Hindawi

Computational Intelligence and Neuroscience

Volume 2022, Article ID 4429286, 11 pages
https://doi.org/10.1155/2022/4429286

Research Article
Drought Assessment Based on Data Fusion and Deep Learning

Yanling Li , Bingyu Wang, and Yajie Gong

School of Mathematics and Statistics, North China University of Water Resources and Electric Power, Zhengzhou 450046, China

Correspondence should be addressed to Yanling Li; liyanling@ncwu.edu.cn

Received 23 April 2022; Revised 30 May 2022; Accepted 6 June 2022; Published 31 July 2022

Academic Editor: Dalin Zhang

Copyright © 2022 Yanling Li et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Drought is a major factor affecting the sustainable development of society and the economy. Research on drought assessment is of
great significance for formulating drought emergency policies and drought risk early warning and enhancing the ability to
withstand drought risks. Taking the Yellow River Basin as the object, this paper utilizes data fusion, copula function, entropy
theory, and deep learning, fuses the features of meteorological drought and hydrological drought into a drought assessment index,
and establishes a long short-term memory (LSTM) network for drought assessment, based on deep learning theory. The results
show that (1) after extracting the features of meteorological drought and hydrological drought, the drought convergence index
(DCI) built on the fused features by copula function can accurately reflect the start and duration of the drought; (2) the drought
assessment indices were effectively screened by judging the causality of the drought system, using the transfer entropy; (3) drawing
on the idea of deep learning, LSTM for drought assessment, which was established on DCI and the drought assessment factors, can
accurately assess the drought risks of the Yellow River Basin.

1. Introduction drought. To make matters worse, the mass construction of

water utilization projects and soil retention projects alters
Drought is a common natural disaster, which is usually the state of the underlying surface and changes the evap-
induced by the abnormal reduction of rainfall. With a high oration, infiltration, runoff yield, and concentration, as well
frequency, a wide range of influence, and a high degree of as the speed of water resource circulation [4].
damage, drought poses a serious threat to agricultural The standardized precipitation index (SPI) is proposed
production, economic development, and social stability [1]. by the World Meteorological Organization (WMO) to
Statistics show that drought has killed 11 million people monitor the severity of drought [5]. The index can be cal-
around the world. In the twentieth century alone, more than culated simply from uncomplex input data, which are easy to
2 billion were affected by drought [2]. The damage of acquire. Applicable to multiple timescales, the SPI can meet
drought is far more serious than that of any other disaster. the needs of different regions and applications, laying the
The economic loss of a drought ranges from 6 to 8 billion basis for monitoring drought at different timescales. As the
USD per year [1]. Drought has obviously affected people’s research goes deeper, scholars were no longer satisfied with
normal production and life, so the study of drought as- the assessment and monitoring of a single type of drought
sessment has become a common concern of researchers. and came up with composite drought indices. For example,
With global warming, drought occurs more frequently Wu et al. [6] constructed a drought index coupling rainfall
and in a wider range across the globe. In many countries and and soil water through fuzzy cloud reasoning and proved the
regions, drought brings far more damage than flood. reliability of the index by historical drought evolution. Won
According to the Sixth Assessment Report (AR6) of the et al. [7] combined the copula function with SPI and the
United Nations (UN) Intergovernmental Panel on Climate evaporation demand drought index into a joint drought
Change (IPCC), from 2010 to 2019, human activities caused index and applied the index to monitor the drought situation
global surface temperature to rise by 0.8°C–1.3°C [3]. The of atmospheric water supply. By kernel entropy component
temperature rise would increase the scale and severity of analysis, Han et al. [8] built a composite drought index
2 Computational Intelligence and Neuroscience

