Hess 28 3305 2024

Hydrol. Earth Syst. Sci.
, 28, 3305–3326, 2024

https://doi.org/10.5194/hess-28-3305-2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.
Machine-learning-constrained projection of bivariate hydrological

drought magnitudes and socioeconomic risks over China
Rutong Liu1 , Jiabo Yin1 , Louise Slater2 , Shengyu Kang1 , Yuanhang Yang1 , Pan Liu1 , Jiali Guo3,4 , Xihui Gu5 ,
Xiang Zhang6 , and Aliaksandr Volchak7
1 StateKey Laboratory of Water Resources Engineering and Management, Wuhan University,
Wuhan, Hubei, 430072, PR China
2 School of Geography and the Environment, University of Oxford, Oxford, UK
3 Hubei Key Laboratory of Construction and Management in Hydropower Engineering, China Three Gorges University,
Yichang 443002, Hubei Province, China

4 College of Hydraulic and Environmental Engineering, China Three Gorges University, Yichang, Hubei 443002, PR China
5 School of Environmental Studies, China University of Geosciences, Wuhan 430074, PR China
6 National Engineering Research Center of Geographic Information System, School of Geography and Information
Engineering, China University of Geosciences, Wuhan 430074, PR China

7 Engineering Systems and Ecology Faculty, Brest State Technical University, Moskovskaya 267, 224017 Brest, Belarus
Correspondence: Jiabo Yin ([email protected])
Received: 22 July 2023 – Discussion started: 6 September 2023

Revised: 31 May 2024 – Accepted: 10 June 2024 – Published: 25 July 2024
Abstract. Climate change influences the water cycle and al- Results showed that our hybrid hydrological–deep-learning
ters the spatiotemporal distribution of hydrological variables, model achieved > 0.8 Kling–Gupta efficiency in 161 out
thus complicating the projection of future streamflow and hy- of the 179 catchments. By the late 21st century, bivariate
drological droughts. Although machine learning is increas- drought risk is projected to double over 60 % of the catch-
ingly employed for hydrological simulations, few studies ments mainly located in southwestern China under SSP5-85,
have used it to project hydrological droughts, not to mention which shows the increase in drought duration and severity.
bivariate risks (referring to drought duration and severity) as Our hybrid model also projected substantial GDP and pop-
well as their socioeconomic effects under climate change. We ulation exposure by increasing bivariate drought risks, sug-
developed a cascade modeling chain to project future bivari- gesting an urgent need to design climate mitigation strategies
ate hydrological drought characteristics in 179 catchments for a sustainable development pathway.
over China, using five bias-corrected global climate model
(GCM) outputs under three shared socioeconomic pathways
(SSPs), five hydrological models, and a deep-learning model.
We quantified the contribution of various meteorological 1 Introduction
variables to daily streamflow by using a random forest model,
and then we employed terrestrial water storage anomalies In a warming world, the change in the global water cycle is
and a standardized runoff index to evaluate recent changes in expected to alter the regional and seasonal distribution of key
hydrological drought. Subsequently, we constructed a bivari- hydrological variables such as precipitation and evapotran-
ate framework to jointly model drought duration and sever- spiration (Allan et al., 2020; Yin et al., 2023b). As precipi-
ity by using copula functions and the most likely realiza- tation patterns are particularly sensitive to changes in atmo-
tion method. Finally, we used this framework to project fu- spheric forcing and local conditions, precipitation extremes
ture risks of hydrological droughts as well as the associated are generally increasing globally, exacerbating the spatial
exposure of gross domestic product (GDP) and population. heterogeneity of precipitation (Donat et al., 2016; Tabari,
2020). A suite of shared socioeconomic pathways (SSPs) has
Published by Copernicus Publications on behalf of the European Geosciences Union.

3306 R. Liu et al.: Machine-learning-constrained projection of bivariate hydrological drought magnitudes
been proposed to simulate different possible future scenar- incompleteness of a single-variable analysis (Ayantobo et al.,
ios of social responses to climate change, and these are em- 2017; Nabaei et al., 2019). At present, studies on hydrolog-
ployed to investigate the possible effects of long-term climate ical drought within a bivariate framework are still lacking.
change (Meinshausen et al., 2020; Zhang et al., 2021). By us- Beyond the choice of approach (univariate or bivariate), the
ing the SSP framework, numerous works have indicated that Gravity Recovery and Climate Experiment (GRACE) and
the redistribution of precipitation may lead to a decline in GRACE Follow-On (GRACE-FO) satellites now provide 2
water storage in some regions and intensify water scarcity in decades of large-scale terrestrial water storage (TWS) data,
arid regions (Sönmez and Kale, 2018; Woolway et al., 2020; which capture the water deficit in various forms on land and
Yao et al., 2023). With increasing atmospheric greenhouse which can be used to monitor droughts (Schmidt et al., 2006).
gases, numerous studies have reported a widespread increase The drought severity index based on TWS (TWS-DSI) can be
in drought events, even in areas with increasing annual runoff used to monitor past drought events and also shows potential
(Dai et al., 2018). The rapidly changing distribution of pre- advantages in drought warning, forecasting, and projection
cipitation and other meteorological elements under climate (Nie et al., 2018; Pokhrel et al., 2021).
change complicates projection of future runoff and drought. In recent decades, many studies have used bias-corrected
China’s socioeconomic development, particularly of its outputs from global climate models (GCMs) to project fu-
agricultural sector, is threatened by the rapid intensification ture hydrological drought scenarios (e.g., Ashrafi et al., 2020;
of extreme hazards under climate change (Piao et al., 2010). Dixit et al., 2022; Kim et al., 2021). The growing applica-
Over the past few years, China has been hit by severe drought tion of machine learning has revealed a high potential for
events which have caused considerable damage to ecosystem improving the accuracy of hydrological simulation and pre-
productivity and socioeconomic growth (Yin et al., 2023a; diction (Mokhtar et al., 2021). In recent years, many machine
Zhai and Zou, 2005). For instance, one extreme drought in learning algorithms have been adopted in drought simulation
Sichuan Province in 2022 resulted in power shortages and and have produced good performance, such as wavelet neural
led to economic losses of USD 669 million. Water shortage networks (WNNs) (Xiujia et al., 2022), support vector ma-
is also a key challenge that hinders the sustainable develop- chines (SVMs) (Zhu et al., 2021), and long short-term mem-
ment of the North China Plain (Chen and Yang, 2013). Over ory (LSTM) neural networks (Dikshit et al., 2021a). These
the period of 1985–2014, drought accounted for about 19 % algorithms can be used to simulate the evolution of future
of economic losses from all meteorological hazards (Chen droughts and construct risk maps for drought contingency
and Sun, 2019). With continuing global warming, the eco- planning (Rahmati et al., 2020). Among the different models,
nomic losses from severe drought events might increase by the LSTM can effectively simulate short-term and long-term
over USD 10 billion per year by the late 21st century, un- streamflow series, and their performances have been vali-
derscoring the importance of projecting future droughts over dated at short temporal scales (Dikshit et al., 2021b; Kang
China (Lu et al., 2023). et al., 2023).
Droughts can be triggered by divergent mechanisms and In this study, we project changes in bivariate hydrologi-
are thus distinguished according to the type of drought, such cal drought characteristics (duration and severity) and their
as meteorological or hydrological drought (Yihdego et al., associated socioeconomic risks under three SSPs (i.e., SSP1-
2019). The majority of studies have focused on meteorolog- 26, SSP3-70, and SSP5-85) over 179 catchments in China.
ical droughts, which can then be translated to a hydrologi- To achieve this, we combine five hydrological models and a
cal drought, while fewer works have focused on hydrolog- deep-learning model (i.e., the LSTM) and then drive the hy-
ical drought, probably due to a lack of measurements like brid models with the five bias-corrected GCM outputs in the
in the standardized runoff index (SRI) (Barker et al., 2016; sixth phase of the Coupled Model Intercomparison Project
Kumar et al., 2016; Tirivarombo et al., 2018). Furthermore, (CMIP6). Then, we employ a machine-learning-based frame-
hydrological droughts are affected not only by the water cy- work (i.e., a random forest (RF) model) to quantify the
cle, but also by human interventions, which makes them dif- sensitivity of daily streamflow to different meteorological
ficult to predict accurately (Wu et al., 2021). Currently, the variables. We employ the run theory and two drought met-
majority of drought impact assessments focus on the inves- rics, the SRI and TWS-DSI, to identify and explore recent
tigation of individual drought variables (i.e., drought dura- changes in drought characteristics. In addition, we use cop-
tion, severity, and intensity) through univariate probabilistic ula functions to build the bivariate model of drought dura-
models and stochastic theory (Byakatonda et al., 2018; My- tion and severity during both the reference and future pe-
ronidis et al., 2018; Zhang et al., 2022). However, univariate riods. After identifying shifts in bivariate drought charac-
drought analysis cannot accurately describe the probability teristics based on the most likely realization approach, we
of drought events, because droughts of either long duration or project the exposure of the gross domestic product (GDP)
severe intensity can lead to substantial socio-ecosystem dam- and population to increasing drought risks in the future. Fi-
ages (Castle et al., 2014; Udall and Overpeck, 2017). There- nally, we decompose the uncertainties arising from different
fore, the bivariate framework based on copula functions has sources by employing the multivariate analysis of variance
been developed for drought projection, compensating for the (MANOVA) method. This study illustrates the used materi-
Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

