Evaluation of the suitability of six drought indices in naturally growing, transitional vegetation zones in Inner Mongolia (China)

Naturally growing vegetation often suffers from the effects of drought. There exists a vast number of drought indices (DI’s) to assess the impact of drought on the growth of crops and naturally occurring vegetation. However, assessing the fitness of these indices for large areas with variable vegetation cover is often problematic because of the absence of adequate spatial information. In this study, we compared six DI’s to NDVI (the normalized difference vegetation index), a common indicator of vegetation occurrence and health based on satellite-acquired reflectance data. The study area covers an aridity gradient from forests to deserts along a 2,400-km-long section across the Inner Mongolia Autonomous Region of China. On an annual timescale, standardized precipitation index (SPI) was the most appropriate in assessing drought in steppes and deserts. On a seasonal timescale, the self-calibrated Palmer drought severity index (scPDSI) displayed the greatest sensitivity during the summer, but not during the other seasons. On a monthly timescale, scPDSI demonstrated the greatest sensitivity to the various vegetation zones (i.e., forests, steppes, and deserts) in June and July. Further analysis indicated that summer drought had a lag-effect on vegetation growth, which varied from one to six months according to the specific vegetation cover. The mixed response of DI’s to NDVI and the lag-effect in transitional vegetation on annual, seasonal, and monthly timescales were ascribed to differences in DI definition and the dominant plant species within the transitional cover. The current study has the potential to inform the drafting of selection criteria of DI’s for the study of drought-related impact on naturally growing vegetation at timescales from month to year.


Introduction
Drought can lead to serious reduction in crop production, possibly affecting the socio-economic sustainability of communities [1]. Drought can lead to plant growth reductions by Remote sensing technologies provide an important source of spatiotemporal data in the study of vegetation dynamics and climate change, particularly because it offers near-continuous images over time and their convenience [18][19][20]. Vegetation response to drought over large, diverse areas is routinely observed with remote sensing techniques [21]. Healthy vegetation reflects more near infrared (NIR) and less red radiation. Normalized difference vegetation index (NDVI) is a normalized ratio of red to NIR spectral reflectance, i.e., (NIR-red)/(NIR +red). Prior research has shown that vegetation productivity, photosynthetic effective radiation, and vegetation coverage are strongly correlated to NDVI [22,23]. In practice, NDVI is often used in assessments of crop yields during periods of water-supply shortages and drought [24][25][26]. Nicholson et al. [27] used AVHRR-based estimates of NDVI and found particularly strong plant response to rainfall in the eastern desert steppes of Africa. Suzuki et al. [28] found a strong relationship between the spatial distribution of NDVI and seasonal change and climate factors in Inner Siberia. Barbosa et al. [22] analyzed NDVI changes in northeastern Brazil and found that values of NDVI in the area were greatly influenced by drought. Meng et al. [29] studied the characteristics of climate variability and NDVI from 1982-2000 in China, and observed strong correlation between NDVI and spatiotemporal variation in several key hydroclimatological variables. The primary objective of the research is to assess the suitability of different DI's for different temporal resolutions, ranging from month to season, to year across a 2,400-km-long aridity gradient from semi-humid (forests) to arid climates (deserts) in northeastern to northcentral China.

Study area
The study area is located in Inner Mongolia Autonomous Region of China (37˚24'-53˚23'N, 97˚12'-126˚04'E), measuring 2,400 km from east to west and 1,700 km from north to south (Fig 1). The climate transitions from semi-humid to arid conditions can be seen from east to west of the study area. Vegetation zones along the gradient show clear transitions from forests, meadow steppes, typical steppes, desert steppes, and deserts, with annual temperature and cumulated precipitation increasing and decreasing from east to west, respectively (Fig 1). Plant species associations for each vegetation zone appear in Table 1. With differences in rooting depths, different plant species respond differently to drought. Owing to their deep roots, trees and some shrubs are largely resistant to long-lasting drought, compared with their shallowrooted counterparts (e.g., grasses) [30].
Intercontinental climate is characterized by cold and long winters, and dramatic swings in temperature during the spring and autumn. Inner Mongolia is one of the most sensitive regions to global climate change, experiencing substantial changes in climate over the past 50 years [31]. Frequency and severity of drought is expected to increase under continued global warming.

