Introduction

Process-based numerical hydrologic models are rapidly evolving to help understand and predict potential impacts of population growth, land use, and climate change on water resources (Athira et al. 2021; Koycegiz and Buyukyildiz 2019; Paniconi and Putti 2015; Arnold et al. 2012). These hydrologic models can be used to develop and evaluate management practices to mitigate the effects of climate change and select the best set of policies for implementation. Hydrologic models are also required to predict extreme events such as floods and droughts that are common in many parts of the world including India. However, the reliability and robustness of hydrological models depend on the availability of appropriate input data and proper model calibration (Chen et al. 2023; Bucak et al. 2017)). In many parts of the world, lack of appropriate input data is a major constraint for hydrologic modeling (Abbaspour et al. 2019). Scarcity of long-term daily climatic and hydrologic data is a challenge for the application of advanced hydrological modeling techniques in water resources management (Alipour and Kibler 2019). Furthermore, heterogeneity in the climatic and hydrologic data restricts the applicability of scientific tools to monitor changes relevant to the hydrologic cycle (WMO 2009). Therefore, reasonably calibrated hydrologic models are still very important tools to predict streamflow for ungauged catchments and to fill data gaps in streamflow records (Alipour and Kibler 2019; WMO 2009). The other challenge of hydrologic modeling is calibration of model parameters under a specific assumption and data scarce conditions (Nguyen et al. 2022; Abbaspour et al. 2019). This indicates that getting unconditional model parameters independent of calibration procedures and objective functions is difficult.

Many model calibration tools and guidelines have been developed for the hydrologic model SWAT (Soil and Water Assessment Tool). Nguyen et al. (2022) introduced an open-source R-based tool called R-SWAT for parameter calibration, sensitivity analysis, uncertainty analysis, and visualization of SWAT outputs. Likewise, Pfannerstill et al. (2014a, b) developed R-scripts for SWAT model parameter calibration, validation, and sensitivity analyses. Their scripts were designed to generate a set of parameter values based on Latin Hypercube Sampling from the R-package FME (Soetaert and Petzoldt 2010) and for each parameter set a model run is performed. In addition, Abbaspour (2005) introduced SWAT–CUP which aims to perform parameter calibration, sensitivity analysis, and uncertainty analysis with the SWAT model. However, SWAT–CUP is proprietary software, and freely available versions have some user restrictions, for example, parallel simulation options are restricted (Nguyen et al. 2022).

Though there are multiple options to calibrate hydrologic models, limiting model uncertainty is still a challenging task. In most cases model outputs for streamflow are compared to corresponding measured data with the assumption that all error variance is contained within the predicted values and that observed values are error free (Moriasi et al. 2007). However, observed data are never error free (Abbaspour et al. 2017). Another common assumption during hydrologic model calibration is the stationarity of processes, i.e., the means and variances of model input data and physical processes do not vary over time. Calibration, therefore, leads to a single set of calibrated model parameters. However, physical processes are dynamic in space and time, so the prediction of outputs with a single parameter set is questionable. The non-stationary behavior of hydrologic processes is often neglected which may result in uncertainties in model predictions based on a single set of parameters derived under the assumption of stationary condition (Mondal et al. 2016). According to Abbaspour et al. (2017), no calibration can lead to a single set of model parameters, because hydrological processes are modeled by a range of mathematical equations. A single set of model parameters could be achieved only for a given set of conditions depending on the calibration procedures and objective functions. Another source of model uncertainty could be inappropriate representation of hydrological model parameters. A single model parameter may influence one or several hydrological processes (Guse et al. 2020a; b), and model outputs could be associated with uncertainty if model parameters are not identified and adopted properly (Pluntke et al. 2014). Finally, hydrologic models that are evaluated against an aggregate measure of a single statistical value are problematic as it is difficult to identify why measured and modeled output variables differ. The use of multi-criteria performance metrics (e.g., with flow duration curve) can improve model evaluation.

