Science of the Total Environment 665 (2019) 338–346

Contents lists available at ScienceDirect

Science of the Total Environment

journal homepage: www.elsevier.com/locate/scitotenv

Meteorological drought forecasting based on a statistical model with

machine learning techniques in Shaanxi province, China
Rong Zhang, Zhao-Yue Chen, Li-Jun Xu, Chun-Quan Ou ⁎
State Key Laboratory of Organ Failure Research, Department of Biostatistics, Guangdong Provincial Key Laboratory of Tropical Disease Research, School of Public Health, Southern Medical Univer-
sity, Guangzhou 510515, China

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• A new modeling strategy was devel-

oped to predict SPEI and droughts at
32 stations in Shaanxi province, China.
• The distributed lag non-linear model
outperformed CCF in selecting the
optimal predictors and specifying their
lag time.
• XGBoost had better prediction accuracy
for droughts with a lead time of 1-6
months (89%~97%) than ANN.

a r t i c l e i n f o a b s t r a c t

Article history: Background: Drought is a major natural disaster that causes severe social and economic losses. The prediction of
Received 2 October 2018 regional droughts may provide important information for drought preparedness and farm irrigation. The existing
Received in revised form 16 January 2019 drought prediction models are mainly based on a single weather station. Efforts need to be taken to develop a
Accepted 16 January 2019 new multistation-based prediction model.
Available online 10 February 2019
Objectives: This study optimizes the predictor selection process and develops a new model to predict droughts
Editor: SCOTT SHERIDAN
using past drought index, meteorological measures and climate signals from 32 stations during 1961 to 2016
in Shaanxi province, China.
Keywords: Methods: We applied and compared two methods, including a cross-correlation function and a distrib-
Drought prediction uted lag nonlinear model (DLNM), in selecting the optimal predictors and specifying their lag time.
Distributed lag nonlinear model Then, we built a DLNM, an artificial neural network model and an XGBoost model and compared their
XGBoost validations for predicting the Standardized Precipitation Evapotranspiration Index (SPEI) 1–6 months
Shaanxi province in advance.
Results: The DLNM was better than the cross-correlation function in predictor selection and lag effect de-
termination. The XGBoost model more accurately predicted SPEI with a lead time of 1–6 months than the
DLNM and the artificial neural network, with cross-validation R 2 values of 0.68–0.82, 0.72–0.89,
0.81–0.92, and 0.84–0.95 at 3-, 6-, 9- and 12-month time scales, respectively. Moreover, the XGBoost
model had the highest prediction accuracy for overall droughts (89%–97%) and for three specific drought
categories (i.e., moderate, severe, and extreme) (76%–94%).

⁎ Corresponding author at: State Key Laboratory of Organ Failure Research, Department of Biostatistics, Guangdong Provincial Key Laboratory of Tropical Disease Research, School of
Public Health, Southern Medical University, Guangzhou, Guangdong 510515, China.
E-mail address: ouchunquan@hotmail.com (C.-Q. Ou).

https://doi.org/10.1016/j.scitotenv.2019.01.431
0048-9697/© 2019 Elsevier B.V. All rights reserved.
 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Conclusion: This study offers a new modeling strategy for drought predictions based on multistation
data. The incorporation of nonlinear and lag effects of predictors into the XGBoost method can signifi-
cantly improve prediction accuracy of SPEI and drought.
1. Introduction
Drought is a natural disaster caused by below-average precipitation
over a long period of time (Han et al., 2016). Droughts have profound
impacts on agriculture, the economy and the environment. For example,
a widespread drought in central and south-west Asia during 1999–2000
resulted in 60 million people affected and an estimated economic loss of
$4.29 billion (Agrawala et al., 2001). With global warming impacting cli-
mates, extreme weather events, including droughts, have occurred
more frequently and severely in the past two decades (Dai, 2011a,
2011b, 2013), causing food security challenges in some regions. The
prediction and early warning signs of drought are particularly impor-
tant in the management and planning of agricultural resources before
the onset of drought.
The first issue addressed in drought prediction is drought iden-
tification. Over the past few decades, drought indexes based on in-
dividual or multiple hydroclimatic variables have been proposed. In
early studies, a drought was defined based on daily/monthly pre-
cipitation (Blumenstock Jr., 1942; Mcguire and Palmer, 1957).
Later, several drought indexes were developed, such as the Palmer
Drought Severity Index (PDSI) (Palmer, 1965), the Standardized
Precipitation Index (SPI) (Palmer, 1965) and the Standardized Pre-
cipitation Evapotranspiration Index (SPEI) (Vicente-Serrano et al.,
2010; Beguería et al., 2014). The PDSI has an inherent time scale,
making it slow to respond to the development and attenuation of
a drought (Hayes et al., 1999). The SPI only considers the effect of pre-
cipitation on droughts. As a new drought index, the SPEI reflects the se-
verity of a drought by constructing a water balance equation of supply
and demand that is based on the difference between precipitation and
evapotranspiration. Evapotranspiration is a temperature-based index cal-
culated using the Thornthwaite equation (Thornthwaite, 1948), but it is
usually underestimated in arid and semiarid areas and is overestimated
in humid regions (Jensen et al., 1990). When this occurs, the Food and Ag-
riculture Organization of the United Nations recommends using the
Penman-Monteith correction method to calculate potential evaporation
based on energy, humidity, and wind speed (Richard, 1998). Chen and
Sun (2015) pointed out that the SPEI with the Penman-Monteith correc-
tion outperformed the SPEI with a Thornthwaite correction in monitoring
droughts, especially in arid regions of China.
Another challenge in predicting droughts is the choice and develop-
ment of an appropriate prediction model. Previously, several models
have been used to predict the occurrence and severity of droughts,
mainly including the ARIMA/SARIMA model (Durdu, 2010), neural net-
work model (Morid et al., 2007; Le et al., 2016) and hybrid model
(Mishra et al., 2007). ARIMA is a simple and common method used to
predict droughts in a single city, which is based on the characteristics
of the sequence itself and does not take into account the effects of
other predictors. The artificial neural network made great contributions
in fitting the nonlinear relationship between predictors and the target
variable in various fields. However, it has many shortcomings and prac-
tical limitations in terms of the robustness of high-dimensional data and
irrelevant descriptors, model interpretability, and computational effi-
ciency (Babajide Mustapha and Saeed, 2016). Specifically, the popular
back propagation (BP) neural networks model is likely to drift towards
the local extremum when solving the global extremum of the complex
nonlinear function, thus causing the training to fail. Chen and Guestrin
(2016) proposed the Extreme Gradient Boosting (XGBoost) algorithm,
which combines all predictors and trains all “weak” learners into
339
© 2019 Elsevier B.V. All rights reserved.