R. Liu et al.: Machine-learning-constrained projection of bivariate hydrological drought magnitudes 3307
als and methods in Sects. 2 and 3, respectively. We compare ature (T2 m ) and dew-point temperature (Td ) are substituted
the SRI and TWS-DSI when assessing drought conditions in into Eq. (1) to calculate RH:
Sect. 4.1. The contributions of meteorological factors to sim-
es (Td ) 1 1 L0
ulating streamflow and calibrating hybrid terrestrial models RH = = exp − . (2)
are shown in Sect. 4.2. The evolution of univariate droughts es (T2 m ) T2 m Td R0
is projected in Sect. 4.3. The bivariate droughts of future sce- Then, the near-surface air pressure (ps) and Td are used to
narios and the associated socioeconomic exposures are eval- deduce the specific humidity (SH), which is mathematically
uated in Sect. 4.4. We discuss the uncertainty of our analysis expressed as follows (Simmons et al., 1999):
and the main limitations of this study in Sect. 5 and finally
summarize our work in Sect. 6. 0.622 × es (Td )
SH = . (3)
ps − 0.378es (Td )
2 Methodology 2.2 Sensitivity analysis of meteorological variables for

runoff
The workflow of this study is divided into four modules
(Fig. 1) described briefly below and detailed in the follow- The RF model (Catani et al., 2013) is used to calculate the
ing sections. In step 1, the hydrological models and LSTM sensitivity of runoff to different meteorological variables, in-
are trained using the ERA5-Land dataset, and then the out- cluding precipitation (pr), air pressure (ps), surface down-
put of hydrological models (HMs) is used as input to feed the welling shortwave and longwave radiation (srsds and srlds),
LSTM. Thus we build hybrid terrestrial models (HTMs). In RH, SH, average temperature, and maximum and minimum
step 2, the trained HTMs are validated using in situ stream- temperature. The contribution of a key variable is derived
flow observations and are then driven by using the outputs by using the pre-established model, the perturbed meteoro-
of five GCMs from CMIP6 to project streamflow and the logical variable, and the remaining (non-perturbed) variables
SRI series. In step 3, monthly drought characteristics (i.e., (Antoniadis et al., 2021; Green et al., 2020). The percentage
drought duration and severity) are defined using run theory change in streamflow is derived from the following equation:
and combined with copula functions to construct a bivari- mean R(i+1SD) − R(all)

ate drought framework. Future bivariate drought change is Si = × 100 %, (4)
evaluated using the most likely realization method. Mean- stdev(Robs )
while, the TWS measurements from GRACE missions are where Si indicates the sensitivity of streamflow to the ith me-
also employed to characterize recent changes in TWS-based teorological variables, which are pr, ps, SH, RH, srlds,
droughts, which are also compared with the hydrological srsds, and temperature. Robs is the observation of streamflow
droughts. In step 4, we employ future scenarios of GDP and (m3 s−1 ). R(i+1SD) is the streamflow simulated by perturb-
population alongside our future drought projections to pro- ing i by +1SD. R(all) is the streamflow simulated by all the
duce a socioeconomic assessment of drought exposure over meteorological variables. stdev(Robs ) represents the standard
China. Finally, we examine the contribution of uncertainty deviation of Robs .
from different sources in projecting drought change and ex-
posure. 2.3 Deep-learning-constrained hydrological modeling
2.1 Derivation of 2 m relative and specific humidity 2.3.1 Conceptual hydrological models
As relative humidity and specific humidity are not directly For preliminary hydrological simulations, we select five hy-
available from the ERA5-Land dataset, we estimate these drological models to represent hydrological characteristics in
two variables based on the physical relationship in the at- different environments. GR4J (Génie Rural à 4 paramètres
mosphere. The Clausius–Clapeyron relationship is used to Journalier) is a lumped model with four parameters devel-
derive saturated vapor pressure (es ) and air temperature (T ) oped by Perrin et al. (2003). It consists of two water storage
and is expressed as follows (Koutsoyiannis, 2012): modules (runoff yielding and routing) and uses daily rainfall
and evapotranspiration as inputs to simulate streamflow se-
1 1 L0
es (T ) = e0 exp − , (1) ries (Kunnath-Poovakka and Eldho, 2019). This model has
T0 T R0 been successfully used to simulate hybrid runoff processes
where T0 , e0 , L0 , and R0 are the freezing temperature in on many continents (Gu et al., 2023; Shin and Kim, 2021).
Kelvin, the saturated vapor pressure at the freezing tem- Additionally, we use a temperature-based method (Oudin
perature, the latent heat of vaporization, and the gas con- et al., 2005) to estimate the potential evapotranspiration of
stant of water vapor (with values of 273.15 K, 611 Pa, the GR4J model.
2.5 × 106 J kg−1 , and 461 J kg−1 K−1 , respectively). The HBV (Hydrologiska Byråns Vattenbalansavdelning)
Since near-surface relative humidity (RH) cannot be di- model was initially developed by the Swedish Meteorologi-
rectly obtained from the ERA5-Land dataset, the 2 m temper- cal and Hydrological Institute for Hydrological Forecasting
https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Figure 1. Schematic flowchart of the method, including machine learning (ML)-constrained hydrological simulations, evaluation of bivariate
hydrological drought characteristics and change, and socioeconomic evaluation of drought exposure under climate change.
(Bergström and Forsman, 1973). This model includes five XAJ model usually shows great performance in simulating
modules and one transform function to quantify hydrological hydrological conditions in China (Hu et al., 2005; Jiang et al.,
variables (i.e., precipitation, snow, soil moisture, runoff, and 2007). However, due to inadequacies in the simulation of arid
baseflow) (Bergström, 1995). It has been employed widely to regions, the results of the XAJ model were not considered the
simulate streamflow, and it has a particularly good capacity best option in northern China.
for simulating snowmelt runoff (Kriauciuniene et al., 2013). We used the SCE (shuffled complex evolution) approach
HMETS (Hydrological Model of École de technologie to maximize the objective function (i.e., the Kling–Gupta
supérieure) contains 21 parameters and two reservoirs (i.e., efficiency) to optimize these models (Duan et al., 1992).
the saturated and vadose zones) and is considered to effi- The most complete 20-year observation period is selected to
ciently complete hydrological simulation on limited scales calibrate five models in each watershed using a daily time
(Martel et al., 2017). The model can simulate six processes step. To calibrate the hydrological models, a cross-validation
in the water cycle, including the accumulation, melt, and re- method developed by Arsenault et al. (2017) is used for cali-
freezing of snow, water infiltration and routing, and evapo- bration, which employs the odd years of the data to calibrate
transpiration (Qi et al., 2020). It has been widely used for the models and the even years of the data to validate them.
streamflow simulation under climate change and has shown As the catchments are located in different climatic regions,
great performance (Chen et al., 2018). the parameters of the models are calibrated for each catch-
The SIMHYD (simple lumped conceptual daily rainfall- ment, which means that the parameters are not universal. Al-
runoff) model is a daily rainfall-runoff model developed by though the uncertainties shown by the hydrological models
Porter and McMahon (1975). There are four types of wa- are ineradicable, the overall uncertainty is acceptable on the
ter fluxes from different sources: impervious areas, infiltra- current scale after optimizing the five hydrological models
tion, interflow, and groundwater storage (Chiew et al., 2002). for each catchment.
Although the model was developed earlier, it has shown
good accuracy in simulating runoff over China (Yu and Zhu, 2.3.2 Hybrid scheme of hydrological model and
2015). machine learning
The XAJ (Xinanjiang) model is a hydrological model,
which can usually achieve better performance in humid and Recurrent neural network (RNN) models have had consid-
semi-humid areas than in arid areas (Ren-Jun, 1992). As erable success in hydrological modeling (Cho et al., 2014;
the model was developed based on the underlying surface Sherstinsky, 2020). However, when considering long input
of the Yangtze River Basin in China, it is composed of a sequences, RNNs struggle to capture the relationships be-
three-layer evapotranspiration module with four parameters tween distant points due to a phenomenon known as “long-
and separates the runoff into four components (i.e., surface term dependencies” (Yu et al., 2019). With the development
water, groundwater, interflow water, and flow routing) (Tian of deep learning, this problem can be successfully avoided
et al., 2013). To date, it has been widely reported that the by using LSTMs.

An LSTM cell includes input, output, and forget gates. The 2.4 Drought indexes and run theory
input gate determines which new information can be stored in
the cell state, and the forget gate identifies which information The TWS-DSI is employed to measure the degree of terres-
will be discarded from the cell state. The output gate controls trial drought severity (Zhao et al., 2017). It is a dimension-
which part of the cell state is selected as the output. The up- less standardized water storage anomaly index that can in-
dated cell state is a combination of the information retailed dicate terrestrial drought conditions when below the mean
and the new information to be added. By using this architec- standard value. The TWS-DSI can be expressed mathemati-
ture, the LSTM can avoid the problem of gradient vanishing cally as follows:
or explosion during backpropagation, especially when a se-
ries is long (Gers et al., 2000). The LSTM can be expressed TWS − DSIx,y = (TWSx,y − TWSy )/σy , (12)
as follows: where TWSx,y is the TWS at year x and month y. TWSy
f g t = σ (Whf hs t−1 + Wxf x t + bf ), (5) and σy represent the means and standard deviation of TWS
at month y.
ig t = σ (Whi hs t−1 + Wxi x t + bi ), (6) The SRI is a measure of the variability of runoff for
ct = tanh(Whec hs t−1 + Wxec x t + bec ),
e (7) a given duration based on the percentage of accumulated
ct = f g t · ct−1 + ig t ·e
ct , (8) runoff (Shukla and Wood, 2008). The hydrological drought
og t = σ (Who hs t−1 + Wxo x t + bo ), (9) classification and ranges indicated by the SRI are shown in
Table S1 in the Supplement. To calculate the SRI, we sim-
hs t = og t tanh(ct ), (10) ulate the retrospective time series of streamflow and fit the
where x t , f g t , ig t , and og t are input variables and forget, sample series to a probability distribution. The SRI is con-
input, and output gates at time t. W· are the weights, where sidered to follow a Pearson type-III distribution (Vicente-
Wi , Wc̃ , Wf , and Wo are the weights of each gate, Wx· are Serrano et al., 2012) and is calculated as follows:
the weights of each gate at time t, and Wh· are the weights 
− r − c0 +c1 r+c2 r 2 0 < F (x) ≤ 0.5,
of each gate at the former time t − 1. The operator is the SRI =
2
1+d1 r+d2 r +d3 r 3
(13)
2
symbol for the dot product of two vectors. ct and hs t are the r − c0 +c1 r+c2 r 0.5 < F (x) ≤ 1,
1+d r+d r 2 +d r 3
1 2 3
cell state of the LSTM and the hidden unit at time t, respec-
tively, and ct−1 and hs t−1 are at the former time t − 1. e ct is q
1
where r = ln[ F (x) 2 ]. F (x) is the cumulative probability
the activation function of the hidden layer. bi , bf , bo , and
bc are bias items, and σ (·) and tanh(·) are the sigmoid func- density of the SRI. c0 , c1 , c2 , d1 , d2 , and d3 are the empiri-
tion and the hyperbolic tangent function, respectively. At the cal constants, taken as 2.516, 0.803, 0.010, 1.433, 0.189, and
initial moment, cell and hidden states are set to zero arrays. 0.001 separately.
The hydrological outputs together with other climate vari- After calculating the two drought indexes, the degree of
ables are used as inputs to feed the LSTM model (i.e., the water deficit can be determined according to the Grades of
LSTM is thus constrained by the HMs). Because changes in Meteorological Drought and the previous classification (Di-
meteorological variables require some time to converge be- kici, 2020). Table S1 presents the drought classification and
fore they are reflected in the runoff, it is essential to calculate thresholds used to identify the drought degrees. The run the-
the lag time caused by the flow convergence for the model. ory is employed to obtain characteristics of drought events
The catchment response lag time d is defined as the time dur- from the time series (Yevjevich, 1967). When the drought in-
ing which precipitation accumulates in the river to generate dex is below a mild drought (i.e., ≤ −0.5 drought index), a
runoff for the gauge downstream, and this is expressed math- drought event is detected (Fig. 2), and then the drought dura-
ematically as follows (Berne et al., 2004; Ganguli and Merz, tion and drought severity are extracted.
2019):
2.5 Socioeconomic exposure assessments based on the
d = 2.51A0.4 0.4
d [h] = 0.11Ad [d], (11) copulas and the most likely realization
where Ad (km2 ) represents the catchment area. Meteorologi- To integrate the assessment of drought change arising from
cal variables from day T − d to day T are employed to drive the duration and severity under climate change, we employed
HTMs. a copula framework by constructing the joint probability dis-
We combine the five hydrological models with LSTM to tribution of two variables. After extracting the drought dura-
construct five HTMs. To compare the performance of these tion (D) and severity (S), we fit their marginal distributions
HTMs, we use 10 HTMs as candidates for streamflow sim- with the seven distributions shown in Table S2. The OR case
ulation in each catchment. The calibrated HTMs are then (i.e., where a bivariate drought event is identified with ei-
driven by the outputs of five GCMs under each SSP (aggre- ther high severity or long duration) of the joint return period
gated to produce a basin average series) during 1985–2100 (JRP) in a copula-based framework is used to quantify the
over 179 catchments to project future daily streamflow. occurrence of drought events (Yin et al., 2020). The joint