encompassing rainfall, runoff, evapotranspiration, and soil ringing a bell of the severity of drought [13]. In fact, the
water content and found that the index is highly sensitive to Yellow River Basin is the largest drought-stricken basin in
mild drought. China [14]. Drought has brought multiple problems, such as
The drought cannot be fully illustrated by a single dried rivers and soil degradation [15]. It is of realistic sig-
drought index alone. Therefore, many scholars have pro- nificance to study the drought assessment of the Yellow
posed considerable works to construct a series compre- River Basin, which would aid ecological balance and social
hensive drought index that cover as many drought variables development.
as possible [9]. Ren et al. [10] combined SPI, PDSI, and SPEI This paper selects the measured monthly runoffs of
into comprehensive drought indices through fuzzy com- 1980–2019 at six major hydrological stations and the
prehensive evaluation. Maji and Kanrar [11] proposed a monthly meteorological data of 1980–2019 at 21 meteoro-
comprehensive drought index by principal component logical stations along the trunk of the Yellow River. On this
analysis (PCA). Yet the combined drought index based on basis, the meteorological and hydrological drought features
the weighting method and the fuzzy combined method has were fused for assessment. The meteorological data were
certain subjectivity in weighting, and it is easy to cause obtained from China Meteorological Data Service Center
errors, and the combined drought index based on the (http://data.cma.cn), and the hydrological data were col-
principal component analysis method combines related lected from the official site of the Yellow River Conservancy
variables linearly and cannot reflect their nonlinear impact Commission of the Chinese Ministry of Water Resources
characteristics. (http://www.yrcc.gov.cn/).
Based on data fusion, DCI combining meteorological
and hydrological factors was constructed, meteorological 3. Methodology
and hydrological drought for the feature layer fusion by
copula function. Copula function is simply a specification of 3.1. Data Fusion
how univariate marginal distributions combine to form
multivariate distribution. There is no limitation in choosing 3.1.1. Selection of Characteristic Parameters. The effective-
the marginal distribution function, and all margin-free ness of the copula function comes from the fact that the
characteristics can be fully maintained. The DCI not only has function can merge random marginal distributions, which
the characteristics of the meteorological drought index that contain the information of all variables, into a joint dis-
can quickly capture the onset of drought but also has the tribution, without losing or distorting any information. In
advantage of the hydrological drought index that can de- 1993, McKee created SPI [16], a multitimescale drought
scribe the duration of drought, fusing meteorological and index. The SPI can effectively assess the features of drought
hydrological characteristics. The original information can be on different scales, using the occurrence probability of
optimized and combined through data fusion, making it precipitation. The standardized runoff index (SRI) measures
possible to turn multisource information into effective hydrological drought by river runoff and its statistic. The
output. This paper extracts the features of hydrological calculation process of the SRI is similar to that of the SPI
drought and meteorological drought from the data collected [17, 18]. The drought/flood grading standards of SPI and SRI
at six major hydrological stations and meteorological sta- follow the Grades of meteorological drought (GB/T20481-
tions along the trunk of the Yellow River and fuses the 2017), which was formulated by National Technical Com-
features with copula function, producing a hybrid drought mittee on the Standardization of Climate and Climate
index composed of both hydrological and meteorological Change of China.
factors. When selecting the drought factors, the influence
between variables is measured by transfer entropy, and 3.1.2. Copula Function. According to Sklar’s theorem
through this influence, a causal relationship between vari- [19, 20], for any two one-dimensional (1D) random variables
ables is established. Moreover, the LSTM network, a deep X and Y, if their distribution functions are FX(x) � P(X ≤ x)
learning model, was adopted to evaluate drought by drought and FY(y) � P(Y ≤ y), and their joint distribution function is
factors. FC(x,y), then there exists a unique copula function C()
satisfying the following equation:
FC (x, y) � C􏼂FX (x), FY (y)􏼃, x, y ∈ R. (1)
2. Study Area and Data Sources
For two-dimensional (2D) variables, the common forms
The 5,464 km-long Yellow River (Figure 1) is the sixth of copula function include Gaussian, Archimedean, and
longest in the world, and the second longest river in China, student T. Among them, the Archimedean copula is widely
flowing through 9 provinces. The basin of the river, covering applied because it can reflect the dependence intensity with a
an area of 795,000 km2, spans across four geomorphic units: parameter. The Archimedean copula can be further divided
the Qinghai-Tibet Plateau, Inner Mongolia Plateau, the into three types: Gumbel, Clayton, and Frank. With
Loess Plateau, and North China Plain. Since the 1980s, the asymmetric structures, Gumbel and Clayton copulas can
temperature rose significantly in the Yellow River Basin, capture the asymmetric properties. With a symmetric de-
while the rainfall dropped slightly [12]. Since the 1990s, the pendence structure, the Frank copula permits negative de-
basin was hit by increasingly serious droughts. In 2009, a pendence by histogram of variables X and Y. The density
wide and severe drought plagued the Central Plains region, function can be defined as
Computational Intelligence and Neuroscience 3

100° E 105° E 110° E 115° E

Hydrological station
Meteorological station
40° N Main stream 40° N

35° N 35° N

Miles
0 60 120 240 360 480

100° E 105° E 110° E 115° E

Figure 1: Yellow River Basin.

CGu (u, v; θ)(ln u · ln v)θ− 1 3.1.3. Feature Layer Fusion. Let random variable X be the
cGu (u, v; θ)  meteorological drought feature SPI, with a marginal dis-
uv(− ln u)θ +(− ln v)θ 
2− 1/θ
(2) tribution of FX(x). The joint distribution probability of
drought features can be calculated by
(− ln u) +(− ln v)  + θ − 1,
1/θ
FC (x, y)  CFX (x), FY (y)
θ θ

