Spatial patterns of long-term vegetation greening and browning are consistent across multiple scales: Implications for monitoring land degradation
Abstract
Understanding spatial–temporal patterns of terrestrial vegetation response to climate change (long-term greening/browning) is important for developing strategies to mitigate degradation. Semiarid rangelands are especially susceptible to degradation, which challenges wildlife conservation and human livelihoods that depend on livestock production. In the cold-arid Trans-Himalayan ecosystem (northern India), temperature is increasing, and it is also becoming progressively wetter. Yet, counter-intuitively, there are widespread concerns over degradation. We evaluated whether greening/browning patterns in long-term satellite-derived vegetation indices (normalized difference vegetation index [NDVI]) are consistent across different spatial and temporal scales using 6 datasets: MODIS (250 m, 500 m, 1 km, and 5.5 km), SPOT 1 km, and GIMMS 8 km. Results indicate browning in the spring and greening in late summer. Location of hotspots of degradation (browning) was broadly consistent across spatial scales (10−2–102 km2) and were found in regions with warmer temperature and at higher elevations. Broadly, the spatial/temporal pattern of browning does not coincide strongly with location and timing of human land use via livestock grazing. This geographical and seasonal context indicates that vegetation response may be more strongly related to climate than to human land use (overgrazing). Importantly, the dynamic nature of greening/browning, across space and time, is not captured by composite annual metrics (sum-NDVI, max-NDVI, and mean-NDVI). This reiterates the importance of both intraannual and interannual assessments. Location of hotspots indicates that degradation occurs in a spatially contiguous manner, but these are not stationary and instead shift with seasons. Overall, the results show that evaluating the consistency of greening/browning trends across different spatial/temporal scales is critical for understanding and managing vegetation degradation.
1 INTRODUCTION
Terrestrial vegetation response to ongoing and projected climate change is of critical importance to global biogeochemical feedbacks, biodiversity, and human livelihoods derived from ecosystem services. In recent decades, significant global alterations in vegetation include poleward latitudinal range shifts (Parmesan & Yohe, 2003), altitudinal shifts (Lenoir, Gégout, Marquet, de Ruffray, & Brisse, 2008), change in biomass (Myneni, Keeling, Tucker, Asrar, & Nemani, 1997), advancing spring phenology (Cleland, Chuine, Manzel, Mooney, & Schwart, 2007), and delayed autumn phenology (Menzel & Fabian, 1999). Of these, change in biomass and phenology are key variables for ecosystem monitoring, particularly for concerns over degradation (Reynolds et al., 2011). Studies on satellite-derived metrics of long-term trends in vegetation status, such as the normalized difference vegetation index (NDVI) or related indices (Pettorelli, 2013; Pettorelli et al., 2005), reveal high heterogeneity in responses: vegetation is increasing in some regions (greening), but declining (browning) in others (Zhu et al., 2016).
Understanding this inherent heterogeneity in greening/browning, globally and regionally, is key for adapting to climate effects on vegetation (Fensholt et al., 2012; Zhu et al., 2016). Vegetation change in arid and semiarid regions raises important concerns over degradation and desertification (Fensholt et al., 2012; Turnbull, Wainwright, & Brazier, 2008), as well as loss of soil fertility (Chandregowda, Murthy, & Bagchi, 2018). Such subhumid ecosystems comprise over 40% of the terrestrial realm, house important biodiversity, and support the livelihoods of over 1 billion people (Briske, 2017). However, there is high uncertainty over the extent and severity of degradation (browning) in many arid and semiarid regions. Conclusions of many studies are either counter-intuitive or are mutually dissonant. Formulating managerial interventions and policy decisions over degradation can be difficult, unless the evidence is more coherent (Pettorelli, 2013; Pettorelli et al., 2005; Turnbull et al., 2008). For example, increased precipitation is expected to induce greening in dry regions, but studies often encounter considerable browning in semiarid ecosystems (Shaohong, Yunhe, Du, & Qinye, 2007; Yu, Luedeling, & Xu, 2010). Similarly, many cold regions show vegetation decline, despite increased temperature (Verbyla, 2008; Yu et al., 2010).
