Vegetation response to extreme climate events on the Mongolian Plateau from 2000 to 2010

Climate change has led to more frequent extreme winters (aka, dzud) and summer droughts on the Mongolian Plateau during the last decade. Among these events, the 2000–2002 combined summer drought–dzud and 2010 dzud were the most severe on vegetation. We examined the vegetation response to these extremes through the past decade across the Mongolian Plateau as compared to decadal means. We first assessed the severity and extent of drought using the Tropical Rainfall Measuring Mission (TRMM) precipitation data and the Palmer drought severity index (PDSI). We then examined the effects of drought by mapping anomalies in vegetation indices (EVI, EVI2) and land surface temperature derived from MODIS and AVHRR for the period of 2000–2010. We found that the standardized anomalies of vegetation indices exhibited positively skewed frequency distributions in dry years, which were more common for the desert biome than for grasslands. For the desert biome, the dry years (2000–2001, 2005 and 2009) were characterized by negative anomalies with peak values between −1.5 and −0.5 and were statistically different (P < 0.001) from relatively wet years (2003, 2004 and 2007). Conversely, the frequency distributions of the dry years were not statistically different (p < 0.001) from those of the relatively wet years for the grassland biome, showing that they were less responsive to drought and more resilient than the desert biome. We found that the desert biome is more vulnerable to drought than the grassland biome. Spatially averaged EVI was strongly correlated with the proportion of land area affected by drought (PDSI <− 1) in Inner Mongolia (IM) and Outer Mongolia (OM), showing that droughts substantially reduced vegetation activity. The correlation was stronger for the desert biome (R2 = 65 and 60, p < 0.05) than for the IM grassland biome (R2 = 53, p < 0.05). Our results showed significant differences in the responses to extreme climatic events (summer drought and dzud) between the desert and grassland biomes on the Plateau.


