Evaluation of satellite-based and reanalysis precipitation datasets by hydrologic simulation in the Chenab river basin

Precipitation data is regarded as an essential driving variable in hydrologic models (Sharannya et al. 2020; Jimeno-Sáez et al. 2021). At high altitudes, the scarcity, non-existence or lack of gauged observations coupled with orographic effects and complex weather systems hinders the efforts to quantify current and future water availability (Nazeer et al. 2021). The four basic principal methods for precipitation estimates such as ground-based gauges, ground-based radars, satellites and reanalysis products vary in results because of their limitations (Michaelides et al. 2009). The lack of ground-based gauge stations in rugged topography creates problems in reflecting true variability and assessing the accuracy of areal rainfall (Andréassian et al. 2001). The precipitation estimates by ground-based radars are also restricted because of their limited coverage at a regional scale (Martens et al. 2013). Satellites have higher spatio-temporal coverage but their precipitation estimates are vulnerable to detect rainfall of low intensity, systematic biases and poor performance over snow-covered areas (Mugnai et al. 2013). The large-scale weather systems can be better described by reanalysis products but fail to distinguish spatial variability due to their low spatio-temporal resolution (Kidd et al. 2013). However, these products are a useful alternative to observe precipitation products in data-scarce regions to fill gaps in data and to support in the assessment of water-related issues (Nazeer et al. 2021). In general, higher spatial variability in precipitation is observed in complex topographic regions over short horizontal distances due to orographic effects than plain areas, which must be solved for better planning and management of water resources (Bookhagen & Burbank, 2006; Amiri Conoscenti & Mesgari 2018; Amiri & Gocic 2021a, 2021b). In some countries, long-term records of precipitation are unavailable due to government indifference, lack of resources, political instability and some other reasons (Tan et al. 2021a, 2021b). In recent years, the availability of data from a variety of sources (historical, observed, satellite and radar) has gained importance for hydrological modeling. In the majority of publications reported from Asia (i.e. China and India, 58%) and the USA (14%), the most popular precipitation products used in the SWAT model were CFSR and TRMM as well as PERSIANN, CMADS, APHRODITE, CHIRPS and NEXRAD (Tan et al. 2021a, 2021b).

Several open-source precipitation products (Meng et al. 2019) are available at different spatial resolutions (0.05°×0.05° to 1°×1°) and time scales (hourly, daily and monthly). The most widely used satellite, gridded and reanalysis products include Modern-Era Retrospective Analysis for Research and Applications (MERRA2) version 2 (Rienecker et al. 2011), Precipitation Estimates from Remotely Sensed Information using Artificial Neural Networks-Climate Data Records (PERSIANN-CDR) (Ashouri et al. 2015), Climate Hazards Group Infrared Precipitation with Stations dataset (CHIRPS) (Funk et al. 2014), European Centre for Medium-Range Weather Forecast (ECMWF) Reanalysis-5 (ERA-5) (Muñoz-Sabater et al. 2021) and National Center for Environmental Prediction-Climate Forecast System Reanalysis (NCEP-CFSR) (Saha et al. 2014). These datasets can be categorized into satellite only (TRMM and GPM), satellite adjusted (e.g. PERSIANN-CDR and CHIRPS) and reanalysis (e.g. ERA5 and CFSR) precipitation products (Rees & Collins 2006). The hydrologic performances of satellite and reanalysis data products are mainly assessed in two ways: (1) by comparison between reanalysis and observed data (Sharp et al. 2015); (2) reanalysis dataset as an input data source to hydrological models and outputs are compared with observed streamflow (Tomy & Sumam 2016; Liu et al. 2018).

