Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

Article 1

Assessing the Robustness of Hydrological Models and Machine 2

Learning Techniques in Predicting Extreme Streamflow Events 3

Muhammad Rashid , Zeeshan Ali , Saif Haider , Muhammad Waqar Naseer , Muhammad Atiq Ur Rehman
1 2 1,5, * 1 4
Tariq1,3,4, Aminjon Gulakhmadov5,6,7, Hussein Almohamad8, Motrih Al-Mutiry9, Hazem Ghassan Abdo10 5

1 Centre of Excellence in Water Resources Engineering, University of Engineering and Technology, Lahore 6
54890, Pakistan; m.rashid.6122@gmail.com; ranasaifhaider@gmail.com; mwaqar55786@gmail.com; 7
2 International Water Management Institute, Pakistan; engrzeshanali@yahoo.com; 8
3 Institute for Sustainable Industries & Liveable Cities, Victoria University, P.O. Box 14428, Melbourne 9
8001, Australia. muhammadatiqurrehman.tariq@cdu.edu.au 10
4 College of Engineering and Science, Victoria University, Melbourne 8001, Australia. 11
5 State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and Geography, Chinese 12
Academy of Sciences, Urumqi 830011, China; aminjon@ms.xjb.ac.cn 13
6 Institute of Water Problems, Hydropower and Ecology of the National Academy of Sciences of Tajikistan, 14
Dushanbe 734042 15
7 Department of Hydraulics and Hydro Informatics “Tashkent Institute of Irrigation and Agricultural 16
Mechanization Engineers”, National Research University, Tashkent 60111496, Uzbekistan 17
8 Department of Geography, College of Arabic Language and Social Studies, Qassim University, Buraydah 18
51452, Saudi Arabia; h.almohamad@qu.edu.sa 19
9 Department of Geography and Environmental Sustainability, College of Humanities and Social Sciences, 20
Princess Nourah bint Abdulrahman University, P.O. BOX 84428, Riyadh 11671, Saudi Arabia; mkal- 21
mutairy@pnu.edu.sa 22
10 Geography Department, Faculty of Arts and Humanities, Tartous University, Tartous P.O. Box 2147, 23
Syria; hazemabdo@tartous-univ.edu.sy 24
* Correspondence: ranasaifhaider@gmail.com 25

Abstract: Accurate stream flow prediction plays a crucial role in water resource management, flood 26
control, and environmental planning for complex watersheds. This research delves into an extensive 27
evaluation of streamflow simulation models—Multilayer Perceptron (MLP), Extreme Learning Ma- 28
chine (ELM), Support Vector Machine (SVM), and the Soil and Water Assessment Tool (SWAT) 29
hydrological model spanning the timeline from 1985 to 2018 in the Astore Basin, Pakistan. The pri- 30
mary objective was to comprehensively assess the predictive performances of these models across 31
distinct time segments and gauge their reliability in simulating streamflow dynamics. The study 32
commenced with examining the SWAT model's performance, utilizing the NSE, PBIAS, R2, and 33
RMSE metrics during calibration (1985–2000) and validation (2001–2009) periods. While the SWAT 34
model effectively approximated streamflow, it exhibited limitations in accurately predicting peak 35
Citation: To be added by editorial
and low-flow conditions. Subsequently, the machine learning models (MLP, ELM, and SVM) were 36
staff during production.
scrutinized concerning their performance metrics—R2, NSE, PBIAS, and RMSE—across training 37
Academic Editor: Firstname Last- (1985–1995), validation (1996–2005), and testing (2005–2009) datasets. ELM displayed superior per- 38
name formance during the training phase, boasting a remarkable R2 of 0.94, followed by SVM and MLP. 39
Received: date MLP showcased consistent strength in validation, maintaining an R2 of 0.73, while SVM led in test- 40
Revised: date ing with an R2 of 0.71. Despite their merits, none of the models precisely replicated observed stream- 41
Accepted: date flow patterns, as evidenced by the discrepancies with observed flow and the SWAT model's simu- 42
Published: date lations. This emphasizes the necessity for ongoing refinement and validation to enhance predictive 43
accuracy and ensure closer alignment with real-world hydrological dynamics. This extensive com- 44
parative analysis offers critical insights into the nuances of MLP, ELM, SVM, and SWAT model 45
Copyright: © 2024 by the authors. performances, highlighting their varied strengths and limitations across distinct temporal seg- 46
Submitted for possible open access ments. It underscores the importance of continual refinement and validation to improve predictive 47
publication under the terms and capabilities, essential for accurate streamflow simulations and effective water resource management 48
conditions of the Creative Commons in the Astore Basin and similar hydrological contexts. 49
Attribution (CC BY) license
(https://creativecommons.org/license Keywords: Upper Indus Basin, Machine learning, Water, SWAT Model, EML, SVM, MLP 50
s/by/4.0/).

Sustainability 2024, 15, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/ Sustainability


Sustainability 2024, 15, x FOR PEER REVIEW 2 of 27

51

1. Introduction 52
The role of streamflow is critical in the fields of hydrology, climatology, dam con- 53
struction, floods, drought hazards, and water resources planning and management [1,2]. 54
Hydrological models have been widely used for stream flow prediction due to their ability 55
to simulate the complex interactions between various hydrological processes, such as pre- 56
cipitation, temperature, infiltration, and runoff [3–5]. The hydrological behavior of UIB in 57
response to climatic change has been studied by using various atmospheric input param- 58
eters. Recent studies [6–10] have focused on improving the accuracy and efficiency of hy- 59
drological models for stream flow prediction. One approach is to use machine learning 60
algorithms, such as artificial neural networks (ANNs) and support vector machines 61
(SVMs), to improve the accuracy of hydrological models. For example, [11] used an 62
SWAT-ANN hybrid model to predict stream flow in a river basin and achieved 63
higher accuracy than traditional hydrological models. Another approach studied by Ku- 64
mar et al, [12] is to incorporate remote sensing data into hydrological models. Remote 65
sensing data, such as satellite-based rainfall estimates and land cover data, can improve 66
the accuracy of hydrological models by providing spatially and temporally distributed 67
information on hydrological processes [13–17]. Zhang et al. [16] used remote sensing data 68
to improve the prediction of stream flow in a watershed in the United States. Garee et al. 69
[18] utilized the Soil and Water Assessment Tool (SWAT) model to evaluate the Hunza 70
River Basin, incorporating elevation band and snowmelt algorithms. Hydrologiska 71
Byrans Vattenbalansavdeling (HBV) model was used by Ali et al. [7] to forecast the 72
streamflow in the Hunza River Basin. 73
Furthermore, recent studies [19–22] have focused on developing new hydrological 74
models that can better simulate the complex interactions between hydrological processes. 75
Li and Hongkai [16,23] developed a new hydrological model that can simulate the effects 76
of soil moisture on stream flow, which is an important factor that is often overlooked in 77
traditional hydrological models. The two main techniques for predicting stream flow are 78
empirical and physical-based. Examples of the physical-based approaches include the 79
conceptual hydrological models. Empirical models include statistics and machine learn- 80
ing models. Some assumptions are necessary for the calibration process in a physical- 81
based approach [24]. Precise modeling and forecasting of runoff in alpine glaciated wa- 82
tersheds have become progressively significant for the comprehensive governance and 83
exploitation of water resources. An essential tool for determining streamflow conditions 84
and analyzing the effects of land-use change and climate change in watersheds is the hy- 85
drological model [16,21,25–27]. 86
Several studies including [21,28–35] have compared the performance of EML and 87
SVMs on different datasets. Some studies have shown that EML can achieve similar or 88
better performance than SVMs on large datasets, while other studies have found that 89
SVMs outperform EML on certain types of data. The choice between EML and SVMs de- 90
pends on the nature of the data and the specific problem being addressed. Muhammad et 91
al. [36] evaluated the predictive power of several machine learning models, such as the 92
multilayer perceptron neural network (MLPNN), extreme learning machine (ELM), sup- 93
port vector regression (SVR), Gaussian process regression (GPR), adaptive network-based 94
fuzzy inference system (ANFIS), and Gaussian process regression (GPR) [21,28,37–43]. 95
The models were trained using only precipitation and streamflow data collected at differ- 96
ent time lags. ELMs are often used for classification and regression tasks and have been 97
applied to a wide range of applications, including image classification, speech recognition, 98
and financial forecasting [20,22]. 99
Sustainability 2024, 15, x FOR PEER REVIEW 3 of 27

For the machine learning models to achieve high accuracy, it is crucial to comprehend 100
the fluctuations of streamflow and the significant factors that affect it, such as precipita- 101
tion (P), temperature, and evapotranspiration, which have direct and indirect impacts. 102
Therefore, it is necessary to understand the impact of climate on streamflow to establish 103
a strategic modelling framework. The selection of the best input variables is also crucial 104
for the success of machine learning models. 105
A comparative evaluation of machine learning and hydrological modeling and 106
choice of the most suitable approach for streamflow prediction in the Astore River Basin 107
depends on various factors such as data availability, computational resources, and the 108
level of accuracy required [44–48]. Machine learning algorithms may be more suitable if 109
the available data is limited, and computational resources are limited [10,49,50]. However, 110
hydrological modeling may be more suitable if a more detailed understanding of the hy- 111
drological processes is required, and there is sufficient data available to calibrate and val- 112
idate the model. 113
However, the choice of the most suitable approach depends on various factors, and 114
a careful analysis of these factors is necessary to make an informed decision. In this work, 115
we investigate and evaluate the efficacy of machine learning approaches in simulating 116
discharge in the Upper Indus Basin's Astore River in the HKH Mountains with hydrolog- 117
ical models. In order to identify an effective discharge strategy for glaciated alpine catch- 118
ments with intricate hydrological mechanisms, the outcomes of the two machine learning 119
techniques (ELM, MLP, and SVM) will be compared with the traditional semi-distributed 120
hydrological model Soil and Water Assessment Tool (SWAT). The overall research results 121
will contribute to a better understanding of the high mountain hydrological process and 122
will be useful in future strategic planning for the region's water resources, management, 123
and sustainability. 124

2. Materials and Methods 125

2.1. Study Area 126


The Astore River basin, which is part of the larger UIB sub-basin and situated be- 127
tween 35°20′49″N and 74°51′22″E has been chosen for hydro-glaciological modeling. The 128
basin spans an area of approximately 3988 km2 in the northwest corner of the Himalayan 129
range, with elevations ranging from 1178 m to 8054 m as shown in Figure 1. Glaciers cover 130
roughly 543 km2 of land, including Nanga Parbat, the ninth-highest mountain globally. 131
During winter, the main section of the watershed is typically covered by snow and glaci- 132
ers, which account for 607 km2 or 15% of the entire area. The Rattu Valley (2718 m) had 133
an average annual temperature of 9.9°C between 1998 and 2012, whereas Rama (3179 m), 134
situated at a higher altitude, averaged 2.9°C during the same period [48]. 135
The Astore watershed is a high-elevation region with a peak altitude of 8000 meters 136
above sea level [50]. The climate within the basin varies, with frigid winters and hot sum- 137
mers at lower altitudes, as well as wide temperature fluctuations [13]. Snow covers about 138
30% of the area during the summer and 80% in the winter [44]. The Astore watershed 139
includes a range of terrain types, such as vegetation, rivers, lakes, and glaciers. The vege- 140
tation comprises perennials that can tolerate freezing temperatures, such as broad and 141
needle mixed woods, coniferous and broadleaf forests, and grasslands, including typical 142
steppe, alpine swamp, and shrub meadows. Different researchers [51–54] have reported 143
that the Astore watershed is home to a diverse array of land types. 144
Sustainability 2024, 15, x FOR PEER REVIEW 4 of 27

145

Figure 1. The location of the Astore Basin, and Meteorological and Hydrological Stations used in the Study. 146