Three aspects can influence our ability to draw inference from greening/browning trends, toward management interventions for degradation. First is the choice of an appropriate spatial scale at which greening/browning is interpreted (Cumming, Cumming, & Redman, 2006; Jackson & Fahrig, 2015). Second is the use of various metrics for estimation of trends, even though they may all be related to a common index (e.g., NDVI; Gaitán et al., 2013). Third is the general difficulty in distinguishing the effects of climatic factors and human-induced land-use change in driving greening/browning trends (Reynolds et al., 2011). It is important to reconcile these underlying effects due to spatial scale, choice of metrics, and nature of land use to facilitate management and policy decisions on degradation.
1.1 Spatial scale
Long-term satellite data provide coverage at different spatial scales that span three orders of magnitude (i.e., from 10−2 km2 to nearly 102 km2). Management strategies can be designed and implemented at the coarse scale (101–102 km2) but are more difficult to prescribe and operationalize at finer scales (10−2–100 km2). For example, management action to arrest and/or reverse degradation can be envisioned for a watershed, but it is difficult to customize them at the scale of individual pastures within a watershed. Interpretations of trends, greening or browning, is not independent of scale (Cumming et al., 2006; Jackson & Fahrig, 2015) because of inherent spatial autocorrelation, and this may allow different studies to reach different conclusions (see Figure S1). Specifically, management inputs will be able to target degradation if browning is sufficiently positively autocorrelated at the fine scale (i.e., high degree of clustering). If patterns are clustered, then it will contain ‘hotspots’ of degradation, which can become targets for management interventions. However, in absence of clustering, or for diffused patterns, there is a risk of mismatch in the spatial configuration of the problem and its purported solution (Bestelmeyer, Goolsby, & Archer, 2011; Briske, 2017; Cumming et al., 2006). Another aspect, related to history of development of satellite technology, is the difference in length of time series data at coarse scale and fine scale. Data at the coarse scale (pixel size 101–102 km2) are available since the 1980s, whereas technology at the finer resolutions (pixel size 10−2–100 km2) appeared only in the early 2000s. So, due to differences in length of their respective time series, conclusions may differ between studies conducted at different scales.
1.2 Data metrics
Different metrics related to NDVI are frequently used to track trajectories of long-term vegetation change. For example, annual mean, annual sum, and annual maxima of NDVI are frequently used metrics (Gaitán et al., 2013). Although these capture essential features of interannual patterns, they exclude important information on intraannual and seasonal aspects of greening/browning (see Figure S2). A survey of the literature (Figure S2), suggests that our understanding of degradation is primarily based on studies that investigate a single spatial scale (i.e., pixel size), or use composite metrics that do not incorporate seasonal or intraannual patterns (i.e., mean, maximum, or sum of NDVI). For example, among 255 NDVI-based studies of vegetation degradation, in 67 prominent journals, only 48 studies incorporate intraannual variability, and only two studies address more than one spatial scale (Figure S2). Increasingly, it is becoming clear that managerial decisions need to be informed by both intraannual and interannual variation (de Jong, de Bruin, de Wit, Schaepman, & Dent, 2011; Pettorelli, 2013).
1.3 Land use
Greening/browning can not only be driven by responses to climatic factors (precipitation and/or temperature) but also through land-use change. Because the effects of these factors can occur simultaneously, it is not always possible to distinguish between them (Turnbull et al., 2008). For example, the role of climate and overgrazing cannot be easily distinguished in driving vegetation decline in drylands of Asia (Zheng, Xie, Robert, Jiang, & Shimizu, 2006). In fact, although vegetation change is widespread across much of Central Asian highlands, only a small fraction of the landscape may actually be overgrazed (e.g., Mongolian steppes; Fernández-Giménez et al., 2017; Gao, Angerer, Fernandez-Gimenez, & Reid, 2015).
Here, we analyse long-term vegetation trends for the Trans-Himalayan region in northern India (Figure S3), with the aim to understand the spatio-temporal patterns of degradation. Trans-Himalaya is part of the larger Central Asian highlands, which extends as the Tibet-Xinjiang plateau (China) and surrounding mountains (Figure S3). This region is of high global significance for its biological, anthropological, hydrological, and climatological importance. For example, nearly a third of all people in the world (China, India, Pakistan, and Bangladesh) depend on rivers originating from this cold-arid ecosystem, and adjacent mountains. Intensity of climate change in this region is unusually high compared with the global average (Figure S4; Shaohong et al., 2007; Turco, Palazzi, von Hardenberg, & Provenzale, 2015; Xu et al., 2009). Previous studies on vegetation trends in this region have been largely restricted to the Tibet-Xinjiang plateau (Yu et al., 2010; Zhang, Zhang, Dong, & Xiao, 2013) although this cold-arid ecosystem covers nearly 105 km2 of mountainous terrain in northern India (an area comparable with size of England and Wales combined).
