{{\text{p}}_i}$ is the summed slopes of greening area in the $i{\text{th}}$ LULC type; $n$ is the length of time series; ${\text{res}}$ is the spatial resolution ($1 \times 1\,{\text{km}}$).

The influence of human activities on the greening patterns was of particular interest in this study. The relation between greening rate and distance to human-dominated areas was investigated. The LULC map in 1990, i.e. the first year of available LULC data, was used as the baseline map to estimate the distance to human-dominated areas. A series of buffers ranging from 1 to 40 km with 1 km increment near cropland and built-up area were designed in a Geographic Information System (GIS) environment. The proportion of area (POA) with greening, stabilization, and browning lands was calculated within each 1 km buffer. Consequently, the POA in greening and browning lands was shown as a function of the distance from the human-dominated area.

The spatial distribution of the mean and the trend values based on the annual data from 1982 to 2019 for NDVI and to 2015 for LAI are shown in figure 2. Compared to LAI, vegetated area was larger according to the mean NDVI data and corresponded well to the GIMMS NDVI result (figure S1 available online at stacks.iop.org/ERL/17/044012/mmedia) based on the threshold 0.13. Greenness according to NDVI appeared mostly in the farming areas nearby rivers or reservoirs mainly in the north and the west of the study area (figures 2(a)–(c)). The greening processes were mostly concentrated in the oases (figures 2(b)–(d); consisting mostly of farming areas, while browning processes, i.e. NDVI trend values lower than zero, were scattered in the northern and the southern grasslands (figures 2(b)–(d)). The area under greening, stabilization (no change), and browning covered, respectively, 52.8%, 40.2%, and 7.1% of the total vegetated areas according to NDVI, and corresponding values according to LAI were 51.8%, 45.1%, and 3.2%.

Zoom In Zoom Out Reset image size

Cropland increased rapidly in the study region at the cost of decreasing mostly grassland, whereas the other LULC types (including land surfaces of sand and desert, saline-alkali land and bare land/rock) were relatively unaffected throughout the past three decades (table 2). The NDVI and the LAI results both showed comparable trends according to both the linear and the LOWESS regression models even during the different periods (figures 3 and S2), with different fluctuations throughout the past 30–40 years. Vegetation coverage increased significantly (P < 0.01) in southern Xinjiang as a whole and a three-section trend could be observed from the LOWESS regression results, i.e. no change from the 1980s to 1990s, significant greening from the mid-1990s to 2010, and a shift to browning after 2010. The greening in cropland appears to have happened much faster compared to the other LULC types, while the greening of the other sparsely vegetated covers was unstable and minor.

Zoom In Zoom Out Reset image size

The relation between vegetation coverage and vegetation change rates showed that areas with higher NDVI and LAI had higher greening rates (figures 4(a)–(d)) and the 'green' LULC types (cropland, forest/shrubs, grassland) had faster greening (figures 4(b)–(e)) compared to the others. However, forest/shrubs and grassland had relatively low greening rates compared to cropland, which was even smaller than others according to the LAI data analysis. The regression coefficient between the slope of NDVI and LAI and their mean value was lower (figure 4(c)), and the significance level was also reduced after excluding the cropland pixels (figure 4(f)). Therefore, the greening in southern Xinjiang could be described with spatial clustering as 'greening in green land', and cropland playing an important role in the rapid greening process.

Zoom In Zoom Out Reset image size

Greening was the main status of cropland and forest/shrubs according to the NDVI analysis while both greening and stabilization covered nearly half of the grassland area according to the NDVI and the LAI analysis (figure 5). Despite certain discrepancies in the RAs, stabilization and browning trends covered >50% areas of grassland and other LULC. Cropland and grassland contributed the most to the increased vegetation indexes. Specifically, grassland played a more important role in NDVI increase, which was well reflected in the two NDVI data (figures 5(b) and S3), while cropland played an important role in LAI increase (figure 5(d)). Given the small area of cropland (around 3%) compared to grassland (26%) among vegetated areas, the former had a notably larger NDVI and LAI increase per unit area compared to the latter. On the other hand, most browning areas were covered by grassland or others (about 99% in NDVI and 87% in LAI in total).

Zoom In Zoom Out Reset image size

The frequency distributions of the long-term slopes according to NDVI and LAI confirmed the distinct patterns for different LULC types and it was fitted using kernel density curves (figure 6). Cropland showed a widespread NDVI and LAI increase with larger slopes, compared to the other vegetation types (figures 6(a)–(e)). Forest/shrubs showed higher variance of the change slopes than the other vegetation types with two peaks for NDVI and the median slope was also positive (figures 6(b)–(f)). In contrast, more slopes of NDVI and LAI were near zero for grasslands (figures 6(c)–(g)) and negative for other areas (figures 6(d)–(h)). The GIMMS NDVI results showed more slopes concentrated near zero for grassland and other areas (figures S4(c) and (d)), and cropland showed a relatively dispersed distribution of slopes greater than zero (figure S4(a)).

Zoom In Zoom Out Reset image size

Since 1990, greening was more evident near human-dominated areas and it decreased with distance (figure 7). According to both the NDVI and the LAI data, greening areas had the largest POA (around 90%) within 2 km of human-dominated area. The proportion of greening decreases dramatically as the distance increases until around 15 km, whereas POA with browning increased slowly with distance up to 15 km for NDVI and 5 km for LAI. Therefore, there was significantly more greening and less browning near human-dominated areas. Despite a 'discrepancy' in minimum points between NDVI and LAI results, which resulted in different distances of minimum points (the first minimum point occurred in 14 and 6 km according to NDVI and LAI data, respectively), the minimum proportion of greening trends was similar of approximately 25%. There were fluctuations after the first minimum point in both NDVI and LAI results, but as a whole, the ratio of greening, stabilization, and browning trends was around 4:5:1 for a long distance ($>$25 km) from the human-dominated area.

