Drought severity and change in Xinjiang, China, over 1961–2013

Global warming, as indicated by the reported increase in global surface temperature by 0.74°C over the past 100 years, is now an accepted fact (IPCC 2007). Water- and climatic-related hazardous events have happened frequently over the past several decades, as have more extreme events. Drought is one of these extreme events and is harmful to agriculture, ecology, hydrology, economy and human life (Núñez et al. 2014). Under the global warming background, slight decreases in severity of drought were observed using the Palmer Drought Severity Index (PDSI) (Palmer 1965) and monthly precipitation (P) (Sheffield et al. 2012; Sun et al. 2012). By the end of the 21st century, 15–44% of the world is anticipated to be affected by drought (IPCC 2012). In China, the crop area affected by drought increased annually from an average 11.6 to 25.1 million hectares between the 1950s and early 2000s, which amounts to an increase of 116%; and the proportion of crop area impacted by drought increased from 8 to 16% (NBSC 2010; MWR 2010; Xu et al. 2014). Frequent severe droughts in many areas of northern China in 1997 and from 1999 to 2002 (Zhang 2003), as well as in southwestern China from 2009 to 2010 (Xu et al. 2014; Zhang et al. 2012a), caused huge economic and social losses. Drought severity has been quantified by different drought indices, such as the dryness index (DI) (Budyko 1974; Arora 2002), the PDSI (Palmer 1965) and the Standardized Precipitation Index (SPI) (McKee et al. 1993), the Crop Moisture Index (Palmer 1968), the Composite Index (CMA 2006), the Joint Deficit Index (Kao & Govindaraju 2010), and so on. Among these, the SPI describes both short- and long-term drought impacts by taking into account P anomalies at different timescales (Mirabbasi et al. 2013). The length of the P record and the nature of probability distributions affect the estimation of SPI values (Mishra & Singh 2010). The SPI is a purely P-based index and does not consider the role of other climatic or hydrologic factors in the development of a drought event. This has led to the development of deuterogenic standardized drought indices, such as the Standardized Runoff Index (Shukla & Wood 2008), the Standardized Precipitation Evapotranspiration Index (SPEI) (Beguería et al. 2010; Vicente-Serrano et al. 2010), the Standardized Alternative to the PDSI (Sheffield et al. 2012), the Standardized Groundwater Index (Bloomfield & Marchant 2013), the Standardized Streamflow Index (Núñez et al. 2014), and the Standardized Snow Melt and Rain Index (Staudinger et al. 2014).

The SPEI, which is the standardized difference between cumulative P and cumulative reference evapotranspiration (ET_o), is well suited to representing the effects of global warming on drought severity (Beguería et al. 2010). Besides being multiscalar like the SPI, the SPEI is easy to calculate and is as sensitive as the PDSI in measuring evapotranspiration demand (Beguería et al. 2013). Although the SPEI is relatively new (developed in 2010), it has been applied in many research contexts, including studies of climate change, drought variability, drought reconstruction, drought atmospheric mechanisms, drought impacts and drought monitoring (Vicente-Serrano et al. 2010). With the increase in the applications of the SPEI, there has been considerable improvement in the methods used to compute it. One of the improvements requires calculating ET_o using the Hargreaves equation (Hargreaves & Samani 1985) or the Penman-Monteith equation (Allen et al. 1998) recommended by the Food and Agriculture Organization (FAO) (Jensen et al. 1990), because there was bias toward the temperature of drought index (Zhang et al. 2016a), of which ET_o was estimated using the Thornthwaite (1948) (TW) equation. By using observational and global gridded data, Beguería et al. (2013) highlighted three different ET_o (or actual evapotranspiration) calculation methods and their effects on the resulting SPEI series. Based on this information, they recommended the most robust PM equation for estimating ET_o, provided the required climatic data are available.

