Intercomparison of SWAT and ANN techniques in simulating streamflows in the Astore Basin of the Upper Indus

The main purpose of the hydrological cycle is to convert rainfall into runoff (Jimeno-Sáez et al. 2018). This is a complicated process that involves a non-linear relationship between rainfall, runoff, and other hydrological processes as well as watershed features. Hydrometeorological processes (interception, evapotranspiration (ET), percolation, and infiltration) impacts surface runoff (Haruna & Abba 2021). These hydrological processes are also influenced by watershed attributes and weather patterns. Moreover, the spatiotemporal land-use dynamics and climate change produce a significant level of uncertainty (Niedda et al. 2014). Rainfall–runoff modeling is used for water resource planning and extreme events assessment (Jehanzaib et al. 2022). The hydraulic structures design and extreme events forecasting mainly depends upon the rainfall–runoff relationship (Rajurkar et al. 2004).

Hydrological models are developed using the fundamental principle of water cycle (Chen et al. 2012). The structure of the aforementioned models varies from simple lumped models to complicated process-based distributed models. There are numerous hydrological models namely SWAT, Systeme Hydrologique European (SHE) model, Institute of Hydrology Distributed Model (IHDM), Hydrological Simulation Program-FORTRAN (HSPF), and Variable Infiltration Capacity (VIC) which are widely used for hydrological simulations (Jehanzaib et al. 2022). Hydrological model selection is purely based on model simulation capabilities and data availability of the project area (Henriksen et al. 2003). SWAT is one of the most sophisticated semi-distributed hydrological model which assess streamflow at watershed scale (Grusson et al. 2017). SWAT requires a large amount of spatiotemporal hydrometeorological and terrestrial data to evaluate streamflow (Pang et al. 2020). The model parameterization and calibration process are complicated as it is based on large number of parameters and their complex interconnections (Noori & Kalin 2016). SWAT-CUP (SWAT Calibration and Uncertainty Procedures) is a program which excessively used to perform SWAT model calibration, validation, uncertainty analysis and sensitivity analysis (Abbaspour et al. 2015).

It is well-founded knowledge that the SWAT model is excessively used for streamflow assessment (Cibin et al. 2010; Grusson et al. 2017). Haleem et al. (2022) used SWAT to isolate the impacts of climate change and anthropogenic activities on streamflow in the Upper Indus Basin (UIB). Wisal et al. (2020) assessed the suitability of gridded precipitation data for streamflow forecasting using the SWAT model. Li et al. (2010) used the SWAT model to simulate streamflow and sediment load. Setegn et al. (2010) used the SWAT model to examine the impact of regional geography, land use, soil, and climate on hydrology. Tibebe & Bewket (2011) computed monthly streamflow and erosion rate in the Keleta region of Ethiopia using the SWAT model. In mountainous basins, SWAT is frequently employed to investigate how snow affects the various components of water cycle (Shahid et al. 2021).

The literature shows that data-driven models are widely used to simulate hydrological processes at global scale (Rajurkar et al. 2004; Wang et al. 2006). The aforementioned models are considered superior owing to their minimal input meteorological variables requirement, less computation time, and ability to deliver highly precise and accurate results. Artificial intelligence (AI) approaches like the Artificial Neural Network (ANN) have emerged as a viable alternative to traditional hydrological models. ANN's capacity to model non-linear interactions between the input and output parameters without the need for understanding the underlying physical processes is perhaps their strongest feature, making them more suitable than other methods (Noori & Kalin 2016). Systems-related data can be analyzed using empirical approaches like AI, soft computing (SC), data mining (DM), computational intelligence (CI), and machine learning without explicitly accounting for the physical behavior of the objective system (Solomatine et al. 2008).

Machine learning algorithms are excessively used in hydrological modeling (Solomatine et al. 2008; Khan et al. 2023). ANN-based models have been used for rainfall–runoff and sediment simulations (Tokar & Johnson 1999). ANN models have been used to predict floods and reservoir inflows (Lohani et al. 2012). Huang et al. (2014) used support vector machine (SVM) to simulate the river's monthly flow at Huaxi station, and the findings demonstrated that the model has a high level of accuracy in predicting the river's monthly flow. SVM models and ANN were used by Ghorbani et al. (2016) to estimate the daily flow of the Cypress River in Texas. Their findings suggested that the SVM model performs better than the ANN in terms of river flow forecasting and precision.