“strong” learners through the additive training strategy. The new ap-
proach has been recently used to achieve state-of-the-art results in var-
ious fields, such as daily solar radiation predictions (Fan et al., 2018),
failures in the banking sector (Carmona et al., 2018), disease diagnosis
and prediction (Livne et al., 2018; Taylor et al., 2018); however, its
drought prediction abilities are thus far unknown.
Many other efforts on drought prediction were taken in recent years,
including the availability and selection of the predictand. For example,
remote sensing products (i.e., Normalized Difference Vegetation Index
(NDVI)) have been applied to drought prediction because of the avail-
ability of consistent temporal and spatial measurements at the regional
and global scales (Funk and Brown, 2006; Tadesse et al., 2014; Akarsh
and Mishra, 2015). Moreover, many human activities which are difficult
to quantify have been gradually incorporated in drought prediction,
such as reservoir operation, irrigation, land use change, and deforest
(Ma et al., 2018; Yuan et al., 2017). In the case of low prediction capabil-
ity, probabilistic drought prediction has been popular for end user's de-
cision making (Demargne et al., 2014). A lot of attention has been put on
acquiring the forecast probability density function (PDF) and the prob-
ability of falling in different drought categories (Hao et al., 2016;
Behrangi et al., 2015). Apart from predicting drought signals based on
different indicators, some studies have focused directly on predicting
the impacts of drought on society or ecosystem (Nadolnyak and
Vedenov, 2013; Shafiee-Jood et al., 2014). Drought impact inventories
that cover drought impact from agriculture to water quality have been
proposed (Stahl et al., 2016; Bachmair et al., 2016). These efforts en-
hanced drought prediction to some extent. However, drought predic-
tion is still a major challenge for climatologists, hydrologists and
decision makers because of its complex origin and occurrence at differ-
ent temporal and spatial scales.
In the north and west regions of China, droughts occur frequently
and cause a high rate and degree of loss in agricultural production
(Zhang et al., 2015). Particularly, in the province of Shaanxi, drought is
the most serious and recurring natural disaster (Min et al., 2010), with
approximately 110 moderate droughts and 18 extreme droughts occur-
ring during 1951–2012 (Jiang et al., 2015). This has led to adverse eco-
nomic losses and social consequences in this region, such as water
shortages, a deterioration of water quality, crop failures and locust
plagues (Wang and Zhai, 2003). In this study, we developed a distrib-
uted lag nonlinear model (DLNM) and two machine learning models
(i.e., multilayer perceptron feedforward neural network model with the
Levenberg-Marquardt optimization algorithm and XGBoost) to predict
SPEI in Shaanxi province with a lead time of 1–6 months based on past
SPEI, meteorological measures and climate signals during 1961–2016.
2. Materials and methods
2.1. Study area
Shaanxi is an inland province located in the northwest part of China
between 31°43′–39°35′ N and 105°29′–111°14′ E and has an area of
205,800 km2 (Fig. 1). The local elevation ranges from nearly 170 m to
3767 m, with values ranging higher in the south and north and lower
in the center. Shaanxi has a subtropical monsoon climate, with an aver-
age annual temperature between 8 and 16 °C and annual precipitation
from 340 mm to 1250 mm. In 2016, the proportion of primary industry
within Shaanxi's GDP was 8.73%, the number of employees in primary
industry accounted for 44.37%, and the GDP per capita was $7680,
340 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Fig. 1. The geographical location of Shaanxi province and the spatial distribution of meteorological stations.
which is slightly lower than the national average ($8126) (Shaanxi
Provincial Bureau Of Statistics, 2018).
2.2. Materials
We obtained daily meteorological data collected at 32 weather
stations in Shaanxi from 1961 to 2016 from the China Meteorological
Data Sharing Service (http://data.cma.cn/), including daily average
pressure (hPa), average temperature (°C), average maximum tem-
perature (°C), average minimum temperature (°C), average relative
humidity (1%), average wind speed (m/s), precipitation (mm) and
sunshine duration (h). Then, we calculated the monthly average for
each measure.
We obtained potential climate predictors, such as Oceanic Niño
Index (ONI), Southern Oscillation Index (SOI), Pacific Decadal Oscilla-
tion (PDO), North Atlantic Oscillation (NAO), Atlantic Multidecadal Os-
cillation (AMO) and Interdecadal Pacific Oscillation (IPO) from the
Climate Prediction Center of the National Oceanic and Atmospheric Ad-
ministration of the United States (http://www.noaa.gov/).
Four normalized SPEI at 3-, 6-, 9- and 12-month time scales (SPEI_3,
SPEI_6, SPEI_9, SPEI_12) were calculated based on the difference be-
tween precipitation and potential evapotranspiration (PET). PET reflects
the quantity of evaporation and transpiration when the water source is
plentiful. A monthly reference for PET is calculated according to the
Penman-Monteith equation, and the default original parameterization
was used with a reference crop height of 0.12 m (Allen et al., 1994).
The SPEI is a standardized variable with a mean of 0 and a standard de-
viation of 1. Hence, SPEI can be compared in different times and spaces.
Moreover, SPEI can be computed at different scales. For example, data of
the current and the past five months can be used to compute the SPEI_6
for a given month. The definition and specific classification of drought
based on SPEI is shown in Table 1, which is acknowledged and has
been widely used in different regions (Chen and Sun, 2015; Ali et al.,
2017; Tirivarombo et al., 2018).
2.3. Methodology
2.3.1. Distributed lag nonlinear model (DLNM)
There was significant correlation between the meteorological and
climate predictors and SPEI (Table S5). The stepwise regression was
used to select significant predictors in our study. Once a new variable
is added, each variable in the model is statistically tested to check if it
can be deleted without appreciably increasing the model R-Square.
We used the variance inflation factor (VIF) to measure the degree of
variance growth due to the high correlation between predictors. To
avoid the multicollinearity problem, which can affect the stability of
the model and produce large errors when predicting data outside the
sample (Wold et al., 1984; Naes and Mevik, 2001), among the highly
correlated variable with VIF N10, we chose the one that made the
model predict best. In this study, we modeled the exposure-lag-
response association between predictors and SPEI using DLNM, which
has been widely used in epidemiological studies of the health effects
of meteorological variables (Gasparrini, 2011; Gasparrini et al., 2010).
