Heliyon 10 (2024) e22942

Contents lists available at ScienceDirect

Heliyon
journal homepage: www.cell.com/heliyon

Deep learning versus hybrid regularized extreme learning machine

for multi-month drought forecasting: A comparative study and
trend analysis in tropical region
Mohammed Majeed Hameed a, b, Siti Fatin Mohd Razali a, c, *,
Wan Hanna Melini Wan Mohtar a, c, Majed Omar Ahmad Alsaydalani d,
Zaher Mundher Yaseen e, f
a
Green Engineering and Net Zero Solution (GREENZ), Department of Civil Engineering, Faculty of Engineering and Built Environment, Universiti
Kebangsaan Malaysia (UKM), 43600 UKM, Bangi, Selangor, Malaysia
b
Department of Civil Engineering, Al-Maarif University College, 31001, Ramadi, Iraq
c
Smart and Sustainable Township Research Centre (SUTRA), Universiti Kebangsaan Malaysia (UKM), 43600, UKM, Bangi, Selangor, Malaysia
d
Department of Civil Engineering, Umm Al-Qura University, Makkah, Saudi Arabia
e
Civil and Environmental Engineering Department, King Fahd University of Petroleum & Minerals, 31261, Dhahran, Saudi Arabia
f
Interdisciplinary Research Centre for Membranes and Water Security, King Fahd University of Petroleum & Minerals, 31261, Dhahran, Saudi
Arabia

A R T I C L E I N F O A B S T R A C T

Keywords: Drought is a hazardous natural disaster that can negatively affect the environment, water re­
Drought forecasting sources, agriculture, and the economy. Precise drought forecasting and trend assessment are
Innovative trend analysis essential for water management to reduce the detrimental effects of drought. However, some
Deep learning
existing drought modeling techniques have limitations that hinder precise forecasting, necessi­
Climate smart agriculture
Drought turning points
tating the exploration of suitable approaches. This study examines two forecasting models, Long
Multivariate standardized streamflow index Short-Term Memory (LSTM) and a hybrid model integrating regularized extreme learning ma­
chine and Snake algorithm, to forecast hydrological droughts for one to six months in advance.
Using the Multivariate Standardized Streamflow Index (MSSI) computed from 58 years of
streamflow data for two drier Malaysian stations, the models forecast droughts and were
compared to classical models such as gradient boosting regression and K-nearest model for
validation purposes. The RELM-SO model outperformed other models for forecasting one month
ahead at station S1, with lower root mean square error (RMSE = 0.1453), mean absolute error
(MAE = 0.1164), and a higher Nash-Sutcliffe efficiency index (NSE = 0.9012) and Willmott index
(WI = 0.9966). Similarly, at station S2, the hybrid model had lower (RMSE = 0.1211 and MAE =
0.0909), and higher (NSE = 0.8941 and WI = 0.9960), indicating improved accuracy compared to
comparable models. Due to significant autocorrelation in the drought data, traditional statistical
metrics may be inadequate for selecting the optimal model. Therefore, this study introduced a
novel parameter to evaluate the model’s effectiveness in accurately capturing the turning points
in the data. Accordingly, the hybrid model significantly improved forecast accuracy from 19.32 %
to 21.52 % when compared with LSTM. Besides, the reliability analysis showed that the hybrid
model was the most accurate for providing long-term forecasts. Additionally, innovative trend

* Corresponding author. Department of Civil Engineering, Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia (UKM),
43600, UKM, Bangi, Selangor, Malaysia.
E-mail address: [email protected] (S.F. Mohd Razali).

https://doi.org/10.1016/j.heliyon.2023.e22942
Received 4 September 2023; Received in revised form 19 November 2023; Accepted 22 November 2023
Available online 28 November 2023
2405-8440/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
M.M. Hameed et al. Heliyon 10 (2024) e22942

analysis, an effective method, was used to analyze hydrological drought trends. The study
revealed that October, November, and December experienced higher occurrences of drought than
other months. This research advances accurate drought forecasting and trend assessment,
providing valuable insights for water management and decision-making in drought-prone
regions.

1. Introduction

Drought, a devastating natural calamity, is defined by an extended period of below-normal precipitation. It ranks among the most
severe worldwide catastrophes due to its far-reaching consequences and substantial economic losses. Droughts are responsible for 22 %
of economic damages resulting from disasters on a global scale, and they account for 33 % of the harm concerning the people impacted
[1,2]. Typically, drought is divided into four primary forms: meteorological, hydrological, agricultural, and socio-economic drought.
Drought is considered a gradual process that begins with a shortfall in precipitation, known as meteorological drought. An extended
meteorological drought may cause a hydrological drought characterized by reduced dam reservoir capacity, decreased river flow, and
lower lake water levels. Furthermore, extended droughts can trigger crop-threatening water shortages, hindering growth and
development due to inadequate water supply.
Various drought indices developed since the 1960s based on meteorological and hydrological variables to identify and quantify
drought. The “standardized precipitation index” (SPI) [3] is widely used for drought quantification, while the surface water supply
index (SWSI) [4], standardized runoff index (SRI) [5], and standardized streamflow index (SSI) [6,7] are more famous indices used for
describing the hydrological drought. The SSI is more popular than the SWSI index because it’s easier to use and requires only
streamflow data, whereas determining the variable weights of SWSI is subjective [8] and demands lots of input data (e.g., temperature,
streamflow, rainfall, reservoir storage levels, and snow). Moreover, the ability of SSI to provide precise assessments of hydrological
droughts is similar to the evaluations carried out by SWSI [9].
Hydrological drought forecasting can aid decision-making, early warning and mitigation strategies, but it is challenging due to
complexity and creeping nature (slow progression). Drought forecasting involves selecting appropriate models, including physical,
conceptual, and artificial intelligence (AI)-based models’ models. While physical and conceptual models are data-intensive and
consider the process of a studied catchment [10], AI models are cost-effective models and require minimal data. Various models for
drought forecasting have been used in literature, including artificial neural networks [11,12], support vector regression [13,14], and
assembling models (e.g., random forest, gradient boost regression) [15,16]. Furthermore, some scholars have turned to the extreme
learning machine (ELM) as an upgraded version of the traditional neural network. While conventional neural networks have some
drawbacks, such as slow performance and poor generalization, the ELM has shown promising results in achieving good forecasting
accuracy, particularly in forecasting droughts [17–19].
Developing models to obtain highly accurate predictions continues to be a primary focus of researchers. It is important to note that
researchers are pursuing two main trends in using advanced AI models for drought forecasting. The first trend involves utilizing
Metaheuristic algorithms (e.g., whale optimization algorithm, butterfly optimization, ant colony algorithm, imperialistic competitive
algorithm, firefly optimization algorithm, genetic algorithm, and others) to enhance the performance of classical models and create
hybrid models that eventually offer superior forecasting accuracy [10,20–29]. These models have been shown to be highly effective in
forecasting droughts compared to classical models. Another direction that researchers have taken is the broad application of advanced
deep learning algorithm (e.g., long short-term memory (LSTM)), which can process large amounts of data and identify complex
patterns [30–37]. This approach has proven to be highly effective in generating accurate drought predictions. Deep learning and
hybrid models are promising for drought forecasting, but more evaluation and comparison are needed to understand their capabilities
and limitations.
Understanding drought trends is crucial for developing effective strategies to mitigate the negative impacts of drought. This
knowledge is useful in water resource management, crop production, disaster risk reduction, and long-term adaptation planning. There
are various methods to analyze drought patterns and trends, including non-parametric and parametric methods like the “Mann-Kendall
test, Sen’s slope, Spearman’s rho test, and linear regression” [38–42]. However, these methods have restrictive assumptions [43], and
the presence of positive serial correlation in time series data can increase the probability of identifying a trend even when there is none
[44]. To address this issue, the pre-whitening technique was proposed, but it can remove a significant portion of data trend [45,46]. To
solve this problem, the innovative trend analysis (ITA) method was proposed by Ref. [47], which does not impose any restrictive
assumptions and has shown greater effectiveness than traditional methods in identifying drought trends (e.g., such as serial correlation
or seasonal cycles) [48]. ITA has been successfully used in diverse climatic region in identifying drought trends [49–53].
The primary objective of this research is to enhance the understanding of drought trend dynamics and provide reliable forecasts for
droughts several months in advance. This knowledge is crucial for improved water management planning and for establishing effi­
cacious advanced early warning systems and drought policy preparations. Previous research efforts have primarily focused on either
forecasting drought using predictive models or analyzing drought patterns and trends. However, they did not integrate both methods
concurrently, creating a gap in the comprehensive understanding of drought dynamics. Therefore, this research aims to bridge that
divide by combining the analysis of drought patterns and trends with predictive models. The study includes a comprehensive analysis
of popular and robust forecasting models in the field of drought, such as DL and hybrid models. This aspect of the research enhances the
depth of the investigation, mainly because previous studies have not explored such a comparison between the two forecasting models.

2
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 1. The illustrates the geographic locations of the streamflow stations investigated in this study.

The Multivariate Standardized Streamflow Index (MSSI) is adopted, considering all drought time scales in contrast to indices like SDI
and SSI [54]. Thus, the performances of a hybrid model that combines Regularized Extreme Learning Machine (RELM) with a novel
snake optimizer (RELM-SO) and a Deep Learning Model (LSTM) are investigated. The models forecast drought for one to six months in
advance at two specifically dry locations in Selangor State, Malaysia [55]. Also, the models are validated against benchmark stand­
alone models, such as k-nearest neighbor and gradient boosting regression, as well as other robust models developed in the literature.
Typically, the traditional statistical metrics may be inadequate for evaluating the effectiveness of drought forecasting models due to
the high autocorrelation in time series data [56–58], and the issue remains a challenge not addressed in previous studies [11,59].
Consequently, this study validates models based on their ability to accurately capture critical and turning points (i.e., critical points
indicate significant changes/events, while turning points signal trend direction shifts in drought time series data), aiming to identify
the most effective model for reliable drought forecasting. Finally, this research also comprehensively analyses drought trends for all
months of the year. The ITA method is utilized to identify the most vulnerable months for severe drought, as it is an effective tool that
does not require numerous variables and assumptions.

2. Statistical methods and case study

2.1. Multivariate standardized streamflow index (MSSI)

The calculation of MSSI involves multiple SSI time series for different aggregation time scales (K) at specific stations. While
studying various time scales in classical indices such as SSI can provide information on different types of droughts, it may also cause
confusion due to its complexity [54]. Therefore, to derive MSSI, it is suggested to reduce the number of SSI time scales and summarize
the variability using principal components analysis (PCA). MSSI is derived PCA to a set of K time series of SSI, with K representing the
drought time scale of SSI. PC1, the first principal component of the PCA analysis, holds essential information on the percentage of
variation in K for the initial variable. As a result of the characteristics of PC1, its values cannot be compared between months, unlike
SSI, which has a specific statistical characteristic (mean is 0 and a standard deviation is 1). Hence, to standardize the PC1 time series, it
is essential to use the average and standard deviation of different months throughout the year [54,60]. Notably, Equation (1) provides

3
M.M. Hameed et al. Heliyon 10 (2024) e22942

Table 1
General information and statistical description of the studied streamflow stations.
Description Streamflow stations

S1 S2

Longitude 101 31″ 24′

◦
101◦ 26″ 35′
Latitude 03◦ 40″ 67′ 03◦ 24″ 10′
District Hulu Selangor Kuala Selangor
River Basin Bernam Selangor
Data duration 1-1-1961 to 31-1-2018 1-1-1961 to 31-1-2018
Maximum value 204.7320 m3/s 29.7832 m3/s
Average value 56.6064 m3/s 8.9139 m3/s
Minimum value 4.4790 m3/s 1.2070 m3/s
Standard deviation 30.1697 m3/s 4.8189 m3/s
Skewness 1.1593 1.1375
Station Code S3615412 S3414421

Fig. 2. Temporal variation of MSSI for both hydrological stations. a) S1 station, and b) S2 station.

the mathematical formula that used for standardizing the drought data.

