Hydrological drought is not only caused by natural hydroclimate variability but can also be directly altered by human interventions including reservoir operation, irrigation, groundwater exploitation, etc. Understanding and forecasting of hydrological drought in the Anthropocene are grand challenges due to complicated interactions among climate, hydrology and humans. In this paper, five decades (1961–2010) of naturalized and observed streamflow datasets are used to investigate hydrological drought characteristics in a heavily managed river basin, the Yellow River basin in north China. Human interventions decrease the correlation between hydrological and meteorological droughts, and make the hydrological drought respond to longer timescales of meteorological drought. Due to large water consumptions in the middle and lower reaches, there are 118–262 % increases in the hydrological drought frequency, up to 8-fold increases in the drought severity, 21–99 % increases in the drought duration and the drought onset is earlier. The non-stationarity due to anthropogenic climate change and human water use basically decreases the correlation between meteorological and hydrological droughts and reduces the effect of human interventions on hydrological drought frequency while increasing the effect on drought duration and severity. A set of 29-year (1982–2010) hindcasts from an established seasonal hydrological forecasting system are used to assess the forecast skill of hydrological drought. In the naturalized condition, the climate-model-based approach outperforms the climatology method in predicting the 2001 severe hydrological drought event. Based on the 29-year hindcasts, the former method has a Brier skill score of 11–26 % against the latter for the probabilistic hydrological drought forecasting. In the Anthropocene, the skill for both approaches increases due to the dominant influence of human interventions that have been implicitly incorporated by the hydrological post-processing, while the difference between the two predictions decreases. This suggests that human interventions can outweigh the climate variability for the hydrological drought forecasting in the Anthropocene, and the predictability for human interventions needs more attention.
Drought is a natural phenomenon occurring due to climate variability that is associated with oceanic and/or terrestrial anomalies (Hong and Kalnay, 2000; Hoerling and Kumar, 2003). As the rainfall deficit reaches a certain threshold, a meteorological drought occurs. If the meteorological drought persists for a period of time, it will decrease the soil moisture and river flow, resulting in agricultural and hydrological droughts. Although the rainfall deficit is a primary driver for agricultural and hydrological droughts, terrestrial hydrological processes (e.g., snow melting, evapotranspiration) and geological and topographic conditions also play a non-trivial role in the drought propagation (Van Loon et al., 2012; Rimkus et al., 2013; Teuling et al., 2013; Stoelzle et al., 2014; Staudinger et al., 2015). Therefore, monitoring and forecasting of agricultural and hydrological droughts not only provide more relevant guidelines for the management of agricultural and water resources but also raise challenging scientific questions on the mechanism and predictability of multi-scale drought processes (Pozzi et al., 2013; Yuan et al., 2013; Wood et al., 2015).
Given that rainfall deficit is a major cause of the hydrological drought, many studies focus on understanding the propagation from meteorological to hydrological drought. For instance, Vicente-Serrano and López-Moreno (2005) investigated the relationship between streamflow and antecedent rainfall by using a correlation analysis, and found that hydrological drought index series have the highest correlation with a 2-month accumulated meteorological drought index series in a mountainous Mediterranean basin. Such a correlation analysis method was then widely used to understand the timescale of hydrological drought (Rimkus et al., 2013; Haslinger et al., 2014; Bloomfield et al., 2015; Folland et al., 2015; Niu et al., 2015; Kumar et al., 2016). A recent study from Barker et al. (2016) comprehensively investigated the relationship between meteorological and hydrological droughts for 121 near-natural catchments in the United Kingdom, where the relationship was found to be associated with natural climate and catchment properties.