 P(x ≤ X, y ≤ Y) (6)
− θe− θ − 1e− θ(u+v)
cFr (u, v; θ)  (3)  p.
2,
e − 1 + e − 1e − 1
−θ − θu − θv
The drought convergence index (DCI) can be expressed
as
u + v− θ − 1
θ− 1 − θ − 2− 1/θ
cCl (u, v; θ)  (1 + θ)(uv)− , (4)
DCI  φ− 1 (p), (7)
where θ is a parameter; u and v are marginal cumulative
probabilities. where φ is the standard normal distribution.
The parameter θ can be solved by the Kendall rank Referring to classification standards of SPI [21] and the
correlation coefficient τ. For the Gumbel copula, the National Climate Center’s Standard, (GB/T20481-2017),
relationship between θ and τ is τ  1 − 1/θ, θ ∈ [1, ∞). For droughts can be divided into different levels (Table 1) by
the Frank copula, that relationship is τ  1+ severity.
4/θ(1/θ 0 t/et − 1dt − 1), θ∈R/{0}. For the Clayton copula, According to the frequency of historical droughts and
θ

that relationship is τ  θ/θ + 2, θ ∈ (0, ∞). the empirical frequency of DCI, DCI  -1 was taken as the
For different objects, the fused feature between SPI and threshold of drought occurrence: if the DCI value remains
SRI can be described by the∧ Euclidean distance between below -1, the drought must have occurred.
empirical copula function C(u, v) and copula function
C(u, v) :
 3.2. Transfer Entropy. Considering the transmissibility be-
 2
n  
d  i1 C ui , vi  − C ui , vi  .
tween information, Schreiber [22] coined the concept of

∧
(5) transfer entropy based on information entropy theory. Based
on mutual information that reflects the correlation between
The smaller the distance, the better the goodness-of-fit of variables, the transfer entropy measures the causality of
copula function for the variables. information transfer in terms of magnitude and direction.
4 Computational Intelligence and Neuroscience

Table 1: Division of drought levels. between variables X and Y, in the light of the information
DCI Drought levels
provided by variable Z. Transfer entropy considers the past
state of variable Y, and the dependence between variables X
(− ∞,− 2.0] Extreme drought
(− 2.0, − 1.5] Severe drought
and Y. Therefore, the following formula can be derived from
(− 1.5, − 1.0] Moderate drought the relationship between conditional mutual information
(− 1.0, ∞) No drought and transfer entropy:
TEX⟶Y � I Yi+1 ; Xi |Yi 􏼁. (11)
3.2.1. Conditional Mutual Information. Conditional mutual
information refers to the amount of mutual information
acquired about event yj based on the known event xi, under 3.3. Deep Learning. In deep learning, the recurrent neural
the given event zk: network (RNN) has been widely and deeply applied in
natural language processing (NLP), such as speech recog-
p􏼐yj |xi , zk 􏼑 nition, language modeling, and machine translation. The
I􏼐yj ; xi |zk 􏼑 � log . (8)
p􏼐yj |zk 􏼑 RNN has a good memory, supports parameter sharing, and
realizes turing completeness. Therefore, it has lots of ad-
Solving the expectations of variables X, Y, and Z, the vantages in learning the nonlinear features of series.
mean conditional mutual information of X relative to Y Proposed by Hochreiter and Schmidhuber [23], the
under Z can be obtained as follows: LSTM overcomes the defects of RNN by adding a memory
unit to the recurrent layer, putting an end to the problem of
I(Y; X|Z) � E􏽨I􏼐yj ; xi |zk 􏼑􏽩
exploding or vanishing gradients. The LSTM relies on the
N N N
functions of the forget gate, the input gate, the output gate,
p􏼐yj |xi , zk 􏼑 and the memory unit to propagate and memorize long- and
� 􏽘 􏽘 􏽘 p􏼐yj , xi , zk 􏼑log
j�1 i�1 k�1 p􏼐yj |zk 􏼑 short-term information. Figure 2 shows the structure of the
LSTM.
N N N During the operation of the LSTM, the forget gate firstly
� 􏽘 􏽘 􏽘 p􏼐yj , xi , zk 􏼑log p􏼐yj |xi , zk 􏼑 (9) determines which information to forget in the current unit
j�1 i�1 k�1 according to the input at the current state. The output state
of the previous unit is controlled by the sigmoid function:
N N N
− 􏽘 􏽘 􏽘 p􏼐yj , xi , zk 􏼑log p􏼐yj |zk 􏼑 ft � σ 􏼐Wf 􏼂ht− 1 , xt 􏼃 + bf 􏼑. (12)
j�1 i�1 k�1
The input gate consists of two parts. Firstly, the sigmoid
� H(Y|Z) − H(Y|X, Z). layer determines which new information to add into the
current unit. Then, the tanh layer obtains the new candidate
state of the unit. Finally, the two are combined to obtain the
state of the unit at the current moment t:
3.2.2. Transfer Entropy. Suppose discrete variables Xi and Yi,
i � 1, 2, . . ., N are of the same length and mutually act on it � σ Wi 􏼂ht− 1 , xt 􏼃 + bi 􏼁, (13)
each other. Then, the transfer entropy from X to Y reflects
the information transfer from X to Y in the past states. 􏽥 t � tanh WC 􏼂ht− 1 , xt 􏼃 + bC 􏼁.
C (14)
P Yi+1 |Xi , Yi 􏼁 The memory unit C stores the memory in the RNN and
TEX⟶Y � 􏽘 P Yi+1 , Xi , Yi 􏼁log
P Yi+1 |Yi 􏼁 represents the long-term memory. The short-term memory
is denoted by h. Both memories are propagated layer by layer
backward to ensure the memory function of the LSTM:
� 􏽘 P Yi+1 , Xi , Yi 􏼁log P Yi+1 |Xi , Yi 􏼁
(10) 􏽥 t.
Ct � ft ⊙ Ct− 1 + it ⊙ C (15)

