Abstract
Rainfall is a vital input to model watershed hydrology, and the availability of numerous gridded and point-observed rainfall datasets poses a major challenge to the modellers to choose the appropriate data. This study compares three gridded rainfall datasets (i.e., 1° × 1°, 0.5° × 0.5°, and 0.25° × 0.25°) and point rainfall observations of the India Meteorological Department (IMD) on the simulation of streamflow of a river basin in the southern Western Ghats (India) using the Soil and Water Assessment Tool (SWAT). The results show that the different datasets lead to different optimal model parameter values and consequent water balance components, significantly in groundwater hydrology. The 0.5° × 0.5° and 0.25° × 0.25° datasets result in comparable SWAT model performances (NSE = 0.75 and 0.70, respectively), probably due to the similarity in the rain gauge network density employed for deriving the datasets and also due to the spatial discretization threshold used for sub-watershed delineation. However, the coarser resolution data (1° × 1°) results in poor performance (NSE = 0.21). The study suggests that the choice of rainfall data depends on the spatial resolution of the data and the spatial discretization threshold while compromising the computational requirement vis-à-vis simulation accuracy.
HIGHLIGHTS
The effects of various IMD gridded rainfall datasets and point rainfall observations on the simulation of streamflow of a river basin of the southern Western Ghats (India) was assessed.
The optimal parameter values for 1° × 1° rainfall data are far deviated from the point observation data.
The 0.25° × 0.25° and 0.5° × 0.5° have better model performances compared to 1° × 1° and point observations.
Graphical Abstract
ACRONYMS
- CRB
Chaliyar River Basin
- CE
Coefficient of Efficiency
- CV
Coefficient of Variation
- FAO
Food and Agricultural Organisation
- HRU
Hydrological Response Unit
- IMD
Indian Meteorological Department
- KGE
Kling-Gupta Efficiency
- NRMSE
Normalised Root Mean Squared Error
- NRSC
National Remote Sensing Centre
- QGIS
Quantum Geographic Information System
- R2
R-squared statistic
- SRTM DEM
Shuttle Radar Topography Mission Digital Elevation Model
- SWAT
Soil and Water Assessment Tool
- VE
Volumetric Efficiency
INTRODUCTION
Watershed hydrological models are effective tools for quantifying watershed responses to changes in different natural and anthropogenic factors (e.g., climate, land use/ land cover, land and water management practices) (Abbaspour et al. 2015; Dile et al. 2016; Woznicki et al. 2016; Yang et al. 2016). Among various watershed hydrological models, the Soil and Water Assessment Tool (SWAT) is one of the most widely used across the globe for decision-making, policy development and assessment for watershed management over a broad spectrum of spatial scales, environmental conditions, and applications (e.g., Nerantzaki et al. 2015; Schmalz et al. 2015; Guse et al. 2016; Malagò et al. 2016; Mannschatz et al. 2016; Valiya Veettil & Mishra 2020; Frizzle et al. 2021; Getachew et al. 2021). Since the accuracy of the SWAT model predictions is largely underpinned by the accuracy of the input data (i.e., rainfall, topography, soil characteristics and land use/ land cover), uncertainties in the input data in space and time hinder effective decision-making and operational applications. Being the most important process driver variable determining the spatial and temporal variation of the hydrological fluxes and state variables in the hydrological models, rainfall data is the largest source of uncertainty in hydrologic simulations (Nijssen & Lettenmaier 2004; Bárdossy & Das 2008; Starks & Moriasi 2009; Tuppad et al. 2010; Schreiner-McGraw & Ajami 2020). Owing to the non-linearity of the hydrological processes, the errors in the rainfall data result in an inappropriate representation of the hydrological processes and potentially lead to undesirable outcomes for watershed management decisions (McMillan et al. 2018).
Point (rain gauge-based) rainfall measurements are the major source of rainfall data in SWAT modelling approaches (Tan et al. 2021). However, the application of spatially distributed (gridded) rainfall products in the SWAT model provides better streamflow estimation in large watersheds compared to point-based observations due to the better spatial representation of the precipitation pattern across the watershed (Moon et al. 2004). Currently, numerous gridded rainfall data products, derived from satellite/radar-based observations, reanalysis, and interpolation of rain gauge data, are available and their applicability in hydrological modelling has been discussed by many researchers (e.g. Ashouri et al. 2015, Musie et al. 2019; Dembele et al. 2020; Le et al. 2020; Setti et al. 2020; Zhang et al. 2020; Talchabhadel et al. 2021; Tan et al. 2021; Wang et al. 2021). Although numerous gridded datasets are available for the southern Asian region, Tan et al. (2021) observed that most of the SWAT modelling applications in India use the gridded rainfall dataset of the India Meteorological Department (IMD) at a spatial resolution of 0.25° × 0.25° and at daily time scale. In addition, different researchers (e.g., Kolluru et al. 2020; Sharannya et al. 2020; Setti et al. 2020) recommended the IMD gridded data as an appropriate choice for SWAT modelling applications in India when point rainfall measurements are unavailable.