The hydrological models mainly in practice are: lumped, semi-distributed and fully distributed. For example, Hydrologiska Byråns Vattenbalansavdelning model (HBV), hydrological model (HYMOD), artificial neural network (ANN)-based data-driven hydrological models, GR4 J (Génie Rural à four paramètres Journalier), Snowmelt Runoff Model (SRM), simplified version of the HYDROLOG (SIMHYD) and hydrological TANK model are lumped models. The Soil and Water Assessment Tool (SWAT), Variable Infiltration Capacity (VIC) and Hydrologic Engineering Center-Hydrologic Modeling System (HEC-HMS) are semi-distributed models, and Visualizing Ecosystem Land Management Assessments (VELMA) and Variant of Système Hydrologique Européen (MIKE SHE) are fully distributed models (Devkota & Gyawali 2015; Sitterrson et al. 2017; Pandey et al. 2020). These are either continuous or event-based flow generating models (Hossain Hewa & Wella-Hewage 2019; Marahatta et al. 2021). The SWAT model was selected as the most popular model from a list of 73 hydrological models (Mannschatz et al. 2016). The reasons behind the selection of the SWAT model are its immense application proficiency, user-friendliness, and the fact that it is well promoted and supported in hydrological processes. The SWAT model has the ability to simulate streamflow in data-limited regions (Dile & Srinivasan 2014; Singh & Saravanan 2020). The SWAT model has been extensively applied in river basins around the world for hydrological simulations among many other hydrological models (Das et al. 2019; Tan et al. 2021a, 2021b). Several publications on the use of SWAT for various purposes (e.g. simulating hydrological processes, water management practices, climate change impact studies, land-use change, soil erosion and contaminant transport) are available in the SWAT repository (CARD 2020).

The use of satellite-based and reanalysis precipitation as an alternative to rain gauge precipitation data in hydrological models has been increased in the recent past (Duan et al. 2019; Muche et al. 2020). The type of precipitation data source affects the results from the SWAT model, especially in complex and heterogeneous topography watersheds. However, the observed precipitation data is limited particularly in developing countries and remote areas. Furthermore, even if precipitation data is available, its access to public use is restricted due to strict data policy (Duan et al. 2019). The accuracy of the rain gauge network is also considerably influenced by flaws in instrument installation methods, losses through instrument by wetting inside walls and evaporation, and the wind effects above the gauge orifice (Jimeno-Sáez et al. 2021; Senent-aparicio et al. 2021). In addition to rain gauge data, the availability of alternative sources for precipitation data will enhance the hydrological modeling efforts, particularly in data-scarce regions. Global satellite, gridded and reanalysis datasets are useful alternatives to poorly gauged or ungauged basins for modeling hydrological processes. In order to overcome the scarcity of rain gauge data, many precipitation sources (observed, satellite and radar data) are merged into gridded datasets (Abatzoglou et al. 2018).

Several studies were conducted in data-scarce or ungauged catchments using satellite-based and reanalysis datasets to simulate streamflow with hydrological models worldwide (Tolera et al. 2018; Le et al. 2020; Tan et al. 2021a, 2021b). For example, Wang & Zeng (2012) assessed reanalysis precipitation products versus measured data and concluded that the performance of NCEP-CFSR was the best among studied precipitation products (ERA-40, ERA-Interim, MERRA, GLDAS, NCEP-NCAR-1 and NCEP-CFSR). In another study conducted to evaluate four precipitation products, NCEP-CFSR and ERA-Interim performed better than NCEP-NCAR reanalysis and ERA-40 in the Tibetan Plateau (Bao & Zhang 2012). Funk et al. (2014) applied precipitation data from the NCEP-CFSR to run a hydrological model on five watersheds of various hydro-climate regimes and found that runoff simulations were as good as or better than traditional rain gauge data. In the Blue Nile River Basin, streamflow simulations from three different hydrological models showed that CFSR data has the ability to simulate streamflow quite similar to streamflow simulated with weather station data (Dile & Srinivasan 2014; Worqlul et al. 2017).

In China, CFSR data was used to simulate streamflow in the Bahe River Basin (Hu et al. 2017), the Kaidu River Basin (Tian et al. 2017) and the Kash River Basin (Gao et al. 2018a, 2018b), and the results were satisfactory. A study conducted to assess the hydrologic performance of CFSR using the SWAT model showed good results compared to using local climate data for streamflow simulation (Cuceloglu & Ozturk 2019). In another study, the simulated streamflow with CFSR showed a poor performance compared to using local climate data (Alemayehu et al. 2018). Similarly, input data from CFSR showed a poor performance in the Kash River Basin compared to ERA-Interim using a SWAT hydrological model (Gao et al. 2017).