Data source
Daily and monthly hydrometeorological data (2000-2015) from 26 weather stations in Inner Mongolia (Fig 1) were downloaded from the China Meteorological Data Network (http://data. cma.cn/, last accessed Nov. 14, 2018). These data and their derivatives served as input in the calculation of DI's (see below). Vegetation evolution and dynamics was examined in terms of MODIS-based estimates of NDVI, downloaded from the Geospatial Data Cloud (http://www. gscloud.cn/, last accessed Nov. 14, 2018).

Drought indices
The six DI's considered here consist of SPI (standardized precipitation index), SPEI (standardized precipitation evapotranspiration index), MI (relative moisture index), Pa (precipitation

PLOS ONE
Evaluation of the suitability of six drought indices in naturally growing, transitional vegetation zones anomaly percentage), K (Sielianinow coefficient), and scPDSI (self-calibrated Palmer drought severity index) ( Table 2). The China Meteorological Administration currently uses some of these DI's in the classification of meteorological drought throughout the country [31]. The Kindex forms the basis for the Climate Regionalization in China and is widely used in assessments of climate change and associated drought, as well as assessments of rates of desertification across China [20,32]. Traits of the six DI's include: SPI is used to quantify precipitation deficits as anomaly percentile on multiple timescales. SPI depends on commonly available precipitation data and is relatively easy to implement in the assessment of drought severity at different timescales [33]. It is computed by fitting a Gamma probability density function to the frequency distribution of precipitation and transforming the Gamma distribution into a standardized normal distribution [34] ( Table 2). The drought severity is defined by the probability of anomaly occurrence. Lloyd et al. [34] found SPI suitable for quantifying meteorological, hydrological, and agricultural drought.
SPEI is derived from SPI by standardizing the difference between water supply (precipitation) and water demand estimated from potential evapotranspiration [35,36] ( Table 2). The potential evapotranspiration is calculated with the Thornthwaite method, which takes into account air temperature and an annual heat index, based on monthly air temperature. The calculation of the index involves a climatic water balance implemented at different

Drought indices Equations Variables
H(x) is the cumulative probability of distribution; G(x) is the gamma distribution; G(x) is undefined for x = 0; P (x = 0) is the probability of zero precipitation, and x is precipitation in mm.
SPEI (Standardized precipitation evapotranspiration index) [5] D i = P i -PET i D i values are aggregated at different timescales, following the procedure for SPI. The difference D k i;j in a given month j and year i depends on the timescale k chosen; D i -data series are subsequently fitted to a log-logistic probability distribution function. P i and PET i are the precipitation and potential evapotranspiration for month i (both in mm); calculations of PET i are based on the Thornthwaite method [46].
P is the rainfall in a given period (mm); PET is potential evapotranspiration (mm) during the same period.

Pa (Precipitation anomaly percentage) [38]
Pa ¼ P-P P � 100% P is a period of precipitation; P is the average rainfall over the same period over successive years.
T and P are accumulated temperature (˚C) and precipitation (mm), when temperature is � 10oC. scPDSI (Self-calibrated Palmer drought severity index) [40] d is the moisture anomaly; P is the total monthly precipitation; P is a 'climatologically appropriate' precipitation amount for existing climatic conditions'; K is a climate coefficient (weight), determined by month and geographic location. https://doi.org/10.1371/journal.pone.0233525.t002

PLOS ONE
Evaluation of the suitability of six drought indices in naturally growing, transitional vegetation zones timescales, and adjusted to a log-logistic probability distribution. Because SPEI is based on an assessment of the water balance, it can be directly compared to scPDSI [35].
MI is calculated by dividing the difference between rainfall and potential evapotranspiration by the potential evapotranspiration during the same period [37] (Table 2); MI has been used to assess the degree of drought across multiple timescales.
Pa is calculated from the occurrence of drought caused by precipitation anomalies within a given period, which is calculated by dividing the difference between observed precipitation and normal (30-year mean) precipitation by the normal precipitation during the same period ( Table 2). The index can distinguish dry events at monthly and annual timescales [38]. The index is simple to calculate and, as a result, is widely used.
K is calculated based on air temperature and rainfall amount during the growing season, whenever air temperature is � 10˚C [39]. The original equation coefficient of 0.10 was adjusted to 0.16 in adapting the method ( Table 2) to China's conditions [31]. [40] is based on a sum of the current soil moisture anomaly and a fraction of the previous index value. The soil moisture anomaly is calculated based on an equation of water supply and demand and, as a result, it accounts for variables of precipitation, runoff, soil water storage, and evaporation [41]. The model requires data for air temperature, precipitation, and soil water content [42]. The resultant values are classed into 11 drought severities. With long-term calibration deficiencies and limited utility to areas of calibration, PDSI was further revised, giving rise to the self-calibrated PDSI (scPDSI; Table 2) [43]. The drought index calculated with scPDSI is generally more spatially compatible and accounts for extreme wet and dry events by taking into account frequencies of rare events.