− 􏽘 P Yi+1 , Xi , Yi 􏼁log P Yi+1 |Yi 􏼁 From the past output, current input, and current unit
state, the output gate drives the current output:
� H Yi+1 |Yi 􏼁 − H Yi+1 |Xi , Yi 􏼁. ot � σ Wo 􏼂ht− 1 , xt 􏼃 + bo 􏼁, (16)

According to the definition of transfer entropy, log ht � ot ⊙ tanh Ct 􏼁, (17)

P(Yi + 1|Yi) is the occurrence probability of Yi+1 under the
condition of state Yi, which excludes the influence of the past where xt is the input of unit state; ft, it, and ot are the ac-
state of Y over its future information. In this way, TE can tivation values of the forget gate, the input gate, and the
accurately measure the information transferred from X to Y. output gate, respectively; C􏽥 t and Ct are the candidate and
Conditional mutual information considers the dependence output states, respectively; ht is the output of the node; Wf ,
Computational Intelligence and Neuroscience 5

ht−1 ht ht+1

Ct−1 Ct
× +
tanh
~ × Ot
A ft it Ct × A
σ σ tanh σ
ht−1 ht

xt−1 xt
xt+1

Figure 2: Structure of the LSTM.

Wi , Wc , and Wo are the weight matrices of the forget gate, 4. Case Study
the input gate, the unit state, and the output gate, respec-
tively; bf, bi, bc, and bo are the biases of the forget gate, the 4.1. Data Fusion. Data fusion intends to combine infor-
input gate, the unit state, and the output gate, respectively. mation in the best possible way to obtain more effective
The LSTM network uses the error backpropagation algo- information. This paper adopts the data fusion to combine
rithm to update the weights. The weight is divided into two meteorological data and hydrological data and evaluate the
parts, one part participates in the output of the previous drought level comprehensively. It is worth noting that the
neuron as Wfh , Wih , Wch , and Woh , and the other part data fusion generally falls into three levels: data layer fusion,
participates in the current input as Wfx , Wix , Wcx , and Wox . feature layer fusion, and decision layer fusion. The fusion by
L is the loss function, and the error at the current moment t copula function is the feature layer fusion because SPI and
is defined as SRI, two characteristic indices, are adopted, which reflect
meteorological and hydrological droughts, respectively.
def zL
δt � . (18)
zht
4.1.1. Selection of Copula Function. Taking the Tangnaihai
The input to the neuron is
Station in the upper reaches of the Yellow River for example,
the frequency histograms and 2D frequency diagrams of SPI
netf,t � Wf 􏼂ht− 1 , xt 􏼃 + bf � Wfh ht− 1 + Wfx xt + bf , (19) and SRI are displayed in Figure 3, and the cumulative
distribution and density function diagrams of the Gumbel
neti,t � Wi 􏼂ht− 1 , xt 􏼃 + bi � Wih ht− 1 + Wix xt + bi , (20) copula are displayed in Figure 4.
The SPI and SRI values of the Tangnaihai Station con-
net􏽥c ,t � Wc 􏼂ht− 1 , xt 􏼃 + bc � Wch ht− 1 + Wcx xt + bc , (21) centrated in [− 2, 2]. The SPI frequency distribution was
higher in the middle than the two sides and symmetric to a
neto,t � Wo 􏼂ht− 1 , xt 􏼃 + bo � Woh ht− 1 + Wox xt + bo , (22) certain extent. The SRI frequency distribution was very
asymmetric: the left side was taller than the right side. Thus,
zL the distribution feature can be illustrated well by the Gumbel
def
δf,t � , (23) copula, which has the shortest Euclidean distance. Using the
znetf,t
Gumbel copula, the authors established the copula function
of the hybrid drought index for the Tangnaihai Station:
def zL
δi,t � , (24) 0.6247
zneti,t CGu (u, v; θ) � exp􏼚− 􏽨(− lnu)1.6008 +(− lnv)1.6008 􏽩 􏼛, (28)

def zL CGu (u, v; θ)(lnu · lnv)0.6008

δ􏽥c,t � , (25) cGu (u, v; θ) �
znet􏽥c,t 1.3753
uv􏽨(− lnu)1.6008 +(− lnv)θ 􏽩
zL 0.6247
def
δo,t � . (26) · 􏼚􏽨(− lnu)1.6008 +(− lnv)1.6008 􏽩 + 0.6008􏼛.
zneto,t
(29)
The error at t-1 is
By the principle of Euclidean distance, the type of copula
zL zL zht zh function was selected for each station. Table 2 shows the
δt− 1 � � � δt t . (27)
zht− 1 zht zht− 1 zht− 1 selection results and the relevant parameters.
6 Computational Intelligence and Neuroscience