Besides natural climate and hydrological processes that affect the development of hydrological drought, human activities such as land-use and land-cover change, irrigation, reservoir operation and groundwater exploitation can also influence hydrological drought significantly (AghaKouchak et al., 2015; Van Loon et al., 2016a). López-Moreno et al. (2009) found that the second largest reservoir in Europe increased the duration and severity of hydrological drought over downstream areas. Similar studies found that reservoir regulation might reduce the drought severity over upstream areas but increase it over downstream areas in Australian and Chinese catchments (Wen et al., 2011; Zhang et al., 2015), as many reservoirs were built for meeting the irrigation demand more reliably. However, there is limited knowledge on the integrated impact of human activities on hydrological drought processes over a large river basin due to the lack of human water-use data, which is one of the major issues that hinders the understanding of hydrological drought in the Anthropocene (Van Loon et al., 2016b). An alternative approach is to use a land surface hydrological model (Wada et al., 2013; Zhou et al., 2016) or a less complicated water balance model to recover the naturalized streamflow by assimilating the reported water-use data. The naturalized streamflow data can be used to calibrate the hydrological model without human components to investigate the natural response of hydrological processes to climate variations (Yuan et al., 2016); it can also be used to investigate the integral anthropogenic impact on hydrological drought by comparing with observed streamflow time series.
The ultimate goal for the understanding of drought processes is to facilitate the development of drought early warning systems for drought adaptation and mitigation. Fortunately, the development of ocean–atmosphere–land coupled general circulation models (CGCMs) provides an unprecedented opportunity to transform advances in seasonal forecasting research (Kirtman et al., 2014; Yuan et al., 2015a) into an integrated drought service. Besides meteorological drought forecasts (Dutra et al., 2012; Yuan and Wood, 2013; Ma et al., 2015), agricultural drought forecasts with dynamical seasonal climate forecast models have also been widely applied and evaluated (Luo and Wood, 2007; Mo et al., 2012; Sheffield et al., 2014; Yuan et al., 2015b; Thober et al., 2015). However, dynamical forecasting of hydrological drought based on the CGCM–hydrology coupled approach (Yuan et al., 2015a) has received less attention (Trambauer et al., 2015; Sikder et al., 2016), although there are many statistical forecasting studies for low flows. The reasons are threefold: (1) a skilful seasonal forecasting of streamflow usually occurs over basins with large storages of snow, surface and/or subsurface water (Wood and Lettenmaier, 2008; Koster et al., 2010), and strong control from initial hydrological conditions limits the added value from climate predictions (Wood et al., 2016; Yuan, 2016); (2) unlike meteorological drought forecasts, both agricultural and hydrological drought forecasts are influenced by the uncertainty in the hydrological model, and the hydrological drought forecasting tends to be more challenging since the errors from upstream areas can be transferred to or even amplified in downstream areas; and (3) many river basins are altered by human activities, where the management impacts are often neglected in most dynamical forecasting systems. In fact, seasonal forecasting of hydrological drought in the Anthropocene raises the questions of how to define the predictability of the anthropogenic processes within a coupled hydroclimate system, how to distinguish the uncertainty from each component, and how to assess the forecast skill of hydrological drought with natural and anthropogenic forcings.
Locations of hydrological stations in the Yellow River basin. Shaded
areas are regional mean annual rainfall (mm day
This paper focuses on the understanding and seasonal forecasting of hydrological drought over a heavily managed river basin in north China, the Yellow River basin. Both naturalized and observed streamflow along the mainstream of the Yellow River will be used to (i) investigate the relationship between meteorological and hydrological droughts under natural and anthropogenic conditions, (ii) to quantify the influence of human activities on the characteristics of hydrological drought (e.g., drought frequency, duration and severity, and seasonality of hydrological drought onset), and (iii) to assess hydrological drought forecasting in the Anthropocene with an experimental seasonal hydrological forecasting system established over the Yellow River basin (Yuan et al., 2016).
