Evaluating the Performance of Soil and Water Assessment Tool (SWAT) in a Snow-Dominated Climate (Case Study: Azna–Aligoudarz Basin, Iran)

Abstract

This study aims to investigate the capability of the SWAT (Soil and Water Assessment Tool) model in hydrologic simulation of a cold and mountainous climate, the Azna–Aligoudarz Basin, Iran. For this purpose, daily climatic data from the Aligoudarz synoptic station, discharge data from the Marbare hydrometric station, soil and land use maps, and a 10 m digital elevation model of the study area were used. The results demonstrated that the model exhibited poor performance due to poor simulation of runoff generated from snowmelt. To enhance the model’s performance, the calibration period was split into warm and cold seasons using a temperature threshold of 3.6 °C. As a result, the model’s performance improved, with the Nash–Sutcliffe Efficiency (NSE) increasing from 0.28 to 0.60 and R² rising from 0.32 to 0.61. The research indicated that refining the conceptual and theoretical framework of the SWAT model is essential to reduce uncertainty and achieve reliable accuracy, particularly in snow-dominated and mountainous areas.

1. Introduction

Particular hydrological and environmental characteristics of cold climates and mountainous regions affect hydrologic processes and seasonal patterns. These cause a shift towards winter and early spring due to earlier snowmelt, which poses significant challenges to communities residing downstream [1,2,3]. Additionally, in areas with low temperatures, rather than observing the formation of ice on lakes and rivers, four processes including snow, permafrost, freeze–thaw cycles, and glaciers have been comprehensively recognized and evaluated in mountainous regions [4,5]. However, not all these processes can be fully addressed using available hydrological models. Therefore, it is necessary to simplify the complication level of watershed systems and corresponding processes. On the other hand, there are growing appeals for having an accurate representation of the physical reality in a watershed, which makes doing research challenging with simplified models. In the past several years, advancements in snow data collection techniques have partially solved the data shortage and assisted for simulating snow processes in cold climates and mountainous regions [6]. This field is maturing and parallel ongoing research has been dealing with the improvement of predictions in hydrologic models to overcome lower accuracy in results due to simplification in analytical methodologies.

Hydrological models are classified as stochastic or deterministic based on their modeling techniques [7]. Stochastic models rely on statistical regressions of observed data, using a series of input parameters. In contrast, deterministic models operate by representing the conservation of mass, momentum, and energy within a framework of partial differential equations and water budget balances in a system [8]. In recent decades, various collaborative hydrological simulation frameworks have been developed to simulate hydrological processes within different climates and terrains [9,10] that sometimes include water management tools, such as the Soil and Water Assessment Tool (SWAT) [11]. SWAT is a watershed-scale, physically-based, semi-distributed model that enables different time series including daily, monthly, and annual. As a distributed hydrological model for river basins, the SWAT model has become an essential and effective tool for water resource management and environmental protection strategies [12]. SWAT consists of several modules with varying functions, offering great flexibility and versatility to adapt to different hydrological components and conditions [13].

The application of SWAT has been widely adopted in numerous studies including ones in cold regions. Jiang et al. (2020) [14] investigated how climate change affects seasonal and annual streamflow in the Nicolet River watershed in southern Quebec. This research revealed a 13% increase in streamflow peak each year, which was linked to the unexpected snowmelt that occurred in early spring. Wu et al. (2018a&b) [15,16] applied SWAT to investigate the impact of snowmelt on soil erosion in an upland watershed at mid-high latitudes. Their research was conducted in the Abujiao River Basin in Heilongjiang Province, China, and examined how temperature and precipitation influenced soil erosion patterns in a freeze–thaw watershed. These findings showed that runoff and snowmelt significantly impacted seasonal soil erosion. Cold temperatures and high rainfall can increase soil erosion in regions with freeze–thaw cycles. One of the major topics to be investigated for SWAT is model calibration in snow-dominated regions, which has been typically performed by using data over a year [17,18,19] and has not given satisfactory results.

SWAT model has weakness in simulating hydrological processes in cold climates and mountainous regions. Zhao et al. (2022) [17] introduced total radiation to the temperature index method to enhance the simulation of snowmelt runoff incorporating remote sensing data. Results showed enhanced simulation of snowmelt runoff, but this study indicated limitations such as uncertainties in the process of determining the snowmelt threshold snowmelt factor as well as in the data derived from remote sensing, which needs to be fully examined and studied. In another study, Myers et al. (2021) [20] incorporated a simple energy-balance ROS model into the SWAT for improved SWAT calibration and simulations of hydrologic extremes. This study demonstrated that by adding a ROS model to SWAT, results improved. However, the study did not explore potential challenges or limitations related to the scalability of the ROS model when applied to smaller basins or more localized hydrological contexts. In other research, Peker et al. (2021) [21] investigated an application of SWAT in cold and mountainous regions in Turkey. Their study showed that SWAT had inadequate performance on high-elevated catchments. To overcome this challenge, a sequential calibration process was implemented; nevertheless, certain point observations have always deviated from their respective simulated values within specific periods. They concluded that such regions demand rigorous and continuous measurements of snow depth. Additionally, Liu et al. (2020) [22] assessed a modification of SWAT in simulating runoff in a mountainous region in China. To improve SWAT, parameters for snowmelt timing and factors that were tailored to North American conditions (default setting in SWAT) were localized to the studied area in China. The results of simulations revealed the improvement in daily runoff accuracy within the adjusted model. However, the results indicated that values were still lower than the actual values during the snowmelt runoff. This inconsistency probably emerged because only the temperature factor is considered in the degree–day factor model, which has to be further examined in future studies.