The daily gridded rainfall data of IMD over India are available at different spatial resolutions, viz., 1° × 1°, 0.5° × 0.5°, and 0.25° × 0.25°. The IMD gridded rainfall data products were generated from the interpolation of point rainfall measurements, but the rain gauge station network density varies significantly among the datasets (Rajeevan et al. 2006; Rajeevan & Bhate 2009; Pai et al. 2014). The 1° × 1° gridded daily rainfall data were developed based on a fixed network of 2,140 rain gauge stations (Rajeevan et al. 2006), whereas the 0.5° × 0.5° data used a variable network of 6,076 rain gauge stations (Rajeevan & Bhate 2009). Daily rainfall data from 6,955 stations were utilized for the development of the 0.25° × 0.25° gridded rainfall dataset (Pai et al. 2014). Thus, the difference in the spatial resolution of these gridded datasets is influenced by the variability of the rain gauge network density considered for the interpolation at various grid points. Nevertheless, a realistic representation of the spatial and temporal variability of rainfall patterns is required for accurate estimation of catchment responses in watershed hydrological models. In this premise, understanding the adequacy and coherence of the IMD gridded data products to reproduce watershed hydrological processes is a prerequisite to the data selection for SWAT modelling.
The SWAT model delineates the watershed into sub-basins, which are then discretized into hydrological response units (HRUs), which are the basic computational units of SWAT, where all the water balance components are estimated. The SWAT model assumes spatially uniform precipitation within a sub-basin, even if the gridded rainfall data are used as the model collects the rainfall input to the HRUs of a given sub-basin from the grid (or station) closest to the centroid of the sub-basin. Hence, the performance of the SWAT model strongly depends on the accuracy of the representation of the rainfall in a sub-basin by the grid closest to the centroid of the sub-basin (Neitsch et al. 2011). Therefore, it is significant to evaluate the suitability of the IMD gridded rainfall datasets in simulating watershed hydrology. Moreover, Dembele et al. (2020) noted that many regional climate datasets outperform the global data, calling for regional evaluation studies for hydrological modelling applications. Many researchers assessed the suitability of different available satellite datasets in simulating watershed hydrology in the Indian context (e.g., Sharannya et al. 2020; Setti et al. 2020). These studies evaluated the satellite data for SWAT modelling by considering the IMD rainfall data at 0.25° × 0.25° resolution as the reference with the assumption that finer resolution spatial data would yield better performance. However, there is a growing need to evaluate the variability in the performance of the SWAT model with the different IMD gridded rainfall datasets (Vema et al. 2018). Therefore, this study focuses on the impact of varying spatial resolutions of the IMD gridded rainfall datasets (i.e., 1° × 1°, 0.5° × 0.5°, and 0.25° × 0.25°) and point rainfall observations on the simulation of streamflow of a tropical watershed (Chaliyar River Basin; CRB) of the southern Western Ghats, Kerala, India. This study critically addresses the research question: how significant is the spatial resolution of the IMD gridded rainfall data for modelling the streamflow using SWAT? It sheds light on the implications of the choice of the gridded rainfall datasets for the representation of spatial patterns in SWAT modelling.
METHODOLOGY
Watershed characteristics
Location map and stream network of Chaliyar River Basin (CRB; Kerala, India).
Input data for SWAT modelling
The primary inputs for setting up the SWAT model are the digital elevation model (DEM), soil map, land use/ land cover and weather data. The SRTM DEM (1-arc second) was used for the representation of the catchment topography of the basin. The soil characteristics of the catchment area draining Kerala were collected from the Soil Survey Organization, Government of Kerala, whereas the soil data of the remaining area of the basin (in Tamil Nadu) were extracted from the FAO soil database. Figure 2 shows the spatial variability and the areal extent of the different soil textural classes of the basin. The land use/ land cover data of the catchment area in Kerala (generated from IRS P6 LISS III images) were collected from the Kerala State Remote Sensing Centre, Government of Kerala, whereas the land use/ land cover data of the watershed area in Tamil Nadu were extracted from the NRSC land use/ land cover data (derived from IRS P6 AWiFS images) (Figure 3). The accuracy of the land use/ land cover data of the CRB was estimated using Google Earth (Rwanga & Ndambuki 2017). The overall efficiency and the Kappa coefficient of the land use/ land cover data were 88% and 0.86, respectively. The land use/ land cover data were reclassified to the corresponding land use/ land cover classes in the SWAT database.
SWAT modelling
In this study, QSWAT, an interface in QGIS, was used to set up the hydrological model for CRB. The spatially distributed geoenvironmental data such as DEM, land use/land cover, soil and weather data (IMD gridded and point observation rainfall and temperature) were used. The CRB was divided into 40 sub-basins (areal extent ranging from 10 to 240 km2) with a total number of 2,537 HRUs. The delineated watershed obtained from the SWAT model is shown in Figure 4. The SWAT model was run for a period from 1980 to 2004, where the first five years of the simulation were selected as the warm-up period. The model was run with four different rainfall datasets (gridded data and point observations) without changing the other input data (i.e., DEM, land use/land cover, soil as well as temperature data).
Calibration and validation of the SWAT model
The calibration and validation of the SWAT model for the basin were carried out using the Sufi2 optimization technique in SWAT-CUP (Abbaspour et al. 2007, 2015). The primary step in the calibration and validation of the SWAT model is to identify the most sensitive parameters of the river basin. In this study, we used one-at-a-time sensitivity analysis in the SWAT-CUP tool to estimate the sensitivity of each model parameter. The sensitivity analysis indicates that CN2, ALPHA_BF, GW_DELAY, GWQMN, SOL_AWC, ESCO, SOL_K, CH_K2, SLSUBBSN, and GW_REVAP were the most sensitive parameters. A few previous studies also support this observation (Raneesh & Thampi 2010, 2011; Sathya & Thampi 2020; Upadhyay et al. 2022). Hence, these parameters were used for the calibration and validation of the SWAT model of the CRB.