A recent study by Lakshmi and Sudheer (2021) attempted to apply a new way of calibration for Cedar Creek and Riesel Y2 watersheds of the USA to address limitations of seasonal calibration based on the SWAT model. The authors partitioned the simulation period into different clusters and performed calibrations for each cluster group. Here, parameters sets were identified for each independent cluster group and the results revealed improved streamflow estimation. However, Lakshmi and Sudheer (2021) applied seasonal clustering on a single simulation period which might not sufficiently depict the effects of multi-year variability of hydrological processes on model parameter optimization.

Therefore, a calibration approach that leverages multiple objective functions, multiple variables, multiple parameter sets, multiple evaluation techniques, using data that includes wet, average, and dry years and many model simulation periods may reduce uncertainty and improve hydrologic model accuracy (Guse et al. 2020a, b; Lu et al. 2015; Ewen 2011; Moriasi et al. 2007). In this study, we intended to calibrate and validate the hydrological model SWAT for the Adyar catchment, India, using daily streamflow data by considering seasonal differences. Due to data scarcity and distinct seasonality the catchment is a challenging case for model calibration. Accordingly, the current study introduces new hydrologic model calibration procedures that address the effect of temporal variability (non-stationarity of hydrologic processes) and seasonal differences on model parameter selection. Specific objectives include: (i) to develop a more robust calibration procedure using different calibration periods and (ii) to investigate if distinct seasons lead to different model performances and model parameter set.

Materials and methods

Study area

The Adyar catchment is part of the Chennai basin and is located in the south of Chennai city, India (Fig. 1). Adyar river enters the Chennai District at Nandambakkam after flowing approximately 9 km through the metropolitan area in a total of its 15 km stretch in the suburbs of Chennai (Kumar et al. 2019). The Adyar river flows along the southern part of Chennai city and enters the Bay of Bengal (Fig. 1). It originates from the three tank groups, Kallampadu, Pillapakkam, and Kavanur. There are three major tanks called Chembarambakkam, Sriperumpudur, and Pillapakkam in the Adyar catchment. The total catchment area is about 828.8 km2 and its land use is characterized by rapid urbanization (Nithila Devi et al. 2020). The upper Adyar catchment comprises about 75% of the catchment and is important with regard to flood control. Chembarambakkam tank has an annual average discharge volume of 89.4 million cubic meters at Kattipara (Nithila 2020), which is roughly 40% of streamflow in the catchment. The Adyar river is seasonal and reaches its peak streamflow for a few days during the northeast monsoon rainy season. Due to urbanization the Adyar River is heavily polluted.

Fig. 1
figure 1

Location map of Adyar catchment within Chennai Basin, Tamil Nadu, India, topography, river network, boundary of modelling catchment and weather stations

Full size image

The Adyar catchment has a tropical wet and dry climate feature (Kumar et al. 2019). The catchment receives rain from South–West and North–East monsoons (Sudheer et al. 2016). The October, November, and December months contribute most of the North–East monsoon rainfall, and the topographic depression of the Bay of Bengal causes rainfall to occur in the form of one or two annual cyclones (Kiram and Kumar 2017). Based on rainfall data analyses from three rain gauge stations (Korattur Anicut, Sriperumbudur, and Chembarambakkam) between the years 2000 and 2019, about 55% of annual average rainfall in the Adyar basin upstream of Chembarambakkam occurred during the North–East monsoon, which lasts from October to December. The annual rainfall for the North–East monsoon over the period 2000–2019 varies from 238 mm (2003) to 1895 mm (2015). During the same period, the annual rainfall for the winter and west monsoon seasons together (between January to September) varies from 281 to 921 mm. The long-term annual average rainfall of the modelling catchment (part of Adyar catchment upstream of Chembrabakkam) over the years is 1260 mm. In addition to rainfall variability, the mean maximum and minimum temperature differs from month to month. As evident from the average monthly maximum and minimum temperature, May was the warmest month over the period between 2000 and 2019. Its average maximum and minimum temperature values were 38.6 °C and 27.2 °C, respectively. The lowest average maximum temperature value occurred in December with a value of 29.4 °C and the lowest average minimum temperature occurred in January with a value of 19.8 °C. The annual average maximum and minimum temperatures values were 33.9 °C and 23.8 °C, respectively. Factors that control seasonal patterns of rainfall and temperature include the Madden–Julian Oscillation (MJO) and the inter-tropical convergence zone (Dey et al. 2022).