A review of existing studies suggests that cold climates and other factors, such as snowmelt, shape hydrological parameters in cold regions. These changes are critical for hydrological trends and water resource management, underscoring the importance of modeling hydrological processes and assessing models in cold climates. This research aims to apply and evaluate the performance of the SWAT model in a mountainous and cold climate region, the Azna–Aligoudarz Basin in western Iran. The modification for the snow-melting process involved using separate calibrations for warm and cold seasons defined by a temperature threshold, which has been further evaluated in the study area.

2. Material and Methods

2.1. Study Area

The study area is located in western Iran and is a part of the Karun catchment basin, which covers a relatively large area of 2189.10 km², with 1585.83 km² (72%) classified as highlands and only 603.27 km² consisting of plains. The main sources of streams are rainfall and snowmelt from high elevations. The primary drainage occurs via the Marbare River, with other notable rivers including the Azna River, Aligoudarz River, Dareh Takht River, and Kamandan River. The basin’s lowest point is 1818 m, found in the outlet areas, while the highest elevation reaches 3746 m above sea level in the northern and southwestern parts. Figure 1 illustrates the geographical location of the study area in Iran.

2.2. Data Description

The input data for this study include daily meteorological information such as precipitation, average relative humidity, minimum and maximum temperatures, solar radiation, and wind speed from the Aligoudarz synoptic station, as well as precipitation data from the Kamandan and Dareh Takht rain gauge stations for 33 years from 1991 to 2023. Additionally, discharge data from the Marbare hydrometric station from 1991 to 2021 along with the soil, land use, and DEM maps of the study area were collected. This dataset was obtained from the Meteorological Organization and the Ministry of Energy of Iran [23]. Climatic data analysis from the Aligoudarz synoptic station indicates that frost occurs on 99 days, with an average of only 2749.7 h of sunshine. A summary of the climatic information from this station is provided in

Table 1. Climatic information from the Aligoudarz synoptic station.

Station	Frost (Day)	Sun (Hour)	Evap. (mm/Day)	Rainfall (mm)	Avg. Humidity (%)	Avg. Temp. (°C)	Climatic Division
Aligoudarz	99	2749.7	2048.2	70.4	40	12.4	Semi-humid summer Very cold winter

.

Figure 2a shows the violin diagram of precipitation and Figure 2b shows the violin diagram of average temperature during the study period. The average monthly precipitation varied from 4.8 to 0, with values of 0–3 mm having the highest occurrence (Figure 2a). The average monthly temperature varied from 27 to −3.6 °C, with values of 5–25 (°C) having the highest occurrence at the Aligoudarz station (Figure 2b).

In this study, four hydrometric stations in the study area were considered. The Marbare hydrometric station, located at the basin’s outlet, was used to calculate the outlet discharge and serves as the basis for model calibration.

Table 2. Characteristics of the hydrometry stations located in the studied area.

Station	Station’s Type	Establishment	Altitude(m)	Latitude	Longitude
Aligoudarz	Synoptic	1985	1980	33°24′	49°42′
Kamandan	Hydrometry Rain Gauge	1967	2050	33°18′14′′	49°25′36″
Dareh Takht	Hydrometry Rain Gauge	1955	1940	33°21′14′	49°22′23″
Marbare	Hydrometry	1958	1820	33°22′52′′	49°24′6″
ChamZaman	Hydrometry	1961	1870	33°23′36′′	49°23′27″
Vazmehdar	Snow Gauge	1974	1912	33°22′35′′	49°22′47″

shows the characteristics of the stations, including geographic longitude and latitude, altitude, and year of establishment. The Vazmehdar snow gauge station was also used to analyze the snow accumulation process.

Figure 3 illustrates the likely distribution of monthly average discharge values at the Cham Zaman, Kamandan, Dareh Takht, and Marbare observation stations. While these stations have similar discharge distributions, their peak values differ. Cham Zaman shows the highest peak at 82 m³/s, with values of 0–5 m³/s most frequent. Dareh Takht peaks at 61 m³/s, with 0.1–5 m³/s most common. Kamandan has a maximum of 5 m³/s and a frequent range of 0.25–5 m³/s. Marbare, as the basin outlet, peaks at 38 m³/s, with values of 0.5–10 m³/s as the most probable.