PC1ym − PC1m
Z1ym = (1)
SD1m

In the above formula, the Zlym refers to the standardized value of PC1 in year y and month m, The term PC1ym is the PC1 value in yth year
an mth month, PC1m , SD1m and are the mean and standard deviation of the PC1 of the mth month. Notably, Z1ym is considered as the
MSSI. Since the PC1m value is statistically insignificant and almost zero, it can be ignored in the equation’s numerator. The MSSI time
series as represented in Z1ym is organized in ascending order, and a graphic of its empirical probability distribution is shown to identify
the categories of drought and wet severity (Appendix A).

2.2. Case study location

The daily streamflow records utilized in this research were collected from two stations located in Selangor, Malaysia, spanning an
extensive period from 1961 to 2018. The first station (S1) is located in Bernam River, which covers an area of 3335 km2 and is a crucial

4
M.M. Hameed et al. Heliyon 10 (2024) e22942

region for agriculture and water supply for Selangor and Perak, mainly for irrigation [61]. Besides, the river basin has relatively high
temperatures, humidity, and an average annual precipitation varies between 2000 and 3500 mm. Recently, urbanization resulting
from changes in the Malaysian government’s economic policies has influenced water quality and quantity in the region. The second
station (S2), is located in the Selangor River and has a catchment area of 2200 km2 [62]. It is a primary water source for several
regional water treatment plants. The Selangor River basin has diverse natural landscapes, including forests, agricultural areas, and
developed regions. The geographical locations of these studied streamflow stations are illustrated in Fig. 1.
The geographical coordinates and statistical characteristics of both study stations are summarized in Table 1. Station S1 has higher
maximum, average, and minimum flow rates compared to station S2, with values of 204.7320 m3/s, 56.6064 m3/s, and 4.4790 m3/s,
respectively, compared to 29.7832 m3/s, 8.9139 m3/s, and 1.2070 m3/s at station S2. However, station S1 also exhibits much higher
variability in flow rate measurements than station S2, with a standard deviation of 30.1697 m3/s compared to 4.8189 m3/s at station
S2.
The MSSI was calculated using PCA, which integrated SSI data collected at various time intervals, ranging from SSI-1 to SSI-48.
According to the hydrological drought depicted in Fig. 2 (a, and b), both selected stations experienced severe drought events be­
tween 1978 and 2018. Station S2 faced more than five such events, while S1 experienced only one during the same time period. It is
worth highlighting that the S1 station experienced severe drought conditions for a longer period between 2010 and 2018. Overall, the
presented result of hydrological drought provides compelling evidence that droughts are an issue in the region, with both stations
being significantly impacted.

2.3. Trend analysis: Sen-innovative trend method

The innovative trend analysis method (ITAM) is a novel approach used to detect the trend of time-series data proposed by Şen [47].
The core concept of this approach involves dividing a time series into two equal parts spanning from the first date to the end date of the
studied data. The two divided segments are sorted in ascending order and exhibited on the X and Y-axis, respectively. In the Cartesian
coordinate system, both divided segments are sorted in ascending order and plotted on the X and Y-axis. The first segment (Xi:i = 1,2,
..., n/2) is plotted on the horizontal axis, while the second segment (Yj:j = n/2 + 1, n/2 + 2, ..., n) is plotted on the vertical axis.
Appendix B illustrates how a line bisects at 1:1 (45◦ ) separates the diagram into two triangles of equal dimensions. According to
Appendix B, an ascending trend is represented by the upper triangle, whereas a descending trend is depicted by the lower one. To
estimate the trend, the slope of the Innovative trend analysis method called SITAM statistic is calculated using Equation (2).

2(Xi − Yi )
SITAM = (2)
n

where, Xi , and Yi are the average value of the first and second half of the time-series data and n is the length of the data.
To determine if there is a statistically significant trend in a time series, the null hypothesis (Ho) is rejected if the calculated slope
value (SITAM ) does not exceed a critical value (Ssci ). On the other hand, if SITAM > Ssci , the alternative hypothesis (Ho) indicating the
presence of a significant trend is applicable. The probability density function (PDF) of the null hypothesis is used to calculate the test
significance, and the confidence limits (CL) of the trend slope can be derived by setting Ssci as the confidence interval of a standard
normal PDF with a mean of zero and a standard deviation of σs at a significance level of α. Thus, the CL can be computed using Equation
(3)
CL(1− a) = 0 ∓ Ssci × σs (3)

In the above equation, the σ s is slop standard deviation. Notably, the test details can be found in Refs. [47,63]. The ITAM was used in
this study to assess the importance of monthly hydrological drought changes. A 5 % significance level was applied, considering the
infrequency of significant changes in drought.

3. Applied artificial intelligence models

3.1. Long short-term memory (LSTM)

LSTM is one of the most popular applications of deep learning neural networks, which has been used widely in literature to model
time series data [64]. It is considered an improved version of recurrent neural network (RNN). Capturing long-term time correlations
can be problematic with a traditional RNN, which commonly encounters gradient disappearance (i.e., vanishing gradient) and ex­
plosion problems [65]. LSTM is a sort of time-cyclic neural network which implemented partially to address the long-term correlation
problem of RNN [66,67]. Besides, it has good architecture enables the forecasting model to learn from training data across numerous
time steps. Furthermore, the memory cell within the hidden layer neural nodes of the cyclic neural network can be utilized to anchor
and retain past information. The model can effectively use the historical data by integrating three-active gate structures (i.e., input,
forget, and output gates).
Based on Appendix C, the sequence input of drought time series data x (x1, x2, x3, …, xt) when inserted into the LSTM model, the
hidden layer states (h1, h2, h3, …, ht) and the output of memory unit can be calculated as using Equations (4)–(8) [68]:
it = sigmoid(Whi ht− 1 + Wxi xt ) (4)

5
M.M. Hameed et al. Heliyon 10 (2024) e22942

( )
ft = sigmoid Whf ht− 1 + Wxf xt (5)

ct = ft ⊕ ct− 1 +ic ⊕ tanh(Wxc xt + Whc ht− 1 ) (6)

ot = sigmoid(Wxo xt + Who xt− 1 + Wco ct ) (7)

ht = ot ⊕ tanh(ct ) (8)

In the above equations, the symbol ⊕ is Hadamard product; ct is the cell state’s vector; ot , it , and ft are representing the output of
different gates; ht− 1 is the output data of the hidden layer unit regarding the previous iteration, while the new state of the memory cell
is ht ; Wc , Wf , and Wh are the corresponding gate’s weights, and the tanh, and sigmoid are activation functions.

3.2. K-nearest neighbor (KNN)

KNN, a supervised machine learning approach, can be employed for conducting several tasks such as classification, clustering, and
regression. Its algorithm is not complex and requires no assumptions on the primary distribution of the data. The KNN is a
nonparametric model, so its concept depends mainly on similarity (neighborhood) measures. To forecast the query point of MSSI, KNN
first starts to gather data observations close to that point. Notably, the neighbors of a query point can be calculated by several distance
functions (e.g., Euclidean approach). Once nearby data points have been identified, the algorithm sorts them based on their distance
from the new data point (query point). The next step involves selecting a specific number of data points with the shortest distances and
assigning their responses. The data points are chosen according to the value of k (number of neighbors) that the user defined. Finally,
assemble the responses of the closed k-points to forecast the response of the query point. For more details on the algorithm, please see
the following reference [69].

3.3. Gradient boosting regression (GBR)

The GBR technique can be defined as a machine learning ensemble method which boosts the predictive accuracy of a conventional
decision tree model. It is achieved by integrating boosting, a sequential statistical process that combines multiple weak predictive
models to produce a single, highly precise model [70]. The method employs an iterative approach, refining the weak learner’s tree
model estimates by incorporating the pseudo residuals of the current learner, disregarding the negative gradient of the loss function
[71]. The iteration process continues until the GBR model’s loss function is minimized at the lowest level and hence, the prediction
accuracy improves. The iterative learning procedure of GBR with K decision tree is explained as follows:
For a given training data D = {(x1 , y1 ), (x2 , y2 ), ..., (xn− 1 , yn− 1 ), ( xn , yn )}, the loss function can be derived using Equation (9)

L(y, f (x)) = (y − f (x))2 (9)

in this context, L denotes the loss function, which relies on the actual value y, while x represents the input values.
Step 1. Initialize the new tree model (weak learner) with an initial constant value, as presented in Equation (10)
∑
N
fo (x) = arg minc L(yi ,c) (10)
i=1

in Eq. (10), N represents the total number of data points, and the parameter c is a constant value.
Step 2. set the iteration number m = 1,2,3, …, K

i. For i = 1,2, 3, …, N. The pseudo residuals of the ith training data are determined using Equation (11).
[ ]
∂L(y, f (xi ))
rm,i = − (11)
∂f (xi ) f (x)=fm− 1 (x)

rm,i is refers to the pseudo-residual associated with the ith training data point at the mth iteration of the model.

ii. Build a regression tree based on rl,i and determine the size of the leaf node, Rm,l , for the mth tree. To obtain an approximate value
of the fitting residual, conduct a prediction for the area of the leaf nodes in the decision tree.
iii. Perform linear search for each value of l ranging from 1 to L within the leaf node range and minimize the loss function using
gradient descent. The best fitting residual fitting values for each blade, resulting from the analysis, are presented in Equation
(12).
∑
N
cm = arg minc L(yi , fm− 1 (xi ) + c) (12)
i=1

6
M.M. Hameed et al. Heliyon 10 (2024) e22942

Where, cm is the best residual fitting value for mth tree [70]

iv. Update the regression tress Using Equation (13).

∑
L
fm (x) = fm− 1 (x) + cml I(x ∈ Rml ) (13)
l=1

Step 3. Obtain the final model using Equation (14).