Database

Daily measured rainfall data at three weather stations (Korattur Anicut, Sriperumbudur, and Chembarambakkam) and gridded temperature (maximum and minimum) data with spatial resolutions of 0.5° by 0.5° from 2000 to 2019 were available from the India Meteorology Department (IMD 2019). Furthermore, from 2004 to 2020, daily streamflow of the Adyar River upstream of the Chembarambakkam reservoir were provided by the Tamil Nadu State Water Resources Department (TWRD 2020) for calibration and validation of the SWAT model simulation. Spatial data including soil, land use land cover maps, and topographic information (digital elevation model) were used to setup the hydrologic model (Fig. 2, Table 1). The shuttle radar topography mission (SRTM) 1 arc-second global digital elevation model (DEM) data with 30 m by 30 m resolution was available from the US Geological Survey website (USGS 2018). A soil map of the scale of 1:50,000 from Tamil Nadu Agricultural University (TNAU 2018) and land use classification for the year 2016 derived from Sentinel-1 and Sentinel-2 data (Fig. 2) (Steinhausen et al. 2018) were used. In general, the spatial, weather, and hydrologic data were used to setup and calibrate the SWAT model. All sources and their basic features are presented in Table 1.

Fig. 2
figure 2

Model input data that includes Slope (A), modelled watershed boundary, and river networks (B), Soil types (C) and land use-cover classes (D) after (Steinhausen et al. 2018) maps of the Adyar catchment

Full size image
Table 1 Spatial and hydrometeorological data sources and availability used this study
Full size table

SWAT model setup

The semi-distributed, time continuous eco-hydrologic model SWAT (Soil and Water Assessment Tool) was used for this study (Arnold et al. 1998; Arnold and Fohrer 2005). The SWAT model is one of the most popular process-based models widely used to simulate catchment hydrological processes globally including the Indian Subcontinent (e.g., Dash et al. (2023); Singh and Saravanan (2020); Wagner et al. (2019); Demirel et al. (2009)). SWAT can provide spatially distributed outputs for water balance components based on hydrological response units (HRUs). For this study, one SWAT model was setup for the Adyar catchment in which Chembarambakkam reservoir is located (Fig. 1). ArcSWAT 2012 was used to delineate the watershed and compile SWAT input files. SWAT2012 version 664 was used to run the simulation. The Food and Agricultural Organization (FAO) guideline was followed in classifying slope classes as 0–2% (foot slope), 2–5% (gentle sloping), 5–8% (sloping), 8–15% (strongly sloping), and > 15% (moderately steep to very steep) (Jahn et al. 2006). The catchment was divided into eleven sub-basins. Then, we used spatially distributed land use (thirteen different classes, Fig. 2D), ten soil (Fig. 2C), and five slope classes (Fig. 2A) to generate hydrologic response units (HRUs): each unique combination of land use, soil type, and slope class defined an HRU within every subbasin (Neitsch et al. 2011). All possible combinations of land use, soil and slope layers were used to setup the HRUs without applying any thresholds to derive a model setup with high spatial precision (Her et al. 2015). This approach resulted in 3066 HRUs.

Calibration and performance matrices

Three years (from 2001 to 2003) were considered a warm-up period to define appropriate initial conditions and to maintain equilibrium conditions for the model (Noori and Kalin 2016). Daily measured inflow discharge to Chembarambakkam Lake were available for the period from 2004 to 2020, and these data were applied using different calibration approaches to yield a reasonable hydrologic representation of the catchment. According to Moriasi et al. (2007), ultimate model calibration should be done with sufficient data length to include wet and dry years and encompass a range of hydrologic events. Thus, model calibration was attempted for different combinations of measured inflow discharge to Chembarambakkam:

  1. 1.

    All inflow discharges considered for daily time scale

  2. 2.