2.3. Methodology

To evaluate the performance of SWAT in the hydrological simulation of a cold climate and mountainous region, the Azna–Aligoudarz Basin, two scenarios were considered for calibration: (A) calibration without considering snow season (our planned approach, hereafter called the warm-period scenario) and (B) calibration with the snow season. To perform a more detailed analysis and set criteria for categorizing the available data into two groups, warm periods (without snowfall) and cold seasons (with snowfall), the snowfall and snowmelt threshold temperatures were determined for the study area. This was conducted by analyzing the minimum and maximum temperatures, precipitation, and snowfall events at the region’s synoptic station over the study period from 1991 to 2023. For this purpose, the data were first filtered based on the recorded phenomenon parameter at the station. Subsequently, the analysis focused on precipitation events that resulted in the occurrence of the snow phenomenon in the region. The analysis identified a snowfall threshold of 3.6 °C based on the minimum temperature (T_min) and a snowmelt threshold of 0 °C based on the maximum temperature (T_max). These threshold values were then verified during the calibration process for snow-related hydrologic parameters.

2.4. Theoretical Background of Snowmelt and Runoff Generation in SWAT

The theoretical background of SWAT was reviewed to understand the hydrologic simulation of snow processes. The SWAT hydrological model utilizes the water balance equation (Equation (1)) to simulate various processes within the hydrological cycle. These processes encompass surface runoff, surface infiltration, evapotranspiration, snowmelt, groundwater flow, deep infiltration, and subsurface flows. A kinematic storage model was employed to forecast the horizontal movement, while the backward movement was mimicked by generating a shallow groundwater system. The Muskingum technique was employed for stream flood routing. Next, discharge from a waterway was modified to account for losses during transportation, evaporation, diversions, and the flow that returns

$${SW}_t={SW}_0+{\displaystyle {\sum }_{i=1}^t}(R_{day}-Q_{sur}-E_a-W_{seep} Q_{gw})$$

(1)

where SW_t is the final value of soil moisture, SW₀ is the initial value of soil moisture, R_day is the amount of precipitation on the ith day, Q_surf is the amount of surface runoff on the ith day, E_a is the amount of evapotranspiration on the ith day, W_seep is the amount of sediment that enters the unsaturated zone from the soil profile in ith day, Q_gw is the return water amount on the ith day and time t.

2.5. Curve Number Method

The curve number method was developed by the United States Natural Resources Service (NRCS) in 1950 to calculate runoff [24]. This method has been widely used worldwide due to its simplicity, stability, and usefulness for ungauged basins as recommended within the SWAT manual [25]. Equation (2) calculates surface runoff based on the curve number method.

$$Q_{surf}={\displaystyle \frac{{(R_{day}-I_a)}^2}{(R_{day}-I_a)+S}}$$

(2)

where Q_surf is the depth of surface runoff (mm), R_day is the depth of daily rainfall (mm), I_a is the initial storage including catchment, the infiltration rate before the start of runoff and surface storage (mm), and S is the storage, which is estimated from Equation (3)

$$S=25.4({\displaystyle \frac{1000}{CN}}-10)$$

(3)

where CN is the catchment curve number that ranges between 0 and 100 and depends on the basin’s physical characteristics.

2.6. Snow Module in SWAT

In the SWAT snow module, the variation in snowpack is determined by sublimation, melting, and snowfall. The amount of water kept in the snowpack is measured in the form of the snow water equivalent (SWE). Snowmelt occurs if there is snow in a sub-basin area. The snowmelt module will be deactivated if the critical snowfall temperatures fall below the threshold value. The snowmelt component in SWAT is primarily split into snowpack and snowmelt sections [26]. Equation (4) shows the snow mass balance equation:

$${SWE}_i={SWE}_{i-1}+R_{dayi}-E_{subi}-{SNO}_{mlti}$$

(4)

where SWE_i is the equivalent of snow water (mm H₂O), R_dayi is precipitation as snowfall (mm H₂O), E_subi snow sublimation with the unit of mm H₂O and SNO_mlti is the amount of snow melting (mm H₂O), SWE_i−1 is the equivalent of snow water on the previous day (mm H₂O).

Snowfall temperature (SFTMP), measured in degrees Celsius, is a threshold to distinguish between snow and rain. If the daily air temperature is below the SFTMP, any precipitation is classified as snowfall and adds to the snow accumulation. Snow sublimation is estimated based on potential evapotranspiration within the evapotranspiration module [27]. The snowpack will not begin to melt until the temperature exceeds a specific threshold. This temperature, known as the snowmelt temperature (SMTMP), is also expressed in degrees Celsius. The snow temperature is determined by Equation (5):

$$T_{snowi}=T_{snowi-1}\left(1-TIMP\right)+\overline{T}avi\times TIMP$$

(5)

where T_snowi is the snow temperature on current day i and T_snowi−1 is the snow temperature on previous day i − 1 (°C), T_avi is the average air temperature on ith day (°C) and TIMP is the snow temperature delay factor. If the snow temperature is higher than SMTMP, it starts to melt. SWAT model uses a simple linear method of degree–day factor to estimate snowmelt runoff. The amount of snow melting is calculated as follows:

$${SNO}_{mlti}=b_{mlti}\times {sno}_{covi}\times \left[{\displaystyle \frac{T_{snowi}+T_{maxi}}{2}}-SMTMP\right]$$

