Performance evaluation of various hydrological models with respect to hydrological responses under climate change scenario: a review

Abstract

Studies reviewed in this paper show anomaly for temperature pertaining to streamflow and rainfall showing different trends, especially in Ethiopia to support the research findings and interpretation. There are many hydrological models, including 54 physically distributed, lumped, and conceptual hydrological models, of which 28 have been used in Ethiopian river basins. The models include the most adaptable and commonly used SWAT model applicable from small areas up to large basins. It is indeed a challenge to use a single hydrological model as the data rely on consistency, limitation-free, and exactly fitted output. The overall performance of individual physically-based, conceptual, and machine learning (ML) models varied at different watersheds. Reasonably, ML performs very well, up to 0.99 for R² and NSE and up to 0.001 for PBIAS. Inopportunely, using a single hydrological model has its limitations; ensemble multi-individual models, coupling or hybridization of physical or conceptual models with machine learning, combining evolutionary optimization algorithms with ML, and also comparisons of multi-single hydrological models, and selecting the best one are recommended options. No single hydrological model is indispensable and can be termed as better than the other for any watershed. Somewhat, ML outperforms SWAT but cannot be considered an absolute substitute. The size of the watershed, the number of data used, and the ratio between calibrations year to validation year do not have a clear correlation with the performance, particularly for the SWAT model accounted for in this review. Optimization algorithms explore multiple options and choosing the right one is a tedious task before a final decision is taken.

Keywords:

• Climate
• hydrological models
• machine learning
• SWAT
• watershed

REVIEWING EDITOR:

• Montemurro Marco, Arts et Métiers Sciences et Technologies, France

SUBJECTS:

• Civil, environmental and geotechnical engineering
• water engineering
• hydraulic engineering

1. Introduction

Water, defined as an essential natural resources, is considered to be finite and subject to change under various circumstances. Droughts and floods also pose hazards to society under unpredictable scenarios (Van Loon, 2015; Cenobio-Cruz et al., 2023; Nivesh et al., 2022). Planning, policymakers, and managers need to know the amount of streamflow to effectively allocate water resources in the watersheds they are responsible for (Alamdari et al., 2017). In most developing and impoverished countries, direct streamflow monitoring is quite challenging due to lack of funds and inadequate equipment. As a result, on the condition where stream flow has not been recorded, the hydrological model is obligatory to estimate the water availability at the relevant location and real time (Sanjay Shekar & Vinay, 2021; Tibangayuka et al., 2022).

The hydrological model can partially or fully represent the water cycle and can interpolate and extrapolate the hydrological cycle components of catchments both spatially and temporally (Fniguire et al., 2022). Events and patterns related to stream flow are hard to predict. This is because, stream flow is incredibly complicated, dynamic, non-stationary, and non-linear (Yaseen et al., 2015). Many water professionals worldwide are meticulous in reducing the uncertainty of hydrological models used for determining stream flow discharge to enhance the sustainable management and development of water resources (Sempewo et al., 2022).

The estimated stream flow or runoff from a catchment is useful for many purposes, such as determining flood peaks, determining the availability of water for municipal needs, designing storage facilities for a range of uses, organizing irrigation operations for agricultural or other industrial uses, protecting wildlife, and securing future supplies of potable water (Ranjan & Singh, 2022). The proficiency and accuracy of flow prediction models can have a direct impact on management decisions about water resources (Noori & Kalin, 2016). Although, there are many hydrological models available to forecast current and future stream flow, choosing a specific model is a challenge because of the uncertainty in model prediction caused by input factors (climate and land use), model parameters, consideration of scale in watershed modeling, and methods of calibration (Veettil et al., 2021).

The ideal model uses the fewest parameters and the least amount of model complexity while producing results that are very similar to reality (Suliman et al., 2015). To avoid needless costs resulting from poorly informed predicted planning, decision-makers must rely on reliable hydrological modeling results (Mubialiwo et al., 2021).

2. Effect of climate on the catchments hydrology response

Climate change, which has laid down attention recently, is one of the major global challenges of the 21st century by altering the catchment hydrological process (Shahid et al., 2018; Merga et al., 2022; Kamilia et al., 2024). Based on scientific pieces of evidence, an increase in greenhouse gas emissions might cause an average temperature rise of 1.4–5.8 °C by the end of the century. There will be noticeable repercussions from climate change, particularly in poorer countries where agriculture is a significant economic factor (Worqlul et al., 2018). Currently, there are many climate effects on catchments across both developing countries, including Ethiopia and developed countries.

Using the Liuxihe (LXH) model, Sheng et al. (2023) evaluated how the hydrology of the Jinshan River (455,000 km²) in China’s upper Yangtze River basin has been affected by climate change. His findings showed that, for various global climate models, rainfall and temperature will rise by 0.7% to 11% and 1.0 to 2.1, respectively, over the next 30 years (up to 2050) in comparison to the base period. Runoff variations, on the other hand, show greater uncertainty and range from −5.6% to 11.0% without any discernible pattern. Not every increase in precipitation causes runoff to alter in the same way.

In the Huaihe River Basin (123,900 km²) in China, Bian et al. (2023) forecasted future changes in the flood regime using the Coupled model inter-comparison project (CMIP6) and the Xinanjiang hydrological model. The results demonstrated that there will likely be an increase in the average intensity, quantity, and extreme precipitation annually, as well as significant future increases in the average, maximum, and lowest temperature values. In the distant future (2071–2100), the annual peak flood is similarly predicted to rise by 46.4% (43.1%) and 45.4% (51.1%), respectively, at Wangjiaba station and Wujiadu station, based on the shared socioeconomic pathways (SSP5-8.5) scenario. Therefore, the findings of this study suggest that a technical body in the country should take necessary measures to minimize the disasters caused at the time of increased peak floods.

N. M. Reddy et al. (2023) used a conceptual lumped model called the Sacramento model with CMIP6 to study the effects of climate change on streamflow in the Godavari basin (5380 km²), India. According to the findings, there will be an increase in annual precipitation ranging from 9% to 55.0% in all SSP126, SSP245, SSP370, and SSP585 scenarios, as well as in all near, mid, and far future years. The SSP585 scenario is probably going to analyze the biggest future increase in precipitation. In the near future, the stream flow will also increase minimally by 47.12% and up to 127.48% in the distant future.

Using HEC-HMS, Lone (2022) determined the effects of hydro-climatic changes in the Sindh River in the Kashmir basin (1560.32 km²) and concluded that, while the maximum temperature in the spring will continue to decline in the near and far future, it will also decline in the near future during the winter. In contrast, the other three seasons will see an increase in the near and far future in both RCP4.5 and RCP8.5. The lowest temperature will decrease in the winter and spring and increase in the summer and autumn for both RCP4.5 and RCP8.5, as well as for both the near and distant future. In both the near and distant future, at RCP4.5 and RCP8.5, the amount of rainfall during the spring will decline while that of the other three seasons will rise. The yearly streamflow will rise in both the near future and long term in both RCP4.5 and RCP8.5, with a range of 2.8% to 18.3% in all three gauge sites. Nevertheless, because this study only used one general circulation model (GCM), the conclusions cannot be fully authenticated.

Using the SWAT model, Nilawar and Waikar (2018) determined how the Purna River basin (15,579 km²) of India is impacted by the response to climate change. The outcome showed that the maximum temperature will rise by 0.05 to 1.46 °C and that in the near and distant future, RCP8.5 will have a value that is comparatively larger than RCP4.5. Under RCP4.5 and RCP8.5, respectively, the anticipated average annual precipitation (PCP) increases by 1.22–5.59% (10.81 mm to 49.47 mm) and 2.79–12.29% (24.67 mm to 108.63 mm). In RCP4.5 and RCP8.5, as well as in all subsequent periods, the rainy season will rise and the dry season will fall. In one of the Congo River Basin’s big watersheds, under RCP4.5, the average percent change in annual discharge for the near, mid, and far futures are −1.8%, −6.1%, and −6.0%, respectively. Under RCP8.5, there will be a corresponding decline of −3.1%, −3.9%, and 0.25% for the near, mid, and far futures, respectively (Karam et al., 2023).

The research evaluated in this paper indicates that the hydrological response varies with percentage change as well as trend change (increasing or decreasing) on the impact of climate change on future hydro-climatological reaction, which includes temperature, streamflow, and rainfall. The temperature will rise in almost every basin, with a few notable exceptions and seasonal variations.

Ethiopia has been considered as an example in this review paper to highlight how the climate may affect future hydrological responses. Research has been conducted across the nation’s river basins to analyze the effects of climate change on the hydrological response in the future (Figure 1a–c). The works of literature considered in this review were; Dibaba et al. (2020) on the Fincha Abay basin, Merga et al. (2022) on the Dhidessa/Upper Blue Nile basin, Gebre and Ludwig (2017) on the Gilgel Abay/Upper Blue Nile, Eshete et al. (2022) on the Megech/Upper Blue Nile, Gizaw et al. (2021) on the Upper Awash river basin, Genale Dawa, Baro river basin, Bekele et al. (2019) on the Keleta/Awash River basin, Gebrechorkos et al. (2020) on the Awash (61,032 km²), Fita and Abate (2022) on the Melka Wakana, Fentaw et al. (2018) on the Tekeze river basin, Orkodjo et al. (2022) on the Omo river basin, Mekonnen et al. (2023) and Akirso et al. (2022) on the Gidabo river/Rift valley lake basin, Edamo et al. (2022) and Kuma et al. (2021) on the Bilate River/Rift valley lake basin, Abdule et al. (2023) on the Yado/Genale Dawa river basin.

Figure 1. (a) Annual maximum temperature ($\mathrm{°C}$); (b) Annual rainfall; (c) Annual streamflow change in Ethiopian river basins respectively.

Based on the review articles taken into consideration for this review, the temperature will rise in all scenarios in both the mid and long-term future with a maximum change of up to 8.2$\mathrm{°C}.$ The yearly rainfall and outflow, however, indicate an inconsistent pattern moving forward. Rainfall and streamflow will increase in certain areas of the watershed, such as the Awash River basin, while they will drop in other areas, such as the Gidabo and Bilate rivers of the Rift Valley Lake Basin.

Furthermore, different researchers have come to differing results about how hydrology will be impacted by climate change; using different hydrological models could be one reason for this. For example, Mengistu et al. and Gizaw et al. (2021) have used SWAT and HSPF models, respectively, and assessed the influence of climate on the hydrological response of the Baro River basin. According to Mengistu et al. (2023), there will be a decline in the annual streamflow in the far and near future. Conversely, Gizaw et al. (2021) predicted an increase in streamflow in the near and far future. Akirso et al. (2022) and Mekonnen et al. (2023) also used HEC-HMS and SWAT to evaluate the impact of climate change on the hydrology of the Gidabo watershed (Rift Valley Lakes Basin), respectively. According to them, there will likely be a future decrease in annual stream flow and precipitation; however, under RCP 8.5 and in the distant future, the magnitude of change remains extremely high (−5.2% and −46.84%, respectively). The results of Kuma et al. (2021) and Edamo et al. (2022) studied in the Billate/Rift Valley using SWAT and HEC-HMS in RCP8.5 and RCP4.5 showed that future annual stream flow and rainfall would decrease; on the other hand, future annual stream flow and rainfall would increase and decrease, respectively (Edamo et al., 2022).

In summary, the anticipated hydrological reaction to climate change varies from one watershed to another, among models, and simultaneously among authors. It is difficult to get reliable interpretations and information because of the variance in predictions between models and researchers. According to the Ethiopian basin’s climatic impact on future hydrology findings, the upper Blue Nile River basin’s Dhidessa watershed will see nearly zero temperature change, while the Wabishebele River basin will have an 8.2 °C temperature shift. Gidabo may have a loss of up to 86% in yearly precipitation, whereas, Upper Awash may see an increase of up to 56% in precipitation. As a result, the Gidabo watershed’s future hydrology may be seriously threatened. The Awash River basin may see a 191% rise in streamflow and will cause flooding problems, while, the Gidabo River may result in a 43% decrease and will cause a water shortage.

