Evaluation of the SWAT model for water balance components under changing climatic conditions in the Mahanadi River Catchment, Chhattisgarh Open Access

Interest in global water issues – such as droughts, floods, eutrophication, and deteriorating water quality – is increasing among scientists, policymakers, and the general public. The hydrological cycle, which governs the distribution and movement of water within a region, is highly complex and influenced by a range of factors. These include natural elements like climate patterns and soil types, as well as human activities such as land use changes and climate change (Dey & Mishra 2017; Wu et al. 2017; Sahu et al. 2021c; Verma et al. 2022b). Understanding these processes and accurately assessing the water balance is crucial for managing water resources effectively (Shawul et al. 2013; Dhami et al. 2018; Pradhan et al. 2022; Sahu et al. 2024c).

In recent years, hydrological models have become essential tools for representing and simulating water flow and hydrological processes. These models are particularly useful for evaluating the impacts of changes in land use and climate variability on water systems (Abubakari et al. 2017; Sahu et al. 2023c). Among these models, the SWAT (Soil and Water Assessment Tool) model is well-regarded for its ability to quantify watershed loads and predict hydrological cycles across various scales (Arnold et al. 1998; Abbaspour et al. 2015; Dong et al. 2015; Gan et al. 2015; Yatheendradas et al. 2008).

However, the calibration of hydrological models like SWAT can be challenging due to various uncertainties. These uncertainties stem from the model's structure, input data quality, parameter settings, and the operational model itself (Singh et al. 2005; Song et al. 2015; Uniyal et al. 2015). Key sources of uncertainty include human error in model operations, oversimplification of complex hydrological processes, limitations in data accuracy, and the wide range of model parameters (Zhang et al. 2009; Kumar et al. 2023). To address these issues, researchers use various methods for uncertainty analysis, such as ‘Generalized Likelihood Uncertainty Estimation (GLUE), Sequential Uncertainty Fitting (SUFI-2), parameter solution (ParaSol), and the Markov Chain Monte Carlo (MCMC) approach’. These techniques help in refining models, reducing uncertainties, and improving their predictive accuracy (Yang et al. 2008; Gong et al. 2011; Abbaspour et al. 2015; Narsimlu et al. 2015; Wu & Chen 2015). SUFI-2 is chosen because it strikes an excellent balance between computational efficiency, accuracy, and uncertainty quantification. Its iterative nature and straightforward implementation make it particularly attractive for large-scale hydrological models where parameter uncertainty is a key concern.

Researchers who have focused on addressing parameter uncertainty and model input concerns in SWAT models include Li et al. (2010), Nie et al. (2011), Tan et al. (2015), and Zhao et al. (2018). Satellites providing data on various spatiotemporal scales have been used to produce input data for places which have inadequate observational data (Thiemig et al. 2013; Serrat-Capdevila et al. 2016; Tramblay et al. 2016; Sahu et al. 2021b, 2022a, 2023a, 2024a; Kumar et al. 2023; Sahu & Mehta 2024). CMIP6 GCM data (available at http://www.nature.com/scientificdataa) and Indian Meteorological Department (IMD) grid data (1985–2017) from the IMD, Pune, have also been incorporated for specific regions to improve runoff simulation in the SWAT model.

The Mahanadi River (sixth largest) serves as the critical water source for Chhattisgarh and Orissa. However, in recent years, faces challenges due to extreme climate changes and human activities that have deteriorated water quantity and quality (Verma et al. 2023; Sahu et al. 2021d, 2022b). Also, the northern part or the portion of Chhattisgarh experienced significant drying trend −7.1 mm/year (Kumar & Bassi 2021), −2.1 mm/year (middle Mahanadi, Sahu et al. 2023a), and −1.7 mm/year (Seonath basin, Sahu et al. 2023a). Therefore, it is critical to accurately anticipate streamflow in order to facilitate effective water resource management. To the best of the author's knowledge, integrating IMD grid data with CMIP6 GCM data has not been applied in the Mahanadi River Basin (MRB). This contrasts with other Indian basins, where IMD grid data has been used in areas with limited traditional weather stations. A few difficulties with scale problems in the input data used for modeling and their effects on modeling outcomes were highlighted by Singh & Kumar (2017). For instance, the study by Verma et al. (2022a) demonstrated that with conventional meteorological data, the streamflow simulation's accuracy was notably poor in the upper Mahanadi watershed.