Figure 2. Drought duration and severity identification based on the run theory, where −0.5 denotes the drought threshold (grey dashed line).
distribution of drought duration and severity is constructed

by using a copula function, which is valuable for describing
correlated hydrological variables (Li, 1999). Unlike univari-
ate drought frequency analysis, the JRP within a bivariate
framework can be represented by an isoline, which contains
infinite combinations of values of these two multivariate ar-
rays of variables. It is important for risk assessments to se-
lect a representative combination along the isoline. Previous
studies have only selected joint design values according to
the same frequency hypothesis that considering two corre-
lated variables follows the same cumulative probability in
their distributions, but this approach lacks a statistical basis
and poorly describes the physical characteristics of droughts
(Yin et al., 2018). In this paper, the joint probability density
is used to optimize the most likely realization (Fig. 3), which
is expressed mathematically as follows:
Figure 3. Joint distribution of drought duration and severity under
a critical Tor . The green lines are two arbitrary values of duration
and severity. The red line is the isoline of two variables under a crit-

∗ ∗
ical Tor , and the blue line denotes the traditional equal-frequency
(d , s ) = arg maxf (d, s) = c[Fd , Fs ] · fd · fs ,
 assumption. The dT and sT are marginal distribution quantiles for
C[Fd , Fs ] = 1 − µ/Tor , (14) a given probability level T . FS and FD are the cumulative proba-

 dC(Fd ,Fs )
c[Fd , Fs ] = d(F , bility densities of severity and duration, respectively. Tor is a given
d )d(Fs ) probability level in the OR case.
pressed at the catchment scale as follows:

where c[Fd , Fs ] is the copula probability density function;
fd and fs are the fitted probability density functions of D Th I (Th − Tf )
EPOP = × POP, (15)
and S, respectively; Fd and Fs are the marginal distributions Tf A d
of D and S, respectively; (d ∗ , s ∗ ) is the most likely realiza- Th I (Th − Tf )
EGDP = × GDP, (16)
tion under a given JRP Tor ; and µ is the mean interarrival Tf A d
time between two consecutive droughts.
The future socioeconomic exposure after the 2020s has where EPOP and EGDP denote the population and GDP ex-
been defined directly as ranging from 0 % to 100 % (Gu et al., posure; Th and Tf denote the historical and future JRPs, re-
2020a), but dynamically shifting climate risks cannot be rep- spectively; I (·) denotes the controlling function, which is 1
resented under this definition without considering fluctuation when Th − Tf < 0 or 0 when Th − Tf ≥ 0; and POP and GDP
in the frequency of hazards. Here, the socioeconomic expo- denote the population and the gross domestic product (USD),
sure is defined by considering the shift in the JRP and is ex- respectively, of a given catchment in the future climate.

2.6 Quantifying the uncertainty contributed by data (Tapley et al., 2004). Many international institutes have
different sources released the TWS mascon products at a monthly scale, in-
cluding the JPL (Jet Propulsion Laboratory of the California
Uncertainties in the future drought projections can arise from Institute of Technology), the GSFC (Goddard Space Flight
the SSPs, GCMs, and HTMs. During both the historical Center of NASA), and the CSR (Center for Space Research
(1985–2014) and future (2071–2100) periods, the combina- of the University of Texas). As these three mascon solutions
tion of three SSPs, five GCMs, and 10 HTMs through the are produced at different spatial resolutions, we generated
impact modeling chain resulted in 150 hybrid combinations. blended TWS data based on the average of the JPL, GSFC,
The overall uncertainty is calculated from the variance of and CSR with a 0.5° × 0.5° resolution from 2002 to 2022 and
the future estimated JRP relative to the historical 50-year fill in the missing data using a linear interpolation approach
droughts. To partition the uncertainty from different sources (Yin et al., 2022).
of data and their interaction effects, the MANOVA is used
and expressed as follows (Weinfurt, 1995):
3.3 ERA5-Land dataset
1yx,y,z = M + Sx + Gy + Hz + Ix,y,z , (17)
ERA5-Land is a dataset that consists of a large volume of
where M denotes the mean change in all the indicators in the
meteorological variables, including precipitation, tempera-
models; Sx , Gy , and Hz denote the impacts on the indicators
ture, and air pressure. The spatial resolution of the dataset
of the xth SSP, yth GCM, and zth HTM, respectively; and
is 9 km, and the temporal resolution is 1 h (Yilmaz, 2023).
Ii,j,k is the overall impact arising from the interactions of
Under the latest global reanalysis and the lapse rate correc-
different sources. The overall variance V is then expressed
tion, the ERA5-Land reanalysis dataset provides a substitute
as follows:
for unavailable observed weather data by taking the effect of
V = VS + VG + VH + VISG + VISH + VIGH + VISGH , (18) altitude on the spatial scheme of climate variables into ac-
count (Pelosi et al., 2020). Six variables are used in the study
where VS, VG, and VH are the variances from the SSPs, (i.e., pr, ps, T2 m , Td , srlds, and srsds) and are aggregated to
GCMs, and HTMs, respectively. VISG , VISH , VIGH , and the daily scale from the hourly scale before conducting data
VISGH denote the variances caused by the coupling between analysis (Muñoz Sabater, 2019).
the different sources of data. The contribution of each source
to the overall uncertainty is quantified by the variance of each
source and the total variance. 3.4 Bias-corrected GCM outputs and socioeconomic
scenarios
3 Data and materials The climate outputs of five GCMs of the historical sce-
nario and three SSPs (i.e., SSP1-26, SSP3-70, and SSP5-
3.1 In situ observation dataset
85) in CMIP6 are used to represent different climate scenar-
We use a gridded meteorological dataset with 0.5° × 0.5° res- ios. Generally, the SSP5-85 configured the highest carbon
olution, including daily temperature (maximum, minimum, emission and human interference with the natural environ-
and average; °) and daily precipitation (mm) from 1961 to ment. SSP3-70 and SSP1-26 have progressively conservative
2018 provided by the National Meteorological Bureau of changes to represent climate change resulting from different
China. The dataset is regarded as the latest gridded meteoro- levels of human activity. The series of bias-corrected vari-
logical dataset in China and has been applied to some stud- ables have been downscaled to 0.5° × 0.5° resolution from
ies (e.g., Wu et al., 2018; Yin et al., 2021a, b). Meanwhile, 1850 to 2100 in the Intersectoral Impact Model Intercompar-
we gathered the daily streamflow of 463 in situ hydrological ison Project 3b (ISIMIP3b) (Lange, 2019). To reduce the sys-
stations spanning different periods during 1961–2018 (Liu, tematic biases of CMIP6 raw outputs, seven variables from
2023). The hydrological stations are densely distributed in the bias-corrected ISMIP3b dataset have been used, i.e., tem-
eastern China, while western China has a sparser distribution. perature (daily average, maximum, and minimum), pr, ps,
Through rigorous data quality checks, 179 unnested basins srsds, srlds, RH, and SH (Lange and Büchner, 2021).
with at least 20 years of data were selected, covering nine Population and GDP data under three SSPs are employed
major watersheds in China. For more details on streamflow to evaluate the potential socioeconomic risks of drought
data processing and catchment screening, please refer to Yin in a warming world. An open-access population dataset is
et al. (2021b). adopted which takes into account the universal two-child
policy, the census results, and the statistical annual report
3.2 GRACE and GRACE-FO measurements (Jiang et al., 2017). The economic index from 2010 to 2100
is estimated based on the Cobb–Douglas and Population–
Temporal variations in Earth’s gravitational field observed Environment–Development model (Jiang et al., 2018). All
by the GRACE satellites have been used to retrieve TWS of the data have been used previously to assess the socioe-