In the last decade, there have been few studies comparing SWAT with data-driven models. Pradhan et al. (2020) compared the accuracy of three ANN models to the SWAT model in predicting streamflow and found that ANN models results are more precise. Koycegiz & Buyukyildiz (2019) compared SWAT, support vector machines (SVM), and ANN in the headwaters of the Carsamba River, which are located in the Konya Closed Basin of Turkey where SWAT was outscored in streamflow simulation by the data-driven models (ANN and SVM). Jajarmizadeh et al. (2015) used SWAT and SVM models to assess monthly stream flow in southern Iran. The SVM results were superior to SWAT model. Daily flow in Portugal's Pracena basin was predicted using ANN and SWAT models by Demirel et al. (2009). According to the findings, the ANN model was more accurate in forecasting peak flow than the SWAT model.

In the current research work, monthly streamflow data is simulated using the ANN and SWAT models for the Astore Basin located at Indus River. Intercomparison of empirical and hydrological model has not been reported in the region. The gap was filled by the current study to assess the performance of ANN and SWAT model in simulating the streamflow. As Pakistan is a data-scarce country, this study will assist the water experts and policy-makers in gaining reliable alternative to data-intensive hydrological models in the stated region.

The UIB is located in the extreme North of Pakistan which lies between 33.67° and 37.20° N, and 70.50° and 77.50° E coordinates, which is mainly covered by snow and glaciers. The elevation of UIB ranges between 8,500 and 200 m above mean sea level. UIB's border is shared by China, India, and Afghanistan. Additionally, Pakistan and Afghanistan both share the Indus River Basin, where Pakistan cover the larger portion 32.48–37.07°N and 67.33–81.83°E (Haleem et al. 2022). UIB contributes to over half of Pakistan's surface water resources and is crucial to the country's long-term economic growth.

Precipitation and temperature (min and max) historical record from 1985 to 2010 were collected from Pakistan's Meteorological Department (PMD). Hydrological information, including monthly and mean flow, was collected from the Water and Power Development Authority (WAPDA). The 26-year span (1985–2010) was split into two phases: the calibration phase (1985–2005) and the validation phase (2006–2010). All monthly streamflow data were verified to be complete and free of any significant gaps. PMD and WAPDA follow the International Meteorological Organization's Guide on Hydrological Practices for monitoring hydrometeorological parameters (Rahman et al. 2022).

SWAT is a process-based semi-distributed watershed hydrological model that was introduced by the United States Department of Agriculture – Agriculture Research Services (USDA-ARS) in 1990. It was designed to assess the impacts of meteorological and terrestrial determinants on streamflow (Wisal et al. 2020). In order to give precise spatial details and increase the model's accuracy, SWAT divides huge basin into smaller sub-basins. The SWAT model was developed to predict and simulate that how land use and agricultural practices will affect water resources in terms of water quantity and quality at the basin scale (Rahman et al. 2022). Furthermore, scenario-based simulations using the SWAT model were mainly used to explore the hydrological response to land-use dynamics and climate change (Yin et al. 2017).

SW₀ and SW_t denote the initial and final levels of soil moisture. P_day denotes daily precipitation, R_Surf denotes the surface runoff, ET_day denotes daily evapotranspiration, W_weep denotes water that has infiltrated into the ground, and R_gw denotes groundwater return flow. All of the variables listed above have ‘mm’ as their units.

The Doyian monitoring station at the Astore Basin was chosen for model calibration to bring in front the most dominant parameters of the streamflow using SWAT-CUP which was automatically calibrated through Sequential Uncertainty Fitting program algorithm (SUFI-2) (Abbaspour et al. 2015). The SUFI-2 algorithm employs 11 distinct objective functions, including the R², Modified Coefficient of Determination (b R²), NSE, Modified Nash–Sutcliffe Efficiency (MNS), Ratio of the Standard Deviation of Observations to the Root-Mean-Square-Error (RSR), Ranked Sum of Squares (SSQR), Kling–Gupta Efficiency (KGE), PBIAS, Multiplicative Form of the Square Error (MSE), Summation Form of the Square Error, and the χ². These objective functions collectively provide a rigorous framework for model assessment and parameter calibration, enabling researchers to make informed decisions and refine models to accurately represent complex real-world processes. While a variety of objective functions are available for optimizing parameter ranges, our study focused on the assessment of a single objective function, the NSE. This decision was driven by NSE's inherent ability to effectively capture temporal dynamics, a critical aspect in hydrologic calibration. Moreover, NSE holds a prominent and well-established position as one of the most widely adopted statistical measures in the field (Gassman et al. 2007). By focusing on NSE, we aimed to ensure a comprehensive and precise evaluation of temporal calibration within the context of our study. Monthly streamflow data from 1985 to 2005 were used to calibrate the SWAT-CUP model. The main purpose of calibration is to determine goodness-of-fit index (GFI) between the measured and modeled data using R², NSE, PBIAS, and RMSE statistical performance indicators. For R² and NSE, the most favorable value is 1, whereas it is 0 for PBIAS and RMSE. After satisfactory calibration process, the model was tested using historical record ranging from 2006 to 2010.