M ∑
∑ L
f (x) = fM (x) = cml I(x ∈ Rml ) (14)
m=1 l=1

3.4. Regularized extreme learning machine (RELM)

ELM is one of the most famous types of Feed-Forward Networks with a Single Hidden Layer introduced by Ref. [72]. According to
various sources, ELM has a straightforward structure, quick training time, strong generalization capabilities, and effective performance
[73–75]. The principles of ELM training and prediction are detailed in Refs. [76,77]. In training ELM, the primary goal is to determine
the weight vector (β) that connects the hidden layer and output layer, which is accomplished through the calculation of β. The
computed weight vector (β) is given by Equation (15) below.
̂
β=H † T (15)

In the above formula, the Moore-Penrose generalized inverse of the hidden layer output matrix H is denoted by H† . Notably, the term
H† is computed using the expression (HT × H)− 1 × HT , where T represents the actual data vector.
Once the structural parameters have been determined, the trained model can forecast the output values. ELM’s main flaw, however,
is that it is unstable and prone to overfitting and interference from data noise and outlier values [78,79]. To address this issue, some
researchers [80] presented a regularization parameter while determining the weight β, significantly improving the ELM model’s
generalization capability and practicality. As per statistical learning theory, the risk associated with prediction during the learning
process is a combination of structural and empirical risks. A model with solid generalization capabilities should efficiently balance
these two risks excellently. Consequently, the overall risk is represented as a weighted sum of the mentioned risks, where the ratio of
empirical risks can be adjusted by introducing a weighting parameter γ [80]. Empirical risk can be computed using the squared error
‖ε‖2 , while the variables ‖β‖2 that maximize the distance to the interface are used to represent the structural risk [81]. The mathe­
matical formulation of the regularized extreme learning machine (RELM) model is presented in Equation (16):
⎧ ( )
⎪ 1 2 1 2
⎪
⎪ min ‖β‖ + γ‖ ε ‖
⎪
⎪ 2 2
⎪
⎪
⎨
∑N
( ) (16)
⎪
⎪ s.t. βi g wi xj +bi − Tj
⎪
⎪
⎪
⎪ i=1
⎪
⎩
̃
j= 1, 2, 3, ...,N

In the above formula, the N is the total number of neurons in the models’ hidden layer, while the N ̃ is the number of observations.
Moreover, the terms j, w, b and x , T are the j th observation, wights, bias of hidden layer, the input and output data, respectively. The
optimization problem presented in the equation above can be solved by creating the Lagrangian equation as illustrated in Equation
(17).
⎧ ⃦ ⃦ ( )
⎪1⃦
⎪ 2 1 ⃦ 2
∑Ñ ∑N
( )
⎪
⎪ ⃦ ⃦β‖ + γ ⃦ε ‖ − a β g w x +b − t − εj
2 ⃦
⎨ j i i j i j
2
(17)
j=1 i=1
L(β, ε, a) =
⎪
⎪ ⃦ ⃦
⎪
⎪
⎩ 1⃦⃦ 2 1 ⃦⃦ 2
= ⃦β‖ + γ⃦ε‖ − a(Hβ − T − ε)
2 2
The aj s the Lagrangian multiplier, which is presented in the above equation. The optimal condition for Karush-Kuhn-Tucker (KKT),
which applies to L(β, ε, a), can be expressed in Equation (18).

7
M.M. Hameed et al. Heliyon 10 (2024) e22942

⎧
⎪ ∂L
⎪
⎪
⎪ = 0 ⇒βT = aH
⎪
⎪ ∂β
⎪
⎨
∂L (18)
⎪ = 0 ⇒γεT +a= 0
⎪
⎪ ∂ε
⎪
⎪
⎪
⎩ ∂L = 0 ⇒Hβ − T − ε= 0
⎪
∂a
By combining the equations provided above, the final expression for β can be calculated using Equation (19).
( )− 1
I
̂
β = H T H+ HT T (19)
γ

When using the equation listed above to determine the weight vector ̂
β between the hidden layer and output layer, the resulting ELM is
known as a RELM. By adjusting the value of γ, which represents the trade-off between empirical and structural risks, the model can
achieve an optimal balance between these two risks [80]. Compared to the traditional ELM, the RELM model exhibits superior
anti-interference ability, more robust neural network generalization, and higher prediction accuracy [82].

3.5. Snake optimizer (SO)

SO is a metaheuristic and sophisticated algorithm suggested by Ref. [83] and imitates the mating actions of snakes. The algorithm is
triggered by conditions similar to those that activate the snake mating, which occurs when there is enough food, and the temperature is
low. Like other metaheuristic methods, the SO algorithm generates random candidate solutions. Moreover, the swarms are divided
equally into male and female groups. To determine the optimal candidate solution during each iteration, the SO algorithm analyzes
each group to identify the best male and female individuals. The parameters for Temperature (T) and Food Quantity (FQ) can be
explained in Equation (20, and 21) [84].
( g)
T = exp − (20)
T
( )
g− T
FQ = C1 exp (21)
T

where the term g refers to the current iteration, while T is the total iterations, and C1 is a constant value (C1 = 0.5). Thus, if FQ less than
C1 , the snakes employ a food search approach in which they randomly select a position and subsequently update their current location.
The mathematical model for the exploration step for both male and female snakes is illustrated below:

a) Male snakes: The exploration step for male is expressed using Equation (22).
( )
xi,j (g+1) = x(rand∈[1,N/2] ,j) (g) ∓ C2 × Ai,male (ub − lb)×rand∈u(0,1) + lb (22)

( )
frand,male
where, Ai,male = exp − fi,mail , and C2 is a constant
In the above formula, N is the number of individuals, a male snake’s location is xi,j while the x(rand∈[1,N/2] ,j) is the location of random
male snake, and rand is a random number between zero and one. The fitness function of the previously nominated male snake for
random search can be represented by frand,male , while fi,mail is the fitness function of ith male in the group. Finally, the upper and lower
bounds are ub,and lb, respectively. The diversity factor, represented by the flag direction operator ±, is utilized to randomly explore all
potential directions in the provided search space model.

b) Female snakes: The exploration step for Female is expressed using Equation (23).
( )
xi,j = x(rand∈[1,N/2] ,j) (g+1) ∓ C2 × Ai,female (ub − lb)×rand∈u(0,1) + lb (23)

( )
frand,female
where, Ai,female = exp − fi,female .
In the exploitation phase, the SO algorithm utilizes two conditions to determine the optimal solutions.

➢ If FQ > Threshold (i.e., T is more than 0.6), the snakes move to search for food only. Thus, the exploitation phase can be represented
using Equation (24)
( )
xi,j (g+1) = xfood ∓ C3 × T × rand × xfood − xi,j (g) (24)

The variable xi,j represents the location or status of individuals, regardless of gender (male or female)., C3 is a constant, and xfood is
the position of the best individuals

8
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 3. Block diagram the main processes that used for establishing the forecasting models.

➢ If FQ is less than the specified Threshold value (less than 0.6), The snakes will exhibit two distinct behaviors, which are either
fighting or mating.

These behaviors can be illustrated using the following models for fighting and mating.

i. Fighting mode: the fighting capability of the male agent (F male ) is expressed in Equation (25) below:
( )
xi,j (g+1) = xi,j (g) + C3 × Fi,male ×rand × xbest,femail − xi,male (g) (25)
( )
f
Where, Fi,male = exp − best,female
fi .
The term xi,j is the ith male position, xbest,femail is the position of the best individual in the female group. Likewise, the male agent’s
fighting capability, Fi,male , can be expressed in Equation (26).
( )
xi,j (g+1) = xi,j (g) + C3 × Fi,female ×rand × xbest,male − xi,female (g+1) (26)
( )
fbest,male
where, Fi,female = exp − fi .

ii. Mating mode: during the mating mode, both male and female agents possess the capability to modify their positions according to
the following criteria (see Equation (27, and 28)):
( )
xi,male (g+1) = xi,m (g) ∓ C3 × MMi,male × FQ × xi,female − xi,male (g) (27)
( )
fi,female
where, MMi,male = exp − fi ,male .
( )
xi,female (g+1) = xi,f (g) ∓ C3 × MMi,female × FQ × xi,male − xi,female (g+1) (28)
( )
fi,male
where, MMi,female = exp − fi ,female .
In the above equation, the xi,f , and xi,m are the positions of ith females and males, respectively. The MMi,female and MMi,male are the
mating ability of females and males.

9
M.M. Hameed et al. Heliyon 10 (2024) e22942

Table 2
The candidate parameters of the applied models.
Models Hyperparameters range

KNN K = [2–200]
GBR Learning rate = [0.0856–0.9670]
Number of trees = [3–56]
LSTM Number of Hidden Nodes = [80− 260]
Max-Epochs = [100− 500]
Gradient Threshold = [0.001–0.952]
Initial Learn Rate = [0.0001–0.01]
Learn Rate Drop Period = [100− 300]
Learn Rate Drop Factor = [0.3–1]
RELM-SO Hidden nodes = [2–15]
Maximum Iteration = [50− 650]
Search Agents = [8–120]
Regularized parameter = [1/N, N]. Where N is the total number of training observations.

3.6. Developing the forecasting models

In the current study, the applied models included KNN, GBR, and LSTM, along with a fourth hybrid model called (RELM-SO) that
combined RELM and SO algorithms. The SO algorithm was used to optimize the RELM’s parameters such as wight and bias values as
well as regularization factor. Furthermore, this work uses root mean square error (RMSE) as the objective function, with the aim of
minimizing it using the SO algorithm. The study’s training methodology involved selecting optimal input lags via the partial auto­
correlation function (Appendix D), splitting data into training and testing sets (75 % for training, and 25 % of testing), normalizing the
data, initializing model parameters, and validating the models using several statistical metrics. Fig. 3 depicts the main process
employed to develop the forecasting models, as illustrated in the accompanying flowchart. Notably, the candidate hyperparameters of
each model are provided in Table 2.

3.7. Statistical performance metrics

The study used the root mean square error (RMSE), Nash-Sutcliffe efficiency index (NSE), Mean absolute error (MAE), Willmott
index (WI), and Correlation of determination coefficient (R2) as the five most commonly used statistical parameters to evaluate the
effectiveness of the developed MSSI forecasting models. These parameters were expressed mathematically in the study using Equations
(29)–(32) [85,86].

i. Root mean square error (RMSE)

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n
( )2
RMSE= MSSIobsi − MSSIpredi (29)
n i=1

ii. Nash-Sutcliffe efficiency index (NSE)

n ⃒
∑ ⃒
⃒MSSIobs − MSSIpred ⃒
i i

NSE= 1− i=1
∑n (30)
|MSSIobsi − MSSIobs |
i=1

iii. Correlation of determination (R2)

n (
∑ )2
MSSIobsi − MSSIpredi
2
R = 1− i=1
∑n ( )2 (31)
MSSIpredi − MSSIpred
i=1

iv. Willmott index (WI)

n (
∑ )2
MSSIobsi − MSSIpredi
WI= 1− ∑
n (⃒ ⃒
i=1
) (32)
⃒MSSIpred − MSSIobs ⃒+|MSSIobs − MSSIobs | 2
i i
i=1

MSSIobsi , and MSSIpredi are observed and forecasted values, while n is the dataset’s total number of drought values. The mean value

10
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 4. MSSI trends detected by innovative trend analysis method during January to December at a station S1. (a) January, (b) February, (c) March,
(d) April (e) May, (f) June, (g) July, (h) August, (i) September, (j) October,(k) November, and (l) December.

for forecasted and calculated droughts are MSSIpred , and MSSIobs . For the model to be considered optimal, it should produce error
measures (RMSE, and MAE) that are as low as possible (approaching zero) and anticipate outcomes that closely match the measured
values. Specifically, the values of R2, WI, and NSE should be closer to one, indicating a greater degree of consistency between the
predicted and actual values.

3.8. Identification and analysis of turning points in drought time series data

In time series data, turning points are points where the trend of the data changes direction, while critical points are points where a

11
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 5. MSSI trends detected by innovative trend analysis method during January to December at a station S2. (a) January, (b) February, (c) March,
(d) April (e) May, (f) June, (g) July, (h) August, (i) September, (j) October,(k) November, and (l) December.

significant change occurs, such as a sudden spike or drop in the values. Both types of points can be identified using various statistical
techniques such as time series regression analysis. Turning points indicate changes in the underlying pattern of the data and can help
identify shifts in the behavior of the phenomenon being studied. Critical points, on the other hand, can indicate the occurrence of an
event or a change in the underlying process generating the data. These points may offer valuable insights into the dynamics of the
system under study and can facilitate informed decision-making across a range of fields, including water resources planning and
management.