For this reason, the SWAT model used meteorological data from IMD Pune and CMIP6 GCMs to anticipate streamflow in the higher reaches of the Mahanadi River. The SUFI-2 technique was used to perform parameter sensitivity and uncertainty evaluations, model calibration, and validation. Based on the results of the simulation, the water balance components were further examined. Highlighting the need for dynamic and adaptive water management strategies, as there were significant differences in how hydrological elements contributed to the water balance between low-flow and high-flow years. Additionally, the study identified spatiotemporal variations in hydrological elements, highlighting the relationship between precipitation and water balance components.

The two primary land use types, which account for 31.05 and 64.96% of land area, respectively, are agriculture and forestry. Other land cover categories include water bodies (2.33%) and urban areas (0.61%). Barren land, which constitutes 1.04% of the total area, is primarily located in mountainous regions affected by extensive coal, iron, and manganese mining activities. The dominant soil types in the catchment are clayey and loamy, covering 43.31 and 50.09% of the land area, respectively. Additionally, there are smaller areas of loamy skeletal (5.73%) and clay skeletal (0.70%) soils (Sahu et al. 2023a). The study area features three major dams: the Dudhawa Dam, constructed in 1964; the Gangrel Dam, built in 1979; and the Hurakud Dam, completed in 1957, arranged in order from upstream to downstream. Over the past 40 years, the catchment has experienced significant soil erosion and nutrient depletion, primarily due to industrialization and rapid economic growth. These factors have contributed to the degradation of soil quality and alterations in the hydrological dynamics of the region.

To set up the SWAT model, prepare input datasets like DEM, land use/land cover (LULC), soil, and weather data (e.g., precipitation and temperature) in the required coordinate system. Create a project in ArcSWAT, delineate the watershed using the DEM, and define hydrologic response units (HRUs) by overlaying LULC, soil, and slope maps. Input weather data, link stations, and configure the simulation settings (e.g., time step and warm-up period). Once the setup is complete, run the model and validate outputs with observed data.

In SWAT Calibration Uncertainty Procedure (SWAT-CUP), export SWAT outputs (e.g., streamflow) and prepare observed data in .txt format. Import the SWAT project and observed data into SWAT-CUP, select calibration parameters, and define their range. Choose a calibration algorithm like SUFI-2, set simulation details, and run the calibration to optimize parameters. Evaluate performance using indices like NSE and R², then validate the model with independent data for reliable application.

The soil resources were digitally classified using multiple sets of IRS satellites imagery at scales ranging from 1:250,000 to 1:12,500, to establish the first soil types ‘https://bhuvan-app3.nrsc.gov.in/data/download/index.php’. The study catchment exhibited four distinct soil classifications (Figure 2), and a lookup table was manually created to represent these classifications. For hydrological and meteorological data, the Mahanadi Catchment Authority provided monthly streamflow and some precipitation datasets. The utility of high-resolution grid-based precipitation datasets for analyzing variability and changes in the MRB has been highlighted in recent studies (Pradhan et al. 2022; Sahu et al. 2021a, 2021c, 2021d, 2022a, 2022b, 2022c, 2023a, 2023b, 2024a; Kumar et al. 2023; Sahu & Mehta 2024). Additionally, the CMIP6 GCM datasets (http://www.nature.com/scientificdata) contributed another subset of precipitation data. The SWAT model used the Weather Generator (WGEN) to create additional meteorological features, with the National Centers for Atmospheric Prediction (NCEP) providing the necessary data for WGEN development (https://globalweather.tamu.edu/). Weather stations were integrated to compute the statistical parameters required for the study catchment using WGEN.