(6)

$$b_{mlti}={\displaystyle \frac{(SMFMX+SMFMN)}{2}}+{\displaystyle \frac{(SMFMX-SMFMN)}{2}}\times sin\left[{\displaystyle \frac{2\pi }{365}}(i-81)\right]$$

(7)

where SNO_mlti is the total of snowmelt on day i (mm H₂O), b_mlti is the melting coefficient on ith day (mm H₂O °C⁻¹ d⁻¹), sno_covi is the part of the HRU area covered by snow, T_maxi is the maximum air temperature on day i, SMFMX is the melting factor for 21 June (mm H₂O °C⁻¹ d⁻¹) and SMFMN is the melting factor for 21 December (mm H₂O °C⁻¹ d⁻¹). SMFMX is the maximum melting factor in the Northern Hemisphere and the minimum melting factor in the Southern Hemisphere [25].

The degree–day factor method, which is used in SWAT, has a limitation in that its default values for melting coefficient are based on the North American region. Additionally, the model only considers temperature as a factor in snowmelt, overlooking other influential parameters [28]. SWAT incorporates the elevation bands method to address the impact of orography on precipitation and temperature in mountainous areas. This technique allows users to define up to 10 elevation bands within each sub-basin, which has proven to be valuable and essential in various snow-dominated alpine catchments [29,30,31].

2.7. Calibration, Validation, Sensitivity Analysis and Uncertainty Analysis Using SWAT-CUP

The SWAT Calibration and Uncertainty Program (SWAT-CUP) is specifically designed for calibration, validation, sensitivity analysis, and uncertainty analysis [32]. It features various capabilities and algorithms, including parameter solutions (ParaSol), Markov chain Monte Carlo (MCMC), generalized likelihood uncertainty estimation (GLUE), sequential uncertainty fitting 2 (SUFI2), and particle swarm optimization (PSO). The SUFI2 algorithm is employed for uncertainty and sensitivity analysis and calibration and validation of the SWAT model. SUFI-2 seeks to determine the most realistic parameter ranges by iteratively adjusting them to optimize the agreement between simulated and observed data while simultaneously quantifying parameter uncertainty. This method follows a sequential refinement process, where each iteration narrows the parameter ranges based on previous results. Recognizing the inherent uncertainty in hydrological modeling, SUFI-2 explicitly incorporates this aspect by employing 95% prediction uncertainty (95PPU) to evaluate the model’s capability to capture observed values within a defined confidence interval. Model performance and uncertainty are assessed using the p-factor (the proportion of observed data within the 95PPU) and the r-factor (the ratio of the simulated standard deviation to that of the observed data). Additionally, SUFI-2 utilizes objective functions, such as the Nash–Sutcliffe efficiency and the coefficient of determination, to quantify the accuracy of the model’s predictions [32].

The study period spans from 1991 to 2021. The period between 1991 and 1992 was used for model warm-up (to initialize all components), while the period from 1993 to 2016 was designated for calibration. The years from 2017 to 2021 were reserved for model validation.

To compare the hydrologic processes in seasons with snow and seasons without snow, two approaches were utilized for the calibration of the model. First, the entire dataset was used for calibration and validation. In the second approach, the dataset was divided into warm (no snow) and cold (with snow) seasons by identifying the snowfall and snowmelt threshold temperatures. This was achieved by analyzing minimum and maximum temperatures, precipitation, and snowfall occurrences at the synoptic station for the study period from 1991 to 2023. A threshold temperature of 3.6 °C for snowfall (based on minimum temperature) and a 0 °C threshold for snowmelt temperature based on maximum temperature, were established for potential improvement of the calibration outputs.

2.8. Application of SWAT to Simulate Study Area

The modeling process began by dividing the main area into several sub-basins using the elevation digital model map. These sub-basins were then further divided into smaller units, known as hydrological response units (HRUs), by integrating three maps: land use (Figure 4a), soil (Figure 4b), and slope (Figure 4d) classification. Initially, the DEM of the study area (Figure 4c) must be provided to the model to establish the hydrological characteristics, streams, and the basin outlet. The study area was divided into 33 sub-basins. Figure 4c also illustrates the designated sub-basins.

The land use was further introduced to the model. The study area consists of 10 types of land use (Figure 4a and

Table 3. Land use coverage in the studied area.