2.2. Datasets 147


2.2.1.Hydro-meteorological data 148
The GMRC-WAPDA and SWHP-WAPDA provided daily streamflow data from the 149
Doyian gauging station in the Astor River Basin, as well as meteorological data collected 150
from two automatic weather stations (AWS) located at Rama and Rattu. Additionally, in- 151
formation on climatic variables at the Astor station (which sits at an elevation of 2168 me- 152
ters) was obtained from the Pakistan Meteorological Department (PMD). 153
2.2.2.Topographic Data 154

The Shuttle Radar Topography Mission (SRTM) digital elevation of 30-meter data 155
was generated by the NASA website (https://search.earthdata.nasa.gov/search). It is a rev- 156
olution in digital mapping of the tropics and other areas of the world. According to the 157
criteria of the hydrological model, the study region was divided into seven different ele- 158
vation ranges. The band ranges from a low elevation of 1180 to 2725 meters (5381-8066m). 159
Arc GIS was used to calculate the mean hypsometric elevation (h) of the Astor watershed 160
for each zone. 161
162
Sustainability 2024, 15, x FOR PEER REVIEW 5 of 27

2.2.3. Soil Data 163

Soil data at a spatial resolution of 30 meters were obtained from Food and Agricul- 164
ture Organization website https://data.apps.fao.org/map/catalog/srv/eng/catalog. A total 165
number of four classes was observed in the Astore basin. More details such that soil clas- 166
ses, counts, and weightage shown in Table 1. As observed that Calcaric Fluvisol has 83% 167
and Lithosols has 0.8%. The detailed map of the soil class is shown in Figure 2. 168

Table 1. Statistics on the soil distribution in the Astore basin 169

Sr. No. Soil Name Texture Count Area (Km2) Weightage of Area
1 Be72-2a-3669 Lithosols Loam 36007 0.3601 0.8117
2 Be78-2c-3679 Lithosols Loam 106992 1.0699 2.4119
3 I-B-U-3712 Calcaric Fluvisol Loam 3694765 36.9477 83.2916
4 Glacier-6998 Gleysols UWB 598179 5.9818 13.4848

2.2.4. Land Use Land Cover 170


Land use data with a spatial resolution of 30 meters were procured from 171
https://due.esrin.esa.int/page_globcover.php and the Landsat 8 database. The total num- 172
ber of seven classes of LULC was observed. In the Astore basin, more area is covered by 173
rangeland, glaciers, and barren land. Figure 2 displays the soil, slope, drainage network, 174
and Landcover map and the proportion of watershed area for each Landcover class. 175

176

Figure 2. (a) Details about Land Use Land Cover, (b) shown the Drainage Network Climate Stations and Elevation (c) 177

Soil Classification and (d) Percentage of Slope in the Astore Basin. 178
Sustainability 2024, 15, x FOR PEER REVIEW 6 of 27

2.3. Machine Learning Model 179


In this work, streamflow is estimated using two machine learning techniques Sup- 180
port Vector Machine (SVM) and Extreme Learning Machine (ELM) based on a variety of 181
input variables, including precipitation and minimum and maximum temperatures. Py- 182
thon is used to implement ELM and SVM using a variety of packages, including sci-kit- 183
learn and Keras. 184

2.3.1. Extreme Learning Machine 185


ELM is a type of machine learning algorithm that falls under the umbrella of "ran- 186
domized neural networks" [55]. Extreme Machine Learning is a powerful and efficient 187
machine learning algorithm that can handle very large datasets with high-dimensional 188
features [56,57]. Since its initial introduction by Huang et al. [58], the scientific community 189
has given it a great deal of attention. A kind of feed-forward neural network called an 190
Extreme Learning Machine, or ELM for short, was created to train single hidden layer 191
feedforward networks quickly and effectively. ELMs may train far quicker than tradi- 192
tional neural networks since they employ a single hidden layer feedforward neural net- 193
work with randomly generated weights. The main advantage of EML is its ability to learn 194
quickly without the need for pre-processing or feature selection [56]. It is also capable of 195
handling non-linear relationships between the input and output variables. 196
The methodology for EML involves randomly generating a large number of neurons 197
in the hidden layer, followed by a simple linear regression to learn the weights of the 198
connections between the input and output layers. This process is repeated several times, 199
and the best model is selected based on its performance on a validation dataset. In ELM, 200
the weights of the input layer are randomly initialized and then the output layer is directly 201
computed using the Moore-Penrose pseudoinverse of the hidden layer [21,36]. ELM is 202
very effective in solving various classification and regression problems, and it is often 203
used as a fast and efficient alternative to traditional neural networks as shown in Figure 204
3. 205

206
Figure 3. A Review on Extreme Machine Learning 207

The ELM (Extreme Learning Machine) is a machine learning model renowned for its 208
efficiency in training and prediction tasks, particularly in supervised learning scenarios. 209
Introduced by Guang-Bin Huang, the ELM stands out due to its unconventional approach 210
to training neural networks [59,60]. Unlike traditional gradient-based methods used in 211
training neural networks, ELM adopts a unique strategy where the hidden layer parame- 212
ters are randomly generated, typically as a single layer of neurons, with fixed weights 213
connecting the input layer. During training, the output weights are calculated in a single 214
step using a simple linear algebra computation, enabling rapid learning and significantly 215
reducing training time. ELM has demonstrated exceptional performance in various fields 216
Sustainability 2024, 15, x FOR PEER REVIEW 7 of 27

such as pattern recognition, regression, and classification tasks, attracting attention for its 217
simplicity, speed, and effectiveness in handling large datasets. Consider output layer as 218
w can be determined by solving a system of linear equations, i.e., w ¼ Hyy:Hy is called 219
the Moore–Penrose generalized inverse of the matrix H: The hidden layer matrix, H, can 220
be expressed as: 221

𝑃(𝑎1 , 𝑏2 , 𝑥1 ) ⋯ 𝑃(𝑎𝐿 , 𝑏𝐿 , 𝑥1 )
H= [ ⋮ ⋱ ⋮ ] (1) 222
𝑃(𝑎1 , 𝑏2 , 𝑥𝑚 ) ⋯ 𝑃(𝑎𝐿 , 𝑏𝐿 , 𝑥𝑚 ) 𝑚∗𝐿

Here, 𝑃(𝑎1 , 𝑏2 , 𝑥1 ) is called as the 𝑙𝑡ℎ hidden layer node with reference to 𝑥𝑖 , where 223
𝑎𝐿 = ( 𝑎𝐿1 , … , 𝑎𝐿𝑚 )𝑡 is the weight vector and 𝑏𝑖 indicates the bias to the hidden nodes. 224

2.3.2. Support Vector Machine 225

SVMs are a type of supervised learning algorithm that attempts to find the best deci- 226
sion boundary that separates data points into different classes[29,61]. Support Vector Re- 227
gression is a popular method for prediction tasks. SVR (Support Vector Regression) is a 228
type of regression analysis that uses support vector machines (SVMs) to perform nonlin- 229
ear regression [62]. The decision boundary is chosen such that it maximizes the margin 230
between the different classes of data points [28]. This margin is the distance between the 231
decision boundary and the closest data points from either class. SVMs can handle high- 232
dimensional data and work well when the number of features is greater than the number 233
of samples. 234
For SVM, the methodology involves identifying the support vectors that can separate 235
the data into different classes. The hyperplane that is identified is used to predict the out- 236
put variable based on the input variables [30]. In SVR, the goal is to find a function that 237
best fits a set of data points while keeping the error between the predicted values and the 238
actual values within a certain tolerance. SVR is very effective in solving various regression 239
problems, and it is often used in applications such as streamflow prediction, weather fore- 240
casting, and medical diagnosis. A general and graphical representation of SVM and As- 241
tore basin is shown in Figure 4. Support Vector Regression (SVR) is a powerful machine 242
learning model primarily used for regression tasks, particularly when dealing with com- 243
plex and nonlinear datasets. Based on the Support Vector Machine (SVM) algorithm, SVR 244
aims to find the optimal hyperplane that best fits the data by minimizing the error be- 245
tween predicted and actual values while allowing for a predefined margin of tolerance. 246
Unlike traditional regression techniques, SVR operates by transforming input data into a 247
higher-dimensional space using a kernel function, enabling it to handle nonlinear rela- 248
tionships between variables. SVR seeks to maximize the margin, which represents the dis- 249
tance between the hyperplane and the closest data points, known as support vectors. This 250
approach allows SVR to effectively handle outliers and capture intricate patterns in the 251
data, making it a popular choice for modeling tasks in finance, economics, and other fields 252
where predictive accuracy in complex datasets is crucial. SVR seeks a regression function 253
f(x) by fitting the input samples 𝑿𝒊 𝝐 𝑹𝒏 ,.where 𝑤̇ 𝝐 𝑹𝒏 and 𝒃 𝝐 𝑹𝒏 as: 254

f(x) = 𝑤 𝑡 x + b (2) 255

The primal problem of SVR can be stated as 256

𝟏
𝐦𝐢𝐧 𝒘𝒕 𝒘 + 𝑪(𝒆𝒕 Ԑ + 𝒆𝒕 Ԑ∗),s.t.𝒚𝒊 − (𝒙𝒕𝒊 + 𝒃) ≤ Ԑ + Ԑ𝒊 , 𝒙𝒕𝒊 𝒘 + 𝒃 − 𝒚𝒊 ≤ Ԑ + Ԑ∗𝒊 . And 257
𝒘,𝒃,Ԑ,Ԑ∗ 𝟐
Ԑ∗𝒊 ≥ 𝟎, Ԑ∗𝒊 ≥ 𝟎 𝒇𝒐𝒓 𝒊 = 𝟏, 𝟐, … . . , 𝒎. (3) 258

where, C > 0,℮ and Ԑ >0 are the trade-off parameter, vector of ones, and input pa- 259
rameters, respectively. Ԑ = (Ԑ𝟏 , … . . , Ԑ𝒎 )𝒕 and Ԑ∗ = ( (Ԑ∗𝟏 , … … , Ԑ∗𝒎 )𝒕 are the slack vectors. 260
The decision regression function of SVR for any input sample 𝑿𝒊 𝝐 𝑹𝒏 is 261
Sustainability 2024, 15, x FOR PEER REVIEW 8 of 27

ƒ(x) = ∑𝑚
𝑖=1(∝1𝑖 -∝2𝑖 )k(x, 𝑥𝑖 ) + b (4) 262

where ∝1 and ∝2 are Lagrangian multipliers. 263

264
Figure 4. (a) General Support Vector Machine Margin, (b) Graphical Representation of Astore Basin 265

2.3.3. Multi-Layer Perceptron (MLP) 266


Multilayer Perceptron (MLP) stands as a pivotal architecture in artificial neural net- 267
works, known for its versatility in handling complex data patterns. This class of neural 268
network comprises multiple layers of interconnected nodes, allowing it to capture intri- 269
cate relationships and nonlinearities within datasets. MLP excels in various domains, from 270
image and speech recognition to financial forecasting and hydrological modeling, owing 271
to its adaptability and capability to learn and generalize from diverse datasets. Its archi- 272
tecture, with input, hidden, and output layers, facilitates robust learning, making MLP a 273
potent tool for tackling intricate problems and achieving high prediction accuracies. De- 274
spite its computational demands, MLP's capacity to model complex relationships in data 275
remains a cornerstone in modern machine learning applications. The three layers that 276
make up the fundamental structure of MLP are the input layer, the hidden layer, and the 277
output layer Figure 5. n-nh1-no, where n, nh1, and no denote the number of neurons in 278
the input layer, first hidden layer, and output layer, respectively, is the primary represen- 279
tation of the MLP layers. The number of hidden layers can be changed by the need for an 280
appropriate architecture that includes a large number of connections and processing com- 281
ponents. A typical MLP's neurons are built with many algorithms in mind to tackle diffi- 282
cult tasks more effectively and satisfactorily. While there are calculation nodes in the out- 283
put and hidden levels, there are none in the input layer. 284
Sustainability 2024, 15, x FOR PEER REVIEW 9 of 27