The precipitation dataset at 0.25
Monthly mean rainfall (mm day
Figure 1 shows the locations of 12 mainstream hydrological gauges used in this
study, with Tangnaihai gauge in the headwater region and Lijin gauge at the
outlet of the entire Yellow River basin that has a drainage area of
Monthly mean naturalized (NAT, blue lines) and observed (OBS, red lines) streamflow
(10
Annual mean naturalized (NAT, blue bars) and observed (OBS, red bars) streamflow
(10
Except for the Tangnaihai gauge in the headwater region, streamflow at the hydrological gauges used in this study was influenced by human interventions (Fig. 3). Therefore, the Yellow River is an ideally large river basin to investigate the hydrological drought processes and predictability in the Anthropocene. In general, human interventions decrease streamflow over upper and middle reaches of the Yellow River during rainy seasons, while increase it during dry seasons (Fig. 3b–h). This suggests that reservoirs in the upper and middle reaches of the Yellow River store rain water in wet season and distribute it in the remaining time of the year according to the need, which is similar to regulations in other parts of the world (Wada et al., 2014). Actually, Fig. 4a shows that the annual mean observed streamflow at upper reaches can be higher than the naturalized streamflow during dry years due to the reservoir water release (e.g., years 2000, 2002, 2006 and 2010 for the Lanzhou gauge). Over the lower reaches, the observed streamflow is significantly lower than the naturalized streamflow during wet season due to heavy water consumption, and riverbed leakages because of groundwater overdraft and possible geomorphology change caused by sediment accumulation (Fig. 3i–l). The observed streamflow is close to the naturalized streamflow during dry seasons because of no significant water consumption or reservoir management. Figure 4 also shows that the magnitudes of reservoir storage changes are quite small as compared with streamflow. In fact, the mean absolute changes of reservoir storage during 1998–2010 are about 14–38 and 12–14 % of observed and naturalized streamflow, respectively. This suggests that other human interventions, such as direct withdrawal of surface water for agricultural, industrial and civil consumptions, account for a large part of streamflow variations in the Yellow River.
The Standardized Precipitation Index (SPI; McKee et al., 1993) was used as the meteorological drought index. The main advantage of SPI is its multiscale nature, i.e., it can be used to represent meteorological drought at different timescales. In this study, sub-basin mean precipitation datasets averaged over antecedent 1 to 24 months were used to calculate SPI-1 to SPI-24. To account for the seasonality, the SPI for each target month and for each timescale was calculated separately by using 50-year data. For example, SPI-6 for October 1982 was calculated by firstly fitting an empirical distribution based on the precipitation averaged between May and October during 1961–2010, and the May–October mean precipitation in 1982 was then used to determine the SPI-6 value for October 1982.
Similarly, monthly naturalized streamflow datasets at 12 hydrological gauges
during 1961–2010 were also standardized by using the same procedure,
resulting in a hydrological drought index named as the Standardized Streamflow
Index (SSI). Note that SSI was similar to the standardized runoff index (SRI)
defined by Shukla and Wood (2008), except that streamflow is used here for a
standardization. For the anthropogenic streamflow, the same parameters of
probabilistic distributions fitted from the naturalized streamflow were used,
and the SSI values were then calculated. Both the 50-year (1961–2010) data
and a series of 30-year transient climatologies (e.g., a 1961–1990 climatology
was used for year 1975, a 1962–1991 climatology was used for 1976, and so on)
were used to calculate the stationary and non-stationary SPI and SSI values,
respectively. A threshold of
Correlations between Standardized Precipitation Index (SPI) at different timescales and monthly naturalized (blue) or observed (red) Standardized Streamflow Index (SSI) for 12 hydrological gauges located from the upper to lower mainstream of the Yellow River. The solid lines are for SPI and SSI calculated based on transient 30-year climatologies with consideration of non-stationarity, while the dashed lines are for those based on the 50-year (1961–2010) climatology.
A number of seasonal hydrological ensemble hindcast datasets created by Yuan (2016)
were used in this study. To have this paper self-contained, the
hindcast experiments are briefly described below. Firstly, a continuous
offline hydrological simulation with calibrated VIC model and river routing
model driven by observed meteorological forcings from 1951 to 2010 was
conducted to generate the initial hydrological conditions (ICs) for the
hydrological hindcasts (Yuan et al., 2016). The observed meteorological
forcing datasets including daily precipitation, daily maximum and minimum
surface air temperature, and surface wind were interpolated from 324 China
Meteorological Administration stations. The VIC model (version 4.0.5) was used
to predict runoff in a water balance mode over the the entire Yellow River basin
with 1321 grid cells at 0.25
The same as Fig. 5 but for each target month for five selected hydrological gauges.
Figure 5 shows the Pearson correlation coefficients between SPI at different
timescales and monthly SSI both for naturalized and observed streamflow.
