Streamflow forecasting holds pivotal importance for planning and decision-making in the domain of water resources management. The Chitral basin in Pakistan is characterized by high altitude and glaciated terrain. Simulating streamflows in this type of region is challenging due to complex orography and uncertain climate data. This complexity persuaded us to explore three frameworks (soil and water assessment tool (SWAT), artificial neural network (ANN), and hybrid of SWAT–ANN (H2)) for simulating the Chitral river under two different climate datasets (observed climatology (OC) and reconciled gridded climatology (RGC)) to give all six model combinations. Model evaluation was done first by indices (Nash–Sutcliff efficiency, Kling–Gupta efficiency, coefficient of determination, percent bias, and root mean square error) based on which we further assigned scores to models reflecting their performance during calibration and validation epochs. The research revealed that ANN-RGC stood first with 53 points, followed by H2-RGC (50 points) and SWAT-RGC (45 points). Trailing behind in the fourth and fifth positions were SWAT-RGC and SWAT-OC (26 points each), respectively, while ANN-OC finished last (22 points). In addition, this study proposed a bias scaling approach for simulation biases resulting in reduction in recession and baseflow biases and specifically improved low-scoring models. Despite ANN's superiority over conventional models, it could be of limited utility in uncertain or data-scarce conditions.

  • Reliable climate data hold pivotal importance in hydrological modeling.

  • Artificial neural networks scored the highest but were also found to be more sensitive to data quantity and quality.

  • The coupling harnessed the capabilities of the parent frameworks and performed well overall.

  • In uncertain data conditions, the soil and water assessment tool and hybrid models could be more suitable choices.

  • Implied linear scaling efficiently removed model biases.

Hydrological modeling of regional basins with the climate change extreme events is crucial for planning water-based strategies and management policies for supporting the economy (Mogano & Mokoele 2019; Nam et al. 2022; Syed et al. 2023). The Chitral river basin is a regional subbasin that is located in northwest Pakistan. It is an important tributary of the Kabul River basin, which itself is a pivotal subbasin of the Upper Indus Basin (UIB). It is one of the least explored basins in the country that has a high potential for water supply and hydropower generation. Most of the region is characterized by the high-altitude, rugged, and glaciated terrain with a complex hydroclimatic regime. Its regional hydrology is influenced by synoptic- and valley-scale climatic circulations (Immerzeel et al. 2015). On the synoptic-scale circulations, the Chitral basin receives the main share of its total annual precipitation through westerly Mediterranean disturbances during the winter season. The winter fall provides major replenishment to glaciers and stores supply for the summer runoff. Unlike the other eastern parts of the UIB, the monsoon has little influence over the Chitral basin, but occasional penetration of moist air during the late spring and summer season sometimes induces heavy downpours, causing flooding (Ahmad et al. 2016). At the valley scale, precipitation distribution is dictated by local orographic events, which are manipulated by the topography and atmospheric conditions (Bookhagen & Burbank 2006; Immerzeel et al. 2014, 2015). This interplay between these factors adds up to the nonlinearity and complex behavior, making it difficult to model hydrological responses of the basin.

Difficult terrain and the presence of large perennial glaciers on the northern side of the Chitral region make it challenging to set up and operate climate stations (Ahmad et al. 2018). Only a few meteorological stations are present, which are operational in lower valleys. These stations inherently depict a biased representation of the region, as lower areas are relatively drier while areas at higher elevations are influenced by orography with a nonlinear gradient (Immerzeel et al. 2015; Dahri et al. 2018; Syed et al. 2022). In addition, the high-altitude climate stations are often exposed to wind-induced systematic biases, which could degrade the accuracy of precipitation records (Dahri et al. 2018). Dahri et al. (2018) attempted to adjust this systematic bias over the UIB and reported to have recovered a significant proportion of undermined precipitation over the Chitral basin. Nevertheless, uncertain precipitation along with paucity of other climatic parameters like temperature are among the major obstacles to a comprehensive understanding of the regional hydrology (Winiger et al. 2005; Hewitt 2009; Palazzi et al. 2013; Immerzeel et al. 2015). Previously, different gridded datasets were used to fill the spatial gaps, which are mainly of three types: observation-based, reanalysis, and satellite-based gridded datasets (Ragettli & Pellicciotti 2012; Baudouin et al. 2020). As the observation-based gridded sets depend upon conventional climate station data, the presence of usual errors could greatly reduce the reliability of these datasets (Syed et al. 2022). Another alternative is a satellite gridded dataset, which indirectly measures precipitation amount via infrared and passive microwave sensing. However, many studies have reported the poor performance of satellite datasets in high-altitude regions (Iqbal & Hussain 2017; Yuan et al. 2017). Moreover, satellite products often require ground observation data for their adjustment, which makes it dependent on the quality of ground data (Syed et al. 2022). Another common method is using reanalysis gridded datasets, which are fundamentally derived from numerical weather predictors. These models use numerical algorithms bounded by physical laws and can generate several climate outputs. However, evaluating its reliability is also highly desirable and crucial for gauging the predictive capacity of hydrological models (Lutz et al. 2016). Previously, Dahri et al. (2021) evaluated 27 widely used precipitation products of different types over all the UIB. Similarly, Baudouin et al. (2020) cross-validated 20 latest gridded datasets in the same area. Both studies showed that the European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA5) (Hersbach et al. 2020) provides comparatively better performance over the UIB region. Recently, Syed et al. (2022) utilized ERA5 Land and reported good performance of the dataset over the Chitral basin.

Hydrological modeling of stream flows in the emerging countries, which mostly depends on the agriculture sector, is crucial for water, energy, and food security (Jimeno-Sáez et al. 2018; Mahmood et al. 2022) and helps us to better prepare and plan for the effects of extreme weather events in the region (Mekonnen et al. 2012). Since the precipitation and temperature are parameters of prime importance for runoff modeling, data quality directly affects model performance (Yuan et al. 2017). However, considering climate complexity over the Chitral region along with uncertain observation records, it is much more challenging to simulate the hydroclimatic responses. In conventional physical modeling, the effects of climate variables like precipitation and temperature do not directly translate into streamflow responses but rather depend upon other inputs like land use, vegetation, etc. These inputs are usually based on remote sensing (RS) and could contain uncertainty due to indirect measurement (Hosseini et al. 2022). Moreover, these data are generally incorporated into the model as a single gridded layer and assumed temporarily stationary, which may introduce uncertainty in the temporal hydrological response (Hosseini et al. 2022). Furthermore, another thorny issue affecting physical modeling is its susceptibility to improper parameterization due to poor or deficient data (Nesterova et al. 2021). This can result in erroneous interpretation of the model results as, during calibration, defective data are compensated by alteration in other parameters. However, regardless of improvement in data acquisition and modeling methodology, limitations do persist and, thus, it is significantly important to focus on enhancing the reliability of the main driving parameters such as precipitation and temperature for better hydrological model output. Overall, in the Chitral region, only a few studies relate climate with stream flows. These studies utilized physical modeling to simulate stream flows (Hasson et al. 2019; Azzam et al. 2022). Most of these focused on the methodology and overlooked the deficiency in observational climate data (Dahri et al. 2018). Recently, Syed et al. (2022) addressed uncertainty in the climate data and utilized (Dahri et al. 2018) reconciled climatology for hydrological modeling. Although the study was able to produce good results, it only considered a single physical modeling methodology. This study emphasizes both modeling methodology and the provision of reliable climate datasets for effective predictive modeling under complex conditions encompassing terrain and climate.

Based on modeling methodology, a hydrological model can be divided into data-driven (DD) models and physical process-based models (Hosseini et al. 2022). Among DD models, the artificial neural network (ANN) is a successful and robust model (Panu et al. 2000; Elshorbagy & Simonovic 2002; Alizadeh et al. 2017). It is being used widely as an alternative to conventional modeling for stream flow simulation and other water resource applications (Jimeno-Sáez et al. 2018; Ali et al. 2019). The ANNs can capture complex hydrological responses by skipping the intermediate catchment processes and directly relating input to output (Singh et al. 2011). ANNs are being effective in capturing nonlinear behavior; therefore, they could perform better in dealing with uncertainty and data biases (Hosseini et al. 2022). Previously, several studies have noted the good performance of ANNs for flow simulation in complex regimes (Santos & Silva 2014; Wang et al. 2015; Shiau & Hsu 2016; Jimeno-Sáez et al. 2018, 2017); However, physical, process-based hydrological modeling conventionally consists of two types of models: conceptual- and physical-based hydrological models (Islam 2011). The conceptual-based model uses simplified mathematical equations to represent the hydrological process, while the physical-based model could either utilize more complex approximations of partial differential equations, representing mass, momentum, and energy, or uses empirically based approximations (Islam 2011). Based on considerations of spatial variability, the physical-based models are further subdivided into lumped, distributed, and semi-distributed models (Islam 2011; Grayson 2022). Most well-known among the physical-based model is the soil and water assessment tool (SWAT) model. Several studies have reported its performance to be better than other modeling frameworks like Hydrologic Engineering Center – Hydrologic Modeling System (HEC-HMS), Variable infiltration Capacity (VIC), Water Evaluation And Planning (WEAP), and Global Water Availability Assessment Model (GWAVA) (Nguyen Khoi 2015; Horan et al. 2021; Touseef et al. 2021). Moreover, various studies also applied the SWAT to simulate complex hydrological conditions (such as potholes, data scarcity, complex mountain topography, and snow melting) (Mekonnen et al. 2012; Grusson et al. 2015; Marahatta et al. 2021; Abbas et al. 2021, 2022; Syed et al. 2022). Some studies also used ANNs in conjunction with the SWAT (Mekonnen et al. 2012; Noori & Kalin 2016; Ali et al. 2019). These studies used fused or hybrid modeling that takes advantage of both types of models by collecting the initial response of subbasins from discretized physical models and, usually, used its output residual to feed ANNs for computing the complex stream flow dynamics (Mekonnen et al. 2012; Ali et al. 2019). Several studies have reported improved performance. For instance, Mekonnen et al. (2012) applied the SWAT–ANN hybrid model and compared its performance with individual ANN and SWAT models. It was revealed that the hybrid arrangement was able to produce better results than the other individual models for overall periods and during low- and high-flow scenarios.