We used DLNM to build a cross-basis for each independent variable
and then predicted SPEI by linearly combining the cross-basis. We de-
termined the optimal lag of each predictor by looping. The optimal lag
was determined according to the criteria of maximizing R2 and mini-
mizing GCV. Moreover, we added spatially random intercepts to the
Table 1
Categorization of drought and wet grade according to the SPEI.
Categorization SPEI
Extreme drought ≦−2
Severe drought −1.99 to −1.50
Moderate drought −1.49 to −1.00
Normal −0.99 to 0.99
Moderate wet 1.00 to 1.49
Severe wet 1.50 to 1.99
Extreme wet ≧2
 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346 341

model to explain the variations among station locations. Some studies learning rate in each iteration and wt is the weight of the ft(cb. x, season).
pointed out that drought is seasonal (Wang et al., 2015), so the season The formulas for objective functions and specific parameters were
was considered in the model as a categorical variable (1:spring, 2:sum- displayed in Appendix.
mer, 3:autumn, 4:winter). The statistical model for all weather stations
together is as follows: t
f t ðcb:x; seasonÞ ¼ arg min obj ðθt Þ
θt
ð3Þ
SPEIt;i ¼ β0;i þ β1 FactorðseasonÞ þ β2 cb:SPEIt−1;i ¼ arg min Lt ðθt Þ þ Ωt ðθt Þ
θt
þ β3 cb:meteorologicalt−1;i þ β4 cb:climatet−1;i ð1Þ


þ ε; ε  N 0; σ 2
We trained the model using XGBLinear in the XGBoost package in R.
A model evaluation was performed using the R package of “CARET”
where the SPEIt, i represents the monthly drought index for the sta-
(Classification And REgression Training) (Kuhn, 2015), which is a
tion i in month t; β0, i denotes the random intercept; Factor(season)
framework for streamlining the process of creating and comparing pre-
denotes the dummy variables of season; cb. SPEIt−1, i denotes the
dictive algorithms. Finally, all tuning parameters in XGBoost were deter-
cross-basis of the SPEI in the previous month for the station i; and
mined by 10-fold cross-validation.
cb. meteorologicalt−1, i and cb. climatet−1, i denote the cross-basis of
the meteorological predictors (average maximum temperature, av-
2.3.3. Model validation
erage pressure, average precipitation, average wind speed, average
To verify the validation and stability of the model for predicting SPEI,
relative humidity and sunshine duration) and climate predictors
we used the 10-fold cross-validation (CV) method. Briefly, data is ran-
(Pacific Decadal Oscillation, Southern Oscillation Index, Interdecadal
domly divided into 10 groups by serial number or time. Nine of the
Pacific Oscillation, Atlantic Multidecadal Oscillation, North Atlantic
groups are used to build a model, which is called a training dataset,
Oscillation, Oceanic Niño Index), respectively, in the previous
and the remaining group, called a test dataset, is used to validate the
month for the station i.
model. This process is repeated 10 times, and the average CV R2 (coeffi-
cient of determination), CV RMSE (root mean square error) and CV MAE
2.3.2. Machine learning models
(mean absolute error) are then obtained.
This study adopted the multilayer perceptron feed-forward neural
Compared with SPEI predictions, the forecasting of a drought cate-
network due to its popularity. A neural network is a complex network
gory is more important in practice because decision making regarding
structure composed of an input layer, hidden layer and output layer.
drought is usually done based on the severity of drought. Therefore, we
The neurons between layers are connected by weight and error, but
calculated the percentage of droughts successfully predicted, that is,
there is no connection between the neurons within layers. The working
the observed SPEI and the predicted SPEI are both less than −1 (Chen
principle of the multilayer perceptron is to minimize the error between
and Sun, 2015; Somorowska, 2016; Yao et al., 2018). Furthermore, we
the output value of the model and the target value by updating the
used the overall accuracy, the producer's accuracy and the user's accu-
weights (Haykin, 1994). The Levenberg-Marquardt algorithm was
racy for each category to assess the accuracy of drought classification
used to update the weights in this study. We used the cross-
(Rhee and Im, 2017; Huang and Chou, 2008). The overall accuracy repre-
correlation function to evaluate potential lagged relationships over 0–-
sents the percentage of actual drought categories that are classified cor-
24 months between meteorological measures, climate signals and the
rectly. The producer's accuracy is the proportion of actual drought
SPEI index. The lag terms with the highest correlation with SPEI were
conditions that are forecasted correctly, and the user's accuracy is the
chosen as predictors in the input layer. All variables in the input layer
proportion of predicted drought conditions that happen in the field.
were transformed using the min-max normalization method, and the
output of the neural network was inversely normalized for calculating
2.3.4. Sensitivity analysis
the correlation with the actual value. We varied the number of neurons
Many covariates may have a potential lag and nonlinear effects on
in the hidden layer from 1 to 2n + 8 to find an optimal number accord-
SPEI, therefore, determining their effect patterns is important. In our
ing to the result of cross-validation, where n is the number of input var-
main analysis, the effects of all covariates were flexibly fitted using a
iables (Le et al., 2016). The learning rate, momentum and number of
cross-basis among three models. In previous studies using machine
iterations were set to 0.1, 0.9 and 5000, respectively. Since the initial
learning models, the cross-correlation function (CCF) was usually used
weights of the neural network are assigned randomly, the results of
to choose the lag term with the highest correlation as the predictor. In
each run are different. To solve this problem, we ran 20 times per set
our sensitivity analysis, to determine whether the cross-basis method
of parameters and then calculated the average. We used MATLAB
is superior to the traditional CCF method, we examined the prediction
2014a software to construct the neural network model.
capacity of ANN and the XGBoost model using the CCF-based predictor
XGBoost stands for “Extreme Gradient Boosting”, where Gradient
selection strategy.