conomic impact of extreme hydrological hazards (Yin et al., station can be attributed to catchments of different sizes, we
2022). only considered the largest catchment scales in our analysis.
We quantified the sensitivity of seven historical mean meteo-
rological variables (i.e., pr, ps, SH, RH, srlds, srsds, and tem-
4 Results perature) to the monthly streamflow in each grid. Due to the
sparse number of observation stations in northwestern China,
4.1 Observed changes in SRI- and TWS-DSI-based
the reliability of the sensitivity analysis for these regions is
drought
lower than that of the dense observed areas. Precipitation typ-
As there are insufficient streamflow observations to compute ically plays a major role in generating runoff in southeastern
the SRI in northwestern China, we also employed the TWS- China, although SH plays the most important role in some re-
DSI as the Supplement. This approach enriches the variety gions, such as central, southwestern, and northeastern China.
of the water storage or flux being evaluated. Based on linear Over 30 % and 38 % of the stations show a SH sensitivity
regression and the least-squares method, trends in drought rate of > 10 % in western and northeastern China, respec-
characteristics (i.e., frequency, duration, and severity) are es- tively, indicating the dominance of SH in these areas. In con-
timated by using the GRACE and GRACE-FO dataset and trast, RH and shortwave radiation have a negative contribu-
observed runoff across China. Figures 4 and 5 show the tion to streamflow, especially shortwave radiation, which has
drought trends based on the TWS-DSI and the SRI, respec- a pronounced negative sensitivity in 394 stations, probably
tively. Overall, the two indexes show similar trends in most due to enhanced evapotranspiration (Ma et al., 2019). These
catchments, suggesting that drought hazards increased from negative contributions mean that enhancement of these two
2002 to 2022. TWS-DSI droughts increased in 54 % of the ar- variables will inhibit the generation of streamflow, showing
eas, which are mainly located in the Qinghai–Tibet Plateau, the potential adverse effects of climate change on streamflow
the North China Plain, and northwestern Xinjiang Province. generation. In general, RH contributes to increasing stream-
Likewise, SRI droughts increased in over 51 % of the studied flow over most regions of China, but the opposite effect is ob-
catchments, which mainly dominate northeastern and south- served in 179 stations mainly located in southwestern China
eastern China. The severity of the droughts measured by the and in the Yellow River and Huaihe River basins. This is a re-
TWS-DSI is twice that of the hydrological drought, primar- sult of the mutual feedback of water and heat dynamics (i.e.,
ily because the TWS-DSI metric incorporates all the vertical saturated vapor pressure increases with warming and inten-
water fluxes, offering a comprehensive view of the shifts in sifies evaporation, leading to a decrease in surface water),
water scarcity. On the other hand, the TWS-DSI can with which was also found by Liu et al. (2017). The temperature
difficulty represent the aquifer recharge processes, which are has a positive contribution to streamflow generation in north-
fundamental physical processes of baseflow and the hydro- eastern China, suggesting potential mitigation for the defi-
logical drought in its entire extension. Therefore, catchments ciency of the surface flow. However, there is interactive feed-
with aquifer recharge and storage capacity will exceed sev- back between hydrological and thermal factors that results in
eral times the time step of the analysis, enlarging the sever- an inability to directly assess the impact of temperature on
ity of the droughts. Some locations exhibit discrepancies de- hydrological droughts (Fig. 6i and f).
pending on the index considered. For instance, droughts in The performances of simulated streamflow by different
the Qinghai–Tibet Plateau and northeastern China show op- HTMs are shown in Fig. 7. The model that has the largest
posite trends. Anomalies in the Qinghai–Tibetan Plateau may Kling–Gupta efficiency (KGE) is considered to be the best-
be explained by the transformation of snowpack melt into performing one in each catchment. In Fig. 7a and b, the GR4J
surface runoff under the influence of climate change, which and GR4J LSTM performed best in 77 out of the 179 stud-
helps to compensate for the lack of surface water in the area ied catchments. The median KGE value of GR4J is higher
(Stewart, 2009). The discrepancy observed in northeastern than 0.83, revealing a superior performance to the other hy-
China could potentially be linked to the rise in soil moisture drological models. Subsequently, the XAJ and XAJ LSTM
from increased infiltration, which causes a higher proportion are the best models in 57 catchments, mainly located in the
of water to be stored within the soil than at the surface, inter- southern Yangtze River. Finally, the HBV and HBV LSTM
fering with the quantification of hydrological drought (Wang performed best in only 10 catchments, where the streamflow
et al., 2017). Finally, both indicators show a consistent posi- was impacted by snowfall in plateaus and northern frozen ar-
tive drought trend in most areas of China, and particularly in eas. All the catchments exhibit KGE values greater than 0.9
the North China Plain and the Pearl River Basin. during the calibration period in Fig. 7c, showing good per-
formance in simulation. During the validation period, only
4.2 Machine-learning-constrained streamflow 18 catchments have KGE values below 0.6, and most of
simulation and model evaluation the catchments have KGE values greater than 0.8 (Fig. 7d).
In summary, the trained models simulate streamflow well
The RF model was used to quantify the sensitivity of stream- in all the studied catchments. Additionally, the KGE val-
flow to different meteorological variables (Fig. 6). Since a ues in the southern region are generally higher than those

Figure 4. Trends in drought frequency, duration, and severity based on the TWS-DSI from 2002 to 2022 using three GRACE and GRACE-FO
products (a–i) and the blended data (j–l).

Figure 5. Trends in drought frequency, duration, and severity from 2002 to 2022 over China. (c) The index of severity is based on the SRI
statistic (Eq. 13).
Figure 6. Sensitivity of meteorological variables to daily streamflow. The figure uses a thin-plate smoothing spline method to interpolate the
point-based station data (circles). The grey areas indicate missing data.

Figure 7. Hydrological simulation performances of all the candidate models. (a) The best-performing model with the highest KGE value.
The catchments are colored according to the best-performing models. (b) Boxplots of all the catchments for 10 HTMs indicated by KGE
values. (c, d) The highest KGE values during the calibration (c) and validation (d) periods, respectively.
in the northern region during the validation period, which is ensemble average severity and duration for the 50-year his-
consistent with previous hydrological simulation works (Gu torical return period (RP).
et al., 2020b, 2021). This phenomenon may be attributed to In western China, we projected a significantly increas-
the higher dependence of streamflow on rainfall in southern ing drought trend under the three SSPs, which indicates po-
China, which is governed by a humid climate pattern (Zheng tential for increased water scarcity and more frequent ex-
et al., 2022). treme drought events. In southeastern China, we projected
that drought events are likely to intensify under SSP3-70 but
4.3 Projected changes in univariate drought not under SSP5-85. It is generally considered that SSP5-85
characteristics is accompanied by higher carbon emissions than SSP3-70
(O’Neill et al., 2016). However, future work should also take
We projected the future daily runoff series by driving the significant action to control the extent of climate change,
HTMs with the bias-corrected CMIP6 variables, and then we combined with strong climate policies under SSP5-85 (Fu-
estimated the monthly SRI to identify the drought duration jimori et al., 2017). As a result, there is no deterioration of
and severity. Based on the maximum Bayesian information drought severity with policy interventions, which emphasizes
criterion (BIC), we selected the best-performing marginal the significance of ensuring the implementation of climate
distributions for duration and severity from seven candidate strategies. In northern China, in contrast, we found that fu-
distributions shown in Table S2, based on the historical data ture drought risks are projected to decrease in the three sce-
for each catchment. Figures 8 and 9 show the multimodel

Figure 8. Multimodel ensemble average design severity (dimensionless) under a 50-year RP for three SSPs, together with relative changes
(%) in 2071–2100 compared to 1985–2014.
narios, which is possibly related to more moisture conver- vealing substantial spatial heterogeneity across China. Over-
gence from the East Asian monsoon circulation as the cli- all, the severity and duration of droughts slightly increase in
mate warms (Chowdary et al., 2019). some catchments and have a risk of extreme intensification
We display the relative change in drought characteristics as a result of global warming.
with a 50-year RP for all the catchments for five GCMs under
the three SSPs using violin plots (Fig. 10). For most catch- 4.4 Bivariate drought changes and corresponding
ments, the relative change in drought duration and severity socioeconomic risks
is negative. However, the relative change in some scenar-
ios reached a maximum of 400 %, highlighting the extreme To capture the complex dependence structure between
change in drought. The median relative change in severity drought severity and duration, we used a copula function
based on IPSL-CM6A-LR under SSP3-70 is 30 %, and 22 % to quantify the bivariate risk of hydrological droughts un-
of the catchments have a relative change of over 200 %, rep- der climate change. Changes in the JRP of the historical
resenting the most severe case of drought evolution. Further- (1985–2014) drought event with the 50-year JRP in the fu-
more, the distributions of the projections based on the MPI- ture (2071–2100) period are shown in Fig. 11. The medians
ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL models are of the projected future JRP are 38.78, 14.52, and 19.24 years
highly skewed and bimodal under SSP3-70 and SSP5-85, re- under SSP1-26, SSP3-70, and SSP5-85, respectively. For the

Figure 9. Multimodel ensemble average design duration (months) of the multiple models for a 50-year RP for three SSPs, together with
relative changes (%) in 2071–2100 compared to 1985–2014.
69 % and 60 % catchments under SSP3-70 and SSP5-85, we The eastern coastal regions have higher significant economic
find that the JRP of the 50-year drought is reduced to less exposure, such as in the Huaihe River Basin, the Yangtze
than 25 years in the future period, suggesting that the risk River Basin, and the Pearl River Basin, which is consis-
of drought increases by more than 2 times in these catch- tent with the distribution of economically developed re-
ments. In addition, we find a marked increase in the num- gions in China. The medians of GDP exposure are USD 5.5,
ber of catchments with increased drought risk compared to USD 9.8, and USD 14.3 million per square kilometer under
the univariate drought assessments. The JRP of catchments the three SSPs, respectively, which indicates the vulnerability
in northeastern and central China tends to decrease, suggest- of economic losses to drought disasters under global warm-
ing higher changes in risks than with univariate assessments. ing. The population affected by drought is mainly located
This result is consistent with previous studies (He et al., in the southern Yangtze River Basin and the Huaihe River
2011; Xu et al., 2015), which indicates that the use of bi- Basin under SSP3-70, as the median exposure is 525 and
variate drought analysis can amplify the individual effects of 205 people per square kilometer under SSP3-70 and SSP5-
two drought characteristics. 85, respectively. This is because the increase in population
Future GDP and population exposures to increasing bi- is higher in Sichuan, Guangdong, and Zhejiang provinces
variate drought risk in three scenarios are shown in Fig. 12. than in other Chinese provinces under SSP3-70 (Chen et al.,