The calibration period was from 1985 to 1998 and the validation period was from 1999 to 2004. All four SWAT simulations obtained using the three gridded rainfall data and point observation data were calibrated and validated independently for the same time period. Maximizing the coefficient of efficiency (CE) was considered the objective function during the calibration of the models. The variability in the optimum parameter values for different models was analyzed. The performance of the SWAT model was evaluated using various indices, viz., normalized root mean squared error (NRMSE), R-squared statistic (R2), coefficient of efficiency (CE) volumetric efficiency (VE) and Kling-Gupta efficiency (KGE) (Dawson et al. 2007). In addition, the percentage error in the predicted high flow (i.e., the ratio of the difference between simulated and observed values to the observed value) was also estimated to assess the model performance at high flows (Sudheer et al. 2003).
RESULTS
Spatial distribution of rainfall across the basin
Spatial variability of mean annual rainfall in the CRB: (a) 0.25° × 0.25°, (b) 0.5° × 0.5°, (c) 1° × 1°, and (d) point observations.
Spatial variability of mean annual rainfall in the CRB: (a) 0.25° × 0.25°, (b) 0.5° × 0.5°, (c) 1° × 1°, and (d) point observations.
The datasets exhibit considerable spatial variability of seasonal rainfall as well (Table 1), probably due to the influence of regional orography (Simon & Mohankumar 2004). The major rain-bearing seasons, such as the Indian summer monsoon and post-monsoon, are characterized by relatively lower spatial variation of rainfall compared to the non-monsoon seasons (i.e., winter and pre-monsoon). For example, pre-monsoon rainfall in the region is mostly convective and caused by mesoscale forcing, which includes local effects, such as urbanization (Mitra et al. 2012). The spatial variability of the point observed rainfall is considerably higher in the non-monsoon seasons compared to the gridded datasets (Table 1). Among the different gridded datasets, 0.5° × 0.5° data exhibit similar spatial variability with the point observations in the annual as well as the Indian summer monsoon and post-monsoon seasonal scales. The rainfall pattern suggests spatial coherence between the point observations and 0.5° × 0.5° gridded data with minor variations in the rainfall magnitude. Although finer in spatial resolution, the 0.25° × 0.25° data underestimate the rainfall in the upstream sub-basins while overestimating in the northern parts and downstream sub-basins compared to the rain gauge-based estimates.
Descriptive statistics of annual and seasonal rainfall in the CRB
. | Rainfall data . | Mean (mm) . | CV (%) . |
---|---|---|---|
Annual | Point | 2327 | 16 |
0.25° × 0.25° | 2259 | 32 | |
0.5° × 0.5° | 2425 | 16 | |
1° × 1° | 2114 | 36 | |
Winter (January-February) | Point | 11 | 64 |
0.25° × 0.25° | 15 | 60 | |
0.5° × 0.5° | 10 | 18 | |
1° × 1° | 11 | 38 | |
Pre-monsoon (March-May) | Point | 250 | 40 |
0.25° × 0.25° | 196 | 27 | |
0.5° × 0.5° | 232 | 23 | |
1° × 1° | 179 | 48 | |
Indian summer monsoon (June-September) | Point | 1636 | 17 |
0.25° × 0.25° | 1634 | 42 | |
0.5° × 0.5° | 1819 | 20 | |
1° × 1° | 1601 | 39 | |
Post-monsoon (October-December) | Point | 393 | 17 |
0.25° × 0.25° | 413 | 14 | |
0.5° × 0.5° | 362 | 17 | |
1° × 1° | 323 | 21 |
. | Rainfall data . | Mean (mm) . | CV (%) . |
---|---|---|---|
Annual | Point | 2327 | 16 |
0.25° × 0.25° | 2259 | 32 | |
0.5° × 0.5° | 2425 | 16 | |
1° × 1° | 2114 | 36 | |
Winter (January-February) | Point | 11 | 64 |
0.25° × 0.25° | 15 | 60 | |
0.5° × 0.5° | 10 | 18 | |
1° × 1° | 11 | 38 | |
Pre-monsoon (March-May) | Point | 250 | 40 |
0.25° × 0.25° | 196 | 27 | |
0.5° × 0.5° | 232 | 23 | |
1° × 1° | 179 | 48 | |
Indian summer monsoon (June-September) | Point | 1636 | 17 |
0.25° × 0.25° | 1634 | 42 | |
0.5° × 0.5° | 1819 | 20 | |
1° × 1° | 1601 | 39 | |
Post-monsoon (October-December) | Point | 393 | 17 |
0.25° × 0.25° | 413 | 14 | |
0.5° × 0.5° | 362 | 17 | |
1° × 1° | 323 | 21 |
Variability in the model parameters
The optimal parameters of the calibrated SWAT models exhibit considerable variability among the different rainfall datasets (Table 2). The optimal parameter values for the 0.5° × 0.5° gridded rainfall are consistent with the optimal parameters for the point observed data, which could be attributed to the close similarity in the representation of the rainfall pattern between these datasets, compared to other datasets. The variation in the model parameter values vis-à-vis various water balance components was assessed to understand the variability in the representation of the watershed hydrological processes due to the variations in the parameter values (Table 3).