12
M.M. Hameed et al. Heliyon 10 (2024) e22942

Table 3
The results of innovative trend test of MSSI for S1 station.
Month Lower High

Slope Up* Lo** Slope Up* Lo**

Jan 0.1235 0.0058 − 0.0058 − 0.0373 0.0017 − 0.0017

Feb 0.1225 0.0053 − 0.0053 − 0.0390 0.0016 − 0.0016
Mar 0.1175 0.0048 − 0.0048 − 0.0402 0.0016 − 0.0016
Apr 0.1135 0.0040 − 0.0040 − 0.0410 0.0019 − 0.0019
May 0.1162 0.0038 − 0.0038 − 0.0396 0.0017 − 0.0017
Jun 0.1172 0.0032 − 0.0032 − 0.0400 0.0016 − 0.0016
Jul 0.1169 0.0025 − 0.0025 − 0.0419 0.0011 − 0.0011
Aug 0.1179 0.0031 − 0.0031 − 0.0429 0.0011 − 0.0011
Sep 0.1238 0.0024 − 0.0024 − 0.0415 0.0011 − 0.0011
Oct 0.1293 0.0046 − 0.0046 − 0.0384 0.0015 − 0.0015
Nov 0.1369 0.0032 − 0.0032 − 0.0367 0.0018 − 0.0018
Dec 0.1437 0.0055 − 0.0055 − 0.0368 0.0018 − 0.0018

The symbols *, and ** are the upper and lower critical bounds of the trend stope at 5%significance level.

Table 4
The results of innovative trend test of MSSI for S2 station.
Lower High

Month Slope Up* Lo** Slope Up* Lo*

Jan 0.0339 0.0047 − 0.0047 − 0.0027 0.0028 − 0.0028
Feb 0.0322 0.0055 − 0.0055 − 0.0057 0.0057 − 0.0057
Mar 0.0330 0.0053 − 0.0053 − 0.0163 0.0072 − 0.0072
Apr 0.0336 0.0037 − 0.0037 − 0.0181 0.0061 − 0.0061
May 0.0371 0.0037 − 0.0037 − 0.0138 0.0052 − 0.0052
Jun 0.0381 0.0025 − 0.0025 − 0.0125 0.0045 − 0.0045
Jul 0.0372 0.0024 − 0.0024 − 0.0171 0.0032 − 0.0032
Aug 0.0357 0.0022 − 0.0022 − 0.0250 0.0024 − 0.0024
Sep 0.0370 0.0015 − 0.0015 − 0.0255 0.0036 − 0.0036
Oct 0.0425 0.0044 − 0.0044 − 0.0256 0.0017 − 0.0017
Nov 0.0409 0.0042 − 0.0042 − 0.0267 0.0018 − 0.0018
Dec 0.0393 0.0045 − 0.0045 − 0.0102 0.0032 − 0.0032

The symbols *, and ** are the upper and lower critical bounds of the trend stope at 5%significance level.

Table 5
Result AI models for S1 station.
Lead Time Models Training Phase Testing Phase

MAE RMSE NSE WI MAE RMSE NSE WI

MSSIt+1 GBR 0.0318 0.0406 0.9403 0.9992 0.6181 0.8977 0.4755 0.8114
KNN 0.0651 0.0856 0.8778 0.9964 0.6370 0.9144 0.4595 0.8015
RELM-SO 0.0513 0.0661 0.9037 0.9978 0.1164 0.1453 0.9012 0.9966
LSTM 0.0809 0.1083 0.8482 0.9936 0.1416 0.1677 0.8799 0.9954
MSSIt+2 GBR 0.0519 0.0668 0.9026 0.9978 0.6750 0.9394 0.4293 0.7950
KNN 0.1756 0.2768 0.6705 0.9467 0.8575 1.1621 0.2750 0.6983
RELM-SO 0.0853 0.1066 0.8400 0.9943 0.2445 0.3044 0.7933 0.9840
LSTM 0.1346 0.1604 0.7475 0.9879 0.3051 0.3450 0.7420 0.9807
MSSIt+3 GBR 0.0400 0.0504 0.9250 0.9987 0.7312 0.9869 0.3818 0.7775
KNN 0.1485 0.1968 0.7217 0.9772 0.8142 1.1021 0.3116 0.7286
RELM-SO 0.1147 0.1409 0.7850 0.9900 0.3233 0.4000 0.7267 0.9710
LSTM 0.1135 0.1415 0.7874 0.9898 0.4180 0.4941 0.6466 0.9548
MSSIt+4 GBR 0.0375 0.0481 0.9298 0.9989 0.8084 1.0658 0.3165 0.7456
KNN 0.2499 0.3817 0.5318 0.8802 0.9598 1.2368 0.1884 0.6471
RELM-SO 0.1338 0.1670 0.7495 0.9858 0.4990 0.6363 0.5780 0.9168
LSTM 0.1099 0.1384 0.7941 0.9904 0.6270 0.6904 0.4699 0.9168
MSSIt+5 GBR 0.1978 0.2532 0.6294 0.9622 0.7972 1.0271 0.3260 0.7221
KNN 0.2816 0.4322 0.4725 0.8328 1.0022 1.2676 0.1526 0.6236
RELM-SO 0.1521 0.1920 0.7151 0.9809 0.5098 0.6338 0.5689 0.9176
LSTM 0.1456 0.1788 0.7272 0.9834 0.6169 0.7337 0.4784 0.8886
MSSIt+6 GBR 0.1334 0.1686 0.7499 0.9853 0.8559 1.0875 0.2789 0.7274
KNN 0.2777 0.4196 0.4793 0.8468 1.0100 1.2720 0.1490 0.6258
RELM-SO 0.1695 0.2122 0.6821 0.9765 0.5350 0.6547 0.5492 0.9119
LSTM 0.1477 0.1811 0.7230 0.9832 0.6278 0.7311 0.4711 0.8947

13
M.M. Hameed et al. Heliyon 10 (2024) e22942

Table 6
Result AI models for S2 station.
Lead Time Models Training Phase Testing Phase

MAE RMSE NSE WI MAE RMSE NSE WI

MSSIt+1 GBR 0.1046 0.1380 0.8730 0.9950 0.1140 0.1502 0.8672 0.9936
KNN 0.1441 0.1918 0.8251 0.9900 0.1342 0.1820 0.8437 0.9906
RELM-SO 0.1040 0.1377 0.8738 0.9951 0.0909 0.1211 0.8941 0.9960
LSTM 0.1048 0.1423 0.8728 0.9947 0.1060 0.1383 0.8765 0.9947
MSSIt+2 GBR 0.0783 0.1023 0.9053 0.9973 0.2023 0.2587 0.7627 0.9807
KNN 0.2414 0.3070 0.7078 0.9717 0.2171 0.2968 0.7454 0.9727
RELM-SO 0.1782 0.2292 0.7843 0.9861 0.1590 0.2125 0.8135 0.9872
LSTM 0.1804 0.2315 0.7818 0.9858 0.1843 0.2404 0.7839 0.9838
MSSIt+3 GBR 0.1563 0.1987 0.8112 0.9895 0.2570 0.3337 0.6985 0.9679
KNN 0.2813 0.3528 0.6601 0.9626 0.2567 0.3523 0.6990 0.9617
RELM-SO 0.2405 0.3051 0.7095 0.9746 0.1590 0.2125 0.8135 0.9872
LSTM 0.2315 0.2956 0.7203 0.9759 0.2415 0.3159 0.7168 0.9713
MSSIt+4 GBR 0.1283 0.1671 0.8453 0.9926 0.3153 0.4138 0.6302 0.9517
KNN 0.3587 0.4451 0.5674 0.9312 0.3253 0.4306 0.6185 0.9355
RELM-SO 0.2881 0.3631 0.6525 0.9629 0.2641 0.3540 0.6902 0.9633
LSTM 0.3007 0.3684 0.6373 0.9628 0.2934 0.3907 0.6559 0.9559
MSSIt+5 GBR 0.2016 0.2624 0.7571 0.9812 0.3594 0.4559 0.5786 0.9383
KNN 0.3856 0.4789 0.5356 0.9203 0.3521 0.4684 0.5871 0.9242
RELM-SO 0.3264 0.4147 0.6068 0.9502 0.3004 0.4060 0.6477 0.9508
LSTM 0.2900 0.3591 0.6506 0.9639 0.3436 0.4258 0.5970 0.9482
MSSIt+6 GBR 0.2858 0.3662 0.6567 0.9611 0.3629 0.4833 0.5716 0.9250
KNN 0.3885 0.4930 0.5334 0.9206 0.3649 0.5023 0.5693 0.9164
RELM-SO 0.3613 0.4580 0.5660 0.9371 0.3373 0.4570 0.6018 0.9353
LSTM 0.2034 0.2541 0.7556 0.9831 0.3696 0.4751 0.5637 0.9353

The study assessed drought forecasting models using performance metrics. However, conventional indicators like RMSE, R2, and
others may not be effective due to high auto-correlation in the drought time series data [56–58]. Thus, most forecasting models provide
very high accuracy with a correlation coefficient of more than 0.9 [87]. Therefore, the study focused on identifying critical and turning
points in time series data to evaluate the model’s ability to simulate them accurately. The process involved detecting the turning points
of original time series data (MSSI) and collecting corresponding values from multiple models, and then calculating the average ab­
solute error of the turning points (AAETP) for each model separately. This process was conducted for both stations to assess the model’s
predictability in forecasting hydrological drought for six months in advance. In this study, the AAETP is expressed using Equation (33)
AAETP = average (|Ei |) (33)
The computation of Ei (forecasting error computed for turning and critical points) for each model results in a vector is carried out by
subtracting the forecasted values of the critical points (i.e., forecasted ones) from their corresponding actual values. The algorithm for
computing turning points in a drought time series data can be summarized as follows:

i. Calculate the first derivative of the time series data (MSSI data) by taking the difference between consecutive data points.
ii. Find the sign of the first derivative using the sign function. The sign indicates whether the data is increasing or decreasing.
iii. Find the locations where the sign of the first derivative changes from positive to negative or vice versa. These locations are the
turning points.
iv. Find the turning points by looking for the points where the absolute difference in signs is equal to 2 (i.e., second derivative).
v. Visualize the data and the turning points using a plot.
vi. Once turning points are identified, compute AAETP.

4. Result and discussions

4.1. Drought trend analysis results

The present study utilized monthly streamflow data spanning 58 years (1961–2018) to calculate the MSSI for two study stations. To
determine MSSI, PCA combined different time scales of SSI, including SSI-1 up to SSI-48. The temporal distribution of the MSSI is
presented in Fig. 2 (a, and b). Additionally, the trend analysis was conducted using the ITAM at a significant level of 5 %. One of the
most notable features of ITAM is its ability to visually explore the trend of the time series data, whether increasing or decreasing.
However, due to the fluctuating nature of drought data, using traditional methods to explore drought patterns can be challenging. To
illustrate this point, let’s take the S1 station as an example and analyze drought patterns for each month separately, as shown in
Appendix E. At first glance, it appears that every month has a dry period followed by a wet period, making it difficult to discern the
drought trend. Therefore, this study adopted a technique of separating low drought values from high values. The monthly MSSI data is
divided into two groups based on optimal percentile value (threshold). Low values characterize the first group, while high values

14
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 6. Line graph showing the calculated and forecasted hydrological drought by the applied models for both stations. A) S1station. B) S2 station.

characterize the second group.