Introduction
Over the past century, extreme climate events (e.g., drought, extreme temperatures) have increased in frequency and severity globally, particularly in central and northern Eurasia (Groisman et  The Mongolian Plateau has experienced several extreme climate events during the past decade, including extreme winters (referred to as dzuds in Mongolia) and summer droughts, sometimes both in the same year (drought-dzud combinations). The frequency and amplitude of these extreme events has increased in the 2000-2010 period compared to the few decades prior to 2000 (supplementary table 1 supplementary table 1 available at stacks.iop.org/ERL/8/035033/mmedia). Droughts on the Mongolian Plateau are characterized by precipitation and soil moisture deficits with Palmer drought severity index (PDSI) values >−2 to >−5 (Davi et al 2010). Dzuds typically feature extreme cold, heavy snowfall, reduced availability of forage and widespread mortality of livestock (Tachiiri et al 2008, Morinaga et al 2003. When a dzud is preceded by a summer drought, the combined summer drought-dzud often leads to even higher mortality of livestock (Fernández-Giménez et al 2012). The consecutive 1999-2002 summer drought-dzud was the worst of the last 50 years when 30% of the national herd perished (Fernández-Giménez et al 2012, Tachiiri et al 2008. The 2009-2010 winter dzud was also very severe when 8.5 million livestock, 20% of the national herd, died during this period (Fernández-Giménez et al 2012). The warming and drying trends in Inner Mongolia were more significant during the last 30 years than in the preceding 20 years . Previous studies also suggested that the five-year dzud cycles are related to the El Nino Southern Oscillation (ENSO) while the decadal cycles are influenced by the Indian monsoon (Morinaga et al 2003). Precipitation variability in the region is a function of various forcings such as the westerly jet stream migrating northwards in summer, the Siberian anticyclone in winter, evapotranspiration and the monsoon (Aizen et al 2001, Iwao andTakahashi 2006). Extreme events like summer droughts and winter dzuds are expected to increase in frequency and magnitude (Fernández-Giménez et al 2012).
Ecosystem responses to these extreme climate events vary by biome on the Mongolian Plateau owing to differences in vegetation cover, precipitation, evapotranspiration and albedo (Donohue et al 2009, Snyder et al 2004, Yi et al 2010. Climate anomalies often show strong spatiotemporal variations (New et al 2000), indicating that any impact assessment should be site-specific (Qi et al 2012). To date, little research has examined the response of vegetation to dzud and combined summer drought-dzud for different biomes on the plateau using long-term satellite observations (Tachiiri et al 2008, Xu et al 2012. Satellite-derived vegetation indices including the normalized difference vegetation index (NDVI) and the enhanced vegetation index (EVI) have been widely used as surrogates of vegetation canopy greenness and primary productivity (Myneni et al 1995, Huete et al 2006. Most vegetation greenness anomaly studies have focused on the effects of summer drought on forest biomes with significant carbon sequestration potential like the Amazon basin (Samanta et al 2010, Lewis et al 2011. The effects of extreme climate events in semiarid biomes, which cover 40-50% of the terrestrial surface and are home to 40% of the world population, have received relatively less attention (Reynolds et al 2007). There are few studies that focus on vegetation anomalies in semiarid regions (e.g., Xu et al 2012). Even fewer studies focus on the impacts of extreme winters unique to the Mongolian Plateau (Tachiiri et al 2008). In contrast with previous research based on vegetation greenness in the region that utilized the NDVI (Xu et al 2012), we made use of the EVI and EVI2 vegetation indices. Huete et al (2002) suggested that the use of NDVI in semiarid regions was questionable owing to its sensitivity to the soil background signature in areas with <50% fractional cover. Consequently, EVI was proposed as it was not as sensitive to bright soil signatures and atmospheric effects.
Here we examined the responses of vegetation to extreme climate events on the plateau using the satellite-derived vegetation indices and land surface temperature. We identified the spatiotemporal dynamics in the anomalies of vegetation greenness using long-term earth observation records and examined the climate drivers of these anomalies using meteorological trends. Our specific questions are: (1) how does vegetation respond to climate extremes, throughout the last decade across the Mongolian Plateau? and (2) were there any significant differences in the vegetation response among the biomes, or between the two political units, Inner and Outer Mongolia, especially during the extreme summer drought of 2001 and winter dzud of 2010?

Satellite data
We used monthly composites of EVI (MOD13A3), land surface temperature (LST, MOD11A2) and shortwave white sky albedo (MCD 43B3) (Collection 5) derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra satellite from 2000-2010. These datasets with a spatial resolution of 1 km represent a range of biophysical variables to identify anomalies during the past decade (figure 2).
While EVI anomalies can be obtained for the period 2000-2010 based on the decadal mean, they cannot be calculated for an earlier period as previous synoptic sensors (e.g., AVHRR) lacked a blue band that is needed for atmospheric correction (Jiang et al 2008). To complement the MODIS EVI, we used the new vegetation index and phenology (VIP) dataset of the NASA MEaSUREs (Making Earth System Data Records for Use in Research Environments) program (http://vip.arizona.edu/viplab data explorer.php#) (Jiang et al 2008). The VIP dataset consists of three decades (1981-2010) of consistent vegetation indices derived from AVHRR and MODIS and has a climate modeling grid (CMG) resolution of 5.6 km (0.05 • ). This new dataset enabled us to compute EVI2 anomalies in the past decade as deviations from a 30-year mean .
For the MODIS products, we only used the pixels with the highest quality in our analysis, through the use of standard quality control (QC bits) and LDOPE tools (https://lpdaac. usgs.gov/tools/ldope tools) (Samanta et al 2012). EVI and EVI2 were used to examine the effects of summer droughts on vegetation greenness whereas the LST product was used to assess the effects of extreme winters. Shortwave white sky albedo was filtered with pixel reliability flag = 0 in the MCD43B2 product, to retain only the pixels that had undergone full BRDF inversion and was used to validate the EVI and EVI2 anomalies to account for false positives from clouds or contaminated pixels.