Optimal model parameters for different rainfall datasets
Parameter . | Lower limit . | Upper limit . | Optimized value . | |||
---|---|---|---|---|---|---|
Point . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . | |||
r_CN2 | −0.20 | 0.20 | −0.07 | 0.02 | −0.07 | −0.19 |
v_ALPHA_BF (days) | 0.0 | 1.0 | 0.89 | 0.96 | 0.89 | 0.98 |
v_GW_DELAY (days) | 30.0 | 450 | 33 | 43.86 | 33 | 350.45 |
v_GWQMN (mm) | 0.0 | 5000 | 2035 | 1045 | 2035 | 4105 |
r_SOL_AWC (mm) | 0.0 | 1.0 | 0.60 | 0.83 | 0.60 | 0.81 |
v_ESCO | 0.0 | 1.0 | 0.33 | 0.45 | 0.33 | 0.85 |
r_SOL_K (mm/hr) | 0.0 | 0.80 | 0.12 | 0.20 | 0.12 | 0.24 |
v_CH_K2 (mm/hr) | 0.0 | 500 | 270.5 | 378.5 | 270.5 | 447.5 |
r_SLSUBBSN (m) | 0.0 | 0.20 | 0.14 | 0.11 | 0.14 | 0.10 |
v_GWREVAP | 0.0 | 0.20 | 0.02 | 0.02 | 0.02 | 0.07 |
Parameter . | Lower limit . | Upper limit . | Optimized value . | |||
---|---|---|---|---|---|---|
Point . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . | |||
r_CN2 | −0.20 | 0.20 | −0.07 | 0.02 | −0.07 | −0.19 |
v_ALPHA_BF (days) | 0.0 | 1.0 | 0.89 | 0.96 | 0.89 | 0.98 |
v_GW_DELAY (days) | 30.0 | 450 | 33 | 43.86 | 33 | 350.45 |
v_GWQMN (mm) | 0.0 | 5000 | 2035 | 1045 | 2035 | 4105 |
r_SOL_AWC (mm) | 0.0 | 1.0 | 0.60 | 0.83 | 0.60 | 0.81 |
v_ESCO | 0.0 | 1.0 | 0.33 | 0.45 | 0.33 | 0.85 |
r_SOL_K (mm/hr) | 0.0 | 0.80 | 0.12 | 0.20 | 0.12 | 0.24 |
v_CH_K2 (mm/hr) | 0.0 | 500 | 270.5 | 378.5 | 270.5 | 447.5 |
r_SLSUBBSN (m) | 0.0 | 0.20 | 0.14 | 0.11 | 0.14 | 0.10 |
v_GWREVAP | 0.0 | 0.20 | 0.02 | 0.02 | 0.02 | 0.07 |
‘v_’ and ‘r_’ imply replacement and a relative change to the initial parameter values, respectively.
Variability in the mean annual water balance components of the CRB for different rainfall datasets
Water balance components . | Point . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . |
---|---|---|---|---|
Precipitation | 2353 | 2333 | 2401 | 2073 |
Surface runoff | 1195 | 1230 | 1292 | 1115 |
Lateral flow | 240 | 241 | 234 | 195 |
Groundwater contribution to streamflow | 717 | 656 | 671 | 578 |
Total aquifer recharge | 757 | 695 | 709 | 611 |
Total water yield | 2189 | 2162 | 2233 | 1920 |
Percolation out of soil | 755 | 696 | 709 | 611 |
Actual evapotranspiration | 127 | 129 | 128 | 119 |
Water balance components . | Point . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . |
---|---|---|---|---|
Precipitation | 2353 | 2333 | 2401 | 2073 |
Surface runoff | 1195 | 1230 | 1292 | 1115 |
Lateral flow | 240 | 241 | 234 | 195 |
Groundwater contribution to streamflow | 717 | 656 | 671 | 578 |
Total aquifer recharge | 757 | 695 | 709 | 611 |
Total water yield | 2189 | 2162 | 2233 | 1920 |
Percolation out of soil | 755 | 696 | 709 | 611 |
Actual evapotranspiration | 127 | 129 | 128 | 119 |
Note: all values are in mm.
The water balance components estimated by the SWAT model with the same parameter values are significantly different between the rainfall datasets (i.e., point data vs. 0.5° × 0.5° data) (Tables 2 and 3). The variations in the water balance components for different rainfall datasets imply that the water balance components for 0.25° × 0.25° and 0.5° × 0.5° datasets have a relatively lower deviation from the optimal parameter values of the point observation data compared to the 1° × 1° rainfall (Table 3). The ratio of the precipitation to evaporation for the different models using 0.25° × 0.25°, 0.5° × 0.5°, and 1° × 1° and point datasets is 18.08, 18.75, 17.42 and 18.52 respectively, which also implies the comparable performances of the 0.25° × 0.25° and 0.5° × 0.5° datasets with point observations. On the other hand, a large deviation of the optimal parameter values for the 1° × 1° data is perhaps the result of the difference in the spatial variability of rainfall across the basin. The CN2 is the initial SCS curve number for moisture condition II and is a function of the soil permeability, land use/ land cover and antecedent soil water conditions (Neitsch et al. 2011). Since the curve number directly influences the surface runoff, the decrease in the curve number (Table 2) leads to a corresponding decrease in the surface runoff (Table 3).
The variability in the optimal values of the different parameters of the SWAT model using different rainfall data indicates that the rainfall at 0.5° × 0.5° represents the rainfall pattern and watershed responses similar to the point observations. The calibrated parameter values for the 1° × 1° gridded rainfall dataset manifest considerable deviation from the point observations, which could be attributed to the inadequate representation of the rainfall for the CRB. Although finer in spatial resolution, the 0.25° × 0.25° gridded data also show deviation from the parameters of the point observations. The analysis implies that the deviation of most of the sensitive parameters of the 1° × 1° and 0.25° × 0.25° gridded datasets is the result of the relatively lower rainfall estimates of these gridded datasets.
Performance evaluation of model simulations
Simulation of daily discharge during calibration period for (a) 0.25° × 0.25° (b) 0.5° × 0.5° (c) 1° × 1° (d) point observation rainfall datasets.