SWAT is a hydrological model designed for watershed-scale analysis that operates on a continuous-time step and a semi-distributed model to simulate various hydrological and environmental processes. The model integrates a wide range of data, including meteorological information (e.g., rainfall and temperature), LULC, soil characteristics, and management practices to predict outcomes such as streamflow, sediment transport, nutrient loading, and pesticide movement (Arnold et al. 1998). In practice, SWAT divides a catchment area into smaller, manageable sub-catchments. Each sub-catchment is further broken down into HRUs, which are areas characterized by uniform soil properties, land cover, and slope conditions. Within each HRU, SWAT calculates key variables such as surface runoff (water flowing over land), soil moisture, sediment transport, nutrient cycles, and crop yields. These calculations are then aggregated across HRUs within a sub-catchment and routed through the river systems to model how water and its components move throughout the entire catchment area. SWAT accounts for four types of water storage: surface runoff, soil water, shallow groundwater (which can contribute to river flow or evaporate), and deep groundwater (which exits the watershed system). This detailed approach helps in understanding and managing water resources effectively. Comprehensive information on the SWAT model can be found in Neitsch et al. (2011) and is also accessible online at http://swatmodel.tamu.edu.

This study used several approaches for hydrological calculations:

The SWAT2012 model utilized a 90-m DEM to define the MRC. To ensure the river networks matched the topographic map, a threshold area of 2,600 ha was used to set the catchment boundaries. Soil and land use maps were adjusted to this boundary, resulting in the definition of 126 sub-catchments. All data layers were projected using the ‘WGS_1984_UTM_Zone_44N’ coordinate system. Land use and soil types were linked through attributes databases and lookup tables. The sub-catchment layer was combined with the HRU layer by reclassifying slope classes, land cover, and soil types. The SWAT model was supplied with threshold values for soil, land use, and slope class percentages, which were then utilized to establish the HRUs. For the land use/soil/slope dataset in this study, a threshold percentage of 20%/20%/10% was applied, resulting in a total of 776 HRUs.

The calibration of the SWAT model is a complex process involving two methods that may be used: an auto-calibration approach and a manual (trial-and-error) way. However, before starting the calibration, it is crucial to conduct a parameter sensitivity and identifiability analysis. This step is important because the sensitivity of model parameters can vary across different catchments. Understanding which parameters significantly impact the model's simulation results helps in focusing the calibration efforts more effectively (Cibin et al. 2010). For this study, the SWAT-CUP tool was utilized. SWAT-CUP assists in parameter sensitivity analysis, model calibration, and validation by employing various optimization techniques. It also facilitates uncertainty analysis, which is essential for understanding the reliability of the model's predictions (Abbaspour 2015). Among the techniques available in SWAT-CUP, the SUFI-2 approach was chosen. SUFI-2 is preferred due to its extensive use in analyzing parameter sensitivities and its capability to address multiple sources of uncertainty, including those related to parameters, input data, and model structure (Yang et al. 2008; Wu & Chen 2015). The uncertainty is measured based on the 95% prediction uncertainty (95PPU) band, using two key statistics: the p-factor and the r-factor. The p-factor represents the percentage of observations covered by the 95PPU band, while the r-factor measures the mean thickness of the 95PPU band divided by the standard deviation of the data. According to Abbaspour et al. (2007; 2015), the simulation results are often regarded as having an identical match with the actual data when the values of the p-factor and r-factor are both equal to 1. However, achieving high p-factor might increase the value of the r-factor; it is therefore preferable to attain a balance between the two. Information on SUFI-2 and other methods may be found in Abbaspour's paper (2015).

The SWAT model was calibrated for streamflow from 1985 to 2003 on a monthly basis based on data availability. This calibration process included a four-year warm-up period (1985–1988) to allow the model parameters to stabilize. The length of this warm-up period is determined by the characteristics of the watershed and the available data. After calibration, the model was validated using data from 2004 to 2017. To accurately simulate streamflow, the model used a water balance approach for the entire catchment area (sub-catchment 126), restoring data to its original state without considering the impact of the construction of the dam on the streamflow dynamics. Streamflow data from an additional upstream station (Kotni, Simga, Rajim, Jondhra, Seorinarayan, Basantpur, Bamnidih, and Kurubhata) was also utilized to calibrate and test the model and evaluate its predictive accuracy. The Nash–Sutcliffe Efficiency (NSE) statistic was selected as the goal or maximum objective function for assessing the model's performance because it effectively measures how well the model fits observed data (Nash & Sutcliffe 1970). Initial parameter ranges were established based on previous studies of similar basins and SWAT guidelines, which are informed by knowledge of the specific watershed hydrological cycle (Me et al. 2015; Liu et al. 2016). Throughout the calibration process, parameters were adjusted, and new ranges were proposed based on model evaluations. It should be noted that some of the recommended ranges were manually adjusted to ensure they did not exceed any range that fell outside physically relevant ranges and remained realistic (Stehr et al. 2010).