This study will test three types of modeling methodologies (ANN, SWAT, and ANN–SWAT hybrid models) under given observed and corrected climate datasets. The key element will be to investigate the performance of each model based on its unique structure, optimal adjustments, and given inputs. This study will also suggest the improvement in techniques for streamflow simulation and forecasting.

Study area

The Upper Chitral basin (Figure 1) is located in the northwest of Pakistan. It is also a transboundary basin shared between Pakistan and Afghanistan with a lower basin lying in Afghanistan. The geographical basin lies between 35°–37°N latitude and 70°–74°E longitude and has a drainage area of 14,639.54 km2 in the Pakistan-administered region. Almost 14.5% of the region consists of perennial glaciers. Furthermore, it is characterized by a highly steep, glaciated, mountainous terrain with very few flat or mild slope surfaces. The mountains of the Chitral basin are part of the Hindu Kush range with elevations ranging from 7,700 to 1,100 m above mean sea level. For most of the year, the Chitral region remains cold and dry. Most of the precipitation occurs due to westerly Mediterranean disturbances between the months of October and April. During these months, the temperature mostly remains subzero, due to which the precipitation falls as snow, which helps to replenish the glaciers. Unlike other large parts of the country, the Chitral region is little influenced by the summer monsoon; however, the occasional penetration of a system could bring heavy downpours. The annual mean precipitation recorded (1992–2017) is 465 mm. The Chitral river originates as the Mastuj river and is fed from the glacial melt of the Hindu Kush Mountain range. The hydrology of the Chitral basin predominantly behaves as a snow basin. The peak flow occurs between June and September, as the temperature rise accelerates the snowmelt. The mean flow of the Chitral river at the main Chitral city gauge is 292 m3/s for the period of 2003–2015. The main tributaries of the Chitral river are Turkho, Ludko, Yarkhon, Turich, and Golen Gol rivers.
Figure 1

Basin map of the Chitral river basin alongside climate station and climate zones.

Figure 1

Basin map of the Chitral river basin alongside climate station and climate zones.

Close modal

Climate datasets

For analysis, the observed data of the Chitral station were acquired from the Pakistan Meteorological Department. It will be hereinafter referred to as observed climatology (OC). However, the climate of the highlands is generally characterized by its immediate spatial- and temporal-based fluctuations, which are difficult to adequately capture based on data from a single station (Waseem et al. 2021; Abbas et al. 2023). Therefore, to compensate for the aforementioned disadvantage, we also used the gridded climate dataset. As stated previously, the studies showed that the ECMWF Re-Analysis products comparatively provide better results over the UIB region; therefore, we selected the ERA5 Land, which is among the latest of their products. It specifically emphasizes land assimilation. Compared to previous products, it has also been considerably improved, in relation to enhanced spatial and temporal resolution (9 km × 9 km × 1 h), altitude-based lapse rate correction, and no data assimilation requirement (Hersbach et al. 2020; Muñoz Sabater et al. 2021; Abbas et al. 2023). The dataset can be acquired from the Copernicus Climate Change Service (Muñoz Sabater 2019). As the raw reanalysis products contain model bias, it is highly advisable to apply correction before its use in any kind of analysis (Pelosi et al. 2020). Usually, observed data serve as a correcting dataset; however, in our case, available observed data already contained biases and systematic errors, e.g., wind-induced (Dahri et al. 2018). As the usage of such data could give inaccurate results, we applied (Dahri et al. 2018) monthly climatologies (1999–2011) to treat the systematic bias. The bias-corrected ERA5 Land climatology will further be referred to as reconciled gridded climatology (RGC).

Moreover, apart from precipitation and temperature, additional external factors such as snow cover and relative humidity might be indirectly linked to streamflow response. Nonetheless, the integration of these variables relies on data availability and accessibility. In our case, the availability of these products was limited to coarser scale, which could degrade the simulation process. We therefore underscored the value of parsimonious approaches and did not include these parameters.

Bias correction

Different studies utilized several types of bias correction techniques. These techniques range from simplified to more complex approaches. Various studies compared the performance of bias correction techniques and mostly related it to the location (Lenderink et al. 2007; Teutschbein & Seibert 2012; Shrestha et al. 2017). Shrestha et al. (2017) compared simpler linear bias correction and more complex quantile delta mapping techniques under high-altitude Himalayan conditions. Their findings show that for a moderate temporal resolution such as on a monthly scale, both simpler linear scaling and quantile delta mapping yield almost similar results. The location of the Chitral basin lies in close proximity to the Himalayan range and both are governed by the orographic condition; therefore, linear scaling seems more pragmatic. This study applies linear scaling on the ERA5 Land-gridded dataset using (Dahri et al. 2018) monthly climatologies (1999–2011). Equations (1)–(4) pertaining to precipitation and temperature linear scaling are shown below.
formula
(1)
formula
(2)
formula
(3)
formula
(4)
where P refers to the precipitation, T refers to the temperature, d means daily, Um refers to the long-term mean, * denotes bias-corrected, his refers to historical raw data, obs stands for observed data, and sim is the raw future data.

Model performance evaluation

Various types of objective functions were used to evaluate the model and the data performances. These functions are Nash–Sutcliff efficiency (NSE), modified Kling–Gupta efficiency (KGE), coefficient of determination (R2), Pearson's correlation (R), percent bias (PBIAS), root-mean-square error (RMSE), mean square error (MSE), and Bayesian information criterion (BIC). Table 1 shows equations and ranges corresponding to each objective function. The NSE, commonly used for evaluation of model results, serves to measure the fitness of observation and simulation plots on a 1:1 line scale. It ranges from − to 1, with the latter value being considered a perfect fit, while a value equal to 0 signifies that the model simulation is no better than the mean of observations (Jackson et al. 2019). The NSE equation can also be reformulated in relation to RMSE where NSE is equal to 1 minus the square of the ratio between RMSE and the standard deviation of observation (Sobs). The range of RMSE is from 0 to ; according to the formulated relation if RMSE is equal to Sobs, it means the NSE is equal to 0, while if the RMSE value is half of the Sobs that means the NSE is equal to 0.75. Therefore, to obtain good modeling results, the RMSE value should be equal to or less than half of the Sobs (Syed et al. 2022). The modified KGE diagnostic model assesses performance by computing three components: correlation coefficient (R), bias term (β), and variability term (γ). Its value ranges from to 1, but the highest is desirable. Modified KGE values express the lower limit of its three components, i.e., the obtained value of KGE is the worst of its components (Jackson et al. 2019). The coefficient of determination (R2) explains the collinearity between the simulated and the measured data; a value ranging from 0 to 1 and closer to 1 is desirable. Pearson's correlation (R) is the square root of R2, and it ranges from −1 to 1 where a value close to 1 is desirable (Jimeno-Sáez et al. 2018). The PBIAS evaluates the mean tendency of simulation to be greater or smaller than observation. The ideal value is 0 (Jimeno-Sáez et al. 2018). The MSE is the square of RMSE, and its ideal value is 0. The BIC is a likelihood function based on the parsimonious approach that penalizes models if it contains excessive parameters. The lowest value of BIC is preferred (Mekonnen et al. 2012).

Table 1

Objective functions

Objective functionsEquationsRange
NSE   to 1 
RMSE  0 to  
KGE ,   to 1 
R2  0 to 1 
R  −1 to 1 
PBIAS   
MSE , RMSE2 0 to  
BIC  0 to  
Objective functionsEquationsRange
NSE   to 1 
RMSE  0 to  
KGE ,   to 1 
R2  0 to 1 
R  −1 to 1 
PBIAS   
MSE , RMSE2 0 to  
BIC  0 to  

N is the number of flow values, Qobs and Qsim are the observed and simulated flows for ith observations, and and are the means of the observed and simulated flow values. Sobs and Ssim are the standard deviation of the corresponding observed and simulated flows. m is the number of parameters.

SWAT model

The SWAT is a physical-based, semi-distributed model, which is developed by the United States Department of Agriculture (USDA). It is a popular tool for analyzing and testing numerous applications related to water resources, land use, climate change, etc. The SWAT has a relatively flexible input requirement and can simulate sub-daily to annual timescales (Arnold et al. 1998; Shrestha et al. 2018). The model's notable characteristic is its use of hydrological response units (HRUs) to generate a unique spatial identity based on land use, soil, and slope. Each HRU uses a water balance equation and considers processes such as runoff, evapotranspiration, infiltration, percolation, and lateral flows.

To simulate snow conditions, the SWAT uses the temperature index method, which assumes that the precipitation will fall as snow in the HRU if the daily temperature is less than a certain threshold temperature (snowfall temperature, SFTMP) and snow will melt if it is more (snowmelt temperature, SMTMP). This principle allows the SWAT to regulate the quantity of liquid water equivalent, which further adjusts the snowpack in the basin. The snow-related parameters and the threshold temperature can be adjusted later during calibration (Neitsch et al. 2011).

For setting up a workable SWAT model, it is required to incorporate two types of primary inputs, namely, physiographic input layers and climate input. The former includes a digital elevation model (DEM), land use/cover, and soil layers. These layers are used in a raster format to generate HRUs and to define an initial set of parameters and assumptions. A higher-quality DEM is preferred as it provides for a better representation of the stream network, slope, and elevation distribution across the basin. The land use and soil data are used to inform the diverse physiographic conditions of the basin. In the case of climate data, the SWAT has the option for several types of input variables, e.g., precipitation, temperature, humidity, sunshine, and wind speed; however, only precipitation and temperature are mandatory. Another option is a weather generator file, which can fulfill the climate input requirement on its own (Neitsch et al. 2011).

