Characteristics of integrated droughts based on a nonparametric standardized drought index in the Yellow River Basin, China

Drought is a natural recurring climate extreme that usually occurs over long time spans and spreads across a large spatial scale (National Research Council 2010). Compared with other natural hazards like flood and hurricane, drought is one of the costliest and least understood hazards, exerting huge adverse influences on ecological and social systems (Wilhite 2000; Hao & Aghakouchak 2014). As the Federal Emergency Management Agency (Federal Emergency Management Agency 1995) reported that annual losses caused by droughts in the USA is $6–8 billion, and the 1987–1989 drought occurring across a large portion of the USA led to the losses of about $39.4 billion in direct and indirect ways, which is obviously larger than that induced by hurricanes and floods (WGA 2004). Global warming is expected to result in intensified water cycles, and extreme events have been becoming more frequent, severe and pronounced (e.g. drought and flood) worldwide (Kunkel 2003; Roy & Balling 2004). Therefore, understanding the characteristics of droughts is very important for hazard adaptation and mitigation in a changing environment (Hao & Aghakouchak 2013; Huang et al. 2014a, 2014b).

Drought indices are important tools for drought characterization and prediction due to their abilities to simplify the complex interactions among various climate and climate-related variables. They allow scholars to quantitatively assess climate anomalies in terms of their frequency, duration, and severity, as well as spatial extent (Wilhite 2000). During the past decades, various indicators have been proposed to characterize droughts, and each has its own advantages and disadvantages (Mishra & Singh 2010). Some of the frequently used indexes are the standardized precipitation index (SPI; McKee et al. 1993, 1995), Palmer drought severity index (PDSI; Palmer 1965), crop moisture index (CMI; Palmer 1968), soil moisture drought index (SMDI; Hollinger et al. 1993), vegetation condition index (VCI; Liu & Kogan 1996), etc. Amongst these drought indicators, SPI is one of the most frequently utilized indices for meteorological drought assessment (Shukla et al. 2011; AghaKouchak & Nakhjiri 2012) due to its simplicity, standardized feature, and flexibility across various time scales (e.g., 1-, 3-, 6-, 12-month) (Hao & Aghakouchak 2014; Farahmand & Aghakouchak 2015).

Nevertheless, SPI has a potential weakness in that an appropriate parametric probability distribution function (PDF) has to be assumed in fitting precipitation time series (Angelidis et al. 2012). In practice, many issues will arise from the assumption that samples should follow a specific distribution. Typically, in constructing the SPI, the precipitation records are fitted to a gamma PDF which, however, may not be the most appropriate selection of distribution (Guttman 1999; Quiring 2009). Quiring (2009) pointed out that the developed SPI values, especially in the tails, were extremely sensitive to the selection of parametric distribution function. In fact, there exists no worldwide-accepted parametric distribution for meteorological and hydrologic variables (Silverman 1986; Smakhtin 2001), and the use of a specific parametric distribution is expected to lead to a large deviation for the low or high quantiles (Sharma 2000). Thus, a nonparametric method, which is distribution-free, is a good alternative to parametric approaches in developing drought indicators.

Furthermore, the existing available indices have some drawbacks including statistical incomparability and temporal inconsistency (Farahmand & Aghakouchak 2015). For instance, since SPI and PDSI have different scales, they cannot be directly compared with each other (Steinemann & Cavalcanti 2006). Additionally, droughts are categorized differently based on different types of shortages (e.g. the deficits of precipitation, streamflow, soil moisture, etc.) (Heim 2002). Univariate drought indicators can only focus on specific aspects of water shortages, leading to varying performances (e.g. in identification of the drought onset and termination). A meteorological drought (deficit in rainfall) may develop and end quickly, while the onset and termination of the corresponding hydrological drought (shortage in runoff) responds to the meteorological drought with some lag time (Pandey & Ramasastri 2001). Therefore, the current consensus amongst scholars is that no indicator based on a single variable is ideal for drought monitoring and prediction (AghaKouchak 2014). A holistic method for drought characterization needs an integrated indicator merging various sources of water shortages (precipitation, runoff, soil moisture, evapotranspiration, etc.). The advantage of standardized indicators is that they provide an opportunity to construct statistically consistent indicators on this basis of precipitation (SPI), runoff (SRI), soil moisture (SSI), etc. (Farahmand & Aghakouchak 2015). Therefore, characterizing droughts based on multiple climate variables using a nonparametric approach is very essential for developing spatially and temporally consistent drought indices, which is the major motivation of this study.