The basis for selecting the optimal threshold is to effectively separate higher and lower drought values, facilitating the analysis of
drought patterns and trends for each category. By establishing a clear separation between the two categories, any overlap or mixing of
hydrological drought data is avoided, enabling a focused examination of each group’s distinct trends. Also, this separation allows for
more effective identification and understanding of the patterns associated with low and high drought values. While determining the
specific threshold value can be complex, this study employed a trial-and-error method to compute the optimal threshold value for each
station that effectively separates the higher drought data from the lower data. Thus, this procedure would help to conduct a
comprehensive analysis of drought patterns and trends. To achieve this, the 60th percentile is used as the threshold for low values for
the S1 station, while the remaining 40th percentile (100-60) is used as the threshold for high values (Fig. 4(a to l)). Similarly, the 35th
percentile is used as the threshold for low values for the S2 station, while the 65th percentile (100-35) is used as the threshold for high
values (Fig. 5(a to l)). Each station has its own threshold value, which differs from the threshold value of other station. This difference
in threshold values is attributed to the variations in drought characteristics observed between the stations. The graphs presented
clearly demonstrate that the S2 stations have experienced a significant decrease in low drought values for all months. However, the
increasing trend in high drought values is only marginal. The drought patterns for the second station (S1) are quite complex. While a
decreasing trend was observed in low drought values, an increasing pattern was detected for higher drought values in all months of the
year. Tables 3 and 4 show the station-wise trend indicator, the slope of the trend indicator, the upper confidence level (Up), and the
lower confidence level (Lo) of ITAM for the monthly MSSI of S1 and S2 stations. The trend analysis results for both stations revealed

15
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. Comparison between calculated MSSI and forecasted values for S1 station through the testing phase: (A) RELM-SO, (B) LSTM, (C) GBR, (D)
KNN models. Comparison between calculated MSSI and forecasted values for S2 station through the testing phase: (E) RELM-SO, (F) LSTM, (G) GBR,
(H) KNN models.

that the months of October, November, and December exhibit the highest slope (i.e., SITAM ) in terms of low values compared to the
other months. This implies that the severity of drought during these months is remarkably high.

4.2. Modeling results

The current research utilized four distinct machine learning models to forecast the hydrological drought at two locations in
Malaysia, with a forecast horizon of up to six months. The capability of the hybrid forecasting model (RELM-SO) and LSTM model as
deep learning algorithms to forecast hydrological drought for two stations in Malaysia up to six months in advance is discussed in this
section of the study. Additionally, two standalone ML-based models (e.g., GBR and KNN) were used as a benchmark to assess the

16
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

efficiency of RELM-SO and LSTM. The quantitative assessment results of the forecasting models throughout the training and testing
phases for the S1 and S2 stations are provided in Tables 5 and 6. Overall, all models perform satisfactorily in the training stage, with the
GBR model exhibiting clear superiority. Notably, during the training stage, a model was given input and output data, whereas only
input data was provided in the testing phase. Accordingly, the latter stage is more precise in evaluating the effectiveness of the model’s
performance [77]. By analyzing the results presented in Tables 5 and 6, it is apparent that the performance of GBR is notably different
between the training and testing stages, unlike the other models (e.g., LSTM and RELM-SO) that exhibited excellent performance.
Besides, the RELM-SO has illustrated the best accuracy in forecasting the MSSI from one to six months ahead, followed by LSTM, GBR,
and KNN. As an illustration, the RELM-SO model demonstrates remarkable accuracy in forecasting a one-month drought in advance
(MSSIt+1) for the S1 station during the testing phase. This model yields lower values of RMSE (0.1453) and MAE (0.1164), as well as
higher values of NSE (0.9012) and WI (0.9966). Nonetheless, the LSTM model also shows good accuracy, lower than the RELM-SO
model, with an RMSE of 0.1677, MAE of 0.1416, NSE of 0.8799, and WI of 0.9954. Similarly, for station S2, the hybrid model’s
performance was more effected than comparable models, reporting fewer forecasting error (RMSE = 0.1211, and MAE = 0.0909), and
higher accuracy (NSE = 0.8941, and WI = 0.9960), followed by LSTM (RMSE = 0.1383, MAE = 0.1060, NSE = 0.8765, and WI =

17
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

0.9947). Generally, there was an outstanding variation in the capacity of the model in forecasting drought for the S1 and S2 stations,
likely due to differences in their respective training data quality. Specifically, the absence of severe drought conditions in the training
data for station S1 resulted in a relatively higher forecasting error during the testing phase. To further illustrate this point, a line graph
(Fig. 6(A, and B)) was created to depict how the models simulated the changes in drought values over time. The results indicate that
classical models such as GBR and KNN could not accurately simulate the most critical drought events, particularly in the S1 station. On
the other hand, advanced models like LSTM and RELM-SO provided highly precise results in simulating these critical events, even
when they were not present in the training data. This suggests that these models have excellent generalization capabilities. Overall, the
RELM-SO has simulated the MSSI for both stations much better than LSTM, reflecting its capability to forecast the hydrological
drought.
The models’ performance in simulating observed MSSI at both stations several months ahead (from one to six) was evaluated using
scatter plots for testing phases (Fig. 7 (A-H)). These plots provide informative visualizations of the deviation between forecasted and

18
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

calculated MSSI and the correlation of determination (R2) between them. The density distribution of real drought data was also
calculated, and the colored points indicate the corresponding projected values. The forecasted results illustrated that the RELM-SO
model outperforms the other models regarding higher R2 values. The R2 ranges for several models at S1 and S2 stations were
determined. At the S1 station, the RELM-SO model had the highest R2 range of 0.987 to 0.920, with LSTM following closely behind at
0.982 to 0.842. GBR and KNN had lower ranges of 0.817 to 0.731 and 0.812 to 0.640, respectively. Similarly, for the S2 station, the
RELM-SO model had the highest R2 values (0.973–0.767), followed by LSTM with a range of 0.967 to 0.760. GBR and KNN had lower
ranges of 0.962 to 0.732 and 0.945 to 0.711, respectively.

4.2.1. Analysis of forecasting models based on turning point detection

The analysis of the model’s capability to capture drought turning points has been assessed. Fig. 8a demonstrates how the applied
algorithm identifies the critical and turning points in the calculated drought data. Based on the results offered in Fig. 8 (b, and c), the

19
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

hybrid model (RELM-SO) demonstrated more accurate capturing of turning points compared to LSTM and other ML-based models, as
evidenced by its lowest value of AAETP. The superiority of RELM-SO was quantified in terms of its ability to reduce the AAETP in­
dicator. The findings indicated a significant improvement in forecasting accuracy of 50.64 %, 58.16 %, and 21.52 % using RELM-SO
compared to GBR, KNN, and LSTM models, respectively, for the S1 station. Similarly, for the S2 station, the results demonstrated a
significant enhancement in forecasting accuracy of 20.82 %, 28.61 %, and 19.32 % when compared to GBR, KNN, and LSTM models,
respectively.
An analysis of the relative errors in forecasting tuning and critical points is conducted to discern the optimal forecasting model. The
best model is characterized by its accuracy in pinpointing these points. Violin plots are used to present the relative error percentages
(Fig. 9(a and b)) due to their capacity to comprehensively display data elements like mean, median, IQR, and distribution. For station
S1, it’s evident that the KNN and GBR models yield high errors, oscillating between − 33 % and 69 %, in forecasting the turning points
for one month ahead, as demonstrated in Fig. 9a. Meanwhile, the LSTM model yields a smaller error range, fluctuating between − 3%
and − 40 %; only two points that account for a mere 3.13 % of total points surpass this range. Besides this, the RELM-SO model registers
the narrowest error range, hovering around − 18 % to − 20 %, only three points slightly exceeding this span. It’s clear that the RELM-SO

20
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

is the most reliable model for forecasting drought at station S1. Similarly, for station S2, the hybrid model manifests the least relative
errors, as shown in Fig. 9b, affirming its superiority over the LSTM and all other models.

4.2.2. Reliability analyses

The study suggests that further statistical assessments are necessary to determine whether the adopted model (RELM-SO) has better
accuracy for long-term drought forecasting than other models. Some researchers propose using reliability assessment (RA), which is an
advanced statistical indicator commonly used in the literature to evaluate model performance and consistency [88]. RA can also
determine whether the forecasting models meet the minimum requirement for acceptable accuracy [73]. Thus, RA metric is a crucial
tool for determining the effectiveness of forecasting models. RA can be mathematically expressed using Equation (34).
( ) Ns
100% ∑
(RA) = jk (34)
Ns k=1

where Ns is the total number of MSSI ji is an RA index calculated by two steps. The term jk is calculated via two steps. First, it is crucial
to determine the Relative Absolute Error (RAE) as a vector of jk component using Equation (35).

21
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

⃒ ⃒
⃒MSWIk(Observed) − MSWIk(Forecasted) ⃒
RAEk = ⃒⃒ ⃒
⃒ (35)
MSWIk(Observed)

In the second step, the equivalent value of jk is determined based on the threshold (φ) of an adequate drought parameter and the RAEk .
If the RAEk is greater than α, then jk is assigned a value of zero. However, if RAEk is less than or equal to φ, then jk is assigned a value of
one. The Chinese Standards recommend an optimal value of 0.2 (20 %) for φ.
Based on the reliability metric, it is evident from Fig. 10 (A, and B) that the RELM-SO model outperforms other models in terms of
dependability. The results demonstrate that, when forecasting MSSI several months ahead (t+1 to t+6), the proposed model out­
performs other comparable models, as evidenced by the highest RA ranging from 87.65 % to 20.18 % for the S1 station (see Fig. 10 A)
and from 79.78 % to 36.25 % for the S2 station (see Fig. 10 B). Thus, the study confirms the significant superiority of the proposed
model. As forecasting accuracy declines with increasing drought forecast lead time, conducting a RA can provide valuable insights into
identifying the optimal number of months in advance for a model to produce high-precision forecasts. Based on illustrated results in
Fig. 10(A, and B), it can be inferred that the RELM-SO can produce reliable forecasts for the first station up to two months in advance,

22
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 7. (continued).

while for the second station, the forecast horizon is limited to three months.