Zoom In Zoom Out Reset image size

The greening in southern Xinjiang study area had a strong and significant correlation with human activities, but is also influenced by other indirect factors (Zeng et al 2017, Piao et al 2020), such as climate change and atmospheric influences (CO₂ fertilizer effect and atmospheric nitrogen deposition; Piao et al 2020). The general greening trend over different LULC types suggested that greening was affected by these indirect, and most likely large scale, factors to different degrees.

Climatic effects (precipitation, temperature, water balance) were, however, investigated and no significant effect of precipitation was found for the past 40 years in the vast majority of the study area (results not shown; see e.g. Pu et al 2011). In addition, the temperature had also little influence on the vegetation coverage due to the limited water resources available throughout most of the year. Atmospheric factors might drive the greening trends on a much larger scale and play an important role in time-series fluctuations. However, the effect of atmospheric factors can be assumed homogeneous on the studied scale (Piao et al 2015). The greening difference among the LULC types will minimize the effects of the large scale and indirect factors as well because LULC types are spatially mixed. The NDVI and LAI time series fluctuations are to some extent caused by atmospheric factors, but the difference in greening rate between adjacent areas should be largely attributed to human activities. Therefore, the effects of human activities can be separated in our analysis from other factors to examine the modulation of the greening patterns.

The greening patterns in southern Xinjiang during the past decades can be summarized as larger greening trends in the 'green land' and greening accelerated by continuously increasing crop production. Cropland was the main vegetation contributing to the greening (figures 2 and 5) and it was the LULC type with the fastest greening rate (figure 4), which reveals the increasing farming intensity in the study area. Southern Xinjiang has experienced significant cropland expansion at the cost of grassland shrinkage from 1990 to 2020, in particular, since 2000 (table 2; Peng et al 2011, Ning et al 2018), and the agricultural center shifted to southern Xinjiang in recent years (Gao et al 2019, Cai et al 2021). The LULC change in the study area was mainly attributed to the transformation of prevailing institutions and policies in China (Lambin et al 2003, Zhang et al 2015), from a 'command' to a market-oriented economy (Tisdell 2009), which intensified and mobilized the agricultural production, alongside a steadily improving socio-economic condition of the local population. Additionally, the vegetation near human-dominated areas showed more greening areas, which indicates that agricultural production had greatly supported the greening process in southern Xinjiang. Therefore, economic development and population increase had a strong relation with the heterogeneous greening patterns, reflected as the rapid greening cropland and mixed grassland changes. Noteworthy is that crop- and grassland was always the 'surrounding' environment of the built-up area (figure 1), which formed agricultural oases dominated by farming and animal husbandry. More browning areas within 10 km (figure 7) from human-dominated areas may indicate natural vegetation degradation, grassland in particular.

The heterogeneous greening in southern Xinjiang showed that non-climatic factors- human activities, affected the ecological processes associated with LULC changes, for example, agricultural management and human-dominated area expansion. Beyond the warmer and wetter climate that provided better conditions for vegetation growth in NW China, the social and economic conditions are the dominant factor of vegetation changes in this hyper-arid region. Greening in human-dominated areas could be implemented by two means, as previous studies argue, i.e. reclaiming or transforming other vegetation covers, in particular grassland (Ning et al 2018, Cai et al 2021). The potential ecological effect of fast agricultural development is causing water pressure in southern Xinjiang. Distinct patterns of greening in cropland and grassland indicate the great disequilibrium of water resources between crops and natural vegetation. Water source varies across cropland and natural vegetation cover, i.e. crops rely on irrigation water, while natural vegetation relies on precipitation, ice/snow-melt water, or deep groundwater. The region under human control shows more greening processes may be related to the larger appropriation of water resources near human-dominated areas. In many arid regions in China and globally, water consumption for irrigation is even higher than that in wet regions (Deng et al 2006, Liang et al 2015). Thus, concentrated greening induced by the increasing farming intensity and expansion of cropland will enhance the water pressure in the hyper-arid southern Xinjiang. In addition, the slight area decrease of forest/shrubs and expansion of the others (table 2) may agree with desertification in the oasis of the Tarim Basin (Jiang et al 2019).

This study is subject to uncertainty, mainly derived from scale issues of the remote sensing datasets and oversimplification of the system complexity, which may be the direction of future research. Firstly, the different resolutions between the NDVI and the LAI datasets, and the LULC datasets induce (a) slight grid discrepancies, i.e. few pixels with various LULC types to have assigned one NDVI or LAI value; and (b) the slight discrepancy of linear trends, in particular, the large decrease of NDVI after 2018. This may introduce uncertainty in the estimates of the slopes of NDVI or LAI of each LULC type. Secondly, different results for grassland and the other LULC types are caused by the differences of NDVI and LAI to reflect vegetation coverage and processes. According to the results, NDVI filtered by a specific threshold may have better vegetation identification under the coarse growing seasons present for grassland and other lowly vegetated areas (figures 2(a) and S1), whereas LAI data may underestimate the vegetation cover in these areas (figure 2(c)). Accordingly, other results had some minor discrepancies due to both distinct periods and different data processing of providers (e.g. the significant discrepancy of turning points for the greening proportions in figures 7 and S5). Finally, a recent study of analyzing spatial and temporal data (such as NDVI and LAI) recommended that auto-correlation induced by the commonly used method may overestimate the vegetation increase (Cortés et al 2021). The standard trend method of linear regression and significance-test can depict the directions and relative rates of the changes well in a long-term analysis compared to the LOWESS regression (figure 3), while the absolute increases of NDVI and LAI were not fitted well.