The Yellow River ranks the second largest river in China and the sixth largest river in the world in terms of its length (Shiau et al. 2007). In northern China, the Yellow River is a primary source of freshwater for nearly 107 million residents within the watershed, accounting for 8.7% of the total population in China (Wang et al. 2006). Meanwhile, the Yellow River Basin has been severely plagued by droughts from ancient times to the present (She & Xia 2013). Historically, one continuous drought hit the Yellow River Basin (YRB) in 1637–1643, and the extremely severe drought contributed to the collapse of the Ming Dynasty (Xie & Fu 2004). Another drought occurring between late 1920s and early 1930s had affected about 20 million people, and it is claimed that more than 3 million people died due to drought-related diseases and famine (Xie & Fu 2004). After the 1970s, the zero-flow phenomena which frequently occurred in the downstream of the Yellow River have gained wide attention. Highly frequent flow interruptions during the past 30 years have led to widespread unfavorable influences on local social and ecological development. Hence, it is of great importance to scientifically and reasonably monitor droughts based on a reliable drought indicator in the YRB.

Therefore, the primary objectives of this study are: (1) to develop a nonparametric multivariate standardized drought index (NMSDI) for comprehensively and effectively characterizing the drought conditions in the YRB; (2) to investigate the drought evolution characteristics including its trend and persistence in the future; (3) to identify possible change points of calculated NMSDI series, thus determining whether drought evolution in the YRB is non-stationary.

The observed gridded monthly precipitation and simulated monthly streamflow data based on the Variable Infiltration Capacity (VIC) model covering 1952–2012 in the YRB were utilized in this study. The VIC model (Liang et al. 1994) was applied to derive the land surface hydrologic fluxes/states. It is a semi-distributed macro-scale hydrologic model characterized by representing sub-grid variability of precipitation, soil moisture storage capacity, topography, vegetation classes, etc. (Liang et al. 1994; Nijssen et al. 1997). Meteorological forcing (e.g. precipitation, wind speed, temperature) for driving the VIC model was acquired from China Meteorological Administration (CMA), which was interpolated into 0.25 degrees. Land surface characteristics like vegetation, soil, and elevation were gained from Nijssen et al. (2001). Six parameters, i.e. the infiltration parameter b, the second and third soil layer depths (d₂, d₃), and the three parameters in baseflow scheme (D_m, D_s, W_s), were calibrated to match the long-term monthly streamflow observations. The dataset offers a current state-of-art estimate of land surface fluxes/states, and it is of great value in assessing the long-term water balance. Model calibration/validation was conducted for each river basin (including the Yellow River Basin) in China. In total, meteorological forcing from about 756 meteorological stations were interpolated into 0.25 degrees as inputs into the model, and streamflow records from 15 hydrological stations were used for model calibration/validations. An optimization algorithm of the multi-objective complex evolution of the University of Arizona (MOCOM-UA) (Yapo et al. 1998) was used to find the optimal parameter set. The Nash–Sutcliffe efficiency and the relative error between the simulated and observed mean annual streamflow were used to assess the model performance. Details for the model calibration/validation should be referred to Zhang et al. (2014).

The initial Mann–Kendall (MK) trend test method recommended by the World Meteorological Organization (Mitchell et al. 1966) is a nonparametric approach. However, the MK test results are expected to be impacted by the persistence of hydro-meteorological series. Hamed & Rao (1998) provided a modified Mann–Kendall (MMK) trend test via taking into account the lag-i autocorrelation to remove the persistence. Daufresne et al. (2009) pointed out the MMK test approach is more robust than the initial MK in capturing the trends of hydro-meteorological series. The detailed procedures of the MMK can refer to Huang et al. (2014a).

The R/S analysis was first developed by Hurst (1956) and applied to investigate the hydrological changes in the Nile River. Hereafter, R/S analysis has been extensively used in morphology, hydrology, cell reproduction, earthquake activity, etc. (Hosking 1984; Wang et al. 2013). The basic idea of the R/S analysis is to change the timescale of a specific series and explore its statistical features at different time scales (Oliver & Ballester 1996).

Different Hurst indexes represent different types of time series. For H = 0.5, it means this series is totally independent; for ⁠, it indicates the future trend of the series may be opposite to the existing series, and a smaller H value exhibits a stronger opposite persistence (Oliver & Ballester 1996). For ⁠, it denotes the future trend of the series may be consistent with the existing series, and a larger H means a stronger persistence. Therefore, this method was adopted to investigate the persistence of drought in the future in the YRB.

Existing statistical test methods like the Mann–Kendell test, sliding T test, sliding F test, and rank sum test are frequently utilized to identify change points. All of them strongly rely on the assumption that input data should be stationary. However, hydrologic series present greatly nonlinear features due to the high variability of the hydrologic process. Thus, these approaches are difficult to capture real change points. The heuristic segmentation method proposed by Pedro et al. (2001) is based on the idea of sliding T test and is modified to divide a non-stationary series into several stationary series, thus overcoming the limitation of the above methods.