Land Use Type	Abbreviations	Area (%)	Area (Km2)
Grassland	GRAS	72.5	1588.5
Shrubland	SHRB	15.7	343.8
Irrigated cropland and pasture	CRIR	8.2	179.5
Cropland/grassland mosaic	CRGR	1.3	30.1
Cropland/woodland mosaic	CRWO	0.1	2.9
Baren or sparsely vegetated	BSVG	0.1	2.9
Savanna	SAVA	0.9	19.8
Dryland cropland and pasture	CRDY	0.7	15.5
Residential-medium density	URMD	0.2	4.8
Mixed forest	FOMI	0.04	0.9

): residential-medium density, shrubland, savanna, grassland, mixed forest, cropland/woodland mosaic, irrigated cropland and pasture, cropland/grassland mosaic, dryland cropland and pasture, and baren or sparsely vegetated. A major part of the area consists of grassland (72%), followed by shrubland (15%), and irrigated cropland and pasture make up just 8%, indicating sufficient rainfall for dryland farming.

The soil class map of the area was created and analyzed by the model. The study area contains four soil classes (Figure 4b and

Table 4. Soil classes in the studied area.

Soil Texture Type	Abbreviations	Area (%)	Area (Km2)
Loam	I-Rc-Yk-c-3508	53.813	1178.3
Clay_loam	Xk5-2-3a-3578	40.3	881.8
Loam	I-Rc-Xk-c-3122	3.8	82.5
Clay_loam	Xh33-3a-3289	2.1	46.3

), characterized by loam and clay-loam textures, and most of the region has loam soil (57%). The region was further categorized by slope based on its topographic features. Given that most of the area is mountainous, five slope classes were defined: 0–5%, 5–15%, 15–30%, 30–45%, and over 45%. Figure 4d illustrates the spatial distribution of slope across the study area.

2.9. Evaluation Criteria

The evaluation of the SWAT model was conducted automatically using SWAT-CUP. The SUFI-2 algorithm was selected for its ability to adjust parameters with minimal iterations. Additionally, this method accounts for both model uncertainty and the uncertainty of SWAT parameters relative to the observed data. The coefficient of determination (R²) and the Nash–Sutcliffe efficiency (NSE) model are commonly used in hydrological research to assess model performance [33]. R² indicates the degree of correlation between the estimated and observed values. It ranges from 0 to 1, where 0 signifies no correlation and 1 indicates perfect correlation. The Nash coefficient varies from −∞ to 1, with values closer to 1 reflecting higher simulation accuracy Equations (8) and (9) are used to calculate R² and NSE

$$R^2={\left({\displaystyle \frac{{\sum }_{i=1}^N\left(X_i-\overline{X}\right)\left(Y_i-\overline{Y}\right)}{\sqrt{{\sum }_{i=1}^N{\left(X_i-\overline{X}\right)}^2{\sum }_{i=1}^N{\left(Y_i-\overline{Y}\right)}^2}}}\right)}^2, 0\le R^2\le 1$$

(8)

$$NSE=1-{\displaystyle \frac{{\sum }_{i=1}^n{(Xi-Yi)}^2}{{\sum }_{i=1}^n{(Xi-\overline{X})}^2}}$$

(9)

where $Y_i$ and $X_i$ are the estimated and observed values in the ith time step, and $N$ number of time steps, $\overline{X}$ and $\overline{Y}$ are the average of the estimated and observed values, respectively. Furthermore, model performance was evaluated using the p coefficient and R coefficient. The p coefficient represents the proportion of data within the 95% prediction uncertainty (95PPU) band, and the R coefficient represents the ratio of the average width of the 95PPU band to the standard deviation of the measured variable.

3. Results

In this section, after meeting the modeling requirements, the observed runoff at the outlet station was compared with the model-generated runoff for the corresponding sub-basin. Figure 5 shows trends in precipitation, observed runoff, and model output runoff, revealing that the model contains errors in simulating runoff at the outlet station. These differentiations are noticeable in the differences between dashed lines and solid lines.

To calibrate the model, the observed discharge from hydrometric stations was first compared to the simulated discharge. According to calibration theory, the key parameters that are present in the SWAT model for calibration are known at the beginning, which allows for efficient model calibration. Figure 6 shows observed flow rate changes over time at four hydrometric stations: Dareh Takht, Cham Zaman, Kamandan, and Marbare, compared with the model’s simulations for each sub-basin. These graphs highlight a significant discrepancy between observed and simulated values, showing low base flow and high surface runoff in simulated values. Consequently, parameters related to infiltration, interflow, and base flow recession have been calibrated first in the SWAT-CUP model.

At this point, the calibration stage of the model was conducted. Despite running the model approximately 120 times with 500 simulations and conducting sensitivity analyses on various parameters, the calibration results remained unsatisfactory. The best results achieved were NS = 0.28 and R² = 0.32. Figure 7 illustrates the calibration results for the 1993–2009 period, where the model struggled to accurately predict observed flow rate changes in most areas.

