The gauged Gurupur watershed, with an area of 699 km², is located in Dakshin Kannada district (“Dakshin” means “southern” in local language), Karnataka, India (Figure 1). The Gurupur stream originates from the Western Ghats (75˚10'42"E; 13˚09'31"N) and flows in the South West direction to join River Nethravathy which later drains into the Arabian Sea near Mangalore (at 74˚49'55"E; 12˚50'43"N).

ASTER 30 m resolution data was used to assess the watershed’s topographical features (Figure 2). Elevation ranges from 2 m amsl (above mean sea level) at the outlet of the watershed to 1872 m amsl in Western Ghats mountains. The mean elevation is 169 m amsl. About 85 percent of the watershed area has elevation less than 200 m amsl; while 5 percent of the watershed area lies above 800 m amsl.

Land use characteristics for the watershed were derived from the Landsat (ETM+) image and updated from the Quick Bird (0.6 m) image (Figure 3).

Dense vegetation cover (evergreen forests) comprises 12 percent of the watershed area and is largely towards its northern part. About 54 percent of watershed area is barren rocky with low and scattered vegetation, 18 percent covered by mixed forests and grasslands, and 15 percent is moderately cultivated with coconut plantations and mixed agriculture.

The soils are predominantly lateritic (84 percent), 11 percent is coastal alluvial soil and 5 percent red sandy soils.

Watershed has elevation-adjusted mean annual rainfall close to 4700 mm. About 90 percent of the total annual rainfall occurs during the period from June-October with coefficient of variation being 18 percent. The probability of the wet day following a wet day ranges from 0.7 to >0.9 during the months of June-October. Minimum and the maximum temperature ranges from 21.8˚C (January) to 32.7˚C (April).

The second watershed, i.e., “ungauged” Upper Payaswini watershed, with an area of about 44.6 km², lies in Kodagu district of Karnataka, which forms a part of Western Ghats with high mountain ranges running north-south as also shown in Figure 1 in sections above. The watershed is drained by Upper Payaswini River, a perennial west flowing stream originating in Brahmagiri hill ranges (75˚30'E; 12˚24'30"N) of the Western Ghats. It traverses through the states of Karnataka and Kerala to eventually join the Arabian Sea near Kasargod (74˚59'5.36"E; 12˚28'49.59"N). The toposheet for the watershed was unavailable and Satellite information (ASTER 30 m resolution) was used to derive the elevation map (Figure 4).

The entire watershed is mountainous with high peaks situated on the southern side. The elevation ranges from 174 m amsl to 1322 m amsl. The mean elevation for the watershed is 771 m amsl. The land use characteristics for the watershed were derived from the Landsat (ETM+) image and updated from the Quick Bird (0.6 m) image (Figure 5). Considerable part of the watershed (85 percent) is covered by evergreen forests, followed by barren rocky land area spread over 14 percent of total watershed area. The built up and habitation area is very less and spread in small clusters. Red sandy soils are predominant and small areas in south western part of the watershed have lateritic soils.

Average annual rainfall is close to 5600 mm. The area receives about 90 percent of rainfall during June to October with coefficient of variation (CoV) being 20 percent. The probability of the wet day following a wet day ranges between 0.7 to >0.9 from June to October. Minimum and the maximum temperature ranges from 14˚C (January) to 29˚C (March).

Soil and Water Assessment Tool (SWAT) is a physically based, computationally efficient, continuous time model with spatially explicit parameterization applicable

to ungauged watersheds [21]. In SWAT, watersheds could be sub-divided into multiple sub-basins connected by a stream network. The model discretization of a sub-basin is according to Hydrologic Response Units (HRUs) consisting of unique soil, slope and land cover combinations. The local HRU water balance is presented by four storage volumes: snow, soil profile (0 - 2 m), shallow aquifer (2 - 20 m), and deep aquifer (>20 m), including canopy interception of precipitation, partitioning of precipitation, snowmelt water, and irrigation water between surface runoff and infiltration, redistribution of water within the soil profile, and return flow from shallow aquifers. Soil profile can be subdivided into multiple layers. Soil water processes include infiltration, evaporation, plant uptake, lateral flow, and percolation to lower layers. Percolation from the bottom of the soil profile and root zone recharges the shallow aquifer (groundwater recharge), which is conceptualized as an unconfined aquifer that contributes to flow in the main channel or reach of the sub-basin [22]. Deep percolation from the shallow aquifer recharges the deep aquifer (deep groundwater recharge).