In recent years, different drought indices, analysis technologies, study durations, and regions in Xinjiang have been investigated. First, the drought indices, i.e. SPI and SPEI, have been frequently utilized with the analysis of T and P. Wang et al. (2013) reported that P-indices values significantly increased in Xinjiang from 1960 to 2009. Using the SPI, Zhang et al. (2012b) inferred that the severity and duration of droughts have decreased in northern Xinjiang (NX), but increased in the southern region of south Xinjiang (SX) and the center of eastern Xinjiang. Yang et al. (2011) conducted an analysis of spatiotemporal changes (1960–2100) in T and P extremes in the Tarim River Basin (located in the southeastern Xinjiang) over 1960–2009. Their results indicated a higher probability of flood occurrence in summer together with frequent occurrence of droughts. Wang et al. (2015) analyzed monthly SPEI during 1960 to 2010 and its teleconnection with atmospheric circulation patterns in northwestern China, including Xinjiang. Their results showed that step changes occurred in 1986, droughts occurred frequently before 1986, and wet periods prevailed after 1986. The periods of 1973–1983 and 1993–1998 showed the highest and lowest drought activity, respectively. Second, analysis technologies, such as tree ring or satellite data, were applied to overcome the limitations in observation data or in spatial scale (Chen et al. 2015; Xu et al. 2015). Cao et al. (2015) evaluated terrestrial water storage changes of Gravity Recovery and Climate Experiment satellite data, derived total deficit index to investigate drought dynamic over northwestern China, including Xinjiang, from 2003 to 2012. They concluded that the study region experienced a severe long-term drought from May 2008 to December 2009. Li & Zhou (2014) analyzed trends in ET_o, P and an aridity index (DI) during 1960–2010 and found a general decline in drought severity in Xinjiang. Previous studies have contributed greatly to the assessments of drought severity, but regional drought severity research is still needed given the differences in the evolution, trends and spatiotemporal distribution of droughts in arid and semi-arid regions.

Previous research has shown that the evolution and trends in drought characteristics in Xinjiang differ based on the drought indices used, and these apparent trends may even conflict with one another (Zhang et al. 2012b; Li & Zhou 2014). In recent years, the SPI and SPEI have been applied widely to assess drought severity, but no comparisons related to the ET_o estimation methods or their effects on SPEI at the site-scale have been conducted for arid and semiarid regions. This study aims to: (1) assess the differences between TW-based ET_o and PM-based ET_o, as well as their effects on SPEI_TW and SPEI_PM estimates in the arid and semi-arid regions of Xinjiang; (2) compare the temporal evolution, trends and spatial distribution of SPEI_TW and SPEI_PM against the SPI at the one-, three-, six- and twelve-month timescales for different subregions; and (3) analyze the correlations between the three drought indices and assess their sensitivity to P and ET_o. Given the extensive applications of the SPI and SPEI indices in drought analysis, the adaptability and utility of the SPEI in arid and semi-arid areas would be evaluated.

The Xinjiang Uygur Autonomous Region is located in northwestern China, interior mainland central Asia. The region encompasses both arid and semi-arid zones. In general, the area is characterized by ‘three mountains and two basins.’ The ‘three mountains’ features include the Kunlun Mountains in the south, the Tianshan Mountains in central Xinjiang, and the Altai Mountains to the northeast. The ‘two basins’ refer to the Tarim Basin between the Tianshan and Kunlun mountains, and the Junggar Basin between the Altai and Tianshan mountains. The two basins both contain vast deserts, including the Takelamakan desert in the south, which is the largest desert in China. The region is thus divided into two (NX, which has 26 sites, and SX, which has 27 of the total 53 sites) by the Tianshan mountain range. The sites in SX and those close to the Taklamakan desert are extremely arid. Because wet marine air rarely reaches inland, the ecosystems in these regions are very fragile. The subsurface of the Gobi desert responds quickly to solar heating, resulting in major evaporation of ground moisture and thus increased drying (Zhang et al. 2012c). Xinjiang is also quite sensitive to global climate changes (Deng et al. 2014).

Since McKee et al. (1993) first proposed the SPI as a meteorological drought index, this metric has been widely used to measure drought across different regions (Hayes et al. 1999; Bonaccorso et al. 2003; Tsakiris & Vangelis 2004; Paulo & Pereira 2007; Mishra & Singh 2010). To compute the SPI at different timescales using the P series in the present study, the procedure suggested by McKee et al. (1993) and Lloyd-Hughes & Saunders (2002) was followed. Drought levels were divided into five groups: extreme drought (SPI ≤ −2.0), heavy drought (−1.5 > SPI ≥ −2.0), moderate drought (−1.0 > SPI ≥ −1.5), mild drought (−0.5 > SPI ≥ −1.0) and normal conditions (0.5 > SPI ≥ −0.5).

In order to assess additional information on drought conditions, the ratios of P to ET_o,PM and ET_o,TW, denoted as DI_TW and DI_PM, i.e., DI_TW=P/ET_o,TW and DI_PM=P/ET_o,PM, were also computed. In these equations, dryness increases as DI decreases. The UNEP (1993) defines DI values in the ranges of [0.05, 0.2], [0.2, 0.5], [0.5, 0.65], and [0.65, 0.8] as arid, semi-arid, dry sub-humid, and semi-humid climate types, respectively.