To effectively evaluate model performance, both graphical and statistical approaches should be utilized (Nyeko 2015; Sahu et al. 2022c; Turkane et al. 2024). In this study, a visual comparison of the discrepancies between simulated and observed hydrographs was conducted, as relying solely on performance indicators can be misleading and may result in unrealistic simulations (Daggupati et al. 2015; Sahu et al. 2023b). Using graphical techniques, such streamflow hydrographs, it is possible to compare simulated and observed information directly and visually, as well as identify patterns in the time and amplitude fluctuations of the flow. Apart from the graphical method, statistical measures such as the ‘percentage bias (PBIAS), RMSE (root mean square error) to observations standard deviation ratio (RSR), and Nash–Sutcliffe efficiency coefficient (NSE)’ were employed to assess the consistency between simulated and observed data.

In this study, 14 streamflow parameters, including curve number (CN2), soil characteristics, and topography, were analyzed for monthly sensitivity across nine catchment outflows. The Latin hypercube sampling method was employed to evaluate parameter variability across different HRUs and sub-catchments. Using the SWAT-CUP program, a t-test based on t-statistics and p-values identified sensitive parameters, with smaller p-values and larger t-stat values indicating higher sensitivity. Table 2 highlights these sensitive parameters, which are italicized, bolded, and ranked from 1 to 14. The results reveal that the goal or maximize objective function (NSE) significantly influences parameters associated with baseflow (ALPHA_BNK), surface runoff (CN2), hydraulic conductivity (CH_K2), groundwater dynamics (GW_DELAY, GW_REVAP), slope length (SLSUBBSN), evapotranspiration (ET) (ESCO), and soil properties (SOL_K, SOL_BD, SOL_AWC). These findings underscore the critical role of both subsurface and overland processes in streamflow simulation and was in line with Singh & Sarvanan 2020's work who studied IB river basin and identified ALPHA_BF, CN2, CH_N2, CH_K2 and RCHRG_DP, the most sensitive parameters. The final ranges of the 10 key parameters were effectively calibrated and validated for use across eight outlets on a monthly basis. Initial parameter ranges for streamflow prediction are chosen based on a combination of hydrological theory, local knowledge, and iterative refinement. This systematic approach ensures an efficient calibration process that produces physically realistic and reliable predictions.

Numerous research works have shown that employing a monthly step instead of a daily one improves SWAT model outcomes in general (Shen et al. 2010; Uniyal et al. 2015; Jang et al. 2018; Sahu et al. 2021a). Therefore, in this study, monthly streamflow data from January 1985 to December 2017 was utilized to validate and calibrate the SWAT model. The SUFI-2 technique was used for calibration after carefully selecting the model's input parameters and their corresponding uncertainty ranges (Cao et al. 2018; Gupta et al. 1999). Using a threshold value of 0.5, the NSE served as the goal function. Instead of precise values, parameter ranges were used to identify the parameters, allowing the calibrated model to be evaluated using the same parameter ranges. Moreover, due to the varied and dispersed characteristics of the catchment, calibration at a single site was inadequate for the broad catchment area. Consequently, for the distributed hydrological model, multisite calibrations yielded better results than single-site calibrations.

The performance of the SWAT model was assessed by comparing monthly statistical indicators – NSE, root mean square error (RSR), and percent bias (PBIAS) – between observed and simulated streamflow data at three sites (Jondhra, Seorinarayan, and Basantpur) during both calibration and validation periods (Tables 3–5).

• Calibration period: NSE = 0.90, RSR = 0.32, PBIAS = –11.4

• Validation period: NSE = 0.86, RSR = 0.38, PBIAS = –5.7

• Observations: The streamflow was often overestimated during calibration (PBIAS = –11.4) compared to the validation period (PBIAS = −5.7).

• Calibration period: NSE = 0.76, RSR = 0.49, PBIAS = 17.2

• Validation period: NSE = 0.90, RSR = 0.32, PBIAS = 1.4

• Observations: streamflow was underestimated more during the calibration period (PBIAS = 17.2) than during the validation period (PBIAS = 1.4).