Surface runoff from daily rainfall is estimated with a modification of the SCS curve number (CN) method and Green-Ampt infiltration method. Lateral sub-surface flow in the soil profile (0 - 2 m) is calculated simultaneously with percolation. A kinematic storage routing technique is used. Groundwater flow condition to total streamflow is simulated by creating shallow aquifer storage [23]. Other components include pumping, withdrawals, and seepage to the deep aquifer. The Muskingum method is used for channel flood routing. Outflow from a channel is adjusted for transmission losses, evaporation, diversions and return flow. The model offers three options for estimating potential evapotranspiration, including Hargreaves, Priestley-Taylor, and Penman-Monteith method. The SWAT model computes evaporation from soil and plants separately. Srinivasan & Arnold [24] describe integration of SWAT into a GIS system for input dataset development and model output visualization.

The hydrologic cycle, as simulated by SWAT, is based on the water balance Equation (1):

S W t = S W 0 + ∑ i = 1 t ( R d a y − Q s u r f − E a − w s e e p − Q l a t − Q g w ) (1)

where SW_t is the final soil water content (mm water), SW₀ is the initial soil water content on day I (mm water), t is the time (days), R_day is the amount of precipitation on day I (mm water), Q_surf is the amount of surface runoff on day I (mm water), E_a is the amount of evapotranspiration on day I (mm water), w_seep is the amount of water entering the vadose zone from the soil profile on day I (mm water), Q_lat is lateral flow from soil to channel and Q_gw is the amount of return flow on day I (mm water).

Major inputs for simulating land phase of the hydrological cycle can be categorized into spatial and non-spatial data. Spatial datasets pertain to topography, land use, and soil type. Non-spatial includes data on weather, soil properties, land use/cover characteristics, and crops. Table 1 provides various input data used for model set up.

The following datasets were prepared for the two watersheds: 1) a Digital Elevation Model (DEM) with a spatial resolution of 30 m (derived from ASTER), 2) land-use map from Landsat (ETM+) image and updated from the Quick Bird (0.6 m resolution), 3) soil map at a scale of 1:100,000 in which the physical soil layer properties (including texture, bulk density, available water capacity, saturated conductivity, soil albedo and organic carbon) were collected mainly from National Bureau of Soil Survey and Land Use Planning (NBSS&LUP) Handbook

and field data, 4) climate data provided from IMD Climatological tables (1951-1980) (mean monthly rainfall, maximum and minimum mean monthly air temperature, mean monthly wind speed, solar radiation, and relative humidity) and rain gauges in Western Ghats (daily precipitation from 6 rain gauges (RNG 1 - 6) in and around Gurupur watershed; 2 rain gauges (RNG1B-2B) in and around Upper Payaswini watershed) provided by the Water Resources Development Organisation, Government of Karnataka (as shown in Figure 1 in sections above and Table 2).

The ArcView Interface for SWAT 2000, was used to delineate the boundaries of the entire watershed and its sub-basins, along with their drainage channel from ASTER 30 m resolution DEM [25]. The boundary of the watershed was superimposed on Google Earth to ensure that delineated watershed boundary and drainage channels closely matched the mountain and hill ridges and drainage network on Google Earth imagery. The Gurupur watershed was finally divided into 26 sub-basins, with sizes varying from 470 hectare to 10,160 hectare (see Figure 2), and the Upper Payaswini watershed into 29 sub-basins with sizes varying from 2.5 to 479.5 hectares (see Figure 4). Further, land use and soil map were used to define multiple HRUs for each of the sub-basins.

The land cover/use map was processed as raster data and included 6 categories consisting of 1) forest (evergreen/mixed), 2) barren and rock outcrop, 3) agriculture, 4) low density habitation/built up, 5) water bodies, 6) grassland/pastures, derived from Landsat (ETM+) image and Quick Bird (0.6 m resolution). These remotely sensed data accounted for the spatial variability in forest and other vegetation characteristics. The vegetation map was finalized after field verification in each of the delineated sub-basins. The barren landcover category, which includes

*RNG refers to raingauge station; STRG refers to stream gauge station.

bare ground, and rock outcrops, was parameterised by modifying the dirt road transportation parameter set in SWAT [10]. The soil database consisting of soil map and attributing databases developed by NBSS&LUP was used to characterize soils in the study areas. Finally, the values for the standard soil and land use parameters used to configure the model were extracted and/or estimated from datasets by the SWAT interface. In SWAT, these parameters are grouped at the levels of watershed, sub-basin, and HRU, and are described in detail by Neitsch et al. [25]. The number and diversity of HRUs can influence model output, and to ensure a high level of resolution, multiple HRUs were defined for each watershed.

Rainfall in mountainous watersheds is influenced by changes in elevation. Fontaine et al. [9] showed that definition of elevation bands within the model’s sub-basins can enhance simulation performance in watersheds with topography having large elevation gradients. To account for a high elevation gain (around 1800 m in Gurupur and around 1000 m in Upper Payaswini watershed), two elevation bands were defined. The data from six raingauge stations has record on daily rainfall from 1991-2009 for Gurupur watershed. Data from these raingauges and from the National Weather Station maintained by Indian Meteorological Department (IMD) located approximately 15 straight line kilometres south west of the south western watershed boundary were used to estimate the local precipitation lapse rate of 25 m/km applicable to elevation range of 1050 - 1800 m amsl and 12 m/km applicable to range of 500 - 1050 m amsl (estimated from rainfall data provided by the Water Resources Development Organisation, State Government of Karnataka, India). Once established, these values were used to parameterize SWAT and then maintained throughout the calibration.