Climate data
We identified temperature and precipitation anomalies following Aragao et al (2007) and Anderson et al (2010). Previous studies have used merged Tropical Rainfall Measuring Mission (TRMM) data (Huffman et al 1995) to interpret vegetation index anomalies when the existing climate station network was sparsely distributed. We used the monthly precipitation rate (3B43 version 7, http://disc.sci. gsfc.nasa.gov/) to characterize precipitation deficits (Aragao et al 2007, Anderson et al 2010. Our study area is at the northern edge of TRMM coverage (50 • N) and includes the dominant land cover types of grassland and desert biomes in IM and OM. We also used the PDSI (Palmer 1965) to characterize drought severity and extent. PDSI, a widely used drought index (Dai et al 2004), is a measure of soil moisture that is currently available as compared to the long-term mean, and incorporates rainfall, soil moisture demand and supply (Palmer 1965). We obtained PDSI data at 0.5 • × 0.5 • resolution from the Numerical Terradynamic Simulation Group at the University of Montana (Zhao and Running 2010). The PDSI values range from −10 (dry) to −10 (wet). Mild drought is represented by values of −1 to −1.9, moderate drought by −2 to −2.9 and severe drought by <−3 (figure 3).

Analyses
We calculated standardized anomalies (sa) of satellite-based data as where sa is the standardized anomaly of a biophysical variable (e.g., vegetation greenness, LST, etc) for the dry season mean of a specific year (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) relative to the long-term growing season mean (x) and standard deviation (σ ) over the past decade. The MODIS monthly composites were averaged for June, July and August (JJA) to represent the dry season mean for each year over the 2000-2010 period. Similarly, the January-February mean of MODIS LST data was used to calculate the standardized anomaly for extreme winters in the Mongolian Plateau as the maximum snow depth typically occurs in January (Morinaga et al 2003).
Precipitation data from the monthly TRMM 3B43 was processed similarly to provide dry season estimates of rainfall for each year over the 2000-2010 period as well as the long-term growing season mean and standard deviation. Standardized anomalies were then calculated at the pixel level for each year for both the MODIS and VIP datasets. In order to track the temporal variation, we plotted frequency histograms of the anomalies for each variable, by year and binned the data in order to quantify negative large anomalies (<−1 std.) as a percentage of vegetated area (e.g. figure 4). Positively skewed frequency distributions are characterized by negative anomalies and signify dry years, whereas negatively-skewed distributions with positive anomalies signify wet years. Finally, the EVI, PDSI and TRMM datasets were spatially averaged over the desert and grassland biomes for IM and OM independently in order to quantify and validate the interannual variability of summer and extreme winter trends at an aggregated level with a higher signal-to-noise ratio (Atkinson et al 2011).
The Mongolian Plateau is highly heterogeneous and can be divided into desert, grassland and forest biomes, with distinctive eco-climatic zones ( (Chen et al 2013). This suggests that any anomaly study of vegetation productivity or land surface phenomenon should first stratify the plateau into its constituent biomes and major administrative divisions to enable the comparison of vegetation anomalies between biomes and to avoid confounding effects in interannual comparisons of anomaly trends. We used the World Wildlife Fund (WWF) biome boundaries (Olson et al 2001) to stratify our study region and delineate desert, grassland and forest biomes. The desert biome in IM includes the eastern Gobi desert steppe and the Alashan Plateau semi-desert while the grassland biome primarily consists of the Mongolian-Manchurian grasslands, which extend into OM and the Ordos Plateau. The desert biome in OM includes the eastern Gobi desert steppe, Junggar basin semi-desert, Gobi lakes valley desert steppe, Great lakes basin desert steppe and Alashan Plateau desert steppe. The OM grassland biome consists mainly of the Mongolian-Manchurian grasslands but also includes the Sayan intermontane steppe, Altai alpine meadow and the Khangai mountains alpine meadow.   (table 1).