The variation in the estimation could be caused by various factors, including different hydrological model types, different amounts of input data, different global climate models, and different calibration approaches used by different authors. Therefore, before any planning, policy-making, or development of water resources for any given watershed, a climate impact assessment for future hydrology should be conducted. Furthermore, a single hydrological model’s output cannot produce a trustworthy result for the comprehensive planning of water resources; rather, the output’s accuracy can be raised by utilizing additional, more sophisticated techniques and methods.

3. Performance evaluation of the hydrological model

Assessing the accuracy of generated data to the measured data and constituent values requires more than just simple professional guidance is unacceptable. Scientifically, the accuracy of the streamflow simulation generated by the hydrological model is evaluated using a goodness-of-fit statistical index (Lamontagne et al., 2020). For this reason, many goodness-of-fit criteria have been developed (Althoff & Neiva, 2021). It is possible to assess the correctness of a hydrological model using a variety of measures, which will also boost the predictability of the results (Jackson et al., 2019). The assessment of model performance is critical for calibration, and evaluation of hydrological models. One goodness-of-fit criterion, which is either maximized or minimized by modifying the model’s parameters, is frequently used as the basis for calibrating hydrological models. A minimum of one dimensionless criterion (maximization) and one error measure (minimization) is required among the numerous studies for assessing the performance of a hydrological model (Gelete et al., 2023).

Three categories are commonly used to evaluate streamflow: hydrological signatures, performance evaluation index, and error measure. While the ideal value of a performance index (dimensionless criterion) is one and should be maximized as an objective function, the ideal value of an error measure is zero and should be minimized (Althoff & Neiva, 2021). There are numerous performance evaluation statistics for the hydrological simulation; nevertheless, different researchers employed different evaluations. It’s unclear why they chose one out of the many and why others were overlooked.

Januário et al. (2022) employed NSE and KGE to assess the simulation’s correctness. KGE was used for the calibration and validation. KGE is preferred over NSE because it overcomes several of NSE's stated faults. Nevertheless, error metrics like PBIAS and RSR were overlooked. Akiso et al. used NSE, R², and RVE for climate impact assessment on streamflow. However, Mekonnen et al. (2023) have used NSE, PBIAS, and R² for the same watershed as Akirso et al. (2022). Evaluating the accuracy of the model with multi-evaluation statistics is better than using minimum-number evaluation statistics. Ang and Oeurng (2019) evaluated the simulation with a performance index (NSE) and an error index (PBIAS and RSR). The simulation’s PBIAS evaluation (1.14) indicates very good performance; on the other hand, the RSR evaluation (0.63) shows satisfactory performance. This result suggested that assessing the model’s accuracy solely using PBIAS or RSR might not offer sufficient and reliable conclusions. KGE is an enhanced NSE evaluation index as well, but the majority of the research has still optimized the parameter values during calibration and validation using the traditional NSE.

Oversensitivity to high flows, overestimation of low flows, and underestimation of high flows are some of the drawbacks of the traditional performance evaluation method, which primarily uses performance indexes like R², NSE, and error measurements like PBIAS, RMSE, and RSR (Althoff & Neiva, 2021; Boisvert et al., 2016). In recent decades, various hydrological simulations have employed hydrologic signature evaluation to address the shortcomings and enhance classical performance evaluation (Topalović et al., 2020; Chilkoti et al., 2018). However, the current studies still use conventional evaluations (NSE, R², PBIAS, and RSR) without considering hydrologic signatures, including Sheng et al. (2023) and Abdule et al. (2023).

Hydrological signatures (indices characterizing hydrologic behavior) are now commonly used to understand space-time variability in hydrological processes and to diagnose weaknesses in hydrological models (Addor et al., 2018). A comparison of signatures obtained from simulated and observed flows can indicate poorly simulated runoff components (peak flow, mean flow, median flow, low flow) that cannot be detected by statistical performance measures alone (Topalović et al., 2020).

Furthermore, the uncertainty analysis method or optimization algorithms employed for the model’s calibration and validation have an impact on its performance. Because of nature’s randomness and our limited knowledge of the hydrological system, uncertainty is an inherent part of all hydrologic models. The use of hydrological modeling uncertainty in climate impact assessments has received far less attention (Bastola et al., 2011). Adapted uncertainty evaluation techniques include Bayesian statics (BS), particle swarm optimization (PSO), parameter solution (PS), sequential uncertainty fitting version-2 (SUFI-2), Monte Carlo (MC), Markov chain Monte Carlo (MCMC), Generalised likelihood uncertainty estimation (GLUE), and Bayesian statics (BS) (L. Zhang et al., 2019).

In the Xinxue and Xiaoqing rivers, L. Zhang et al. (2019) compared the SWAT performance using the SUFI-2 and GLUE uncertainty evaluation methodologies. The Xinxue River basin showed a higher applicability for the SUFI-2 method because of its more logical and accurate analysis results, whereas the Xiaoqing River basin was proven to have a higher application for the GLUE approach. Tang et al. (2021) used SUFI-2, GLUE, PS, and PSO to analyze the uncertainty of SWAT modeling for the Lancang River Basin. NSE and R² of SUFI-2 (0.89, 0.89), GLUE (0.87, 0.89), PS (0.90, 0.92), and PSO (0.86, 0.84) respectively for the whole time simulation. The findings demonstrated that PS outperformed the others and also SUFI-2 remains well performing. As a result, the two authors’ findings demonstrated that a single uncertainty analysis strategy may also reduce the model’s performance; as a result, it is preferable to compare several approaches and select the best one.

Performance evaluation statistics that are employed as objective functions in calibration and validation, including NSE, R², PBIAS, and RSR, also have an impact on overall performance (Table 1). The objective function used in peak flow and low flow simulation also varied (Fowler et al., 2018). According to Fowler et al. (2018), the low flow is better fitted by using the split KGE and index of agreement than using NSE and KGE. Using various objective functions also results in various values for the optimal parameter values, which influences the estimation of water resources (Kouchi et al., 2017). Based on Paul and Negahban-Azar (2018) the performance index for R², NSE, PBIAS, and KGE was 0.91, 0.82, −53.36, and 0.56 for R², 0.91, 0.84, −40.45, and 0.62 for NSE, 0.80, 0.77, −14.93, and 0.70 for PBIAS, and 0.89, 0.85, −32.93, and 0.74 for KGE was used as objective function. KGE objective function has performed best out of the four objective functions. For improved performance, NSE, R², and PBIAS were ranked second, third, and fourth respectively. According to Lee and Choi (2021) for calibration and validation, the RMSE of the extremely high flow and extremely low flow estimated using NSE, KGE, and adjusted KGE as objective function showed that, respectively, they were (0.86, 0.79), (0.74,0.74), and (0.74,0.72) and (0.04, 0.04), (0.03, 0.01), and (0.06, 0.04). For high flow adjusted KGE is better performance whereas for low flow KGE is better performance than the other objective function. Thus, various objective functions and optimization procedures must also be taken into consideration to increase the models’ simulation accuracy.

Table 1. Performance evaluation types and description.

Performance evaluation	Description	References
Coefficient of determination (R^2)	Measures the degree to which the 45-degree line drawing is spread from the observed and simulated runoff values	(Fanta & Sime, 2022)
	The degree to which the actual and simulated data are related or fit together	(Ismail Dhaqane et al., 2023)
	Overly susceptible to elevated flow levels	(Althoff & Neiva, 2021)
	Ranges from 0 to 1; 1 represents a properly suited simulation, and 0 represents poor simulations.	(Ismail Dhaqane et al., 2023)
	$\frac{{({\sum }_{i=1}^n{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})(Q_{\mathit{\text{sim}}}-{\overline{Q}}_{\mathit{\text{sim}}}))}^2}{{\sum }_{i=1}^n{{(Q}_{\mathit{\text{ob}}}-{Q^{-}}_{\mathit{\text{ob}}})}^2{{\sum }_{i=1}^n{(Q}_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2};$ Q is discharge, sim is simulation, ob is observation, $\overline{Q}$ is mean discharge	(Hu et al., 2018)
	>0.85 (very good); (0.75,0.85) (good); (0.60,0.75) (satisfactory); $\le$0.60 (not satisfactory)	(Ferreira et al., 2021)
Nash-Sutcliffe efficiency (NSE)	The most widely applied standard	(Aliye et al., 2020)
	Calculate the degree to which the simulated and observed values are associated.	(Dehua et al., 2017)
	Underestimating high flows and overestimating low flows	(Boisvert et al., 2016)
	Overly sensitive to large flows	(Althoff & Neiva, 2021)
	Ranges from $-\infty$ to 1; $-\infty$ (bad simulations), 1 (perfectly fitted)	(Boisvert et al., 2016)
	1−$\frac{{\sum }_{i=1}^N({Q_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2}{{\sum }_{i=1}^N{{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}^2}$	(Hu et al., 2018)
	>0.75 (very good); (0.60,0.75) (good); (0.50,0.60) (satisfactory); $\le$0.50 (not satisfactory)	(Edamo et al., 2022)
Kling-Gupta Efficiency (KGE)	An alternate NSE metric	(Januário et al., 2022)
	Improve a number of the NSE's alleged flaws
	examines several forms of model errors, including variability, dynamics, and mean error
	Ranges from $-\infty$ to 1; $-\infty$ (bad simulations), 1 (perfectly fitted)
	1−$\sqrt{({r-1)}^2+({\frac{{\sigma }_{\mathit{\text{sim}}}}{{\sigma }_{\mathit{\text{ob}}}}-1)}^2+({\frac{{\mu }_{\mathit{\text{sim}}}}{{\mu }_{\mathit{\text{ob}}}}-1)}^2}$ r = $\frac{\frac{1}{n}{\sum }_{i=1}^n{(Q}_{\mathit{\text{sim}}}-{\overline{Q}}_{\mathit{\text{sim}}}){(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}{\sqrt{\frac{1}{n}{\sum }_{i=1}^n{{(Q}_{\mathit{\text{sim}}}-{\overline{Q}}_{\mathit{\text{sim}}})}^2\frac{1}{n}{\sum }_{i=1}^n{{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}^2}}$	(Januário et al., 2022)
	r is a correlation between observed and simulated, $\sigma$ is the standard deviation, $\mu$ is the mean >0.75 (good); (0.50,0.75] (intermediate); (0.0,0.50) (poor); $\le$0.0 (very poor)
Modified NSE	Suggested for minimal flow	(Althoff & Neiva, 2021)
	$\frac{{\sum }_{i=1}^N{(Q}_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}{{\sum }_{i=1}^N{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}$
Log NSE	Utilized primarily the emphasis is on modeling modest flows.	(Lamontagne et al., 2020)
Index of Agreement (DI)	Indicates the level of error in the model’s forecast.
	It is designed to rectify the R^2's deficiencies.	(Althoff & Neiva, 2021)
	Recognized for its insensitivity to systematic under- and overestimations of the model	(Mubialiwo et al., 2021)
	1 − $\frac{{\sum }_{i=1}^N({Q_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2}{{\sum }_{i=1}^N{(|Q_{\mathit{\text{sim}}}-{\overline{Q}}_{\mathit{\text{ob}}}|+|Q_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}}|)}^2}$	(Ferreira et al., 2021)
	>0.9 (very good); (0.85,0.90) (good); (0.75,0.85) (satisfactory); <0.75 (not satisfactory)
Correlation coefficient (CC)	Quantifies the degree to which the measured and anticipated or estimated datasets have a linear connection $\frac{\sum {((Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})(Q_{\mathit{\text{sim}}}-{\overline{Q}}_{\mathit{\text{sim}}}))}{\sqrt{{\sum }_{i=1}^n{{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}^2{{\sum }_{i=1}^n{(Q}_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2}}$	(Al-Mukhtar & Al-Yaseen, 2019)
Error measure
Percent of Bias (PBIAS)	To gain more insight into the model’s propensity to overestimate or underestimate streamflow	(Bouadila et al., 2023)
	When this criterion’s value is positive, it undervalues flows; conversely, negative values suggest that flows are overestimated
	Varies from −∞ to +∞; 0 represents a perfectly fitted model, and ±∞ represents a poorly calibrated model $\frac{{\sum }_{i=1}^n{Q}_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}}}{{\sum }_{i=1}^n{Q}_{\mathit{\text{ob}}}}$*100
	$(\pm 10,\pm 10)$ (very good); $(\pm 10,\pm 15$) (good);$(\pm 15,\pm 10$) (satisfactory); $\ge \pm 25$ (not satisfactory)	(Ferreira et al., 2021)
Root mean square-standard deviation ratio (RSR)	utilized to calculate the discrepancies between observed data and values projected by a model	(Bouadila et al., 2023)
	Varies from 0 (perfectly fitted) to ∞ (poor simulation).
	$\sqrt{\frac{{\sum }_{i=1}^N({Q_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2}{{\sum }_{i=1}^N{{(Q}_{\mathit{\text{ob}}}-{\overline{Q}}_{\mathit{\text{ob}}})}^2}}$
	Very good (<0.5); good (0.50,0.60); satisfactory (0.60,0.70); not satisfactory (≥0.70)	(Ang & Oeurng, 2019)
Root mean square error(RMSE)	The idea behind RSR is the same
	Varies from 0 (perfectly fitted) to ∞ (poor simulation). $\sqrt{\frac{{\sum }_{i=1}^N({Q_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}})}^2}{n}}$	(Al-Mukhtar & Al-Yaseen, 2019)
Mean bias error (MBE)	Describe if a model has a negative or positive value (i.e. underestimates or overestimates the observation)	(Fouchal & Souag, 2019)
	$\sum \frac{Q_{\mathit{\text{ob}}}-Q_{\mathit{\text{sim}}}}{n}$
	Varies between - $\infty ;$ +$\infty$; 0 (perfect fitted), $\infty$ (bad simulation)
Relative volumetric error(RVE)	The strength of the disparity between the discharge that was seen and that was simulated	(C. Gao et al., 2016)
	$\frac{{\sum }_{i=1}^N{(Q}_{\mathit{\text{sim}}}-Q_{\mathit{\text{ob}})}}{Q_{\mathit{\text{ob}}}}$ *100; 0 (perfect simulation), $\infty$ (badly simulated)	(G. Wang et al., 2017)
P-factor	Computes the data seen using a 95% prediction uncertainty Interval varies from 0 to 1, where 1 indicates that the simulations have fully enclosed the observed data. Performance less than 0.7 is weak; performance greater than 0.7 is well simulated. $\frac{N_{\mathit{\text{in}}}}{N}$*100; Where Nin is number of data enclosed in 95ppu, N is total number of data	(Chilkoti et al., 2018) (Tang et al., 2021)
R-factor	Relative width of a band with a 95% probability Greater than 1.5 poorly fitting, less than 1.5 properly fitted $\frac{\overline{W}}{\sigma n};$ Where $\overline{W}$ is the average width of 95%ppu, $\sigma n$ is the standard deviation of the observed	(Chilkoti et al., 2018) (Tang et al., 2021)
	Hydrologic signature
Flow duration curves	T is the percent of flow exceedance (5, 20, 50, 80, 95); QT is the T^th percentile. Range ($-\infty$ +$\infty$) and $(-0.25,0.25)$ is well fitted $\frac{Q_{T,S}-Q_{T,O}}{Q_{T,O}};$ s is a simulation, and o are observations; T stands for percent exceedance.	(Topalović et al., 2020)