Following the unsuccessful calibration, available data including temperature, precipitation, and discharge have been analyzed to find the relationships between these variables and further investigate reasons for failure. Results showed that observed values from four hydrometric stations alongside corresponding sub-basin data. This analysis revealed instances where, despite rainfall, negative temperatures (black circles in graph) indicated snowfall. Meteorological data from the synoptic station also confirmed these snow events. During such periods, discharge from the hydrometric stations did not increase, reflecting the impact of snowfall. Over time, however, discharge rose as runoff from snowmelt occurred. Analysis of the modeled data showed that in snow-affected areas, the model incorrectly converted all precipitation into runoff, treating it as rain and failing to account for snow. Further analysis of the aquifer and underground streams revealed no impact on surface water during the event. Hence, the model’s inability to simulate runoff from snowmelt was identified as the primary source of error.

To support this argument, Figure 8a shows observed rainfall, temperature, and discharge at the Marbare station, representative of the four study stations. Figure 8b shows the model’s calculated values. At Marbare (the basin outlet), the model overestimated flow during high rainfall periods and underestimated flow during medium and low rainfall periods.

Using the 3.6 °C temperature threshold, observed data have been categorized into cold (with snowfall) and warm (without snow) seasons. The model calibration continued using only warm-season data. Initially, the 3.6 °C threshold temperature was used to attribute months to relevant seasons. It was observed that temperatures dropped below this threshold in multiple months, attributing from November to May as the cold season. Snowmelt events, based on maximum temperature, also occurred during this period, particularly in March and April. This further justified the classification into seven cold months (November–May) and five warm months (June–October). Calibration proceeded using only warm-season data, with a total of 500 simulations, sensitivity analysis, and testing of various parameters. Ultimately, 22 parameters were selected, with TLAPS and SMFMN having the greatest impact on runoff. Chanda et al. (2025) [34], in a study that simulated streamflow in a mountainous region in India, found that 22 hydrological parameters are effective on streamflow and TLAPS has the most impact. The minimum, maximum, variable type (including Multiply, which multiplies the previous value by the suggested value, which is less than 1, and Replace, which replaces the previous value by the suggested value), and optimal values for these parameters are reported in

Table 5. Effective parameters and their optimal values in model calibration.

Parameter	Opt.	Max. Min.		Var. Type	Description
CN2.mgt	−0.49	−0.48	−0.50	Multiply	SCS runoff curve number (−)
ALPHA_BF.gw	0.00	0.01	0.00	Replace	Base flow alpha factor (1/days)
GW_DELAY.gw	306.10	310.14	305.68	Replace	Groundwater delay time (days)
GWQMN.gw	0.92	0.95	0.91	Replace	Threshold depth in shallow aquifer for return flow (mm)
GW_REVAP.gw	0.15	0.15	0.15	Replace	Coefficient for groundwater revap (days)
CH_K2.rte	103.56	103.64	103.49	Replace	Effective hydraulic conductivity in main channel alluvium
SOL_AWC(..).sol	0.88	0.89	0.88	Multiply	Available water capacity of the soil layer (mmH2O/mm soil)
SOL_K(..).sol	0.26	0.26	0.26	Multiply	Saturated hydraulic conductivity (mm/h)
REVAPMN.gw	1.02	1.02	1.02	Replace	Threshold depth in shallow aquifer for revap/percolation (mm)
OV_N.hru	−0.01	−0.01	−0.01	Multiply	Manning’s “n” value for overland flow (−)
SLSUBBSN.hru	0.21	0.21	0.21	Multiply	Average slope length (m)
PLAPS.sub	−13.34	−13.34	−13.34	Replace	Precipitation lapse rate
SURLAG.bsn	14.75	14.75	14.75	Replace	Surface runoff lag time
TLAPS.sub	−9.72	−9.72	−9.72	Replace	Temperature lapse rate
SFTMP.bsn	10.47	10.47	10.46	Replace	Snowfall temperature
SMTMP.bsn	−9.89	−9.89	−9.89	Replace	Snowmelt base temperature
SMFMX.bsn	4.58	4.59	4.58	Replace	Maximum melt rate for snow during year
SMFMN.bsn	0.88	0.88	0.87	Multiply	Minimum melt rate for snow during the year
SNOEB(..).sub	310.65	310.66	310.64	Replace	Initial snow water content in elevation bands
SNOCOVMX.bsn	−35.24	−35.23	−35.45	Replace	Snow water content that corresponds to 100% snow cover
ALPHA_BNK.rte	0.03	0.03	0.03	Replace	Baseflow alpha factor for bank storage (day)
SOL_BD(..).sol	−0.52	−0.51	−0.52	Multiply	Moist bulk density

.

Using the optimal parameters and statistical criteria established for the SWAT model in this research, the calibration of monthly runoff data from 1993 to 2016 resulted in R² and NSE values of 0.61 and 0.60, respectively (

Table 6. Accuracy of the model in the calibration and validation in Scenario A (calibrated based on the complete dataset) and Scenario B (calibrated based on only the warm-period dataset).