Artificial neural network

ANNs are based on a group of DD system identification approaches, which are inspired by the human nervous system. The ANN can map complex relationships between the input and the output without prior knowledge of the physical processes at work (Noori & Kalin 2016; Hosseini et al. 2022). A typical ANN comprises processing units called neurons, configured in single or multiple layers, also called shallow and deep networks with embedded hidden layers. Each neuron in a layer is interlinked with neurons of the neighboring layers. For sequential analysis, independent input and output consist of an equal number of elements. As a prerequisite to the execution of ANN, it is advised to normalize input features as in most cases, the inputs are different in magnitude and, therefore, the normalization serves to bring the values to a common scale. The operation of a single neuron on the modeling framework is as follows. First, each neuron in an input layer multiplies the normalized input array with weight and adds sums of weighted inputs to bias. This is followed by the step in which the activation function is applied, which produces an output. The weights are responsible for adjusting the mathematical relation between the input and the network; positive and negative weights represent excitatory and inhibitory signals, while zero weights show no signals. The activation or transfer function serves to introduce the nonlinearity in the ANN, without which ANNs would only be a set of linear functions. The activation function needs to be differentiable to its parameters (i.e., w and b), which can then be used for optimization using the gradient of Jacobian-based methods (Hosseini et al. 2022). There are several types of activation functions, e.g., linear, sigmoid, hyperbolic tangential, and rectified linear unit. These functions can be changed for each layer. In a multilayer ANN, the neuron of the previous layer becomes an element of the input array, and that layer acts as an input array for the next layer and so forth. Consequently, the last layer consists of a single neuron that calculates the single output from previous layers (Jimeno-Sáez et al. 2018; Hosseini et al. 2022). To get the actual output, values of the output layer must be translated back to the actual scale. The basic mathematical function of a single neuron are represented in Equation (5).
formula
(5)

In the above equation, y is the resultant of neuron j, f is the activation function of neuron, x is the th input element (where ), w is the associated scalar weight, and b is the scalar bias associated with a neuron.

In the ANN model, the error reduction is through the backpropagation algorithm. This process, at first, is initiated from randomly assigned weights. The weights are then repeatedly adjusted to minimize the loss function (given as an equation of model residual, MSE, etc.) at each iteration (Jimeno-Sáez et al. 2018; Ali & Shahbaz 2020; Hosseini et al. 2022). The process is continuously repeated until the system fulfills the given rule or a limit, i.e., the gradient becomes zero or close to zero, the number of epochs reaches a set limit, or the model meets the early stopping rule. The backpropagation algorithm works in two passes, forward and backward. In the forward pass, the input array passes through the layers to the output layer. At the output layer, the loss is calculated, and the error is transmitted back for adjustment, which is called a backward pass. In this way, the model learns by adjusting the weights to give output as near to the target variable as possible (Ali & Shahbaz 2020).

Hybrid modeling

The hybrid modeling approach integrates more than one modeling paradigm. It is based on the assumption that the modeling can benefit by taking advantage of different approaches. In the past, a few studies have been done on the hybridization of hydrological modeling (Mekonnen et al. 2012). Nevertheless, fewer considered SWAT and ANN hybridization (Mekonnen et al. 2012; Noori & Kalin 2016; Ali et al. 2019). Conventional hybrid modeling is a two-step process. The first step consists of setting up the SWAT model and allowing it to simulate unadjusted output. This usually poorly represents the observed data; however, it considers a wide variety of basin characteristics that influence the pattern of output. The interpretation of the hybrid model, in this regard, is related to a preliminary parameterization of the SWAT model; therefore, it is advisable to use a representative dataset. Next, the residual or error series is calculated by taking the difference between the model output and observed data. In the second step, the ANN is employed to refine the SWAT inferences by training the network to simulate the residual. During the training of ANN, the SWAT output together with other hydroclimatic parameters can serve as input features, while the residual will be used as target data. During testing, simulated outcomes can be added to the SWAT output to get hybridized inferences.

In addition to the conventional hybrid methodology, we also proposed to use the simpler and direct methodology in which the simulated output of the SWAT model will be treated as a primary predictive variable of the ANN model alongside other climate variables. The simulated flow variables will be adjusted with a suitable time lag to optimize the relationship with the original observed flow. The time lag of the simulated variable is meant to account for the unadjusted basin retention and routing time. In this context, the primary difference between the traditional hybrid (H1) and proposed simple hybrid modeling (H2) is the choice of a different target variable/output used for ANN modeling and the use of lag time-adjusted simulated variables.

SWAT model setup

For setting up the SWAT model, the primary inputs, i.e., physiographic layers and climate data, were preprocessed. First, we incorporated Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Digital Elevation Model (DEM) of 30 m resolution, which was acquired from the United States Geological Survey (USGS), web platform in the model. Next, a custom land use input map was independently created by applying a supervised learning classification technique on Landsat-7, 30 m resolution imagery. For simplicity, the land use was classified into five SWAT-compatible land use classes (barren: BARR, forest: FRST, rangeland: RNGL, agriculture: AGRL, and water {glaciers/snow}: WATR). Furthermore, soil data of 1 km resolution was acquired from the open-source Food and Agriculture Organization (FAO) web platform (FAO and International Institute for Applied Systems Analysis (IIASA), 2023). All the physiographic layers were clipped over the study area before their input in the models. Based on climatology, two independent models were configured: SWAT-OC and SWAT-RGC. The SWAT-OC model only considered a single station, while the SWAT-RGC took into account the 220 gridded stations. Models were allowed to run from 2003 to 2015 (13 years), and 2 years from 2001 to 2002 were reserved for warming up the model.

Parameter adjustment is a necessary step in any hydrological modeling endeavor as it allows us to fit the simulation output to the observations, which is known as calibration. The SWAT calibration and uncertainty program (SWAT-CUP) allows us to automatically calibrate and validate a model using multiple calibration algorithms for output optimization. We selected the sequential uncertainty fitting (SUFI-2) algorithm, which is a popular tool for determining the uncertainty between observed and simulated values using a limited number of simulations (Yang et al. 2008; Ghoraba 2015). Before calibration, a sensitivity analysis was performed to identify the most influential parameters. The SWAT-CUP provides two options to perform the sensitivity analysis, i.e., global and one-at-a-time (OAT). In this study, we utilized OAT as it has a light computational cost. In OAT, a set of selected parameters is held constant and a single parameter is allowed to change to assess the variation in the output. Usually, three to five simulations are sufficient to determine the sensitivity of a parameter (Mikayilov et al. 2007). For both the models, we initially selected 20 parameters for the sensitivity analysis, of which 11 and 14 parameters (shown in Table 2) were identified as sensitive for SWAT-OC and SWAT-RGC models during calibration. Furthermore, during the calibration process, the chosen parameters were organized into two distinct sets. The first set comprised parameters related to snow and elevation bands, while the second set included the remaining known as hydrological parameters. This separation into two groups was intended to prevent potential identification issues during the calibration process as explained in the parameterization section of the study by Abbaspour et al. (2017). Briefly, the rationale behind conducting separate calibrations lies in the sensitivity of different parameters concerning the alteration of the water budget within a catchment. Specifically, parameters related to snow and elevation are recognized for their significant impact on the overall budget dynamics. However, the hydrological parameters play a fundamental role in regulating factors like time and concentration and do not directly affect the budget.

Table 2

Selected sensitive parameters and their calibrated values corresponding to OC and RGC climate datasets

ParametersDescriptionRGCOC
Hydrological parameters    
 r__CN2.mgt SCS runoff curve number 0.00044 −0.03 
 v__ALPHA_BF.gw Base flow alpha factor (days) 0.91 0.79 
 v__GW_DELAY.gw Groundwater delay (days) 88 – 
 v__GWQMN.gw Threshold in the shallow aquifer for return flow to occur (mm) 89 62 
 v__GW_REVAP.gw Groundwater ‘revap’ coefficient 0.03 – 
 v__SLSUBBSN.hru Average slope length (m) 52 58 
 v__HRU_SLP.hru Average slope steepness (m/m) 0.03 0.03 
 v__OV_N.hru Manning's ‘n’ value for overland flow 17 16 
Snow and elevation band parameters   
 v__SFTMP.bsn Snowfall temperature (°C) 2.56 3.12 
 v__SMTMP.bsn Snowmelt base temperature (°C) −1.8 0.4 
 v__SMFMX.bsn Maximum snowmelt rate during the year of summer solstice (mm/°C-day) 2.1 3.6 
 v__TIMP.bsn Snowpack temperature lag factor 0.46 – 
 v__TLAPS.sub Temperature lapse rate (°C/km) −5.7 −5.5 
 v__PLAPS.sub Precipitation lapse rate (mm/km) 301.25 390.8 
ParametersDescriptionRGCOC
Hydrological parameters    
 r__CN2.mgt SCS runoff curve number 0.00044 −0.03 
 v__ALPHA_BF.gw Base flow alpha factor (days) 0.91 0.79 
 v__GW_DELAY.gw Groundwater delay (days) 88 – 
 v__GWQMN.gw Threshold in the shallow aquifer for return flow to occur (mm) 89 62 
 v__GW_REVAP.gw Groundwater ‘revap’ coefficient 0.03 – 
 v__SLSUBBSN.hru Average slope length (m) 52 58 
 v__HRU_SLP.hru Average slope steepness (m/m) 0.03 0.03 
 v__OV_N.hru Manning's ‘n’ value for overland flow 17 16 
Snow and elevation band parameters   
 v__SFTMP.bsn Snowfall temperature (°C) 2.56 3.12 
 v__SMTMP.bsn Snowmelt base temperature (°C) −1.8 0.4 
 v__SMFMX.bsn Maximum snowmelt rate during the year of summer solstice (mm/°C-day) 2.1 3.6 
 v__TIMP.bsn Snowpack temperature lag factor 0.46 – 
 v__TLAPS.sub Temperature lapse rate (°C/km) −5.7 −5.5 
 v__PLAPS.sub Precipitation lapse rate (mm/km) 301.25 390.8 

In the case of the hybrid model, calibration was not applied, and the SWAT model's uncalibrated outputs/simulations were directly carried forward to the subsequent step. These outputs were then employed for the assimilation process within the ANN model.