4. Types of hydrological models

Hydrological systems are complicated and full contentment in simulation is challenging, researchers are searching for more advanced ways to anticipate (Haznedar & Kilinc, 2022) better approach. The simulation of catchments’ hydrological processes by hydrological models is considered a significant concern nowadays. Consequently, a large number of hydrological models have been developed (Tran et al., 2018). Many investigations are being conducted on questions such as which model will yield better results that are consistent with observed discharges. They also conclude that most investigations probably just need a simple to moderately complicated model. It was said that even highly advanced modeling does not yield better results (Devi et al., 2015).

Many hydrological models from various countries at different eras have been developed to model the hydrology of the watersheds. Accurately determining the most practical and cost-effective hydrological models for a given watershed, however, becomes difficult (Doliso & Hatiye, 2021). The models’ results vary greatly depending on the site, and no single model is better under all hydrologic conditions (Aliye et al., 2020; Banda et al., 2022). The following criteria can be used to classify hydrological models; spatial (lumped, distributed) (Elsanabary & Gan, 2015); temporal (static and dynamic) (Tanksali & Soraganvi, 2021; Nesru, 2023); the dynamic portion is further divided into event-based and continuous (Ayalew et al., 2022; De Luca et al., 2022); an aspect of randomness (deterministic and stochastic) (Fanta & Sime, 2022; Gebremichael et al., 2023); process-based (physical, conceptual, and empirical) (Devi et al., 2015; Jehanzaib et al., 2022); distributed model (partially distributed, fully distributed) (Onyutha et al., 2021; Usman et al., 2021). From all, parts of 54 physically-based, distributed, conceptual, and lumped models reviewed in this paper (Table 2), 28 hydrological models were also utilized in Ethiopia mostly in the upper Blue Nile river basin (Table 3).

Table 2. Types of some physically distributed, conceptual, and lumped hydrological models.

Models full name	Abbreviation	Lumped	Distributed	Physically-based	Conceptual	References	Number of parameters	References
Agricultural Policy Extender	APEX		√	√		(Golmohammadi et al., 2014; White et al., 2017)	32	Wang & Jeong, 2019
A Telier Hydrologique Spatialise	ATHYS		√	√		(Khattati et al., 2016)	8	(Khattati et al., 2016)
Australian Water Balance Mode	AWBM	√			√	(Mubialiwo et al., 2021)	8	(Kunnath-Poovakka & Eldho, 2019)
–	CEQUEAU		√	√		(Mai et al., 2020)	29	(Boisvert et al., 2016; Fniguire et al., 2022)
Coupled Routing and Excess Storage	CREST		√	√		(J. U. N. Ma et al., 2018)	12	(J. U. N. Ma et al., 2018)
Distributed Hydrology Soil Vegetation Model	DHSVM		√	√		(Hasan et al., 2023)	30	(Huang et al., 2019)
Data-Based Mechanistic	DBM	√			√	(S. K. Singh & Marcy, 2017)	6	(S. K. Singh & Marcy, 2017)
Dynamic Watershed Simulation Mode	DWSM		√	√		(X. Zhang et al., 2013)	30	(X. Zhang et al., 2013)
Genie Rural a 4 Parameters	GR4J	√			√	(Kunnath-Poovakka & Eldho, 2019)	4	(Kunnath-Poovakka & Eldho, 2019)
–	GRSD		√		√	(Thirel et al., 2015)	8	(Lavenne et al., 2016)
Gridded Surface Subsurface Hydrologic Analysis	GSSHA		√	√		(Al-Areeq et al., 2021)	16	(Al-Areeq et al., 2021)
Hydrologiska Byrans Vattenbalansavdelning	HBV		√		√	(Uysal et al., 2023)	21	(Uysal et al., 2023)
Hydrological Engineering Centre-Hydrologic Modeling System	HEC-HMS		√	√	√	(Sempewo et al., 2022)	20	(Sempewo et al., 2022)
Hydro Informatic Modeling System	HIMS		√		√	(Jiang et al., 2015)	9	(Jiang et al., 2015)
Hydrological Model of É cole de technologie superieure	HMETS	√			√	(Shen et al., 2018; Arsenault et al., 2019)	21	(Shen et al., 2018; Arsenault et al., 2019)
Hydrological Model Focusing on SUB-flows’ Variation	HMSV	√			√	(Lafdani et al., 2013; Onyutha, 2019)	10	(Lafdani et al., 2013; Onyutha, 2019)
	HSAMI	√			√	(Arsenault et al., 2015; Martel et al., 2017)	23	(Martel et al., 2017)
Hydrological Simulation Program—FORTRAN	HSPF		√		√	(Xie et al., 2019; Sahraei et al., 2020)	19	(Sahraei et al., 2020)
	HYDROTEL		√	√		(Lucas‐Picher et al., 2020)	7	(Lucas‐Picher et al., 2020)
HYdrological MODel	Hymod	√			√	(S. K. Singh & Marcy, 2017; Ávila et al., 2022)	5	(Ávila et al., 2022)
HYdrological Prediction of the Environment	HYPE		√	√	√	(Santos et al., 2022)	22	(Santos et al., 2022)
Identification of unit Hydrographs and Component flows from Rainfall, Evaporation, and Stream	IHACRES	√			√	(Mubialiwo et al., 2021)	11	(Mubialiwo et al., 2021)
Kinematic Runoff and Erosion Model	KINEROS2		√	√		(Saran et al., 2021)	12	(Saran et al., 2021)
Large Area Runoff Simulation Model	LARSIM		√	√	√	(Ludwig et al., 2023)	34	(Beg et al., 2019)
Liuxihe Model	LXH		√	√		(Osman et al., 2021; Xu et al., 2021)	12	(Xu et al., 2021)
McMaster University Hydrologiska Byråns Vattenbalansavdelning	MACHBV	√			√	(Buddika & Coulibaly, 2020)	15	(Buddika & Coulibaly, 2020)
MIKE- Nedbor Afstromnings Model	MIKE11-NAM	√			√	(Nannawo et al., 2022)	9	(Nannawo et al., 2022)
MIKE-Système Hydrologique Européen	MIKE-SHE		√	√		(L. Ma et al., 2022)	23	(Kaleris & Langousis, 2017)
Modified Interaction Soil Biosphere Atmosphere	MISBA		√	√		(Elsanabary & Gan, 2015)	26	(Elsanabary & Gan, 2015; Islam & Gan, 2015)
Modello Idrologico Semi-Distribuito in continuo	MISDc	√			√	(Mohammadi et al., 2022)	11	(Mohammadi et al., 2022)
Modèle Hydrologique Simplifiéàl’ Extrême	MOHYSE	√			√	(Arsenault et al., 2015; Arsenault et al., 2019)	10	(Arsenault et al., 2015; Arsenault et al., 2019)
MOdèle à Réservoirs de Détermination Objective du Ruisselleme- totally specialized	MORDOR-TS		√		√	(Laura et al., 2017)	22	(Laura et al., 2017)
Parameter-Efficient Semi-Distributed Watershed Model	PED-WM		√		√	(Moges et al., 2017)	9	(Moges et al., 2017)
SWAT Integrated with QGIS	QSWAT		√	√		(Munoth & Goyal, 2020)	28	(Aqnouy et al., 2023)
Research Center for Climate Change Water Balance Model	RCCC-WBM		√		√	(Ning et al., 2022)	4	(Ning et al., 2022)
Sacramento Soil Moisture Accounting	SAC-SMA	√			√	(Buddika & Coulibaly, 2020)	19	(Kunnath-Poovakka & Eldho, 2019; Buddika & Coulibaly, 2020)
Simplified Hydrolog	SIMHYD	√			√	(Yu & Zhu, 2015)	9	(Muhury et al., 2023)
Soil and Water Integrated Model	SWIM		√	√		(Cui et al., 2018)	32	(Krysanova & Wechsung 2000)
Modified Soil Moisture Accounting and Routing	SMARG	√			√	(Darbandsari & Coulibaly, 2020)	10	(Darbandsari & Coulibaly, 2020)
Soil Water Assessment Tool	SWAT		√	√	√	(Noori & Kalin, 2016; Thapa et al., 2017; K. Singh et al., 2022)	28	(Aqnouy et al., 2023)
Soil Water Balance Model	SWBM		√	√		(Sirigu & Montaldo, 2022)	27	(Sirigu & Montaldo, 2022)
Storm Water Management Model	SWMM		√	√		Nayeb et al., 2019)	10	(Nayeb et al., 2019)
	TANK	√			√	(Okiria et al., 2022; N. M. Reddy et al., 2023)	16	(Okiria et al., 2022)
	TETIS		√		√	(Barrientos et al., 2020)	22	(Barrientos et al., 2020)
Topographic, Kinematic, Approximation, and Integration	TOPKAPI		√	√		(Atif et al., 2019)	14	(Atif et al., 2019)
TOPography-Based Hydrological Model	TOPMODEL		√		√	(J. Gao et al., 2017; Januário et al., 2022)	10	(Januário et al., 2022)
	TopNet		√	√		(Bastola et al., 2011; S. K. Singh & Marcy, 2017)	15	(S. K. Singh & Marcy, 2017)
Variable Infiltration Capacity	VIC		√	√		(Naha et al., 2021)	16	(Naha et al., 2021)
Water Balance Simulation Mode	WaSiM-ETH		√	√		(S. K. Singh & Marcy, 2017)	9	(S. K. Singh & Marcy, 2017)
Water And Snow balance Modeling System	WASMOD	√			√	(Nuri & Abdüsselam, 2020)	6	(Nuri & Abdüsselam, 2020)
	WATFLOOD		√	√		(Shrestha et al., 2022)	19	(Bingeman et al., 2006)
Water Evaluation and Planning	WEAP		√		√	(Ibrahim et al., 2022)	9	(Ismail Dhaqane et al., 2023; Ruiz-García et al., 2023)
Water and Energy Transfer between Soil, Plant and Atmosphere	WetSpa		√	√		(Mohammady et al., 2018; Kavian et al., 2020)	11	(Kahsay et al., 2017)
Xinanjiang	XAJ	√			√	(Touseef et al., 2020)	16	(Ismail Dhaqane et al., 2023)

Table 3. Some streamflow simulation hydrological models used in Ethiopia.