Degradation and desertification across the larger Trans-Himalayan landscape are serious concerns that are often linked with livestock overgrazing (Shaohong et al., 2007) but remain to be adequately addressed (Bagchi, Gupta, Murthy, & Singh, 2017). Loss of ecosystem function and services, declining wildlife populations, negative impacts on human livelihoods with increase in ecological refugees, intensified soil erosion, and severe dust-storms, as well as devastating locust outbreaks, are some consequences linked to altered vegetation and climate in this ecoregion (Feng, Guo, Luo, & Qin, 2012; Neumann et al., 2015; Piao, Friedlingstein, Ciais, Viovy, & Demarty, 2007; Wang, Dong, Zhang, & Liu, 2004; Wang et al., 2017). Gradual warming and progressively wetter conditions in this cold and arid ecoregion are expected to result in greening (Fensholt et al., 2012; Zhu et al., 2016). But, previous studies do not offer much support for greening, instead browning is frequently reported and is attributed to overgrazing (Shen et al., 2013; Yu et al., 2010; Zhang et al., 2013). But, rising livestock numbers alone do not always explain observed patterns of degradation (Akiyama & Kawamura, 2007; Fernández-Giménez et al., 2017; Gao et al., 2015), and vegetation response to climate change could also be fairly important (Zheng et al., 2006). Experimental studies in this ecosystem indicate that climate change, particularly warming, can have strong negative impacts on vegetation (Klein, Harte, & Zhao, 2007), and moderate intensity grazing actually lends positive feedbacks on vegetation (Bagchi & Ritchie, 2010; Klein et al., 2007). But it has not been possible to disentangle the role of climatic drivers and purported overgrazing, in driving degradation patterns across much of Central Asia (Akiyama & Kawamura, 2007; Fernández-Giménez et al., 2017; Gao et al., 2015).
Here, we analyse long-term time series data from the beginning of the satellite era, and investigate greening/browning trends in the Trans-Himalayan region (Figure S3). We include available data at different spatial scales, spanning nearly four orders of magnitude in pixel-size from 10−2 to nearly 102 km2, to assess the degree of consistency across spatial scales. Next, we analyse intraannual and interannual trends of greening/browning and compare seasonal results against the commonly used composite metrics of annual mean-NDVI, peak-NDVI, and sum-NDVI. Finally, to address the influence of human land use, that is, degradation due to overgrazing, qualitatively, we compare vegetation trends in subregions where livestock numbers have generally declined in recent decades with subregions where they have increased. We also examine whether greening/browning is differently associated with environmental variables (elevation, temperature, and precipitation). From this, we identify locations that may require management attention (i.e., hotspots of browning), and whether these coincide with the broad distribution of human land use across the Trans-Himalaya. Importantly, this study area covered distinct subregions where livestock numbers have either increased or decreased, in the last three to four decades. The Spiti region has generally experienced a decline in livestock, particularly goat–sheep (Singh, Sharma, & Babu, 2015). Conversely, in Changthang, increased market demand for pashmina wool has been linked to rise in livestock numbers (Namgail, Bhatnagar, Mishra, & Bagchi, 2007). So, if overgrazing is the primary driver of vegetation trends, then we expect corresponding differences in greening/browning trends in Spiti and Changthang (Figure S3).