ANN and hybrid setting

In total, six models of different combinations were independently configured. Two models were exclusively of ANN type, i.e., ANN-RGC and ANN-OC. The remaining four were hybrid models: two are the traditional hybrid RGC-based model (H1-RGC) and the hybrid OC-based model (H1-OC), while the other two are the proposed hybrid RGC-based model (H2-RGC) and the hybrid OC-based model (H2-OC).

There are no explicit guidelines regarding the number of inputs for an ANN model; however, it should be carefully selected as the input variables are instrumental in mapping the underlying nonlinear processes. More than necessary inputs can reduce the learning rate of ANN, while a deficient number of inputs could result in the formation of a weak model (Jimeno-Sáez et al. 2018). Our study employed three variables for the exclusively ANN-based models (ANN-RGC and ANN-OC), namely, precipitation, temperature, and flow. The former two were the input variables, while the latter was the target variable. The number of input variables was increased to four as a new derived variable was added to account for the lag. The details of lagged and non-lagged variables are as follows: lag-adjusted variables are ‘N-days lagged precipitation’ (P), ‘N-days lagged accumulated precipitation’ (AP), and ‘N-days lagged mean temperature’ (Tm), while the non-lagged variable is temperature (T). Furthermore, RGC consists of a large number of data points and incorporating it all into the ANN model would be computationally intensive; thus, we discretized the basin into six climate zones, and climate data points were then averaged over that area for each zone. In this way, climate variability could be preserved for each climate zone. The climate zones are shown in Figure 1.

For hybrid modeling, flow-type variables were derived by simulating unadjusted stream flows using SWAT models corresponding to RGC and OC datasets. For the traditional hybrid models (H1-RGC and H1-OC), the target variable residual flow (Rf) was calculated by taking the difference between observed flow and simulated flow, while the raw simulated flow (Fsim) was selected as the primary input variable for the model. These variables are considered as non-lagged variables. However, in the case of the novel hybrid models (H2-RGC and H2-OC), Fsim was modified into two new variables to account for the lag. The latter are ‘N-days lagged SWAT simulated flow’ (Fswat) and ‘N-days lagged mean SWAT simulated flow’ (Fmswat). These variables will be adjusted for the time lag in the analysis that follows.

The ANN framework takes the training set as a source of parameterization and verifies the performance using test datasets. During training, a small proportion of the dataset is used for in-process validation called the validation dataset. This dataset does not directly take part in parameterization but is rather used as a preliminary check for model performance and to control overfitting via early stopping (Hosseini et al. 2022). Generally, a larger proportion of datasets are reserved for training ANN. In our study, we considered the daily dataset from 2003 to 2010 as the training and validation datasets. This is the same as the calibration period of the SWAT model, while the dataset from 2011 to 2014 was taken as the testing set, which corresponds to the validation period of the SWAT model. For modeling, we used the Levenberg–Marquardt (LM) algorithm as the backpropagation algorithm. Previous studies suggested that the LM algorithm showed better performance and a faster convergence rate as compared to other algorithms (Nayebi et al. 2006; Kişi 2007). Moreover, studies involving stream flow forecasting have reported on the good performance of the LM algorithm (Kişi 2007; Yaseen et al. 2015). For the activation function of the hidden layer, we opted for the rectified linear unit, as it is considered computationally less expensive than other popular activation functions like tanh and sigmoid (Xu et al. 2015; Ali & Shahbaz 2020).

Time lag adjustment

The basin hydrological response for a variable may differ as it depends upon several factors, e.g., heterogeneity, the format of climatic variables (snow), basin characteristics (slope, area, shape), ground roughness, absorption, and obstructions (reservoirs). These factors dampen or delay the hydrological response, which can obscure the relationship between the input and output variables. To get the most relatable inputs, variables are configured based on different lag times. For that reason, the time lags for each variable were sequentially increased from one to several days until the highest correlation was achieved against the output. The optimum lag time was determined for the following variables: P, AP, Tm, Fswat, and Fmswat. It is to be noted that in the case of ANN-RGC, the most influential variables in each climate zone were assessed based on the highest correlation. First, the time lags were configured to give the highest correlation with the output variable, and then the correlations of variables presented in six climate zones were compared to each other and the highest correlated variable was selected for further analysis.

The cross-correlation analysis, for both RGC and OC, is depicted in Figure 2. It showed that the precipitation variables (P and AP) have a longer delay and weak correlation with the flow. However, comparatively speaking, AP, being an improved parameter of P, showed better correlation and relatively low lag time. Furthermore, the Tm shared a strong positive correlation with the flow and a low delay time. The analysis seems to suggest that most of the winter precipitation that accumulates as snowpack was unable to directly translate its response into the flow. This justifies the larger delay time and weaker correlation with the flow. However, melting occurs as the temperature rises, which highlights a positive correlation. In the case of RGC, variables with the highest correlated lag time were zone-1 precipitation with 146 days lag (Pzone1t = 146), zone-2 accumulated precipitation with 60-day lag (APzone2t=60), and zone-1 N-days lagged mean temperature with 11 days lag (Tmzone1t=11). For OC, the following variables are determined to be highly correlated: precipitation with 135 days lag (Pt=135), accumulated precipitation with 42 days lag (APt=42), and N-days lagged mean temperature with 13 days lag (Tmt=13).
Figure 2

Pearson cross-correlation analysis of P, AP, Tm, Fswat, and Fmswat time-lagged variables corresponding to RGC and OC climate datasets.

Figure 2

Pearson cross-correlation analysis of P, AP, Tm, Fswat, and Fmswat time-lagged variables corresponding to RGC and OC climate datasets.

Close modal

The cross-correlation of hybrid variables showed contrasting results for RGC and OC. The H2-RGC time-adjusted simulated flow variables provided a decent relation with the output flow. However, the H2-OC simulated flow variables were unable to track the trend of the observed flow. This might be credited to a better distribution and basin representative climate for SWAT simulation. In the case of RGC, maximum correlated variables were N-days lagged SWAT flow simulation with 40 days lag (Fswatt=40) and N-days lagged mean SWAT flow simulation with 88 days lags (Fmswatt=88). In the case of OC, the N-days lagged SWAT flow simulation with 135 days lag (Fswatt=40) and N-days lagged mean SWAT flow simulation with 40 day lags (Fmswatt=88) were found to be highly correlated.

Input configurations

After determining the highest correlated variables, all the variables (lagged and non-lagged) are then configured in several configurations to investigate the combined relation with the target. The influence of these combinations of inputs is then investigated and ranked according to the highest value of correlation. For this, we adapted the methodology of Mekonnen et al. (2012) for the modeling, which considered a basic single hidden layer with one neuron to link multiple input configurations. The MSE was used as the training metric with LM as the backpropagation algorithm, while Pearson's correlation was used to evaluate each combination.

The result analysis as presented in Table 3 seems to suggest that, in the case of ANN-RGC, the maximum correlation was achieved when all the given climate variables are combined. Moreover, compared to ANN-OC, ANN-RGC seems to provide a slightly better correlation with the flow. The traditional hybrid models (H1-OC and H1-RGC), however, show less affiliation with their target variable, while H2-OC and H2-RGC provided a reasonably better correlation with the flow. Due to low performance, H1 models were discarded for further analysis

Table 3

Input combination and performance of models (ANN-RGC, ANN-OC, H1-RGC, H1-OC, H2-RGC, and H2-OC) corresponding to their target variables