285
Figure 5. An Illustration of the MLP Model Consists of Three Layers: The Input Layer, the Hidden Layer, and the Out- 286

put Layer (left). A Representation of a Common Processing Element with Signal Flow (Right). 287

For a given network architectural need, the number of hidden layers and neurons in 288
each hidden layer can be changed. Converting numerous inputs into one or more outputs 289
is one of MLP's most important functions. Assuming that xi (i=1, 2..., m) are the inputs for 290
a pre-organized model, the corresponding weights for which are wi (i=1, 2..., m) are vari- 291
ables that will be modified subsequently by error algorithms. Equation 5 is used to define 292
the net input to a single node in such circumstances. Equation 6 states that the activation 293
function "f" transforms the net input into an output. Any node's output will function as 294
an input for the computational node after it. 295
𝑁𝑒𝑡 = ∑𝑚𝑖=1 𝑥𝑖 𝑤𝑖 (5) 296

𝑦 = 𝑓(𝑁𝑒𝑡) (6) 297

Two learning mechanisms unsupervised and supervised have been used to establish 298
the network's weights. The current study uses supervised learning, where the user is fully 299
aware of the input and output patterns of the network's usual design. In MLP, the mean 300
square error between known and unknown outputs is minimized by adjusting the 301
weights in each iteration using the backpropagation algorithm (BPA). 302
The dataset's observed values are the known output, while the network-computed 303
values from the same dataset are the unknown output. Two computations are used to 304
adjust weights in each iteration or epoch: feed-forward calculations and back-propagation 305
of errors, which is sometimes referred to as the mean square error (MSE). Each layer, 306
which is represented by Equations 7 through 10, performs the feed-forward computation. 307
The hidden layer's the jth node's net input is provided by: 308

𝑁𝑒𝑡ℎ𝑗 = ∑𝑛𝑖=1 𝑤ℎ𝑗𝑖 𝑥𝑖 (7) 309

There is always a connection weight between two nodes. Not every node has to be a 310
computational node to build connection weights. Even though there is no computation on 311
the input layer nodes, Equation 7 contains the connection weight, which is the weight 312
between the ith node of the input layer and the jth node of the hidden layer. Equation 8 313
now provides the output of this specific node in the hidden layer. 314
Sustainability 2024, 15, x FOR PEER REVIEW 10 of 27

ℎ𝑖 = 𝑓(𝑁𝑒𝑡ℎ𝑗 ) (8) 315

Moreover, Equation 9 provides the net input to the kth node of the output layer. 316

𝑛
𝑁𝑒𝑡𝑦𝑘 = ∑𝑖=1
𝑘
𝑤𝑜𝑘𝑗 ℎ𝑗 (9) 317

where wokj is the weight of the link between the computational jth hidden layer node 318
and the computational kth output layer node. Equation 10, the output layer's ultimate 319
result, is obtained from the kth node; 320
𝑦𝑘 = 𝑓(𝑁𝑒𝑡𝑦𝑘 ) (10) 321

2.4. Hydrological Model 322


The Agricultural Research Service of the United States Department of Agriculture 323
created the process-based, semi-distributed, spatially explicit, time-continuous, Hydro- 324
logical Response Unit (HRU)-based SWAT model [63]. The major basins are divided into 325
smaller sub-basins by the SWAT model, which increases the model's accuracy and de- 326
pendability by providing more precise geographical features [64]. The SWAT model was 327
created to anticipate and simulate how decisions and practices related to land manage- 328
ment and agriculture might affect water resources across a range of basin sizes, both in 329
terms of quantity and quality [64]. Furthermore, scenario-based simulations utilizing the 330
SWAT model are primarily used to investigate the hydrological responses to land use and 331
climate change [65,65]. Because of the SWAT model's computational efficiency, simulating 332
huge basins or various management options is simple [66]. In this study, streamflow in 333
UIB was simulated using Arc SWAT version 2012, revision 664. Generally speaking, the 334
SWAT model is widely utilized in soil erosion, evapotranspiration, rainfall-runoff process 335
simulation, and water quality monitoring. SWAT can evaluate the effects of climate 336
change on soil, meteorological, and water resources as well as the movement of nutrients 337
and sediments under different land use and cover (LULC) conditions [67–73]. The SWAT 338
model operates by first dividing whole basins into many sub-basins, and then, using a 339
digital elevation model (DEM), into HRUs, which are unique combinations of soil type, 340
slope, and LULC [74,75]. Two approaches are used by the SWAT model to create HRUs: 341
one way uses soil and LULC data to generate HRUs for each sub-basin, and another 342
method generates numerous HRUs based on threshold values, as described by Arnold 343
et al. [63]. The current study employed 10%, 20%, and 10% criteria for land, soil, and slope, 344
respectively, as advised by Setegn et al. [76]. The number of sub-basins (HRUs) created by 345
the SWAT model for Doyian 5 (8) following the successful overlaying of the soil, slope, 346
and LULC datasets. The land and routing phases are the names of the two separate stages 347
that make up the SWAT paradigm. During the land phase hydrological component, the 348
SWAT model uses the Soil Conservation Service (SCS) approach to simulate surface run- 349
off for each HRU based on the daily precipitation. An alternate technique to SCS in the 350
SWAT model for simulating surface runoff is the green and ampt infiltration approach. 351
Precipitation inputs are needed for the Green and Ampt infiltration methods on a less 352
frequent basis. Using Muskingum or Variable storage approaches, simulated streamflow 353
is directed through the network of streams and rivers to the basin outflow during the 354
routing phase. 355
LULC and soil maps, elevation data, and soil attributes are required as input data for 356
the SWAT model before streamflow simulation. Downloaded from USGS Earth Explorer 357
(https://earth explorer.usgs.gov/), a DEM with a resolution of 30 m obtained from a shuttle 358
radar topographic mission (SRTM) was used for the current study. DEM provided the 359
basin delineation and topographic parameter retrieval needed for the SWAT model. The 360
supervised classification approach was used to create the LULC map for 2020, which can 361
be observed in Figure 1d. The LULC map was created using imagery from the Landsat-8 362
Sustainability 2024, 15, x FOR PEER REVIEW 11 of 27

satellite. The SVM approach was used to create the LULC map. Rahman et al. [77] pro- 363
vide a comprehensive explanation of the SVM algorithm. Based on Rahman et al. (2020a), 364
the Food and Agriculture Organization (FAO) produced a soil map 365
(http://www.fao.org/soils-portal/soil-survey/ soil-maps-and-databases/faounesco-soil- 366
map of-the-world/ en/) with a resolution of 1: 5,000,000 from which the soil data for the 367
study region was taken. The retrieved soil map from FAO and the Harmonized World 368
Soil Database v1.2 were utilized to get the necessary soil parameters for SWAT. Addition- 369
ally, the current study simulates streamflow using the SCS technique, Penman-Monteith 370
equation, and variable storage method the default settings of the SWAT model. A water 371
balance technique is used to simulate streamflow using the SWAT model. This strategy is 372
dependent on temperature and precipitation inputs and may be shown as follows: 373

𝑆𝑊𝑡 = 𝑆𝑊0 + ∑𝑡𝑖=1(𝑃𝑑𝑎𝑦 − 𝑄𝑆𝑢𝑟𝑓 − 𝐸𝑎 − 𝑊𝑠𝑒𝑒𝑝 − 𝑄𝑔𝑤 ) (11) 374

Where; 𝑆𝑊𝑡 final soil water (mm), 𝑆𝑊0 initial soil water (mm), 𝑃𝑑𝑎𝑦 precipitation 375
at the time of i (mm), 𝑄𝑆𝑢𝑟𝑓 surface runoff (mm), 𝐸𝑎 evapotranspiration (mm), 𝑊𝑠𝑒𝑒𝑝 376
water flow to the unsaturated zone (mm) and 𝑄𝑔𝑤 water flow of watershed from under- 377
ground (mm). 378

2.4.1. Calibration and validation of SWAT model 379


The Sequential Uncertainty Fitting program method (SUFI-2) was used to automati- 380
cally calibrate the SWAT-CUP model, which was then used to bring the most prominent 381
streamflow characteristics to the fore utilizing the Doyian monitoring station at the Astore 382
Basin [68]. The R2, Modified Coefficient of Determination (b R2), NSE, Modified Nash– 383
Sutcliffe Efficiency (MNS), Ratio of the Standard Deviation of Observations to the Root- 384
Mean-Square-Error (RSR), Ranked Sum of Squares (SSQR), Kling–Gupta Efficiency 385
(KGE), PBIAS, Multiplicative Form of the Square Error (MSE), Summation Form of the 386
Square Error, and the χ2 are among the eleven distinct objective functions used by the 387
SUFI-2 algorithm. Altogether, these objective functions offer a strict framework for pa- 388
rameter calibration and model evaluation, empowering researchers to decide wisely and 389
improve models to faithfully capture intricate real-world processes. Although there are 390
other objective functions available for optimizing parameter ranges, the NSE was the only 391
objective function that was the subject of our investigation. The fundamental capability of 392
NSE to accurately represent temporal dynamics—a crucial component of hydrologic com- 393
putation was the impetus for this choice. Furthermore, NSE is one of the most extensively 394
used statistical measures in the area, enjoying a well-established and important position. 395
Our aim in concentrating on NSE was to make sure that temporal calibration was accu- 396
rately and thoroughly assessed within the parameters of our investigation. To calibrate 397
the SWAT-CUP model, monthly streamflow data from 1985 to 2005 were utilized. The 398
primary goal of calibration is to use the statistical performance metrics R2, NSE, PBIAS, 399
and RMSE to calculate the goodness-of-fit index (GFI) between the measured and mod- 400
eled data. The most advantageous value is 1 for R2 and NSE and 0 for PBIAS and RMSE. 401
Table 2. Best-suited Calibrated Parameters of SWAT for Astore Basin 402

Fitted Min/Max
Sr # Parameters Names Fitted Value Min/Max value
Value value
1 R__CN2.mgt Curve Number -0.290551 -0.29552/0.29 -0.309214 -0.44792/-0.29
2 V__ALPHA_BF.gw Base Flow Alpha Factor 0.299644 0.296438/0.3 0.259477 0.258026/0.26
3 V__GW_DELAY.gw Groundwater Delay Time 456.670135 453.4113/486 474.558502 467.4962/486
4 V__GW_REVAP.gw Groundwater "revap" coefficient 0.14879 0.145967/0.15 0.147758 0.143239/0.15
5 V__GWQMN.gw Threshold Depth Shallow Aquifer 1.453138 1.224427/1.78 1.178502 1.098336/1.9
6 V__EPCO.hru Plant Uptake Compensation Factor 0.999923 0.999234/1 0.992477 0.863216/1
Sustainability 2024, 15, x FOR PEER REVIEW 12 of 27