The data on daily rainfall from rainfall gauges were pre-processed into database files within the SWAT required format for a simulation period extending from 1^st January 1991 to 31^st December 2009 in case of Gurupur watershed and 1^st January 1990 to 31^st December 2010 in case of Upper Payaswini. The long term (1951-1980) mean monthly values on rainfall, solar radiation, wind speed, relative humidity, and number of precipitation days in a month obtained from IMD Climatological tables (Station Mangalore in case of Gurupur–12˚52'N; 74˚51'E and station Madikeri in case of Upper Payaswini–12˚25'N; 75˚44'E) were used as inputs into the weather generator (a stochastic engine incorporated in the deterministic SWAT software package). The missing values on daily rainfall were simulated by the calibrated and validated weather generator using the procedures described by Neitsch et al. [25]. Further, during the simulation period records on daily streamflows were available in gauged Gurupur watershed for Polali gauge station STRG (refer to Table 2 and Figure 1) from 1991 to 1999 years and intermittent data for select days in the months of June and July 2009 and 2010 were available from a mini hydropower project STRGB (refer to Table 2 and Figure 1) in the ungauged Upper Payaswini watershed, which makes the model evaluation possible.

The original design objective of the SWAT model was to operate in large-scale, ungauged basins with little or no calibration efforts [26]. Studies have demonstrated that SWAT input parameter values can be successfully estimated without or with minimum calibration in a wide variety of hydrologic systems and geographic locations using readily available GIS databases that have been developed based on prior knowledge [27] [28].

The calibration has been undertaken by conducting a sensitivity analysis to identify the surface runoff and lateral flow related parameters that are sensitive for simulation. Guidance for identifying input parameters for the calibration and sensitivity analysis was provided by prior research in mountainous forested catchments [29] [30] [31]. Research indicated that watershed runoff generation should incorporate rapid lateral flow from soils in wet humid regions as a preferential mechanism of total flow generation, which increases as the degree of wetness of the soil increases. They identified three parameters to reflect local environmental conditions to improve streamflow predictions: the soil infiltration capacity (permeability), soil profile, and soil saturated conductivity.

The sensitivity analysis method implemented in SWAT is Latin Hypercube One-factor-At-a-Time (LH-OAT) [32] [33]. Latin Hypercube (LH) sampling is computationally efficient, and One-At-a-Time (OAT) procedures ensure that a change in model output is unambiguously attributed to the change in an input parameter [10]. LH-OAT starts with taking N Latin Hypercube sample points for N intervals, it then varies each LH sample point P times by changing each of the parameters one at a time. The method operates by loops and each loop starts with a Latin Hypercube point. Around each Latin Hypercube point j, a partial effect S_ij, for each parameter e_i is calculated as in Equation (2):

S i j = | 100 * ( M ( e 1 , ⋯ , e i ∗ ( 1 + f 1 ) , ⋯ , e p ) − M ( e 1 , ⋯ , e i , ⋯ , e p ) [ M ( e 1 , ⋯ , e i ∗ ( 1 + f 1 ) , ⋯ , e p ) + M ( e 1 , ⋯ , e i , ⋯ , e p ) ] 2 ) f i | (2)

where M(...) refers to model functions, f_i is the fraction by which the parameter e_i is changed (a predefined constant) and j refers to a LH point. A final effect is calculated by averaging the partial effects of each loop for all LH points. The final effects can be ranked with the largest effect being given rank 1 and the smallest effect given rank equal to the number of the parameters. Thus, the impact of each parameter on the model results can be quantified and the most sensitive parameters can be identified [34].

For this study the daily measured streamflow datasets were available for the period 1991 to 1999 years (for gauged Gurupur watershed). The period from January 1, 1991 to December 31, 1991 served as a warm-up period for the model, allowing state variables to assume realistic initial values for the calibration period. Data from 1992 to 1995 has been used for calibration and 1996 to 1999 has been used for validation purposes.

Evaluation of model performance during the calibration and validation periods has been done following the methods, namely percent bias (PBIAS), coefficient of determination (R²), and Nash-Sutcliffe efficiency (E_NS) [35].