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4
f(x)

f(y)
0.3 0.3

0.2 0.2

0.1 0.1

0 0
−3 −2 −1 0 1 2 3 4 −2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
SPI SRI

(a) (b)

6
Frequency

0
1
1
0.5 0.8
0.6
0.4
0.2
0 0
SRI SPI

(c)

Figure 3: Histogram of SPI and SRI.

1 8

0.8
6

0.6
C(u,v)

c(u,v)

4
0.4

2
0.2

0 0
1 1
1 1
0.8 0.8
0.5 0.6 0.5 0.6
0.4 0.4
0.2 0.2
SRI 0 0 SRI 0 0
SPI SPI

(a) (b)

Figure 4: Cumulative distribution function and probability density function of the Gumbel copula.
Computational Intelligence and Neuroscience 7

Table 2: Copula functions for different stations.

Location Station Gaussian T Gumbel Frank Clayton Selected function and parameter θ
Tangnaihai 1.2443 1.2681 0.0490 0.1151 0.4732 Gumbel copula θ  1.6008
Upper reaches
Shizuishan 3.6693 3.6773 0.0468 0.0548 0.0812 Gumbel copula θ  1.0816
Longmen 2.8618 2.8965 0.1403 0.0736 0.0639 Clayton copula θ  0.4531
Middle reaches
Sanmenxia 2.3053 2.3308 0.0545 0.0603 0.0672 Gumbel copula θ  1.2932
Huayuankou 3.8682 3.8848 0.0619 0.0542 0.0617 Frank copula θ  0.6412
Lower reaches
Lijin 3.9394 3.9606 0.1140 0.0959 0.0878 Clayton copula θ  0.1540

Comparison of interannual drought indexes at Tangnaihai Station

1
Drought index

-1

-2

1980 1990 2000 2010 2020

Year
SRI–12 DCI–12
SPI–12 Drought threshold

Figure 5: Comparison of interannual drought indexes at the Tangnaihai Station.

4.1.2. Comparison of Drought Indices. Considering the sit- time of drought as the SRI. Therefore, the DCI combines the
uation of long-term drought, this paper computes the in- merits of the SPI and the SRI: The DCI can sensitively
terannual SPIs and SRIs and conducts the feature layer capture the start time of drought as the SPI, and effectively
fusion of meteorological and hydrological droughts by depict the drought duration. By fusing meteorological in-
copula function, producing a hybrid drought index coupling formation with hydrological information, the DCI can
long-term meteorological and hydrological information. characterize both meteorological and hydrological droughts
Taking the Tangnaihai Station for example, the interannual simultaneously. The advantages of the DCI include high
SPI, SRI, and DCI were compared (Figure 5). sensitivity, strong recognizability, and wide applicability.
The threshold of moderate drought was set to -1. This research further validated the theory of Kimaru et al.
Normally, drought begins when the drought index falls [21] and meteorological drought occurs and terminates very
below -1 and ends when the index rises above -1. The quickly, while hydrological drought begins and ends with a
drought occurrence directly depends on rainfall. If the certain delay in response to meteorological drought.
rainfall is consistently abnormal for a period, it will affect the
confluence between surface water and groundwater through
the natural water circulation and thus impact the hydro- 4.2. Selection of Assessment Factors. When selecting the
logical condition. The SPI responds sensitively to the mo- drought factors, the influence between variables is measured
ment of the drought occurrence, while the SRI can accurately by transfer entropy, and through this influence, a causal
determine the duration and end time of drought. As shown relationship between variables is established. Based on
in Figure 5, the DCI trend was overall similar to the trends of conditional mutual information, the transfer entropy is an
SPI and SRI. The three indices were consistent. The drought asymmetric measure [24]. Let TEX⟶Y be the information
start time recognized by the DCI was earlier than that transfer from X to Y; TEY⟶X be the information transfer
recognized by the SRI, and roughly the same as that iden- from Y to X. If TEX⟶Y > TEY⟶X, the influence of X
tified by the SPI. Besides, the DCI recognized the same end over Y is stronger than that of Y over X. Then, X would be
8 Computational Intelligence and Neuroscience