Figure 10. Violin plots of the relative changes (%) in severity and duration during the historical drought event with a 50-year RP under three
SSPs. The white circles are the median values of the relative changes.
2020). Overall, the exposures of GDP and population show bias-corrected GCM outputs for accurate projections. Fur-
large heterogeneity in their sensitivities to different scenar- ther, the contributions of the SSPs to population exposure are
ios, and the distribution of the affected catchments is consis- 1.5 times those of GDP exposure, which shows that the effect
tent with economic and social development. of climate change is greater for POP exposure than for GDP
exposure. In particular, the independent factors (i.e., SSPs,
GCMs, and HTMs) contribute over 50 % of the uncertainty
5 Discussion of GDP and population exposures, suggesting that GDP and
population exposures are less responsive to complex cou-
5.1 Uncertainty decomposition pling. In contrast, the coupled factors (i.e., the combination
of SSPs, GCMs, and HTMs) mainly contribute to the uncer-
The overall uncertainty in our projections arises from the tainty of the JRP, accounting for 82.63 % of the overall un-
different SSPs, GCMs, and HTMs as well as their inter- certainty, especially SSM–GCM–HTM, which accounts for
actions. We assemble these seven sources using MANOVA 36.97 % of the uncertainty. Finally, the relatively low contri-
(Fig. 13). For GDP and POP exposure, we find HTMs to be bution of the choice of SSP, SSP–GCM, and SSP–HTM to
the main source of uncertainty; they contribute uncertainties the JRP uncertainty indicates that the future risk projection
of 27.55 % and 26.14 %, respectively. This indicates that the uncertainty is relatively stable in future risk projections.
quality of an HTM is important for the accuracy of socioe-
conomic predictions. Likewise, the GCM and GCM–HTM
provide over 30 % of the uncertainty in GDP and popula-
tion exposures, which indicates the critical importance of

Figure 11. The future multimodel ensemble mean JRP of the historical drought with a 50-year Tor based on the bivariate approach. The
future JRPs of 179 catchments under three SSPs are presented in panels (a–c), while panel (d) displays rain cloud plots of the projected JRP
under each SSP.
5.2 Limitations and future work and Forecasting hydrological modeling system, the Soil and
Water Assessment Tool, or the hydrological modules of land
surface process models) to reduce uncertainty in future re-
The uncertainty caused by the underlying surface situation
search. Finally, the GDP and population projections cannot
and the coupling relationships behind interrelated variables
reflect future economic development and population migra-
remains unexplained in this study. Therefore, revealing in-
tion well, especially governmental intervention in immigra-
teractions between multisource data is important for under-
tion and economic policies. It is better to consider the dy-
standing how the drivers affect the water cycle under climate
namic impact of human management on socioeconomic de-
change. Here, only five GCM outputs and one in situ ob-
velopment, which is essential for the construction of a more
servation dataset were used to drive our HTMs. The sparse
reliable projection framework.
dataset may undermine the robustness of the approach. Pro-
viding a larger number of GCMs and observational data to
assemble a more sophisticated model might be an effective 5.3 Suggestions for drought mitigation in China
approach for improving accuracy and reliability. Although
the catchments gathered in this study cover nine major wa- In order to curb global warming and mitigate the threats of
tersheds in China, there is still a requirement for stream- climate change, the Chinese government is striving to reach
flow data with a more uniform spatial density. Considering its carbon peak before 2030, achieve carbon neutrality before
geospatial sampling techniques, a homogeneous density of 2060, and bolster efforts in disaster reduction (Kundzewicz
catchments is significant for revealing the spatial distribu- et al., 2019; Z. Liu et al., 2022). China has nonetheless ex-
tion of drought. On the other hand, due to the heterogene- perienced several extreme drought events during the past
ity of the different climatic regions in China, we would like 5 years, threatening the population’s health and economic de-
to expand hydrological models (e.g., the Weather Research velopment (Ding and Gao, 2020; Y. Liu et al., 2022; Mallap-

Figure 12. The multimodel ensemble mean exposure of GDP (a–c) and population (d–f) to bivariate drought characteristics under different
SSPs in the future period.
aty, 2022). The Intergovernmental Panel on Climate Change by severe drought events that are expected to significantly in-
(IPCC) has emphasized that projections of future climate tensify with climate change. We strongly highlight the impor-
trends can equip policymakers with the scientific insight tance of strictly implementing carbon emission reduction ini-
needed to navigate the challenges of climate change (Pörtner tiatives and developing prevention programs to limit poten-
et al., 2022). The results of this study aim to alert policymak- tial drought losses. Preserving local ecological balance and
ers to drought risk in southwestern China, which was just hit employing rational use of water resources could be the key

Figure 13. The fractional uncertainty contributions of all the sources to the GDP exposure, population exposure, and JRP estimate for all
179 catchments (a, c, e) and the average fractional contribution of each source (b, d, f).
to mitigating potential losses from extreme droughts (Chang extreme droughts (Chang et al., 2019; Sohn et al., 2016). Fi-
et al., 2019; Sohn et al., 2016). Although China has con- nally, this work highlights the importance of strictly imple-
structed hydraulic structures with a total water storage capac- menting carbon emission reduction initiatives and develop-
ity of over 7064 × 109 m3 , current irrigation facilities need ing prevention programs to limit potential drought losses.
to expand to mitigate the challenge of drought under climate
change (Cai et al., 2015; Xiao-jun et al., 2012). In addition, it
is beneficial for policymakers to establish a drought informa- 6 Conclusions
tion system to get a comprehensive collection of drought im-
pacts from all potential sectors, which can link government In this study, hybrid LSTM-constrained hydrological mod-
and research organizations (Wilhite et al., 2007). els show high efficiency in studied catchments over China,
The IPCC has emphasized that projections of future cli- demonstrating that machine learning can effectively con-
mate trends can equip policymakers with the scientific in- strain hydrological simulation. Projected changes in 50-year
sight needed to navigate the challenges of climate change bivariate drought characteristics, expressed as a JRP, indicate
(Pörtner et al., 2022). The results of this study aim to alert that the risk of hydrological drought is likely to more than
policymakers to drought risk in southwestern China, which double in over 60 % of catchments by the end of the 21st cen-
is expected to intensify with climate change. Preserving lo- tury under SSP5-85. The spatial distribution of change re-
cal ecological balance and employing rational use of water veals that the catchments with severely increased drought
resources could be the key to mitigating potential losses from risk are mainly located in southwestern China. Notably, the

exposure of GDP and population varies greatly across dif- Financial support. This research has been supported by the Na-
ferent SSPs. The median GDP exposure under SSP5-85 is tional Natural Science Foundation of China (grant nos. 52009091,
1.5 times that of SSP3-70, but the median population expo- 52242904, 52261145744, and 52179018) and UK Research and In-
sure is just 40 % that of SSP3-70. The higher population ex- novation (grant no. MR/V022008/1).
posure under SSP3-70 can be attributed to rapid population
growth. Finally, we find that the interaction between multiple
sources of data explains more than 80 % of the uncertainty in Review statement. This paper was edited by Lelys Bravo de Guenni
and reviewed by Federico Gómez-Delgado and three anonymous
future changes in JRPs, showing the importance of consider-
referees.
ing the relationships between model components. Our find-
ings demonstrate that China will face higher drought risks in
a warmer future, emphasizing the urgency of implementing
strategies to reduce carbon emissions. Our study is insuffi- References
cient in the revelation of drought hazard drivers and needs
to expand datasets and hydrological models to promote the Allan, R. P., Barlow, M., Byrne, M. P., Cherchi, A., Douville,
reliability of simulations in future studies. We would also H., Fowler, H. J., Gan, T. Y., Pendergrass, A. G., Rosen-
feld, D., Swann, A. L. S., Wilcox, L. J., and Zolina, O.: Ad-
like to take governmental interference in economic and de-
vances in understanding large-scale responses of the water cy-
mographic policies into consideration. cle to climate change, Ann. NY Acad. Sci., 1472, 49–75,
https://doi.org/10.1111/nyas.14337, 2020.
Antoniadis, A., Lambert-Lacroix, S., and Poggi, J.-M.:
Data availability. The ISIMIP3b data can be downloaded Random forests for global sensitivity analysis: A se-
from https://doi.org/10.48364/ISIMIP.842396.1 (Lange and lective review, Reliab. Eng. Syst. Safe., 206, 107312,
Büchner, 2021). The ERA5-Land data can be downloaded https://doi.org/10.1016/j.ress.2020.107312, 2021.
from https://doi.org/10.24381/cds.e2161bac (Muñoz Sabater, Arsenault, R., Essou, G. R., and Brissette, F. P.: Improving hydro-
2019). The gridded meteorological data for China and the logical model simulations with combined multi-input and mul-
streamflow simulations used in this study are available at timodel averaging frameworks, J. Hydrol. Eng., 22, 04016066,
https://doi.org/10.17605/OSF.IO/FVYSE (Liu, 2023). https://doi.org/10.1061/(ASCE)HE.1943-5584.0001489, 2017.
Ashrafi, S. M., Gholami, H., and Najafi, M. R.: Uncertainties
in runoff projection and hydrological drought assessment over
Supplement. The supplement related to this article is available on- Gharesu basin under CMIP5 RCP scenarios, J. Water Clim.
line at: https://doi.org/10.5194/hess-28-3305-2024-supplement. Change, 11, 145–163, 2020.
Ayantobo, O. O., Li, Y., Song, S., and Yao, N.: Spatial comparability
of drought characteristics and related return periods in mainland
Author contributions. JY and RL conceived this study. RL, SK China over 1961–2013, J. Hydrol., 550, 549–567, 2017.
and YY performed the modeling and data processing. PL, JG, XG Barker, L. J., Hannaford, J., Chiverton, A., and Svensson, C.:
and XZ helped with the data analysis. RL, LS and AV drafted the From meteorological to hydrological drought using standard-
manuscript. All the authors reviewed the paper and approved the ised indicators, Hydrol. Earth Syst. Sci., 20, 2483–2505,
submission. https://doi.org/10.5194/hess-20-2483-2016, 2016.
Bergström, S.: The HBV model, in: Computer models of watershed
hydrology, 443–476, ISBN 978-0-918334-91-6, 1995.
Competing interests. At least one of the (co-)authors is a member Bergström, S. and Forsman, A.: DEVELOPMENT OF A CON-
of the editorial board of Hydrology and Earth System Sciences. The CEPTUAL DETERMINISTIC RAINFALL-RUNOFF MODEL,
peer-review process was guided by an independent editor, and the Hydrol. Res., 4, 147–170, https://doi.org/10.2166/nh.1973.0012,
authors also have no other competing interests to declare. 1973.
Berne, A., Delrieu, G., Creutin, J.-D., and Obled, C.: Temporal and
spatial resolution of rainfall measurements required for urban hy-
drology, J. Hydrol., 299, 166–179, 2004.
Disclaimer. Publisher’s note: Copernicus Publications remains
Byakatonda, J., Parida, B. P., Moalafhi, D. B., and Kenabatho, P.
neutral with regard to jurisdictional claims made in the text, pub-
K.: Analysis of long term drought severity characteristics and
lished maps, institutional affiliations, or any other geographical rep-
trends across semiarid Botswana using two drought indices, At-
resentation in this paper. While Copernicus Publications makes ev-
mos. Res., 213, 492–508, 2018.
ery effort to include appropriate place names, the final responsibility
Cai, X., Zeng, R., Kang, W. H., Song, J., and Valoc-
lies with the authors. Regarding the maps used in this paper, please
chi, A. J.: Strategic Planning for Drought Mitigation un-
note that Figs. 4–9 and 11–12 contain disputed territories.
der Climate Change, J. Water Res. Pl., 141, 04015004,
https://doi.org/10.1061/(ASCE)WR.1943-5452.0000510, 2015.
Castle, S. L., Thomas, B. F., Reager, J. T., Rodell, M., Swenson, S.
Acknowledgements. The numerical calculations in this paper were C., and Famiglietti, J. S.: Groundwater depletion during drought
performed on the supercomputing system in the Supercomputing threatens future water security of the Colorado River Basin, Geo-
Center of Wuhan University. phys. Res. Lett., 41, 5904–5911, 2014.