The rainfall estimates from PERSIANN-CDR have been extensively applied in many studies (Ashouri et al. 2015; Zhu et al. 2016; Liu et al. 2017, 2018; Jimeno-Sáez et al. 2021). The performance of the PERSIANN-CDR precipitation product was poor in the detection of precipitation events and the amount of daily precipitation over Columbia. However, the streamflow simulated with PERSIANN-CDR data as input to the hydrological model was quite similar to the observed flow in the Upper Yangtze and Yellow River Basins of the Tibetan Plateau (Liu et al. 2017).

Numerous studies from CHIRPS precipitation data as input to hydrological models concluded that its performance was generally good in basins across the globe (Zambrano et al. 2017; Baez-Villanueva et al. 2018). The CHIRPS dataset has greater importance among other precipitation products because of its finer resolution (0.05). Tuo et al. (2018) described that CHIRPS data showed a satisfactory performance at a monthly scale in the SWAT model. Duan et al. (2019) assessed the performance of open-source precipitation products (i.e. CHIRPS, CFSR and TRMM) as input data to the SWAT model. The results revealed that CHIRPS yielded the best performance among the precipitation products. In the Upper Blue Nile Basin, CHIRPS outperformed other precipitation products at daily, monthly and seasonal scales (Bayissa et al. 2017). Gao et al. (2018a, 2018b) investigated the performance of CHIRPS and PERSIANN-CDR with in situ measurements from 1983 to 2014 in Xinjiang, China. The results showed that CHIRPS was more accurate with gauge observations than PERSIANN-CDR. Similarly, CHIRPS performed well when compared with PERSIANN-CDR over Chile (Zambrano et al. 2017). Baez-Villanueva et al. (2018) conducted a study using six satellite and reanalysis products and found that CHIRPS performance was superior compared to TRMM 3B42v7, TRMM 3B42RT, CMORPH, PERSIANN-CDR and MSWEPv2 in the Chile and Colombia basins. However, CHIRPS performance was the worst among the six precipitation products over the Yellow River, China (An et al. 2020).

This study evaluates the performance of different precipitation products for hydrological applications in poorly gauged Chenab River Basin (CRB). In addition to the assessment of precipitation products, this study also examines the performance of the SWAT model with satellite-based and reanalysis precipitation products in a mountainous watershed. The results of this study will contribute to the selection of a more accurate precipitation product for streamflow simulation in the CRB. Especially, the goals of this study are: (1) to compare observed with simulated streamflow driven by precipitation products in the SWAT model and (2) to assess the performance of each precipitation product as forcing data to the hydrological model. The novelty of this study lies in the capability of the SWAT model to simulate streamflow in the ungauged catchment to better understand the estimation capabilities of satellite and reanalysis precipitation products. The simulation of flow using freely available precipitation data is yet to be explored in the CRB. ERA5 is a recently released global reanalysis dataset that is yet untested in hydrological modeling across the world (Hersbach et al. 2020). This study would help to identify the most appropriate precipitation product for the hydrological application.

The CRB is located in the foothills and very high mountains of the western Himalayas in the south and north, respectively. About half of the total supplies of water come from the eastern Hindukush, Karakoram and western Himalayas (Winiger et al. 2005). The Chenab River originates in the Kulu and Kangra districts in the western (Punjab) Himalayas in India's Himachal Pradesh. The confluence of two major tributaries, i.e., the Bhaga and the Chandra, formed the River Chenab which flows through the Siwalik Range and continues to enter Pakistan reaching 974 km in length (Luqman et al. 2017). The CRB lies between 73 to 78° E and 32 to 35° N which covers 26,000 km² up to Marala Barrage (Figure 1). The average annual water flow is 918 m³ s⁻¹ with a 20% snowmelt contribution (Singh et al. 1997). The lowest elevation point (235 m) is near the Marala barrage, and the highest (7,103 m) point lies in the snow-covered area (Shahzad et al. 2018). Singh et al. (1995) studied the spatial and seasonal change of precipitation with respect to altitude division. They described that rainfall mainly occurs during monsoon (about 75%) and pre-monsoon (about 65%), while 15–26% of rainfall occurs during winter as snowfall in the Greater Himalayan and Middle Himalayan ranges, respectively. In the outer Himalayan ranges, winter rainfall occurs in the liquid form instead of solid precipitation due to lower altitudes. A fair proportion of the flow is from the snowmelt in the mid or later summer season which is enhanced later by pre-monsoon and monsoon rainfall. This combined effect of snowmelt and seasonal rainfall results in peak flows during June–September in the Chenab catchment (Singh et al. 1997).