There is an increase in correlation as the SPI timescale increases,
suggesting that streamflow is not only influenced by concurrent
precipitation but also by antecedent precipitation up to a few months.
Similar to Vicente-Serrano and López-Moreno (2005), the timescale with
the maximum correlation is considered as the timescale of SPI that
streamflow responds to. For the naturalized streamflow, it responds to 6–12 months SPI over the upper and middle
reaches of the Yellow River, and to about 4 months SPI over the lower reaches. The correlations for observed streamflow
are significantly lower than for naturalized streamflow for gauges from
Xunhua down to Huayuankou, with
Time series of naturalized (blue) and observed (red) 1-month
Standardized Streamflow Index (SSI) for five selected hydrological gauges.
SSI is calculated based on transient 30-year climatologies. The horizontal
black lines represent the threshold of
In order to analyze the relation during different seasons, five gauges from upper to lower reaches are selected and the correlations between SPI at different timescales and monthly SSI for different target months are plotted in Fig. 6. It is found that streamflow responds to the shorter timescale of SPI in wet and warm seasons and to the longer timescale in dry and cold seasons. Again, the correlations for observed streamflow for different target months are lower than that for naturalized streamflow, except for the headwater region (i.e., Tangnaihai gauge) without significant human interventions (Fig. 6). The differences are larger during summer seasons than winter seasons, which is consistent with the seasonality of human water use as shown by different streamflow in Fig. 3. Again, non-stationarity basically weakens the relation between meteorological and hydrological droughts over lower reaches, regardless of season.
To demonstrate the effect of human interventions on streamflow variations directly, both the time series of naturalized and observed SSIs calculated based on transient climatologies are plotted for the five selected gauges in Fig. 7. It is found that human interventions sometimes have a positive influence on hydrological drought over the upper reaches. For example, observed SSI can be larger than naturalized SSI at the Lanzhou gauge (Fig. 7b). However, those increases mostly occur in winter seasons, while they do decrease in summer season when the water demand and water consumption are high. For the middle and lower reaches, observed SSIs are basically lower than naturalized SSIs (Fig. 7d–e), indicating that human interventions exacerbate hydrological drought conditions in the lower reaches of the Yellow River basin. The results based on 50-year climatology generally have lower SSI in 1990s–2000s and higher SSI in 1960s–1970s (not shown), suggesting there are decreasing trends in streamflow over the Yellow River basin that is consistent with previous studies (e.g., Piao et al., 2010).
By following the definition of hydrological drought event (it should last for at least 3 continuous months) in Sect. 2.2, the frequency, duration and severity of hydrological droughts under natural and anthropogenic conditions are calculated and shown in Fig. 8. Under natural conditions, seasonal hydrological drought occurs 8–16 times during the 50-year (1961–2010) period, where the frequency of hydrological drought is not necessarily higher over lower reaches than that over upper reaches (Fig. 8a). This is partly because there is more precipitation over lower reaches, and the naturalized hydrological drought basically represents the response to the meteorological drought. In contrast, the observed hydrological drought frequency shows quite different characteristics, where the human interventions increase drought frequency by up to 65 % from the upper gauges down to the Xiaheyan gauge (the 5th gauge in Fig. 8a), but they increase the drought frequency by 118 or even 262 % over the lower reaches (e.g., Lijin gauge at the outlet).
Characteristics of hydrological drought events based on streamflow
time series with or without human influences during 1961–2010 for
12 hydrological gauges. A hydrological drought event is selected when the
Standardized Streamflow Index (SSI) is continuously below
Ratios of the number of hydrological drought onsets occurring in
different seasons to the total number of hydrological drought events for the
12 hydrological gauges during 1961–2010. Drought events are classified the
same as those in Fig. 8. The left panels
For the drought duration in natural conditions, it is generally longer over upper reaches, which is again due to a drier climate (Fig. 8b). With human interventions, there is a slight decrease in drought duration for the upper gauges down to Xiaheyan gauge. This suggests that human interventions reduce the persistency of drought over the upper reaches of the Yellow River basin. From Shizuishan gauge (the 6th gauge in Fig. 8b) down to the outlet, the duration of hydrological drought increases by 21–99 % under human interventions. There is no significant change in drought severity down to Xiaheyan gauge (the 5th gauge in Fig. 8c), but the severity increases by up to 8 times over the lower reaches (Fig. 8c). Therefore, human interventions not only increase drought frequency over most areas of the Yellow River basin but also increase drought severity dramatically over the lower reaches. As compared with the results based on a constant climatology (cyan and pink bars in Fig. 8), the non-stationarity generally reduces human influence on drought frequency, but increases the human impact on drought duration and severity.