Boosting is a machine learning technique for regression and classifica-
tion problems, which produces a prediction model in the form of an en-
3. Results
semble of weak prediction models. As an efficient and scalable variant of
the Gradient Boosting Machine (Friedman, 2001), XGBoost has won
During 1961–2016 in Shaanxi province, China, the mean SPEI at 32
several machine learning competitions in recent years based on its con-
stations for four different time scales, namely SPEI_3, SPEI_6, SPEI_9
venience, parallelism and impressive predictive accuracy (Adam-
and SPEI_12, had an average of 0.002, 0.002, 0.003, and 0.005 and stan-
Bourdarios et al., 2015). XGBoost is used for supervised learning prob-
dard deviations of 0.987, 0.985, 0.984, and 0.980, respectively. For the
lems, we use the training data (cb.x in DLNM and season) to predict a
four different time scales, there were 10.14%–10.81% of all months dur-
target variable SPEI. The model is described by the following:
ing the study period having moderate droughts on average at 32 sta-
tions, 4.68%–5.33% of months having severe droughts, and 1.23%–
X
t
^ði t Þ ¼
y f k ðcb:x; seasonÞ ¼ y
ðt−1Þ
^i þ εwt f t ðcb:x; seasonÞ ð2Þ 1.42% of months having extreme droughts, respectively (Table 2). The
k¼1 occurrence frequency and proportion of drought categories differed
with time scales at the same station and varied by stations at the same
where the parameters θt of model ft(cb. x,season) were selected by opti- time scale (Tables S1–S4).
mizing the objective function objt(θt) which contains the loss function Lt We made the correlation analysis between the predictors and
(θt) and the regularizing term Ωt(θt) used in each iteration t, ε is the drought indicators in Table S5. The partial R-square of every predictor
342 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Table 2
Average SPEI and the average number of three drought categories at 32 weather stations in Shaanxi province, 1961–2016.
SPEI Max Mean Min SD
SPEI_3 2.653 0.002 −2.578 0.987
SPEI_6 2.535 0.002 −2.599 0.985
SPEI_9 2.438 0.003 −2.583 0.984
SPEI_12 2.358 0.005 −2.560 0.980
in DLNM was displayed in Tables S6 and S7. The significant predictors
and their corresponding optimal lags based on the cross-basis generated
in the DLNM were listed in Table S8. Three models (i.e., DLNM, ANN and
XGBoost model) were applied in 1–6-month predictions of multistation
SPEI in Shaanxi. The cross-validation R2 values of the DLNM model were
0.21–0.57, 0.26–0.62, 0.37–0.67 and 0.43–0.92 for SPEI_3, SPEI_6,
SPEI_9 and SPEI_12, respectively. The cross-validation R2 values of the
ANN model were 0.412–0.680, 0.46–0.74, 0.64–0.81 and 0.57–0.88, re-
spectively. The XGBoost model achieved the highest correlation values,
with cross-validation R2 values of 0.68–0.82, 0.72–0.89, 0.81–0.92 and
0.84–0.95, respectively. The R2 values of one-month predictions of
SPEI_12 were up to 0.955. Fig. 2 shows the R2, RMSE and MAE values
of the three models for the 1–6 month predictions. For each model, the
prediction capacity became demonstrably less accurate (i.e., smaller R2
and greater RMSE and MAE) with the extension of lead time because
of accumulation of prediction error at each time step. In addition, the
prediction capacity became better with the increase in the time scale,
probably because of the increase of filter length used in the calculation
of SPEI which more effectively reduces the noise.
Among the three models, the XGBoost model had the highest predic-
tion accuracy for future droughts. Based on SPEI_12, the XGBoost suc-
cessfully predicted 94%–97% droughts with a 1–6-month lead time.
The successful prediction proportions of droughts were 93%–96%,
91%–95% and 89%–94% for SPEI_9, SPEI_6 and SPEI_3, respectively.
Moreover, the overall prediction accuracies for specific drought catego-
ries were 76%–85% for SPEI_3, 81%–89% for SPEI_6, 87%–92% for SPEI_9,
and 89%–94% for SPEI_12. Fig. 3 shows the producer's accuracy and
user's accuracy of each specific drought category (i.e., moderate, severe
and extreme drought) using the three different methods. In all time
scales and lead times, the XGBoost model showed the highest
producer's accuracy for each category compared to the DLNM and
ANN models. Using the XGBoost model, the producer's accuracy of the
three drought categories was up to 88% to 95% for SPEI_12, 88% to 93%
for SPEI_9 78.55%, 74% to 90% for SPEI_6, and 60% to 87% for SPEI_3.
For user's accuracy, the XGBoost model performed best in the identifica-
tion of moderate droughts and severe droughts, but the ANN model had
a 100% user's accuracy in the classification of extreme droughts. Using
the XGBoost model, the user's accuracy of three types of drought ranged
from 91% to 99% for SPEI_12, 89% to 99% for SPEI_9, 84% to 96% for
SPEI_6, and 82% to 96% for SPEI_3 (detailed values listed in Table S9).
In the sensitivity analysis, we found that the predictors and their op-
timal lags selected by the cross-basis (Table S8) were different from those
of the CCF (Table S10). Table S11 shows the results for the ANN and
XGBoost models using CCF-based predictors as the input dataset. Gener-
ally, the CCF-based method was less accurate than the cross-basis. In ad-
dition, we compared the results of simultaneous prediction of 32 stations
with the individual prediction of each station (Table S12). The results
showed that the former had much better prediction performance.
4. Discussion
Accurately predicting future droughts has useful implications for
risk management, drought preparedness and farm irrigation. Instead
of predicting droughts at a single station or predicting the average
droughts of multiple stations, we simultaneously predicted droughts
in 32 weather stations in Shaanxi, China. Compared with the individual
prediction of each station, the multistation-based prediction has greater
Moderate drought %(n) Severe drought %(n) Extreme drought %(n)
10.80 (73.84) 5.33 (36.44) 1.23 (8.41)
10.81 (73.97) 5.18 (35.47) 1.25 (8.56)
10.21 (69.84) 4.98 (34.06) 1.33 (9.09)
10.14 (69.41) 4.68 (31.38) 1.42 (9.75)
practical significance, because we can provide drought warning for a
large region by simultaneously modeling, which are in favor of the de-
velopment of the regional integrated drought management program.
Moreover, simultaneously modeling indeed improve the prediction be-
cause the machine learning needs a large sample to train the model and
to guarantee the model robust and simultaneously modeling has obvi-
ous advantages in this aspect. This paper compared three procedures
(i.e., DLNM, ANN and XGBoost model) in drought forecasting. The
final results show that the XGBoost model performed the best in SPEI
predictions with a lead time of 1–6 months. The XGBoost model also
achieved higher producer's and user's accuracies for the three drought
categories.