Soil Evaporation Compensation


7 V__ESCO.hru 0.956995 0.856651/1 0.991113 0.966466/1
Factor
8 V__OV_N.hru Manning's "n" 0.021365 0.019294/0.04 0.037785 0.036494/0.04
9 V__HRU_SLP.hru Average slope steepness 0.039696 0.036963/0.04 0.041175 0.042702/0.04
10 V__SLSUBBSN.hru Average slope length 51.611473 40.67941/150 112.416672 40/150
Manning's "n" value for the main
11 V__CH_N2.rte 0.255184 0.252427/0.28 0.214776 0.184971/0.25
channel
Effective hydraulic conductivity in
12 V__CH_K2.rte 82.565231 74.250473/7 90.842766 68.5389/90.79
main channel
Available Water Capacity of Soil
13 R__SOL_AWC.sol 0.063963 0.047926/0.08 0.050149 0.037049/0.08
Layer
14 R__SOL_BD.sol Moist bulk density 2.431435 1.814353/2.5 0.956 0.9/2.5
15 V__SURLAG.bsn Surface Runoff Lag Coefficient 9.479791 8.849303/9.75 8.127467 7.628319/8.75
Maximum melt rate for snow
16 V__SMFMX.bsn 15.321431 15.2314/15.36 14.290265 14.0714/14.36
during year
17 V__SFTMP.bsn Snowfall temperature 7.103765 7.079665/7.16 7.446956 7.259322/7.46
18 V__CANMX.hru Maximum Canopy Storage 7.991878 7.602683/8.9 7.813922 7.78036/8.1

3. Results 403

3.1. Streamflow Simulation using SWAT Model 404


A semi-distributed model with a hydrological physical basis, SWAT was first cali- 405
brated to lower uncertainty in the modeled values. At the Doyian monitoring station of 406
the Astore Basin, the 26-year historical record, which spans from 1985 to 2010, was divided 407
into calibration (1985–2000) and validation (2001–2009) periods. The calibration technique 408
employed the NSE as the goal function. Sensitivity analysis was done after the model was 409
calibrated to identify the parameters that most affect streamflow. Abbaspour et al. [78] 410
provide a thorough description of model calibration techniques. Your reference is sug- 411
gested. Table 2 displays the 18 most crucial parameters used for model calibration. Table 412
3 shows how the model's performance in forecasting the water-shed conditions was as- 413
sessed using four statistical performance indicators: R2, NSE, PBIAS, and RMSE. In this 414
work, four statistical indicators R2, NSE, PBIAS, and RMSE were used to evaluate the 415
SWAT model's performance. For the calibration and validation periods, the values of R2, 416
NSE, PBIAS, and RMSE were, respectively, 0.80 and 0.71, 0.73 and 0.66, 15.7 and 17.3, and 417
79.81 and 106.26. Figure 6 makes it clear that the SWAT model was able to monitor the 418
observed streamflow during the pertinent periods. Nevertheless, during calibration, sev- 419
eral peaks were underestimated, but this was less of an issue during validation. Further- 420
more, the SWAT model overestimated the low-flow circumstances and underestimated 421
the peak flow. Furthermore, the flow under general circumstances was accurately approx- 422
imated using the SWAT model. In summary, the SWAT model fared well in the Astore 423
Basin understudy. 424

Table 3. Performance evaluation of SWAT Model during Calibration and Validation Period using NSE, PBIAS, R2, and 425

RMSE Statistical Indices 426

R2 NSE PBIAS RMSE


Calibration(1985-2000) 0.8 0.73 15.7 79.81
Validation (2001-2009) 0.71 0.66 17.3 106.26
427
Sustainability 2024, 15, x FOR PEER REVIEW 13 of 27

428
Figure 6. Calibration and Validation of SWAT Model for Astore Basin 429

3.2. SWAT Model Results 430


It was clear that the Soil and Water Assessment Tool (SWAT) was reliable in glaciated 431
alpine catchments when it comes to calculating discharge levels. Through the integration 432
of critical spatial data related to soil properties, slope, and land use, SWAT proved that it 433
could generate outputs for runoff that were remarkably accurate. The accuracy of the 434
model's streamflow estimation was notably enhanced by its thorough approach, which 435
included a hydrological response unit (HRU) study and watershed delineation. However, 436
the model's complex parameter calibration procedures and dependence on extremely spe- 437
cific geographical data presented difficulties, especially when handling large-scale appli- 438
cations or areas with sparse data. It is noteworthy to emphasize that the average entire 439
basin flow in SWAT's simulations was 130.83 cumecs, which deviates somewhat from the 440
145.8 cumecs measured flow rate. The variation in estimation points to possible con- 441
straints on the model's ability to accurately forecast discharge volumes in this particular 442
situation. A comparison of the observed and SWAT-simulated flow during one year is 443
shown in Figure 7. 444

Observed flow SWAT

250
200
150
100
50
0

445
Figure 7. Observed and Simulated flow of the Astore Basin 446
Sustainability 2024, 15, x FOR PEER REVIEW 14 of 27

3.3. Performance Metrics of Training Data (1985-1995) 447


When trained on data from 1985 to 1995, the evaluation of three machine learning 448
models Multilayer Perceptron (MLP), Extreme Learning Machine (ELM), and Support 449
Vector Machine (SVM)—reveals interesting trends in how well they perform. To assess 450
the models' predicted accuracy and training dataset fit, a variety of performance indica- 451
tors were employed in a thorough evaluation process. Starting with their coefficient of 452
determination (R2), which is a metric that shows how much variation in the observed data 453
the model was able to capture, ELM was the clear winner with an outstanding R2 value of 454
0.94. The extraordinary alignment between the ELM model's predictions and the actual 455
observed values in the training data is indicated by the high R2 score. SVM trails some- 456
what behind ELM but has an excellent predictive fit with an R2 of 0.85, following closely 457
behind ELM. MLP has a reasonable R2 value of 0.73, however, it is not as strong as ELM 458
and SVM, suggesting a worse fit for the training set. 459
Furthermore, ELM's superiority is further supported by the Nash-Sutcliffe Efficiency 460
(NSE) scores, which illustrate the models' predictive power by contrasting the actual val- 461
ues with the model predictions. ELM has the best accuracy in predicting the observed 462
values in the training dataset, with an NSE score of 0.46. With an NSE of 0.41, SVM is not 463
far behind, suggesting a respectable prediction performance. Compared to ELM and SVM, 464
MLP falls significantly behind with an NSE of 0.4, suggesting a little lower ability to accu- 465
rately predict observed values. Additional information is provided by the Percent Bias 466
(PBIAS) values, which show the average propensity of the models to either overestimate 467
or underestimate observed values. In this case, MLP reports the highest PBIAS of 872.3, 468
indicating a significant bias in its training data predictions. On the other hand, compared 469
to MLP, ELM, and SVM show less bias, with PBIAS values of 643.23 and 726.45, respec- 470
tively, indicating a more balanced tendency in their predictions. 471
Furthermore, ELM's superiority is further supported by the Root Mean Squared Er- 472
ror (RMSE) values, which measure the average size of errors between predicted and ob- 473
served values. With an RMSE of 576.27, ELM has the lowest prediction errors on the train- 474
ing set. With an RMSE of 929.16, SVM comes in second, while MLP has the highest RMSE 475
of 1264.8 out of all the models, showing more prediction errors than ELM and SVM. Even 476
with these strong conclusions drawn from the training set, care must be used when ex- 477
trapolating these conclusions to brand-new, untested data. To choose a model that not 478
only performs well on training data but also generalizes well for accurate predictions in 479
real-world circumstances, more assessment and validation using independent datasets 480
are essential. 481

Table 4. Performance Metrics for various Models on the Training Data. 482

Training Data (1985-1995)


Machine Learning Techniques R2 NSE PBIAS RMSE
MLP 0.73 0.4 872.3 1264.8
ELM 0.94 0.46 643.23 576.27
SVM 0.85 0.41 726.45 929.16

3.4. Performance Metrics of Validation Data 483


Table 5 displays the performance characteristics of three different machine learning 484
models, which were evaluated using validation data from 1996 to 2005: Multilayer Per- 485
ceptron (MLP), Extreme Learning Machine (ELM), and Support Vector Machine (SVM). 486
Metrics including R2, Nash-Sutcliffe Efficiency (NSE), Percent Bias (PBIAS), and Root 487
Mean Squared Error (RMSE) were used to assess these models. With the highest R2 value 488
of 0.73 among these models, MLP stands out as having a reasonably good fit for the vali- 489
Sustainability 2024, 15, x FOR PEER REVIEW 15 of 27

dation data. SVM performs competitively in terms of fitting the data, coming a close sec- 490
ond with an R2 of 0.71. On the validation dataset, ELM shows a little lower R2 of 0.66 but 491
maintains a decent degree of prediction accuracy. 492
With an NSE of 0.55, MLP leads the group in terms of predicted accuracy on the val- 493
idation data. NSE scores are a measure of predictive power. SVM comes in second place 494
behind MLP, exhibiting a respectable prediction performance with an NSE of 0.53. Com- 495
paring ELM to MLP and SVM, it shows a somewhat poorer prediction accuracy (NSE of 496
0.48), putting it significantly behind. With a PBIAS of 481.77, MLP has the lowest bias 497
among the three models, which is measured as the average prediction bias. With a PBIAS 498
of 502.34, SVM exhibits a similar degree of balanced predictions, placing it a close second. 499
In comparison to MLP and SVM, ELM has a higher PBIAS of 578.91, suggesting a com- 500
paratively bigger bias in its predictions. 501
Additionally, RMSE analysis, which quantifies the average magnitude of prediction 502
errors, reveals that ELM has the lowest RMSE (287.91), suggesting that the validation data 503
had fewer prediction mistakes. With an RMSE of 1034.28, MLP comes in second, while 504
SVM, with an RMSE of 1177.53, is the highest of the three models, indicating more predic- 505
tion errors than ELM and MLP. While these performance indicators offer insightful infor- 506
mation about how well the models perform on the validation data, one should be cautious 507
when thinking that the efficacy of this dataset will transfer to fresh, untested data. To de- 508
termine the model's resilience and applicability for actual prediction tasks, more testing 509
and validation on other datasets such as an alternative test set are essential. 510

Table 5. Performance Metrics for various Models on the Validation Data. 511

Testing Data (1996-2005)


Machine Learning Techniques R2 NSE PBIAS RMSE
MLP 0.73 0.55 481.77 1034.28
ELM 0.66 0.48 578.91 1177.53
SVM 0.71 0.53 502.34 287.91

3.5. Performance Metrics of Testing Data 512


Using testing data from 2005 to 2009, Table 4 presents the performance characteristics 513
of three different machine learning models: Multilayer Perceptron (MLP), Extreme Learn- 514
ing Machine (ELM), and Support Vector Machine (SVM). A number of measures, includ- 515
ing R2, Nash-Sutcliffe Efficiency (NSE), Percent Bias (PBIAS), and Root Mean Squared Er- 516
ror (RMSE), were used to evaluate the models. With the highest R 2 value of 0.71 among 517
these models, SVM stands out as having a reasonably good fit to the test data. With a 518
competitive R2 of 0.69, which indicates a commendable performance in fitting the testing 519
dataset, MLP trails closely behind. Despite having a little lower R2 of 0.62, ELM is still able 520
to predict the testing results with a decent degree of accuracy. 521
The models' capacity to forecast is shown by their NSE scores, of which MLP tops the 522
group with an NSE of 0.58, indicating better predictive accuracy on the testing data. With 523
an NSE of 0.60, SVM has a respectable predictive performance, trailing closely behind. 524
With an NSE of 0.54 lagging slightly behind MLP and SVM, ELM has comparatively 525
poorer predicting accuracy. With a PBIAS of 402.88, which represents the average bias in 526
predictions, SVM has the lowest bias of the three models. With a PBIAS of 430.12, MLP 527
comes in just behind, suggesting a similar degree of balanced forecasts. When compared 528
to SVM and MLP, ELM has a somewhat higher PBIAS of 491.65, suggesting a compara- 529
tively greater bias in its predictions. 530
Furthermore, analyzing RMSE a metric that quantifies the average magnitude of pre- 531
diction mistakes shows that SVM has the lowest RMSE of 339.79, suggesting that the test- 532
ing data had fewer prediction errors. Of the three models, ELM has the highest RMSE 533
(390.78), indicating higher prediction errors than SVM and MLP, which come in second 534
Sustainability 2024, 15, x FOR PEER REVIEW 16 of 27

and third, respectively. It's important to understand that, despite these insights into the 535
models' performance on the testing data, the models' efficacy on this dataset may not al- 536
ways translate to performance on fresh, untested data. The resilience and applicability of 537
the model for practical forecasting tasks must be determined through additional valida- 538
tion and testing on separate datasets. 539