    Extreme inflow values excluded for daily time scale

  3. 3.

    Daily time scale inflows only for wet season (Oct to Dec)

  4. 4.

    Monthly inflows (wet season)

  5. 5.

    Only wet season of wet years (annual rainfall > long-term mean) for daily time scale

  6. 6.

    Dry years of wet season (annual rainfall < long-term mean) for daily time scale

Because the land use/cover input data used to setup the SWAT model was from 2015, the calibration period included years between 2015 and 2019 and validation was carried out for the period 2004–2008. Prior to calibration, key SWAT model input parameters that cause changes on the hydrograph of inflow to the reservoir were identified through parametric sensitivity runs. Here, we applied changes on the value of a parameter, ran the model, and evaluated the responses of the streamflow under different values. Based on this, ten parameters were considered for model calibration. Five thousand different parameter combinations were sampled using a Latin Hypercube sampling approach in R environment using package FME (Soetaert and Petzoldt 2010). Then, the default SWAT model input parameter values were altered based on values sampled using Latin Hypercube sampling for each calibration run inside the R environment (R Core Team 2013). Model performances for all 5000 model iterations were evaluated based on measured streamflow using the hydroGOF package in R (Zambrano-Bigiarini 2017). The numerical goodness-of-fit of the simulated versus measured streamflow were evaluated in different segments of the flow duration curve based on Pfannerstill et al. (2014a, b); and Haas et al. (2017). Kling Gupta Efficiency Coefficient (KGE) was the objective function used to identify the parameter combination that yielded the best fit between simulated and observed streamflow data. Furthermore, the values of Nash–Sutcliffe Efficiency (NSE), Percent Bias (PBIAS), and Standardized Root Mean Square Error (RSR) were also considered in evaluating the performance of the model. KGE was used as the main objective function instead of NSE for model calibration as it improves the bias, the over/under estimation of simulations, and the variability measures significantly (Gupta et al. 2009). For this reason, daily and monthly model performances were evaluated based on the optimum KGE values.

In most hydrologic models, selecting optimum statistical values for the given objective functions is based on the general assumption of time-invariant model parameters and catchment characteristics (Abebe et al. 2010). However, catchment characteristics and model parameters are time-variant, and uncertainty associated with model parameters are assumed to vary with time. Thus, model validation, land use change, and climate change impact studies using parameters that were determined during the calibration process may not be enough to produce reliable predictions. Antecedent and climatic conditions of the basin cause variability of hydrologic processes through time, and precipitation is one of the primary hydrologic model inputs that influences the parameter optimization processes (Lakshmi and Sudheer 2021). Moriasi et al. (2007) also reported that data used to simulate hydrologic processes affect calibration and validation of hydrologic models. Therefore, we applied successive calibration using different time periods (1-year time shift) of simulated and observed time series discharge to see how model parameters vary for different time periods. Moreover, the best calibrated model was also validated for different time periods by moving the validation period 1 year ahead to see the changes in model efficiency. Figure 3 summarizes all step-by-step modelling steps.

Fig. 3
figure 3

Flowchart summarizing the methodological approach

Full size image

Results

Model parameters and performance for different calibration data sets

Due to seasonal variability and the influence of peak flows, the daily base calibration resulted in relatively poor performance (KGE = 0.56, NSE = 0.19, and PBIAS = 14.5%) compared to the wet monsoon and monthly values (Table 2, Figs. 4, 5). Particularly, the NSE value for the daily base calibration indicates a weak performance, but the positive NSE value indicates the model is at least better than using the mean value of all measurements (Moriasi et al. 2007). Similarly, KGE and PBIAS are in the acceptable range.