Catani, F., Lagomarsino, D., Segoni, S., and Tofani, V.: Landslide Duan, Q., Sorooshian, S., and Gupta, V.: Effective and efficient
susceptibility estimation by random forests technique: sensitivity global optimization for conceptual rainfall-runoff models, Water
and scaling issues, Nat. Hazards Earth Syst. Sci., 13, 2815–2831, Resour. Res., 28, 1015–1031, 1992.
https://doi.org/10.5194/nhess-13-2815-2013, 2013. Fujimori, S., Hasegawa, T., Masui, T., Takahashi, K., Herran, D. S.,
Chang, J., Guo, A., Wang, Y., Ha, Y., Zhang, R., Xue, L., and Tu, Z.: Dai, H., Hijioka, Y., and Kainuma, M.: SSP3: AIM implementa-
Reservoir operations to mitigate drought effects with a hedging tion of shared socioeconomic pathways, Global Environ. Chang.,
policy triggered by the drought prevention limiting water level, 42, 268–283, 2017.
Water Resour. Res., 55, 904–922, 2019. Ganguli, P. and Merz, B.: Trends in compound flooding in north-
Chen, H. and Sun, J.: Increased population exposure to ex- western Europe during 1901–2014, Geophys. Res. Lett., 46,
treme droughts in China due to 0.5 °C of additional warming, 10810–10820, 2019.
Environ. Res. Lett., 14, 064011, https://doi.org/10.1088/1748- Gers, F. A., Schmidhuber, J., and Cummins, F.: Learning to forget:
9326/ab072e, 2019. continual prediction with LSTM, Neural Comput., 12, 2451–71,
Chen, J., Li, C., Brissette, F. P., Chen, H., Wang, M., and Essou, G. 2000.
R.: Impacts of correcting the inter-variable correlation of climate Green, J. K., Berry, J., Ciais, P., Zhang, Y., and Gentine,
model outputs on hydrological modeling, J. Hydrol., 560, 326– P.: Amazon rainforest photosynthesis increases in response
341, 2018. to atmospheric dryness, Science Advances, 6, eabb7232,
Chen, Y., Guo, F., Wang, J., Cai, W., Wang, C., and Wang, K.: https://doi.org/10.1126/sciadv.abb7232, 2020.
Provincial and gridded population projection for China under Gu, L., Chen, J., Yin, J., Sullivan, S. C., Wang, H.-M., Guo,
shared socioeconomic pathways from 2010 to 2100, Scientific S., Zhang, L., and Kim, J.-S.: Projected increases in magni-
Data, 7, 83, https://doi.org/10.1038/s41597-020-0421-y, 2020. tude and socioeconomic exposure of global droughts in 1.5 and
Chen, Z. and Yang, G.: Analysis of drought hazards in North 2 °C warmer climates, Hydrol. Earth Syst. Sci., 24, 451–472,
China: distribution and interpretation, Nat. Hazards, 65, 279– https://doi.org/10.5194/hess-24-451-2020, 2020a.
294, https://doi.org/10.1007/s11069-012-0358-3, 2013. Gu, L., Chen, J., Yin, J., Xu, C.-Y., and Zhou, J.: Responses
Chiew, F. H. S., Peel, M. C., and Western, A. W.: Application and of precipitation and runoff to climate warming and impli-
testing of the simple rainfall-runoff model SIMHYD, in: Math- cations for future drought changes in China, Earths Future,
ematical models of small watershed hydrology and applications, 8, e2020EF001718, https://doi.org/10.1029/2020EF001718,
335–367, ISBN 1-887201-35-1, 2002. 2020b.
Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Gu, L., Yin, J., Zhang, H., Wang, H.-M., Yang, G., and Wu, X.: On
Bougares, F., Schwenk, H., and Bengio, Y.: Learning phrase rep- future flood magnitudes and estimation uncertainty across 151
resentations using RNN encoder-decoder for statistical machine catchments in mainland China, Int. J. Climatol., 41, E779–E800,
translation, arXiv [preprint], arXiv:1406.1078, 2014. 2021.
Chowdary, J. S., Hu, K., Srinivas, G., Kosaka, Y., Wang, L., and Gu, L., Yin, J., Wang, S., Chen, J., Qin, H., Yan, X., He, S., and
Rao, K. K.: The Eurasian jet streams as conduits for East Asian Zhao, T.: How well do the multi-satellite and atmospheric reanal-
monsoon variability, Curr. Clim. Change Rep., 5, 233–244, 2019. ysis products perform in hydrological modelling, J. Hydrol., 617,
Dai, A., Zhao, T., and Chen, J.: Climate Change and Drought: a Pre- 128920, https://doi.org/10.1016/j.jhydrol.2022.128920, 2023.
cipitation and Evaporation Perspective, Curr. Clim. Change Rep., He, B., Lü, A., Wu, J., Zhao, L., and Liu, M.: Drought hazard as-
4, 301–312, https://doi.org/10.1007/s40641-018-0101-6, 2018. sessment and spatial characteristics analysis in China, J. Geogr.
Dikici, M.: Drought analysis with different indices for the Asi Basin Sci., 21, 235–249, 2011.
(Turkey), Sci. Rep., 10, 20739, https://doi.org/10.1038/s41598- Hu, C., Guo, S., Xiong, L., and Peng, D.: A modified Xinanjiang
020-77827-z, 2020. model and its application in northern China, Hydrol. Res., 36,
Dikshit, A., Pradhan, B., and Huete, A.: An improved SPEI 175–192, 2005.
drought forecasting approach using the long short-term mem- Jiang, T., Chen, Y. D., Xu, C., Chen, X., Chen, X., and
ory neural network, J. Environ. Manage., 283, 111979, Singh, V. P.: Comparison of hydrological impacts of cli-
https://doi.org/10.1016/j.jenvman.2021.111979, 2021a. mate change simulated by six hydrological models in the
Dikshit, A., Pradhan, B., and Alamri, A. M.: Long lead time drought Dongjiang Basin, South China, J. Hydrol., 336, 316–333,
forecasting using lagged climate variables and a stacked long https://doi.org/10.1016/j.jhydrol.2007.01.010, 2007.
short-term memory model, Sci. Total Environ., 755, 142638, Jiang, T., Jing, Z., Cheng, J., Lige, C., Yanjun, W., Hemin, S.,
https://doi.org/10.1016/j.scitotenv.2020.142638, 2021b. Anqian, W., Jinlong, H., Buda, S., and Run, W.: National and
Ding, T. and Gao, H.: The record-breaking extreme drought in Yun- provincial population projected to 2100 under the shared so-
nan Province, Southwest China during spring-early summer of cioeconomic pathways in China, Advances in Climate Change
2019 and possible causes, J. Meteorol. Res.-PRC, 34, 997–1012, Research, 13, 128, 28–137 https://doi.org/10.12006/j.issn.1673-
2020. 1719.2016.249, 2017.
Dixit, S., Atla, B. M., and Jayakumar, K. V.: Evolution and Jiang, T., Jing, Z., Li-Ge, C. A. O., Yan-Jun, W., Bu-Da, S.
drought hazard mapping of future meteorological and hydrolog- U., Cheng, J., Run, W., and Chao, G. A. O.: Projection of
ical droughts using CMIP6 model, Stoch. Env. Res. Risk A., 36, national and provincial economy under the shared socioeco-
3857–3874, 2022. nomic pathways in China, Advances in Climate Change Re-
Donat, M. G., Lowry, A. L., Alexander, L. V., O’Gorman, P. A., and search, 14, 50–58, http://www.climatechange.cn/EN/10.12006/j.
Maher, N.: More extreme precipitation in the world’s dry and wet issn.1673-1719.2017.161 (last access: 20 July 2024), 2018.
regions, Nat. Climate Change, 6, 508–513, 2016.