Table 6. Performance Metrics for Various Models on the Testing Data. 540

Validation Data (2005-2009)


Machine Learning Techniques R2 NSE PBIAS RMSE
MLP 0.69 0.58 430.12 350.89
ELM 0.62 0.54 491.65 390.78
SVM 0.71 0.6 402.88 339.79

3.6. Extreme Learning Machines (ELM) Results 541


Prospective outcomes were obtained when Extreme Learning Machines (ELM) were 542
utilized to evaluate discharge in glaciated alpine catchments. Without requiring any data 543
processing, ELM was able to accurately estimate an average entire basin flow of 140.02 544
cumecs, demonstrating its efficiency. This capacity to manage sophisticated, multi-dimen- 545
sional information effectively demonstrated the model's quick recognition and learning 546
of complex patterns found in the data. The model's predictions, which showed remarkable 547
accuracy and dependability, nearly matched the measured flow rates, which averaged 548
145.8 cumecs. ELM's exceptional dexterity in interpreting nonlinear relationships between 549
input and output variables allowed for quick model training and very accurate predic- 550
tions. However, even while ELM is adept at managing large datasets and intricate inter- 551
actions, it could have trouble capturing the subtle intricacies present in intricate hydro- 552
logical systems. This difference might explain the little difference between the observed 553
and projected flow rates. A comparison of the annual time frame's observed and ELM- 554
simulated flow is shown in Figure 8. 555

Observed flow ELM

250
200
150
100
50
0

556

Figure 8. Performance Comparison of Observed Flow with ELM Models. 557

558
Sustainability 2024, 15, x FOR PEER REVIEW 17 of 27

3.7. Support Vector Machine (SVM) Results 559


Support Vector Machine (SVM) and Support Vector Regression (SVR) were evalu- 560
ated, and the results showed how well they performed in terms of discharge level predic- 561
tion. These models demonstrated exceptional skill in handling high-dimensional, compli- 562
cated data and determining complex correlations between variables. SVM, which is espe- 563
cially well-known for its skill in determining the best decision limits, effectively calculated 564
discharge in difficult hydrological situations. With an average flow rate of 145.8 cumecs, 565
the model's projections astonishingly closely matched the measured flow rate, showing 566
an average entire basin flow of 134.18 cumecs. The precision of SVM's discharge estima- 567
tion is shown by this precise alignment. It's crucial to remember that SVM's processing 568
requirements might be a limitation, especially when working with large datasets or real- 569
time applications. The contrast between the observed and SVM-simulated flow through- 570
out the year is shown in Figure 9. 571

SVM Observed flow

250
200
150
100
50
0

572
Figure 9. Performance Comparison of Observed Flow with SVM Models. 573

3.8. Multilayer Perceptron (MLP) Results 574

The evaluation of the Multilayer Perceptron (MLP) model demonstrated how well it 575
predicted discharge levels. These models demonstrated exceptional proficiency in man- 576
aging high-dimensional, complicated data, and successfully identifying complex relation- 577
ships between variables. MLP, which is well-known for its capacity to draw the best pos- 578
sible decision boundaries, proved effective in determining discharge levels even in com- 579
plex hydrological environments. The model's estimates, which averaged 145.8 cumecs, 580
nearly matched the average flow rate for the whole basin, which was 136.35 cumecs. This 581
is significant. This exceptional consistency confirms MLP's dependability for hydrological 582
forecasts and highlights its accuracy in discharge estimation. However, it's important to 583
recognize that the processing requirements of MLP may provide challenges, particularly 584
when working with large datasets or in real-time applications. 585
A further indication of the model's accuracy in simulating real-world flow dynamics 586
is the comparison of observed and MLP-simulated flow during one year, as shown in the 587
picture. This convergence demonstrates that MLP may be a useful tool for forecasting dis- 588
charge levels in hydrological scenarios, offering insightful information for the manage- 589
ment of water resources and decision-making procedures. 590
Sustainability 2024, 15, x FOR PEER REVIEW 18 of 27

MLP Observed flow

250
200
150
100
50
0

591
Figure 10. Performance Comparison of Observed Flow with MLP Models. 592

3.9. Comparative Analysis 593


When streamflow simulations from 1985 to 2018 were analyzed, interesting differ- 594
ences in the predictive performances of the Soil and Water Assessment Tool (SWAT) hy- 595
drological model, Multilayer Perceptron (MLP), Extreme Learning Machine (ELM), Sup- 596
port Vector Machine (SVM), and other models were found. Upon examining the training 597
data spanning from 1985 to 1995, ELM was the most exceptional performer, with an 598
astounding R2 of 0.94. SVM came in second at 0.85 and MLP third at 0.73. Additionally, 599
ELM and SVM had lower RMSE values than MLP, indicating that the predictions made 600
during this phase were rather accurate. 601
Moving on to the later validation data, which covered the years 1996 to 2005, MLP 602
continued to be strong with an R2 of 0.73. SVM trailed closely behind in parallel, with an 603
R2 of 0.71. ELM, on the other hand, showed larger RMSE values, indicating somewhat 604
worse prediction accuracy, even with its good R2 of 0.66. Testing data from 2005 to 2009 605
was evaluated, and SVM emerged as the winner with the greatest R2 of 0.71, closely fol- 606
lowed by MLP at 0.69. Both versions demonstrated reliable performance. Meanwhile, 607
ELM showed an R2 of 0.62, which suggests somewhat lower prediction accuracy at this 608
stage. Using the mean streamflow data from the model across the whole duration, the 609
averages produced by MLP, ELM, and SVM were 136.35, 140.02, and 134.18 cumecs, re- 610
spectively. Comparatively speaking, the measured flow was 145.8 cumecs, whereas the 611
SWAT hydrological model predicted a flow value of 130.83 cumecs. 612
However, none of the models accurately reproduced the observed streamflow pat- 613
terns when compared to the observed flow and SWAT model. This variation highlights 614
the shortcomings of the models and highlights the need for continuous improvement, 615
such as the use of ensemble techniques or substantial parameter adjustment, to improve 616
their forecast accuracy and more closely mimic the dynamics of streamflow in the real 617
world. This thorough comparison analysis highlights the subtle differences in MLP, ELM, 618
and SVM performances over various periods. From 1985 to 2018, the constant pursuit of 619
improvement and rigorous verification was crucial for enhancing the models' forecasting 620
skills, which are necessary for precise streamflow simulations in hydrological forecasting 621
and efficient water resource management. 622
Sustainability 2024, 15, x FOR PEER REVIEW 19 of 27

MLP SVM ELM SWAT Model Observed flow

2000
2018 250 2001
2017 200 2002

2016 150 2003


100
2015 2004
50
0
2014 2005

2013 2006

2012 2007
2011 2008
2010 2009
623

Figure 11. Performance Comparison of Observed Flow with MLP, SVM, ELM, and SWAT Model. 624

4. Discussion 625
A crucial component of planning and managing sustainable water resources is accu- 626
rate hydrological process modeling. For simulating hydrological processes, data-driven 627
and hydrological models are typically employed. In this study, machine learning methods 628
outperform the SWAT model for forecasting streamflow under both normal and abnormal 629
circumstances. The literature [52,72,73,79–82] supports the findings of the current study 630
on the estimation of the flow forecast using the SWAT model. The methods used to gen- 631
erate streamflow differ greatly at low-, medium-, and high-flow times. High-flow events 632
are mostly brought on by heavy storm rainfall, whereas low-flow events are mainly pro- 633
duced by base flow [52,83]. SWAT models are poor at properly simulating very high 634
stream flows, but they excel at estimating low flows. As Kim et al. [84] have shown, ma- 635
chine learning approaches, on the other hand, may be less accurate in forecasting the low- 636
est values but may produce superior results for extremely high values. Consequently, ba- 637
sin streamflow simulation may be effectively carried out using these models. It is advised 638
that in research involving extreme hydrologic events like floods, machine learning models 639
be used to mimic high-flow occurrences. In hydrological management studies, however, 640
when low-flow episodes are of higher importance, the SWAT model would be used. The 641
following factors could be the cause of the SWAT model's poor performance: (1) the most 642
sensitive factors identified may have an impact on the model's performance [85,86] (Cibin 643
et al. 2010); (2) the critical snow-specific characteristics identified may also lead to uncer- 644
tainty in the model's output; and (3) the flow from each sub-basin accumulates within 645
itself before it reaches the Indus River, making it challenging to track the hydro peak of 646
the stream flow precisely. Several studies have revealed that the observed peak-flow in- 647
efficiency can be attributed to the formulation of SWAT [87–92]. Consequently, the ma- 648
chine learning (ML) models outperform the SWAT model in terms of streamflow simula- 649
tion. 650
Although there is still a propensity to underestimate high streamflow estimates, the 651
study's findings indicate that applying machine learning models can help reduce mistakes 652
in this regard. This issue originates from the fact that training datasets mostly contain 653
medium and low flow values, while high flow values are rare. Comparable problems have 654
been documented in studies conducted by Jimeno-Sáez et al. [92]. These methods have a 655
few shortcomings, which are addressed below: The physical features of the watershed are 656
unrelated to ML approaches. The lumped technique, which is the foundation of the ML 657
Sustainability 2024, 15, x FOR PEER REVIEW 20 of 27

approach, overlooks the variability of sub-catchment factors. Moreover, watershed dy- 658
namics such as soil moisture, runoff production, pollutants export, etc., cannot be in- 659
cluded in empirical models that use machine learning approaches. The SWAT model is a 660
process-based, highly data-intensive model that needs a variety of input factors, including 661
soil type, hydrometeorological data, land use/cover, and DEM, for analysis. Unfortu- 662
nately, due to inadequate station distributions, socioeconomic and political challenges, 663
and restricted information interchange between governments in transboundary basins, 664
the aforementioned metrics are not readily available everywhere. Such circumstances 665
lend themselves to the employment of machine learning models such as MLP, ELM, and 666
SVM, which need a small number of input parameters. This study shows that even with 667
fewer resources and input parameters needed, ELM, SVM, and MLP perform better than 668
the SWAT model. As demonstrated in other previous works [36,93,94], this promising ef- 669
ficiency opens up potential for future advancements, such as applying the ANN technique 670
to mimic water quality processes. 671