Typically, dataset is split into segments while modeling through ANN technique. The datasets used in this study were split into two phases: the training phase (80% of the data) and the testing phase (20% of the data). So, the 26-year (1985–2010) historical record was split into training (1985–2005) and testing (2006–2010) periods. In order to train the network, we applied a backpropagation approach which compares modeled values to observed data and then computes the network error. The output error was repeatedly recycled backwards through the network until the network's parameters were adjusted to desirable values (Javan et al. 2015). The training phase comes to an end when the computed error reduces to its minimum level in the validation datasets (Gholami et al. 2015). There are numerous backpropagation approaches for network training. It is well-founded knowledge that the Levenberg–Marquardt (LM) algorithm is superior to other algorithms in terms of higher performance (smaller estimated error) and faster convergence (when calculating epoch size) (Nayebi et al. 2006). Literature shows that streamflow is widely forecasted using the LM algorithm (Yaseen et al. 2015).

The parameter m is equal to 1, and W_n represents the actual value of a statistical performance measure, whereas indicates the perfect value of the performance measure, obtained when model simulations fully match with observed data. The L_p metric is always positive. Lower L_p values are preferred since they imply higher model performance.

SWAT is a hydrological physically based semi-distributed model which was initially calibrated which will reduce uncertainties in the modeled values. The 26-year historical record which ranges from 1985 to 2010 was split into calibration (1985–2005) and validation (2006–2010) periods at the Doyian monitoring station of the Astore Basin. The NSE was used as objective function during the calibration process. After model calibration, sensitivity analysis was carried out to bring in front the parameters which mainly impacts streamflow. The reader is referred to Abbaspour et al. (2015) for a detailed explanation of model calibration procedures. Table 1 illustrates the 28 most significant parameters chosen for model calibration.

As demonstrated by Table 2, four statistical performance indicators like R², NSE, PBIAS, and RMSE were used to evaluate that how well the model performed in forecasting the watershed conditions. In this study, the performance of the SWAT model was assessed using four statistical indicators, i.e., R², NSE, PBIAS, and RMSE. The R², NSE, PBIAS, RMSE value for both calibration and validation periods were (0.80 and 0.71), (0.73 and 0.66), (15.7 and 17.3), and (79.81 and 106.26), respectively.

Performance metrics have been calculated for all ANN architecture, and the resulting rankings obtained through the CP approach are shown in Table 4. According to their respective rankings, the top three architectures are (2,27,1), (2,19,1), and (2,28,1). Using the CP method, the architecture with the lowest ranking was (2,4,1). These findings highlight how well the CP technique evaluates and ranks ANN architectures, providing important information for our research's decision-making.

Accurate hydrological process simulation plays vital role in sustainable water resources planning and management. Data-driven and hydrological models are usually used for hydrological processes simulation. In this research work, ANN technique overweighs the SWAT model in predicting the streamflow in general and extreme conditions. The results of the current study regarding underestimation of the peak flow prediction by the SWAT model is supported by the literature (Demirel et al. 2009; Makwana et al. 2017). Streamflow generation processes vary significantly during low-, medium-, and high-flow periods. Low-flow events are primarily caused by base flow, whereas high-flow events are caused by intense storm rainfall (Wu et al. 2009). SWAT models are good at estimating low flows but struggle to simulate very high streamflow accurately. In contrast, a single ANN can achieve better results for very high values but may have limited accuracy in predicting the lowest values, as demonstrated by Kim et al. (2015). As a result, these models are well-suited for simulating basin streamflow. An ANN model is recommended to simulate high-flow events in studies involving extreme hydrologic events such as floods. The SWAT model, on the other hand, would be preferable for hydrological management studies where low-flow events are of greater interest.