Table 3: Transfer entropies between drought influencing factors and the drought index.
Location Upper reaches Middle reaches Lower reaches
Transfer entropy TEX⟶Y TEY⟶X TEX⟶Y TEY⟶X TEX⟶Y TEY⟶X
Air temperature 0.1414 0.0740 0.1590 0.1022 0.1555 0.1290
Runoff 0.1304 0.1034 0.1394 0.1222 0.1410 0.0880
Rainfall 0.1252 0.1633 0.1717 0.2066 0.1215 0.1947
Humidity 0.1525 0.1059 0.1731 0.1537 0.1790 0.1083
Air pressure 0.1026 0.1049 0.1663 0.1724 0.1535 0.1542
Vapor pressure 0.1412 0.1611 0.1564 0.1638 0.1540 0.1647
Sunshine hours 0.1740 0.1348 0.1494 0.1353 0.1418 0.0838
Wind velocity 0.1360 0.0714 0.1564 0.0874 0.1925 0.1204

Table 4: Evaluation index of the LSTM drought assessment model.

Station
Evaluation index
Tangnaihai Shizuishan Longmen Sanmenxia Huayuankou Lijin
MSE 0.0015 0.0031 0.0028 0.0021 0.0019 0.0029
R2 0.9602 0.9738 0.9917 0.9626 0.9639 0.9934

regarded as a driver of Y. Then, two variables were defined as selected as the activation function. The Adam optimizer was
Y � {DCI12} and X � {Drought factors}. Table 3 shows the employed to optimize the model. Drawing on the features of
values of TEX⟶Y and TEY⟶X. momentum and root mean square propagation (RMSProp),
Table 3 shows that TEX⟶Y < TEY⟶X held for adaptive learning rates were designed from parameters like
rainfall, air pressure, and vapor pressure. Thus, these factors the first and second order moment estimations of gradients,
are largely constrained by drought and are not the causes of aiming to effectively update the weights.
drought. In the long run, drought is mainly affected by
constantly changing factors. The action of drought will affect
4.3.3. Model Evaluation. The model performance was
the natural water circulation, thereby constraining the
evaluated by MSE and R2 . The evaluation results are listed in
formation of clouds, rains, and fogs. As a result, meteoro-
Table 4.
logical factors like rainfall, and air pressure will be influ-
As shown in Figure 6 and Table 3, the LSTM network can
enced by drought. Rainfall, as an instantaneous factor, drive
accurately assess the drought risks. The MSE was always
short-term drought more significantly than long-term
smaller than 0.0032 and minimized at 0.0015. The R2 was
drought. Drought will be influenced only if the rainfall
always greater than 0.96, peaking at 0.9934. The interannual
increases or decreases continuously and breaks the balance
drought fluctuation was gentle, and the drought occurrence
of water circulation. Overall, this paper chooses the fol-
probability was low and highly predictable. Thus, the LSTM
lowing factors to evaluate drought: air temperature, runoff,
achieved a high prediction accuracy with a strong stability.
humidity, sunshine hours, and wind velocity.

4.3.4. Drought Assessment. Using the meteorological and

4.3. LSTM Drought Evaluation Model hydrological data of 1980–2019, the multivariate LSTM
4.3.1. Data Preprocessing. To speed up model training, min- network was called to assess the drought risks. As shown in
max normalization was chosen to preprocess the data: Figure 7, the upper reaches of the Yellow River would be
humid, the lower reaches would be normal, and the middle
x − xmin
x∗i � i , (30) reaches would be dry in 2020. The prediction agrees with the
xmax − xmin actual situation in 2020.
where xi and x∗i are the original data and the normalized Inspired by the working mode of the human brain, deep
data, respectively, and xmax and xmin are the maximum and learning analyzes and learns the data by an artificial neural
minimum of the data, respectively. network (ANN) with multiple deep layers, which mimics the
human brain. The long short-term memory (LSTM) net-
work, as a deep learning neural network, can extract abstract,
4.3.2. Model Training. To ensure the consistency between representative features from the original data. Capable of
the data distribution of input and output variables, the first sequential and directional circulation, the LSTM is very
90% of the original data were taken as the training set; the suitable for analyzing the correlated time series. Poornima
number of hidden layer nodes was set to 65; the batch size and Pushpalatha [25] found that the LSTM outperformed
was set to 88; the number of iterations was set to 2,600; the the autoregressive integrated moving average (ARIMA), and
mean squared error (MSE) was taken as the loss function. improved the prediction accuracy of the LSTM by adding
On this basis, an LSTM network was established. To prevent humidity and temperature to training, both of which are
vanishing gradients, a rectified linear unit (ReLU) was positively correlated with the drought index. Dikshit et al.
Computational Intelligence and Neuroscience 9