Anomalies of vegetation indices and LST in the desert biome
Negative anomalies of EVI and EVI2 were larger for the IM desert biome than for the OM.   (2001,2005,2009) and a prominently negative skew in a wet year like 2003 ( figure 4(b)). Furthermore, the frequency distributions of desert biome anomalies were more evenly distributed and showed greater variation for IM (figure 4(a)) than for OM ( figure 4(b)), with the notable exception of 2003 with a peak centered on +2 that denoted an extremely wet year. The VIP EVI2 anomalies in the desert biome were similar to the EVI  (2000-2001, 2005, 2009) are characterized by the majority of negative anomalies with peak values between −1.5 and −0.5 std. and are statistically different (p < 0.001) from relatively wet years (2003,2004,2007). VIP EVI2 is based on a longer-term mean

Anomalies of vegetation indices and LST for the grassland biome
There was no significant difference between EVI and EVI2 anomalies in the IM and OM grassland biome. The proportion of area covered by negative anomalies in EVI and EVI2 was generally lower than 10% for both IM and OM (figure 2, table 2). The frequency distributions of both EVI and EVI2 anomalies were generally indicative of positive anomalies (figure 5).
The MODIS EVI anomalies for the IM and OM grassland biomes (figures 5(a)-(d)) showed strong negative skews (i.e., a majority of positive anomalies) with peak values of +0.5 to +1. The frequency distributions of EVI anomalies were tightly grouped for both the IM and OM grassland biomes with wet years not significantly different from dry years. In comparison, the VIP EVI2 anomalies, based on a longer-term mean , showed greater variation and spread for the frequency distributions (figure 5). Dry years such as 2000 had peak values centered on zero for the IM grassland biome (figures 5(a) and (c)), whereas they were centered on +0.5 for OM grassland biome (figures 5(b) and (d)). Wet years in both IM and OM grassland biomes were centered on +1 (figure 5).
In IM, negative anomalies in the LST covered 50-80% of the area in 2001, 2003, 2005, 2006, 2008 and 2010. Similarly, the negative LST anomalies in OM covered 50-85% in 2001, 2004, 2005, 2008 and 2010. The dzud extreme winter event of 2009-2010 was more prominent for OM and IM grassland biomes than in the desert biome showing a strong positive skew in 2010, reaching a peak of −1.5 std. (supplementary figure 1(c) available at stacks.iop. org/ERL/8/035033/mmedia). In 2001, summer negative LST anomalies in IM occupying 3.7% and 6.8% in EVI and EVI2, respectively, corresponded to negative winter anomalies in January-February covering 52% of the area (table 2). Though negative LST anomalies in the winter of 2000-2001 covered 72% of OM, the preceding summer did not show evidence of anomalies from the extreme summer for the grassland biome compared to the desert biome (table 2).

Spatially averaged PDSI, TRMM and LST
The proportion of land area affected by drought (PDSI < −1) for the desert biome reached 60-80% for 2001, 2005 and 2009, showing that the desert biome was largely

Greening and extreme event trends
Northern and northwestern China have experienced changes in climate with an increase in precipitation, leading to enhanced vegetation growth in the 1980s and 1990s . This is also corroborated by a decrease in PDSI < −2 over IM in the 1980s and 1990s (Dai et al 2004). Our analyses of EVI and EVI2 anomalies showed an increase in positive anomalies in non-drought years (e.g., 2003 and 2004; tables 1 and 2) and were consistent with previous studies (Xu et al 2012). A recent NDVI-based study found an increasing trend in vegetation greenness between 1982 and 2009 for the whole of China, with a corresponding increase in percentage area under positive anomalies from 27% in the late 1980s to 61% in the late 2000s (Peng et al 2011). This greening trend was generally consistent with earlier studies that suggested increasing greenness and primary productivity during the last two decades of the 20th century not only in China (Xiao and Moody 2004, Park and Sohn 2010 but also more broadly across the middle and high latitudes of the northern hemisphere (Myneni et al 1997, Zhou et al 2001, Xiao and Moody 2005. This greening trend, however, has been partly offset by extreme climate events such as the 2000-2002 droughts (Davi et al 2009(Davi et al , 2010. PDSI in China and the Mongolian  (Davi et al 2009). These contradictory trends present a problem for studying anomalies at the scale of the entire plateau. In addition, there are two climate gradients with increasing precipitation and a decrease in temperature from the southwest to northeast in IM and from the south to north in OM. Our analysis stratified the anomalies by the two dominant biome types as well as by political boundaries owing to divergent land cover/land use trajectories (Chen et al 2013).