4.2.3. Discussion
Multi-month ahead drought forecasting is imperative in establishing strategies to mitigate the impacts of drought, manage water
resources efficiently, and develop an early warning system. This study employed multiple models to predict one to six-month-ahead
droughts for two particular dry regions in Malaysia. The RELM-SO model proved to be more accurate than the LSTM model, benchmark
models like KNN, GBR, and others that were developed in earlier works. Quantitative results indicate that the RELM-SO model per­
forms better in multi-month drought forecasting. This can be attributed to the model’s structure, which amalgamates the advantages of
two robust algorithms: RELM and SO. Significantly, SO was utilized to train the RELM due to its proven effectiveness in solving a
variety of real-world engineering problems and surpassing other traditional and new algorithms (e.g., Coyote Optimization Algorithm,
Moth-flame Optimization, Harris Hawks Optimizer, Whale Optimization Algorithm, etc.) [83]. A noteworthy element is the RELM’s
regularization parameter, which enhances the models’ ability to yield more accurate forecasts, particularly with testing data [82]. This

23
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 8. Turning point detection and analysis: a) illustration of the algorithm’s capability in detecting critical and turning points in drought time
series data. b) AAETP results for station S1.c) AAETP results for station S2.

parameter also augments the model’s capability by striking a balance between two forms of risk: structural risk (related to the models’
coefficients) and empirical risk (related to forecasting errors).
Further analysis is necessary to ensure the superiority of the suggested RELM-SO model in forecasting drought one month in
advance in Malaysia. This involves evaluating the accuracy of the RELM-SO model by comparing it with other models established in
previous studies. One of the studies used advanced models based on a combination of wavelet transform (WA) with ANFIS, creating the
WANFIS model [89]. Similarly, the second model was established based on a combination of WA, ARIMA, and ANN, creating the
WAANN model. Quantitative analysis showed that the latter model had higher forecasting accuracy for drought, with an R2 value of
0.9603. Another study investigated the capacity of a model established through a combination of WA and ANN, which achieved a
higher result with an R2 of 0.938 [90]. Moreover, other researchers presented advanced models using Fuzzy-SVR, Boosted-SVR, and
classical SVR to forecast drought and found that the Fuzzy-SVR model performed best with an R2 value of 0.903 [91]. Additionally,
literature has investigated single models such as ANN and found that it can provide a good value of R2, reaching 0.833 [92]. Finally, an
advanced combining model was developed based on the integration of WA, ARIMA, and ANN, which provided higher results with an
R2 of 0.872 [92]. Furthermore, some researchers have used a novel model based on ANN coupled with the Firefly Algorithm to forecast
drought at various temporal scales using the SPI index [93]. The best results obtained in these studies yielded an R2 value of 0.887.
Overall, the reviewed models provided satisfactory accuracy, with R2 values ranging from 0.9603 to 0.833. However, the accuracy
forecasting of these models is still less than that obtained in the RELM-SO model, which provides a higher forecasting accuracy ranging
from 0.987 to 0.973.
The poor forecasting results of the GBR and KNN models for the S1 station from 2012 to 2018, specifically for peak drought values,

24
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 9. Depicts violin diagrams illustrating the relative error percentages. Panel a) represents the S1 station, while panel b) corresponds to the
S2 station.

suggest that these models may not effectively capture the complex patterns and dynamics of hydrological drought data. This indicates
that they might oversimplify the relationships between the input features and the target variable, making them unsuitable for accu­
rately forecasting drought values. Moreover, the relatively high correlations between the features and outputs negatively affect
ensemble models like GBR, leading to poor forecasting results [94].
The analysis of turning points using the AAETP indicator reveals that the RELM-SO model performs the best, followed by LSTM. The
Violin diagram presentation provides valuable information on the quality forecasting of these models. Observing Fig. 9, it is evident
that the distribution of relative error for RELM-SO closely resembles a normal distribution, unlike other models, in both studied
stations. Notably, all models, except RELM-SO, exhibit noticeable skewness towards positive and negative values in both stations.
Furthermore, in station SI, the relative error distribution of the LSTM model is less accurate compared to its competitor, RELM-SO. The
LSTM model tends to provide underestimating forecasts, with most relative error records above zero. Consequently, based on the
lowest forecasting error at the turning points for both stations and the normality of their distribution, it is confirmed that RELM-BWO is
the superior model.
In the studied area, the achieved results may have a notable impact on water resource management, agriculture, and disaster
preparedness. Accurately anticipating drought conditions and utilizing advanced drought trend analysis may provide stakeholders and
decision-makers with significant information on drought dynamics. This enables them to implement ecosystem protection measures,
sustainable water resource management strategies, and advanced drought warning systems. Such a comprehensive and precise
analysis may effectively mitigate the environmental impacts of drought.

5. Conclusion

Accurate trend analysis and reliable drought forecasting are both crucial for understanding drought variability, informing stake­
holders about consequences, and facilitating effective water resource management to mitigate the negative impacts of drought. The
hydrological drought at two hydrological stations in West Malaysia was computed using MSSI based on streamflow data spanning 58

25
M.M. Hameed et al. Heliyon 10 (2024) e22942

Fig. 10. Results of reliability analysis for assessing model efficiencies in forecasting MSSI for one to six months in advance. A) S1 station, and B)
S2 station.

years (1961–2018). This study compares the performance of four models - the hybrid model (RELM-SO), deep learning model (LSTM),
GBR, and KNN - in forecasting droughts over a forecasting horizon of one to six months ahead. The results highlight that the hybrid
model outperforms the comparable models and several novel and recent models developed in the literature. Considering the presence
of significant autocorrelation in the drought time series data, the models were evaluated based on their capacity to capture critical and
turning drought points. The RELM-SO demonstrates superior performance in capturing critical and turning points, as AAETP values
indicate. The hybrid model also exhibits substantial improvements in forecast accuracy compared to GBR, KNN, and LSTM models,
with reductions in AAETP ranging from 21.52 % to 58.16 % for the S1 station and improvements of 19.32 %–28.61 % for the S2 station.
Furthermore, the analysis reveals that RELM-SO can provide reliable forecasts up to two months in advance for the S1 station, while the
forecast horizon for the S2 station is limited to three months in advance. These findings highlight the effectiveness of the hybrid model
in drought forecasting, surpassing other models in both capturing turning points and improving forecast accuracy.
Additionally, this study addresses the need for comprehensive drought trend analysis using the ITAM method. Trend analysis
revealed that the S2 station experienced a significant decrease in low drought values for all months, while the increasing trend in high
drought values was marginal. The drought patterns for the S1 station were complex, with a remarkable decreasing trend in low drought
values but an increasing trend in higher drought values throughout the year has been observed. Besides, the months of October,
November, and December exhibited the highest slope (i.e., SITAM ), showing a higher frequency of drought events than in other months.
Overall, this study showcases the effectiveness of the hybrid model (RELM-SO) in accurately forecasting droughts, capturing critical
turning points, and providing valuable insights through effective drought trend analysis, thereby enhancing water resource man­
agement and decision-making.

Data availability

The supporting data for this study are owned by the Department of Irrigation and Drainage Malaysia. Due to restrictions, the

26
M.M. Hameed et al. Heliyon 10 (2024) e22942

authors are not authorized to share the data publicly.

CRediT authorship contribution statement

Mohammed Majeed Hameed: Writing - original draft, Visualization, Software, Methodology, Conceptualization. Siti Fatin Mohd
Razali: Validation, Supervision, Project administration, Data curation, Conceptualization. Wan Hanna Melini Wan Mohtar: Vali­
dation, Investigation, Formal analysis, Data curation. Majed Omar Ahmad Alsaydalani: Visualization, Funding acquisition, Data
curation. Zaher Mundher Yaseen: Writing - review & editing, Validation, Investigation.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgements

The authors appreciate the Department of Irrigation and Drainage Malaysia for providing the streamflow data used in this study.

Appendices A to E. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.heliyon.2023.e22942.

References

[1] Q.B. Pham, T.C. Yang, C.M. Kuo, H.W. Tseng, P.S. Yu, Coupling singular spectrum analysis with least square support vector machine to improve accuracy of SPI
drought forecasting, Water Resour. Manag. 35 (2021) 847–868, https://doi.org/10.1007/s11269-020-02746-7.
[2] Y. Zhang, W. Li, Q. Chen, X. Pu, L. Xiang, Multi-models for SPI drought forecasting in the north of haihe River Basin, China, stoch, Environ. Res. Risk Assess.
(2017), https://doi.org/10.1007/s00477-017-1437-5.
[3] T.B. Mckee, N.J. Doesken, J. Kleist, The relationship of drought frequency and duration to time scales, AMS 8th Conf. Appl. Climatol. (1993) 179–184.
[4] K.A. Tareke, A.G. Awoke, Comparing surface water supply index and streamflow drought index for hydrological drought analysis in Ethiopia, Heliyon 8 (2022),
e12000, https://doi.org/10.1016/j.heliyon.2022.e12000.
[5] I.F. Hasan, R. Abdullah, Multivariate index for monitoring drought (case study, Northeastern of Iraq), Nat. Hazards (2023), https://doi.org/10.1007/s11069-
023-05837-x.
[6] S. Shamshirband, S. Hashemi, H. Salimi, S. Samadianfard, E. Asadi, S. Shadkani, K. Kargar, A. Mosavi, N. Nabipour, K.W. Chau, Predicting Standardized
Streamflow index for hydrological drought using machine learning models, Eng. Appl. Comput. Fluid Mech. (2020), https://doi.org/10.1080/
19942060.2020.1715844.
[7] N. Van Hieu, N. Van Tuan, N.K. Bang, P.H. Hai, T.T. Hoa, Assessment of hydrological drought using the standardized streamflow index (SSFI): a case study of the
tien yen River Basin of quang ninh province, vietnam, J. Geosci. Environ. Protect. 10 (2022) 309–326.
[8] M. Dehghani, B. Saghafian, F. Nasiri Saleh, A. Farokhnia, R. Noori, Uncertainty analysis of streamflow drought forecast using artificial neural networks and
Monte-Carlo simulation, Int. J. Climatol. 34 (2014) 1169–1180, https://doi.org/10.1002/joc.3754.
[9] P. Aghelpour, H. Bahrami-Pichaghchi, V. Varshavian, Hydrological drought forecasting using multi-scalar streamflow drought index, stochastic models and
machine learning approaches, in northern Iran, Stoch. Environ. Res. Risk Assess. 35 (2021) 1615–1635, https://doi.org/10.1007/s00477-020-01949-z.
[10] N. Nabipour, M. Dehghani, A. Mosavi, S. Shamshirband, Short-term hydrological drought forecasting based on different nature-inspired optimization algorithms
hybridized with artificial neural networks, IEEE Access 8 (2020) 15210–15222, https://doi.org/10.1109/ACCESS.2020.2964584.
[11] K.A. Tareke, A.G. Awoke, Hydrological drought forecasting and monitoring system development using artificial neural network (ANN) in Ethiopia, Heliyon 9
(2023), e13287, https://doi.org/10.1016/j.heliyon.2023.e13287.
[12] Z. Ali, I. Hussain, M. Faisal, H.M. Nazir, T. Hussain, M.Y. Shad, A. Mohamd Shoukry, S. Hussain Gani, Forecasting drought using multilayer perceptron artificial
neural network model, Adv. Meteorol. 2017 (2017), 5681308, https://doi.org/10.1155/2017/5681308.
[13] R.C. Deo, O. Kisi, V.P. Singh, Drought forecasting in eastern Australia using multivariate adaptive regression spline, least square support vector machine and
M5Tree model, Atmos. Res. (2017), https://doi.org/10.1016/j.atmosres.2016.10.004.
[14] M. Borji, A. Malekian, A. Salajegheh, M. Ghadimi, Multi-time-scale analysis of hydrological drought forecasting using support vector regression (SVR) and
artificial neural networks (ANN), Arabian J. Geosci. (2016), https://doi.org/10.1007/s12517-016-2750-x.
[15] A. Dikshit, B. Pradhan, A.M. Alamri, Short-term spatio-temporal drought forecasting using random forests model at new south wales, Australia, Appl. Sci. 10
(2020), https://doi.org/10.3390/app10124254.
[16] M. Lotfirad, H. Esmaeili-Gisavandani, A. Adib, Drought monitoring and prediction using SPI, SPEI, and random forest model in various climates of Iran, J. Water
Clim. Chang. (2021), https://doi.org/10.2166/wcc.2021.287 jwc2021287.
[17] G.C. Wang, Q. Zhang, S.S. Band, M. Dehghani, K. wing Chau, Q.T. Tho, S. Zhu, S. Samadianfard, A. Mosavi, Monthly and seasonal hydrological drought
forecasting using multiple extreme learning machine models, Eng. Appl. Comput. Fluid Mech. 16 (2022) 1364–1381.
[18] Z.M. Yaseen, M. Ali, A. Sharafati, N. Al-Ansari, S. Shahid, Forecasting standardized precipitation index using data intelligence models: regional investigation of
Bangladesh, Sci. Rep. 11 (2021), https://doi.org/10.1038/s41598-021-82977-9.
[19] M. Jamei, I. Ahmadianfar, M. Karbasi, A. Malik, O. Kisi, Z.M. Yaseen, Development of wavelet-based kalman online sequential extreme learning machine
optimized with boruta-random forest for drought index forecasting, Eng. Appl. Artif. Intell. 117 (2023), 105545.
[20] A. Danandeh Mehr, A. Torabi Haghighi, M. Jabarnejad, M.J. Safari, V. Nourani, A new evolutionary hybrid random forest model for SPEI forecasting, Water 14
(2022), https://doi.org/10.3390/w14050755.
[21] A. Kaur, S.K. Sood, Cloud-Fog based framework for drought prediction and forecasting using artificial neural network and genetic algorithm, J. Exp. Theor. Artif.
Intell. 32 (2020) 273–289, https://doi.org/10.1080/0952813X.2019.1647563.
[22] O. Kisi, A. Docheshmeh Gorgij, M. Zounemat-Kermani, A. Mahdavi-Meymand, S. Kim, Drought forecasting using novel heuristic methods in a semi-arid
environment, J. Hydrol. 578 (2019), 124053, https://doi.org/10.1016/j.jhydrol.2019.124053.