If r_j falls inside the confidence limits, the hypothesis that r_j is zero is accepted using a two-tailed test and a maximal lag j with temporal-dependence in x_L, i.e. j_TD, is determined.

The observed D_ET in Xinjiang resulted to different DI values for different months and at the annual scale (see Table 2; certain high DI_TW values for the months of February, March and November should be omitted because they were overestimated as a result of very low ET_o,TW values). Because the ET_o,TW values were lower than the ET_o,PM values, higher DI_TW values than DI_PM values were obtained for each respective month. The DI_TW implied a wetter climate in Xinjiang than the DI_PM did, resulting in a ‘semi-arid’ climate classification for EX rather than ‘arid’. This also showed the deviation of ET_o,TW from ET_o,PM and that of DI_TW from DI_PM. Using DI_TW resulted in an unreasonable climate type classification for Xinjiang.

The large D_ET values found resulted from the use of different equations to estimate ET_o values. The TW equation only considers air temperature-related climatic variables; therefore, increasing air temperature (T) should result in increased ET_o. On the other hand, the PM equation considers T as well as u₂, relative humidity and sunshine hour, and the estimated ET_o,PM values reflect the comprehensive interactions between these factors. Therefore, ET_o,PM values are thought to better characterize atmospheric evaporation ability in Xinjiang. Beguería et al. (2013) emphasized the more important role of using proper ET_o equations in moisture-limited areas than in high-P areas. As with the DI_TW versus DI_PM variations, the variations in D_ET likely contribute to corresponding differences between the subsequent estimates of SPEI_TW and SPEI_PM values in Xinjiang.

The r values of the three drought indices for different timescales were much smaller and more variable than the r values found in Beguería et al. (2013). The r values in their study were greater than 0.93 for three different ET_o equation-based SPEI series and 13 sites in the Netherlands and Spain. This indicates that the performance of different ET_o equations affects SPEI more in Xinjiang, China, than in the Netherlands or Spain. This is likely because Xinjiang is located in an arid and semi-arid region and there are over three months when T is below zero there, which results in large D_ET values. On the other hand, negative T values have less impact on D_ET in the Netherlands and Spain because these countries belong to different (wetter and warmer) climate types than Xinjiang. Therefore, it is not surprising that larger deviations between SPEI_TW and SPEI_PM were found in Xinjiang than in the Netherlands and Spain. This emphasizes the importance of choosing a proper ET_o equation when calculating SPEI.

The Sen's slope (10³b) and MMK statistic (Z_m) values were compared for the SPI, SPEI_TW and SPEI_PM over twelve months at the one-, three-, six- and twelve-month timescales (Figure 10). Figure 10 shows the following.

• (1) Values of b and Z_m for both SPI and SPEI_PM in different months were generally positive regardless of the temporal scale. This indicates that drought severity for each month in Xinjiang generally decreased from 1961 to 2013 according to the SPI and SPEI_PM, although seasonality was a factor. However, there were negative Z_m values for SPEI_TW at months 4, 6, 4 and 2 given the one-, three-, six- and twelve-month scales, respectively. Additional negative values were found for the SPEI_TWb value. Trends in SPEI_TW varied alternately between increasing and decreasing, revealing the complicated nature of drought severity. These results indicate that droughts in Xinjiang were alternately lighter or harsher over the period 1961–2013 according to the SPEI_TW. By comparing the b and Z_m values for the SPI, SPEI_TW and SPEI_PM, we were able to identify great consistency between the SPI and SPEI_PM trends but little consistency between the SPI-SPEI_TW and SPEI_TW-SPEI_PM trends.

• (2) The differences in b values over twelve months decreased consistently as the timescale increased for each index. The j_TD between SPI and SPEI_PM increased as the timescale increased, but for SPEI_TW it was opposite. Corresponding to one-, three-, six- and twelve-month timescales, the j_TD values of SPI in 2, 4, 6 and 12 months were larger than 0, the j_TD values of SPEI_TW in 3, 2, 0 and 0 months were larger than 0, and the j_TD values of SPEI_PM in 7, 8, 8 and 12 months were larger than 0. The j_TD values of SPI from one- to twelve-month timescale increased from 1 and 8 to 11 and 12, respectively. The j_TD values of SPEI_TW were generally smaller than 8, indicating temporal-independence of SPEI_TW. The j_TD values of SPEI_PM ranged between 1 and 12 and become more and more close for each month as the timescale increased to 12-month. The general increase in j_TD values at the twelve-month timescale affected the significance of the twelve-month SPI and SPEI_PM trends, shifting them from significant to insignificant in different months, which was more suitable for SPI. The SPEI_TW exhibited fewer self-correlations and was less correlated with time for most of the months analyzed at the different timescales; as a result, this index was inapplicable in this location because climatic variables are time-dependent to different extents. In general, temporal trends in the SPEI_PM agreed well with those for the SPI over a twelve-month period, indicating that drought severity in Xinjiang had declined.