Table 2 SWAT model performances when calibration was carried out by comparing observed against simulated daily streamflow (extreme values included, extreme values excluded, monsoon season records, wet years, dry years, and monthly time steps are considered). The NSE and PBIAS values are acceptable based on Moriasi et al. (2007) except for the NSE value of Scen-6
Full size table
Fig. 4
figure 4

Comparison of observed and simulated streamflow based on calibration carried out considering all records for the period (2004–2019) on daily time scale with reference to rainfall

Full size image
Fig. 5
figure 5

Flow duration curves and hydrographs of simulated and observed monsoon season daily streamflow (monsoon season wet) values for calibration and validation periods of the Adyar catchment

Full size image

The model performance for the daily streamflow was improved when the calibration was carried out for North–East monsoon season (October to December). Both the (KGE = 0.67) and (NSE = 0.58) values were higher as compared to values from the daily base calibration. Likewise, the PBIAS and RSR values were improved for the seasonal calibration. In addition to statistical indices, the good fit between observed and simulated streamflow was also evaluated graphically based on Pfannerstill et al. (2014a); Pfannerstill et al. (2014b) and Haas et al. (2017) using the flow duration curves and hydrographs (Fig. 5).

Although modeled and observed streamflow matched well for the wet monsoon season during the calibration and validation periods, model efficiency tended to decrease for dry seasons and years (Table 2). Therefore, the performance of the SWAT model was separately evaluated for dry and wet years. Dry years are those years that received a total annual rainfall less than the long-term mean annual rainfall, and in wet years the total annual rainfall is greater than the long-term mean (Fig. 6). The performance of the model during dry years was worse than during wet years, NSE (dry years = − 0.59, wet years = 0.58). Figure 7 shows portions of model performances from the 5000 iterations for the 6 scenarios mentioned in Table 2. Simulations of streamflow for the north–east monsoon season, wet years, and monthly time steps better reproduced observed streamflow for both calibration and validation periods. It is not surprising that streamflow in dry years and dry seasons is more difficult to model due to water withdrawal and insufficient observations of streamflow and the data are in general relatively hard to handle. Aggregation of the data to monthly values resulted in a better performance, although there were differences for some months (Table 2).

Fig. 6
figure 6

Annual rainfall anomaly (annual rainfall–long-term mean) for the Adyar catchment for the period 2000–2019. The years showing positive anomalies are defined as wet years and the years with negative anomalies are dry years

Full size image
Fig. 7
figure 7

Model performance tendencies of different scenarios when model parameters were optimized based on Scen-1 (entire daily raw streamflow records considered), Scen-2 (all-time streamflow records after the outliers were excluded), Scen-3 (wet season records (October to December)), Scen-1(monthly calibration), Scen-5 (only wet years daily records) and Scen-6 (only dry years daily records). The red dots for each scenario represent the best calibrated values of objective functions (refer Table 2)

Full size image

Temporal variability of the SWAT model performance

One of the objectives for this study was to investigate the variability of hydrologic model parameters and the efficiency of the SWAT model under time-variant conditions. Model parameter selection and performance indices differ when different time periods were used for streamflow calibration. When the SWAT model was calibrated for the period 2004–2009, the simulated streamflow was overestimated by 8.2% compared to observed data, whereas the overestimations were 12.10%, 18.3%, and 20.0%, for the periods 2004–2010, 2004–2011, and 2004–2012, respectively. However, fitted model parameters for all time periods were the same. Likewise, the calibration periods of 2004–2013 and 2004–2014 had the same model parameters and simulated streamflow was overestimated by 20.7% and 22.1%, respectively. These results indicate that model uncertainties such as overestimation/underestimation could be attributed to the temporal dynamics of model input variables. We also found that the length of calibration period influenced model efficiency. The efficiency of the model was improved when the calibration period was increased from 6 to 8 years. On average, the KGE value was 0.27 for 6-year calibration period, whereas it was 0.4 for 8-year calibration. It is worth remarking that the variability of model performance reflected the time shifts of hydrologic response variables in the SWAT model, but these changes in model performance are not statistically significant.