Simulation of daily discharge during calibration period for (a) 0.25° × 0.25° (b) 0.5° × 0.5° (c) 1° × 1° (d) point observation rainfall datasets.
Simulation of daily discharge during the validation phase for (a) 0.25° × 0.25° (b) 0.5° × 0.5° (c) 1° × 1° (d) point observation rainfall datasets.
Simulation of daily discharge during the validation phase for (a) 0.25° × 0.25° (b) 0.5° × 0.5° (c) 1° × 1° (d) point observation rainfall datasets.
Daily rainfall and discharge time series of the 1° × 1° rainfall dataset during 2001–2003. The shaded region indicates the Indian summer monsoon season, which is the dominant rainy season of the basin. However, the rainfall values are remarkably lower during the season.
Daily rainfall and discharge time series of the 1° × 1° rainfall dataset during 2001–2003. The shaded region indicates the Indian summer monsoon season, which is the dominant rainy season of the basin. However, the rainfall values are remarkably lower during the season.
Evaluation matrices used for SWAT model performance assessment
Evaluation Matrices . | IMD Rainfall Data . | |||||||
---|---|---|---|---|---|---|---|---|
Point data . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . | |||||
Calibration . | Validation . | Calibration . | Validation . | Calibration . | Validation . | Calibration . | Validation . | |
NRMSE | 0.95 | 0.86 | 0.74 | 0.96 | 0.82 | 0.88 | 1.22 | 1.57 |
R2 | 0.74 | 0.73 | 0.84 | 0.70 | 0.80 | 0.77 | 0.57 | 0.27 |
CE | 0.73 | 0.70 | 0.83 | 0.70 | 0.79 | 0.75 | 0.56 | 0.21 |
VE | 0.53 | 0.52 | 0.62 | 0.55 | 0.62 | 0.49 | 0.42 | 0.25 |
KGE | 0.71 | 0.65 | 0.81 | 0.74 | 0.81 | 0.71 | 0.51 | 0.27 |
Evaluation Matrices . | IMD Rainfall Data . | |||||||
---|---|---|---|---|---|---|---|---|
Point data . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . | |||||
Calibration . | Validation . | Calibration . | Validation . | Calibration . | Validation . | Calibration . | Validation . | |
NRMSE | 0.95 | 0.86 | 0.74 | 0.96 | 0.82 | 0.88 | 1.22 | 1.57 |
R2 | 0.74 | 0.73 | 0.84 | 0.70 | 0.80 | 0.77 | 0.57 | 0.27 |
CE | 0.73 | 0.70 | 0.83 | 0.70 | 0.79 | 0.75 | 0.56 | 0.21 |
VE | 0.53 | 0.52 | 0.62 | 0.55 | 0.62 | 0.49 | 0.42 | 0.25 |
KGE | 0.71 | 0.65 | 0.81 | 0.74 | 0.81 | 0.71 | 0.51 | 0.27 |
Scatter plots between observed and simulated discharges of point observation rainfall dataset during (a) calibration (b) validation.
Scatter plots between observed and simulated discharges of point observation rainfall dataset during (a) calibration (b) validation.
Scatter plots between observed and simulated discharges of 0.25° × 0.25° rainfall dataset during (a) calibration and (b) validation.
Scatter plots between observed and simulated discharges of 0.25° × 0.25° rainfall dataset during (a) calibration and (b) validation.
Scatter plots between observed and simulated discharges of 0.5° × 0.5° rainfall dataset during (a) calibration and (b) validation.
Scatter plots between observed and simulated discharges of 0.5° × 0.5° rainfall dataset during (a) calibration and (b) validation.
Scatter plots between observed and simulated discharges of 1° × 1° rainfall dataset during (a) calibration and (b) validation.
Scatter plots between observed and simulated discharges of 1° × 1° rainfall dataset during (a) calibration and (b) validation.
Table 4 shows the results of different evaluation metrics used for assessing the performance of the SWAT model simulations among the various rainfall datasets. In general, the SWAT model performance using 0.25° × 0.25° and 0.5° × 0.5° data is comparable with the point observed data. Further, these gridded data products offer better accuracy (in terms of different metrics) compared to the point data. However, all the evaluation metrics indicate the poor performance of the 1° × 1° rainfall dataset. The model performances using 0.5° × 0.5° as well as 0.25° × 0.25° data are more consistent, implying the capability of these gridded rainfall products for the application of hydrological modelling at the watershed scale.
Table 5 provides the percentage error in annual peak flow estimates for different rainfall datasets during the validation period. The positive value of percentage error implies the overestimation of simulation, whereas the negative value indicates underestimation. It is noted that the peak flows are underestimated in most of the years during the validation period for all rainfall datasets (Table 5). However, the percentage error is found to be minimum for 0.5° × 0.5° data in most of the years compared to other rainfall products. Large underestimation of peak flow values can be seen for 1° × 1° rainfall data except for the years 1999 and 2000.