Input configurationsTargetR
ANN-RGC   
Pzone1t=146 Q 0.210 
Pzone1t=146, APzone2t=60 Q 0.502 
Pzone1t=146, APzone2t=60, Tzone4 Q 0.790 
Pzone1t=146, APzone2t=60, Tzone4, Tmzone1t=11 Q 0.834 
 APzone2t=60, Tzone4, Tmzone1t=11 Q 0.834 
ANN-OC   
Pt=135 Q 0.168 
Pt=135, APt=42 Q 0.394 
Pt=135, APt=42, T Q 0.792 
Pt=135, APt=42, T, Tmt=13 Q 0.801 
APt=42, T, Tmt=13 Q 0.804 
H1-RCG   
 Fsim Rf 0.483 
 Fsim, APzone2t=60 Rf 0.483 
 Fsim, APzone2t=60, Tmzone1t=11 Rf 0.491 
Fsim, APzone2t=60, Tzone4, Tmzone1t=11 Rf 0.512 
 Fsim, Tzone4, Tmzone1t=11 Rf 0.511 
H1-OC   
 Fsim Rf 0.428 
 Fsim, APt=42 Rf 0.432 
Fsim, APt=42, Tmt=13 Rf 0.433 
 Fsim, APt=42, T, Tmt=13 Rf 0.431 
 Fsim, T, Tmt=11 Rf 0.429 
H2-RCG   
 Fswatt=40 Q 0.841 
 Fswatt=40, Fmswatt=88 Q 0.865 
 Fswatt=40, Fmswatt=88, APzone2t=60 Q 0.876 
Fswatt=40, Fmswatt=88, APzone2t=60, Tmzone1t=11 Q 0.881 
 Fswatt=40, Fmswatt=88, APzone2t=60, Tzone4, Tmzone1t=11 Q 0.876 
 Fswatt=40, Fmswatt=88, Tzone4, Tmzone1t=11 Q 0.873 
H2-OC   
 Fswatt=135 Q 0.110 
 Fswatt=135, Fmswatt=40 Q 0.322 
 Fswatt=135, Fmswatt=40, APt=42 Q 0.359 
 Fswatt=135, Fmswatt=40, APt=42, Tmt=13 Q 0.772 
Fswatt=135, Fmswatt=40, APt=42, T, Tmt=13 Q 0.793 
 Fswatt=135, Fmswatt=40, T, Tmt=11 Q 0.778 
Input configurationsTargetR
ANN-RGC   
Pzone1t=146 Q 0.210 
Pzone1t=146, APzone2t=60 Q 0.502 
Pzone1t=146, APzone2t=60, Tzone4 Q 0.790 
Pzone1t=146, APzone2t=60, Tzone4, Tmzone1t=11 Q 0.834 
 APzone2t=60, Tzone4, Tmzone1t=11 Q 0.834 
ANN-OC   
Pt=135 Q 0.168 
Pt=135, APt=42 Q 0.394 
Pt=135, APt=42, T Q 0.792 
Pt=135, APt=42, T, Tmt=13 Q 0.801 
APt=42, T, Tmt=13 Q 0.804 
H1-RCG   
 Fsim Rf 0.483 
 Fsim, APzone2t=60 Rf 0.483 
 Fsim, APzone2t=60, Tmzone1t=11 Rf 0.491 
Fsim, APzone2t=60, Tzone4, Tmzone1t=11 Rf 0.512 
 Fsim, Tzone4, Tmzone1t=11 Rf 0.511 
H1-OC   
 Fsim Rf 0.428 
 Fsim, APt=42 Rf 0.432 
Fsim, APt=42, Tmt=13 Rf 0.433 
 Fsim, APt=42, T, Tmt=13 Rf 0.431 
 Fsim, T, Tmt=11 Rf 0.429 
H2-RCG   
 Fswatt=40 Q 0.841 
 Fswatt=40, Fmswatt=88 Q 0.865 
 Fswatt=40, Fmswatt=88, APzone2t=60 Q 0.876 
Fswatt=40, Fmswatt=88, APzone2t=60, Tmzone1t=11 Q 0.881 
 Fswatt=40, Fmswatt=88, APzone2t=60, Tzone4, Tmzone1t=11 Q 0.876 
 Fswatt=40, Fmswatt=88, Tzone4, Tmzone1t=11 Q 0.873 
H2-OC   
 Fswatt=135 Q 0.110 
 Fswatt=135, Fmswatt=40 Q 0.322 
 Fswatt=135, Fmswatt=40, APt=42 Q 0.359 
 Fswatt=135, Fmswatt=40, APt=42, Tmt=13 Q 0.772 
Fswatt=135, Fmswatt=40, APt=42, T, Tmt=13 Q 0.793 
 Fswatt=135, Fmswatt=40, T, Tmt=11 Q 0.778 

Best configurations are highlighted in bold.

Number of hidden neurons

The determination of the number of hidden neurons was computed using a detailed sensitivity analysis. The selection criteria were based on the model structure complexity, optimum simulative accuracy, and training quality. For both ANN and hybrid models, the number of hidden neurons was increased from 2 to 15. A single model was then retrained 1,000 times using the LM backpropagation algorithm. In this study, a parsimonious approach was adapted for simplicity and to facilitate straightforward model replication for industry practitioners. For selecting the most parsimonious model in terms of structural complexity, we used the BIC criterion. For investigating the simulative accuracy and training quality, the median RMSE of 1,000-time retrained models was calculated, and discrepancies between the quantiles were observed. The testing period (the validation period of the SWAT) was used as the evaluation period to detect and avoid overfitting.

The results are shown in Figure 3, synoptically suggesting that an increment in the number of hidden neurons also leads to an increase in simulative accuracy of the model and reduces the discrepancy between the range. For almost all the models, the rate of improvement starts settling down after the 10th neuron. The BIC analysis showed that the modeling structure complexity also behaves analogously, except that a slight rise was observed beyond the 11th neuron. This seems indicative of the fact that the range for optimal models lies beyond the 10th neuron. Similarly, for ANN-RGC, H2-RGC, and H2-OC, simulation accuracy as represented by median RMSE (red line) and variability (lower and upper quantiles) seem to attain a limit state at the 12th neuron, while the structure complexity as depicted by BIC showed a slight rise from there on. However, in the case of ANN-OC, the analysis showed that simulative accuracy's rate of reduction settles around the 11th neuron. The variability also seemed to be in the appropriate range but could be better at the 13th neuron. However, as the BIC value began to show a slight increase at the addition of about the 10th to 11th neuron, after careful examination, we settled for 12 numbers of neurons for ANN-RGC, H2-RGC, and H2-OC models and 11 numbers of neurons for the ANN-OC model.
Figure 3

Models simulative accuracy, training quality, and structure complexity performance evaluation as the number of hidden neurons increases from 2 to 15.

Figure 3

Models simulative accuracy, training quality, and structure complexity performance evaluation as the number of hidden neurons increases from 2 to 15.

Close modal

Model bias removal

Model simulations normally inherit model biases that tend to induce repetitive discrepancies between observed and simulated data. The hydrological studies underline this fact, which could lead to the over- or underestimation of the flow during a particular period (month, season, etc.). For our study, we proposed to treat such simulation biases using simple linear scaling techniques. The basic procedure to apply linear bias correction to flow is similar to that of applying bias correction on precipitation datasets. The observed and simulated flows of the calibration period were taken as reference for calculating the 12 mean monthly factors, which were then multiplied with the daily simulation (calibration and validation period) datasets. More details of linear bias correction are explained in Section 2.3.

After the determination of suitable parameters for ANN models, the latter were trained on data spanning the period of 2003–2010 (8 years) and allowed us to simulate flow for the period of 2011–2015 (5 years). These models were then compared with SWAT-calibrated simulations using statistical objective functions. In total, five objective functions (NSE, KGE, R2, PBIAS, and RMSE) listed in Table 1 were used to evaluate the model performance for calibration and validation periods. The cumulative scores of calibration, validation, and for the whole period were calculated for each model. The performances of the six models were then graded based on model scores. The scoring scheme is devised in such a manner as to award the same score to two models if their performance is the same, while the rest of the models are ranked out of the remaining number of models. For instance, if the two top-ranked models showed the same performance, then they will be awarded a score of 6 each, while the next model will be awarded a score of 4 and not 5.

The model performance and scores in Table 4 reveal that the RGC climate dataset highly influenced the model inferences and distinctively improved the simulation. This could be attributed to the good quality of the dataset as well as its being well distributed. In addition, ANN-RGC and H2-RGC were able to provide better performance during validation rather than the calibration period. This might be due to the large quantity and diversity of the training pool for learning and the less quantity and diversity of the testing pool for forecasting. This leads to reduced variability and improved trend formation during validation, thereby demonstrating enhancements in metrics such as NSE, KGE, and R2. During the calibration phase, the performance for both ANN-RGC and SWAT-RGC models was nearly identical. ANN-RGC provided better PBIAS and RMSE values, while SWAT-RGC showed slightly better R2. The two models provided superior results compared to the other models and scored 26 and 25 points, respectively. The performance of H2-RGC was also closely behind ANN-RGC and SWAT-RGC as it scored 22 points. Among the OC-based models, SWAT-OC was able to outperform other ANN and hybrid-based frameworks and scored 14 points. In the validation phase, both ANN-RGC and H2-RGC fostered a good relationship with the observed flow as compared to SWAT-RGC and secured outstanding results. Between H2-RGC and ANN-RGC, the H2-RGC model provided better values of NSE, R2, and RMSE. H2-RGC and ANN-RGC obtained 28 and 27 points, respectively. The performance of SWAT-RGC, however, remained moderately decent in the validation phase with 20 points. The scores of SWAT-OC (12 points) and H2-OC (13 points) remained near to each other and, relatively speaking, higher compared to ANN-based models (9 points).

Table 4

Statistical performance and scores of each model on calibration and validation periods, without bias correction for flow

 
 

Generally, the scores of all the models fall in a suitable range. However, the ANN-RGC model was found to be more robust in simulating the flow conditions. In contrast, ANN-OC was found to be the lowest performer among all the other models for the OC climate dataset. This shows that the ANN is more sensitive to the quality and quantity of available data compared to other types of modeling frameworks (SWAT and hybrid models). Overall, the ANN obtained 53 and 22 points for ANN-RGC and ANN-OC models, respectively. The process-based SWAT model gives a decent performance under all climate conditions and was found to be the highest performer alongside the hybrid model for the OC climate conditions. The SWAT-RGC and SWAT-OC overall secured 45 and 26 points, respectively. On the other hand, H2-RGC was the second highest performer. The hybrid model also showed suitable performance when it was tested with the OC climate dataset. This balanced performance also indicates that the hybrid framework has taken advantage of both modeling frameworks (ANN and SWAT). Total scores of H2-RGC and H2-OC stood at 50 and 26 points, respectively.

Furthermore, the results of bias-corrected flow simulations of models are shown in Table 5. It shows considerable improvement in all model inferences during calibration; however, examining the effect during the validation period is more crucial. The results for the validation period show improvement in most of the model simulations. It further revealed that the bias correction generally led to improvement in results pertaining to the low-scoring models, for instance, ANN-OC, where considerable improvements were observed to occur. Similarly, improvement can also be observed for SWAT-RGC, SWAT-OC, and H2-OC. However, no considerable improvement was observed in the case of ANN-RGC and H2-RGC; this might be either because calibration results were lower than the validation period, so bias scaling was ineffective to improve validation results, or the linear bias correction did not significantly change the already good modeling results.