First, we made improvements in the selection of predictors and the
identification of their optimal lag time in drought prediction. For ma-
chine learning methods, the variable selection is a particularly critical
task, because using unrelated variables or redundant variables as part
of the training process can increase the cost and the runtime of the pre-
diction system and lead to a less accurate generalization of the model. In
previous studies on drought predictions, there were mainly three ways
to select variables and lags. First, the ARIMA/SARIMA model or some hy-
brid models only used the past drought index, such as SPI or SPEI, as the
predictor. Autocorrelation and partial-correlation functions (ACF and
PACF) were calculated to identify the lags that can explain the largest
variance (Durdu, 2010; Mishra et al., 2007; Mouatadid et al., 2018).
Since only the past lagged SPI or SPEI signals were used as model inputs,
this approach ignores the effects of other covariates and cannot simulta-
neously predict droughts in multiple regions. Second, the significant
predictors that had been previously reported in the literature were con-
sidered, and their lags were determined by combing and trying (Morid
et al., 2007). This approach would be difficult to perform with many pre-
dictors. Last, the cross-correlation function was widely used to select
significant predictors and appropriate lags (Tian et al., 2017; Deo et al.,
2017; Le et al., 2016). Because the predictors are mutually influenced
and work together, there is a deviation in determining the optimal lag
of each predictor based on CCF separately. To solve this problem, we
proposed the use of a stepwise regression with a cross basis in the selec-
tion of predictors and lags in the first drought prediction for the first
time. The cross-basis combines the basis matrices of the predictors
and lags and then specifies the dependency simultaneously in the two
dimensions (Tian et al., 2017; Deo et al., 2017; Le et al., 2016). The step-
wise regression was initially used to select significant predictors, the
distributed lag nonlinear model was used to construct the cross-basis
for each predictor, and the optimal lag was identified through a loop ac-
cording to the fitness of the model. Our results revealed that compared
with using the CCF method, using the cross-basis improved the R2 of the
ANN model by 37–428%, 18–197%, 15–103% and 9–42% with a lead time
of 1–6 months for SPEI_3, SPEI_6, SPEI_9 and SPEI_12, respectively. For
XGBoost, using the cross-basis also generally produced a more accurate
prediction.
Several kinds of methods have been used to predict droughts in pre-
vious studies. ARIMA made great contributions to early single-station
drought predictions due to its simplicity (Durdu, 2010; Mishra et al.,
2007), but the unsatisfactory prediction accuracy and the restriction to
a single site limited its widespread use. Recently, ANNs have found
many applications including drought prediction because of their ability
to reproduce and fit nonlinear relationships (Mishra et al., 2007; Morid
et al., 2007; Mishra and Desai, 2006), but superior and effective neural
 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Fig. 2. Trends of the cross-validation results for different models of predicting multistation SPEI with the increase in lead time. The red line represents the XGBoost model, the orange line represents the ANN model, and the green line represents the
DLNM model.

343
 344
R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Fig. 3. The stacked bar chart of producer's and user's accuracies for different models of predicting multistation SPEI with the increase in lead time. The four figures in the first row represent the producer's accuracy for SPEI_3, SPEI_6, SPEI_9 and
SPEI_12. The four figures in the second row represent the user's accuracy for SPEI_3, SPEI_6, SPEI_9 and SPEI_12. The 6 clusters in each figure represent the lead time of 1–6 months, and the three columns in each cluster in the figure represent
the DLNM, ANN, and XGBoost models from left to right. The three sections in each column represent the producer's or user's accuracies of the moderate drought, severe drought, and extreme drought from bottom to top.
 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
networks usually consume extreme amounts of computing resources
(Edwards, 2015). Improving the training efficiency and convergence
ability of a neural network has always been a research hot spot. To
avoid the dilemmas posed by the ANN model, we developed a
gradient-boosting XGBoost model, which has a higher computational
efficiency and a better capability to solve overfitting problems com-
pared with other machine learning models (Fan et al., 2018). In contrast,
with the simplicity of single decision trees, extreme gradient boosting
produces an intricate structure with hundreds or even thousands of
trees (Capatina et al., 2017). This complexity may bring challenges in
interpreting the final model. We used an extreme gradient boosting
with linear booster XGBlinear to simplify it to an extent. Our research
results proved that the XGBoost method was much more accurate
than the ANN model, especially for SPEI at a 3-month time scale
(i.e., SPI_3). This prediction time scale is essential for reducing the im-
pacts of agricultural drought because the growth cycle of crops is usu-
ally four months. The producer's and user's accuracy in the XGBoost
model outperformed the ANN model, and the running speed of the
XGBoost model was approximately 30 times faster than the ANN
model in our study. Recent studies conducting drought predictions
using many machine learning models, such as a least support vector re-
gression (LSSVR) and an extreme learning machine (ELM) model used
by Mouatadid et al. (2018) and decision trees and extremely random-
ized trees used by Rhee and Im (2017). There is no consistent conclusion
about which model most accurately predicts drought indices, but Fan
et al. (2018) and Shimoda et al. (2018) maintained that the XGBoost
model had higher prediction accuracies than the support vector ma-
chine (SVM) and random forest (RF) methods in predicting daily solar
radiation and nonparticipation in a health check-up scheme. In addition,
we found that the prediction capacity of the XGBoost model was more
accurate with an increase in the time scale, but became less accurate
with an increase in the lead time; these results were consistent with
previous studies (Mishra et al., 2007; Mishra and Desai, 2006).
Climate signals play important roles in drought predictions. Our
study found that, in addition to the well-known impacts of temperature
and precipitation on droughts, other meteorological measures also
played a role in drought prediction, including air pressure, wind speed,
relative humidity and sunshine duration. Moreover, the role of climate
signals in multistation SPEI predictions cannot be ignored. Consistent
with these considerations, Le et al. (2016) has pointed out that climate
signals can also be used as predictors alone to explain some of the
changes in SPEI. Liu et al. (2016) explored the temporal and spatial pat-
terns of droughts and the relationship between droughts and the ENSO
cycle in northwestern China; they argued that the ENSO cycle in different
evolutionary stages influenced the wet and dry conditions in the north-
west region by affecting the East Asian summer monsoon. The East Asian
summer monsoon system has important effect on the summer rainfall
patterns in China, especially in the monsoon regions. In addition, other
circulation systems (i.e., the westerly flow, the southwesterly flow asso-
ciated with Indian summer monsoon, and the northerly flow stemming
from Siberia and Mongolia in winter and spring) could have an impact
on the variability of wet and dry conditions in the study area (Qian
et al., 2007; Yu and Kao, 2007; Shi et al., 2007; Li et al., 2008). Therefore,
the impact of climate signals cannot be ignored in drought predictions.