2 MATERIALS AND METHODS
2.1 Study area
Trans-Himalaya is a cold and arid mountainous landscape with average altitude of approximately 4,500 m (Figure S3). This ecosystem is highly seasonal with average temperature of −18 °C in winter, and 12 °C in the summer (Figure S4). Precipitation occurs primarily as winter snow, but there is a short wet season in late summer, due to the South Asian monsoon. The region is experiencing gradual warming, and also becoming increasingly wetter, both in winter and summer (Figure S4). These trends in temperature and precipitation, toward progressively warmer and wetter conditions, are consistent with adjacent regions of the Tibet-Xinjiang plateau (Ding et al., 2007; Wang et al., 2017; Xu, Knudby, Ho, Shen, & Liu, 2017). The broad vegetation type is alpine shrub-steppe (also called desert-steppe). Vegetation consists of grasses (e.g., Stipa, Elymus, Festuca, and Leymus) and sedges (Carex, and Kobresia), alongside shrubs and forbs (Caragana, Artemisia, Lonicera, and Potentilla). Trees are rare (Salix, Populus, and Juniperus) and are often restricted to river valleys where people plant them for timber and fuelwood. Vegetation growth period is short, between May and August (Bagchi, Gupta, et al., 2017). Grazing, by livestock and wild herbivores, is the primary land use across the Trans-Himalaya, alongside subsistence-based agriculture in human settlements near river valleys. Livestock usually graze near these settlements for most months of the year and visit higher elevations only in the summer months.
2.2 NDVI datasets
We used three sources of NDVI data for our analysis: (1) GIMMS NDVI3g.v1 dataset for 1982–2015 (Global Inventory Monitoring and Modelling Study); (2) SPOT data for 1998–2014 (Satellite Pour l's Observation de la Terre vegetation sensor); (3) different MODIS datasets for 2002–2015 (Moderate Resolution Imaging Spectroradiometer). GIMMS data, at 8-km pixel size and temporal frequency of 15 days, are the longest available time series. SPOT data are at 1-km pixel size and temporal frequency of 10 days. MODIS data are a collection of four datasets with spatial resolution of 250 m, 500 m, 1 km, and 5.5 km, with temporal frequency of 16 days.
2.3 Data processing
Satellite data from mountainous regions require preprocessing and quality checks due to effects of clouds, snow, and shadows on data quality (Figure 1). GIMMS data are available after corrections for clouds and changes in zenith, and also for potential discontinuities due to upgrades in sensor technology. For SPOT data, we used the status map layer information, and for MODIS data we used pixel quality layer information, to screen for pixels with reliable data. In a second step, we screened areas that were vegetated (grasslands and shrub-steppes), because large regions of the study area are under permanent snow or ice, or are barren scree, and otherwise lack natural vegetation cover. Through previous ground surveys in Ladakh and Spiti regions, we found that annual peak NDVI of typical vegetated patches ranges between 0.18 and 0.55 (Bagchi, Gupta, et al., 2017). We applied this as a selection criterion; pixels with annual peak NDVI value >0.18 for at least 70% of the time series were included in further analysis. This way, pixels with infrequent, missing, or poor-quality information were not included in analyses of trends (Figure 1). In general, 40%–70% of the area was sufficiently vegetated and contained reliable information for analysis (Figure S5); and this is consistent with other mountainous regions (e.g., Andes and Patagonia; Gaitán et al., 2013). Vegetated areas included Lahaul, Spiti, Changthang, and River Indus catchment in Ladakh; large tracts in the Karakoram region was deemed to be under ice and snow, yielded low quality data, or were otherwise barren (Figures S3 and S5).
2.4 Data analysis
We addressed intraannual and interannual trends by examining the linear trend (Figure 1) for each pixel, for each of the dates (Julian Day) in their respective time series (Bagchi, Gupta, et al., 2017; Verbyla, 2008). For this, we used Theil–Sen estimate for slope of regression (i.e., Sen's slope; Fernandes & Leblanc, 2005), as it is a more robust estimator of trend than ordinary least squares regression in the presence of outliers in the data. In absence of outliers, Sen's slope converges with ordinary least squares. This analytical approach (Figures 1, S6, and S7) allows inference over seasonal changes in greening/browning without additional steps of data-smoothening (Pettorelli et al., 2005). We repeated this analysis for each of the NDVI datasets (MODIS, SPOT, and GIMMS). A positive slope would indicate a greening trend at the particular date, whereas negative slope would indicate browning (Figures 1, S6, and S7). We used the statistical significance of estimated slope to identify pixels, which either had greening/browning trend or no consistent directional change in NDVI through time, with α = 0.10. We estimated spatial autocorrelation of trends, in greening/browning for each day, with Moran's I index (Lichstein, Simons, Shriner, & Franzreb, 2002). When a spatial pattern is random, then Moran's I is close to zero. Positive values indicate spatial clustering, whereas negative values indicate spatial dispersion or a checkerboard pattern (Figure S1). This helps judge the opportunity for implementing any management interventions to arrest or reverse degradation, as it improves when spatial patterns are clustered, rather than dispersed.