Scenarios	Calibration				Validation
	p-Factor	r-Factor	NS	R2	p-Factor	r-Factor	NS	R2
A	0.14	0.00	0.28	0.32	-	-	-	-
B	0.13	0.07	0.6	0.61	0.12	0.06	0.56	0.78

). Additionally, this table shows that the R² and NSE values for the validation stage were 0.78 and 0.56, respectively, which have been greatly improved in comparison with using the complete dataset for the whole period in years. These values show a 91% improvement in R² and more than twice the improvement in NSE for the calibration stage of both scenarios.

Figure 9a illustrates the calibration process for the period from 1993 to 2016, while Figure 10b shows the validation process from 2017 to 2021. These figures indicate that the SWAT model produced satisfactory results in both stages, with a relatively good agreement between simulated runoff and observed values. In some parts of the simulation period, the rising and falling limbs of the simulated hydrographs closely match the observed values. However, in other sections, the model overestimated runoff in most months, with the slopes of the simulated rising and falling limbs being less steep than the observed values. Similarly, Santra and Das (2013) [35] simulated monthly runoff in India’s Chilika basin using the SWAT model, achieving accuracy coefficients of R² = 0.72 and NSE = 0.54 for calibration, and R² = 0.88 and NSE = 0.61 for validation.

Based on the statistical indicators obtained in this study and using warm-period data for model calibration, it was found that the SWAT hydrological model effectively simulated runoff in the studied locality of the Karun watershed with satisfactory accuracy. Nevertheless, the evaluation of SWAT modeling in this region highlighted some weaknesses, particularly related to the area’s mountainous and cold climate, as well as snowmelt events. Of the results obtained from this study and findings from calibration and validation of the model, it can be concluded that while the model performs well in dry and non-mountainous climates, its performance is weaker in cold and mountainous regions since snow has dominated the hydrologic behavior of the regions.

To compare the performance of the model and highlight the importance of considering snow-related processes in runoff simulation, the simulated snowmelt and snowpack values in two calibration scenarios (Scenario A: calibration using complete information in the year and Scenario B: calibration using only warm-season data) in the outlet sub-basin of the basin were compared with the observed data at the station located at the outlet of the basin. The heat map (Figure 10) shows the snowpack values of the two calibration scenarios and observed data at the basin outlet. Based on Figure 10a, Scenario A predicted value of more than 110 mm, which significantly differs from the range in the observed values of snowpack. The performance of the model is improved in Scenario B, where the simulated snowpack is closer to the actual snowpack of the basin.

To statistically examine the series of observational and simulated snowpack data under the two calibration scenarios, the violin plot of this parameter is also presented in three cases (Figure 10b). According to this figure, the actual snowpack at the observatory station in the outlet of the basin, varied from a maximum of 0 to 22 mm, with the most frequent values between 5 and 15. In Scenario A, the snowpack had a wider range of values, with a maximum of 110 and a minimum of 8.5 mm, with values of 10–90 being the most frequent, while in Scenario B the values varied from 5 to 91 mm, with values of 5–45 being the most frequent in the series.

The same analysis was conducted for the simulated snowmelt in Scenarios A and B (Figure 11a). The simulated snowmelt values in Scenario A are lower than those in Scenario B. The model underestimated snowmelt when the entire dataset was used in Scenario A. Figure 11b shows the probability distribution of the snowmelt under two scenarios. The values of the snowmelt in Scenario B fluctuated around 2500 mm, while in Scenario A the values fluctuated around 1300 mm.

4. Discussion

In this study, overestimated flow occurred between November and April (late autumn to early spring), while underestimation mainly occurred in summer. The model’s overestimation was linked to high precipitation and its assumption that all precipitation directly becomes runoff, neglecting snow effects, particularly from the snow-covered Oshtorankouh mountains. On the other hand, underestimation during low rainfall periods resulted from the model’s inability to predict runoff from snowmelt at the end of the rainy season. Similar studies also confirmed this shortcoming in the snow process modeling in the SWAT model and the other hydrologic models (e.g., the variable infiltration capacity (VIC) model). Zhao et al. (2022) [17] evaluated SWAT in simulating runoff in a cold region and demonstrated that in low-flow regimes, the snowmelt runoff is largely underestimated by the SWAT model setting for both the calibration and validation periods. This shows the weakness of SWAT in snowmelt runoff simulation especially in small snowmelt runoffs. Myers et al. (2021) [20], in a study that used SWAT in runoff simulating for a cold climate, showed that the SWAT model simulated almost no snowmelt from December through February. Also, the SWAT model tended to have later peak mean monthly snowmelt timing in April and May. However, based on observational data from the region, April and May have less snowmelt than the SWAT since the snowpack has already melted. Liu et al. (2020) [22], in a study to improve SWAT for snowmelt runoff simulation, showed that the model has an underestimation of maximum runoff values and a lag in the simulated times compared to the observed time series, especially during the snowmelt runoff period from March to April, when the daily runoff simulation value is relatively low and sometimes no runoff is generated at all. Naha et al. (2016) [36] used the VIC model for hydrological modeling in a snow-covered region and found that the model tends to overestimate runoff generated from snowmelt. This supports the claim regarding the existing difficulties in hydrological models addressing snow melting in cold climate regions, as it is distinct from our findings that SWAT underestimates peak flows and exhibits a delay in the periodic runoff pattern. Therefore, it is necessary to explore approaches to enhance the model’s performance.