In hydrological studies, digital elevation models (DEMs) are often used for watershed delineation, catchment boundaries, stream networks, area slope and aspect. DEM with a resolution of 1 arc-second (∼30 m) is employed to delineate a catchment into sub-basins. The ASTER GDEM V2 dataset was downloaded from the Geospatial Data Cloud site (http://www.gscloud.cn). The impact of DEM resolution (from 5 to 90 m) on watershed delineation results showed that ASTER DEM (30 m) was more appropriate to use especially to save time compared to higher resolution DEM (Buakhao & Kangrang 2016).

Tan et al. (2021a, 2021b) proposed CFSR temperature data with the precipitation data as input to the hydrological model. The key information about the five precipitation products (CFSR, ERA5, CHIRPS, PERSIANN-CDR and MERRA2) is shown in Table 1. The discharge data at the Marala barrage was available from 2015 to 2020; therefore, meteorological data collected from 2010 to 2020 with the initial 5 years were used for the warm-up period in the SWAT model. These precipitation products were used as deriving variables to the SWAT model. Other climate data such as temperature, relative humidity, wind speed and solar radiation were simulated with a weather generator using CFSR data.

According to Tarek et al. (2020), ERA5 has performed much better than previous ERA-Interim in hydrological modeling almost equal in efficiency to the hydrological model that used observed data across North America. ERA5 is the best performing reanalysis product (Gelaro et al. 2017; Hersbach et al. 2020; Jiang et al. 2020; Tarek et al. 2020) among several recently released global atmospheric reanalysis products such as MERRA-2, JRA-5 and CFSR. ERA5 has finer spatial resolution than ERA-Interim, with much improved precipitation and tropospheric representation over space and time. ERA5 has higher quality; a number of output parameters and level of details replaced the ERA-Interim reanalysis data (Uppala et al. 2008). Hersbach et al. (2020) observed that precipitation data of a new reanalysis product (ERA5) performed better than ERA-Interim globally. The increased accuracy of ERA5 over ERA-Interim is documented by some studies for many variables, areas and time scales. For example, Wang et al. (2020) estimated precipitation data of ERA5 and ERA-Interim over the Arctic sea ice. The performance of ERA5 for simulating soil moisture, evaporation and river discharge was improved compared to ERA-Interim (Albergel et al. 2018). However, the performance of ERA5 precipitation data is inconsistent in different regions (Wang et al. 2018; Zhang et al. 2019; Nogueira 2020).

The CFSR is a global third-generation reanalysis product that was designed and implemented for the provision of the best estimates of climatic variables (Saha et al. 2014). The satellite and in situ observations are merged to output product having ∼38 km spatial resolution and hourly time resolution at a global scale. CFSR is derived from the Global Forecast System (Fuka et al. 2013) that was developed by the National Center for Environmental Prediction (NCEP). CFSR data is widely used in hydrological modeling because it is readily available in SWAT format at high spatial and longer time series.