• Calibration period: NSE = 0.93, RSR = 0.27, PBIAS = 4.4

• Validation period: NSE = 0.92, RSR = 0.27, PBIAS = –2.6

• Observations: The flow was underestimated during calibration (PBIAS = 4.4) and overestimated during the validation period (PBIAS = −2.6).

These results align within the satisfactory range established by Moriasi et al. (2007), confirming the effectiveness of the model simulations for both periods. The analysis highlights some overestimation and underestimation trends in specific periods and locations, suggesting areas for further refinement. Nonetheless, for the Jondhra gauging station, predicted stream flows during calibration (PBIAS of −11.4) were often more overestimated than during validation periods (PBIAS of −5.7). When the Jondhra streamflow combined with the Mahanadi streamflow at Seorinarayan, the model drastically changed to underestimation, with an efficiency of 0.76. The simulated streamflow is underestimated more during the calibration period (PBIAS of 17.2) than it is during the validation period (PBIAS of 1.4). Similarly, it was discovered that the flow at Basantpur that merged Bamnidih streamflow with the flow from Seorinarayan was overestimated during the validation period (PBIAS of −2.6) and underestimated during calibration (PBIAS of 4.4). The result appears significant compared to Singh & Saravanan 2020’s work who found (PBIAS, −19.1 and − 24.8), respectively, for Cal/Val. Consequently, the study concludes that reliable results for the examined watershed can be achieved by integrating data from CMIP6 GCMs and regular gauging stations into the SWAT model.

During the calibration period, the p-factor improved significantly as the flow moved through the critical intersections: Jondhra, Seorinarayan, and Basantpur, with values of 0.91, 0.89, and 0.97, respectively. This suggests that the 95PPU band bracketed 91–97% of the recorded streamflow data. Additionally, according to Tables 3–5, the average r-factor was 0.84 during validation. The 95PPU band bracketed 65% (p-factor = 0.65) of the observed data for the validation period, and the r-factor was 0.91. It saw a considerable improvement when it approached Basantpur, where the r-factor was 0.97 and 98% of the observations was captured (p-factor = 0.98). According to previous studies, a p-factor of >0.75 and an r-factor of <1.5 indicate acceptable uncertainty in streamflow simulations (Schuol et al. 2008; Abbaspour et al. 2015). Thus, the SUFI-2 method appears to capture observed streamflow adequately during both calibration and validation periods. Nevertheless, a close look at the uncertainty analysis revealed that peak flows during the 80th, 92nd, and 117th weeks from 1989 were not bracketed under the 95PPU range (Figure 3, Basantpur). This might be due to uncertainties and limits in the SWAT model. In a similar vein, Basantpur's peak flow during the 117th week from 1989 was improved and bracketed under the 95PPU band.

The study underscores the importance of parameters related to surface runoff formation and groundwater recharge in the study catchment. The parameters, those that deals with groundwater recharging and the formation of surface runoff are key hydrological factors. Furthermore, an analysis of the correlations between these parameters found them to be weak (correlation values below 0.001), indicating that these correlations are negligible and do not significantly influence the model outcomes.

When using the SWAT model for watershed studies, water balance serves as the foundation and primary driving force. Ensuring that simulation findings are close to measurements is the aim of calibration. In order to identify the water balance components for the outlet sites, the SWAT model was rerun at the final catchment outlet (the recently installed Kalma gauging site) using the optimal parameters (obtained using spatial proximity-based regionalization) for the entire period (1985–2017). The mean annual contributions of the water balance are listed in Table 6.