The changes, correlations and trends in SPEI_TW did not agree well with those of SPI at the four studied timescales, but those of SPEI_PM did. Beguería et al. (2013) reported that the magnitude of a trend increased with the timescale for SPEI but declined at the global scale, i.e., drought severity was increasing worldwide. In spite of their conclusion, local and regional drought severity and trends may differ from our results. Given that SPI has been accepted worldwide as a spatiotemporal drought index (Guttman 1998; Hayes et al. 1999), SPEI_PM is a more acceptable and applicable index than SPEI_TW in Xinjiang because of its high consistency with SPI at different timescales. Therefore, the following section will compare SPI and SPEI_PM but not SPEI_TW.

On the one hand, ET₀ began increasing in the early 1990s in northwestern China, including Xinjiang (Li et al. 2014), which has resulted in increased drought risks because of higher atmospheric evaporative demand. However, on the other hand, the increase in mean annual P in Xinjiang (Wang et al. 2015) has resulted in upward trends for both the SPI and SPEI indices in recent decades. The increasing trend of SPI in Xinjiang or northwestern China is mainly due to P increase. This result has been reported by other studies (Zhang et al. 2012b; Wang et al. 2013; Chen et al. 2015; Xu et al. 2015), also shown in this research in Tables 1 and 3. Correspondingly, the weather in Xinjiang has become slightly wetter than before, despite the occurrence of major droughts from 1974 to 1981 and 2009 to 2010.

There were many studies on the evolutions, spatial-temporal distributions, and frequencies of drought in Xinjiang. Many applied observed data and calculated some drought indices, such as SPI (Zhang et al. 2012b), evaporative wet index (Zhang et al. 2016b), T and P extremes (Yang et al. 2011), and SPEI (Wang et al. 2015). Other studies also applied tree-ring or satellites technologies, and the deduced data were further used for estimating some drought indices (Cao et al. 2015; Chen et al. 2015; Xu et al. 2015; Zhang et al. 2016b). This study used the observation data at 53 sites in EX over 1961–2013. In general, there is some consistency between this and other studies, i.e., both revealed decreasing trends of drought, but differ in some severe drought period during the study period. For example, the drought period was 1973–1983 during the whole period of 1960–2010 from Wang et al. (2015), but was 1974–1981 from Wen & Shi (2006) and this research. The differences in the results existed because of many factors. These factors include the following. (i) The drought indices were estimated using different mathematic methods. The proposed indices were used for different goals, e.g., the T and P extremes focused on extreme drought events, SPI and SPEI revealed multi-scalar drought evolutions, and DI denoted the ratio of P to ET_o. (ii) The tree ring sampling revealed a long history of droughts, but at a small spatial scale at single or a few sites. The satellite data provided drought distribution information in a larger spatial scale, but with a shorter duration.

This work is an important contribution to the understanding of the spatiotemporal variations in drought severity within arid and semi-arid regions. The SPI, SPEI_TW and SPEI_PM must be selectively applied based on regional conditions. The SPI is applicable when the region's water balance is not a contributing factor to drought. The SPEI accounts for the possible effects of temperature variability and temperature extremes beyond the context of global warming (Vicente-Serrano et al. 2010), but this index requires more climate data than the SPI to be reliable. The SPEI_PM performed better than the SPEI_TW when compared to the SPI, indicating that the PM equation is a better estimation of ET_o. Therefore, the SPEI_PM is the most preferred index for denoting drought severity in Xinjiang when climatic data are available. SPEI_TW is applicable when only temperature data are available. Of the three drought indices, SPEI_PM was selected as the best index for denoting drought severity in Xinjiang, China.