This study has some limitations. First, due to the difficulty of obtaining
solar radiation data, the observed sunshine hours were used instead of the
solar radiation in our SPEI calculations. Second, if there are too many pre-
dictors considered, it would take a long time to determine the optimal lag
of each independent variable in the cross-basis through the loop. Last, to
simplify the model, we used an extreme gradient boosting with a linear
booster rather than tree booster that may perform better.
5. Conclusions
Reliable and efficient variable selection strategies and prediction
models are useful in understanding and predicting drought events. In
345
this study, the CCF and cross-basis approaches were applied to deter-
mine predictors and lags in an input dataset. A traditional statistical
model, DLNM, and two machine learning algorithms, ANN and XGBoost,
were used to predict the SPEI in four time scales and six lead times from
January 1961 to December 2016. The performances of the models in
predicting SPEI and classifying three drought categories were evaluated
using a 10-fold cross-validation method. We found that by using the
predictor selection method of cross-basis, the XGBoost model showed
the best predictive skills. This study provides a useful modeling strategy
for drought prediction, which may therefore assist in making guidelines
for mitigation plans, such as development policies for sustainable
water-use and the management of water supply systems.
Acknowledgements
This study was supported by National Nature Science Foundation of
China [81573249], Nature Science Foundation of Guangdong Province
[2016A030313530].
Declarations of interest
None.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.scitotenv.2019.01.431.
References
Adam-Bourdarios, C., Cowan, G., Germain-Renaud, C., Guyon, I., Kégl, B., Rousseau, D.,
2015. The Higgs machine learning challenge. J. Phys. Conf. Ser. 664, 72015. https://
doi.org/10.1088/1742-6596/664/7/072015.
Agrawala, S., Barlow, M., Cullen, H., Lyon, B., 2001. The drought and humanitarian crisis in
central and southwest Asia: a climate perspective. IRI Special Report, No 01–11. Pal-
isades, International Research Institute for Climate and Society.
Akarsh, A., Mishra, V., 2015. Prediction of vegetation anomalies to improve food security
and water management in India. Geophys. Res. Lett. 42, 5290–5298. https://doi.org/
10.1002/2015GL063991.
Ali, Z., Hussain, I., Faisal, M., Nazir, H.M., Moemen, M.A., Hussain, T., Shamsuddin, S., 2017.
A novel multi-scalar drought index for monitoring drought: the standardized precip-
itation temperature index. Water Resour. Manag. 31, 4957–4969. https://doi.org/
10.1007/s11269-017-1788-1.
Allen, R.G., Smith, M., Perrier, A., Pereira, L.S., 1994. An update for the definition of refer-
ence evapotranspiration. J. Environ. Sci. Health A Tox. Hazard. Subst. Environ. Eng. 43,
1–35.
Babajide Mustapha, I., Saeed, F., 2016. Bioactive molecule prediction using extreme gradi-
ent boosting. Molecules 21, 983. https://doi.org/10.3390/molecules21080983.
Bachmair, S., Svensson, C., Hannaford, J., Barker, L.J., Stahl, K., 2016. A quantitative analysis
to objectively appraise drought indicators and model drought impacts. Hydrol. Earth
Syst. Sci. 20, 2589–2609. https://doi.org/10.5194/hess-20-2589-2016.
Beguería, S., Vicente-Serrano, S.M., Reig, F., Latorre, B., 2014. Standardized Precipitation
Evapotranspiration Index (SPEI) revisited: parameter fitting, evapotranspiration
models, tools, datasets and drought monitoring. Int. J. Climatol. 34, 3001–3023.
https://doi.org/10.1002/joc.3887.
Behrangi, A., Nguyen, H., Granger, S., 2015. Probabilistic seasonal prediction of meteoro-
logical drought using the bootstrap and multivariate information. J. Appl. Meteorol.
Climatol. 54, 1510–1522. https://doi.org/10.1175/JAMC-D-14-0162.1.
Blumenstock Jr., G., 1942. Drought in the United States analyzed by means of the theory of
probability. Tech. Bull.. 819. U.S. Dep. Agric., Washington, D. C.
Capatina, A., Bleoju, G., Matos, F., Vairinhos, V., 2017. Leveraging intellectual capital
through Lewin's force field analysis: the case of software development companies.
J. Innov. Knowl. 2, 125–133. https://doi.org/10.1016/j.jik.2016.07.001.
Carmona, P., Climent, F., Momparler, A., 2018. Predicting failure in the U.S. banking sector:
an extreme gradient boosting approach. Int. Rev. Econ. Financ. 2018. https://doi.org/
10.1016/j.iref.2018.03.008.
Chen, T., Guestrin, C., 2016. XGBoost: a scalable tree boosting system. ACM, 785–794
https://doi.org/10.1145/2939672.2939785.
Chen, H.P., Sun, J.Q., 2015. Changes in drought characteristics over China using the stan-
dardized precipitation evapotranspiration index. J. Clim. 28, 5430–5447. https://doi.
org/10.1175/JCLI-D-14-00707.1.
Dai, A., 2011a. Characteristics and trends in various forms of the Palmer drought severity
index during 1900–2008. J. Geophys. Res. 116, D12115. https://doi.org/10.1029/
2010JD015541.
Dai, A., 2011b. Drought under global warming: a review. Wiley Interdiscip. Rev. Clim.
Chang. 2, 45–65. https://doi.org/10.1002/wcc.81.
346 R. Zhang et al. / Science of the Total Environment 665 (2019) 338–346
Dai, A., 2013. Increasing drought under global warming in observations and models. Nat.
Clim. Chang. 3, 171. https://doi.org/10.1038/nclimate1811.
Demargne, J., Wu, L., Regonda, S.K., Brown, J.D., Lee, H., He, M., Seo, D.J., Hartman, R., Herr,
H.D., Fresch, M., Herr, D., A. M. F., 2014. The science of NOAA's operational hydrologic
ensemble forecast service. Bull. Am. Meteorol. Soc. 1, 79–98.
Deo, R.C., Kisi, O., Singh, V.P., 2017. Drought forecasting in eastern Australia using multi-
variate adaptive regression spline, least square support vector machine and M5Tree
model. Atmos. Res. 184, 149–175. https://doi.org/10.1016/j.atmosres.2016.10.004.