Model type	River basin (example)	Area of study	References
APEX	Upper Blue Nile	Water resources assessment on farm vegetable production	(Worqlul et al., 2019)
AWBM	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
CREST	Upper Blue Nile	Streamflow simulation	(Lakew, 2020)
GR4J	Upper Blue Nile	Climate change on extreme flow	(Meresa & Gatachew, 2019)
HBV	Upper Blue Nile, Rift Valley	Stream flow regionalization of the eastern part of the upper Blue Nile	(Gelete et al., 2023 ; S. Mekonnen et al., 2023)
HEC-HMS	Rift valley	Climate change on streamflow	(Akirso et al., 2022)
HMETS	Blue Nile	Climate change on extreme flow	(Meresa & Gatachew, 2019)
HMSV	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
IHACRES	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
SACRAMENTO	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
SIMHYD	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
TANK	Upper Blue Nile	Extreme flow analysis	(Onyutha, 2019)
HSPF	Tekeze, Baro, Awash, Genale-Dawa	Climate impact assessment	(Gizaw et al., 2021)
MIKE-NAM	Genale Dawa	Simulation of rainfall-runoff	(Aredo et al., 2021b)
MIKE-SHE	Genale Dawa	Effect of land use land cover on stream	(Aredo et al., 2021a)
MISBA	Upper Blue Nile	Climate anomalies impact on upper Blue Nile basin	(Elsanabary & Gan, 2015)
PED-W	Upper Blue Nile	Spatio-temporal rainfall-runoff simulation	(Moges et al., 2017)
SMAR	Upper Blue Nile	rainfall-runoff	(Dessalegn et al., 2017)
SWAT	Valley	Climate impact on streamflow	(Ketema & Dwarakish, 2021)
SWMM	Wabe Shebele	Climate and land use effect on town stormwater	(Bibi, 2022)
SWIM	Upper Blue Nile	Climate change impact on streamflow	(Teklesadik et al., 2017)
TOPAKPI	Omo-Gibe	Flood frequency analyses	(Rabba, 2023)
TOPMODEL	Upper Blue Nile	Land use land cover change	(Gumindoga et al., 2014)
VIC	Upper Blue Nile	Climate change on streamflow and evapotranspiration	(Teklesadik et al., 2017)
Wasim-ETH	Rift valley	Land use land cover change	(Sulamo et al., 2021)
WEAP	Rift valley	Simulating sub-basin hydrology	(Abera & Ayanew, 2021)
WetSPA	Tekeze	Climate change impact on streamflow	(Gebremeskel & Kebede, 2018)
XAJ	Upper Blue Nile	Climate change effect on hydrology	(Atanaw et al., 2023)

4.1. Physically-based, distributed, lumped, and conceptual model

Despite the hydrologic model’s classification as distributed and physically based, based on process and space, respectively, it should be noted that the physically based model is also a distributed hydrological model (Table 2). In augmenting Table 2, the Liuxihe model, which Osman et al. (2021) refer to as a physically based, distributed hydrological model (PBDHM), is one instance to show that physically based models can also be distributed models. Additionally, as indicated in Table 2, the model classified as lumped is also classified as a conceptual model. There are two types of models in the distributed model; full-distributed and partially-distributed. The partial-distributed model can be both a conceptual model and a physically based model, like SWAT and HEC-HMS. SWAT is semi-distributed (Singh et al., 2022); physically-based (Thapa et al., 2017); and conceptual (Noori & Kalin, 2016). HEC-HMS is physically based; semi-distributed (Gebre, 2015; and conceptual hydrological model (Gebre, 2015).

Concerning model parameters, they varied from 4 parameters in the GR4J model to high parameters greater than 25, like SWAT and DHSWM. However, a few researchers only put the sensitive parameters when listing the model’s parameters, which prevents this review paper from inserting the precise number of parameters in a few models. As a result, it is advised that one must be aware that the model’s parameters have their significance (Figure 2), which may not be consistent and precise for a few models. The parameter count presentation in this review paper aims to highlight the models with comparatively high and low counts. It is also possible to easily check whether the input data and the parameters are easily accessible from the study area before starting the model’s analysis, to show that future researchers will find it easy to identify which model requires a large number of parameters, as well as to check the list of parameters using the references (Table 2). This review study focuses on compiling a few alternative journals as a backup plan and then selecting the most appropriate one out of all of them.

Figure 2. Some currently used hydrological models and their parameter counts.

For instance, Khattati et al. (2016) utilized eight parameters for their stream flow simulation using the ATHYS hydrological model; Maref and Seddini used six parameters; and Bouadila et al. (2023) used five parameters. This review used the maximum number of parameters (8) (Table 2) from the 8, 6, and 5 parameters of the ATHYS model was utilize.

The physically based distributed hydrological model, including MIKE-SHE, MISBA, SWBM, CEQUEAU, QSWAT, SWAT, DHSVM, APEX, LARSIM, TOPKAPI, and DWSM, has a relatively large number of parameters, which is greater than 23, whereas, the conceptual model, including GR4J, RCC-WBM, DBM, IHACRES, WASMOD, AWBM, GRSD, HYMOD, HIMS, MIKE NAM, PED-WM, SIMHYD, WEAP, HMSV, MOHYSE, and SMARG, has a relatively low number of parameters. From the distributed models (Table 2), CEQUEAU, DHSVM, GSSHA, MIK-SHE, MISBA, TOPKAPI, and WATFLOOD are fully distributed models, whereas, the remaining distributed models are semi-distributed.

The physically based distributed model is generally flexible and can be used for most of the types of hydrological responses. For example, one of the versatile SWAT models is used for long-term continuous simulations of flow (Haleem et al., 2022); soil erosion (Mosbahi & Benabdallah, 2020); sediment simulation (Ricci et al., 2018); movement of nutrients in watersheds (Zeiger & Hubbart, 2016); event-based flood simulation (Jodar-Abellan et al., 2019); streamflow drought forecasts (Sohoulande Djebou, 2019); streamflow regionalization to ungauged sites (Kanishka & Eldho, 2020); urban water stormwater simulation (Glick et al., 2023); and water quality evaluation (Olaoye et al., 2021) with different sizes and hydrologic, geological, and climatic variables; and evaluating the effects of climate change, land use and cover, and watershed management (Nilawar & Waikar, 2018). However, specifically, some models have been developed for particular purposes including the APEX model for small watersheds, fields, and plots at a higher spatial resolution, VIC for macro-scale watershed, and SWMM for urban watershed flow simulation. The lumped conceptual hydrological model including the simplest GR4J model can used for streamflow regionalization (Arsenault et al., 2019); operational flood forecasting (Buddika & Coulibaly, 2020); land use land cover change effect on streamflow (Mouche et al., 2014; Troy et al., 2007; Leila et al., 2024); climate and human impact on streamflow (Y. Liu et al., 2021).

4.2. Machine learning-based streamflow modeling

Streamflow modeling is currently receiving greater attention with machine learning-based hydrological models (Table 4) than from other data-driven hydrological models (Hu et al., 2018). The primary driving force behind the shift from physically and conceptually based hydrological modeling to machine learning modeling is the lack of observational data, and relatively high performance which makes it more difficult for these models to be applied in locations with poor gauges (He et al., 2021) as summerised in Table 5. Multilayer perceptron (MLP), support vector regression (SVR) (Hussain & Khan, 2020); long-short-term memory (LSTM) (Hu et al., 2018); Gaussian linear regression model (GLM), Gaussian generalized additive model (GAM), multivariate adaptive regression splines (MARSs); artificial neural network (ANN), random forest (RF), and 1D convolutional neural network (1DCNN) (D. Singh et al.,2023); recurrent neural network (RNN), and adaptive fuzzy neural inference system (ANFIS) (M. Khan et al., 2023) are examples of machine learning streamflow modeling.

Table 4. Physically based-distributed and lumped-conceptual hydrologic model summary.

Key characteristic	Physically based-distributed hydrological model	Conceptual-lumped model	References
Catchment represent	Heterogonous unit	Single unit	(Wakigari, 2017),
Data requirement	Huge data	Relatively less	(Darbandsari & Coulibaly, 2020; Birocchi et al., 2023)
Parameter value determined	From the physical behavior of the catchment	From calibration	(Wakigari, 2017)
Complexity	High	Relatively Less	(Darbandsari & Coulibaly, 2020; Birocchi et al., 2023),
Cost requirement	High	Less
Climate impact analysis	Applicable	Applicable	(Dehua et al., 2017)
Time requirement	High	Relatively less	(Darbandsari & Coulibaly, 2020)
Physical laws of conservation	Known	Simplified concept	(Elsanabary & Gan, 2015)
Stream flow simulation Without calibration	Can simulate	May not possible	(S. K. Singh & Marcy, 2017)
Applicability for land use land cover impact	Highly Applicable	Applicable	(Al-Areeq et al., 2021); Kunnath-Poovakka & Eldho, 2019; Janabi et al., 2021)
Output reliability at different spatial and temporal resolution	High	Less	(Darbandsari & Coulibaly, 2020; Mohammadi et al., 2022)
Stream flow regionalization	Applicable	Applicable	(Ocio et al., 2020)
Estimate variables at the interior of the catchment	Applicable	Less applicable	(Darbandsari & Coulibaly, 2020)
Spatial-temporal impact analysis	Highly applicable	Less applicable	(Darbandsari & Coulibaly, 2020)
Power of representing real-world	Highly powerful	Less powerful	(Janabi et al., 2021)
For detail assessment	Highly recommended	Less recommended	(S. K. Singh & Marcy, 2017)
Scope of application	High capability	Limited	(Didovets et al., 2017)
Number of parameters versatile to many application	High High	Less Relatively Less	(Kunnath-Poovakka & Eldho, 2019; Aqnouy et al., 2023) (Al-Areeq et al., 2021; Aliye et al., 2020)
Event-based flood modeling	Applicable	Applicable	(Jodar-Abellan et al., 2019; Buddika & Coulibaly, 2020)

Table 5. Summary of the physical and conceptual model with a machine learning model.

Key-characteristic	Knowledge-based hydrological model	Machine learning model	References
Catchment represent	White or Grey box	Black box	(Hu et al., 2018)
Rainfall-runoff process	Practicable	Not practicable	(Hu et al., 2018)
Preferable for data-limited watershed	Less	Better	(Khodakhah et al., 2022)
Data type required	Metrological, hydrological, spatial data, and related data	Numerically recorded metro-hydrological data	(Khodakhah et al., 2022)
Runoff prediction	With calibration of parameter	With self-learning from the data	(Al-Mukhtar & Al-Yaseen, 2019)
Hydrological significance and robustness	High	Relatively Less	(J. Wang et al., 2017)
Error minimization	Some extent less	Better	(Khodakhah et al., 2022)
Capability for land use land cover change impact, climate impact evaluation	Applicable	Applicable	(C. Gao et al., 2016)
Performance	Better	Relatively better	(C. Gao et al., 2016)

To assess the effectiveness of machine learning-based streamflow modeling with the impact of climate change on future hydrology, the following studies have been conducted: Lin et al. (2019) studied the response of surface runoff to climate change in the Upper Minjiang River (22,900 km²), Southwest China; M. Khan et al. (2023) evaluated the effects of climate change on stream flow dynamics in the Hunza watershed (13,567 km²), which is shared by Pakistan, China, and Afghanistan; D. Singh et al. (2023) studied machine learning and deep learning-based streamflow prediction in a hilly catchment for future scenarios using CMIP6 GCM data in Satluj basin (2457 km²).