5. Conclusions 672
The current study used Multilayer Perceptron (MLP), Extreme Learning Machine 673
(ELM), and Support Vector Machine (SVM) models to simulate streamflow data in the 674
Astore Basin, Pakistan. These models' outputs were subjected to a thorough comparison 675
process with the widely-used SWAT model utilizing statistical performance metrics to 676
assess their modeling skills. During the calibration period (1985-1995), the SWAT model 677
had reasonable performance, as shown by R2 values of 0.80 and NSE values of 0.73. PBIAS 678
values of 15.7 indicated a marginal underestimating, whilst RMSE values of 79.81 demon- 679
strated a respectable level of prediction accuracy. The R2 dropped to 0.71 throughout val- 680
idation (1996–2005), while the NSE was lower at 0.66, suggesting decreasing accuracy. The 681
RMSE climbed to 106.26, suggesting higher prediction errors, while the PBIAS increased 682
to 17.3, indicating an overestimation. 683
On the other hand, throughout the testing, validation, and training phases, the ma- 684
chine learning models demonstrated encouraging performances. MLP demonstrated an 685
R2 of 0.73, an ELM of 0.94, and an SVM of 0.85 during calibration. Conversely, the NSE 686
values for MLP, ELM, and SVM indicated modest predictive ability at 0.4, 0.46, and 0.41, 687
respectively. The RMSE values of 1264.8, 576.27, and 929.16, respectively, corresponded 688
to the PBIAS values of 872.3, 643.23, and 726.45. The models' performances were consistent 689
according to the validation findings, with R2 values for MLP, ELM, and SVM being 0.73, 690
0.66, and 0.71, respectively. The NSE values of 0.55, 0.48, and 0.53 suggest that the predic- 691
tion skills are moderate to good. PBIAS values showed values of 481.77, 578.91, and 502.34 692
with different RMSE values, indicating a modest decline for all models. The models' abil- 693
ities were further validated by testing data, which showed R2 values of 0.69, 0.62, and 0.71 694
for MLP, ELM, and SVM, respectively. NSE values showed acceptable predictive capabil- 695
ities, remaining within a modest range of 0.54 to 0.60. PBIAS values were consistent, rang- 696
ing from 402.88 to 491.65, while RMSE values were similar, from 339.79 to 390.78. 697
The machine learning models demonstrated competitive results when compared to 698
the observed flow of 145.8 cumecs and the SWAT simulated flow of 130.83 cumecs. The 699
average modeled streamflow values were 136.35, 140.02, and 134.18 cumecs for MLP, 700
ELM, and SVM, respectively. Interestingly, in contrast to the SWAT model, which had 701
trouble overestimating low flows and underestimating high flows, these models demon- 702
strated the ability to estimate both general and severe flow situations. These results imply 703
that for streamflow prediction in the Astore Basin, machine learning models specifically, 704
MLP, ELM, and SVM are superior to the SWAT model. The study's implications include 705
helping water managers develop climate-resilient water infrastructure, allocate water to 706
different sectors, and provide strategies for flood adaptation. To improve accuracy, more 707
research combining new meteorological data is advised despite the machine learning 708
Sustainability 2024, 15, x FOR PEER REVIEW 21 of 27

models' encouraging results. Additionally, investigating alternative models based on ma- 709
chine learning and doing comparison studies with the SWAT model may yield a thorough 710
understanding, facilitating the implementation of sophisticated and precise techniques for 711
managing water resources in the area. 712

6. Implications for Policy and Water Resource Management 713


It is impossible to overestimate the importance of the Indus River to Pakistan's food 714
production and economic development it provides 90% of the country's water for agricul- 715
tural purposes and accounts for around 25% of Pakistan's GDP. The World Bank's re- 716
search (2020–2021) projects a startling 32% reduction in water availability by 2025, which 717
might lead to a catastrophic 70 million-ton food deficit in the nation. With this anticipated 718
drop, the current difficulties in allocating river water across provinces are expected to 719
worsen, underscoring the need for pragmatic solutions. 720
The findings of the study have important ramifications for managing water resources 721
and formulating policies. The potential of machine learning models, specifically Multi- 722
layer Perceptron (MLP), Extreme Learning Machine (ELM), and Support Vector Machine 723
(SVM), for a cohesive interprovincial National Water Policy, is highlighted by their supe- 724
rior performance in simulating monthly streamflow data during calibration and valida- 725
tion periods when compared to the SWAT model. Significant improvements in water dis- 726
tribution across provinces, improved flood control techniques, and more thorough plan- 727
ning for the construction of water-related infrastructure become achievable by utilizing 728
MLP, ELM, and SVM models for streamflow forecasting. Furthermore, the suggestion to 729
investigate the influence of extra meteorological factors creates opportunities for improv- 730
ing the precision and comprehensiveness of forecasts produced by MLP, ELM, and SVM 731
models. 732
These machine learning-based models provide a chance to tackle Pakistan's most 733
pressing water management difficulties, opening the door to more effective resource allo- 734
cation and improved resistance to future water scarcity problems. 735

Supplementary Materials: Not applicable. 736

Author Contributions: All authors were involved in the intellectual elements of this paper. M.R., 737
Z.A., and S.H. designed the research. S.H. and M.R. conducted the research and wrote the manu- 738
script. M.R., M.W.N., and M.A.R.T-Writing original draft, writing review, and editing with data 739
arrangement, and analysis. A.G.-Writing review and Editing, formal analysis. H.A.- Writing review 740
and Editing, formal analysis, M.A.M- Writing review and Editing, formal analysis, H.A.B- Writing 741
review and Editing, formal analysis. All authors have read and agreed to the published version of 742
the manuscript. 743

Funding: This project was funded by Princess Nourah bint Abdulrahman University Research Sup- 744
porting Project Number PNURSP2024R675, Princess Nourah bint Abdulrahman University, Ri- 745
yadh, Saudi Arabia. 746

Data Availability Statement: The data used in this study can be available from the authors upon 747
reasonable request. 748

Acknowledgments: The authors are thankful to Princess Nourah bint Abdulrahman University for 749
funding. The authors are also thankful to state key laboratory in China for conducting and helping 750
in this research. 751

Conflicts of Interest: The authors declare no conflict of interest. 752

References 753

1. Ahmed, N.; Wang, G.; Booij, M.J.; Xiangyang, S.; Hussain, F.; Nabi, G. Separation of the Impact of Landuse/Landcover 754

Change and Climate Change on Runoff in the Upstream Area of the Yangtze River, China. Water Resour. Manag. 2022, 36, 755

181–201, doi:10.1007/s11269-021-03021-z. 756

2. Wu, J.; Yuan, X.; Yao, H.; Chen, X.; Wang, G. Reservoirs Regulate the Relationship between Hydrological Drought Recovery 757
Sustainability 2024, 15, x FOR PEER REVIEW 22 of 27

Water and Drought Characteristics. J. Hydrol. 2021, 603, 127127, doi:10.1016/j.jhydrol.2021.127127. 758

3. Soriano, E.; Mediero, L.; Garijo, C. Selection of Bias Correction Methods to Assess the Impact of Climate Change on Flood 759

Frequency Curves. Water (Switzerland) 2019, 11, doi:10.3390/w11112266. 760

4. Iqbal, M.; Wen, J.; Wang, X.; Lan, Y.; Tian, H.; Anjum, M.N.; Adnan, M. Assessment of Air Temperature Trends in the Source 761

Region of Yellow River and Its Sub-Basins, China. Asia-Pacific J. Atmos. Sci. 2018, 54, 111–123, doi:10.1007/s13143-017-0064-x. 762

5. Iqbal Jun; Masood, Muhammad; Masood, Muhammad Umer; Adnan, Muhammad, M.W. Impacts of Climate and Land- 763

Use Changes on Hydrological Processes of the Source Region of Yellow River, China. Sustainability 2022, 14, 14908, 764

doi:10.3390/su142214908. 765

6. Mohammed, I.N.; Bomblies, A.; Wemple, B.C. The Use of CMIP5 Data to Simulate Climate Change Impacts on Flow Regime 766

within the Lake Champlain Basin. J. Hydrol. Reg. Stud. 2015, 3, 160–186, doi:10.1016/j.ejrh.2015.01.002. 767

7. Ali, A.F.; Xiao, C. de; Zhang, X. peng; Adnan, M.; Iqbal, M.; Khan, G. Projection of Future Streamflow of the Hunza River 768

Basin, Karakoram Range (Pakistan) Using HBV Hydrological Model. J. Mt. Sci. 2018, 15, 2218–2235, doi:10.1007/s11629-018- 769

4907-4. 770

8. Yan, R.; Cai, Y.; Li, C.; Wang, X.; Liu, Q. Hydrological Responses to Climate and Land Use Changes in a Watershed of the 771

Loess Plateau, China. Sustain. 2019, 11, doi:10.3390/su11051443. 772

9. Ikram, R.M.A.; Goliatt, L.; Kisi, O.; Trajkovic, S.; Shahid, S. Covariance Matrix Adaptation Evolution Strategy for Improving 773

Machine Learning Approaches in Streamflow Prediction. Mathematics 2022, 10, 1–30, doi:10.3390/math10162971. 774

10. Adnan, R.M.; Mostafa, R.R.; Elbeltagi, A.; Yaseen, Z.M.; Shahid, S.; Kisi, O. Development of New Machine Learning Model for 775

Streamflow Prediction: Case Studies in Pakistan; Springer Berlin Heidelberg, 2022; Vol. 36; ISBN 0123456789. 776

11. Kassem, A.A.; Raheem, A.M.; Khidir, K.M.; Alkattan, M. Predicting of Daily Khazir Basin Flow Using SWAT and Hybrid 777

SWAT-ANN Models. Ain Shams Eng. J. 2020, 11, 435–443, doi:10.1016/j.asej.2019.10.011. 778

12. Kumar, D.N.; Reshmidevi, T. V. Remote Sensing Applications in Water Resources. J. Indian Inst. Sci. 2013, 93, 163–188. 779

13. Arfan, M.; Lund, J.; Hassan, D.; Saleem, M.; Ahmad, A. Assessment of Spatial and Temporal Flow Variability of the Indus 780

River. Resources 2019, 8, doi:10.3390/resources8020103. 781

14. Sawunyama, T.; Senzanje, A.; Mhizha, A. Estimation of Small Reservoir Storage Capacities in Limpopo River Basin Using 782

Geographical Information Systems (GIS) and Remotely Sensed Surface Areas: Case of Mzingwane Catchment. Phys. Chem. 783

Earth 2006, 31, 935–943, doi:10.1016/j.pce.2006.08.008. 784

15. Ringard, J.; Becker, M.; Seyler, F.; Linguet, L. Temporal and Spatial Assessment of Four Satellite Rainfall Estimates over 785

French Guiana and North Brazil. Remote Sens. 2015, 7, 16441–16459, doi:10.3390/rs71215831. 786

16. Gao, H.; Feng, Z.; Zhang, T.; Wang, Y.; He, X.; Li, H.; Pan, X.; Ren, Z.; Chen, X.; Zhang, W.; et al. Assessing Glacier Retreat 787

and Its Impact on Water Resources in a Headwater of Yangtze River Based on CMIP6 Projections. Sci. Total Environ. 2021, 788

765, 142774, doi:10.1016/j.scitotenv.2020.142774. 789

17. Journal, I.; Sensing, R. Identification of Drought Events from Multi Years Temporal SPOT NDVI Data for Potohar Region in 790

Pakistan. Int. J. Remote Sens. GIS 2014. 791

18. Garee Xi; Bao, Anming; Wang, Yu; Meng, Fanhao, K.C. Hydrological Modeling of the Upper Indus Basin: A Case Study 792

from a High-Altitude Glacierized Catchment Hunza. Water 2017, 9, 17-NA, doi:10.3390/w9010017. 793

19. Adnan Ikram, R.M.; Jaafari, A.; Milan, S.G.; Kisi, O.; Heddam, S.; Zounemat-Kermani, M. Hybridized Adaptive Neuro-Fuzzy 794

Inference System with Metaheuristic Algorithms for Modeling Monthly Pan Evaporation. Water (Switzerland) 2022, 14, 795

doi:10.3390/w14213549. 796

20. Adnan, R.M.; Dai, H.L.; Mostafa, R.R.; Islam, A.R.M.T.; Kisi, O.; Heddam, S.; Zounemat-Kermani, M. Modelling 797

Groundwater Level Fluctuations by ELM Merged Advanced Metaheuristic Algorithms Using Hydroclimatic Data. Geocarto 798

Int. 2023, 38, doi:10.1080/10106049.2022.2158951. 799


Sustainability 2024, 15, x FOR PEER REVIEW 23 of 27

21. Zhu, R.; Yang, L.; Liu, T.; Wen, X.; Zhang, L.; Chang, Y. Hydrological Responses to the Future Climate Change in a Data 800

Scarce Region, Northwest China: Application of Machine Learning Models. Water (Switzerland) 2019, 11, 1–19, 801

doi:10.3390/w11081588. 802

22. Ikram, R.M.A.; Hazarika, B.B.; Gupta, D.; Heddam, S.; Kisi, O. Streamflow Prediction in Mountainous Region Using New 803