Huayuankon Station Tangnaihai Station

2.5

0.50 2.0

0.25 1.5

0.00 1.0

DCI–12
DCI–12

0.5
-0.25

0.0
-0.50

-0.5
-0.75
-1.0

-1.00

2016 2017 2018 2019 2020 2016 2017 2018 2019 2020

Year Year

REAI
REAI
LSTM
LSTM

(a) (b)
Shizuishan Station Sanmenxia Station

3
1.0

2 0.5
DCI–12

DCI–12

1 0.0

-0.5
0

-1.0
-1

2016 2017 2018 2019 2020 2016 2017 2018 2019 2020

Year Year

REAI REAI
LSTM LSTM

(c) (d)
Longmen Station Lijin Station
2.5

1.5
2.0
1.0
1.5
0.5
1.0
0.0
DCI–12

DCI–12

0.5
-0.5

0.0
-1.0

-0.5
-1.5

-1.0
-2.0

2016 2017 2018 2019 2020 2016 2017 2018 2019 2020

Year Year

REAI REAI
LSTM LSTM

(e) (f )

Figure 6: Comparison of assessment effects of the LSTM model.

10 Computational Intelligence and Neuroscience

100° E 105° E 110° E 115° E

40° N 40° N

35° N 35° N

High : 2.5

Miles
Low : -0.5
0 60 120 240 360 480

100° E 105° E 110° E 115° E

Figure 7: LSTM assessment of droughts for 2020.

[26] achieved a high prediction accuracy with the LSTM (2) The transfer entropy was adopted to analyze the
neural network. Wu et al. [27] described monthly rainfall causality between climate and hydrological factors
with wavelet transform, ARIMA, and LSTM, and demon- and the drought index, and judge the direction of
strated that the proposed W-A-L composite model can information flow. In this way, the air temperature,
predict rainfall accurately, providing a good reference for runoff, humidity, sunshine hours, and wind velocity
further research into drought prediction. were selected to assess the interannual drought in the
Yellow River Basin.
5. Conclusions (3) A deep learning model was established for drought
assessment. The assessment accuracy of the LSTM
The Yellow River Basin is frequently hit by drought, owing to network was proved by comparing the predicted
the complex natural conditions, unique climate features, and values with the actual values, and by computing the
special geographical conditions. The drought index and evaluation metrics.
evaluation are important issues in the understanding and
prevention of drought. Taking rainfall and runoff as me- Drought research is a hot field among researchers. More
teorological and hydrological factors, this paper explores the scientific studies are expected for the driving mechanism,
nonlinear relationship between meteorological and hydro- evolution, and evaluation of drought. In the era of the big
logical series by the copula function, and constructs a hybrid data, data fusion and deep learning could be combined to
drought index through the feature layer fusion. Then, the open a bright new arena of drought research.
transfer entropy was computed for drought factors and
indices, the causality was judged by the direction of infor- Data Availability
mation flow, and multiple factors were chosen for drought
assessment. Further, the data information was mined The data used to support the findings of this study are
through deep learning and used to build a multivariate available from the corresponding author upon request.
LSTM network for drought evaluation. The main conclu-
sions are as follows:
Conflicts of Interest
(1) Following the idea of copula function, the hybrid
drought index DCI combines the features of rainfall The authors declare that they have no conflicts of interest.
and runoff, and integrates the merits of SPI and SRI.
The index can detect the drought occurrence as the Acknowledgments
SPI, and capture the drought duration as the SRI,
providing an effective tool to depict the start, end, This research was supported by the Key Science and
and progression of drought. Technology Projects in Henan Province (212102310306) and
Computational Intelligence and Neuroscience 11