According to the data, from 2019 to 2050, the mean annual precipitation was recorded at 1250.9 mm, which included the same amount of rainfall and 192.57 mm of surface runoff. ET accounted for 35.35% of this total precipitation, amounting to 474.6 mm, making it the primary form of water loss. This high ET rate is attributed to the fact that 96.01% of the land use in the area consists of agriculture, forestry, and other green cover (Figure 2(a)). In addition, the increased ET was indicative of the mean temperature of 31 °C. The total water yield at the catchment outlet, which includes lateral soil flow, shallow groundwater (baseflow), and overland flow, was calculated at 800.52 mm. After subtracting deep groundwater recharge of 23.7 mm (based on the SWAT model's assumption that deep aquifer recharge exits the catchment (Arnold et al. 1993), the effective water yield was 776.82 mm. The contributions to this water yield were 59.29% from baseflow (464.6 mm), 13.76% from lateral subsurface runoff (110.15 mm), and 24.06% from overland flow (192.57 mm). The proportion of baseflow contribution in this catchment is relatively high compared to other watersheds with a similar percentage of forest cover. The catchment's soil characteristics reveal a dominance of Type C soil (96.01%), which is characterized by low infiltration, clay loam texture, and a Curve Number range of 66–79 for natural vegetation, while the remaining 4.00% comprises Type B soil, known for its moderate infiltration, loamy texture, and a Curve Number range of 46–65 for natural vegetation. According to Neitsch et al. (2011), this suggests a lower infiltration rate and greater surface runoff being generated. Additionally, the average gradient is 1 in 590, indicating that perched water flows out and lateral subsurface runoff is easily generated. A baseflow separation approach, sourced from the SWAT website (https://swat.tamu.edu/software/), indicated that the baseflow to total runoff ratio is nearly 70%. Therefore, the simulation results are considered reasonable and reflective of the catchment's hydrological processes.

It's important to remember that annual fluctuations in precipitation influence water yield components, such as baseflow, lateral flow, and overland flow. On the other hand, the variance in the overland runoff contribution contrasts with the precipitation contribution in 2010 (Figure 3). Further analysis indicates that this was due to the timing of internal precipitation distribution. Due to heavy rains at the beginning of 2010, the initial soil water content was higher. Thus, more surface runoff was produced as long as the precipitation continued. However, there was a 22% ET discrepancy between the minimum ET recorded at the Rajim gauging site (Mahanadi basin) and the Mand river basin (Kurubhata gauging site).

The study analyzed monthly water balance components, revealing that ET, a key factor in water loss, varies seasonally – decreasing in winter and increasing in summer. Variations in rainfall are correlated with seasonal changes in groundwater, lateral subsurface runoff, and overland flow. The findings show that 85.18% of the total annual precipitation occurs between June and September (4 months), contributing to 68.10% water yield of the annual runoff volume. Additionally, there were notable differences in the contributions of hydrological components between wet and dry seasons. On a monthly time step model, the proportion of overland runoff to precipitation varied between 30 and 65%, and an obvious distinction was observed between the wet and dry seasons. It is a fact of nature that groundwater moves more slowly than precipitation.

The analysis of the relationships between individual water balance components and monthly precipitation is summarized in Table 7. Furthermore, N-W, S-W, and some regions of S-E of the study region covering the range of Maikal Hill lying in the minimum average precipitation (Figures 1 and 4) were consistent with the mean ET/precipitation ratio, i.e., 0.41 and supported by with 0.40. Similarly, the N and N-E (Bamnidih, Kurubhata, and Kalma) regions defined as Central Tablelands were heterogeneous in their mean ET/precipitation ratio (0.36) along with the lowest surface runoff/total flow (0.25). This region regulates the water balance ratios: percolation/precipitation and baseflow/total flow, with mean magnitudes of 0.41 and 0.75, respectively. This is because the main soil type is deltaic soil, mixed red and black soil, red and yellow soil, and laterite soil. The results indicate that land use, land cover, and soil type have a significant impact on surface runoff and lateral subsurface flow. The above statistics (Table 7) indicate a significant correlation and imply moderate precipitation, which quickly increases overland flow and interflow in the N-W and S-W regions, thus causing flooding. Moreover, intense precipitation over the N and N-E regions was significant to percolation/precipitation and baseflow/total flow, which ultimately contribute to groundwater.

The CMIP6 GCM data and the IMD Grid data from nine gauging stations were the two sources of meteorological data used in this work to drive the SWAT model in the upper and middle reaches of the Mahanadi River Catchment. Moreover, the model underwent simultaneous validation and calibration at the following three different locations: Jondhra, Seorinarayan, and Basantpur. Comparing the constructed SWAT model with traditional gauging stations as input data, the findings demonstrate that the model accurately replicates monthly streamflow. The following are the primary findings:

• Based on a sensitivity analysis of the monthly streamflow simulation, the most sensitive parameters were hydraulic conductivity (CH_K2) and baseflow (ALPHA_BNK), which together covered 89% (8/9) of the gauging sites out of the 14 parameters. Similarly, surface runoff generation (CN2) with 78% (7/9) and soil characteristics (SOL_AWC) with 45% (4/9) were the second and third most sensitive parameters, respectively. Last but not the least, ET (ESCO), which covered 33% (3/9), was sensitive to a few gauging sites. The correlation between these parameters was high. The parameters exhibit minimal uncertainty, as evidenced by the appropriate p- and r-factor values for both the calibration and validation periods. In addition, the model demonstrated good performance in the graphical and statistical indicators when streamflow was simulated on a monthly scale.