27
M.M. Hameed et al. Heliyon 10 (2024) e22942

[23] A. Malik, Y. Tikhamarine, D. Souag-Gamane, P. Rai, S.S. Sammen, O. Kisi, Support vector regression integrated with novel meta-heuristic algorithms for
meteorological drought prediction, Meteorol. Atmos. Phys. 133 (2021) 891–909, https://doi.org/10.1007/S00703-021-00787-0.
[24] P. Aghelpour, B. Mohammadi, S. Mehdizadeh, H. Bahrami-Pichaghchi, Z. Duan, A novel hybrid dragonfly optimization algorithm for agricultural drought
prediction, Stoch. Environ. Res. Risk Assess. 35 (2021) 2459–2477, https://doi.org/10.1007/s00477-021-02011-2.
[25] A. Danandeh Mehr, B. Vaheddoost, B. Mohammadi, ENN-SA: a novel neuro-annealing model for multi-station drought prediction, Comput. Geosci. 145 (2020),
104622, https://doi.org/10.1016/j.cageo.2020.104622.
[26] O. Rahmati, M. Panahi, Z. Kalantari, E. Soltani, F. Falah, K.S. Dayal, F. Mohammadi, R.C. Deo, J. Tiefenbacher, D. Tien Bui, Capability and robustness of novel
hybridized models used for drought hazard modeling in southeast Queensland, Australia, Sci. Total Environ. 718 (2020), 134656, https://doi.org/10.1016/j.
scitotenv.2019.134656.
[27] M. Achite, F.B. Banadkooki, M. Ehteram, A. Bouharira, A.N. Ahmed, A. Elshafie, Exploring Bayesian model averaging with multiple ANNs for meteorological
drought forecasts, Stoch. Environ. Res. Risk Assess. (2022), https://doi.org/10.1007/s00477-021-02150-6.
[28] R.M. Adnan, R.R. Mostafa, A.R.M.T. Islam, A.D. Gorgij, A. Kuriqi, O. Kisi, Improving drought modeling using hybrid random vector functional link methods,
Water 13 (2021), https://doi.org/10.3390/w13233379.
[29] F. Ahmadi, S. Mehdizadeh, B. Mohammadi, Development of bio-inspired- and wavelet-based hybrid models for reconnaissance drought index modeling, Water
Resour. Manag. 35 (2021) 4127–4147, https://doi.org/10.1007/s11269-021-02934-z.
[30] A. Danandeh Mehr, A. Rikhtehgar Ghiasi, Z.M. Yaseen, A.U. Sorman, L. Abualigah, A novel intelligent deep learning predictive model for meteorological
drought forecasting, J. Ambient Intell. Humaniz. Comput. (2022) 1–15.
[31] A. Docheshmeh Gorgij, M. Alizamir, O. Kisi, A. Elshafie, Drought modelling by standard precipitation index (SPI) in a semi-arid climate using deep learning
method: long short-term memory, Neural Comput, Appl 34 (2022) 2425–2442, https://doi.org/10.1007/s00521-021-06505-6.
[32] A. Dikshit, B. Pradhan, A. Huete, An improved SPEI drought forecasting approach using the long short-term memory neural network, J. Environ. Manage. 283
(2021), https://doi.org/10.1016/j.jenvman.2021.111979.
[33] A. Anshuka, R. Chandra, A.J. V Buzacott, D. Sanderson, F.F. van Ogtrop, Spatio temporal hydrological extreme forecasting framework using LSTM deep learning
model, Stoch. Environ. Res. Risk Assess. 36 (2022) 3467–3485, https://doi.org/10.1007/s00477-022-02204-3.
[34] Y.-S. Ham, K.-B. Sonu, U.-S. Paek, K.-C. Om, S.-I. Jong, K.-R. Jo, Comparison of LSTM network, neural network and support vector regression coupled with
wavelet decomposition for drought forecasting in the western area of the DPRK, Nat. Hazards 116 (2023) 2619–2643, https://doi.org/10.1007/s11069-022-
05781-2.
[35] A. Kaur, S.K. Sood, Deep learning based drought assessment and prediction framework, Ecol. Inform. 57 (2020), 101067, https://doi.org/10.1016/j.
ecoinf.2020.101067.
[36] A. Dikshit, B. Pradhan, A.R. Huete, Spatial meteorological drought forecasting using deep learning for New South Wales, Australia, in: AGU Fall Meet, Abstr,
2020. H178-14.
[37] S. Saha, B. Kundu, A. Saha, K. Mukherjee, B. Pradhan, Manifesting deep learning algorithms for developing drought vulnerability index in monsoon climate
dominant region of West Bengal, India, Theor. Appl. Climatol. 151 (2023) 891–913.
[38] H. Ouatiki, A. Boudhar, A. Ouhinou, A. Arioua, M. Hssaisoune, H. Bouamri, T. Benabdelouahab, Trend analysis of rainfall and drought over the oum Er-rbia
River Basin in Morocco during 1970–2010, Arabian J. Geosci. (2019), https://doi.org/10.1007/s12517-019-4300-9.
[39] D.R. Mahajan, B.M. Dodamani, Trend analysis of drought events over upper krishna basin in Maharashtra, Aquat. Procedia. 4 (2015) 1250–1257, https://doi.
org/10.1016/j.aqpro.2015.02.163.
[40] Ü. Güner Bacanli, Trend analysis of precipitation and drought in the Aegean region, Turkey, Meteorol. Appl. 24 (2017) 239–249, https://doi.org/10.1002/
met.1622.
[41] F. Tosunoglu, O. Kisi, Trend analysis of maximum hydrologic drought variables using mann–kendall and şen’s innovative trend method, River Res. Appl. 33
(2017) 597–610, https://doi.org/10.1002/rra.3106.
[42] A.R. Zarei, M.M. Moghimi, M.R. Mahmoudi, Parametric and non-parametric trend of drought in arid and semi-arid regions using RDI index, water resour,
Manag 30 (2016) 5479–5500, https://doi.org/10.1007/s11269-016-1501-9.
[43] T. Caloiero, Evaluation of rainfall trends in the South Island of New Zealand through the innovative trend analysis (ITA), Theor. Appl. Climatol. (2019), https://
doi.org/10.1007/s00704-019-02988-5.
[44] H. von Storch, in: H. von Storch, A. Navarra (Eds.), Misuses of Statistical Analysis in Climate Research BT - Analysis of Climate Variability, Springer Berlin
Heidelberg, Berlin, Heidelberg, 1999, pp. 11–26.
[45] S. Yue, P. Pilon, B. Phinney, G. Cavadias, The influence of autocorrelation on the ability to detect trend in hydrological series, Hydrol. Process. 16 (2002)
1807–1829, https://doi.org/10.1002/hyp.1095.
[46] E.M. Douglas, R.M. Vogel, C.N. Kroll, Trends in floods and low flows in the United States: impact of spatial correlation, J. Hydrol. 240 (2000) 90–105, https://
doi.org/10.1016/S0022-1694(00)00336-X.
[47] Z. Şen, Innovative trend analysis methodology, J. Hydrol. Eng. (2012), https://doi.org/10.1061/(ASCE)HE.1943-5584.0000556.
[48] A.E. Harka, N.B. Jilo, F. Behulu, Spatial-temporal rainfall trend and variability assessment in the Upper Wabe Shebelle River Basin, Ethiopia: application of
innovative trend analysis method, J. Hydrol. Reg. Stud. 37 (2021), 100915, https://doi.org/10.1016/j.ejrh.2021.100915.
[49] S. Gul, J. Ren, Application of non-parametric innovative trend analysis of different time scale precipitation during (1951–2016) in Khyber Pakhtunkhwa,
Pakistan, Acta Geophys. 70 (2022) 485–503, https://doi.org/10.1007/s11600-021-00703-5.
[50] T. Caloiero, SPI trend analysis of New Zealand applying the ITA technique, Geosci. (2018), https://doi.org/10.3390/geosciences8030101.
[51] J. Li, W. Wu, X. Ye, H. Jiang, R. Gan, H. Wu, J. He, Y. Jiang, Innovative trend analysis of main agriculture natural hazards in China during 1989–2014, Nat.
Hazards 95 (2019) 677–720, https://doi.org/10.1007/s11069-018-3514-6.
[52] A. Elouissi, B. Benzater, I. Dabanli, M. Habi, A. Harizia, A. Hamimed, Drought investigation and trend assessment in Macta watershed (Algeria) by SPI and ITA
methodology, Arabian J. Geosci. 14 (2021) 1329, https://doi.org/10.1007/s12517-021-07670-7.
[53] İ. Dabanlı, Z. Şen, M.Ö. Yeleğen, E. Şişman, B. Selek, Y.S. Güçlü, Trend Assessment by the Innovative-Şen Method, Water Resour, Manag, 2016, https://doi.org/
10.1007/s11269-016-1478-4.
[54] J. Bazrafshan, S. Hejabi, J. Rahimi, Drought monitoring using the multivariate standardized precipitation index (MSPI), water resour, Manag 28 (2014)
1045–1060, https://doi.org/10.1007/s11269-014-0533-2.
[55] H.H. Hasan, S.F. Mohd Razali, N.S. Muhammad, A. Ahmad, Hydrological drought across peninsular Malaysia: implication of drought index, Nat. Hazards Earth
Syst. Sci. Discuss. 2021 (2021) 1–28, https://doi.org/10.5194/nhess-2021-249.
[56] S. Madadgar, H. Moradkhani, A bayesian framework for probabilistic seasonal drought forecasting, J. Hydrometeorol. 14 (2013) 1685–1705, https://doi.org/
10.1175/JHM-D-13-010.1.
[57] A. Malik, A. Kumar, Meteorological drought prediction using heuristic approaches based on effective drought index: a case study in Uttarakhand, Arabian J.
Geosci. 13 (2020) 276, https://doi.org/10.1007/s12517-020-5239-6.
[58] M. Özger, E.E. Başakın, Ö. Ekmekcioğlu, V. Hacısüleyman, Comparison of wavelet and empirical mode decomposition hybrid models in drought prediction,
Comput. Electron. Agric. 179 (2020), https://doi.org/10.1016/j.compag.2020.105851.
[59] P. Ghasemi, M. Karbasi, A. Zamani Nouri, M. Sarai Tabrizi, H.M. Azamathulla, Application of Gaussian process regression to forecast multi-step ahead SPEI
drought index, Alex. Eng. J. 60 (2021) 5375–5392, https://doi.org/10.1016/j.aej.2021.04.022.
[60] S.F.Z. Mustafa, M. Mat Deris, T.S.B. Abd Manan, S. Beddu, N.L. Mohd Kamal, D. Mohamad, S. Yavari, S. Qazi, Z. Hanafiah, S. Omar Abu Nassar, K.L. Yeoh,
I. Sheriff, W.H.M. Wan Mohtar, M.H. Isa, M.S. Yusoff, H. Abdul Aziz, Modelling of similarity characteristics of polycyclic aromatic hydrocarbons (PAHs) in
Sungai Perak, Malaysia via rough set theory and principal component analysis (PCA), Chem. Phys. Lett. 828 (2023), 140721, https://doi.org/10.1016/j.
cplett.2023.140721.