Percentage error in peak flow estimation during validation years
Year . | Point data . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . |
---|---|---|---|---|
1999 | −67.54 | 49.96 | −41.47 | −45.58 |
2000 | −43.98 | −25.71 | −46.87 | 14.51 |
2001 | −27.98 | −42.08 | −15.37 | −81.69 |
2002 | −49.93 | −53.10 | −19.28 | −96.46 |
2003 | 11.67 | −4.27 | 35.92 | −93.45 |
2004 | −55.95 | 54.26 | −25.15 | −68.53 |
Year . | Point data . | 0.25° × 0.25° . | 0.5° × 0.5° . | 1° × 1° . |
---|---|---|---|---|
1999 | −67.54 | 49.96 | −41.47 | −45.58 |
2000 | −43.98 | −25.71 | −46.87 | 14.51 |
2001 | −27.98 | −42.08 | −15.37 | −81.69 |
2002 | −49.93 | −53.10 | −19.28 | −96.46 |
2003 | 11.67 | −4.27 | 35.92 | −93.45 |
2004 | −55.95 | 54.26 | −25.15 | −68.53 |
The evaluation metrics were estimated for different seasons to understand the effect of differences in the precipitation pattern and spatial variability of rainfall on SWAT model simulations (Tables 6 and 7). The NRMSE values obtained for the Indian summer monsoon season for 0.25° × 0.25°, 0.5° × 0.5°, 1° × 1° and point rainfall datasets are 0.26, 0.21, 0.45 and 0.30, respectively during calibration and 0.31, 0.29, 0.76 and 0.36, respectively during validation. Therefore, the error between the observed and simulated flows is relatively lower for the Indian summer monsoon season (June-September) for all the rainfall data during both calibration and validation periods. The model performance indicators, such as R2, CE, VE and KGE obtained for 0.25° × 0.25° data are 0.81, 0.79, 0.8 and 0.77, respectively during calibration and 0.71, 0.69, 0.76 and 0.68 during validation for the Indian summer monsoon season. The 0.5° × 0.5° data also show comparable model performance during the Indian summer monsoon season. The Indian summer monsoon contributes most of the annual rainfall of the basin, followed by the post-monsoon and pre-monsoon seasons. The variability of rainfall among the grids is relatively lower during the Indian summer monsoon and post-monsoon, while the non-monsoon season rainfall is characterized by significant spatial variability (Table 1). The convectional and frontal mechanisms produce storm events characterized by rainfall properties that vary significantly across space and time. This is underpinned by the observation of Watts & Calver (1991), who suggest that rainfall data at finer spatial scales are required for modelling convective rainfall events, whereas low-resolution rainfall data would be sufficient for modelling spatially and temporally more homogenous stratiform events. Loritz et al. (2021) also noted that the use of distributed rainfall in distributed models yields improved performances only during certain periods. While comparing eight gridded rainfall products, Duan et al. (2016) noted that the errors associated with spatial and temporal distribution of precipitation are significantly higher in winter months (December-February), where precipitation is low. Hence, the relatively larger error values during the non-monsoon season are the reflection of the lack of competency of these rainfall products (particularly at coarser spatial scales) to represent the rainfall variability at the watershed scale. Thus, the variability in the error values among the different seasons shows that the spatially distributed rainfall at finer spatial scales may be required to model the streamflow response of the catchment during the non-monsoon season.
Evaluation matrices based on seasonal flow during calibration period
. | Winter . | Pre-monsoon . | Indian summer monsoon . | Post-monsoon . | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | |
NRMSE | 1.85 | 0.99 | 0.53 | 2.64 | 1.89 | 2.44 | 1.69 | 3.85 | 0.30 | 0.26 | 0.21 | 0.45 | 0.45 | 0.41 | 0.38 | 0.55 |
R2 | 0.52 | 0.74 | 0.72 | 0.28 | 0.51 | 0.56 | 0.55 | 0.55 | 0.74 | 0.81 | 0.75 | 0.47 | 0.67 | 0.87 | 0.77 | 0.53 |
CE | 0.49 | 0.67 | 0.71 | 0.15 | 0.50 | 0.52 | 0.51 | 0.50 | 0.71 | 0.79 | 0.73 | 0.30 | 0.57 | 0.77 | 0.71 | 0.41 |
VE | 0.33 | 0.44 | 0.58 | 0.10 | 0.55 | 0.47 | 0.56 | 0.45 | 0.76 | 0.80 | 0.77 | 0.63 | 0.63 | 0.69 | 0.71 | 0.61 |
KGE | 0.38 | 0.40 | 0.38 | 0.18 | 0.47 | 0.52 | 0.52 | 0.50 | 0.71 | 0.77 | 0.75 | 0.61 | 0.70 | 0.68 | 0.79 | 0.71 |
. | Winter . | Pre-monsoon . | Indian summer monsoon . | Post-monsoon . | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | |
NRMSE | 1.85 | 0.99 | 0.53 | 2.64 | 1.89 | 2.44 | 1.69 | 3.85 | 0.30 | 0.26 | 0.21 | 0.45 | 0.45 | 0.41 | 0.38 | 0.55 |
R2 | 0.52 | 0.74 | 0.72 | 0.28 | 0.51 | 0.56 | 0.55 | 0.55 | 0.74 | 0.81 | 0.75 | 0.47 | 0.67 | 0.87 | 0.77 | 0.53 |
CE | 0.49 | 0.67 | 0.71 | 0.15 | 0.50 | 0.52 | 0.51 | 0.50 | 0.71 | 0.79 | 0.73 | 0.30 | 0.57 | 0.77 | 0.71 | 0.41 |
VE | 0.33 | 0.44 | 0.58 | 0.10 | 0.55 | 0.47 | 0.56 | 0.45 | 0.76 | 0.80 | 0.77 | 0.63 | 0.63 | 0.69 | 0.71 | 0.61 |
KGE | 0.38 | 0.40 | 0.38 | 0.18 | 0.47 | 0.52 | 0.52 | 0.50 | 0.71 | 0.77 | 0.75 | 0.61 | 0.70 | 0.68 | 0.79 | 0.71 |
Evaluation matrices based on seasonal flow during validation period
. | Winter . | Pre-monsoon . | Indian summer monsoon . | Post-monsoon . | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | |
NRMSE | 0.72 | 0.77 | 0.65 | 1.50 | 2.60 | 2.72 | 1.52 | 4.89 | 0.36 | 0.31 | 0.29 | 0.76 | 0.39 | 0.34 | 0.31 | 0.95 |
R2 | 0.50 | 0.73 | 0.71 | 0.27 | 0.48 | 0.55 | 0.51 | 0.22 | 0.64 | 0.71 | 0.71 | 0.25 | 0.65 | 0.82 | 0.75 | 0.20 |
CE | 0.42 | 0.65 | 0.68 | 0.18 | 0.46 | 0.50 | 0.50 | 0.16 | 0.57 | 0.69 | 0.67 | 0.13 | 0.52 | 0.70 | 0.69 | 0.18 |
VE | 0.32 | 0.42 | 0.50 | 0.15 | 0.50 | 0.45 | 0.52 | 0.17 | 0.72 | 0.76 | 0.73 | 0.40 | 0.60 | 0.61 | 0.65 | 0.29 |
KGE | 0.37 | 0.38 | 0.32 | 0.16 | 0.46 | 0.51 | 0.45 | 0.13 | 0.54 | 0.68 | 0.66 | 0.25 | 0.65 | 0.64 | 0.73 | 0.32 |
. | Winter . | Pre-monsoon . | Indian summer monsoon . | Post-monsoon . | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | Point . | 0.25 × 0.25 . | 0.5 × 0.5 . | 1 × 1 . | |
NRMSE | 0.72 | 0.77 | 0.65 | 1.50 | 2.60 | 2.72 | 1.52 | 4.89 | 0.36 | 0.31 | 0.29 | 0.76 | 0.39 | 0.34 | 0.31 | 0.95 |
R2 | 0.50 | 0.73 | 0.71 | 0.27 | 0.48 | 0.55 | 0.51 | 0.22 | 0.64 | 0.71 | 0.71 | 0.25 | 0.65 | 0.82 | 0.75 | 0.20 |
CE | 0.42 | 0.65 | 0.68 | 0.18 | 0.46 | 0.50 | 0.50 | 0.16 | 0.57 | 0.69 | 0.67 | 0.13 | 0.52 | 0.70 | 0.69 | 0.18 |
VE | 0.32 | 0.42 | 0.50 | 0.15 | 0.50 | 0.45 | 0.52 | 0.17 | 0.72 | 0.76 | 0.73 | 0.40 | 0.60 | 0.61 | 0.65 | 0.29 |
KGE | 0.37 | 0.38 | 0.32 | 0.16 | 0.46 | 0.51 | 0.45 | 0.13 | 0.54 | 0.68 | 0.66 | 0.25 | 0.65 | 0.64 | 0.73 | 0.32 |
DISCUSSION
The present study evaluated the different IMD gridded rainfall datasets (of varying spatial resolutions) to assess the variability in the performance of the SWAT model in a humid tropical watershed of the southern Western Ghats. The critical question addressed in this study is how significant the spatial resolution of the IMD gridded rainfall data is for modelling watershed hydrology. The results of the hydrologic simulation in SWAT show notable variability in the watershed response with varying spatial resolution of the IMD precipitation data products (i.e., 1° × 1°, 0.5° × 0.5°, 0.25° × 0.25° and point observation), which is manifested as the variability in the model performances. In addition, the optimal parameter values of most of the sensitive parameters and water balance components were changed with respect to the different IMD rainfall products used in this study. Moriasi & Starks (2010) observed wide variability in the optimal parameter values for the sensitive parameters while using different rainfall products. The changes in the optimal parameter values were insignificant in some cases, while in other cases the range of values of certain parameters (e.g., ESCO, EPCO) vary between the upper and lower limits allowed by the model. The variability in the optimal parameter values in the CRB, indicating the differences in the hydrological processes/mechanisms, is also reflected in the water balance components generated for different rainfall data products. This variability is evident in the components representing the subsurface and groundwater hydrology (e.g., water percolation, lateral flow, groundwater contribution to streamflow and aquifer recharge). According to Pang et al. (2020), the differences in the water balance components could be related to the parameter adjustment as well as the variabilities in the rainfall characteristics, such as rainfall intensity. Muche et al. (2020) observed that rainfall data characteristics affect both sensitive parameters and their corresponding ranges of values. Tuo et al. (2016) also suggest that the rainfall inputs affect the best estimate of the parameters and their uncertainty range and are watershed-specific.
Apart from the differences in the rainfall pattern, the variability in the model performances for different datasets is also the result of the choice of the precipitation data (in the SWAT model), wherein the SWAT uses the rainfall data from a single precipitation grid (or gauging station), which is located to the centroid of each sub-basin, without considering the spatial heterogeneity. Although multiple numbers of precipitation gauging stations are located within a given sub-basin, the SWAT uses data from a single grid. Hence, the accuracy of the simulation depends on how well the grid, which is being used by SWAT, represents the precipitation pattern across the sub-basin. This is a critical issue in watersheds characterized by a strong orographic gradient and watersheds developed in rain shadow regions. The grade of orographic precipitation (across the Western Ghats) is influenced by the variability of the topography, altitude as well as steepness of the mountain barrier (Gunnell 1997; Tawde & Singh 2015). The catchments developed across the rain shadow region are characterized by upstream areas with significantly higher amounts of rainfall, which decreases downstream (Thomas 2012). Similarly, rainfall pattern at the watershed scale is also influenced by the orographic gradient and complexity. In such cases, the selection of the rainfall grid or station nearest to the centroid may generate results that deviate significantly from the observed flows.