PBIAS is calculated according to Equation (3):

P B I A S = ( ∑ t = 1 T ( Q s , t − Q m , t ) / ∑ t = 1 T Q m , t ) 100 (3)

where Q_s,t is the model simulated value at time unit t. Q_m,t is the observed data value at time unit t, and t = 1, 2, ..., T. PBAIS measures the average tendency of the simulated data to be larger or smaller than their observed data. Values with small magnitude are preferred, positive values indicate model overestimation bias while negative values indicate underestimation. The optimal value of PBIAS is 0.0, with low-magnitude values indicating accurate model simulation. In short, PBAIS reflects the goodness of model’s simulation in respect of the observed data.

R² is calculated according to Equation (4).

R 2 = { ∑ t = 1 T ( Q m , t − Q ¯ m ) ( Q s , t − Q ¯ s ) / [ ∑ t = 1 T ( Q m , t − Q ¯ m ) 2 ] 0.5 [ ∑ t = 1 T ( Q s , t − Q ¯ s ) 2 ] 0.5 } 2 (4)

where Q ¯ m is the mean observed data value, Q ¯ s is the mean simulated data value for the entire evaluation time period. R² is equal to the square of Pearson’s product-moment correlation coefficient. It represents the proportion of the total variance in the observed data that can be explained by the model. R² ranges between 0.0 and 1.0. Higher values mean better performance.

E_NS is calculated according to Equation (5).

E N S = 1.0 − ∑ t = 1 T ( Q m , t − Q s , t ) 2 / ∑ t = 1 T ( Q m , t − Q ¯ m ) 2 (5)

E_NS indicates how well the plot of the observed values versus simulated values fits the 1:1 line and ranges from −∞ to 1 [36]. Larger E_NS values (close to 1) are equivalent with better model performance.

Additionally, flow duration curves (FDCs) depicting percentile flow availability have been derived and compared both for observed and simulated flows, summarizing relationship between the magnitude and frequency of streamflows [37].

Sensitivity analysis is implemented to identify sensitive parameters for model calibration using LH-OAT. The selected parameters were adjusted over a range of values through a stepwise process that utilized both automated methods, and manual refinement until an acceptable parameter set was obtained [38]. The sensitivity analysis resulted in a list of parameters from most to least sensitive (Table 3) that includes soil, land use and groundwater parameters as discussed in subsequent sections.

CN curve number, which is related to both soil and vegetation, is found to be an important sensitive parameter in the model. It is a function of the soil’s permeability, land use and antecedent soil water conditions. A moderate to low initial value of CN2 number for such highly forested sub-basins is desirable. Another important parameter is Soil Available Water Content (AWC) [26]. With increasing AWC, the soil storage increases resulting in decrease of surface runoff, and or percolation and water available for soil evapotranspiration increases. In addition to AWC, other parameters that were found to be sensitive include Soil Evaporation Compensation factor (ESCO), soil depth (Z) and soil texture including silt (SILT), sand (SAND), and clay (CLAY) content. ESCO adjusts depth distribution for evaporation from the soil to account for capillary action, crusting and cracking. As the value of ESCO is decreased it enables lower layers of soil matrix to compensate for the water deficit. Percentage decrease in ESCO values to match the flows is mostly felt in sub-basin characterized by cultivable area and trees and presence of good soils. Alpha factor for groundwater, ALPHA_BF, is a function of the overall topography, drainage pattern, soil, and geology of the watershed [13]. It is a direct index of the intensity with which the groundwater outflow responds to changes in recharge [39]. Increases in alpha factor cause the simulated recession curve to be much faster. Varying this parameter does the adjustment of recession curves. Variations in values of various other parameters such as CH_K2, GW_DELAY, and SURLAG did not further improve the results and showed low sensitivity in terms of adjusting flows in this study area. Table 3 shows the results of sensitivity analysis of parameters.

*Figures in brackets refer to number of soil layers.

Calibration indicates five soil-land use related parameters to be sensitive for simulation namely, soil evaporation compensation factor (ESCO), soil available water capacity (SOL_AWC), soil depth (SOL_Z) and soil layers, groundwater baseflow (ALPHA_BF) and curve number (CN2).

The measured and simulated daily streamflow for calibration and validation periods are presented. Scatter plots that compare observed and simulated flows are shown. Model performance is evaluated using Nash-Sutcliffe coefficient (E_NS), percent bias (PBIAS) and coefficient of determination (R²). R², and E_NS values above 0.6 are generally acceptable and values near unity indicate a close relationship between predicted and measured yields [40]. After the calibration of several parameters (sensitivity analysis), it was also estimated if SWAT successfully captured the study area’s hydrologic characteristics and reproduced acceptable daily runoff simulations during validation period.

Model validation is conducted using an additional 3 years of data for gauged watershed. FDCs have been derived both for observed and simulated flows in gauged watershed. For ungauged watershed, FDC has been derived based on the simulated flows obtained from SWAT for the watershed. Finally, different runoff components simulated by SWAT are represented for capturing the hydrologic cycle.