Durdu, O.F., 2010. Application of linear stochastic models for drought forecasting in the
Büyük Menderes river basin, western Turkey. Stoch. Env. Res. Risk A. 24, 1145–1162.
https://doi.org/10.1007/s00477-010-0366-3.
Edwards, C., 2015. Growing pains for deep learning. Commun. ACM 58, 14–16.
Fan, J.L., Wang, X.K., Wu, L.F., Zhou, H.M., Zhang, F.C., Yu, X., Lu, X.H., Xiang, Y.Z., 2018.
Comparison of support vector machine and extreme gradient boosting for predicting
daily global solar radiation using temperature and precipitation in humid subtropical
climates: a case study in China. Energy Convers. Manag. 164, 102–111. https://doi.
org/10.1016/j.enconman.2018.02.087.
Friedman, J., 2001. Greedy function approximation: a gradient boosting machine. Ann.
Stat. 29, 1189–1232.
Funk, C.C., Brown, M.E., 2006. Intra-seasonal NDVI change projections in semi-arid Africa.
Remote Sens. Environ. 101, 249–256. https://doi.org/10.1016/j.rse.2005.12.014.
Gasparrini, A., 2011. Distributed lag linear and non-linear models in R: the package dlnm.
J. Stat. Softw. 43, 1–20.
Gasparrini, A., Armstrong, B., Kenward, M.G., 2010. Distributed lag non-linear models.
Stat. Med. 29, 2224–2234. https://doi.org/10.1002/sim.3940.
Han, L., Zhang, Q., Ma, P., Jia, J., Wang, J., 2016. The spatial distribution characteristics of a
comprehensive drought risk index in southwestern China and underlying causes.
Theor. Appl. Climatol. 124, 517–528. https://doi.org/10.1007/s00704-015-1432-z.
Hao, Z., Hao, F., Singh, V.P., Sun, A.Y., Xia, Y., 2016. Probabilistic prediction of hydrologic
drought using a conditional probability approach based on the meta-Gaussian
model. J. Hydrol., 772–780 https://doi.org/10.1016/j.jhydrol.2016.09.048.
Hayes, M.J., Svoboda, M.D., Wilhite, D.A., Vanyarkho, O.V., 1999. Monitoring the 1996
drought using the standardized precipitation index. Bull. Am. Meteorol. Soc. 80,
429–438. https://doi.org/10.1175/1520-0477(1999)080b0429:MTDUTSN2.0.CO;2.
Haykin, S., 1994. Neural networks: a comprehensive foundation. Neural Networks a Com-
prehensive Foundation, pp. 71–80.
Huang, W., Chou, C., 2008. Risk-based drought early warning system in reservoir operation.
Adv. Water Resour. 31, 649–660. https://doi.org/10.1016/j.advwatres.2007.12.004.
Jensen, M.E., Burman, R.D., Allen, R.G., 1990. Evapotranspiration and irrigation water re-
quirements: a manual. Am. Soc. Civil Eng. (70), 332.
Jiang, R., Xie, J., He, H., Luo, J., Zhu, J., 2015. Use of four drought indices for evaluating
drought characteristics under climate change in Shaanxi, China: 1951–2012. Nat.
Hazards 75, 2885–2903. https://doi.org/10.1007/s11069-014-1468-x.
Kuhn, M., 2015. Caret: classification and regression training. Astrophysics Source Code Li-
brary. 129, pp. 291–295.
Le, M.H., Perez, G.C., Solomatine, D., Nguyen, L.B., 2016. Meteorological drought fore-
casting based on climate signals using artificial neural network - a case study in
Khanhhoa Province Vietnam. Procedia Eng. 154, 1169–1175. https://doi.org/
10.1016/j.proeng.2016.07.528.
Li, W.L., Wang, K.L., Fu, S.M., Jiang, H., 2008. The interrelationship between regional west-
erly index and the water vapor budget in Northwest China. J. Glaciol. Geocryol. 30,
28–34.
Liu, Z., Menzel, L., Dong, C., Fang, R., 2016. Temporal dynamics and spatial patterns of
drought and the relation to ENSO: a case study in Northwest China. Int. J. Climatol.
36, 2886–2898. https://doi.org/10.1002/joc.4526.
Livne, M., Boldsen, J.K., Mikkelsen, I.K., Fiebach, J.B., Sobesky, J., Mouridsen, K., 2018. Boosted
tree model reforms multimodal magnetic resonance imaging infarct prediction in acute
stroke. Stroke 49, 912–918. https://doi.org/10.1161/STROKEAHA.117.019440.
Ma, F., Luo, L., Ye, A., Duan, Q., 2018. Seasonal drought predictability and forecast skill in
the semi-arid endorheic Heihe River basin in northwestern China. Hydrol. Earth
Syst. Sci. 22, 5697–5709. https://doi.org/10.5194/hess-22-5697-2018.
Mcguire, J.K., Palmer, W.C., 1957. The 1957 drought in the eastern United States. Mon.
Weather Rev. 85, 305–314. https://doi.org/10.1175/1520-0493(1957)085b0305:
TDITEUN2.0.CO;2.
Min, A.C., Han, Q.H., Jia, J.K., 2010. Situation Analysis of Shaanxi Province. Cpesap Net.
Mishra, A.K., Desai, V.R., 2006. Drought forecasting using feed-forward recursive neural
network. Ecol. Model. 198, 127–138. https://doi.org/10.1016/j.ecolmodel.2006.04.017.
Mishra, A.K., Desai, V.R., Singh, V.P., 2007. Drought forecasting using a hybrid stochastic
and neural network model. J. Hydrol. Eng. 12, 626–638. https://doi.org/10.1061/
(ASCE)1084-0699(2007)12:6(626).
Morid, S., Smakhtin, V., Bagherzadeh, K., 2007. Drought forecasting using artificial neural
networks and time series of drought indices. Int. J. Climatol. 27, 2103–2111. https://
doi.org/10.1002/joc.1498.
Mouatadid, S., Raj, N., Deo, R.C., Adamowski, J.F., 2018. Input selection and data-driven
model performance optimization to predict the standardized precipitation and evap-
oration index in a drought-prone region. Atmos. Res. 212, 130–149. https://doi.org/
10.1016/j.atmosres.2018.05.012.
Nadolnyak, D., Vedenov, D., 2013. Information value of climate forecasts for rainfall index
insurance for pasture, rangeland, and forage in the Southeast United States. J. Agric.