Conventionally, hydrologic model calibration and validation are carried out under the assumption that the hydrologic parameters and the degree of uncertainties are consistent over time. However, the results of this study indicate that the degree of uncertainty and the model parameters could vary through time due to the temporal variability of the input parameters. When calibration was carried out for different periods with a 1-year time lag, but the same length of records, the hydrologic model parameters and evaluation criteria showed differences. Efficiencies of the SWAT model were decreasing when the time lag was increasing. For example, initially the calibration was based on the period of 2015–2019, and the optimum value of the objective function KGE and NSE were 0.68 and 0.59, respectively, and the simulated streamflow was overestimated. Even though statistical indices computed for different validation periods differ, the model performances are reasonable in most cases (Table 3). When a time lag of 1 year was applied for the validation periods, there were slight differences on the KGE, NSE, PBIAS and RSR values. The overall analyses indicated that model performance for different validation periods show increasing variability when there is a larger time lag between two different periods. For example, the KGE values were 0.68 and 0.60 when we validated the model for the period 2004 to 2007 and 2009 to 2012, respectively (Table 3). From this, we conclude that model uncertainty could not be stationary over time and this non-stationarity could be related to temporal differences of observed streamflow, precipitation, and air temperature between earlier and later periods.

Table 3 Performance of a calibrated SWAT model for monsoon season (October to December) streamflow when it is validated for different periods (model parameters remain uniform)
Full size table

Variability of streamflow for different validation periods

Due to the temporal differences of the model input variables, the quality of the streamflow simulation over the study catchment showed considerable differences. These differences were reflected by the statistical indices that were calculated when the model was validated for different periods. The calibrated SWAT model well-represented the recent monsoon season streamflow from the year 2012 onwards. Although reasonable model performances were achieved for the years before 2012 to reproduce the monsoon season streamflow, in most cases, the statistical indices were relatively lower when compared to the values calculated for the periods after the year 2012 (Table 3). The lower performance indices for the years before 2012 may be associated with the land use/cover data and the time of calibration. Static land use/cover data, generated in 2015 by Steinhausen et al. 2018, was used to construct the SWAT model. As hydrologic processes are highly sensitive to land use changes, streamflow simulated under such static land use/land cover conditions can show disparities from the observed one (Wagner et al. 2019). Güse et al. (2015) also concluded that dynamic modeling leads to a more realistic temporal development of land use change and its impacts on hydrologic processes. In addition, the SWAT model was calibrated for the period 2015 to 2019, which might not be sufficient to represent the temporal hydrologic variability that caused differences in the model performance for the periods before and after 2012. Our results indicated that the monsoon season streamflow was underestimated for the period 2005 to 2011, overestimated during 2011 to 2014, and again underestimated during 2015 to 2019.

General water balance

Modelling the water balance components of a catchment is of particular importance in catchments, where measured data are scarce. This study presented simulated water balance components for the upstream part of Adyar catchment. According to the simulated outputs of a calibrated SWAT model, the dominant hydrologic process is evapotranspiration. About 61% of the annual rainfall is lost to the atmosphere as actual evapotranspiration, while the annual catchment water yield accounts for 23% the annual rainfall. Due to the high evapotranspiration, the surface runoff share is only 11% of the total rainfall. Baseflow contributes about 40% to the streamflow. Our results revealed that water losses are more pronounced during the North− East monsoon season (October to December). About 56%, 85%, and 62% of the annual rainfall, surface runoff, and water yield, respectively, occurred in the monsoon season. Thus, availability of surface water flow during the other months of the year (January to September) is very limited.

Impacts of parameter variability on simulated water balance components

Water balance components including surface runoff, groundwater contribution to the streamflow, actual evapotranspiration, and water yield differences were analyzed for different parameter sets. Regardless of the influence of time period, the choice of model parameter influences the simulation of the water balance components. All the major water balance components responded differently for different optimum parameter sets. This is important, because these components were not measured, rather we inferred them from simulations. In this regard, water balance components for the first five best runs (Table 4) from a total of 5000 were analyzed to understand changes of annual water balances with different parameter combinations. Changes in the model parameter values influence the annual water balance components. For example, reducing the CN2 value from − 12 to − 14 decreases the surface runoff and water yield by 5.6% and 11%, respectively. Likewise, a 0.2 increase on the value of EPCO causes an increase of actual evapotranspiration by 4%, and an increase of gw_revap (groundwater revap coefficient) by 0.02 resulted a reduction of groundwater recharge by 7.5%. Groundwater discharge from the shallow aquifer to the streamflow was decreased when the gw_revap values were increased, whereas the actual evapotranspiration was increased. Thus, model parameter values that were optimized during calibration were largely responsible for the differences in the simulated water balance components. In addition, the seasonal variability on climate input variables also causes differences of water balance components. These different calibrations of model parameters are worthy to propose multiple water management planning options, because most of the hydrologic processes are not often measured directly but are instead simulated as part of the water balance components. Through these simulations it becomes clear that the water balance components are varying with parameter changes, even though the objective functions are on a similar level. Figure 8 illustrates the variability of annual mean water balance components for the five best runs from 5000 iterations.