Although there were deviations for ET_o,TW when compared to ET_o,PM in this research, and the TW method has several well documented limitations and biases (e.g., Jensen 1973; Amatya et al. 1995), the method has been extensively applied (Beguería et al. 2013; Yu et al. 2014), because it requires only temperature-related climatic variables, and is more accurate in the regions with positive temperature in most days of a year. Mavromatis (2007) showed that the PM equation is not necessarily superior to the TW when used in a drought index. Yu et al. (2014) computed SPEI_TW for the whole China and revealed that severe and extreme droughts have become more serious since the late 1990s for all of China. ET_o,TW were also applied in drought evaluation of other regions (such as California drought) with the index PDSI (Robeson 2015). SPEI_TW and SPI were effective to monitor the effects of rapid onset of drought at Mead, in the north eastern United States (Hunt et al. 2014). Except for the TW and PM equations, other methods were used, such as Haude (1952) formula and Hargreaves (Hg) equation (Hargreaves & Samani 1985) for estimating SPEI in the Czech Republic (Potop et al. 2012) and global droughts (Beguería et al. 2013), respectively. Estimation of ET_o could be conducted with different methods, considering the user's objectives, the studied regions, and the efficacy that could be achieved.

ET_o values in Xinjiang differed depending on whether they were estimated by the TW or PM equation. Positive D_ET (= ET_o,PM–ET_o,TW) values occurred either for different months or at inter-annual scales. Annual D_ET values in EX reached 430 mm. Under the circumstances, the TW equation proved unsuitable for estimating ET_o. Generally large D_ET values resulted in correspondingly different D_i,TW and D_i,PM values, as well as different SPEI_TW and SPEI_PM values. The PM equation was more reliable than the TW equation in accurately estimating ET_o and subsequent SPEI in Xinjiang. The computed SPEI_TW was more consistent with the SPEI_PM but less consistent with the widely-accepted SPI at the four studied timescales, not only based on the r values between the different indices, i.e., SPI, SPEI_TW and SPEI_PM, but also based on their appropriateness in denoting the historically severe drought from 1974–1981. Moreover, correlations between the three types of drought indices and their related climatic variables (P, ET_o,TW, or ET_o,PM) showed that both the SPI and SPEI_PM were highly sensitive to P, but the SPEI_PM and SPEI_TW were not very sensitive to ET_o.

Using the MMK method, increasing trends in SPI and SPEI_PM were detected for different months, although seasonality was a factor. There were much more downward trends in SPEI_TW which were largely inconsistent with the upward trends in SPI and SPEI_PM. Trends in the SPI and SPEI_PM for different timescales increased at most sites, although the number of sites with insignificant trends increased as the timescale assessed increased. A greater number of sites had insignificant SPEI_PM trends than SPI trends at the same timescales because the SPEI_PM series had stronger self-correlations than the SPI series. The generally increasing trends in the drought indices implied a decrease in drought severity across the arid and semi-arid regions of Xinjiang. The drought evolution patterns in Xinjiang contradicted those at the global scale, as noted in Beguería et al. (2013).

In the meantime, regional differences were also shown between NX and SX through a comparison of different D_ET ranges, different correlations between the three drought indices or drought indices vs climatic variables, different trends and significances, and different spatial distributions. The overall comparison between the three different drought indices indicated that the SPI and SPEI_PM were both applicable in Xinjiang; however, the SPEI_PM better denotes drought severity in Xinjiang in the context of climate change.

Overall, there were three important points from this research which demonstrated new insights. First, this study clearly quantified the relative accuracy difference between ET_o,TW and the ET_o,PM in time and space for different sub-regions of Xinjiang, China. This knowledge may serve as a reference for the other arid regions. Second, our results revealed how the SPEI changed when ET_o was estimated by two different methods, i.e., the Thornthwaite (1948) and the FAO56 Penman-Monteith equation. The performances of the drought indices, i.e., SPI, SPEI_TW & SPEI_PM, may change in different climate regimes, e.g., humid or tropical conditions, but the results from this study are particularly useful for arid or semi-arid regions. Third, this study determined and ranked the sensitivities of the drought indices to different climatic variables over Xingjian. It also examined drought severity, evolution, distribution, and trend over the two sub-regions, i.e., NX and SX. This knowledge is critical for regional and local water resource management.

Drought is a complex phenomenon and difficult to pinpoint in time and space (Vicente-Serrano et al. 2011). This research provides a detailed comparison between three drought indices and their applications in Xinjiang. The outcome of this investigation, particularly new information of regional drought evolution, is essential for disaster prevention and management.

We acknowledge the financial support of the China National Science Foundation (U1203182, 51579213), the State Foundation for Studying Abroad (201506305014), and the China 111 project (B12007). Weather data were supplied by the Meteorological Data Sharing Service Network in China. The anonymous reviewers proposed very helpful comments for improving the paper.