28
M.M. Hameed et al. Heliyon 10 (2024) e22942

[61] S.A. Kadhum, M.Y. Ishak, S.Z. Zulkifli, Evaluation and assessment of baseline metal contamination in surface sediments from the Bernam River, Malaysia,
Environ. Sci. Pollut. Res. 23 (2016) 6312–6321, https://doi.org/10.1007/s11356-015-5853-0.
[62] F.M. Kusin, S.N. Muhammad, M.S.M. Zahar, Z. Madzin, Integrated River Basin Management: incorporating the use of abandoned mining pool and implication on
water quality status, Desalin. Water Treat. 57 (2016) 29126–29136, https://doi.org/10.1080/19443994.2016.1168132.
[63] Z. Zhou, L. Wang, A. Lin, M. Zhang, Z. Niu, Innovative trend analysis of solar radiation in China during 1962–2015, Renew. Energy (2018), https://doi.org/
10.1016/j.renene.2017.12.052.
[64] A. Bhattarai, D. Qadir, A.M. Sunusi, B. Getachew, A.R. Mallah, Dynamic sliding window-based long short-term memory model development for Pan evaporation
forecasting, Knowledge-Based Eng. Sci. 4 (2023) 37–54.
[65] S. Poornima, M. Pushpalatha, Drought prediction based on SPI and SPEI with varying timescales using LSTM recurrent neural network, Soft Comput. (2019),
https://doi.org/10.1007/s00500-019-04120-1.
[66] S. Xu, W. Li, Y. Zhu, A. Xu, A novel hybrid model for six main pollutant concentrations forecasting based on improved LSTM neural networks, Sci. Rep. 12
(2022), 14434, https://doi.org/10.1038/s41598-022-17754-3.
[67] S. Elsayed, M. Gupta, G. Chaudhary, S. Taneja, H. Gaur, M. Gad, M. Hamdy Eid, A. Kovács, S. Péter, A. Gaagai, U. Schmidhalter, Interpretation the influence of
hydrometeorological variables on soil temperature prediction using the potential of deep learning model, Knowledge-Based Eng. Sci. 4 (2023) 55–77, https://
doi.org/10.51526/kbes.2023.4.1.55-77.
[68] G. Fayer, L. Lima, F. Miranda, J. Santos, R. Campos, V. Bignoto, M. Andrade, M. Moraes, C. Ribeiro, P. Capriles, A temporal fusion transformer deep learning
model for long-term streamflow forecasting: a case study in the funil reservoir, southeast Brazil, knowledge-based, Eng. Sci. 4 (2023) 73–88.
[69] M.M. Hameed, F. Khaleel, D. Khaleel, Employing a robust data-driven model to assess the environmental damages caused by installing grouted columns, in:
2021 Third Int. Sustain. Resil. Conf. Clim. Chang., 2021, pp. 305–309, https://doi.org/10.1109/IEEECONF53624.2021.9668027.
[70] M.K. Alomar, F. Khaleel, M.M. Aljumaily, A. Masood, S.F.M. Razali, M.A. AlSaadi, N. Al-Ansari, M.M. Hameed, Data-driven models for atmospheric air
temperature forecasting at a continental climate region, PLoS One 17 (2022), e0277079.
[71] P. Nie, M. Roccotelli, M. Pia, Z. Ming, Z. Li, Prediction of home energy consumption based on gradient boosting regression tree, Energy Rep. 7 (2021)
1246–1255, https://doi.org/10.1016/j.egyr.2021.02.006.
[72] G. Bin Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: theory and applications, Neurocomputing (2006), https://doi.org/10.1016/j.
neucom.2005.12.126.
[73] M.K. AlOmar, F. Khaleel, A.A. AlSaadi, M.M. Hameed, M.A. AlSaadi, N. Al-Ansari, The influence of data length on the performance of artificial intelligence
models in predicting air pollution, Adv. Meteorol. 2022 (2022), 5346647, https://doi.org/10.1155/2022/5346647.
[74] M.M. Hameed, M.K. Alomar, S.F. Mohd Razali, M.A. Kareem Khalaf, W.J. Baniya, A. Sharafati, M.A. Alsaadi, Application of artificial intelligence models for
evapotranspiration prediction along the southern coast of Turkey, Complexity (2021), https://doi.org/10.1155/2021/8850243.
[75] P. Parisouj, H. Mohebzadeh, T. Lee, Employing machine learning algorithms for streamflow prediction: a case study of four river basins with different climatic
zones in the United States, water resour, Manag 34 (2020) 4113–4131, https://doi.org/10.1007/s11269-020-02659-5.
[76] Y. Jiang, X. Bao, S. Hao, H. Zhao, X. Li, X. Wu, Monthly streamflow forecasting using ELM-IPSO based on phase space reconstruction, Water Resour. Manag. 34
(2020) 3515–3531, https://doi.org/10.1007/s11269-020-02631-3.
[77] M.M. Hameed, F. Khaleel, M.K. AlOmar, S.F. Mohd Razali, M.A. AlSaadi, Optimising the selection of input variables to increase the predicting accuracy of shear
strength for deep beams, Complexity 2022 (2022), 6532763, https://doi.org/10.1155/2022/6532763.
[78] W. Xie, J. Wang, C. Xing, S.-S. Guo, M. Guo, L. Zhu, Adaptive hybrid soft-sensor model of grinding process based on regularized extreme learning machine and
least squares support vector machine optimized by golden sine Harris hawk optimization algorithm, Complexity 2020 (2020), 6457517, https://doi.org/
10.1155/2020/6457517.
[79] H. Sanikhani, R.C. Deo, P. Samui, O. Kisi, C. Mert, R. Mirabbasi, S. Gavili, Z.M. Yaseen, Survey of different data-intelligent modeling strategies for forecasting air
temperature using geographic information as model predictors, Comput. Electron. Agric. 152 (2018) 242–260.
[80] W. Deng, Q. Zheng, L. Chen, Regularized extreme learning machine, in: 2009 IEEE Symp. Comput. Intell. Data Min., 2009, pp. 389–395, https://doi.org/
10.1109/CIDM.2009.4938676.
[81] V. Vapnik, The Nature of Statistical Learning Theory, Springer science & business media, 1999.
[82] M.M. Hameed, M.K. AlOmar, A.A.A. Al-Saadi, M.A. AlSaadi, Inflow forecasting using regularized extreme learning machine: haditha reservoir chosen as case
study, Stoch. Environ. Res. Risk Assess. (2022), https://doi.org/10.1007/s00477-022-02254-7.
[83] F.A. Hashim, A.G. Hussien, Snake Optimizer: a novel meta-heuristic optimization algorithm, Knowledge-Based Syst. 242 (2022), 108320, https://doi.org/
10.1016/j.knosys.2022.108320.
[84] I. Al-Shourbaji, P.H. Kachare, S. Alshathri, S. Duraibi, B. Elnaim, M. Abd Elaziz, An efficient parallel reptile search algorithm and snake optimizer approach for
feature selection, Mathematics 10 (2022), https://doi.org/10.3390/math10132351.
[85] N. Mamat, S.F.M. Razali, Comparisons of various imputation methods for incomplete water quality data: a case study of the langat river, Malaysia, J. Kejuruter.
35 (2023) 191–201.
[86] O. Kisi, B. Choubin, R.C. Deo, Z.M. Yaseen, Incorporating synoptic-scale climate signals for streamflow modelling over the Mediterranean region using machine
learning models, Hydrol. Sci. J. 64 (2019) 1240–1252.
[87] Y. Zhang, H. Yang, H. Cui, Q. Chen, Comparison of the ability of ARIMA, WNN and SVM models for drought forecasting in the sanjiang plain, China, Nat. Resour.
Res. 29 (2020) 1447–1464, https://doi.org/10.1007/s11053-019-09512-6.
[88] F. Saberi-Movahed, M. Najafzadeh, A. Mehrpooya, Receiving more accurate predictions for longitudinal dispersion coefficients in water pipelines: training
group method of data handling using extreme learning machine conceptions, water resour, Manag 34 (2020) 529–561.
[89] Y.W. Soh, C.H. Koo, Y.F. Huang, K.F. Fung, Application of artificial intelligence models for the prediction of standardized precipitation evapotranspiration index
(SPEI) at Langat River Basin, Malaysia, Comput. Electron. Agric. 144 (2018) 164–173, https://doi.org/10.1016/j.compag.2017.12.002.
[90] M. Khan, N. Muhammad, A. El-Shafie, Wavelet-ANN versus ANN-based model for hydrometeorological drought forecasting, Water 10 (2018) 998, https://doi.
org/10.3390/w10080998.
[91] K.F. Fung, Y.F. Huang, C.H. Koo, M. Mirzaei, Improved svr machine learning models for agricultural drought prediction at downstream of langat river basin,
Malaysia, J. Water Clim. Chang. (2020), https://doi.org/10.2166/wcc.2019.295.
[92] M.M.H. Khan, N.S. Muhammad, A. El-Shafie, Wavelet based hybrid ANN-ARIMA models for meteorological drought forecasting, J. Hydrol. 590 (2020), 125380,
https://doi.org/10.1016/j.jhydrol.2020.125380.
[93] B. Mohammadi, Modeling various drought time scales via a merged artificial neural network with a firefly algorithm, Hydrology 10 (2023), https://doi.org/
10.3390/hydrology10030058.
[94] M.M. Hameed, S.F. Mohd Razali, W.H.M. Wan Mohtar, Z.M. Yaseen, Improving multi-month hydrological drought forecasting in a tropical region using
hybridized extreme learning machine model with Beluga Whale Optimization algorithm, Stoch. Environ. Res. Risk Assess. 37 (2023) 4963–4989, https://doi.
org/10.1007/S00477-023-02548-4/METRICS.

29