We explored whether greening/browning trends were related to background environmental variables, namely, elevation, mean temperature, maximum temperature, minimum temperature, and precipitation. Precipitation data, at monthly frequency from 1981 onwards at 0.05° spatial resolution, were obtained from CHIRPS database of the Climate Hazards Group (http://chg.ucsb.edu/). Temperature data, at monthly frequency at 0.5° spatial resolution, were obtained from NOAA-ESRL (https://www.esrl.noaa.gov/). Since there is mutual inter-dependence among these environmental variables (e.g., temperature and elevation are inversely related), we used a multivariate approach by ordination with nonmetric multidimensional scaling (NMDS). We used NMDS to reduce five environmental variables to a two-dimensional ordination space. NMDS is suitable for our approach because it represents distance between all pairs of observations in conceptual space (Legendre & Legendre, 1998). Next, we checked whether greening/browning pixels occupy different regions of this NMDS space (Axis 1 and Axis 2). This multivariate analysis is most meaningful at the coarse spatial scales, and for this, we used GIMMS results only. Subsequently, to aid interpretation of the NMDS ordination, we evaluated univariate differences (elevation, mean temperature, and precipitation) in greening and browning trends through mixed-effects repeated measures analysis of variance. To check for covariation with temperature and precipitation we used date (Julian Day) as a categorical variable and pixel identity as a random effect, and included their interaction term. Because data are as time series, we used a first order serial autocorrelation through time, to account for potential temporal autocorrelation (Lichstein et al., 2002). But, for elevation, we used a simpler statistical model without an interaction term because elevation does not vary by date. These geospatial and statistical analyses were conducted using R 3.2.2 for Unix environment (Ubuntu 14.04). We used R packages ‘raster,’ ‘rgdal,’ ‘RCurl,’ ‘stringr,’ ‘nlme,’ ‘mblm,’ and ‘ggplot2’ for the different steps in the analyses.
3 RESULTS
Greening and browning trends (i.e., Sen's slope) varied by date (Julian Day), and this was consistent across all datasets (p < .001 for all six datasets; Figure 2 and Table S1). There were two key seasonal patterns in greening and browning, and these were broadly and qualitatively consistent across the different datasets (Figure 2). All four MODIS datasets (from 250 m to 5.5 km) showed browning (negative slope) during spring and early summer, and greening (positive slope) in during late summer and into autumn (Figure 2). SPOT data (1 km) were also broadly consistent with browning during spring, and greening during summer and autumn. While browning in spring was weak in GIMMS data, it showed substantial browning in autumn (Figure 2).
There was clear geographical variation in greening and browning trends across seasons, and these were broadly consistent across the different datasets (Figure 3). An important pattern was that regions that showed browning in the spring and early summer (e.g., Spiti) underwent recovery and net greening in late summer. Such hotspots of spring degradation followed by late-summer recovery were detected in the southern region (Spiti) and in the central region (River Indus valley), at nearly all spatial scales (Figure 3; see Figures S8–S13). These patterns of greening and browning revealed by our analyses of intraannual trends were largely overlooked by composite or aggregate metrics (Figure 4). Annual mean-NDVI, maximum-NDVI, and sum-NDVI suggested that majority of the study area was greening at 250-m scale, and instead showed a mix of greening and browning at the larger scale (8 km), without clear agreement across the scales for hotspots of degradation (Figure 4). Percentage area under greening or browning showed high level of agreement across many scales (Figure 5). Spatial autocorrelation for degradation (browning pixels) as Moran's I was consistently positive, irrespective of the spatial resolution or season (Figure 6). This depicts the presence of high spatial autocorrelation for degrading areas, that is, true hotspots, as Moran's I index was high at the fine scale during the spring (F3,42 = 13.87, p < .001, for browning; F3,42 = 6.74, p < .001, for greening; Figure 6).