Table 5

Statistical performance and scores of each model on calibration and validation periods, with bias correction for flow

 
 

Flow series assessment

The statistical evaluation only highlights the overall performance of a model and tends to overlook the sequential variation in flows. Figures 46 depict the comparison of the model flow simulation, bias-corrected flow simulation, and the observed flow during the validation period (2011–2015).
Figure 4

Comparison of observation, SWAT simulation, and bias-corrected SWAT simulation on validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Figure 4

Comparison of observation, SWAT simulation, and bias-corrected SWAT simulation on validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Close modal
Figure 5

Comparison between observation, ANN simulation, and bias-corrected ANN simulation in validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Figure 5

Comparison between observation, ANN simulation, and bias-corrected ANN simulation in validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Close modal
Figure 6

Comparison of observation, H2 simulation, and bias-corrected H2 simulation on validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Figure 6

Comparison of observation, H2 simulation, and bias-corrected H2 simulation on validation period simulation. (a) Model based on RGC climate data and (b) model based on OC climate data.

Close modal

In the case of the SWAT model (Figure 4), we note that both SWAT-RGC and SWAT-OC were able to adequately capture the flow trends. However, simple model simulations by both models showed consistent overestimation during the recession period (September onward). Furthermore, the model simulation also seems to exaggerate the peak flow of the year 2015. On the other hand, bias correction of model simulations was effectively able to reconcile the recession and the baseflow biases but was unable to reduce the exaggeration in peak flow, adding slightly more to it. The ANN models (ANN-RGC and ANN-OC) are depicted in Figure 5. The ANN-RGC seems to provide a good fit with observed flow. The ANN-OC model showed a large mismatch with the observed flow data in the year 2011. The rest of the years seem to be normal; however, slight underestimation of the peak flow and overestimation during the recession period was observed. After bias correction, ANN-OC was able to reduce the error; however, the discrepancy relating to simulation results for the year 2011 persisted. Overall, the hybrid models (H2-RGC and H2-OC) shown in Figure 6 provide a representative simulation, but following ANN-OC, H2-OC also depicted uncertain flow conditions in the year 2011. Similarly, bias correction was only slightly able to reduce the flow uncertainty.

Recently, various physical state-of-art models were deployed for improving the understanding of the climate variability of over the Chitral river catchment. By comparing the findings of this study with these contemporaneous research efforts, we can gain a more thorough understanding of the progress made as well as the implications and potential avenues for future research. One recent study that utilized the hydrological model based on least action principle (HyMoLAP), along with a modified snowmelt model named HoMALP-SM, was employed to study the Chitral river basin (Hassan & Khan 2022). The results indicate that HoMALP-SM excels in assimilating flow conditions, yielding impressive metrics such as RMSE, NSE, and CC (Correlation Coefficient) of 118.5, 0.811, and 0.911, respectively. In the context of our study, a comparison between the baseline model and these findings reveals that the HoMALP-SM model by Hassan & Khan (2022) is on par with the SWAT-RGC and SWAT-OC models. However, ANN-RGC and H2-RGC exhibit even better simulation capabilities. Similarly, Usman et al. (2020) applied the conceptual HBV (Hydrologiska Byråns Vattenbalansavdelning)-light model to simulate Chitral river flow. Their model demonstrated promising results, achieving NSE, R2, and PBIAS scores of 0.91, 0.91, and 3.7% during calibration and 0.81, 0.82, and −2% during validation, respectively. By comparison, these outcomes outperformed the results of the SWAT-RGC and SWAT-OC models overall and showed slight improvements over the ANN-RGC and H2-RGC models during the calibration phase. However, notably, the ANN-based model consistently performed well and produced superior results in the validation period. These findings underscore the significance of improved climate data and a more profound understanding of ANN models in advancing baseline models, thereby demonstrating the need for the use of ANN for enhancing simulations.

In this study, we proposed to use a combination of modeling strategies to model the hydrological response of the Chitral river basin in northwestern Pakistan under complex physiographic and climatic conditions. Practical applications of the present study extend from flood management to irrigation management, reservoir operation, and hydroclimatic modeling.

The strategies utilized in this study comprised of physical-based hydrological modeling (SWAT), ANN, and SWAT–ANN hybrid hydrological modeling. In order to fully capture the climate variability, the hydrological model was built using two types of climate sets, namely, the OC and the RGC (Dahri et al. 2018) to assess the model performance under uncertain and corrected climate conditions. The SWAT models were configured and auto-calibrated using SWAT-CUP, whereas the uncalibrated SWAT flow simulation was used for the hybrid modeling. In the case of ANN and hybrid modeling, we assumed that for robust inference, the input parameters must be adjusted to give a strong correlation with the output variables; as a result, the climate and flow simulating parameters were time lag-adjusted corresponding to the flow. The correlation results suggested that the RGC-based configurations seem to provide a better affiliation with the flow. In the case of hybrid modeling, we also explored traditional hybrid modeling (H1) and proposed a hybrid approach (H2); however, traditional hybrid modeling gave poor results and was not further investigated. Furthermore, we also analyzed for the optimal number of hidden neurons after adjusting for the time lag and assessed on three criteria, namely, simulating accuracy, training quality (using RMSE discrepancy evaluation), and model complexity (BIC). In total, we considered 15 hidden neurons, and each model was retrained 1,000 times to determine the variability of the modeling yield. It was concluded that the optimal number of hidden neurons is 12 for ANN-RGC, H2-RGC, and H2-OC, while this number stands at 11 for ANN-OC.

The different models were trained on data pertaining to years 2003–2010, i.e., 8 years, using previously determined configurations and the number of neurons and allowed us to simulate flow from 2011 to 2015 (5 years). For model evaluation, we used five objective functions (NSE, KGE, R2, PBIAS, and RMSE), and for qualifying the overall performance, a scoring scheme was used to calculate model scores using the aforementioned objective functions. The results demonstrated that the climate data highly influenced the model inferences and distinctly improved the simulation compared to OC. The ANN-RGC model was found to be more robust in simulating the flow conditions, followed by H2-RGC and SWAT-RGC. In the case of OC models, SWAT-OC and H2-OC performed equally well; however, ANN-OC slightly underperformed the other OC-based models and stood last in the rankings. This research also highlighted the fact that the ANN model seems to be most sensitive to data quality compared to other modeling frameworks (SWAT and H2). The SWAT performance remained moderate and balanced, while H2 was among the high-performing models and even performed relatively well in OC climate conditions. This study also attempted to further refine the models, especially against model biases using a linear scaling for bias removal. The bias removal has a beneficial effect on results as statistical evaluation demonstrated considerable improvements in validation simulation. Furthermore, the bias correction brought considerable improvements in low-scoring models. The flow simulation examination of the validation/forecasting period reveals that RGC-based models are more effective in capturing the annual trends. Although some slight exaggeration was observed during the recession (September onwards and base flows of SWAT-RGC simulation results, it can be treated by applying a simple linear bias correction. On the other hand, the OC-based flow simulations were not able to match for all the years and showed significant flow exaggeration; for instance, SWAT-OC showed an overestimation of peak flows in 2015, while ANN-OC and H2-OC showed uncertain flow conditions in the year 2011. Furthermore, the comparison of recent advanced models in the Chitral river catchment study underscores the value of improved climate data and emphasizes on the utility of ANN for deeper understanding as these are pivotal for advancing baseline models, enhancing simulation accuracy.

Generally, while this study employs a justifiable methodology backed by peer-reviewed literature, it is important to acknowledge lingering limitations. It is notable that despite the convenience of ANNs, they necessitate abundant data for effective training, unlike physically based models that lack such a requirement. In scenarios with limited data, physically based models might be more appropriate. Another challenge pertains to hydroclimatology's nonstationarity, especially evident in the era of climate change. ANNs struggle with predicting values beyond their training scope, necessitating training with longer records and extreme flow variables. In addition, as highlighted earlier, streamflow responds to climate variables, impacting predicted values. Hence, forecast model validity could be contingent on the climate change intensity.

It is concluded that a reliable climate dataset holds primary importance for reliable hydrological modeling for all types of modeling (process-based, DD, etc.). The climate distribution and the quality of the dataset play a pivotal role in generating hydrological signals for the model to capture the flow variations and trends. This research also laid emphasis on data adjustment (bias correction, time lag adjustment, configuration selection, etc.) and proper parameterization for improvements in model inferences. Furthermore, despite ANN's superiority over the conventional models, it could be susceptible to limited performance, especially in uncertain or limited data conditions. Under these conditions, conventional or hybrid models could be a better option.

S.H. and S.A. contributed to the conceptualization of the article and to acquiring the resources needed for the article. Z.S. performed the data curation. Z.S. and P.M. carried out the formal analysis and the investigation. S.A. contributed to formulating the methodology. S.H. and Z.S. contributed to writing the original draft of the article.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.