Similar to the findings in our study, improving the simulation of snow processes in hydrologic models has improved the simulation of runoff and the performance metrics. Zhao et al. (2022) [17] in a study to improve the SWAT model in simulations of snowmelt runoff, used a −0.54 °C threshold as the snowmelt temperature threshold. They added solar radiation to the degree–day factor of SWAT in a cold region, and showed that 0.07% and 0.16% improvement in R² and NSE, respectively. Myers et al. (2021) [20] incorporated an energy-balance rain-on-snow model into the SWAT to improve snowmelt runoff simulation in a snow-dominated area and showed that adding the rain-on-snow model helped the SWAT to simulate snowmelt runoff more than unmodified SWAT and improved the monthly delay of peak snowmelt runoff. Liu et al. (2020) [22] localized snowmelt date and snowmelt factor parameters in SWAT that are set according to North American values. Their results showed better accuracy, but this improvement is not large because it still demonstrates a low runoff simulation value. As previously stated, Naha et al. (2016) [36] found an underestimation of VIC in snowmelt runoff and elaborated on feeding snowmelt runoff from the MODIS sensor to the model which resulted in better performance. Using a different hydrologic model, HPV (Hydrologiska Byråns Vattenbalansavdelning), Lopez et al. (2020) [37] examined the temperature index in snow-dominated regions inside Europe and showed that applying an exponential snowmelt function coupled with no refreezing in the model enhanced model accuracy in snowmelt estimating. Similarly, Santra and Das (2013) [35] simulated monthly runoff in India’s Chilika basin using the SWAT model, achieving accuracy coefficients of R² = 0.72 and NSE = 0.54 for calibration, and R² = 0.88 and NSE = 0.61 for validation after improving simulation of snow hydrology. The stress on the calibration phase for most hydrologic models including SWAT can be also seen in several studies such as Tahmasebi Nasab et al. (2018) [38], which highlighted that incorporating separate calibrations for warm and cold seasons as well as topographic characteristics, such as surface depressions, into SWAT modeling have improved runoff simulations and overall model performance, emphasizing the need for model refinements to better capture hydrological dynamics under varying conditions. These studies confirm the findings of our research and highlight the necessity of improving snow hydrology in the hydrologic models.

This study has certain limitations that should be acknowledged. First, the analysis was conducted in a region with a relatively small number of meteorological stations, which were also unevenly distributed. A higher density of well-distributed stations would enhance the accuracy and reliability of the results by providing more comprehensive spatial coverage. Second, temperature was considered the primary factor influencing snowmelt; however, insolation plays a significant role in this process. In mountainous areas, snowmelt modeling should incorporate the angle of solar incidence, which depends on both the slope’s steepness and its orientation (e.g., north- or south-facing). Furthermore, in addition to direct solar radiation, other forms of radiation such as diffuse radiation passing through cloud cover, as well as reflected and scattered radiation contribute to surface heating. Incorporating these factors in future studies can improve model performance and yield a stronger correlation with observed snowmelt patterns. Further research is needed to enhance the snowmelt module of the SWAT model, making it more applicable to snow-affected catchment areas.

5. Conclusions

This study aimed to evaluate the suitability of the SWAT model for hydrological simulations in the study area, a region characterized by its mountainous terrain and snow-dominated climate. The research covered the statistical period from 1991 to 2021, with two years allocated for warm-up, calibration conducted from 1993 to 2016, and validation from 2017 to 2021. The modeling results revealed a significant shortcoming in the SWAT model’s ability to simulate basin outflow in this region, which was traced back to insufficient consideration of snowmelt runoff. To address this, the data were divided into cold seasons (with snowfall) and warm seasons (without snowfall) to enhance modeling accuracy. The results showed an improvement in model performance when focusing on warm-season data. The NSE improved from 0.28 to 0.60 after seasonal calibration. Our findings also indicated that the SWAT model tends to underestimate peak runoff, failing to capture extreme flow events accurately. For example, the SWAMT model underestimated the peak flows at the Marbare station and the Cham Zaman station by 36.36 m³/s and 94.38 m³/s, respectively. Additionally, the model exhibits a delay in simulating the timing of periodic runoff peaks, highlighting challenges in representing seasonal variations and snowmelt dynamics. SWAT also struggles with hydrologic simulations in cold climate regions, particularly in accurately modeling snow accumulation and melt processes. These limitations suggest the need for model enhancements, including improved snowmelt representation and refined hydrological process calibration, to increase predictive accuracy in snow-covered areas.