In CRB, the meteorological forcing using the 0.25° × 0.25° and 0.5° × 0.5° rainfall data results in improved predictive capability compared to the gridded data with coarser resolution (1° × 1°) as well as scattered point observation (rain gauge) data. Given the results of this study, it is inferred that incorporation of the spatial variability of rainfall at finer scales can improve the predictive capability of hydrological models (e.g., Woods & Sivapalan 1999; Moriasi & Starks 2010; Euser et al. 2015). It is demonstrated that the uncertainty in discharge simulations of a small rural lowland catchment (6.5 km2) in the Netherlands would be reduced from 20% to 10% as the rain gauge increased four times (Terink et al. 2018). On the other hand, it is also demonstrated that the hydrological systems efficiently dissipate the spatial gradients across watersheds (Obled et al. 1994). Similar findings were reported by various researchers (e.g., Reed et al. 2004; Das et al. 2008), where they demonstrated that fine-resolution rainfall data in space and time does not always yield better model performance. It is also illustrated that the impact of the spatial resolution of rainfall data is reduced with increasing catchment size (Fu et al. 2011). The SWAT-simulated streamflow in the CRB shows comparable performances while using 0.5° × 0.5° and 0.25° × 0.25° rainfall data, which is possibly a result of (1) the more or less similar rain gauge network density between the datasets and (2) the spatial discretization threshold used for sub-watershed delineation. Development of the 0.5° × 0.5° gridded product used rainfall data from 6,076 rain gauge stations across India (Rajeevan & Bhate 2009), and 0.25° × 0.25° data were generated from rainfall data from 6,955 stations (Pai et al. 2014). On the other hand, the 1° × 1° gridded rainfall data was developed using a fixed network of 2,140 rain gauge stations (Rajeevan et al. 2006). Given the characteristic selection of the rainfall grid in the SWAT model, the spatial discretization (or the number of sub-watersheds) might also have a critical role as the variability among the rainfall grids may not be captured efficiently in SWAT unless the size of the sub-watershed is optimal enough to represent the variability of rainfall pattern. Rouhani et al. (2009) suggested that the streamflow is moderate to slightly sensitive to the sub-watershed delineation, and many researchers illustrated marked variation in the water balance components except for runoff. For instance, Tripathi et al. (2006) observed insignificant variability of runoff among different watershed subdivision schemes, while variations were significant for evapotranspiration, percolation and soil water content. Jha et al. (2004) also demonstrated that variation in the number of sub-watersheds had very little effect on streamflow, while significant for sediment and nutrient fluxes. While Jha et al. (2004) showed that a decrease in sub-watershed size below a certain threshold does not improve the simulations, Gong et al. (2010) observed better simulation accuracy at moderate thresholds than the simulation under the finest delineation scheme.
Hence, the findings of this study underscore the significance of a cautious approach for the selection of the spatial resolution of the IMD rainfall data for watershed hydrological modelling using SWAT.
CONCLUSION
This study assessed the bearing of the input rainfall data in different spatial resolutions on the SWAT hydrological model simulations. The spatial rainfall patterns of the watershed (CRB in the southern Western Ghats, Kerala, India) were consistent across all the IMD datasets, except for locations where orographic rainfall is dominant. However, the spatial resolution of the data exhibited variability in the seasonal rainfall. The input rainfall to the SWAT model with varying spatial resolution resulted in different values for the optimal parameters. The simulated values of water balance components, particularly those representing the subsurface and groundwater hydrological components (e.g., water percolation, lateral flow, groundwater contribution to streamflow, and aquifer recharge) were largely influenced by the variability in parameters. Finer resolution of rainfall data (0.5° × 0.5°, and 0.25° × 0.25°) resulted in better SWAT model performance (NSE = 0.75 and 0.70, respectively) as compared to coarse resolution (1° × 1°) data (NSE = 0.21). However, the advantage of using a fine-resolution rainfall dataset in the SWAT model is feasible only when the spatial discretization of the sub-watersheds captures the spatial variability of rainfall, but at the cost of increased time for simulations. Therefore, we recommend a trade-off between the spatial resolution of the input rainfall data and the spatial discretization threshold for optimizing the SWAT modelling framework. The results of this study highlight the significance of a cautious approach to the selection of the spatial resolution of the IMD precipitation data for watershed hydrological modelling using SWAT. Although the significance of the selection of the spatial resolution of rainfall data for watershed hydrological modelling is demonstrated using the IMD gridded data, it applies to other datasets as well.
AUTHOR CONTRIBUTIONS
S. Senan: Data Curation, Formal Analysis, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review and editing. J. Thomas: Conceptualization, Methodology, Validation, Visualization, Writing – review and editing. V. K. Vema: Writing – review and editing. P. J. Jainet: Conceptualization, Methodology. S. Nizar: Visualization, Writing – review and editing. S. Sivan: Visualization, Writing – review and editing. K. P. Sudheer: Conceptualization, Methodology, Funding Acquisition, Supervision Writing – review and editing.
ACKNOWLEDGEMENTS
This work was conducted under the DST project ‘Setting up of State Climate Change Knowledge Cell Under National Mission’ at the Institute for Climate Change Studies, Kottayam. The authors acknowledge funding from the Department of Science and Technology (DST) (sanction number DST/SPLICE/CCP/NMSKCC/PR-62/2016 (G)), the Government of India.
FUNDING
This work was supported by the Department of Science and Technology (DST) (sanction number DST/SPLICE/CCP/NMSKCC/PR-62/2016 (G)), the Government of India.
DATA AVAILABILITY STATEMENT
Data cannot be made publicly available; readers should contact the corresponding author for details.
CONFLICT OF INTEREST
The authors declare there is no conflict.