Vulnerability and resilience of desert and grassland biomes
We found that the desert biomes were slightly more vulnerable and less resilient to drought than the grassland biome on the plateau. Sui et al (2013) showed that while the desert steppe had a lower carbon sink capacity than a typical/meadow steppe, it had the highest standard deviation (108 g cm −2 ). Teuling et al (2010) suggested that grasslands were less resilient than forests during heat wave events in which high temperature and evapotranspiration led to soil moisture depletion and further suppressed plant productivity. This is probably only true of short-term droughts, as other studies suggested a quick recovery in grasslands following summer droughts in the past decade (Xu et al 2012, Ponce-Campos et al 2013. The meadow steppe within the grassland biome had low interannual variability in annual precipitation with coefficient of variation (CV) less than 0.21 (Yu et al 2003); in comparison, the desert biome exhibited higher interannual variability in annual precipitation (CV > 0.40) (Ellis 1992). Bai et al (2004) suggested that ecosystem stability as measured by the CV of aboveground biomass increased from the species to plant functional group and to the community level for IM grasslands with precipitation from January to July explaining up to 49% of the variation. We argue that some herbaceous desert steppe species were less resilient than species of typical and meadow steppes within the grassland biome owing to greater per cent cover and precipitation (John et al 2008). Cheng et al (2007) suggests that dominant herbaceous grass species like Stipa bungeana were being replaced by xerophyte shrubs like Artemisisa ordosica and Cynachum komarovi on the Ordos Plateau-one of the ecoregions within the desert biome. Stipa and Cynachum spp. had shallow roots and depended on light rain events, while Artemisia spp had deeper roots and depended on larger rain events (Cheng et al 2007).

Ecological implications of drought and dzud
Few studies have used satellite data to help characterize extreme winters which are a major climate phenomenon unique to the plateau. For example, Tachiiri et al (2008) used a snow water equivalent to measure snow cover in conjunction with summer NDVI in order to predict livestock mortality. We used the percentage of land area covered by negative LST anomalies to characterize the extreme winter dzud events in 2001 and 2010. Since dzud events are often influenced by the previous summer (Tachiiri et al 2008), we found that negative EVI/EVI2 summer anomalies in 2009 preceded negative LST anomalies (>−1 std.) in January-February 2010.
Semiarid ecosystems like the desert steppe on the Mongolian Plateau (>150 mm annual rainfall) are driven by low precipitation with high variability to operate in a non-equilibrium manner, i.e., abiotic control on plant biomass (Sullivan 1996). Here, the influence of abiotic factors like precipitation variability is more important than densitydependent feedback from grazing herds (Fernández-Giménez and Allen-Diaz 1999). In comparison, the meadow-mountain (350-400 mm annual precipitation) steppe conforms to the more conventional range control model of grazing density dynamics where increased grazing leads to a decrease in grass species and an increase in forbs and annual weeds (Fernández-Giménez and Allen-Diaz 1999). The typical steppe (200-300 mm annual precipitation) lies between the desert steppe and meadow steppe. Grass and other vegetation cover in the typical steppe respond interactively in an intermediate fashion to a combination of annual precipitation dynamics and grazing pressure, with changes similar to the range control model in species composition (Fernández-Giménez and Allen-Diaz 1999).
The Mongolian Plateau received above average rainfall in the mid-1990s that led to an increase in livestock numbers (Retzer and Reudbach 2005). This increased grazing pressure was compounded by socio-economic changes following the collapse of collectivization. These changes included decreased migration, fewer functioning wells (previously maintained by the government) and private ownership of livestock by sedentary herders (Sternberg 2008). In this context, Retzer and Reudbach (2005) suggested that the uncontrolled increase in livestock, in numbers more suited for equilibrium conditions (grazing density control), could not be sustained in a non-equilibrium system (with abiotic factor control) and led to high livestock mortality. Reduced grazing pressures following a combined summer drought-dzud event will have positive ecological feedback in terms of increased grass biomass and canopy cover relative to forbs, shrubs and weedy annuals.