Furthermore, for continuous long-term streamflow simulation, Tibangayuka et al. (2022) simulated streamflow for a high-humid tropical catchment of the Upper Ruvu (7663 km²) in Tanzania; S. Khan et al. (2023) investigated the inter-comparisons of ANN and SWAT in the Astore watershed, upper Indus basin (45, 400 km²), India; Jimeno-Sáez et al. (2018) compared SWAT and ANN Models for Daily Runoff Simulation in Different Climatic Zones of Ladra river (843 km²), Peninsular Spain; Khazaee Poul et al. (2019) compared MLR, KNN, ANN, and ANFIS Models with Wavelet Transform in Monthly StreamFlow Prediction in the Clair river (579,000 km²), Canada; Haznedar et al. (2023) forecast streamflow using a hybrid LSTM PSO approach in the Seyhan river Basin (22.04 km²), Eastern Mediterranean Region of Turkey.

In summary, the model has been used to streamflow modeling based on machine learning, ranging in size from small watersheds (22.04 km²) to macro watersheds (579,000 km²). As seen in Figure 3, there isn’t a single model that performs flawlessly in every watershed or during calibration and validation. As an example of an ANN, the model fits very well in some watersheds and satisfactorily or unsatisfactorily in others. Thus, it is not advised to choose a single machine learning model at random for streamflow analysis; rather, comparing multiple models and choosing the best one could be a better way to increase performance and ensure that the model’s output is accepted for future applications.

Figure 3. Performance of machine learning models with R² at different watershed.

4.3. Event-based hydrological model

Hydrologists frequently utilize incident-grounded models to simulate hypothetical and actual flood disasters. The timescale required by event hydrologic models is sub-daily (Xie et al., 2019). Event-based models consider a single rainfall event separately, as opposed to continuous models that simulate flow over a longer period (Stephens et al., 2018). Regarding the input rainfall for the event-based model, it is a design storm, but the extended synthetic rainfall scenario pertains to the continuous model (Ayalew et al., 2022). Soil Conservation Service Lag-and-Route (SCS-LR) (Nguyen & Bouvier, 2019), Watershed Boundary Network Model (WBNM) (Stephens et al., 2018), Event-Based Approach for Small and Ungauged Basins (EBA4SUB) (Ayalew et al., 2022), Sacramento Soil Moisture Accounting (SAC-SMA) (Bournas & Baltas, 2021), Hydrological Simulation Program–Fortran (HSPF) (Xie et al., 2019), HEC-HMS (Nourali et al., 2016), SWAT (Yaduvanshi et al., 2018), TOPKAPI-X (Janabi et al., 2021), TOPMODEL (Januário et al., 2022), GSSHA (Al-Areeq et al., 2021) are some of the event based hydrologic model.

4.4. Ensemble of hydrological model

Due to inadequate input data, model structure, and model output uncertainty in large-scale exercises, using a single hydrologic model typically results in simulation uncertainties. When modeling hydrological processes, the ensemble model technique may be a more effective tool for making decisions (Himanshu et al., 2023). Chathuranika et al. (2022) have compared the streamflow simulation using SWAT and HEC-HMS, and the simulation result showed that the HEC-HMS model caught medium flows better than the SWAT model did for high flows. Additionally, the SWAT model accurately predicted both dry and wet seasonal flows within seasonal time bases, with modest under predictions of 2.12% and 13.52% relative to the observed values. In comparison to observed flows, the HEC-HMS model underpredicted the dry and wet seasonal flows by 10.76% and 18.54%, respectively. The results showed that the model ensemble had a great capacity to fill in the gaps created by individual models. Hydrology research using ensemble forecasting is presently a well-established field that covers a wide range of operating scenarios (Troin et al., 2021). By using ensemble approaches to improve the existing model’s capability, predictive performance can be enhanced. Consequently, it is thought that by leveraging the advantages of each model, an ensemble approach that integrates the output of several models could increase the accuracy of the individual models (Gelete et al., 2023).

In the Cauvery catchment of the Indian Peninsula, Horan et al. (2018) compared the performance of a single hydrological model with a multi-model ensemble. The results show that in catchments containing reservoirs, the ensemble mean performs better in terms of prediction than the individual models. Teklesadik et al. (2017) assessed the inter-model comparison of the hydrological impacts of climate change on the Upper Blue Nile basin using global climate models and an ensemble of hydrological models. They concluded that future studies on the effects of climate change ought to take into account the use of an ensemble of hydrological models instead of basing assessments on a single hydrological model, which could produce biased results.

The simulation performance of three physically-based hydrological models (SWAT, VIC, and BTOPMC) was assessed by Z. Li et al. (2018). The individual models SWAT (0.77, 0.66), VIC (0.75, 0.65), and BTOPMC (0.82, 0.75), evaluated with NSE for calibration and validation, respectively, are less fitted than the ensemble of the models (0.95, 0.90). Tian et al. (2017) evaluated the impact of climate change on low flows in the Xiangjiang river basin using the Bayesian averaging method with the GR4J, HBV, and Xinanjiang hydrological model. The Bayesian averaging strategy improved the ensemble’s NSE value, and all simulated streamflows fell inside the 95% confidential interval.

The performance is impacted by the ensemble technique employed when employing the ensemble model. Neural network-based averaging, weighted averaging, and simple averaging are among the various ensemble averaging techniques. The neural network performs the best out of the three ensembles when using the two-ensemble technique. Accurate model simulation results can also be achieved by comparing multiple ensemble techniques and choosing the best ensemble technique (Gelete et al., 2023). The ensemble hydrological model performs better than the individual model simulation, to put it briefly. The ensemble can be carried out using a data-driven model or a distributed hydrological model, or it can be lumped and distributed, lumped and lumped, or distributed. The accuracy of the model simulation output can also be influenced by the ensemble technique used.

The ensemble model is superior to the individual model simulation, as demonstrated by the numerical value comparison of the goodness of fit values conducted by Gelete et al. (2023) and Himanshu et al. (2023) in the Katar watershed, Ethiopia, using neural network ensemble (NNE) and the Moral watershed, India, respectively, as a sample shown in Figure 4. According to the results of both authors, the ensemble model performs well across all NSE, R², and PBIAS metrics.

Figure 4. Comparisons of single model performance with ensemble model.

4.5. Hybridization of hydrological model

It has been established that machine learning (ML) is a promising method for correctly predicting streamflow. Unfortunately, there is a lack of transparency because artificial intelligence hydrological modeling is a black box and its parameters hardly ever explain the physical significance of catchment characteristics (Poonia & Tiwari, 2020). Hydrologists have been using hybrid modeling more and more in recent years (Haznedar & Kilinc, 2022).

When multiple hydrological models are integrated, coupled, embedded, and combined for hydrological component simulations, the term ‘hybrid hydrological model’ is used (Kilinc, 2022; Okkan et al., 2021). Hydrological framework in comparison to competitors, hybridized models can combine the advantages of both the physical hydrological model and the data-driven model, or the strengths of each data-driven model individually, resulting in more accurate streamflow predictions and realistic uncertainty intervals (Ren et al., 2018). The hybrid technique is a multi-model strategy that uses the time series produced by the conceptual or physically based hydrological model as inputs of a machine learning system, known as a nested hybrid, to boost the accuracy of flow predictions (Farfán et al., 2020; Okkan et al., 2021; Haznedar et al., 2023).

Examples include the following: merging HBV with a Bayesian neural network (Ren et al., 2018); hybridizing lumped GR4J with a deep learning machine (Kapoor et al., 2023; Kumar et al., 2023); combining the physically based SWAT model with ANN and the HEC-HMS model with ANN for flood forecasting (Kurian et al., 2020; Gebremichael et al., 2023). The hybrid model offers a significant improvement in predictive performance over the component models. According to all of the study findings of the hybridized physically and conceptually hydrological model with ML reviewed in this paper.

Moreover, the hybrid approach combines machine learning and evolutionary algorithms to predict streamflow. This includes integrating an adaptive neuro-fuzzy inference system (ANFIS) with a genetic algorithm (GA) (ANFI-GA) (Kurian et al., 2020); combining particle swarm optimization (PSO) and long short-term memory networks (LSTM) (Haznedar et al., 2023); combining feedforward neural networks (FFNNs) with genetic algorithm (GA) (Zakhrouf et al., 2020); hybrid ANFIS models with GA, particle swarm optimization (PSO), and ant colony optimization for continuous domain (ACOR) optimization algorithms (Azad et al., 2018).

To bolster the hybrid model’s capabilities, this review paper employed numerical comparisons of individual models with the hybrid. These comparisons were carried out by taking Kapoor et al. (2023) in Australia for 222 unregulated catchments taken from the CAMELS-AUS data set by hybridizing GR4J with LSTM and CNN; Kassem et al. (2020) in the northern Iraqi Khazir basin using Asmawa and Khanis stations using SWAT and ANN; Kilinc and Haznedar (2022) in the Euphrates River at Yazikvy and Kilay gauge stations using LSTM and genetic algorithms (GA). The NSE value in the Kapoor et al. (2023) study represents the 95th percentile of the entire catchment.

To support numerical comparisons of individual models with the hybrid: a study conducted by Kapoor et al. (2023) in Australia for 222 unregulated catchments taken from CAMELS-AUS data set by hybridize of GR4J with LSTM and CNN; Kassem et al. (2020) in Khazir basin which located in the north of Iraq using Asmawa and Khanis stations using SWAT and ANN; Kilinc and Haznedar (2022) in the longest river in Western Asia is the Euphrates at Yazikvy and Kilay gauge station using LSTM and genetic algorisms (GA) have used as shown in Figure 5. The NSE value of Kapoor et al. (2023) study is not the single watershed instead the 95 percentile of the whole catchment. 95 percentile is the statistical measure that indicates the value is equal to or greater than 95 percent of the total population. There were differences in the 5 percentile performance across CNN (−0.001), GR4J (0.17), and LSTM (0.23) concerning NSE values. This finding suggests that not all watersheds have uniformly high performance from machine learning. The hybrid hydrological model thus performs better than the single model.

Figure 5. Comparisons of individual models and their hybridized model.

5. Performance of some individual physically based-distributed hydrological models

Table 6 shows that SWAT can estimate streamflow in all sizes of watersheds, ranging from small scales (Maybar, 1.12 km³) and Megech River (170 km³) in Ethiopia’s upper Blue Nile basin to enormous scales (860,000 km² in China’s upper Yangtze and 3.7 million km² in the Congo River). The SWAT models able to simulate streamflow are very good, satisfactory, and, to a small extent, not satisfactory when measured using R², NSE, and PBIAS at various watersheds, as shown in Figure 6. As a result, the SWAT model cannot function well in every watershed. The performance rating status of the other models, such as HEC-HMS, QSWAT, MIKE-SHE, and VIC, is identical to that of the SWAT model. The proportion of very good and good performance ratings is comparatively higher than the percentage of satisfactory and not satisfactory performance ratings, even though the percentage of performance ratings is not consistent as the number of review papers varied. While MIKE-SHE performs very well in the Shaya watershed in Ethiopia, the performance in the Chauhe watershed at the Dage stream flow station in China is both inadequate and satisfactory. As a result, it is uncertain if MIKE-SHE, SWAT, or any other model will be chosen based on performance. Therefore, there isn’t a single model that works superior to the others every time and everywhere. It could be more advantageous to compare multiple hydrological models rather than applying a single model to each watershed.

Figure 6. Overall Performance rating capability of individual model.

Table 6. The various watershed performance index values of some models.