Moreover, in order to further verify the reliability of the calculated NMSDI, the historical severe droughts which occurred in the middle of the YRB were collected from census data to further confirm the performance of NMSDI. As shown in Figure 2(b), the occurrence time of droughts matched well with the low NMSDI values. Two typical drought events which occurred in 2005 and 2006 were picked out (YRCC 2006, 2007), which are perfectly reflected by NMSDI exhibited in the area covered by the large black rectangle in Figure 2(b). The consistence between the timing of low NMSDI values and the years when historical drought events occurred demonstrates the reliability of the developed NMSDI in characterizing drought.

The comparison between NMSDI, SPI, and SSI in other sub-basins in the entire period also exhibits the same characteristics with sub-basin 26. For brevity, the related figures are omitted. The correlation coefficients between monthly NMSDI and SPI/SSFI spanning 1953–2012 in the eight subzones are shown in Table 1. It can be found that the correlation coefficients between monthly NMSDI and SPI/SSFI are very high, and all of them are significant at the 99% confident level. The results further verify the reliability and validity of the developed NMSDI. In view of the comprehensiveness and effectiveness of NMSDI in characterizing drought, it was applied to study the drought characteristics of the YRB in this study.

The average annual precipitation in the WRB decreased by 11.2% from 625 mm (before 1991) to 555 mm (after 1991). The decrease of precipitation availability has a certain contribution to the change point in the WRB. In addition, Huang et al. (2014c) pointed out that the increasing air temperature and the initially decreasing then increasing potential evaporation had a certain effect on the change points of the relationship between rainfall and runoff in the WRB. Since the NMSDI is calculated based on precipitation and runoff, the variations of temperature and potential evaporation may partly lead to the change point of annual NMSDI series in the WRB. However, the quantitative impact of temperature and potential evaporation on the change point of annual NMSDI series need to be further investigated in future studies.

In addition to climate change, anthropogenic interferences may play an important role in leading to the change point. Soil loss is greatly severe in the WRB. Hence, the soil conservation practices including grass-planting, afforestation, building check dams, and creation of level terraces have been performed since 1960. The detailed areas of soil conservation practices are shown in Table 2.

The area of soil conservation is increasing, and the growth rate has rapidly increased after 1990, indicating that local anthropogenic interferences on underlying surface have greatly intensified since 1990. The soil conservation practices alter local microtopography, intercept precipitation, facilitate infiltration rate of water flow, and retain or slow down streamflow, thus reducing streamflow (Chang et al. 2015). Furthermore, with the deepening of China's reform and opening up, a rapid economic development has been achieved since 1990. The mean economic growth speed in the WRB is about 10.5%. The high-speed economy requires numerous water resources to support its sustainable development. The average industrial and domestic water consumption in the WRB after 1990 is about 4.3 billion m³/year, which increased by 52.6% compared with that before 1990 (Huang & Fan 2013), directly resulting in the decrease of runoff in the WRB. Therefore, the coupled impacts of climate change and human activities contribute to the change point in the WRB.

It is well demonstrated in the literature that future drought will become more frequent, severe, and pronounced in most places with global warming. However, the declining drought tendency as shown in this study are based on the Hurst index of the R/S analysis which is conducted by changing the timescale of a specific series and exploring its statistical features at different time scales (Oliver & Ballester 1996). Therefore, the calculated persistence is a projection of the future possible conditions in a probabilistic way in nature. Although the R/S analysis has been extensively applied in morphology, hydrology, cell reproduction, earthquake activity, etc. (Hosking 1984; Wang et al. 2013), the persistence based on this method cannot be used directly for future drought projections. Instead, to better project the future droughts, more physically based approaches such as the global climate model (GCM) should be adopted.

A reliable drought indicator is important for drought mitigation and water resources management. This study introduces a nonparametric multivariate drought index (NMSDI) coupled with the information of precipitation and streamflow, which is distribution free and can eliminate the limitation of parametric approaches. The integrated drought index was used to characterize drought in the YRB on seasonal and annual scales. The primary results are as follows:

1. The variations of 1-month NMSDI is well consistent with that of 1-month SPI and SSI, indicating the reliability and effectiveness of the nonparametric method in developing an integrated drought index. Generally, NMSDI is more sensitive to identify drought onset, persistence and termination.

2. The upstream and northern portion of the basin are drier than other areas. The seasonal NMSDI pattern is similar to the annual NMSDI pattern to a certain degree. Overall, the drought risk in spring and summer is larger than autumn and winter.

3. The YRB is dominated by a decreasing trend of NMSDI at seasonal and annual scales, indicating an increasing risk of drought during the past several decades. However, the decreasing trend of NMSDI may be the opposite in the future, which suggests that drought risk may reduce in the basin.

4. Annual NMSDI series in the YRB is relatively stable, except for the WRB in which a change point (1991) was identified mainly due to the combined impacts of climate change and human activities.

This research was supported by the National Major Fundamental Research Program (2011CB403306-2), the National Natural Fund Major Research Plan (51190093) and the Natural Science Foundation of China (51179148, 51179149, and 51309188).