Data analysis
Monthly and seasonal values of NDVI in each vegetation zone were averaged from the original 16-day and monthly data, respectively. Annual NDVI was, in turn, derived by averaging the growing-season data from May to August of each year. Values of NDVI were extracted within a 50-km radius circle around each weather station. Monthly DI's were calculated directly from equations referenced in Table 2. To investigate lag-effects (or hysteresis), pairwise-correlations between the six DI's over two-to seven-month delays and NDVI of the current month were assessed.
Seasonal DI's were calculated according to the four seasons, namely spring (March-May), summer (June-August), autumn (September-November), and winter (December-February of the following year). Considering precipitation in winter has the potential to impact vegetation growth in the following year, a year (or hydrological year) was defined here as the period from September to August of the following year. Annual and seasonal indices of K, Pa, SPI, and SPEI were calculated directly from corresponding total precipitation, mean air temperature, and total potential evapotranspiration, whereas annual and seasonal indices of MI and scPDSI were calculated by averaging their monthly values.
Indices of SPI, SPEI, and scPDSI were calculated with a python-script developed for the National Integrated Drought Information System of the USA (www.drought.gov/drought/ climate-and-drought-indices-python, last accessed Nov. 14, 2018). Potential evapotranspiration in MI (  Table 3 provides the annual pairwise-correlations between the six DI's and NDVI for the five vegetation zones. Pairwise-correlations between the six DI's and NDVI, from highest to lowest, were associated with typical steppes, followed by meadow steppes, desert steppes, deserts, and forests. In forests, pairwise-correlations between all DI's and NDVI were generally low (r < 0.46) and statistically not significant (p-value > 0.05). For the three steppes, pairwise-correlations were generally high, i.e., maximum r = 0.90, 0.92, and 0.84, respectively, with pvalues < 0.01. In desert zones, pairwise-correlations were generally greater than those for forests and lower for steppes (maximum r = 0.68, p-value < 0.01). From averaged pairwise-correlations for the three steppes, the greatest pairwise-correlation was 0.85 for SPI, followed by 0.84 for SPEI, and 0.83 for Pa.

Seasonal pairwise-correlations
Seasonal pairwise-correlations between the six DI's and NDVI for the five vegetation zones are given in Table 4. Generally, NDVI had low correlation with all DI's during the spring and autumn, compared to their corresponding values in summer. In some cases, correlations were negative in spring and autumn.
In forests, pairwise-correlations were generally low (i.e., r < 0.4; Table 4) and statistically not significant. In meadow steppes, SPI and SPEI had the greatest correlation in summer, i.e., r = 0.68. In typical and desert steppes, scPDSI had the greatest correlation with NDVI, yielding 0.88 and 0.82, respectively, followed by SPEI in summer, with r = 0.81 and 0.69. In deserts, scPDSI had the greatest correlation with NDVI in summer (r = 0.71).

Monthly pairwise-correlations
Fig 2 provides pairwise-correlations of current-month DI's with current-month NDVI. Across months, irrespective of vegetation zone, correlations were least in May. Similar to the annual timescale, DI's were the least correlated with NDVI for forests during the growing season. For meadow steppes, correlation was lowest in August, compared to September. A similar trend was found for typical and desert steppes when based on SPI, SPEI, MI, and Pa. Only scPDSI

PLOS ONE
Evaluation of the suitability of six drought indices in naturally growing, transitional vegetation zones produced correlations greater in August than in September. In deserts, correlations based on all DI's were greatest in June, and subsequently decreased from July to September. Among DI's and irrespective of vegetation zone, scPDSI had the strongest correlation with NDVI in June and July (Fig 2).