SWAT
Author	Country	Watershed	Area (km^2)	Warming (year)	No.data (year)	No.year (cal/val)	Ratio (cal/val)	R^2	NSE	PBIAS	Time base
Akhter et al. (2022)	India	Aripal	266.5	–	(2005–2018)	9	1.8	0.90	0.72	–	Monthly
						5		0.76	0.81	–
Nilawar and Waikar (2018)	Maharashtra	Purna	15579	3	(1980–2004)	12	1.2	0.91	0.91	–	Monthly
						10		0.83	0.82	–
Fentaw et al. (2018)	Ethiopia	Tekeze	43000	0	1994–2008	9	1.5	0.71	0.73	0.53	Monthly
						6		0.79	0.80	0.45
Ketema and Dwarakish (2021)	Ethiopia	Tikurwuha	681	3	1980–2002	13	1.8	0.77	0.68	10.43
						7		0.85	0.64	1.79
Eshete et al. (2022)	Ethiopia	Megech	170	3	1987–2016	15	1.1	–	0.74	0.81	Monthly
						13		–	0.68	0.77
Diriba (2021)	Ethiopia	Dabus	13830	2	1992–2011	10	1.1	0.82	0.70	–	Monthly
						9		0.78	0.71	–
Malik et al. (2022)	India	Aru	385	3	2007–2016	4	1.3	0.84	0.80	−3.06	Monthly
						3		0.85	0.72	−17.6
		Sheshnag	224			4	1.3	0.89	0.83	−5.90
						3		0.86	0.76	−3.13
		Akura	907			4	1.3	0.85	0.55	−21.13
						3		0.80	0.56	−11.87
Narsimlu et al. (2015)	India	Kuwani river	6821	–	1990–2005	10	1.7	0.77	0.71	–	Monthly
						6		0.74	0.69	–
Haleem et al. (2022)	India	Upper Indus	45400	2	2000–2013	9	1.8	0.7	0.69	9.7	Monthly
						5		0.66	0.61	21.3
Su et al. (2017)	China	upper Yangtze	860000	–	1971–1990	10	1	–	0.81	−5.10	Monthly
						10		–	0.79	−4.1
Merga et al. (2022)	Ethiopia	Dhidhessa	19629	3	1991–2020	18	2	0.88	0.76	2.41	Monthly
						9		0.80	0.77	9.2
(Aqnouy et al., 2019)	Morocco	Laou	940	–	2000–2011	5	1.2	0.74	0.76	12.24	Daily
						4		0.82	0.84	8.93
(Costa et al., 2015)	Brazil	Itapemirim	2237	–	1991–2000	5	1	0.76	0.75	11	Daily
						5		0.72	0.67	22
(Teshome et al., 2020)	India	Bina	2822	2	1989–1896	3	1	0.66	0.65	–	Daily
						3		0.65	0.58	–
						3	1	0.96	0.94	–	Monthly
						3		0.72	0.72	–
Vilaysane, et. al. (2015)	Laos	Xedone	7224	3	1990–2008	8	1	0.82	0.82		Daily
						8		0.73	0.71
						8		0.93	0.91		Monthly
						8		0.92	0.86
Sinha et al.	India	Payaswani	1295	–	1991–2010	10	1	0.96	0.96	−0.67	Monthly
						10		0.93	0.92	6.37
Karam et al. (2023)	Congo	Congo	3700000	–	1984–2005	19	6	–	0.64	1.2
						3		–	0.88	−3.6
Sitotaw et al. (2020)	Ethiopia	Blue Nile /Kessie	65784	3	2001–2014	9	1.8	0.81	0.68	−10.8	Monthly
						5		0.89	0.88	8.3
		Blue Nile @boarder	174166	3		9		0.85	0.83	−4.71
						5		0.93	0.89	9.70
Sime et al. (2020)	Ethiopia	Katar/Rift Valley	3360	2	1986–2004	10	1.4	0.82	0.70	–	Monthly
						7		0.78	0.71	–
Bizuneh et al. (2021)	Ethiopia	Maybar/upper Blue Nile	1.12	2	1990–2014	17 6	2.8	0.69 0.57	0.61 0.51	−48.00 –6.10	Daily
Ndhlovu and Woyessa (2021)	South Africa	Kabompo River /Zambezi	72,087	4	1978–2005	16 8	2.0	0.730.70	0.73 0.64	−2.2 9.4	Monthly
HEC-HMS
Author	Country	Watershed	Area (km^2)	Warming (year)	No.data (year)	No. year (cal/val)	Ratio (cal/val)	R^2	NSE	PBIAS	Time base
Akirso et al. (2022)	Ethiopia	Gidabo/Rift Valley	3390	–	1990–2006	11 6	1.8	0.78 0.78	0.53 0.53		Monthly
Lone (2022)	India	Sindih	1560	–	1992–2017	12 9	1.4	0.88 0.86	0.86 0.79	2.71 2.76	monthly
Ranjan and Singh (2022)	India	Punpun	7055	–	2005–2017	8 5	1.6	0.85 0.82	0.89 0.86	8.11 9.42	Monthly
								0.88 0.78	0.69 0.68	7.23 6.86	Daily
Aliye et al. (2020)	Ethiopia	Katar	3354	–	1990–2007	11 7	1.5	0.88 0.87	0.75 0.73		Monthly
QSWAT
Author	Country	Watershed	Area (km^2)	Warming (year)	No. data (year)	No. year (cal/vval)	Ratio (cal/val)	R^2	NSE	PBIAS	Time base
Tanksali and Soraganvi (2021)	India	Ghataprabha	637.11	–	(1982–2013)	27 5	5.4	0.96 0.83	0.89 0.79	–	Monthly
Munoth and Goyal (2020)	India	Upper Tapi	10600	3	(1988–2013)	15 8	1.9	0.75 0.90	0.75 0.89	1.1 –7.7	Monthly
N. N. Reddy et al. (2018)	Lowa	Nishnabotna	4164	4	(1980–1999)	8 8	1.0	0.81 0.61	0.85 0.68	– –	Monthly
462	Ethiopia	Upper Awash	11,420		1986–2005	18 10	1.8	0.86 0.72	0.64 0.60
MIKE-SHE
Author name	Country	Watershed	Area (km^2)	R^2	NSE	PBIAS	Stream (year)	NO. year (cal/val)	Ratio (cal/val)	Time base
Jing (2021)	China	Jialing Beibie	160000	0.85 0.84	0.66 0.69	14.7 1.8	2009–2020	4 3	1.33	Monthly
(Aredo et al., 2021a)	Ethiopia	Shaya	457.6	0.89 0.88	0.8 0.78	2.48 3.90	2009–2020	6 4	1.5	Daily
C. Li et al. (2022)	China	Anning	11150	0.85 0.87	0.72 0.78	– –	1961–2010	7 3	2.3	Daily
				0.84	0.70	–		2	3.5
S. Wang et al. (2012)	China	Chaohe/Xiahua Chaohe/Dage	4854	0.72 0.72	0.69 0.72		1991–1999	5 4	1.25	Monthly
				0.64 0.59	0.59 0.53			5 4	1.25

Note: In this table No. year (cal/val), is the number of data in years used for calibration and validation respectively. The first and second raw of R², NSE, and PBIAS values are the calibration and validation values respectively.

Using the SWAT model as an example, the review’s findings also attempted to show whether or not the watershed area, the number of streamflow data, and the ratio of calibration to validation years have an impact on the goodness of fit value. As seen in Figures 7–9, respectively, the results showed that there isn’t a clear relationship. The size of watershed, numbers of data used, and the numbers of data used in calibration and validation have not indicated an influence on the performance of the SWAT model. Nonetheless, the choice of time step as daily or monthly directly affects the R² and NSE values (Figure 11). The monthly time step exhibits higher NSE and R² values when compared with daily time bases. However, it maynot be always true. Generally, less than 25 years of data are used, the lowest and maximum numbers of recorded streamflow data points are 5 and 30, respectively. This implies a scarcity of long-term historical data for most rivers.

Figure 7. Area of the watershed versus NSE and R² of SWAT model.

Figure 8. The number streamflow data (year).

Figure 9. The ratio of calibration numbers of data to validation data (year).

Not only does the model’s performance differ from one watershed to the other, but it also varies amongst authors who utilized the same model in the same watershed. For instance, SWAT has been utilized for streamflow modeling for various objectives in the Katar watershed, Rift Valley Lake Basin, Ethiopia by Aliye et al. (2020), Sime et al. (2020), Gelete et al. (2023), Balcha et al. (2023). Gelete et al. (2023) simulated very good or comparatively better suited in daily time bases compared from the results of the four author’s performance index. The three authors did a good simulation performance. Based on the findings of the four authors’ performance index, Gelete et al. (2023) simulated with very good or somewhat better matched on a daily time basis (Figure 10).

Figure 10. The SWAT model performance with authors in the case of Katar watershed, Ethiopia.

Neither Sime et al. (2020) nor Aliye et al. (2020) used error measure performance evaluation. Rather, Gelete et al. (2023) and Balcha et al. (2023) employed PBIAS as an error measuring index. The PBIAS index error measure demonstrated satisfactory performance on both a daily and monthly basis, with approximately equal magnitudes. About eighteen years’ worth of streamflow data were utilized by all writers in a comparable quantity. A primary distinction is that Gelete et al. (2023) employed the validation data (1st January 2000 to 31st December 2005) from portions of older observed data and the calibration (1st January 2006 to 31st December 2017) from the relatively current data. While most researchers conclude that the monthly output simulation of streamflow performs better than the daily time bases, a study carried out in Katar revealed that the daily time bases had a comparatively higher performance index (Figure 10). This summary demonstrated that streamflow simulation is a challenging task that necessitates careful examination to produce consistent results that do not vary with time and watersheds. Some studies in this review paper have not used the warming year, as indicated in Table 6, even if it might be that the authors did not write the warming year in the journal as they used it in their analyses. The warming year that authors utilized varied as well (from 2 years up to 4 years). Further research may be necessary, particularly for single watersheds, to determine whether or not the warming year, the ratio of calibration to validation data, the size of the watershed, and the sequence of validation and calibration data affect the model’s performance.

5.1. Comparison of daily and monthly time step simulation performance

The SWAT model produces precise simulation results for the daily and monthly time steps (Figure 11), which are beneficial for the basin’s water resource management (Vilaysane et al., 2015). The outcomes showed that the daily time step streamflow simulation is not as accurate as the monthly flow. The error statistics like PBIAS, RMSE, and others haven’t been examined, though, and it’s unclear whether the time step has the greatest or lowest bias among the simulated and actual data. Himanshu et al. (2023) have compared the daily and monthly simulation performance of the individual model with the ensemble model. The monthly time-based strong performances are based on the same findings as Vilaysane et al. (2015) for the index measures of NSE in all models (SWAT, VIC, and ensemble of VIC and SWAT). On the other hand, in all models for PBIAS, the daily time base performs better than the monthly time base. The results of Chathuranika et al. (2022) also supported the monthly time bases’ superior fit in terms of NSE and R², with the PBIAS for the HEC-HMS model being the same for both monthly and daily time bases.

Figure 11. Comparisons of Monthly and daily time bases simulation performance (R² and NSE).

Tibangayuka et al. (2022) evaluated the HBV, HEC-HMS, and ANN models’ daily and monthly streamflow simulation performance for a high-humidity tropical watershed in Tanzania. The study’s conclusions showed that in every R², NSE, and PBIAS of the three models taken into consideration, the monthly time bases performed better than the daily time bases. HBV (0.74, 0.64), HEC-HMS (0.76, 0.76), ANN (0.67, 0.52); HBV (0.73, 0.64), HEC-HMS (0.76, 0.68), ANN (0.66, 0.55); HBV (0.39, −3.7), HEC-HMS (0.07, 20.61), ANN (−12.73, −16.93) are the daily time bases for R², NSE, and PBIAS, respectively. Figure 11 supports the comparisons between the daily and monthly streamflow simulation output performance evaluation interpretation as an example. It shows that, in terms of performance index (NSE and R²), the monthly time bases outperformed the daily time bases; however, the error measure (PBIAS) does not indicate better performance.

5.2. Performance of some lumped conceptual hydrologic model

The lumped-conceptual hydrological model (MIKE-NAM) performance findings also revealed that different watersheds have different performance ratings, ranging from very good performance to not satisfactory, with the same performance rating status as the physically based-distributed model, as shown in Figure 12. Even if the percentage of performance ratings fluctuated when the quantity of review papers changed, the proportion of very good and good performance ratings is still significantly greater than the percentage of satisfactory and not satisfactory performance ratings (Table 7).

Figure 12. (a) MIKE-NAM overall performance rating; (b) NSE performance rating.

Table 7. MIKE-NAM performance index value.