The SWAT model weak performance may be due to the following reasons: (1) The identification of the most sensitive factors may affect the SWAT model performance (Cibin et al. 2010), (2) The identified critical snow-specific characteristics may also cause uncertainty in the model output, (3) Each sub-basin's flow accumulate within one another before reaching the Indus River, making it hard to precisely track the hydropeak of the streamflow. The formulation of SWAT is considered responsible for the observed peak-flow inefficiency, as indicated in various studies (Demirel et al. 2009; Jimeno-Sáez et al. 2018). Therefore, in terms of simulating streamflow, the SWAT model's performance is notably weaker in comparison to the ANN model.

The results of this study suggest that using ANN models can be beneficial in reducing errors when estimating high streamflow values, although there is still a tendency to underestimate them. This problem arises because high-flow values are scarce in training datasets, while medium and low values predominate, as shown in Figure 6. Similar challenges have been reported in the research of Talebizadeh et al. (2010) and Jimeno-Sáez et al. (2018). There are few drawbacks of ANN technique which are discussed below: ANN technique has no relationship with the physical characteristics of the watershed. ANN approach is based on the lumped method which ignores the heterogeneity of sub-catchment parameters. Moreover, empirical models which include ANN technique cannot take into account watershed dynamics like soil moisture, runoff generation, pollutants export, etc. The SWAT model is a process-based model which is data-intensive requiring many input parameters for its analysis like DEM, land use/cover, soil type, hydrometeorological data, etc. Unfortunately, the aforementioned parameters are not easily available everywhere due to poor station distributions, socioeconomic and political difficulties, and limited information exchange between governments in transboundary basins. Machine learning models, like ANN, are helpful in such conditions because they do not require a lot of input parameters. This study demonstrates that ANN outperforms the SWAT model despite requiring fewer resources and input parameters. This promising efficiency opens up possibilities for future developments, such as using the ANN approach to simulate water quality processes, as seen in some related studies (Kuo et al. 2007; Huo et al. 2013).

The current research work was carried out to simulate monthly streamflow data using ANN and SWAT models at the Astore Basin, Pakistan. The results of both models were compared via statistical performance indicators to assess its modeling capabilities. The best ANN architecture (2,27,1) simulated streamflow was compared with the SWAT model. For the SWAT model, the R² values were 0.80 and 0.71, NSE values were 0.73 and 0.66, PBIAS values were 15.7 and 17.3, and RMSE values were 79.81 and 106.26 for calibration and validation, respectively. On the other hand, for the ANN model with architecture (2,27,1), the R² values were 0.88 and 0.86, NSE values were 0.82 and 0.95, PBIAS values were 0.008 and 0.10, and RMSE values were 70.34 and 75.92 for calibration and validation, respectively. The results suggested that the ANN model is superior to the SWAT model for streamflow prediction. The box plot analysis further demonstrated that the SWAT model under-estimated the high flow while over-estimated the low flow. Comparatively, the ANN model performed well in estimating the general and extreme flow conditions. The findings of the study can be utilized in formulating improved water policies which will help the water managers in suggesting flood adaptation strategies, water allocation for various sectors and climate resilient water infrastructure design in the region. The impacts of additional meteorological variables are recommended to enhance the accuracy of ANN model. Despite the excellent performance of the ANN model, it is essential to acknowledge potential limitations or uncertainties associated with the modeling approaches. Therefore, further studies could explore the potential of other machine learning-based models and compare their results with the SWAT model in the Astore Basin. Such a comparative analysis would provide a comprehensive understanding of different modeling techniques and support the adoption of advanced and accurate methods in water resource management in the region. Overall, the findings emphasize the superiority of the ANN model and its potential to contribute to better water resource management decisions in the Astore Basin.

The Indus River plays a vital role in Pakistan's economic growth and food production as it contributes approximately 25% to the country's gross domestic product and provides 90% water for food production. As per the World Bank's report (2020–2021), 32% projected decrease in water by 2025 will cause approximately 70 million tons of food shortage in the country. River water allocation has always been a big problem among provinces. The projected decline in water will worsen the situation. In this regard, the findings of this research hold significant implications for policy and water resource management. The superior performance of the ANN model compared to the SWAT model in simulating monthly streamflow data during both calibration and validation periods highlights the potential of machine learning-based models in formulating cohesive inter-provincial National Water Policy. Utilizing ANN models for streamflow forecasting can lead to improved water allocation among provinces, more effective flood management strategies, and better planning for water-related infrastructure development. Additionally, the suggestion to investigate the effects of additional meteorological variables opens up opportunities for further enhancing the accuracy and comprehensiveness of the ANN model's predictions.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.