The recent Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA-2) (Gelaro et al. 2017) uses the latest V5 Goddard Earth-observing System Model (GEOS) with a spatial resolution of 50 km and the temporal coverage from 1980 to the near present. It also includes recent upgrades to the original MERRA dataset (Rienecker et al. 2011). Multi-source gauge data are used to improve the accuracy of the MERRA-2 model-generated precipitation which is also employed in CFSR and MERRA-Land products (Reichle et al. 2011). The performance of the MERRA-Land product is much better than the original MERRA to represent soil moisture conditions. The performance reanalysis precipitation products are more reliable and accurate in relatively flatter regions than complex topographic areas with drastic elevation changes (Hamal et al. 2020). However, MERRA2 datasets show more consistent performance in complex topographical regions than other products such as ERA-Interim and CFSR (Chen et al. 2014; Hu et al. 2017).

CHIRPS is open-source data available at daily and monthly scales with a higher spatial resolution of 4.8-km grid (1/20°) and quasi-global coverage (50°S–50°N) from 1981 to the near present (Funk et al. 2015). The latest version of this product can be downloaded from http://chg.geog.ucsb.edu/data/chirps/. The satellite data blended with in situ gauge measurements to generate a time series of gridded precipitation. The databases used to develop CHIRPS data include rainfall observations collected from GHCN and FAO, geostationary thermal infrared satellite observations, the NOAA Climate Forecast System (CFS), monthly rainfall climatology Climate Hazards Group Rainfall Climatology (CHPClim), Tropical Rainfall Measuring Mission (TRMM) B42 rainfall product and the rain gauge stations data from multiple sources (Funk et al. 2015). CHIRPS is categorized as a satellite-gauge dataset and described in detail by Melo et al. (2015).

PERSIANN-CDR satellite-based data was developed by the Center for Hydrometeorology and Remote Sensing (CHRS) at the University of California (Nguyen et al. 2018) using infrared brightness temperature imagery with the gridded satellite by the PERSIANN algorithm. PERSIANN-CDR is accessible with a spatial resolution of 0.25, at a daily scale and quasi-global coverage of 60°S–60°N. The rainfall estimation bias was rectified using the monthly Global Precipitation Climatology Project (GPCP) to increase the reliability of the PERSIANN-CDR data (Ashouri et al. 2015). The data is available in a consistent long-term time series of precipitation datasets which can be helpful for studying the extreme precipitation events. More detailed information about the product is available and accessible at http://chrs.web.uci.edu.

The discharge data during 2015 to 2020 at daily and monthly scales was acquired from the Pakistan Meteorological Department (PMD). The discharge data was used to calibrate (2015–2018) and validate (2019–2020) the SWAT model at the Marala outlet. The selection period is based on the continuous availability of the discharge data.

The land use/land cover (LULC) data was employed to define land use cover in the SWAT model. The dominant classes of LULC in the study area were Urban (0.01%), Agriculture 24.99%), Grassland (26%), Shrubland (24.31%) Deciduous Forest (0.19%), Evergreen Forest (1.03%), Mixed Forest (0.53%), Water (12.62%), Wetland Forest (0.1%) and Barren or Sparsely vegetated (10.14%). The European Space Agency Climate Change Initiative Land Cover (ESA CCI LC) product at 300 m spatial resolution during 1992–2018 is available on the ESA webpage (http://maps.elie.ucl.ac.be/CCI/viewer) (Figure 2).

Besides LULC data, SWAT requires a soil map with soil information on soil properties as input data to SWAT. The world soil map developed by the Food and Agriculture Organization (FAO) was downloaded from http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/faounesco-soil-map-of-the-world/en/ (Figure 2).

The SWAT model was set up using open-source data such as DEM, LULC, soil and precipitation products (Table 2). The CRB was divided into 19 sub-basins with a threshold drainage area of 1,000 ha and three slope classes (0–3, 3–10 and 10–100%). The sub-basins were further divided into 92 HRUs (Hydrological Response Units) by SWAT 2012 interface in ArcGIS 10.5 environment (Figure 2). The topography of the watershed report shows that more than 50% of the catchment area lies above 3,500 m with a mean elevation of 3,182 m. The model was run at daily and monthly scales from 2010 to 2020 for which the initial 5 years (2010–2014) were taken as a warm-up period to lessen the effects of users' estimates of initial state variables.