Machine Learning and Data Preprocessing Methods: A Case Study. Neural Comput. Appl. 2022, 8, doi:10.1007/s00521-022- 804

08163-8. 805

23. Zhou Mohamed; Takeuchi, Kuniyoshi; Koike, Toshio; Selvarajah, Hemakanth; Ao, Tianqi, L.R. Adequacy of Near Real- 806

Time Satellite Precipitation Products in Driving Flood Discharge Simulation in the Fuji River Basin, Japan. Appl. Sci. 2021, 11, 807

1087-NA, doi:10.3390/app11031087. 808

24. Kisi, O.; Heddam, S.; Keshtegar, B.; Piri, J.; Adnan, R.M. Predicting Daily Streamflow in a Cold Climate Using a Novel Data 809

Mining Technique: Radial M5 Model Tree. Water (Switzerland) 2022, 14, doi:10.3390/w14091449. 810

25. Ávila, L.; Silveira, R.; Campos, A.; Rogiski, N.; Gonçalves, J.; Scortegagna, A.; Freita, C.; Aver, C.; Fan, F. Comparative 811

Evaluation of Five Hydrological Models in a Large-Scale and Tropical River Basin. Water (Switzerland) 2022, 14, 1–21, 812

doi:10.3390/w14193013. 813

26. Syed, Z.; Mahmood, P.; Haider, S.; Ahmad, S. Assessing the Utility of Hybrid Hydrological Modeling over Complex 814

Conditions of the Chitral Basin, Pakistan. J. Water Clim. Chang. 2023, 14, 4444–4464. 815

27. Syed, S.; Syed, Z.; Mahmood, P.; Haider, S.; Khan, F.; Syed, M.T.; Syed, S. Application of Coupling Machine Learning 816

Techniques and Linear Bias Scaling for Optimizing 10-Daily Flow Simulations, Swat River Basin. Water Pract. Technol. 2023, 817

18, 1343–1356. 818

28. Nouri, M.; Sihag, P.; Kisi, O.; Hemmati, M.; Shahid, S.; Adnan, R.M. Prediction of the Discharge Coefficient in Compound 819

Broad-Crested-Weir Gate by Supervised Data Mining Techniques. Sustain. 2023, 15, doi:10.3390/su15010433. 820

29. Sahoo, A.; Ghose, D.K. Application of Hybrid Support Vector Machine Model for Streamflow Prediction in Barak Valley, 821

India. IOP Conf. Ser. Earth Environ. Sci. 2022, 1032, doi:10.1088/1755-1315/1032/1/012016. 822

30. Adnan, R.M.; Dai, H.L.; Mostafa, R.R.; Parmar, K.S.; Heddam, S.; Kisi, O. Modeling Multistep Ahead Dissolved Oxygen 823

Concentration Using Improved Support Vector Machines by a Hybrid Metaheuristic Algorithm. Sustain. 2022, 14, 824

doi:10.3390/su14063470. 825

31. Eskandari, M.; Khotanlou, H. Data Stream Classification Using a Deep Transfer Learning Method Based on Extreme Learning 826

Machine and Recurrent Neural Network. Multimed. Tools Appl. 2024, 1–29, doi:10.1007/s11042-023-18075-x. 827

32. Starzec, M.; Kordana-Obuch, S. Evaluating the Utility of Selected Machine Learning Models for Predicting Stormwater Levels 828

in Small Streams. Sustainability 2024, 16, 783, doi:10.3390/su16020783. 829

33. Yaseen, Z.; Jaafar, O.; Deo, R.; Kisi, O.; Adamowski, J.; Quilty, J.; El-Shafie, A. Boost Stream-Flow Forecasting Model with 830

Extreme Learning Machine Data-Driven: A Case Study in a Semi-Arid Region in Iraq. J. Hydrol. 2016, 542, 831

doi:10.1016/j.jhydrol.2016.09.035. 832

34. Isa, Z.; Abdussalam, A.; Sawa, B.; Ibrahim, M.; Isa, U.; Abu-hanifa, B. Identifying Major Climate Extreme Indices Driver of 833

Stream Flow Discharge Variability Using Machine Learning and SHaply Additive Explanation. Sustain. Water Resour. Manag. 834

2023, 9, doi:10.1007/s40899-023-00897-0. 835

35. Du, J.; Zhou, G.; Tang, H.; Turowski, J.; Cui, K.F. Classification of Stream, Hyperconcentrated, and Debris Flow Using 836

Dimensional Analysis and Machine Learning. Water Resour. Res. 2023, 59, doi:10.1029/2022WR033242. 837

36. Muhammad, R.; Liang, Z.; Heddam, S. Least Square Support Vector Machine and Multivariate Adaptive Regression Splines 838

for Streamflow Prediction in Mountainous Basin Using Hydro- Meteorological Data as Inputs. J. Hydrol. 2020, 586, 124371, 839

doi:10.1016/j.jhydrol.2019.124371. 840

37. Pentoś, K.; Mbah, J.T.; Pieczarka, K.; Niedbała, G.; Wojciechowski, T. Evaluation of Multiple Linear Regression and Machine 841
Sustainability 2024, 15, x FOR PEER REVIEW 24 of 27

Learning Approaches to Predict Soil Compaction and Shear Stress Based on Electrical Parameters. Appl. Sci. 2022, 12, 842

doi:10.3390/app12178791. 843

38. Chattopadhyay Deepak; Chattopadhyay, Goutami, S.J. Univariate Modelling of Monthly Maximum Temperature Time 844

Series over Northeast India: Neural Network versus Yule–Walker Equation Based Approach. Meteorol. Appl. 2011, 18, 70–82, 845

doi:10.1002/met.211. 846

39. Mattar, M.A.; Roy, D.K.; Al-Ghobari, H.M.; Dewidar, A.Z. Machine Learning and Regression-Based Techniques for 847

Predicting Sprinkler Irrigation’s Wind Drift and Evaporation Losses. Agric. Water Manag. 2022, 265, 107529, 848

doi:10.1016/j.agwat.2022.107529. 849

40. Ditthakit, P.; Pinthong, S.; Salaeh, N.; Weekaew, J.; Thanh Tran, T.; Bao Pham, Q. Comparative Study of Machine Learning 850

Methods and GR2M Model for Monthly Runoff Prediction. Ain Shams Eng. J. 2023, 14, 101941, doi:10.1016/j.asej.2022.101941. 851

41. Vishwakarma Alban; Abed, Salwan Ali; Kishore, Gottam; Al-Ansari, Nadhir; Pandey, Kusum; Kumar, Pravendra; 852

Kushwaha, N L; Jewel, Arif, D.K.K. Forecasting of Stage-Discharge in a Non-Perennial River Using Machine Learning with 853

Gamma Test. Heliyon 2023, 9, e16290–e16290, doi:10.1016/j.heliyon.2023.e16290. 854

42. Wu Haixing; Wei, Guozhen; Song, Tianyu; Zhang, Chi; Zhou, Huicheng, J.L. Flash Flood Forecasting Using Support Vector 855

Regression Model in a Small Mountainous Catchment. Water 2019, 11, 1327-NA, doi:10.3390/w11071327. 856

43. Torres-Sanchez, R.; Navarro-Hellin, H.; Guillamon-Frutos, A.; San-Segundo, R.; Ruiz-Abellón, M.C.; Domingo-Miguel, R. A 857

Decision Support System for Irrigation Management: Analysis and Implementation of Different Learning Techniques. Water 858

(Switzerland) 2020, 12, doi:10.3390/w12020548. 859

44. Atif, I.; Iqbal, J.; Mahboob, M.A. Investigating Snow Cover and Hydrometeorological Trends in Contrasting Hydrological 860

Regimes of the Upper Indus Basin. Atmosphere (Basel). 2018, 9, doi:10.3390/atmos9050162. 861

45. Hayat, H.; Akbar, T.A.; Tahir, A.A.; Hassan, Q.K.; Dewan, A.; Irshad, M. Simulating Current and Future River-Flows in the 862

Snowmelt-Runo Ff Model and RCP Scenarios. Water 2019, 11, 1–19. 863

46. Latif Ma; Yaseen, Muhammad, Y.Y. Spatial Analysis of Precipitation Time Series over the Upper Indus Basin. Theor. Appl. 864

Climatol. 2016, 131, 761–775, doi:10.1007/s00704-016-2007-3. 865

47. Atif, I.; Iqbal, J.; Su, L.J. Modeling Hydrological Response to Climate Change in a Data-Scarce Glacierized High Mountain 866

Astore Basin Using a Fully Distributed TOPKAPI Model. Climate 2019, 7, doi:10.3390/cli7110127. 867

48. Tahir, A.A.; Adamowski, J.F.; Chevallier, P.; Haq, A.U.; Terzago, S. Comparative Assessment of Spatiotemporal Snow Cover 868

Changes and Hydrological Behavior of the Gilgit, Astore and Hunza River Basins (Hindukush–Karakoram–Himalaya 869

Region, Pakistan). Meteorol. Atmos. Phys. 2016, 128, 793–811, doi:10.1007/s00703-016-0440-6. 870

49. Farhan, S. Bin; Zhang, Y.; Ma, Y.; Guo, Y.; Ma, N. Hydrological Regimes under the Conjunction of Westerly and Monsoon 871

Climates: A Case Investigation in the Astore Basin, Northwestern Himalaya. Clim. Dyn. 2015, 44, 3015–3032, 872

doi:10.1007/s00382-014-2409-9. 873

50. Baig, S.; Sayama, T.; Takara, K. Hydrological Modeling of the Astore River Basin, Pakistan, by Integrating Snow and Glacier 874

Melt Processes and Climate Scenarios. J. Disaster Res. 2021, 16, 1197–1206, doi:10.20965/jdr.2021.p1197. 875

51. Adnan, M.; Nabi, G.; Kang, S.; Zhang, G.; Adnan, R.M.; Anjum, M.N.; Iqbal, M.; Ali, A.F. Snowmelt Runoff Modelling under 876

Projected Climate Change Patterns in the Gilgit River Basin of Northern Pakistan. Polish J. Environ. Stud. 2017, 26, 525–542, 877

doi:10.15244/pjoes/66719. 878

52. Khan, S.; Khan, A.; Khan, M.; Khan, F.; Khan, S.; Khan, J. Intercomparison of SWAT and ANN Techniques in Simulating 879

Streamflows in the Astore Basin of the Upper Indus. Water Sci. Technol. 2023, 88, doi:10.2166/wst.2023.299. 880

53. gul, rahman; Khan, S.; Ullah, R.; Ali, M. Spatio-Temporal Change in the Glaciers of Astore Basin (North-Western Himalaya), 881

between 2016 and 2021 by Using Sentinel-2 Satellite Data; 2023; 882

54. Hayat, H.; Akbar, T.; Tahir, A.; Hassan, Q.; Dewan, A.; Irshad, M. Simulating Current and Future River-Flows in the 883
Sustainability 2024, 15, x FOR PEER REVIEW 25 of 27

Karakoram and Himalayan Regions of Pakistan Using Snowmelt-Runoff Model and RCP Scenarios. Water 2019, 11, 761, 884

doi:10.3390/w11040761. 885

55. Gallicchio, C.; Micheli, A.; Olivas, E.S. Randomized Machine Learning Approaches : Recent Developments and Challenges 886

Randomized Machine Learning Approaches : 2017. 887

56. Wang, J.; Lu, S.; Wang, S.; Zhang, Y. PUTATIONAL INTELLIGENCE. 2022, 41611–41660. 888

57. Sarker, I.H. Machine Learning : Algorithms , Real ‑ World Applications and Research Directions. SN Comput. Sci. 2021, 2, 1– 889