• The mean annual actual ET was 474.6 mm, or almost 38% of the mean annual precipitation (1,250.9 mm), according to the model output files at the final outflow (Kalma). Moreover, the catchment's whole outflow had a mean annual streamflow of 776.82 mm. Additionally, groundwater baseflow contributed the remaining 38.97% of runoff, with surface runoff accounting for roughly 24.8% and lateral subsurface flow accounting for approximately 14.18%. This breakdown of runoff is satisfactory when compared to the 30% result of the baseflow separation method.

• The spatial distributions of overland flow and lateral subsurface flow were consistent with the distribution of precipitation. The highest ET was found in areas with significant agricultural land cover and larger river regions. The ET levels remained constant from 1985 to 2017. Regression analysis revealed that surface runoff, lateral flow, and ET were highly correlated with precipitation, with groundwater being the exception. The results from this water balance study are expected to be useful for protecting the water environment and ensuring the sustainable management of the Mahanadi River Catchment.

• The study estimates that 23.7 mm of annual precipitation contributes to deep aquifer recharge, which appears reasonable but slightly conservative given the geological and hydrological conditions of the Mahanadi Basin. While hard rock formations limit recharge, deltaic regions with higher infiltration potential and substantial baseflow movement support the estimate. Comparisons with other studies (CGWB 2019; Nayak et al. 2023; Sahu et al. 2025) indicate that this value falls within the lower range of reported deep recharge rates ( ∼ 25–40 mm/year) but remains consistent with observed regional patterns.

For effective water management, prioritize the parameters most sensitive in streamflow modeling. Monitor these parameters closely due to their high correlation and minimal uncertainty. Although ET (ESCO) was less sensitive across sites, it remains important in areas where it plays a role. Manage water balance components, noting that actual ET constitutes 38% of annual precipitation, with surface runoff and groundwater baseflow contributing significantly. The runoff breakdown aligns well with expectations, emphasizing the need to balance groundwater recharge and surface runoff management. Utilize spatial analysis to focus on areas with high agricultural cover and large rivers for targeted interventions. This approach will aid in sustainable management and protection of the Mahanadi River Catchment.

Practical recommendations for Water Resources Management which include, optimize reservoir operations (i.e., dynamic reservoir and flow management); increase managed aquifer recharge (i.e., groundwater recharge and sustainable extraction); promote water-efficient irrigation techniques (i.e., climate adaptive irrigation practice); floodplain zoning regulations (i.e., flood and draught preparedness); promote rainwater harvesting (i.e., soil and watershed conservation); last but not the least expansion of gauging station (i.e., enhancing monitoring and data collection).

Future scope/Limitations: The model evaluation can be better refined if the additional procedure for optimizing the selection of CMIP6 emission scenarios is adopted. The scope can further be extended by refining the resolution of weather data (solar radiation, wind speed, relative humidity) similar to precipitation data. For hydrological simulations, the study only uses the SWAT model; however, using several models (such as VIC, HEC-HMS, or MIKE SHE) might enhance dependability and offer comparative insights. The study primarily focuses on hydrological components and does not integrate socioeconomic factors, water governance policies, or stakeholder-driven decision-making processes, which are crucial for practical water resource management.

The precipitation data used in this study were provided by the Indian Meteorological Department (IMD), Pune, and is highly appreciated. Suggestions and comments from reviewers are greatly acknowledged.

R.T.S., N.K., and C.R.G.R. conceptualized the study; N.K., R.T.S., and A.P. supervised and reviewed the study; C.R.G.R., N.K., and A.P. prepared the methodology; N.K. wrote the original draft; R.T.S. and N.K. wrote and edited the article.

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

The authors declare there is no conflict.