To investigate the preference of the occurrence of seasonal hydrological drought, the onset seasons are identified based on individual hydrological drought events. And the ratios of number of drought onsets during different seasons to the total number of drought events are plotted in Fig. 9. Without human interventions, most seasonal hydrological droughts start in summer, especially for the upper and lower gauges (Fig. 9a–b). While for the gauges over middle reaches, autumn is also a preferred season for drought onset (Fig. 9a–b). However, Fig. 9 shows that human intervention changes the seasonality of hydrological drought onset significantly, where the spring is the preferred season for the drought onset for the lower reaches, and the ratio for summer season is also increased for the upper reaches. With human interventions, the hydrological drought onset is earlier, no matter what climatology (transient or constant) you use (Fig. 9c–d).
To demonstrate the capability of predicting hydrological drought in the
Anthropocene, a drought case of 2001 is selected to verify the ensemble
forecasting of SSI. In terms of meteorological condition, 2001 is a moderate
dry year, with precipitation less than the climatology by 9.4 %. However,
2001 is a severe hydrological drought year, with observed streamflow less
than the climatology by 26–37 % over the upper reaches and by
51–86 % over the lower reaches (
Seasonal ensemble hindcast of the 2001 Yellow River hydrological
drought from upstream to downstream gauges by using a climatology method
(ESP/VIC) and the climate-model-based approach (NMME/VIC). Vertical axes are
Standardized Streamflow Index (SSI), where SSI <
The results shown in Fig. 10 neglect the errors or uncertainties in
the hydrological model because the offline simulated hydrological drought index is
used as reference to verify the forecasts. To compare with observed SSI, the
forecast streamflow series are post-processed (see Sect. 2.3 for details),
and the results for 2001 are shown in Fig. 11. Both ESP/VIC and NMME/VIC
can predict a drought condition with ensemble mean SSIs much lower than
To assess the probabilistic forecast skill for hydrological droughts during
the hindcast period of 1982–2010, the Brier score (BS; Wilks, 2011) is used.
It is defined as
The same as Fig. 10 but for the post-processed Standardized Streamflow Index (SSI) hindcasts from ESP/VIC and NMME/VIC (see Sect. 2.3 for details) as verified by observed SSI.
Figure 12 shows the BS for ensemble hydrological drought forecasts at different hydrological gauges and at different lead times. The statistics are based on a set of comprehensive hindcasts, where for each month during the 29-year (1982–2010) period there is a 6-month hindcast. This consists of 348 hindcast cases, each one is 6-months long, and has 28 or 99 realizations for ESP/VIC and NMME/VIC, respectively. For the natural condition (upper panels of Fig. 12), the BS values increase as the forecast leads increase, indicating that the performance generally decreases over leads. The performance of probabilistic hydrological drought forecasting is better over lower reaches than that over upper reaches both for ESP/VIC and NMME/VIC, suggesting the influence of catchment memory. As compared with ESP/VIC, NMME/VIC has a better performance, with BS decreased by 11–26 % in the first month and by 3–14 % in the second and third months.