Table 4 Parameter values for the first five best simulations (SIM1, SIM2, SIM3, SIM4, and SIM5) out of 5000 simulation and the corresponding model performances
Full size table
Fig. 8
figure 8

Annual water balance variability and changes with different model parameterization as compared to the best simulation (SIM1). SIM2, SIM3, SIM4, and SIM5 represent the simulations based on the parameter combinations in Table 4. The y-axis values represent the absolute values and differences of actual evapotranspiration (actual ET), groundwater discharge to the stream flow (GWQ), surface runoff (SURQ), and water yield (WYLD) with reference to sim1 (best simulation)

Full size image

Figures 9 and 10 show the spatial variations of simulated ET and water yield under different parameters sets (Table 4). With reference to the simulation outputs under SIM1, changes in water yield showed high variability, while changes in ET were small, varying between − 8% and 10% (Fig. 10A). Changes in water yield varied between − 130% and 20%. In particular, SIM4 showed a strong deviation in water yield from the other simulations (Fig. 10B).

Fig. 9
figure 9

Spatial variability of ET and water yield simulated based on the best (SIM1) parameter set (Table 4) and rainfall

Full size image
Fig. 10
figure 10

Spatial variability of annual ET (A) and water yield (B) changes of SIM2, SIM3, SIM4, and SIM5 with reference to the best simulation (SIM1) based on the parameter sets in Table 4

Full size image

Moreover, the average annual simulated water balance components varied when model parameters were calibrated for different lengths of calibration periods. This was verified by comparing model outputs for the entire simulation period (2004–2019) after applying sequential calibrations for different periods. The differences were due to the impact of temporal variability of the model input and output variables in selecting model parameters values that resulted from the best optimum KGE. For example, there were 15% decreases in the annual average surface runoff and water yield, 13% and 122% increases on actual evapotranspiration and lateral flow, respectively, and 18% decrease on total aquifer recharge when we applied model parameter sets calibrated (1) for the 2004–2014 period and (2) for the 2004–2015 period. These results indicate that a 1-year difference of calibration period caused a large variation in the selection of best parameter combination, which resulted in considerable changes to major water balance components. Therefore, model calibration and validation periods should include as many years as possible to depict variability in the data set and model.

Discussion

The accuracy of hydrological models depends on accurate representation of model parameters and use of appropriate input data. In this study the influence of seasonality and temporal variability in hydrologic model calibration were analysed by testing different modelling periods and seasonal decomposition. Considering different modelling periods can help to understand how non-stationarity of hydrologic processes influences model calibration and parameters selection procedures, while seasonal decomposition often minimizes model uncertainties that could be masked by aggregating seasonal processes. Our results show that model parameter values and model efficiencies obtained under non-stationary conditions are considerably different compared to calibration under stationary conditions. It is worth remarking that the variability of model performance was reflected by modelled water balance components in the SWAT model, though changes in model performance are not statistically significant. These differences in the values of model parameters and model efficiency are attributed to temporal dynamics of model input variables. This implies that multiple calibration and validation options would be necessary to account for non-stationarity of hydrologic processes. Our results showed remarkable differences in model performance when a 1-year shift was applied to the calibration and validation periods. The changes of the model efficiency increased when the time shift increased between two validation periods (Table 4). Not only the model efficiency varied but also the spatial distribution of water fluxes varied (Figs. 9, 10). This indicates that the selection of calibration and validation periods influence the selection of model parameters as well as the performance and results of the model, which agrees with Myers et al. (2021).