Abbas
A.
,
Waseem
M.
,
Ullah
W.
,
Zhao
C.
&
Zhu
J.
2021
Spatiotemporal analysis of meteorological and hydrological droughts and their propagations
.
Water
13
,
2237
.
doi:10.3390/w13162237
.
Abbas
A.
,
Ullah
S.
,
Ullah
W.
,
Waseem
M.
,
Dou
X.
,
Zhao
C.
,
Karim
A.
,
Zhu
J.
,
Hagan
D.
,
Bhatti
A.
&
Ali
G.
2022
Evaluation and projection of precipitation in Pakistan using the Coupled Model Intercomparison Project Phase 6 model simulations
.
International Journal of Climatology
.
doi:10.1002/joc.7602
.
Abbas
A.
,
Ullah
S.
,
Ullah
W.
,
Zhao
C.
,
Karim
A.
,
Waseem
M.
,
Bhatti
A.
,
Ali
G.
,
Jan
M.
&
Ali
A.
2023
Characteristics of winter precipitation over Pakistan and possible causes during 1981–2018
.
Water
2023
,
2420
.
doi:10.3390/w15132420
.
Ahmad
B.
,
Rana
A. S.
,
Abuzar
K.
,
Kiran
R.
,
Mansoor
R.
&
Kubra
S.
2016
Diagnostic study of heavy downpour in 2015 flash floods over Chitral Area, Northern Pakistan
.
Pakistan Journal of Meteorology
12
(
24
).
Ahmad
S.
,
Israr
M.
,
Liu
S.
,
Hayat
H.
,
Gul
J.
,
Wajid
S.
,
Ashraf
M.
,
Baig
S.
&
Tahir
A.
2018
Spatio-temporal trends in snow extent and their linkage to hydro-climatological and topographical factors in the Chitral River Basin (Hindukush, Pakistan)
.
Geocarto International
35
(
September
),
1
35
.
https://doi.org/10.1080/10106049.2018.1524517
.
Ali
S.
&
Shahbaz
M.
2020
Streamflow forecasting by modeling the rainfall–streamflow relationship using artificial neural networks
.
Modeling Earth Systems and Environment
6
(
3
),
1645
1656
.
https://doi.org/10.1007/s40808-020-00780-3
.
Ali
A.
,
Raheem
A.
,
Khidir
K.
&
Alkattan
M. Q. M.
2019
Predicting of daily Khazir Basin flow using SWAT and hybrid SWAT-ANN models
.
Ain Shams Engineering Journal
11
(
November
).
https://doi.org/10.1016/j.asej.2019.10.011
.
Alizadeh
M. J.
,
Kavianpour
M.
,
Kisi
O.
&
Nourani
V.
2017
A new approach for simulating and forecasting the rainfall-runoff process within the next two months
.
Journal of Hydrology
548
.
doi:10.1016/j.jhydrol.2017.03.032
.
Arnold
J. G.
,
Srinivasan
R.
,
Muttiah
R. S.
&
Williams
J. R.
1998
Large area hydrologic modeling and assessment part I: Model development
.
Journal of the American Water Resources Association
34
(
1
),
73
89
.
https://doi.org/10.1111/j.1752-1688.1998.tb05961.x
.
Baudouin
J.-P.
,
Herzog
M.
&
Petrie
C.
2020
Cross-validating precipitation datasets in the Indus River Basin
.
Hydrology and Earth System Sciences
24
(
January
),
427
450
.
https://doi.org/10.5194/hess-24-427-2020
.
Bookhagen
B.
&
Burbank
D.
2006
Topography, relief, and TRMM-derived rainfall variations along the Himalaya
.
Geophysical Research Letters
33
(
April
).
https://doi.org/10.1029/2006GL026944
.
Dahri
Z.
,
Moors
E.
,
Ludwig
F.
,
Ahmad
S.
,
Khan
A.
,
Ali
I.
&
Kabat
P.
2018
Adjustment of measurement errors to reconcile precipitation distribution in the high-altitude Indus basin
.
International Journal of Climatology
38
(10), 3842–3860.
https://doi.org/10.1002/joc.5539
.
Dahri
Z.
,
Ludwig
F.
,
Moors
E.
,
Ahmad
S.
,
Ahmad
B.
,
Ahmad
S.
,
Riaz
M.
&
Kabat
P.
2021
Climate change and hydrological regime of the high-altitude Indus basin under extreme climate scenarios
.
Science of the Total Environment
768
(
January
),
144467
.
https://doi.org/10.1016/j.scitotenv.2020.144467
.
Elshorbagy
A.
&
Simonovic
S.
2002
Estimation of missing streamflow data using principles of chaos theory
.
Journal of Hydrology
255
(
January
),
123
133
.
https://doi.org/10.1016/S0022-1694(01)00513-3
.
Food and Agriculture Organization (FAO) and International Institute for Applied Systems Analysis (IIASA)
.
2023
Harmonized World Soil Database Version 2.0. Rome and Laxenburg. https://doi.org/10.4060/cc3823en.
Ghoraba
S. M.
2015
Hydrological modeling of the Simly Dam watershed (Pakistan) using GIS and SWAT model
.
Alexandria Engineering Journal
54
(
3
),
583
594
.
https://doi.org/10.1016/j.aej.2015.05.018
.
Grayson
R. B.
,
Blöschl
G.
,
Western
A. W.
&
McMahon
T. A.
2002
Advances in the use of observed spatial patterns of catchment hydrological response
.
Advances in water resources
25
(
8-12
),
1313
1334
.
Grusson
Y.
,
Sun
X.
,
Gascoin
S.
,
Sauvage
S.
,
Raghavan
S.
,
Anctil
F.
&
Sáchez-Pérez
J.-M.
2015
Assessing the capability of the SWAT model to simulate snow, snow melt and streamflow dynamics over an alpine watershed
.
Journal of Hydrology
531
(
December
),
574
588
.
https://doi.org/10.1016/j.jhydrol.2015.10.070
.
Hasson
S. u.
,
Saeed
F.
,
Böhner
J.
&
Schleussner
C.-F.
2019
Water availability in Pakistan from Hindukush-Karakoram-Himalayan Watersheds at 1.5°C and 2°C Paris Agreement Targets
.
Advances in Water Resources
131
(
June
).
https://doi.org/10.1016/j.advwatres.2019.06.010
.
Hersbach
H.
,
Bell
B.
,
Berrisford
P.
,
Hirahara
S.
,
Horányi
A.
,
Muñoz-Sabater
J.
,
Nicolas
J.
,
Peubey
C.
,
Radu
R.
,
Schepers
D.
,
Simmons
A.
,
Soci
C.
,
Abdalla
S.
,
Abellan
X.
,
Balsamo
G.
,
Bechtold
P.
,
Biavati
G.
,
Bidlot
J.
,
Bonavita
M.
&
Thépaut
J. N.
2020
The ERA5 global reanalysis
Quarterly Journal of the Royal Meteorological Society
146
(
730
),
1999
2049
.
Hewitt
K.
2009
The Karakoram anomaly? Glacier expansion and the ‘elevation effect,’ Karakoram Himalaya
.
Mountain Research and Development
25
(
January
),
332
340
.
https://doi.org/10.1659/0276-4741(2005)025[0332:TKAGEA]2.0.CO;2
.
Horan
R.
,
Gowri
R.
,
Wable
P.
,
Baron
H.
,
Keller
V.
,
Garg
K.
,
Mujumdar
P.
,
Houghton-Carr
H.
&
Rees
G.
2021
A comparative assessment of hydrological models in the Upper Cauvery Catchment
.
Water
13
(
January
),
151
.
https://doi.org/10.3390/w13020151
.
Hosseini
H.
,
Hashemi
H.
,
Fard
A. F.
&
Berndtsson
R.
2022
Areal precipitation coverage ratio for enhanced AI modelling of monthly runoff: A new satellite data-driven scheme for semi-arid mountainous climate
.
Remote Sensing
14
(
January
),
270
.
https://doi.org/10.3390/rs14020270
.
Immerzeel
W. W.
,
Petersen
L.
,
Ragettli
S.
&
Pellicciotti
F.
2014
The importance of observed gradients of air temperature and precipitation for modeling runoff from a glacierized watershed in the Nepalese Himalayas
.
Water Resources Research
50
(
March
).
https://doi.org/10.1002/2013WR014506
.
Immerzeel
W. W.
,
Wanders
N.
,
Lutz
A.
,
Shea
J.
&
Bierkens
M. F. P.
2015
Reconciling high altitude precipitation in the upper Indus basin with glacier mass balances and runoff
.
Hydrology and Earth System Sciences Discussions
12
(
May
),
4755
4784
.
https://doi.org/10.5194/hessd-12-4755-2015
.
Iqbal
M.
&
Hussain
A.
2017
Validation of satellite based precipitation over diverse topography of Pakistan
.
Atmospheric Research
201
(
October
).
https://doi.org/10.1016/j.atmosres.2017.10.026
.
Islam
Z.
2011
A Review on Physically Based Hydrologic Modeling. Technical Report. https://doi.org/10.13140/2.1.4544.5924
.
Jackson
E. K.
,
Roberts
W.
,
Nelsen
B.
,
Williams
G. P.
,
James Nelson
E.
&
Ames
D. P.
2019
Introductory overview: Error metrics for hydrologic modelling – A review of common practices and an open source library to facilitate use and adoption
.
Environmental Modelling and Software
119
(
September
),
32
48
.
https://doi.org/10.1016/j.envsoft.2019.05.001
.
Jimeno-Sáez
P.
,
Senent-Aparicio
J.
,
Pérez-Sánchez
J.
,
Pulido-velazquez
D.
&
Cecilia
J.
2017
Estimation of instantaneous peak flow using machine-learning models and empirical formula in peninsular Spain
.
Water
9
(
May
),
347
.
https://doi.org/10.3390/w9050347
.
Jimeno-Sáez
P.
,
Senent-Aparicio
J.
,
Pérez-Sánchez
J.
&
Pulido-velazquez
D.
2018
A comparison of SWAT and ANN models for daily runoff simulation in different climatic zones of peninsular Spain
.
Water
10
(
February
).
https://doi.org/10.3390/w10020192
.
Kişi
Ö.
2007
Streamflow forecasting using different artificial neural network algorithms
.
Journal of Hydrologic Engineering
12
(
5
),
532
539
.
https://doi.org/10.1061/(ASCE)1084-0699(2007)12:5(532)
.
Lenderink
G.
,
van Ulden
A.
,
van den Hurk
B.
&
Keller
F.
2007
A study on combining global and regional climate model results for generating climate scenarios of temperature and precipitation for The Netherlands