Anomalies explained by land cover/use changes
Some of the positive anomalies in drought and non-drought years (e.g., in north-central OM) could be explained by the recent increase in cropland area following the major, persistent drought of -2002(Regdel et al 2012. Mongolia initiated a program called Atar 3, aka, the 'Third Campaign to Reclaim Abandoned Agriculture Lands', which increased government spending between 2005 and 2009 in order to improve food security and prevent future food crises (Pederson et al 2013). Most of the increase in cropland cover is in the Bulgan, Selenge and Tov aimags (first-level administrative subdivision) around the capital, Ulaanbaatar (Pederson et al 2013). A similar increase in agricultural cover could also partly explain the increase in positive anomalies during drought and non-drought years in IM (figure 2). However, there was a reduction of cropland cover in some regions of IM during the same period. Wang et al (2012) suggested that northwestern China experienced a decrease in croplands during 1996-2000 when IM lost 3.3 × 10 4 ha due to the 'Grain for Green' program. The policy was designed to shift 15 million hectares of low-yield cropland to forest as well as the afforestation of barren hillsides by offering grain and cash to farmers as compensation for land conversion (Feng et al 2005, Long et al 2006. EVI/EVI2 anomalies were inversely related to white sky albedo anomalies during the summer droughts of 2001 and 2009. Negative anomalies (<−1) in EVI/EVI2 for both desert and grassland biomes corresponded to positive albedo anomalies (>1) (figure 2). These positive albedo anomalies, indicative of reduced per cent cover, occurred along the typical steppe-desert steppe interface in the grassland-desert ecotone. Conversely, positive VI anomalies (>1) for both desert and grassland biomes corresponded to negative albedo anomalies (<−1), which could be explained in part by irrigated agriculture in IM and cropland expansion in north-central OM. This suggested that the vegetation anomalies were not false signals caused by clouds and/or aerosol-contaminated pixels.

Conclusions
The 2000-2002 combined summer drought-dzud and 2009-2010 dzud were severe and affected most of the Mongolian Plateau. This region has seen an increased frequency of extreme climatic events in the last decade as compared to the previous 20 years. We assessed the extent and severity of drought and extreme winters using precipitation and PDSI data and examined the vegetation response to these extreme events using anomalies of vegetation indices and LST derived from MODIS and AVHRR. We found significant differences in biome response which was unique in both the decadal (MODIS) and multi-decadal (AVHRR) satellite-based greenness records. Our results showed that the grassland biome on the plateau was more resilient to drought than the desert biome. Finally, the longer-term EVI2 record validated our 11-year MODIS EVI record showing that there was a differential response in the desert and grassland biome and that it was observed at both time scales (deviations from an 11-year and 30-year mean). There is a need for further investigation of vegetation response due to possible ecological as well as socio-economic impacts such as grazing, urban expansion and increase in cropland cover. A reduction in grazing pressure following extreme climatic events such as dzud might have positive feedback in terms of increased grass biomass and canopy cover as compared to shrubs and weedy annuals.