Conventionally, hydrologic model calibration and validation are carried out under the assumption that hydrologic parameters and the degree of uncertainties are consistent over time. In this perspective, hydrologic models are calibrated and validated under fixed periods that end with a single set of parameters. However, single set of parameters could be valid to represent processes that are stationary in nature. Abbaspour et al. (2017) also argued that no calibration can lead to a single parameter set or a single model output as physical processes are modeled based on continuum equations and one can only measure a limited amount of data. Therefore, multiple sets of calibrated model parameters and evaluation are needed to simulate multiple processes (Moriasi et al. 2007). In this study, we also found that the length of calibration period influenced the model efficiency. The SWAT model efficiency was improved when the calibration period was increased from 6 to 8 years. On average, the KGE value was 0.27 for 6-year calibration period, whereas it was 0.4 for 8-year calibration.

When calibration was carried out for different periods with a 1-year time lag, but the same length of records, the hydrologic model parameters and evaluation criteria showed differences due to temporal effect. Efficiencies of the SWAT model decreased when the time lag increased. Similarly, there were slight differences in the KGE, NSE, PBIAS, and RSR values when a time lag of 1 year was applied for the validation periods (Table 4). The overall analyses indicated that the model performances for different validation periods show high variability when there is a larger time lag between two different periods.

We found improved model performances of the SWAT model simulations for the wet season (North–East monsoon) and wet years, while our model performances were poor for the wet and dry seasons and years together (Table 2). Furthermore, the SWAT model performance was pronouncedly improved when it was calibrated for wet years only. Our findings are, therefore, in agreement with Gao et al. (2018) who found that separating wet and dry years of monthly runoff strongly improved the efficiency of the SWAT model compared to their calibration results for the wet years and dry years. Similarly, Lakshmi and Sudheer (2021) applied discretization and clustering of simulation periods into different seasons to improve streamflow simulation with SWAT. Their findings showed that calibrating SWAT based on distinct seasons improved streamflow simulation. The authors applied clustering of the simulation periods into sub-annual periods based on the local climatic conditions and calibrated model parameters for the distinct clusters. Their results showed that the SWAT model performance based on NSE was improved from 0.68 to 0.72.

Conclusion

In this study, we examined a new methodological approach to calibrate and improve the efficiency of the SWAT model in a data scarce environment with a strong seasonality. We applied seasonal decomposition (wet and dry seasons) and successive time shifts of the simulation periods based on measured streamflow and performed multiple model efficiency evaluation of the SWAT model for the Adyar catchment, India. The main aim of this research was to investigate how time invariant and seasonal variability of hydrologic processes affect model parametrizations. Based on our findings, the following conclusions are drawn:

  1. 1.

    Optimization of hydrologic model parameters for successively shifted periods of time resulted in different sets of model parameters. This indicates that non-stationary behaviors of hydrologic processes influence parameter selection during calibration and multiple sets of hydrologic model parameters might be necessary rather than using only a single set of parameter values that are calibrated under fixed calibration and validation periods. We recommend researchers test different calibration periods both in length and timing to assess whether timing has a strong effect on model performance in the studied catchment. At least a simple switch of calibration and validation period should be tested.

  2. 2.

    The degree of uncertainty associated with model parameters can be minimized when a calibration or validation period includes more years.

  3. 3.

    Seasonality of hydrologic processes has considerable effects on the SWAT model capability to represent hydrologic processes in the Adyar catchment. As streamflow in the Adyar catchment is dominated by winter monsoon rainfall, calibrating the SWAT model for wet and dry seasons together was not efficient. However, model calibration for winter monsoon and wet years resulted in a reasonable model performance. This is the season that dominates the hydrologic cycle and is most important to be accurately depicted in a model. Therefore, we conclude that understanding the seasonal behavior of a catchment is important prior to calibrating a hydrologic model.