the National Natural Science Foundation of China [16] T. B. McKee, N. J. Doesken, and J. Kleist, “The relationship of
(51709221). drought frequency and duration to time scales,” Eighth
Conference on Applied Climatology, vol. 17, no. 22, pp. 179–
183, 1993.
References [17] R. S. Arnold, “Mapping meteorological drought hazard in the
Philippines using SPI and SPEI,” Spatial Information Re-
[1] J. Park, Y.-J. Lim, B.-J. Kim, and J. H. Sung, “Erratum to: search, vol. 29, pp. 949–960, 2021.
appraisal of drought characteristics of representative drought [18] Y. Y. Xiang, Y. Wang, Y. N. Chen, Y. Bai, L. Zhang, and
indices using meteorological variables,” KSCE Journal of Civil Q. Zhang, “Hydrological drought risk assessment using a
Engineering, vol. 22, no. 7, p. 2650, 2018. multidimensional copula function approach in arid inland
[2] E. Etienne, N. Devineni, R. Khanbilvardi, and U. Lall, “De- basins, China,” Water, vol. 12, no. 7, pp. 1888.1–1888, 2020.
velopment of a demand sensitive drought index and its ap- [19] A. Sklar, “Fonctions de repartition an dimensions et leurs
plication for agriculture over the conterminous United marges,” Publications De L’institut Statistique De L’universite
States,” Journal of Hydrology, vol. 534, pp. 219–229, 2016. De Paris, vol. 8, pp. 229–231, 1959.
[3] X. Fan, Y. Y. Qin, and X. Gao, “Interpretation of the main [20] H. Razmkhah, A. Fararouie, and A. R. Ravari, “Multivariate
conclusions and suggestions of IPCC AR6 working group I flood frequency analysis using bivariate copula functions,”
Report,” Environmental Protection, vol. 49, no. 17, pp. 44–48, Water Resources Management, vol. 36, no. 2, pp. 729–743,
2021. 2022.
[4] M. N. Sattar, J.-Y. Lee, J.-Y. Shin, and T.-W. Kim, “Proba- [21] A. N. Kimaru, J. M. Gathenya, and C. K. Cheruiyot, “The
bilistic characteristics of drought propagation from meteo- temporal variability of rainfall and streamflow into lake
rological to hydrological drought in South Korea,” Water Nakuru, Kenya, assessed using SWAT and hydrometeoro-
Resources Management, vol. 33, no. 7, pp. 2439–2452, 2019. logical indices,” Hydrology, vol. 6, no. 4, p. 88, 2019.
[5] A. Meteorological Society, “Meteorological drought policy [22] Y. Peng, F. D. Li, N. Xu et al., “Spatial-temporal variations in
statement,” Bulletin of the American Meteorological Society, drought conditions and their climatic oscillations in Central
vol. 78, pp. 847–849, 1997. Asia from 1990 to 2019,” Chinese Journal of Eco-Agriculture,
[6] C. Wu, S. Ning, J. Jin et al., “Construction and application of vol. 29, no. 2, pp. 312–324, 2021.
comprehensive drought index based on uncertainty cloud [23] S. Hochreiter and J. Schmidhuber, “Long short-term mem-
reasoning algorithm,” Science of the Total Environment, ory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.
vol. 779, Article ID 146533, 2021. [24] Z. M. Wang, L. L. Cai, and L. Fan, “Emotion recognition
[7] J. Won, J. Choi, O. Lee, and S. Kim, “Copula-based Joint method based on key causal connection of transfer entropy,”
Drought Index using SPI and EDDI and its application to Application Research of Computers, vol. 38, no. 9, pp. 2614–
climate change,” The Science of the Total Environment, 2618, 2021.
vol. 744, Article ID 140701, 2020. [25] S. Poornima and M. Pushpalatha, “Drought prediction based
[8] X. Y. Han, F. R. Du, Q. F. Li et al., “Application of stan- on SPI and SPEI with varying timescales using LSTM re-
dardized comprehensive drought index in drought evaluation current neural network,” Soft Computing, vol. 23, no. 18,
in Beiru River Basin,” Water Resources and Power, vol. 37, pp. 8399–8412, 2019.
no. 6, pp. 19–22, 2019. [26] A. Dikshit, B. Pradhan, and A. Huete, “An improved SPEI
[9] D. Rajsekhar, V. P. Singh, and A. K. Mishra, “Multivariate drought forecasting approach using the long short-term
drought index: an information theory based approach for memory neural network,” Journal of Environmental Man-
integrated drought assessment,” Journal of Hydrology, agement, vol. 283, Article ID 111979, 2021.
vol. 526, pp. 164–182, 2015. [27] X. Wu, J. Zhou, H. Yu et al., “The Development of a hybrid
[10] Y. Ren, Y. M. Wang, J. X. Chang, and Q. Huang, “The spatial wavelet-ARIMA-LSTM model for precipitation amounts and
and temporal distribution of drought in Shaanxi Province,” drought analysis,” Atmosphere, vol. 12, no. 1, p. 74, 2021.
Journal of Natural Resources, vol. 32, no. 1, pp. 137–151, 2017.
[11] S. Maji and S. Kanrar, “SpliceCombo: a hybrid technique
efficiently use for principal component analysis of splice site
prediction,” Ingénierie des Systèmes d’Information, vol. 24,
no. 1, pp. 67–75, 2019.
[12] F. Wang, Z. Wang, H. Yang, and Y. Zhao, “Study of the
temporal and spatial patterns of drought in the Yellow River
basin based on SPEI,” Science China Earth Sciences, vol. 61,
no. 8, pp. 1098–1111, 2018.
[13] S. Huang, P. Li, Q. Huang, G. Leng, B. Hou, and L. Ma, “The
propagation from meteorological to hydrological drought and
its potential influence factors,” Journal of Hydrology, vol. 547,
pp. 184–195, 2017.
[14] Y. Wang, W. X. Shang, and S. M. Peng, “Yellow River basin
drought response system based on the prediction and oper-
ation of cascade reservoirs,” Advances in Water Science,
vol. 30, no. 2, pp. 175–185, 2019.
[15] T. T. Huang, Q. X. Lin, Z. Y. Wu, and Y. Q. Wang, “Spatial
and temporal characteristics of drought in the Yellow River
basin and their correlation with ENSO,” Yellow River, vol. 43,
no. 11, pp. 52–58, 2021.