Author name	Watershed	Country	Area (km^2)	R^2	NSE	RMSE	Stream (year)	Warming up year	NO. year (cal/val)	Ratio (cal/val)
Nannawo et al. (2022)	Bilate/Rift Valley	Ethiopia	5402	0.79	0.79	4	1995–2020	3	15	1.87
				0.76	0.73	5.2			8
Ghosh et al. (2022)	Bhagirathi–Hooghly	India	1690	0.93	–	4.21	2005–2018	–	10	2.2
				0.95	–	5.24			4
R. Kumar et al. (2022)	Kashmir	India	5490	0.95 0.89	0.91 0.96	– –	2006–2013	–	4 4	1.0
K. S. Kumar et al. (2021)	Shipra	India	5600	0.68 0.64	0.78 0.70	0.78 0.70	1994–2005	–	6 3	2.0
Sajadi Bami et al. (2020)	Gonbad	India	4.47	– –	0.80 0.72	– –	1994–2005	–	6 3	1.0
Aredo et al. (2021b)	Shaya/Ganaledawa	Ethiopia	457.6	0.93 0.92	0.86 0.83	3.16 4.04	1990–2015	–	19 7	2.7

Using the GR4J model, Li et al. (2014) have looked into how runoff in the Yarlung Tsangpo River basin on the Tibetan Plateau is being impacted by climate change. Eight streamflow gauge station watersheds in the basin were the subject of independent investigations, with watershed areas ranging from 3761 km² to 49,739 km². When the model’s performance was assessed using the NSE, it was found to be independent of the watershed’s area, the number of data points per year, the number of streamflow data points (years) used for both calibration and validation, and the ratio of the number of calibration points to the number of validation points (Figures 7–9). Of the eight distinct watersheds in the basin, five exhibit very good performance; the other three exhibit satisfactory, poor, and good performance, respectively, and are represented by the Shigatse Station Watershed, Lhatse Station Watershed, and Gyantse Station Watershed as shown in Figure 13. Tian et al. (2013) also evaluated the streamflow’s climate impact using GR4J. Based on the NSE performance evaluation, it performed exceptionally well (0.93 and 0.91 for validation and calibration, respectively).

Figure 13. (a) GR4J NSE versus watershed area; (b) Overall performance rating; (c) Number of data with NSE.

6. Performance comparisons of the SWAT model with some model

As seen in the summary of Figure 14 and in Tables 8 and 9, the overall single hydrological model performance comparisons showed that neither model is unquestionably superior to the other comparative model. It is interesting to note that hydrologic models were created in certain environments and then applied to watersheds with different climatic and watershed characteristics (Chathuranika et al., 2022). The watershed’s hydrological features determine whether a model’s structure is appropriate. It’s possible that a model that faithfully captures the main hydrological processes in one watershed won’t work as well in another with distinct features. Watersheds differ in terms of soil types, land cover, and topography hetrogeniety, which can affect the ideal parameter values. There’s a chance that a model that’s been calibrated for one watershed using the right parameter values won’t translate well to another (Ávila et al., 2022). Due to variations in hydrological features, data accessibility and quality, model complexity and structure, parameterization, human activity, climate variability and change, and model uncertainty, the performance of hydrological models can differ between watersheds. To get dependable and accurate results, hydrological modelling techniques must be specifically tailored to each watershed’s unique characteristics (Sempewo et al., 2022).

Figure 14. Comparisons of the SWAT model with some hydrological models.

Table 8. Comparisons of the SWAT model with HBV, HEC-HMS, GR4J, and ML.

SWAT and HEC-HMS
Author	Country	Watershed	Simulation	R^2	NSE	PBIAS
Bizuneh et al. (2021)	Ethiopia	Gilgel Abay	Calibration	HBV best	SWAT best	HBV best
			Validation	SWAT best	SWAT best	HBV best
		Anger	Calibration	SWAT best	SWAT best	SWAT best
			Validation	SWAT best	SWAT best	SWAT best
		Maybar	Calibration	SWAT best	SWAT best	SWAT best
			Validation	HBV best	SWAT best	SWAT best
Temesgen (2019)	Ethiopia	Geba	Calibration	SWAT best	SWAT best	HBV best
			Validation	SWAT best	SWAT best	HBV best
Faiz et al. (2018)	China	Songhua	Calibration	HBV best	SWAT best
			Validation	HBV best	HBV best
Bugaets et al. (2023)	Russia	Medvezhy	Calibration	HBV best	HBV best
		Elovy	Calibration	HBV best	HBV best
Thapa et al. (2017)	Nepal	Kathmandv	Calibration	HBV best	HBV best	HBV best
			Validation	SWAT best	SWAT best	HBV best
Aliye et al. (2020)	Ethiopia	Katar	Calibration	SWAT best	SWAT best	SWAT best
			Validation	SWAT best	SWAT best	SWAT best
Koch et al. (2020)	China	Cuntan	Calibration	SWAT best	SWAT best
			Validation	SWAT best	SWAT best
		Pingashan	Calibration	SWAT best	SWAT best
			Validation	SWAT best	SWAT best
		Shigu	Calibration	SWAT best	HBV best
			Validation	SWAT best	HBV best
		Zhimenda	Calibration	SWAT best	SWAT best
			Validation	SWAT best	SWAT best
SWAT and HEC-HMS
Author	Watershed	Analysis	R^2	NSE	PBIAS
Doliso and Hatiye (2021	Bilate/Rift Valle	Calibration	SWAT best	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best	SWAT best
	Gidabo/Rift Valley	Calibration	HEC-HMS best	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best	HEC-HMS best
	Bilate/Rift Valley	Calibration	HEC-HMS best	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best	SWAT best
	Gidabo/Rift Valley	Calibration	HEC-HMS best	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best	SWAT best
Aliye et al. (2020)	Katar/Rift valley	Calibration	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best
Gelete et al. (2023)		Calibration	SWAT best	SWAT best	SWAT best
		Validation	SWAT best	SWAT best	HEC-HMS best
Sempewo et al. (2022)	Manafuwa	Calibration	HEC-HMS best	HEC-HMS best	HEC-HMS best
		Validation	SWAT best	HEC-HMS best	HEC-HMS best
	Sezibwa	Calibration	HEC-HMS best	HEC-HMS best	HEC-HMS best
		Validation	HEC-HMS best	HEC-HMS best	HEC-HMS best
Chathuranika et al. (2022)	Huai bang	Calibration	SWAT best	HEC-HMS best	SWAT best
		Validation	SWAT best	HEC-HMS best	SWAT best
Fanta and Sime (2022)	Toba/Abay basin	Calibration	HEC-HMS best	HEC-HMS best	SWAT best
		Validation	HEC-HMS best	HEC-HMS best	HEC-HMS best
Sanjay Shekare and Vinay (2021)	Hemavathi river	Calibration validation	SWAT best SWAT best	SWAT best SWAT best
D. Kumar & Bhattacharjya, (2020)	Tehri/India	Calibration validation	SWAT best HEC-HMS best	SWAT best SWAT best	HEC-HMS best HEC-HMS best
		Calibration Validation	SWAT best HEC-HMS best	HEC-HMS best HEC-HMS best	SWAT best SWAT best
(Khoi, 2016)	Srepok/Vietnam	Calibration	SWAT best	SWAT best	SWAT best
		Validation	SWAT best	SWAT best	HEC-HMS best
SWAT and GR4J
Author (Anshuman et al., 2019); Country (India), Basin (Upper Godavari; 21,000 km^2)
Model	Sub Watershed	Analyses	R^2	NSE	PBIAS	R^2	NSE	PBIAS
SWAT	Adhala	calibration	0.82	0.81	−5.12	equal	SWAT best	SWAT best
		validation	0.63	0.56	−52.9	GR4J best	GR4J best	GR4J best
GR4J		calibration	0.82	0.79	−13.31
		validation	0.71	0.61	−25.47
SWAT	Alandi	calibration	0.88	0.82	13.4	SWAT best	SWAT best	GR4J best
		validation	0.71	0.57	35.8	GR4J best	GR4J best	GR4J best
GR4J		calibration	0.81	0.72	−1.3
		validation	0.81	0.76	−5.68
SWAT	Bahandra	calibration	0.97	0.75	48.6	SWAT best	SWAT best	SWAT best
		validation	0.75	0.52	28.83	GR4J best	GR4J best	GR4J best
GR4J		calibration	0.65	0.6	50.69
		validation	0.81	0.75	28.83
SWAT	Kadwa	calibration	0.89	0.88	1.3	SWAT best	SWAT best	SWAT best
		validation	0.6	0.59	0.61	GR4J best	GR4J best	SWAT best
GR4J		calibration	0.82	0.79	−2.07
		validation	0.75	0.72	−12.31
SWAT	Mula	calibration	0.78	0.78	−1.7	GR4J best	GR4J best	GR4J best
		validation	0.43	0.42	25.6	GR4J best	GR4J best	SWAT best
GR4J		calibration	0.84	0.82	−1.22
		validation	0.61	0.57	37.85
Tegegne et al. (2017)	Gilgel Abay	Ethiopia		calibration		SWAT best	SWAT best	SWAT best
				validation		SWAT best	SWAT best	SWAT best
	Gumara			calibration		GR4J best	SWAT best	GR4J best
				validation		SWAT best	SWAT best	GR4J best
	Megech			calibration		GR4J best	SWAT best	SWAT best
				validation		GR4J best	GR4J best	GR4J best
	Ribb			calibration		GR4J best	GR4J best	SWAT best
				validation		GR4J best	GR4J best	GR4J best
	Model	Watershed	Analysis	R^2	NSE	PBIAS	R^2	NSE	PBIAS
Jimeno-Sáez et al. (2018)	SWAT	Ladra	Calibration	0.59	0.59	2.55	ANN best	ANN best	ANN best
			Validation	0.58	0.57	7.18	ANN best	ANN best	SWAT best
	ANN		Calibration	0.63	0.64	2.03
			Validation	0.61	0.59	12.94
	SWAT	Segura	Calibration	0.53	0.52	4.37	ANN best	ANN best	SWAT best
			Validation	0.49	0.48	8.74	ANN best	ANN best	ANN best
	ANN		Calibration	0.57	0.54	9.03
			Validation	0.52	0.49	5.28
Ahmadi et al. (2019)	SWAT	Kan	Calibration	0.79	0.77	2.08	ANN best	ANN best	ANN best
			Validation	0.68	0.65	3.33	ANN best	ANN best	ANN best
	ANN		Calibration	0.88	0.86	1.6
			Validation	0.87	0.85	2.2
S. Khan et al. (2023)	SWAT		Calibration	0.8	0.73	15.7	ANN best	ANN best	ANN best
			Validation	0.71	0.66	17.3	ANN best	ANN best	ANN best
	ANN		Calibration	0.88	0.82	0.008
			Validation	0.86	0.95	0.1
(Rahman et al., 2022)	SWAT		Calibration	0.86	0.84	−6.98	ML best	ML best	ML best
		Gilgit	Validation	0.79	0.77	−7.9	ML best	ML best	ML best
	MLP		Calibration	0.99	0.99	0.024
			Validation	0.98	0.98	0.033
	SWAT		Calibration	0.83	0.85	3.51	ML best	ML best	ML best
		Hunza	Validation	0.76	0.73	8.4	ML best	ML best	ML best
	MLP		Calibration	0.98	0.98	−0.037
			Validation	0.98	0.98	−0.046
	SWAT		Calibration	0.8	0.83	8.61	ML best	ML best	ML best
		Shatial	Validation	0.74	0.71	14.2	ML best	ML best	ML best
	MLP		Calibration	0.92	0.9	0.095
			Validation	0.91	0.87	0.105

Table 9. Comparisons of some lumped conceptual model.