Appl. Econ. 01, 109–124. https://doi.org/10.1017/S1074070800004612.
Naes, T., Mevik, B., 2001. Understanding the collinearity problem in regression and dis-
criminant analysis. J. Chemom. 15, 413–426. https://doi.org/10.1002/cem.676.
Palmer, W.C., 1965. Meteorologic Drought. US Department of Commerce, Weather Bu-
reau, Washington DC (No. 45).
Qian, W., Lin, X., Zhu, Y., Xu, Y., Fu, J., 2007. Climatic regime shift and decadal anomalous
events in China. Clim. Chang. 84, 167–189. https://doi.org/10.1007/s10584-006-
9234-z.
Rhee, J., Im, J., 2017. Meteorological drought forecasting for ungauged areas based on ma-
chine learning: using long-range climate forecast and remote sensing data. Agric. For.
Meteorol. 237–238, 105–122. https://doi.org/10.1016/j.agrformet.2017.02.011.
Richard, G., 1998. Crop evapotranspiration-guidelines for computing crop water
requirements-FAO irrigation and drainage paper 56. FAO, Rome, Italyhttp://www.
fao.org/docrep/X0490E/x0490e00.htm.
Shaanxi Provincial Bureau Of Statistics, 2018. http://www.shaanxitj.gov.cn/upload/2016/
tongjinianj2016/2016/indexch.htm, Accessed date: 13 August 2018.
Shafiee-Jood, M., Cai, X., Chen, L., Liang, X., Kumar, P., 2014. Assessing the value of seasonal
climate forecast information through an end-to-end forecasting framework: applica-
tion to U.S. 2012 drought in central Illinois. Water Resour. Res. 50, 6592–6609.
https://doi.org/10.1002/2014WR015822.
Shi, Y., Shen, Y., Kang, E., Li, D., Ding, Y., Zhang, G., Hu, R., 2007. Recent and future climate
change in Northwest China. Clim. Chang. 80, 379–393. https://doi.org/10.1007/
s10584-006-9121-7.
Shimoda, A., Ichikawa, D., Oyama, H., 2018. Using machine-learning approaches to predict
non-participation in a nationwide general health check-up scheme. Comput.
Methods Prog. Biomed. https://doi.org/10.1016/j.cmpb.2018.05.032.
Somorowska, U., 2016. Changes in drought conditions in Poland over the past 60 years
evaluated by the standardized precipitation-evapotranspiration index. Acta Geophys.
64, 2530–2549. https://doi.org/10.1515/acgeo-2016-0110.
Stahl, K., Kohn, I., Blauhut, V., Urquijo, J., De Stefano, L., Acácio, V., Dias, S., Stagge, J.H.,
Tallaksen, L.M., Kampragou, E., Van Loon, A.F., Barker, L.J., Melsen, L.A., Bifulco, C.,
Musolino, D., de Carli, A., Massarutto, A., Assimacopoulos, D., Van Lanen, H.A.J.,
2016. Impacts of European drought events: insights from an international database
of text-based reports. Nat. Hazards Earth Syst. Sci. 16, 801–819. https://doi.org/
10.5194/nhess-16-801-2016.
Tadesse, T., Demisse, G.B., Zaitchik, B., Dinku, T., 2014. Satellite-based hybrid drought
monitoring tool for prediction of vegetation condition in Eastern Africa: a case
study for Ethiopia. Water Resour. Res. 50, 2176–2190. https://doi.org/10.1002/
2013WR014281.
Taylor, R.A., Moore, C.L., Cheung, K., Brandt, C., 2018. Predicting urinary tract infections in
the emergency department with machine learning. PLoS One 13, e194085. https://
doi.org/10.1371/journal.pone.0194085.
Thornthwaite, C.W., 1948. An approach toward a rational classification of climate. Geogr.
Rev. 38, 55–94. https://doi.org/10.2307/210739.
Tian, Y., Xu, Y.P., Wang, G., 2017. Agricultural drought prediction using climate indices
based on support vector regression in Xiangjiang River basin. Sci. Total Environ.
622–623, 710–720.
Tirivarombo, S., Osupile, D., Eliasson, P., 2018. Drought monitoring and analysis:
Standardised Precipitation Evapotranspiration Index (SPEI) and Standardised Precip-
itation Index (SPI). Phys. Chem. Earth 106, 1–10. https://doi.org/10.1016/j.
pce.2018.07.001.
Vicente-Serrano, S.M., Beguería, S., López-Moreno, J.I., 2010. A multiscalar drought index
sensitive to global warming: the standardized precipitation evapotranspiration
index. J. Clim. 23, 1696–1718. https://doi.org/10.1175/2009JCLI2909.1.
Wang, Z.W., Zhai, P.M., 2003. Climate change in drought over northern China during
1950–2000. Acta Geograph. Sin. 58, 61–68.
Wang, H., Chen, Y., Pan, Y., Li, W., 2015. Spatial and temporal variability of drought in the
arid region of China and its relationships to teleconnection indices. J. Hydrol. 523,
283–296. https://doi.org/10.1016/j.jhydrol.2015.01.055.
Wold, S., Ruhe, A., Wold, H., Dunn, I.W.J., 1984. The collinearity problem in linear regres-
sion. The Partial Least Squares (PLS) approach to generalized inverses. SIAM J. Sci.
Stat. Comput. 5, 735–743. https://doi.org/10.1137/0905052.
Yao, J., Zhao, Y., Yu, X., 2018. Spatial-temporal variation and impacts of drought in
Xinjiang (Northwest China) during 1961–2015. PeerJ 6, e4926. https://doi.org/
10.7717/peerj.4926.
Yu, J., Kao, H., 2007. Decadal changes of ENSO persistence barrier in SST and ocean heat
content indices: 1958–2001. J. Geophys. Res. Atmos. 112, D13106. https://doi.org/
10.1029/2006JD007654.
Yuan, X., Zhang, M., Wang, L., Zhou, T., 2017. Understanding and seasonal forecasting of
hydrological drought in the Anthropocene. Hydrol. Earth Syst. Sci. 21, 5477–5492.
https://doi.org/10.5194/hess-21-5477-2017.
Zhang, Q., Gu, X., Singh, V.P., Kong, D., Chen, X., 2015. Spatiotemporal behavior of floods
and droughts and their impacts on agriculture in China. Glob. Planet. Chang. 131,
63–72. https://doi.org/10.1016/j.gloplacha.2015.05.007.