In NMDS ordination, the first axis included majority of the variance
, and the second axis had lower information content
. Greening and browning pixels occupied distinct regions of the NMDS ordination space (Figure 7), and this difference persisted across seasons. Subsequent analysis of univariate differences revealed that browning occurred in pixels that were at higher elevations, and greening occurred at lower elevations (Figure 8), and this pattern was persistent across seasons. Similarly, pixels in cooler areas during spring showed greening, whereas pixels in warmer regions showed browning (Figure 8). But, in summer and autumn, this association was reversed; warmer regions showed greening and cooler regions showed browning (Figure 8).
4 DISCUSSION
Our results help address important aspects of ecosystem monitoring and vegetation degradation. They identify degradation (browning) within a temporal (Figure 2) and spatial context (Figure 3), that is overlooked by commonly used composite metrics (Figure 4). This offers clear targets for management interventions: geographical hotspots of browning or degradation (Figure 3), and how they shift over seasons (Figure 5). When these hotspots are viewed within their geographical and seasonal context, it can help make qualitative distinctions between the effect of climatic drivers and human land use (overgrazing). These have implications for developing management strategies to arrest/reverse degradation (Joyce & Marshall, 2017). Now, management interventions might not be able to fully address the drivers of vegetation change—altered climate due to global greenhouse gas emissions. Instead, they may target measures that help adapt to vegetation change, reduce negative impacts of land use on ecosystem services, and foster ecosystem stewardship (Akiyama & Kawamura, 2007; Bestelmeyer et al., 2011; Reid, Fernández-Giménez, & Galvin, 2014). Because grazing by livestock is the primary human land use in the Trans-Himalaya, management interventions can include improved grazing policies (Akiyama & Kawamura, 2007; Briske, 2017; Briske et al., 2011; Joyce & Marshall, 2017 Papanastasis, 2009), because traditional instruments are rapidly deteriorating under pressure from external markets (Namgail et al., 2007; Singh, Bhatnagar, Lecomte, Fox, & Yoccoz, 2013). Previous experimental studies in the Trans-Himalaya have determined how grazing mediates plant–soil feedbacks, especially in Spiti region (Bagchi & Ritchie, 2010; Bagchi, Roy, Maitra, & Sran, 2017), which emerged as a degradation hotspot. These feedbacks can potentially be incorporated into policy decisions. Likewise, experimental climate manipulations have also revealed that negative impacts of warming can be ameliorated through moderate intensity grazing (Klein et al., 2007), because it alters nutrient cycling feedback in soil (Bagchi, Roy, et al., 2017) and aids compensatory vegetation regrowth. The underlying mechanisms revealed by the experimental studies can help interpret the long-term vegetation trends of greening/browning, and together they offer insights for grazing management.
First, it is important to account for intraannual seasonality of greening/browning as composite metrics of annual sum, mean, or maximum NDVI can mask key details, and sometimes be misleading (Figures 3 and 4). Second, it is important to evaluate agreement between different spatial scales, in order to inform management interventions that can attempt to arrest/reverse degradation. If degradation occurs in a spatially contiguous manner, with high autocorrelation and clustering at the fine scale, then it is possible to address it through interventions that are designed for coarser scales (Bestelmeyer et al., 2011; Jackson & Fahrig, 2015). Management action becomes increasingly more difficult if the spatial patterns are random or dispersed (i.e., checkerboard). We found that spatial clustering of greening/browning was consistently higher at 250-m scale compared with 5.5 and 8 km (Figure 6), and this helps delineating the spatial dimension of vegetation degradation (Cumming et al., 2006; Jackson & Fahrig, 2015; Lichstein et al., 2002). Furthermore, spatial clustering was generally strongest in the spring (Figure 6), when there was considerable browning (Figure 2). Together, these patterns indicate that hotspots of degradation can be clearly identified for management interventions, at a relevant spatial scale (Akiyama & Kawamura, 2007; Briske et al., 2011; Papanastasis, 2009).