Lag-effect of drought
Pairwise-correlations between NDVI and lagged DI's were much greater than between NDVI and DI's during the current-month for all vegetation zones (Fig 3). Lag times among DI's varied.
In forests, lagged scPDSI had the greatest correlation in June and July among the six DI's, with maximum correlation occurring after about six months of lag in June and about one month of lag in July. For the other DI's, however, maximum correlation was found to occur after about one month of lag in June and August. In meadow steppes, lagged SPI and Pa had high correlations with NDVI, after about two months of lag in June and about three months in July. In August, however, maximum correlation existed between NDVI and all DI's lagged by one month. In typical steppes, greatest pairwise-correlations with SPI, SPEI, MI, and Pa and NDVI occurred when the DI's were lagged by about five months in June, three months in July, and one month in August. Generally, all correlations were greatest for meadow and typical steppes. In desert steppes and deserts, lagged scPDSI had the greatest correlation with NDVI, when compared to the other DI's. For all vegetation zones, NDVI in August were more strongly correlated with the DI's when lagged by one month, than when not lagged (Fig 3).

Occurrence of drought over the study period
Located in the interior of Eurasia, Inner Mongolia is affected by continental climate and frequently affected by drought. During the study period from 2000-2015, drought affected many areas of the study region differently. Over the study period, six out of the 16 years (i.e., 2002, 2003, 2012, 2013, 2014, and 2015) were found unaffected by drought [44]. The dataset amassed for this study provides a strong basis for the current examination of drought and suitability of the various DI's.

Implications of timescale
On an annual timescale, SPI and SPEI were shown to have a much stronger correlation with NDVI in grasslands (steppes). This is consistent with results from a study that assessed the impact of drought on wheat yield with the same two DI's [5]. If annual drought is the primary focus, SPI is the preferable DI for steppes and deserts.
Pairwise-correlations between DI's and NDVI were highest in summer and lowest in spring and fall in all vegetation zones, as a result of low temperatures prevailing during the latter two seasons, i.e., low temperatures, not available soil water, limit vegetation growth during these seasons [8,45]. Comparing different DI's, scPDSI had stronger correlations with NDVI in the summer for most vegetation zones, except the meadow steppes. In these steppes, SPI and SPEI had stronger correlation with NDVI in summer (r = 0.68), about equivalent to the correlation between scPDSI and NDVI (r = 0.65). Based on these results, scPDSI is a superior DI for summer conditions. This conclusion applies equally well to monthly data, where correlation between scPDSI and summer NDVI were higher than any other pairwise-comparisons across all vegetation zones, with the exception of forests (August only, Fig 2). Current-month scPDSI has added advantages of being a good measure of monthly drought and being based solely on current-month information, offers a more generalized indicator of drought.

Lag-effect of drought
The lag-effect suggests that the effects of drought on vegetation potentially become more important as time elapses. This coincides with observations that the response of NDVI lags changes in temperature and precipitation [46][47][48][49]. This lag-effect occurs not only as a function of leaf accumulation, but also because of the amount of water stored in the soil complex. Leaf biomass, especially in grasses including in annual and perennial plant species, is an accumulated value and determined by past drought conditions. Rainfall and meltwater infiltration

PLOS ONE
Evaluation of the suitability of six drought indices in naturally growing, transitional vegetation zones into the deeper soil is generally available to vegetation over extended periods, causing plant growth, and therefore NDVI, to lag the incidence of infiltration into the soil.
Lag times vary as a function of vegetation zone. Pairwise-correlations between lagged DI's (by 2-7 months) and NDVI were consistently highest for desert steppes and deserts and lowest for forests, meadows, and typical steppes. Grasses in steppes are adapted to short-term drought, whereas shrubs with their various adaptations to dry air and soil conditions are much more resistant to long-term drought [50]. This plant association to drought is consistent with our belief that grasses in steppes are more sensitive to seasonal and monthly drought, and shrubs in deserts are more sensitive to annual drought.
Capacity to store water in soils may help explain the observed lag-effect of drought. Drought indices of SPI, SPEI, Pa, and MI account for only the water supply and demand components of the water budget (i.e., precipitation and evapotranspiration), resulting in the indices being skillful at only accounting for present-time drought response. This may explain why SPI, SPEI, Pa, and MI provide improved correlations with NDVI when lagged. This introduced lagging of the indices effectively compensates for the absence of soil water storage and potential carry-over effects in their explanation of drought (Table 2). Unlike the two-component indices, scPDSI incorporates soil water storage in its definition of drought (Table 2). Due to its three-component formulation, scPDSI is regarded as the better measure of monthly drought across vegetation zones. This conclusion is consistent with findings from prior studies on the utility of DI's; for instance, the study by Chang et al. [51]. This finding has significant implication for gauging the timing, development, and persistence of drought.