Author (Anshuman et al., 2019); Country (India), Basin (Upper Godavari; 21,000 km^2)
Model	Sub watershed	Analyses	R^2	NSE	PBIAS	R^2	NSE	PBIAS
AWBM	Adhala	Calibration	0.88	0.78	−16	AWBM best	GR4J best	GR4J best
		Validation	0.69	0.58	−12.24	GR4J best	GR4J best	AWBM best
GR4J		Calibration	0.82	0.79	−13.31
		Validation	0.71	0.61	−25.47
AWBM	Alandi	Calibration	0.79	0.68	13.4	GR4J best	GR4J best	GR4J best
		Validation	0.75	0.69	5.25	GR4J best	GR4J best	AWBM best
GR4J		Calibration	0.81	0.72	−7.49
		Validation	0.81	0.76	−5.68
AWBM	Bahandra	Calibration	0.62	0.56	48.6	GR4J best	GR4J best	AWBM best
		Validation	0.75	0.73	28.83	GR4J best	GR4J best	GR4J best
GR4J		Calibration	0.65	59.33	50.69
		Validation	0.81	44.64	28.83
AWBM	Kadwa	Calibration	0.87	0.84	6.52	AWBM best	AWBM best	GR4J best
		Validation	0.92	0.92	−14.17	AWBM best	AWBM best	GR4J best
GR4J		Calibration	0.82	0.79	−2.07
		Validation	0.75	0.72	−12.31
AWBM	Mula	Calibration	0.86	0.85	3.59	AWBM best	AWBM best	GR4J best
		Validation	0.74	0.73	2.45	AWBM best	AWBM best	AWBM best
GR4J		Calibration	0.84	0.82	−1.22
		Validation	0.61	0.57	37.85
GR4J and IHACRES
Authors (Tegegne et al., 2017); country Ethiopia; River basin (Upper Blue Nile)
Model	Sub watershed	Analyses	R^2	NSE	PBIAS	R^2	NSE	PBIAS
IHACRES	Gilgel Abay	Calibration	0.61	0.61	0.95	IHACRES best	IHACRES best	IHACRES best
		Validation	0.66	0.62	1.06	IHACRES best	IHACRES best	IHACRES best
GR4J		Calibration	0.59	0.57	6.26
		Validation	0.64	0.58	6.64
IHACRES	Gumara	Calibration	0.79	0.79	7.73	IHACRES best	IHACRES best	GR4J best
		Validation	0.8	0.78	6.76	IHACRES best	IHACRES best	GR4J best
GR4J		Calibration	0.77	0.75	−0.55
		Validation	0.77	0.75	2.44
IHACRES	Megech	Calibration	0.34	0.3	27.82	IHACRES best	IHACRES best	IHACRES best
		Validation	0.44	0.3	−19.3	IHACRES best	IHACRES best	GR4J best
GR4J		Calibration	0.32	0.27	30.6
		Validation	0.42	0.27	13.2
IHACRES	Ribb	Calibration	0.58	0.19	−94.5	IHACRES best	GR4J best	GR4J best
		Validation	0.58	0.36	−6.55	GR4J best	GR4J best	GR4J best
GR4J		Calibration	0.57	0.56	7.4
		Validation	0.6	0.56	6.02

As a result, choosing models for a particular watershed based only on the work of one author or one watershed’s results may introduce uncertainty and reduce the acceptability of the model’s output. A well-fitting model will result in worse performance in a different watershed when used as a comparative model. The most flexible model, SWAT, was reviewed in this study along with conceptual lumped models (GR4J), spatially distributed physically-base models (HEC-HMS, MIKE-SHE, HBV, and VIC), and machine learning models (Figure 14).

A comparison of the effectiveness of the SWAT and MIKE-SHE models was conducted by J. Liu et al. (2016) and Golmohammadi et al. (2014). The two authors’ results diverge. J. Liu et al. (2016) state that using the NSE, R², and PBIAS assessment indexes, respectively, SWAT (0.76, 0.78, and 0.22%) performs better than MIKE-SHE (0.71, 0.70, and 8. 12%). In contrast, Golmohammadi et al. (2014) found that using NSE, R², and PBIAS assessment indices for calibration and validation, respectively, MIKE-SHE (0.84, 0.64; 0.79, 0.71; and 6.67, 3.57) performs better than SWAT (0.74, 0.64; 0.73, 0.7; and −3.14, −12.5).

Himanshu et al. (2023) and Z. Li et al. (2018) have compared the SWAT model with the macro watershed model VIC. All three authors’ results indicated that the SWAT model has better performance than the VIC model. However, Wu et al. (2021) results indicated VIC is outperformed than SWAT. Overall, the HEC-HMS outperforms the SWAT model in various watersheds, according to comparisons between the two models. Nevertheless, the percentage also fluctuated when the number of evaluating papers changed, and the percentage of SWAT might have improved or decreased. This review’s conclusions unmistakably showed that SWAT performs better than HEC-HMS in some watersheds and that HEC-HMS performs better than SWAT in others. Furthermore, at different study times, Gelete et al. (2023) and Aliye et al. (2020) compared SWAT and HEC-HMS in the Katar watershed, Ethiopia. Gelete’s (2023) results suggest that SWAT performs better, but Aliye et al. (2020) found that HEC-HMS performs better. This finding demonstrates the complexity of the hydrological system and the difficulty of simulating it.

The results of comparing machine learning (ML) and SWAT showed that ML performed significantly better than SWAT. Planning and managing sustainable water resources depends on accurate hydrological process simulation. The better performance of machine learning-based streamflow simulation than the SWAT model is due to various reasons. The reasons given by Rahman et al. (2022) for the high performance of machine learning than SWAT model are choosing sensitive parameters in SWAT modeling, possible drawbacks of the SCS-CN method for simulating streamflow, and SWAT's inability to record the hydro peaking. The low accuracy of the snowmelt runoff simulations in the SWAT model may be attributed in part to the assumption regarding the snowmelt factor not being well supported by the actual processes (Zhao et al., 2022). In mountainous areas, SWAT frequently performed poorly and needed extra procedures to be effective (K. I. Islam et al., 2023). Numerous types of input data are needed for the SWAT model, including elevation, data on land use and cover, drainage, geology, and climate. Unfortunately, due to unavoidable issues like uneven observed station distribution, social and political unrest, and limited data sharing amongst nations in transboundary basins, some of these data are not entirely accessible in many regions. Since this method doesn’t require a lot of data, machine learning models are helpful in this situation Rahman et al. (2022). Therefore, due to the limited input data especially in poor countries, the processed-based model (SWAT) had limitations to topography like mountainous areas, peak flow, and snow areas making the SWAT model low-performance.

Comparing the two lumped models (GR4J with AWBM and GR4J with IHACRES) also showed that while GR4J performs better than both AWBM and IHACRES in some watersheds, its performance is inferior in other watersheds as shown in Figure 14 and Table 9. The models’ performance comparisons also reveal that GR4J performs worse during validation and better during calibration with R² or vice versa. Furthermore, GR4J performs worse in the NSE or PBIAS and is superior with R². This review concludes that there is a lack of consistency in the performance of the hydrological models.

7. Conclusion

Climate change is currently the biggest issue affecting water resources. The interpretation of this review paper is bolstered by the use of a relatively water-resource-rich country in Africa (Ethiopia). The majority of the studies in this paper indicate that future temperatures would rise in most basins of the country, with the Wabishebele River basin experiencing an 8.2 °C change in annual maximum temperature. Its future streamflow and rainfall prediction vary with trend (increase or decrease) and also the rate of change between watershed, season, models, and researchers. For instance, Gidabo may suffer an 86% yearly decrease in precipitation, but Upper Awash might receive a 56% annual increase in precipitation. The Gidabo River may witness a 43% decline in stream flow, which may result in water scarcity, while the Awash River basin may see a 191% increase in stream flow, which may lead to flooding issues. Because of this stark disparity in the trend of hydrological change, one needs to conduct a study on climate impact effect evaluation for each watershed or basin to minimize the havoc of water and maximize the benefits associated with it. Therefore, through a simplified and continually upgraded hydrologic model, the streamflow simulation accuracy is evaluated with the model performance evaluation index.

The model’s performance in the simulation is the primary criterion after taking into account all other factors including cost, time, simplicity, and study aim when choosing a model. At least one performance index and one error measure are required for performance evaluation. However, the classical performance evaluation has limitations for oversensitivity to high flows, overestimation of low flows, and underestimation of high flows. Therefore, currently, the NSE-based shortcomings improved with KGE, and Log NSE; the over sensitivity to high flow, underneath valuation of peak flow, and excess valuation of low flow problem can be improved by using hydrological signature. The strength of the hydrologic signature is to diagnose weaknesses in hydrological models.

Additionally, the performance of the model is also affected by the uncertainty analysis technique used in the calibration and validation. The comparisons of the uncertainty evaluation techniques include GLUE, MC, MCMC, SUFI-2, BS, PSO, and PS, and selecting the one to provide high-performance simulations will help to enhance the model’s functionality. Even though most studies underestimate this process evaluating the streamflow simulation using uncertainty measure indices like the relative width of a 95% probability band (R factor) and 95% prediction uncertainty (95PPU) (P factor) shows the strength of the simulation accuracy.

Numerous hydrological models have been developed worldwide, ranging from the basic empirical rational model to the lumped conceptual model and then upgraded to the physically based and distributed model. Choosing one hydrological model for investigation from a variety of models is a difficult issue for a researcher because both knowledge-based hydrological models and machine learning model components are numerous.

Machine learning can approximate NSE and R² quite well, up to 0.99, though this isn’t always the case for all watersheds. For instance, the results of performance evaluations for lumped conceptual models (MIKE-NAM and GR4J) and physically and distributed models (SWAT, HEC-HMS, MIKE-SHE, VIC, and QSWAT) in this review showed that all models can perform well (not satisfactory up to very good).

The most flexible and versatile SWAT model has been compared with MIKE-SHE, VIC, HEC-HMS, HBV, GR4J, and ML to determine which model can perform better than the others. The review’s results showed that ML performed well with SWAT, with notable variation. Nevertheless, neither SWAT nor any other model outperformed the other model. There are study areas where SWAT performed very well and also areas where SWAT performed poorly. A single hydrological model has limits, such as certain models that overestimate low flow and perform well for mean flow; other models may perform well for peak flow, but because of this, the model’s total performance may be reduced. Thus, the shortcomings of the individual model can be improved by the ensemble multi-single hydrological model. The ensemble strategy used, which includes the weighted average, simple averaging, and Bayesian averaging method based on the output of the ensemble of single hydrological models, also affects the ensemble model’s performance. It is therefore necessary to compare the results of the various ensemble techniques to ascertain which is superior.

Relatively, ML modeling has high performance, however its hydrological significance is low. Instead, especially the physically based-distributed has relatively higher hydrological significance and considering the heterogeneity of the watershed though has relatively low performance even if not applicable to all watersheds. Therefore, from this review, it can be depicted that hybridizing the conceptual models based on physical aspects along with the machine learning model is important to maximize the model’s performance. Using this approach, the water resources potential in the river basins and watersheds for a long period of time with out the existence of extreme events can be managed. Based on reviewing the overall summary of future research directions, this review recommends that single-hydrological model cannot provide satisfactory information. Therefore, selecting the best ensemble of conceptual & physical based hydrological models coupled with machine learning and evolutionary algorithms to compare objective functions like NSE and KGE as hydrological signatures and choosing the best uncertainty evaluation techniques, especially for automatic calibration can be a healthy approach.

Authors contributions

Yilak Taye Bihon and Tarun Kumar Lohani conceived of the presented idea, drawn the graphs and figures, and drafted the manuscript. Bogale Gebremariam Neka and Abdella Kemal Mohammed developed the theory, prepared the tables, and performed the computations. Elias Gebeyehu Ayele and Abebe Temesgen Ayalew verified the analytical methods and supervised the findings of the entire work. Getachew Bereta Geremew revised it critically for intellectual content. All authors discussed the results contributed to the final manuscript and agreed to be accountable for all aspects of the work.

Disclosure statement

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

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Additional information

Notes on contributors

Yilak Taye Bihon

Yilak Taye Bihon is a PhD scholar in Hydraulic Engineering, Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Tarun Kumar Lohani

Tarun Kumar Lohani is a Professor, Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Abebe Temesgen Ayalew

Abebe Temesgen Ayalew is an Assistant Professor cum Deanin Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Bogale Gebremariam Neka

Bogale Gebremariam Neka is an Associate Professor in Faculty of Hydraulic & Water Resources Engineering cum Scientific Director,Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Abdella Kemal Mohammed

Abdella Kemal Mohammedis an Associate Professor in Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Getachew Bereta Geremew

Getachew Bereta Geremewis an Assistant Professor in Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.

Elias Gebeyehu Ayele

Elias Gebeyehu Ayeleis an Assistant Professor in Faculty of Hydraulic & Water Resources Engineering, Water Technology Institute, Arba Minch University, Arba Minch, Ethiopia.