The SWAT model is an efficient tool for flow simulation which is a challenging task in ungauged catchments especially in developing countries (Swain & Patra 2017). The SWAT model was designed to manage water resources in large catchments by the US Department of Agriculture-Agricultural Research Services (USDA-ARS) (Arnold et al. 1998). The model is employed widely for hydrological modeling in ungauged catchments (Boongaling et al. 2018; Li et al. 2018). The hydrological processes (e.g. lateral flow, infiltration, evapotranspiration and snowmelt) within the SWAT model are simulated on the water balance equation (Neitch et al. 2009). The soil conservation services (SCS) curve method, the kinetic storage model, Manning's equation and the Hargreaves method were used to calculate surface discharge, lateral flow, channel flow rate and velocity, and evapotranspiration, respectively. Furthermore, a detailed description of the SWAT model is available in the SWAT user manual (Neitch et al. 2009).

The prediction uncertainty in SWAT-CUP is quantified by P- and R-factors (Abbaspour et al. 2007). The percentage of observed data bracketed by 95% prediction band is known as 95% prediction uncertainty (95 PPU). The values of 95 PPU ranges from 0 to 1, in which 1 represents 100% enveloping of the observed data by 95 PPU band. R-factor refers to the width of 95 PPU band enveloped by observed data and varies from 0 to 1 (Abbaspour et al. 2015). P- and R-factors are closely tied. The higher P-value can be attained at the expense of a higher thickness of 95 PPU band (R-factor). Therefore, ideal calibration and validation results can be achieved by balancing P- and R-factors (Figure 2). P-factor of >0.7 and R-factor of <1.5 are acceptable for the calibration of discharge (Abbaspour et al. 2015). For instance, the P- and R-factor values were well within the acceptable range (P-factor of >0.7 and R-factor of <1.5) for the ERA5 precipitation-driven calibration and validation (Figure 3).

The sensitivity/uncertainty analysis was conducted in SWAT-CUP with a widely used auto-calibration algorithm (SUFI-2) developed by Abbaspour (2012). The parameter sensitivities are calculated by different parameter combinations using Latin hypercube. Several objective functions (e.g. R², Nash–Sutcliffe efficiency (NSE), χ² and root mean square) have been tested so far to estimate model performance. In this study, the NSE objective function was chosen to estimate model performance, and global sensitivity analysis (GSA) was employed for parameter sensitivity analysis.

For parameter sensitivity, GSA was performed at the outlet of the CRB (i.e. Marala barrage) to identify the most sensitive parameters among the initial parameters. GSA was applied to choose 10 parameters from the initial 16 parameters (Table 3). On the basis of the larger value of t-stat and lower P-values, the most sensitive parameters were identified. V_ refers to the substitution by a value, while R_ refers to the relative change where values are multiplied by one plus a factor from the parameter range (Abbaspour 2012). Ten most sensitive parameters were found including ALPAH_BF, SFTMP, CN₂, CH_N₂, GW_REVAP, SOL_BD, SURLAG, SOL_K, CH_K₂ and GW_DELAY.

In previous studies (Zhu et al. 2016; Gao et al. 2018a, 2018b), two strategies are commonly adopted for model calibration: (1) the SWAT model calibrated separately forced with each precipitation dataset; (2) the hydrological model forced with gauge precipitation data and then the acquired best-fitted parameters are used in the SWAT model to simulate streamflow driven with each precipitation dataset. In an ungauged catchment, the first strategy is essential. In this study, the SWAT model calibrated separately with each precipitation dataset (i.e. CFSR, ERA5, CHIRPS, MERRA2 and PERSIANN-CDR) is used to assess the hydrologic performance of each precipitation product. Some studies showed an improved performance of hydrological models when satellite-based precipitation datasets are employed for calibration (Bitew et al. 2012; Peng et al. 2021).