Kang, S., Yin, J., Gu, L., Yang, Y., Liu, D., and Slater, L.: poral Analyses, J. Geophys. Res.-Atmos., 124, 4326–4351,
Observation-constrained projection of flood risks and socioeco- https://doi.org/10.1029/2018JD029850, 2019.
nomic exposure in China, Earth’s Future, 11, e2022EF003308, Mallapaty, S.: China’s extreme weather challenges scientists study-
https://doi.org/10.1029/2022EF003308, 2023. ing it, Nature, 609, 888, https://doi.org/10.1038/d41586-022-
Kim, J. H., Sung, J. H., Chung, E.-S., Kim, S. U., Son, M., 02954-8, 2022.
and Shiru, M. S.: Comparison of Projection in Meteorolog- Martel, J., Demeester, K., Brissette, F., Poulin, A., and Arsenault,
ical and Hydrological Droughts in the Cheongmicheon Wa- R.: HMETS-A simple and efficient hydrology model for teach-
tershed for RCP4.5 and SSP2-4.5, Sustainability, 13, 2066, ing hydrological modelling, flow forecasting and climate change
https://doi.org/10.3390/su13042066, 2021. impacts, Int. J. Eng. Educ., 33, 1307–1316, 2017.
Koutsoyiannis, D.: Clausius–Clapeyron equation and saturation Meinshausen, M., Nicholls, Z. R. J., Lewis, J., Gidden, M. J.,
vapour pressure: simple theory reconciled with practice, Eur. Vogel, E., Freund, M., Beyerle, U., Gessner, C., Nauels, A.,
J. Phys., 33, 295, https://doi.org/10.1088/0143-0807/33/2/295, Bauer, N., Canadell, J. G., Daniel, J. S., John, A., Krummel,
2012. P. B., Luderer, G., Meinshausen, N., Montzka, S. A., Rayner,
Kriauciuniene, J., Jakimavicius, D., Sarauskiene, D., and Kaliatka, P. J., Reimann, S., Smith, S. J., van den Berg, M., Velders, G.
T.: Estimation of uncertainty sources in the projections of Lithua- J. M., Vollmer, M. K., and Wang, R. H. J.: The shared socio-
nian river runoff, Stoch. Env. Res. Risk A., 27, 769–784, 2013. economic pathway (SSP) greenhouse gas concentrations and
Kumar, R., Musuuza, J. L., Van Loon, A. F., Teuling, A. J., Barthel, their extensions to 2500, Geosci. Model Dev., 13, 3571–3605,
R., Ten Broek, J., Mai, J., Samaniego, L., and Attinger, S.: https://doi.org/10.5194/gmd-13-3571-2020, 2020.
Multiscale evaluation of the Standardized Precipitation Index as Mokhtar, A., Jalali, M., He, H., Al-Ansari, N., Elbeltagi, A., Al-
a groundwater drought indicator, Hydrol. Earth Syst. Sci., 20, safadi, K., Abdo, H. G., Sammen, S. S., Gyasi-Agyei, Y., and
1117–1131, https://doi.org/10.5194/hess-20-1117-2016, 2016. Rodrigo-Comino, J.: Estimation of SPEI meteorological drought
Kundzewicz, Z., Su, B., Wang, Y., Xia, J., Huang, J., and Jiang, T.: using machine learning algorithms, IEEE Access, 9, 65503–
Flood risk and its reduction in China, Adv. Water Resour., 130, 65523, 2021.
37–45, https://doi.org/10.1016/j.advwatres.2019.05.020, 2019. Muñoz Sabater, J.: ERA5-Land hourly data from 1950 to present,
Kunnath-Poovakka, A. and Eldho, T. I.: A comparative study of Copernicus Climate Change Service (C3S) Climate Data Store
conceptual rainfall-runoff models GR4J, AWBM and Sacra- (CDS) [data set], https://doi.org/10.24381/cds.e2161bac, 2019.
mento at catchments in the upper Godavari river basin, India, Myronidis, D., Ioannou, K., Fotakis, D., and Dörflinger, G.: Stream-
J. Earth Syst. Sci., 128, 33, https://doi.org/10.1007/s12040-018- flow and hydrological drought trend analysis and forecasting in
1055-8, 2019. Cyprus, Water Resour. Manag., 32, 1759–1776, 2018.
Lange, S.: Trend-preserving bias adjustment and statistical down- Nabaei, S., Sharafati, A., Yaseen, Z. M., and Shahid, S.: Copula
scaling with ISIMIP3BASD (v1.0), Geosci. Model Dev., 12, based assessment of meteorological drought characteristics: re-
3055–3070, https://doi.org/10.5194/gmd-12-3055-2019, 2019. gional investigation of Iran, Agr. Forest Meteorol., 276, 107611,
Lange, S. and Büchner, M.: ISIMIP3b bias-adjusted atmo- https://doi.org/10.1016/j.agrformet.2019.06.010, 2019.
spheric climate input data (v1.1), ISIMIP Repository [data set], Nie, N., Zhang, W., Chen, H., and Guo, H.: A global hydrological
https://doi.org/10.48364/ISIMIP.842396.1, 2021. drought index dataset based on gravity recovery and climate ex-
Li, D. X.: On default correlation: A copula function approach, periment (GRACE) data, Water Resour. Manag., 32, 1275–1290,
SSRN Electr., 9, 187289, https://doi.org/10.2139/ssrn.187289, 2018.
1999. O’Neill, B. C., Tebaldi, C., van Vuuren, D. P., Eyring, V., Friedling-
Liu, J., Zhang, Q., Singh, V. P., and Shi, P.: Contribu- stein, P., Hurtt, G., Knutti, R., Kriegler, E., Lamarque, J.-F.,
tion of multiple climatic variables and human activities to Lowe, J., Meehl, G. A., Moss, R., Riahi, K., and Sander-
streamflow changes across China, J. Hydrol., 545, 145–162, son, B. M.: The Scenario Model Intercomparison Project (Sce-
https://doi.org/10.1016/j.jhydrol.2016.12.016, 2017. narioMIP) for CMIP6, Geosci. Model Dev., 9, 3461–3482,
Liu, R.: The streamflow data simulated by 10 hybrid terrestrial https://doi.org/10.5194/gmd-9-3461-2016, 2016.
models under CMIP6, and TWSA trends data, OSF [data set], Oudin, L., Hervieu, F., Michel, C., Perrin, C., Andréassian, V., Anc-
https://doi.org/10.17605/OSF.IO/FVYSE, 2023. til, F., and Loumagne, C.: Which potential evapotranspiration in-
Liu, Y., Hu, Z.-Z., Wu, R., and Yuan, X.: Causes and predictabil- put for a lumped rainfall–runoff model?: Part 2—Towards a sim-
ity of the 2021 spring southwestern China severe drought, Adv. ple and efficient potential evapotranspiration model for rainfall–
Atmos. Sci., 39, 1766–1776, 2022. runoff modelling, J. Hydrol., 303, 290–306, 2005.
Liu, Z., Deng, Z., He, G., Wang, H., Zhang, X., Lin, J., Qi, Pelosi, A., Terribile, F., D’Urso, G., and Chirico, G. B.: Compar-
Y., and Liang, X.: Challenges and opportunities for carbon ison of ERA5-Land and UERRA MESCAN-SURFEX reanaly-
neutrality in China, Nat. Rev. Earth Environ., 3, 141–155, sis data with spatially interpolated weather observations for the
https://doi.org/10.1038/s43017-021-00244-x, 2022. regional assessment of reference evapotranspiration, Water, 12,
Lu, R., Xu, K., Chen, R., Chen, W., Li, F., and Lv, C.: Heat waves 1669, https://doi.org/10.3390/w12061669, 2020.
in summer 2022 and increasing concern regarding heat waves in Perrin, C., Michel, C., and Andréassian, V.: Improvement of a parsi-
general, Atmospheric and Oceanic Science Letters, 16, 100290, monious model for streamflow simulation, J. Hydrol., 279, 275–
https://doi.org/10.1016/j.aosl.2022.100290, 2023. 289, 2003.
Ma, N., Szilagyi, J., Zhang, Y., and Liu, W.: Complementary- Piao, S., Ciais, P., Huang, Y., Shen, Z., Peng, S., Li, J., Zhou, L.,
Relationship-Based Modeling of Terrestrial Evapotranspiration Liu, H., Ma, Y., Ding, Y., Friedlingstein, P., Liu, C., Tan, K., Yu,
Across China During 1982–2012: Validations and Spatiotem- Y., Zhang, T., and Fang, J.: The impacts of climate change on