.
Climate Dynamics
29
(
2–3
),
157
176
.
https://doi.org/10.1007/s00382-007-0227-z
.
Lutz
A.
,
Immerzeel
W. W.
,
Kraaijenbrink
P.
,
Shrestha
A.
&
Bierkens
M. F. P.
2016
Climate change impacts on the upper Indus hydrology: Sources, shifts and extremes
.
PLoS One
11
(
November
).
https://doi.org/10.1371/journal.pone.0165630
.
Marahatta
S.
,
Devkota
L.
&
Aryal
D.
2021
Application of SWAT in hydrological simulation of complex mountainous river basin (part I: model development)
.
Water
13
(
May
),
1546
.
https://doi.org/10.3390/w13111546
.
Mahmood
P.
,
Haider
S.
,
Gabriel
H. F.
,
Shahid
M.
&
Syed
Z.
2022
Acceleration of flow modeling using a freeware 2D GPU-shallow water equations code
.
Arabian Journal of Geosciences
15
(
20
),
1584
.
Mekonnen
B.
,
Nazemi
A.
,
Elshorbagy
A.
,
Mazurek
K.
&
Putz
G.
2012
Hybrid modelling approach to prairie hydrology: Fusing data-driven and process-based hydrological models
.
Hydrological Sciences Journal
60
(
April
),
6562
.
https://doi.org/10.1080/02626667.2014.935778
.
Mikayilov
F.
,
Vejdani
M.
&
Haghighat
S.
2007
SWAT-CUP calibration and uncertainty programs for SWAT
. In:
Proceedings of the International Congress on Modelling and Simulation (MODSIM'07)
,
January
.
Mogano
P.
&
Mokoele
N.
2019
South African climate change adaptation politics: Urban governance prospects
.
International Journal of Social Sciences and Humanity Studies
11
(
1
),
68
83
.
Muñoz Sabater
J.
2019
ERA5-Land Hourly Data from 1950 to Present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). https://doi.org/10.24381/cds.e2161bac
.
Muñoz Sabater
J.
,
Dutra
E.
,
Agusti-Panareda
A.
,
Albergel
C.
,
Arduini
G.
,
Balsamo
G.
&
Boussetta
S.
,
Choulga
M.
,
Harrigan
S.
,
Hersbach
H.
,
Martens
B.
,
Miralles
D. G.
,
Piles
M.
,
Rodríguez-Fernández
N. J.
,
Zsoter
E.
,
Buontempo
C.
&
Thépaut
J. N.
2021
ERA5-land: A state-of-the-art global reanalysis dataset for land applications
.
Earth System Science Data
13
(
September
)
4349
4383
.
https://doi.org/10.5194/essd-13-4349-2021
.
Nam
L. P.
,
Dang Que
N.
,
Van Song
N.
,
Hoang Mai
T. T.
,
Minh Phuong
N. T.
,
Xuan Huong
N. T.
,
Cong Tiep
N.
&
Uan
T. B.
2022
Rice farmers' perception and determinants of climate change adaptation measures: a case study in Vietnam
.
AgBioForum
,
24
(
1
),
13
29
Nayebi
M. M.
,
Khalili
D.
,
Amin
S.
&
Zand-Parsa
S.
2006
Daily stream flow prediction capability of artificial neural networks as influenced by minimum air temperature data
.
Biosystems Engineering
95
(
December
),
557
567
.
https://doi.org/10.1016/j.biosystemseng.2006.08.012
.
Neitsch
S. L.
,
Arnold
J. G.
,
Kiniry
J. R.
&
Williams
J. R.
2011
.
Soil and water assessment tool theoretical documentation version 2009
.
Texas Water Resources Institute
.
Nguyen Khoi
D.
2015
Comparison of the HEC-HMS and SWAT hydrological models in simulating the streamflow
.
Tạp Chí Khoa Học và Công Nghệ
53
(
December
),
189
195
.
Noori
N.
&
Kalin
L.
2016
Coupling SWAT and ANN models for enhanced daily streamflow prediction
.
Journal of Hydrology
533
(
February
),
141
151
.
https://doi.org/10.1016/j.jhydrol.2015.11.050
.
Palazzi
E.
,
Hardenberg
J.
&
Provenzale
A.
2013
Precipitation in The Hindu-Kush Karakoram Himalaya: Observations and future scenarios
.
Journal of Geophysical Research
118
(
January
),
85
100
.
https://doi.org/10.1029/2012JD018697
.
Panu
U. S.
,
Khalil
M.
,
Elshorbagy
A.
,
2000
Streamflow data infilling techniques based on concepts of groups and neural networks
. In:
Artificial Neural Networks in Hydrology
(
Govindaraju
R. S.
&
Rao
A. R.
, eds).
Water Science and Technology Library, vol. 36
.
Springer
,
Dordrecht
, pp.
235
258
.
https://doi.org/10.1007/978-94-015-9341-0_13
.
Santos
C.
&
Silva
G.
2014
Daily streamflow forecasting using a wavelet transform and artificial neural network hybrid models
.
Hydrological Sciences Journal
59
,
312
324
.
doi:10.1080/02626667.2013.800944
.
Shiau
J.-T.
&
Hsu
H.-T.
2016
Suitability of ANN-based daily streamflow extension models: A case study of Gaoping River Basin, Taiwan
.
Water Resources Management
30
(
March
).
https://doi.org/10.1007/s11269-016-1235-8
.
Shrestha
M.
,
Acharya
S. C.
&
Shrestha
P. K.
2017
Bias correction of climate models for hydrological modelling – Are simple methods still useful?
Meteorological Applications
24
(
3
),
531
539
.
https://doi.org/10.1002/met.1655
.
Shrestha
S.
,
Shrestha
M.
&
Shrestha
P. K.
2018
Evaluation of the SWAT model performance for simulating river discharge in the Himalayan and tropical basins of Asia
.
Hydrology Research
49
(
3
),
846
860
.
https://doi.org/10.2166/nh.2017.189
.
Singh
A.
,
Imtiyaz
M.
,
Isaac
R.
&
Denis
D.
2011
Comparison of soil and water assessment tool (SWAT) and multilayer perceptron (MLP) artificial neural network for predicting sediment yield in the Nagwa Agricultural Watershed in Jharkhand, India
.
Agricultural Water Management
104
(
January
).
https://doi.org/10.1016/j.agwat.2011.12.005
.
Syed
Z.
,
Ahmad
S.
,
Dahri
Z.
,
Azmat
M.
,
Shoaib
M.
,
Inam
A.
,
Qamar
M.
,
Hussain
S.
&
Ahmad
S.
2022
Hydroclimatology of the Chitral River in the Indus Basin under changing climate
.
Atmosphere
13
(
February
),
295
.
https://doi.org/10.3390/atmos13020295
.
Syed
Z.
,
Mahmood
P.
,
Haider
S.
,
Ahmad
S.
,
Jadoon
K. Z.
,
Farooq
R.
,
Syed
S.
&
Ahmad
K.
2023
Short–long-term streamflow forecasting using a coupled wavelet transform–artificial neural network (WT–ANN) model at the Gilgit River Basin, Pakistan
.
Journal of Hydroinformatics
,
25
(
3
),
881
-
894
.
Teutschbein
C.
&
Seibert
J.
2012
Bias correction of regional climate model simulations for hydrological climate-change impact studies: Review and evaluation of different methods
.
Journal of Hydrology
456–457
(
August
),
12
29
.
https://doi.org/10.1016/j.jhydrol.2012.05.052
.
Touseef
M.
,
Chen
L.
&
Yang
W.
2021
Assessment of surface water availability under climate change using coupled SWAT-WEAP in Hongshui River Basin, China
.
International Journal of Geo-Information
10
(
May
).
https://doi.org/10.3390/ijgi10050298
.
Usman
M.
,
Pan
X.
,
Penna
D.
&
Ahmad
B.
2020
Hydrologic alteration and potential ecosystemic implications under a changing climate in the Chitral River, Hindukush region, Pakistan
.
Journal of Water and Climate Change
12
.
doi:10.2166/wcc.2020.073
.
Wang
Y.
,
Guo
S.
,
Xiong
L.
,
Liu
P.
&
Liu
D.
2015
Daily runoff forecasting model based on ANN and data preprocessing techniques
.
Water
2015
(
July
),
4144
4160
.
https://doi.org/10.3390/w7084144
.
Waseem
M.
,
Khursheed
T.
,
Abbas
A.
,
Ahmad
I.
&
Javed
Z.
2021
Impact of meteorological drought on agriculture production at different scales in Punjab, Pakistan
.
Journal of Water and Climate Change
13
.
doi:10.2166/wcc.2021.244
.
Winiger
M.
,
Gumpert
M.
&
Yamout
H.
2005
Karakorum-Hindukush-Western Himalaya: Assessing high-altitude water resources
.
Hydrological Processes
19
(
August
),
2329
2338
.
https://doi.org/10.1002/hyp.5887
.
Xu
B.
,
Wang
N.
,
Chen
T.
&
Li
M.
2015
Empirical evaluation of rectified activations in convolutional network
. arXiv preprint arXiv:1505.00853.
Yang
J.
,
Reichert
P.
,
Abbaspour
K. C.
,
Xia
J.
&
Yang
H.
2008
Comparing uncertainty analysis techniques for a SWAT application to the Chaohe Basin in China
.
Journal of Hydrology
358
(
1–2
),
1
23
.
https://doi.org/10.1016/j.jhydrol.2008.05.012
.
Yaseen
Z. M.
,
El-shafie
A.
,
Jaafar
O.
,
Afan
H. A.
&
Sayl
K. N.
2015
Artificial intelligence based models for stream-flow forecasting: 2000–2015
.
Journal of Hydrology
530
(
November
),
829
844
.
https://doi.org/10.1016/j.jhydrol.2015.10.038
.
Yuan
F.
,
Zhang
L.
,
Win
K.
,
Ren
L.
,
Zhao
C.
,
Zhu
Y.
,
Jiang
S.
&
Liu
Y.
2017
Assessment of GPM and TRMM multi-satellite precipitation products in streamflow simulations in a data-sparse mountainous watershed in Myanmar
.
Remote Sensing
9
(
March
),
302
.
https://doi.org/10.3390/rs9030302
.

Author notes

These authors contributed equally to this work.

This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).