The SWAT model parameters were calibrated and validated with observed daily and monthly streamflow along with input data from daily and monthly precipitation products. The performance of the SWAT model was determined by the coefficient of determination (R²), the NSE and the percentage bias (PBIAS) (Nash & Sutcliffe 1970). The strength of linear correlation between simulated and observed streamflow is determined by R² and its values range from 0 to 1 (Krause & Boyle 2005). The best simulation is considered close to 1, while its values of >0.5 are acceptable (Moriasi et al. 2007). NSE depicts how close the simulated and observed streamflow match the 1:1 line and ranges from −∞ to 1 (Moriasi et al. 2007). The model performance based on NSE values is classified as unsatisfactory (NSE ≤0.50), satisfactory (0.50≤NSE≤0.65), good (0.65≤NSE≤0.75) and very good (0.75≤NSE≤1.0). The underestimation or overestimation of simulated streamflow is calculated by PBIAS. Under optimal conditions, PBIAS is 0 where positive values indicate overestimation and negative values indicate underestimation (Gupta et al. 1999). The PBIAS is classified as good (±20%), satisfactory (<±40%) and unsatisfactory (>±40%).

The simulation of streamflow with a hydrological model is mainly controlled by precipitation data. Precipitation greatly influences the output of a hydrological model by its volume, spatial and temporal distribution properties. In this study, the capability of five satellite-based and reanalysis precipitation data as input to the SWAT model was evaluated to distinguish the best precipitation product for simulating streamflow in the CRB. In this section, the effect of different precipitation products on simulated streamflow through the SWAT model is quantified and analyzed at daily and monthly scales.

The hydrologic performance of the precipitation products was evaluated using three satellite-based (i.e. PERSIANN-CDR, MERRA2 and CHIPRS) and two reanalysis data (i.e. ERA5 and CFSR) as input to the SWAT model. The performance indices (R², NSE and PBIAS) are used to assess the hydrologic performance of precipitation data using the SWAT model. The calibration and validation of SWAT driven by each precipitation data were accomplished with the best-fitted parameters separately. The evaluation statistics (R², NSE and PBIAS) of the SWAT model using precipitation products during calibration indicate that ERA5 outperformed all other precipitation products at a daily scale (Table 4). The hydrologic performance of ERA5 was good with the highest R² (>0.83), NSE (0.81) and lower PBIAS (−6%). According to guidelines by Moriasi et al. (2007), MERRA2 and CHIRPS yielded unsatisfactory performance (NSE<0.5), while PERSIANN-CDR performance was satisfactory (NSE>0.5) at a daily scale.

Figure 4 illustrates a comparison between observed and simulated daily streamflow using satellite-based and reanalysis precipitation products as input data to the SWAT model during calibration (2015–2018) and validation (2019–2020) periods. The CFSR performance was the second-best with the lowest PBIAS (−3.9%). All satellite-based precipitation products showed the average tendency of considerable underestimation in streamflow simulations. Only the ERA5- and CFSR-driven SWAT model showed PBIAS<±10%, indicating very good performance on average. According to Moriasi et al. (2007), these results indicate that ERA5 and CFSR data-driven SWAT model performance was good at a daily scale in the CRB. The performance of the SWAT model using satellite-based precipitation data was unsatisfactory (PBIAS>30%) during calibration. Based on R², the performance of satellite-based precipitation products was satisfactory (R²>0.5), while the results of NSE (<0.5) showed unsatisfactory performance. Reanalysis precipitation products (i.e. ERA5 and CFSR) outperformed satellite-based precipitation products with higher NSE, R² and the lowest PBIAS. The CHIRPS data as input to the SWAT model showed unsatisfactory performance with the lowest NSE (0.33) and the highest PBIAS (42.4%) during calibration at a daily scale.

The evaluation indices (R², NSE and PBIAS) of ERA5 indicate that ERA5 performance was the best among precipitation products during validation. The performance of the SWAT model using reanalysis precipitation data was very good (PBIAS<±10%) during validation. This implies that both ERA5 and CFSR datasets are suitable for hydrologic simulation. PERSIANN-CDR and MERRA2 have shown relatively lower NSE (<0.6) and higher PBIAS (>30%). Based on R² (>0.65) results, the hydrologic performance of satellite-based precipitation was good during validation.