Vegetation response to drought
In forests, all DI's had weaker correlations with NDVI across all timescales, even after taking into account the lag-effect of drought. Forests, located near mountain ranges, receive additional water from snowmelt, improving overall growing conditions for forests downslope [50]. The deep roots of trees can access water from deep in the ground and large quantities of carbohydrates and nutrients stored in the roots can cause forests to be less vulnerable to severe, prolonged meteorological drought [32]. In forests, water is normally not a limiting factor.
The steppes are located in the semiarid to arid parts of the aridity gradient. With their shallow rooting systems, grasses are fundamentally more sensitive to reductions in precipitation [52]. Grasses in steppes absorb water from the upper soil layers, and as a result rely on frequent, short pulses of meteoric water input [53]. In steppes, the six DI's were effective at capturing drought across the three timescales and month-long lag times considered. The lageffect of drought at the monthly timescale indicated that longer timescales were necessary to impact vegetation growth in the steppes. All told, lagged SPI, SPEI, MI, and Pa provided reasonably good measures of monthly drought. In considering lag on NDVI response, Pa, SPEI and SPI lagged by one to three months, provided the strongest measure of drought in meadow and typical steppes, whereas current-month scPDSI provided a more reliable measure of ongoing drought in desert steppes.
In deserts, shrubs are the dominant vegetation (Table 1). Desert plants generally have lower transpiration rates, with enormous capacity to absorb water from the ground [53]. Shrubs in deserts adapt to dry conditions with their unique features to limit water losses by, for example: (i) curling their leaves or growing hairs on their surfaces; (ii) storing water in their fleshy or succulent parts; and (iii) accessing water with their vast network of horizontal and vertical roots [54]. Many shrubs in deserts utilize water from the deeper soil and rely on moisture that is replenished with each successive rain pulse [50].
Besides the impact of abiotic variables, growth of natural vegetation can be adversely affected by non-abiotic factors, such as human, livestock, and insect activity. During the 13 years from 2001-2013, NDVI predicted with temperatures and precipitation data from Inner Mongolia was significantly correlated with MODIS-based estimates of NDVI, which suggest that human, grazing, and other biological effects on plant growth were less significant than the effects associated with prevailing hydrometeorological factors [55]. In grassland and desert regions of Inner Mongolia, livestock is the primary factor affecting grass development and growth. Based on results from the Century model [56], growth of grass was shown to be inversely proportional to the severity of grazing.

Conclusions
Until now, most of the studies of this kind have focused on drought effects in agricultural crops; few studies have examined these effects on naturally growing, transitional vegetation. In this study, we analyzed the pairwise-correlations between six common DI's and NDVI, as a measure of vegetation occurrence and response. This study found that: At an annual timescale, there were strong pairwise-correlations between the six DI's and NDVI for semiarid steppes, weak but statistically significant correlations for deserts, and no correlation for forests. Compared with the different DI's, SPI is the most appropriate to assess drought in steppes and deserts, followed by SPEI and Pa.
At a seasonal timescale, there were strong pairwise-correlations between the six DI's and NDVI during the summer, but not during the spring or autumn. The scPDSI-index was most fitting for typical steppes, desert steppes, and deserts during the summer, followed by SPEI; SPI and SPEI were most appropriate for meadow steppes. None of the six DI's considered were appropriate for forests.
At a monthly timescale, there were significantly stronger correlations between NDVI and scPDSI of the current month, compared with the other DI's in June and July. During the summer, NDVI had stronger correlation with DI's of previous months (accounting for a lag-effect of drought) than their corresponding current-month values. The monthly lag-effect is obvious for all DI's, except scPDSI. The lag ranged from one to six months and was distinct for the different vegetation zones, varying from month-to-month. The lag-effect was most obvious in August, irrespective of vegetation zone.
Based on our analysis, SPI could be viewed as the most appropriate to monitor drought on an annual basis and scPDSI, for summer drought in semiarid steppes and deserts. We recommend that users of these indices pay attention to the lag-effect of drought and the functional response of local vegetation at a monthly timescale. With climate change, drought has the potential to affect naturally growing vegetation more severely than in the past. With the 16-year dataset provided here, it is fairly difficult to conclude about the overall suitability of the DI's over longer time horizons.