water resources and agriculture in China, Nature, 467, 43–51, Tian, Y., Xu, Y.-P., and Zhang, X.-J.: Assessment of Climate
https://doi.org/10.1038/nature09364, 2010. Change Impacts on River High Flows through Comparative Use
Pokhrel, Y., Felfelani, F., Satoh, Y., Boulange, J., Burek, P., Gädeke, of GR4J, HBV and Xinanjiang Models, Water Resour. Manag.,
A., Gerten, D., Gosling, S. N., Grillakis, M., and Gudmundsson, 27, 2871–2888, 2013.
L.: Global terrestrial water storage and drought severity under Tirivarombo, S., Osupile, D., and Eliasson, P.: Drought monitoring
climate change, Nat. Climate Change, 11, 226–233, 2021. and analysis: standardised precipitation evapotranspiration index
Porter, J. W. and McMahon, T. A.: Application of a catchment (SPEI) and standardised precipitation index (SPI), Phys. Chem.
model in southeastern Australia, J. Hydrol., 24, 121–134, 1975. Earth Pt. A/B/C, 106, 1–10, 2018.
Pörtner, H.-O., Roberts, D. C., Poloczanska, E. S., Mintenbeck, K., Udall, B. and Overpeck, J.: The twenty-first century Colorado River
Tignor, M., Alegría, A., Craig, M., Langsdorf, S., Löschke, S., hot drought and implications for the future, Water Resour. Res.,
and Möller, V.: IPCC: Summary for policymakers, Cambridge 53, 2404–2418, 2017.
University Press, Cambridge, UK and New York, NY, USA, Vicente-Serrano, S. M., López-Moreno, J. I., Beguería, S., Lorenzo-
https://doi.org/10.1017/9781009157926.001, 2022. Lacruz, J., Azorin-Molina, C., and Morán-Tejeda, E.: Accurate
Qi, W., Chen, J., Li, L., Xu, C., Li, J., Xiang, Y., and Zhang, S.: computation of a streamflow drought index, J. Hydrol. Eng., 17,
A framework to regionalize conceptual model parameters for 318–332, 2012.
global hydrological modeling, Hydrol. Earth Syst. Sci. Discuss. Wang, Z., Li, J., Lai, C., Zeng, Z., Zhong, R., Chen, X., Zhou, X.,
[preprint], https://doi.org/10.5194/hess-2020-127, 2020. and Wang, M.: Does drought in China show a significant decreas-
Rahmati, O., Falah, F., Dayal, K. S., Deo, R. C., Moham- ing trend from 1961 to 2009?, Sci. Total Environ., 579, 314–324,
madi, F., Biggs, T., Moghaddam, D. D., Naghibi, S. A., https://doi.org/10.1016/j.scitotenv.2016.11.098, 2017.
and Bui, D. T.: Machine learning approaches for spatial Weinfurt, K. P.: Multivariate analysis of variance, in: Reading and
modeling of agricultural droughts in the south-east region understanding multivariate statistics, American Psychological
of Queensland Australia, Sci. Total Environ., 699, 134230, Association, Washington, DC, US, 245–276, ISBN: 1-55798-
https://doi.org/10.1016/j.scitotenv.2019.134230, 2020. 273-2, 1995.
Ren-Jun, Z.: The Xinanjiang model applied in China, J. Hydrol., Wilhite, D. A., Svoboda, M. D., and Hayes, M. J.: Understanding
135, 371–381, 1992. the complex impacts of drought: A key to enhancing drought mit-
Schmidt, R., Schwintzer, P., Flechtner, F., Reigber, C., Güntner, igation and preparedness, Water Resour. Manage., 21, 763–774,
A., Döll, P., Ramillien, G., Cazenave, A., Petrovic, S., and https://doi.org/10.1007/s11269-006-9076-5, 2007.
Jochmann, H.: GRACE observations of changes in continental Woolway, R. I., Kraemer, B. M., Lenters, J. D., Merchant, C.
water storage, Global Planet. Change, 50, 112–126, 2006. J., O’Reilly, C. M., and Sharma, S.: Global lake responses
Sherstinsky, A.: Fundamentals of recurrent neural network (RNN) to climate change, Nat. Rev. Earth Environ., 1, 388–403,
and long short-term memory (LSTM) network, Physica D, 404, https://doi.org/10.1038/s43017-020-0067-5, 2020.
132306, https://doi.org/10.1016/j.physd.2019.132306, 2020. Wu, J., Chen, X., Yao, H., and Zhang, D.: Multi-timescale
Shin, M.-J. and Kim, C.-S.: Component combination test to inves- assessment of propagation thresholds from meteorological
tigate improvement of the IHACRES and GR4J rainfall–runoff to hydrological drought, Sci. Total Environ., 765, 144232,
models, Water, 13, 2126, https://doi.org/10.3390/w13152126, https://doi.org/10.1016/j.scitotenv.2020.144232, 2021.
2021. Wu, X., Guo, S., Yin, J., Yang, G., Zhong, Y., and Liu, D.: On the
Shukla, S. and Wood, A. W.: Use of a standardized runoff index event-based extreme precipitation across China: Time distribu-
for characterizing hydrologic drought, Geophys. Res. Lett., 35, tion patterns, trends, and return levels, J. Hydrol., 562, 305–317,
L02405, https://doi.org/10.1029/2007GL032487, 2008. 2018.
Simmons, A. J., Untch, A., Jakob, C., Kållberg, P., and Undén, P.: Xiao-jun, W., Jian-yun, Z., Shahid, S., ElMahdi, A., Rui-min, H.,
Stratospheric water vapour and tropical tropopause temperatures Zhen-xin, B., and Ali, M.: Water resources management strategy
in Ecmwf analyses and multi-year simulations, Q. J. Roy. Me- for adaptation to droughts in China, Mitig. Adapt. Strat. Gl., 17,
teor. Soc., 125, 353–386, 1999. 923–937, https://doi.org/10.1007/s11027-011-9352-4, 2012.
Sohn, J. A., Saha, S., and Bauhus, J.: Potential of forest thinning to Xiujia, C., Guanghua, Y., Jian, G., Ningning, M., and Zihao, W.:
mitigate drought stress: A meta-analysis, Forest Ecol. Manag., Application of WNN-PSO model in drought prediction at crop
380, 261–273, https://doi.org/10.1016/j.foreco.2016.07.046, growth stages: A case study of spring maize in semi-arid re-
2016. gions of northern China, Comput. Electron. Agr., 199, 107155,
Sönmez, A. Y. and Kale, S.: Climate change effects on annual https://doi.org/10.1016/j.compag.2022.107155, 2022.
streamflow of Filyos River (Turkey), J. Water Clim. Change, 11, Xu, K., Yang, D., Yang, H., Li, Z., Qin, Y., and Shen, Y.: Spatio-
420–433, https://doi.org/10.2166/wcc.2018.060, 2018. temporal variation of drought in China during 1961–2012: A cli-
Stewart, I. T.: Changes in snowpack and snowmelt runoff matic perspective, J. Hydrol., 526, 253–264, 2015.
for key mountain regions, Hydrol. Process., 23, 78–94, Yao, F., Livneh, B., Rajagopalan, B., Wang, J., Crétaux, J.-F.,
https://doi.org/10.1002/hyp.7128, 2009. Wada, Y., and Berge-Nguyen, M.: Satellites reveal widespread
Tabari, H.: Climate change impact on flood and extreme precipita- decline in global lake water storage, Science, 380, 743–749,
tion increases with water availability, Sci. Rep., 10, 1–10, 2020. https://doi.org/10.1126/science.abo2812, 2023.
Tapley, B. D., Bettadpur, S., Ries, J. C., Thompson, P. F., and Yevjevich, V. M.: An objective approach to definitions and in-
Watkins, M. M.: GRACE measurements of mass variability in vestigations of continental hydrologic droughts, PhD thesis, Li-
the Earth system, Science, 305, 503–505, 2004. braries, Colorado State University, https://doi.org/10.1016/0022-
1694(69)90110-3, 1967.

Yihdego, Y., Vaheddoost, B., and Al-Weshah, R. A.: Drought Yu, B. and Zhu, Z.: A comparative assessment of AWBM and
indices and indicators revisited, Arab. J. Geosci., 12, 69, SimHyd for forested watersheds, Hydrolog. Sci. J., 60, 1200–
https://doi.org/10.1007/s12517-019-4237-z, 2019. 1212, 2015.
Yilmaz, M.: Accuracy assessment of temperature trends from Yu, Y., Si, X., Hu, C., and Zhang, J.: A Review of Recurrent Neu-
ERA5 and ERA5-Land, Sci. Total Environ., 856, 159182, ral Networks: LSTM Cells and Network Architectures, Neural
https://doi.org/10.1016/j.scitotenv.2022.159182, 2023. Comput., 31, 1235–1270, 2019.
Yin, J., Guo, S., He, S., Guo, J., Hong, X., and Liu, Z.: A copula- Zhai, P. M. and Zou, X. K.: Changes in temperature and precipi-
based analysis of projected climate changes to bivariate flood tation and their impacts on drought in China during 1951–2003,
quantiles, J. Hydrol., 566, 23–42, 2018. Advances in Climate Change Research, 1, 16–18, 2005.
Yin, J., Guo, S., Gu, L., He, S., Ba, H., Tian, J., Li, Zhang, F., Deng, X., Xie, L., and Xu, N.: China’s energy-
Q., and Chen, J.: Projected changes of bivariate flood related carbon emissions projections for the shared socioe-
quantiles and estimation uncertainty based on multi- conomic pathways, Resour. Conserv. Recy., 168, 105456,
model ensembles over China, J. Hydrol., 585, 124760, https://doi.org/10.1016/j.resconrec.2021.105456, 2021.
https://doi.org/10.1016/j.jhydrol.2020.124760, 2020. Zhang, G., Gan, T. Y., and Su, X.: Twenty-first century drought
Yin, J., Guo, S., Gu, L., Zeng, Z., Liu, D., Chen, J., analysis across China under climate change, Clim. Dynam., 59,
Shen, Y., and Xu, C.-Y.: Blending multi-satellite, atmo- 1665–1685, 2022.
spheric reanalysis and gauge precipitation products to fa- Zhao, M., A, G., Velicogna, I., and Kimball, J. S.: Satellite Obser-
cilitate hydrological modelling, J. Hydrol., 593, 125878, vations of Regional Drought Severity in the Continental United
https://doi.org/10.1016/j.jhydrol.2020.125878, 2021a. States Using GRACE-Based Terrestrial Water Storage Changes,
Yin, J., Guo, S., Gentine, P., Sullivan, S. C., Gu, L., He, J. Climate, 30, 6297–6308, https://doi.org/10.1175/JCLI-D-16-
S., Chen, J., and Liu, P.: Does the hook structure con- 0458.1, 2017.
strain future flood intensification under anthropogenic cli- Zheng, J., Wang, H., and Liu, B.: Impact of the long-term precipita-
mate warming?, Water Resour. Res., 57, e2020WR028491, tion and land use changes on runoff variations in a humid sub-
https://doi.org/10.1029/2020WR028491, 2021b. tropical river basin of China, Journal of Hydrology: Regional
Yin, J., Guo, S., Yang, Y., Chen, J., Gu, L., Wang, J., He, Studies, 42, 101136, https://doi.org/10.1016/j.ejrh.2022.101136,
S., Wu, B., and Xiong, J.: Projection of droughts and their 2022.
socioeconomic exposures based on terrestrial water storage Zhu, Q., Luo, Y., Zhou, D., Xu, Y.-P., Wang, G., and Tian, Y.:
anomaly over China, Sci. China Earth Sci., 65, 1772–1787, Drought prediction using in situ and remote sensing products
https://doi.org/10.1007/s11430-021-9927-x, 2022. with SVM over the Xiang River Basin, China, Nat. Hazards, 105,
Yin, J., Gentine, P., Slater, L., Gu, L., Pokhrel, Y., Hanasaki, N., 2161–2185, 2021.
Guo, S., Xiong, L., and Schlenker, W.: Future socio-ecosystem
productivity threatened by compound drought–heatwave events,
Nat. Sustain., 6, 259–272, https://doi.org/10.1038/s41893-022-
01024-1, 2023a.
Yin, J., Guo, S., Wang, J., Chen, J., Zhang, Q., Gu, L., Yang,
Y., Tian, J., Xiong, L., and Zhang, Y.: Thermodynamic driving
mechanisms for the formation of global precipitation extremes
and ecohydrological effects, Sci. China Earth Sci., 66, 92–110,
https://doi.org/10.1007/s11430-022-9987-0, 2023b.

Hess 28 3305 2024

Uploaded by

Copyright:

Available Formats

Hess 28 3305 2024

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hess 28 3305 2024

Uploaded by

Copyright:

Available Formats

Hydrol. Earth Syst. Sci.

, 28, 3305–3326, 2024

Machine-learning-constrained projection of bivariate hydrological

Yichang 443002, Hubei Province, China

Engineering, China University of Geosciences, Wuhan 430074, PR China

Correspondence: Jiabo Yin ([email protected])

Received: 22 July 2023 – Discussion started: 6 September 2023

Published by Copernicus Publications on behalf of the European Geosciences Union.

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

2 Methodology 2.2 Sensitivity analysis of meteorological variables for

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

distribution of drought duration and severity is constructed

pressed at the catchment scale as follows:

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

https://doi.org/10.5194/hess-28-3305-2024 Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024

Hydrol. Earth Syst. Sci., 28, 3305–3326, 2024 https://doi.org/10.5194/hess-28-3305-2024

You might also like