Second, hotspots of degradation can be identified only after the geographical and seasonal context has been determined (Figure 3). The hotspots are themselves not stationary, but move over time; they appear and disappear in different locations in different seasons (Figure 3). For example, Spiti region emerges as a degradation hotspot in the spring and early summer, but it recovers later in the growing season (southern/southwestern study area; Figure 3); this is consistent across all spatial scales, from 250 m to 8 km. Because livestock numbers in Spiti region may have declined over the last few decades (Singh et al., 2015), it is unlikely that overgrazing is the only driver of degradation, and climate effects may also be responsible (see Figure S16). Additionally, a persistent hotspot is also evident in the Changthang region in the GIMMS data (southeastern study area; Figure 3). Here, unlike Spiti, livestock numbers have increased due to influence of external markets for pashmina wool from goats in Changthang (Namgail et al., 2007; Singh et al., 2013). Perhaps rising grazing pressures require monitoring and interventions to prevent degradation in Changthang (Figure S16). Generally, human settlements, and consequently much of the livestock grazing pressure, occur in the lower elevations during much of the year, chiefly along river valleys. But browning appears more severe at higher elevations. Livestock do graze in the higher elevations, but mostly during summer and autumn, and not during spring when there is a stronger evidence for browning. Higher elevations, incidentally, also show stronger warming trend (Xu et al., 2017). Collectively, these inconsistencies between human activity and the geographical/seasonal context of greening/browning indicate that the vegetation trends are perhaps more strongly related to climatic factors than to overgrazing alone (Akiyama & Kawamura, 2007; Fernández-Giménez et al., 2017; Gao et al., 2015; Zheng et al., 2006).
Overall, we find that there has been no advancement in spring phenology, instead it may have regressed, but autumn phenology was extended (Shen et al., 2013). Degradation appears most severe during the spring season (Figures 3 and 5). About a quarter of the landscape appears to be browning in the spring (Figure 5), and these areas require most urgent attention (Figure 3). Rising temperature and precipitation is expected to favour greening in cold-arid ecosystems (Fensholt et al., 2012; Zhu et al., 2016). So evidence of browning (Shen et al., 2013; Yu et al., 2010; Zhang et al., 2013) appears counter-intuitive. The underlying mechanisms that could explain browning instead of greening may include altered snowmelt patterns under warmer climate (Musselman, Clark, Liu, Ikeda, & Rasmussen, 2017), declining water availability in soil (Schlaepfer et al., 2017), and accelerated erosion of topsoil with loss of soil organic matter and reduction in soil fertility (Wang et al., 2017). If these indirect mechanisms are at play (Turnbull et al., 2008), then their effects should be incorporated into land use policy, as they can be influenced by grazing mediated feedbacks at the plant–soil interface (Bagchi & Ritchie, 2010; Bagchi, Roy, et al., 2017; Klein et al., 2007).
From another viewpoint, it is also important to evaluate the evidence for greening/browning across different data sources—MODIS, SPOT, and GIMMS—because they differ in their time scales covered. In general, we found broad agreement between the different datasets (Figures 2 and 3). Such congruence between datasets that otherwise differ in their time scales of coverage imply that the greening/browning trends are fairly strong, and slight differences in intercept do not overwhelm the inference over slopes. It is reassuring for managers to know that spatial and temporal patterns are consistent across different satellites, and they do not contradict one another (Figures 5 and 6). However, there was one key difference between the GIMMS data and the rest. Although the data from 250 m to 5.5 km resolution generally showed browning in spring and greening in late summer, GIMMS data show considerable browning in autumn as well (Figures 3 and 5). As, discussed above, these differences could be due to joint effects of difference in scale, and difference in length of the time series. To separate these two aspects, we reanalysed GIMMS data for the period of 2002–2015, as this controls for difference in lengths of the time series. As expected, reanalysis of the GIMMS data from 2002 onwards showed a closer match with the MODIS and SPOT results (Figures S14 and S15).
In conclusion, defining the geographical and seasonal context of greening/browning is critical for addressing degradation. Conventional methods that use composite metrics may be inadequate for understanding vegetation change within this necessary context. The role of climatic drivers in long-term vegetation change appears stronger than human land use and overgrazing. Our results support that overall warmer and wetter conditions are not advancing spring phenology, rather they are extending the growing season later in summer–autumn. This indicates eco-hydrological processes are likely involved, and the Trans-Himalayan ecosystem is responsive to reduced influence of snowmelt in spring and rising influence of summer rainfall.
ACKNOWLEDGMENTS
Our work was supported by STC-IISc, DST-SERB, DBT-IISc, MoEFCC, and MHRD (Government of India). We thank B. U. Reddy, N. J. Singh, D. Barua, and M. Sankaran for discussions and suggestions.