As verified against observed SSI (lower panels of Fig. 12), the performances
for ESP/VIC and NMME/VIC are surprisingly better than the natural cases, and
the performances do not necessarily degrade over forecast lead times. This
is because human interventions increase the occurrence and severity of
hydrological drought, and outweigh the climate variations in many cases. The
hydrological post-processing imparts the first-order control in the
forecasting, and many post-processed forecasts can represent drought
conditions (SSI <
Brier score (BS) for ensemble hydrological drought forecasts at
different lead times from the climatology method (ESP/VIC) and the
climate-model-based approach (NMME/VIC) for different hydrological gauges as
verified against VIC offline simulated
This paper investigates the effects of human interventions on hydrological
drought processes and forecasting. Naturalized and observed monthly
streamflows are standardized to calculate the hydrological drought index, the
Standardized Streamflow Index (SSI). Comparison between naturalized and
observed SSI at 12 hydrological gauges along the mainstream of the Yellow
River basin (the second largest river basin in China with a drainage area of
7.52
The naturalized streamflow datasets are used to calibrate the VIC land surface hydrological model and the routing model, and both climatological forcings (ESP) and climate model predicted forcings (NMME) are used to drive the hydrological models to provide seasonal streamflow forecasts. For a severe hydrological drought event that occurred over the Yellow River in 2001, ESP/VIC does not capture it while NMME/VIC has some skill when they are verified against naturalized SSI. The added values from climate-model-based seasonal hydrological drought forecasting are decreased in the Anthropocene, where both methods can predict a drought condition after the hydrological post-processing but underestimate the severity. Unlike those in naturalized hydrological drought forecasting, the ensemble spreads do not necessarily increase over forecast leads in the Anthropocene because of the seasonality of human interventions that have been implicitly incorporated in the hydrological post-processing. Based on the assessment of all hindcasts during 1982–2010, it is found that NMME/VIC decreases (improves) the BS against ESP/VIC by 11–26 % in the first month and by 3–14 % in the second and third months for the probabilistic hydrological drought forecasting in the naturalized conditions. In the Anthropocene, the performances of both forecast methods become better in terms of BS, and the forecast skill does not necessarily decrease over forecast leads due to the dominant influence of human water consumption on the hydrological drought processes.
While the effects of human interventions on hydrological drought processes have been studied in the past, to our knowledge, this study is among the first to investigate seasonal hydrological drought forecasting in the Anthropocene. Intensive and direct influence of human water use challenges our understanding of hydrological drought predictability. Traditionally, hydroclimate predictability usually refers to the struggle between deterministic and chaotic physical processes. In the Anthropocene, the human influence sometimes outweighs those natural hydroclimate variations and variability, and can be a major source of predictability (or uncertainty if it is not fully understood) for hydrological drought. Current hydrological post-processing procedures account for the seasonality of human water use by adding or deducting water to or from the predicted streamflow, and adjusting the forecast results based on the Bayesian theory. Another popular method is to parameterize the human interventions directly in the hydrological models (e.g., Wada et al., 2014; Zhou et al., 2016 among others), where the irrigated area can be estimated for crop water demand, and the diversion and return flows can also be simulated in those models. This modeling framework could be further pushed forward by the availability of Surface Water and Ocean Topography (SWOT) satellite data in the near future, where surface water storage in reservoirs and rivers can be monitored at 250 m resolution (Biancamaria et al., 2016). However, it is difficult to account for the interannual variability of human water use in a “real” forecasting mode. During severe hydrological drought events, human water use tends to be more intensive to adapt to drought conditions. How to quantify and model the variability of human water use is an interdisciplinary question for both physical and social sciences. In addition, the interdisciplinary collaboration is also indispensable to objectively quantify and to accurately predict drought impact, as the drought impact and drought are quite different.
Seasonal climate hindcast data from NMME models can be
accessed at the IRI website
(
The authors declare that they have no conflict of interest.
This article is part of the special issue “Observations and modeling of land surface water and energy exchanges across scales: special issue in Honor of Eric F. Wood”. It is a result of the Symposium in Honor of Eric F. Wood: Observations and Modeling across Scales, Princeton, New Jersey, USA, 2–3 June 2016.
We would like to thank Dennis Lettenmaier and three anonymous reviewers for their helpful comments. This work was supported by the National Natural Science Foundation of China (no. 91547103), China Special Fund for Meteorological Research in the Public Interest (Major projects) (GYHY201506001), National Key R&D Program of China (2016YFA0600403), and the Thousand Talents Program for Distinguished Young Scholars. This paper is submitted to the special issue “Observations and modeling of land surface water and energy exchanges across scales” in honor of Eric F. Wood. The first author would like to thank Eric for his inspiring guidance. Edited by: Dennis Lettenmaier Reviewed by: three anonymous referees