21, doi:10.1007/s42979-021-00592-x. 890

58. Huang, G.-B.; Zhu, Q.-Y.; Siew, C. Extreme Learning Machine: Theory and Applications. Neurocomputing 2006, 70, 891

doi:10.1016/j.neucom.2005.12.126. 892

59. Huang, G.-B. An Insight into Extreme Learning Machines: Random Neurons, Random Features and Kernels. Cognit. Comput. 893

2014, 6, 376–390. 894

60. Huang, G.-B.; Zhu, Q.-Y.; Siew, C.-K. Extreme Learning Machine: A New Learning Scheme of Feedforward Neural Networks. 895

In Proceedings of the 2004 IEEE international joint conference on neural networks (IEEE Cat. No. 04CH37541); Ieee, 2004; 896

Vol. 2, pp. 985–990. 897

61. Evgeniou, T.; Pontil, M. Support Vector Machines : Theory and Applications WORKSHOP ON SUPPORT VECTOR 898

MACHINES : THEORY AND APPLICATIONS. 2014, doi:10.1007/3-540-44673-7. 899

62. Gaye, B.; Zhang, D.; Wulamu, A. Improvement of Support Vector Machine Algorithm in Big Data Background. 2021, 2021. 900

63. Arnold Raghavan; Muttiah, Ranjan S.; Williams, James, J.G.. S. Large Area Hydrologic Modeling and Assessment Part i: 901

Model Development. J. Am. Water Resour. Assoc. 1998, 34, 73–89, doi:10.1111/j.1752-1688.1998.tb05961.x. 902

64. Schilling Manoj; Zhang, You-Kuan; Gassman, Philip W.; Wolter, Calvin F., K.E.. J. Impact of Land Use and Land Cover 903

Change on the Water Balance of a Large Agricultural Watershed: Historical Effects and Future Directions. Water Resour. Res. 904

2008, 44, NA-NA, doi:10.1029/2007wr006644. 905

65. Yang, L.; Feng, Q.; Yin, Z.; Deo, R.C.; Wen, X.; Si, J.; Li, C. Separation of the Climatic and Land Cover Impacts on the Flow 906

Regime Changes in Two Watersheds of Northeastern Tibetan Plateau. Adv. Meteorol. 2017, 2017, doi:10.1155/2017/6310401. 907

66. Mukundan David E.; Risse, L.M., R.R. Spatial Resolution of Soil Data and Channel Erosion Effects on SWAT Model 908

Predictions of Flow and Sediment. J. Soil Water Conserv. 2010, 65, 92–104, doi:10.2489/jswc.65.2.92. 909

67. Basheer Haishen; Omer, Abubaker; Ali, Abubaker B.; Abdelgader, Abdeldime M. S., A.K.. L. Impacts of Climate Change 910

under CMIP5 RCP Scenarios on the Streamflow in the Dinder River and Ecosystem Habitats in Dinder National Park, Sudan. 911

Hydrol. Earth Syst. Sci. 2016, 20, 1331–1353, doi:10.5194/hess-20-1331-2016. 912

68. Shafeeque Abid; Basit, Abdul; Mohamed, Abdelmoneim Zakaria; Rasheed, Muhammad Waseem; Khan, Muhammad 913

Usman; Buttar, Noman Ali; Saddique, Naeem; Asim, Mohammad Irfan; Sabir, Rehan Mehmood, M.S. Quantifying the Impact 914

of the Billion Tree Afforestation Project (BTAP) on the Water Yield and Sediment Load in the Tarbela Reservoir of Pakistan 915

Using the SWAT Model. Land 2022, 11, 1650, doi:10.3390/land11101650. 916

69. Tuo Zheng; Disse, Markus; Chiogna, Gabriele, Y.D. Evaluation of Precipitation Input for SWAT Modeling in Alpine 917

Catchment: A Case Study in the Adige River Basin (Italy). Sci. Total Environ. 2016, 573, 66–82, 918

doi:10.1016/j.scitotenv.2016.08.034. 919

70. Masood, M.U.; Khan, N.M.; Haider, S.; Anjum, M.N.; Chen, X.; Gulakhmadov, A.; Iqbal, M.; Ali, Z.; Liu, T. Appraisal of Land 920

Cover and Climate Change Impacts on Water Resources: A Case Study of Mohmand Dam Catchment, Pakistan. Water 921

(Switzerland) 2023, 15, doi:10.3390/w15071313. 922

71. Haider, S.; Masood, M.U.; Rashid, M.; Hassan, S.; Saleem, J. Evaluation of the Effects of Climate and Land Use Variations on 923

the Groundwater Dynamics of the Bari Doab Canal System in Punjab , Pakistan. 2023, 3390. 924

72. Masood, M.U.; Haider, S.; Rashid, M.; Aldlemy, M.S.; Pande, C.B.; Đurin, B.; Homod, R.Z.; Alshehri, F.; Elkhrachy, I. 925
Sustainability 2024, 15, x FOR PEER REVIEW 26 of 27

Quantifying the Impacts of Climate and Land Cover Changes on the Hydrological Regime of a Complex Dam Catchment 926

Area. Sustainability 2023, 15, 15223. 927

73. Haider, S.; Masood, M.U.; Rashid, M.; Alshehri, F.; Pande, C.B.; Katipoğlu, O.M.; Costache, R. Simulation of the Potential 928

Impacts of Projected Climate and Land Use Change on Runoff under CMIP6 Scenarios. Water 2023, 15, 3421. 929

74. Welde, K.; Gebremariam, B. Effect of Land Use Land Cover Dynamics on Hydrological Response of Watershed: Case Study 930

of Tekeze Dam Watershed, Northern Ethiopia. Int. Soil Water Conserv. Res. 2017, 5, 1–16, 931

doi:https://doi.org/10.1016/j.iswcr.2017.03.002. 932

75. Di Luzio Raghavan; Arnold, Jeffrey G., M.S. Integration of Watershed Tools and Swat Model Into Basins. J. Am. Water 933

Resour. Assoc. 2002, 38, 1127–1141, doi:10.1111/j.1752-1688.2002.tb05551.x. 934

76. Setegn, S.G.; Rayner, D.; Melesse, A.M.; Dargahi, B.; Srinivasan, R.; Wörman, A. Climate Change Impact on Agricultural 935

Water Resources Variability in the Northern Highlands of Ethiopia. Nile River Basin Hydrol. Clim. Water Use 2011, 241–265. 936

77. Rahman, K.U.; Pham, Q.B.; Jadoon, K.Z.; Shahid, M.; Kushwaha, D.P.; Duan, Z.; Mohammadi, B.; Khedher, K.M.; Anh, D.T. 937

Comparison of Machine Learning and Process-Based SWAT Model in Simulating Streamflow in the Upper Indus Basin. Appl. 938

Water Sci. 2022, 12, 1–19, doi:10.1007/s13201-022-01692-6. 939

78. Abbaspour, K.C.; Rouholahnejad, E.; Vaghefi, S.; Srinivasan, R.; Yang, H.; Kløve, B. A Continental-Scale Hydrology and 940

Water Quality Model for Europe: Calibration and Uncertainty of a High-Resolution Large-Scale SWAT Model. J. Hydrol. 2015, 941

524, 733–752. 942

79. Pradhan, P.; Tingsanchali, T.; Shrestha, S. Evaluation of Soil and Water Assessment Tool and Artificial Neural Network 943

Models for Hydrologic Simulation in Different Climatic Regions of Asia. Sci. Total Environ. 2020, 701, 134308. 944

80. Sofi, M.S.; Rautela, K.S.; Muslim, M.; Bhat, S.U.; Rashid, I.; Kuniyal, J.C. Modeling the Hydrological Response of a Snow-Fed 945

River in the Kashmir Himalayas through SWAT and Artificial Neural Network. Int. J. Environ. Sci. Technol. 2024, 21, 3115– 946

3128. 947

81. Makwana, J.J.; Tiwari, M.K. Hydrological Stream Flow Modelling Using Soil and Water Assessment Tool (SWAT) and Neural 948

Networks (NNs) for the Limkheda Watershed, Gujarat, India. Model. Earth Syst. Environ. 2017, 3, 635–645. 949

82. Makwana, J.J.; Tiwari, M.K.; Pampaniya, N.K. Impact Assessment of Land Use and Cover Changes on Water Resources Using 950

RS and GIS Techniques in Watershed Area. J. Soil Water Conserv. 2019, 18, 11–21. 951

83. Wu, C.-L.; Chau, K.-W.; Li, Y.-S. Predicting Monthly Streamflow Using Data‑driven Models Coupled with Data‑ 952

preprocessing Techniques. Water Resour. Res. 2009, 45. 953

84. Kim, C.; Kim, C.-S. Comparison of the Performance of a Hydrologic Model and a Deep Learning Technique for Rainfall- 954

Runoff Analysis. Trop. Cyclone Res. Rev. 2021, 10, 215–222. 955

85. Guse, B.; Reusser, D.E.; Fohrer, N. How to Improve the Representation of Hydrological Processes in SWAT for a Lowland 956

Catchment–Temporal Analysis of Parameter Sensitivity and Model Performance. Hydrol. Process. 2014, 28, 2651–2670. 957

86. Cibin, R.; Sudheer, K.P.; Chaubey, I. Sensitivity and Identifiability of Stream Flow Generation Parameters of the SWAT Model. 958

Hydrol. Process. An Int. J. 2010, 24, 1133–1148. 959

87. Raaj, S.; Gupta, V.; Singh, V.; Shukla, D.P. A Novel Framework for Peak Flow Estimation in the Himalayan River Basin by 960

Integrating SWAT Model with Machine Learning Based Approach. Earth Sci. Informatics 2024, 17, 211–226. 961

88. Oo, H.T.; Zin, W.W.; Kyi, C.C.T. Analysis of Streamflow Response to Changing Climate Conditions Using SWAT Model. Civ. 962

Eng. J. 2020, 6, 194–209. 963

89. Demirel, M.C.; Venancio, A.; Kahya, E. Flow Forecast by SWAT Model and ANN in Pracana Basin, Portugal. Adv. Eng. Softw. 964

2009, 40, 467–473. 965

90. Hassan, S.; Masood, M.U.; Haider, S.; Anjum, M.N.; Hussain, F.; Ding, Y.; Shangguan, D.; Rashid, M.; Nadeem, M.U. 966

Investigating the Effects of Climate and Land Use Changes on Rawal Dam Reservoir Operations and Hydrological Behavior. 967
Sustainability 2024, 15, x FOR PEER REVIEW 27 of 27

Water 2023, 15. 968

91. Haider, S.; Masood, M.U.; Rashid, M.; Alshehri, F. Simulation of the Potential Impacts of Projected Climate and Land Use 969

Change on Runoff under CMIP6 Scenarios. 2023. 970

92. Senent-Aparicio, J.; Jimeno-Sáez, P.; Bueno-Crespo, A.; Pérez-Sánchez, J.; Pulido-Velázquez, D. Coupling Machine-Learning 971

Techniques with SWAT Model for Instantaneous Peak Flow Prediction. Biosyst. Eng. 2019, 177, 67–77. 972

93. Yaseen, Z.M.; Faris, H.; Al-Ansari, N. Hybridized Extreme Learning Machine Model with Salp Swarm Algorithm: A Novel 973

Predictive Model for Hydrological Application. Complexity 2020, 2020, 1–14. 974

94. Ouma, Y.O.; Okuku, C.O.; Njau, E.N. Use of Artificial Neural Networks and Multiple Linear Regression Model for the 975

Prediction of Dissolved Oxygen in Rivers: Case Study of Hydrographic Basin of River Nyando, Kenya. Complexity 2020, 2020, 976

1–23. 977
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual au- 978
thor(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to 979
people or property resulting from any ideas, methods, instructions or products referred to in the content. 980

You might also like