The hydrograph of simulated flow using reanalysis precipitation data as input to the SWAT model is highly consistent with the observed flow (Figure 4). However, the hydrograph of simulated flow using satellite-based precipitation data indicates that peaks and base flows are underestimated. MERRA2 and PERSIANN-CDR capture peaks and base flows well than CHIRPS. The performance of CHIRPS is the worst (PBIAS>±40) in simulating streamflow in the CRB.

A comparison of simulated streamflow using three satellite-based (i.e. PERSIANN-CDR, MERRA2 and CHIRPS) and two reanalysis (i.e. ERA5 and CFSR) precipitation products and measured streamflow is shown in Figure 5.

The SWAT model simulated monthly streamflow aggregated from daily streamflow was fairly in agreement with the observed streamflow. The evaluation indices (R², NSE and PBIAS) of the SWAT model using input data from ERA5 indicate that ERA5 outperformed all other precipitation products at a monthly scale during the calibration period (Table 4). The monthly simulated streamflow using reanalysis precipitation data (ERA5 and CFSR) showed very good performance with R² (>0.85), NSE (>0.85) and PBIAS (<±10%). The performance of the model using satellite-based precipitation data (PERSIANN-CDR and MERRA2) was good as indicated by the results of R² (0.81 and 0.83, respectively) and satisfactory as shown by NSE (0.65 and 0.54, respectively). The evaluation indices (R², NSE and PBIAS) of the SWAT model using input data from ERA5 show that ERA5 performance was very good among precipitation products during validation. The results of R² (>85) and NSE (>75) show that the performance of the SWAT model using PERSIANN-CDR and MERRA2 was good at a monthly scale. Using CHIRPS data, the SWAT model performance was unsatisfactory (NSE<0.5 and PBIAS>40) during validation.

There was a significant improvement in evaluation indices of the SWAT model using satellite precipitation products at a monthly scale compared to a daily scale, whereas PBIAS values were almost the same at daily and monthly scales. For instance, the increases in R² values using PERSIANN-CDR data for simulating streamflow were from 0.65 to 0.81 at daily to monthly scales, respectively, during calibration and from 0.71 to 0.93 at daily to monthly scales, respectively, during validation. Similarly, the increases in NSE values were from 0.52 to 0.65 at daily to monthly scales, respectively, during calibration and from 0.59 to 0.78 at daily to monthly scales, respectively, during validation.

The hydrologic performance of ERA5 was very good compared to other precipitation products in terms of R² (0.92), NSE (0.91) and PBIAS (−3.3%). The hydrologic performance of CFSR was the second-best based on evaluation indices such as R² (0.81), NSE (0.81) and PBIAS (0.5%). The hydrologic performance of PERSIANN-CDR was good based on R² (>65) and NSE (>0.65), but its performance was satisfactory based on PBIAS (>30%) during the calibration and validation periods. The SWAT model using ERA5 and CFSR data captured the observed streamflow reasonably well at daily and monthly scales. Based on evaluation indices, ERA5 is considered to be perfect for hydrologic simulation in the CRB. These results indicate that the SWAT model can predict discharge in the CRB quite accurately using ERA5 precipitation data.

The SWAT model using CHIRPS data shows extremely large values for PBIAS (>±40), while NSE was the lowest (e.g. 0.42 during calibration and 0.47 during validation). Similarly, MERRA2 and PERSIANN-CDR underestimated peaks but their performance was also satisfactory based on R² (0.73 and 0.81, respectively) and NSE (0.51 and 0.68, respectively). Larger values of PBIAS (>30%) using PERSIANN-CDR and MERRA2 also result in unsatisfactory performance. The differences in the simulated and observed flows can be attributed to many factors: (i) errors in precipitation estimations, (ii) higher spatial variability in the precipitation due to orographic effects, (iii) errors in stream gauge observations and (iv) possible errors in model structure or combinations of the above (Manfreda et al. 2020; Marahatta et al. 2021). These results indicate that satellite precipitation products are suitable for hydrological simulations at a monthly scale in a complex topographic region. However, CHIRPS performance was the worst both at daily and monthly scales.

Satellite-based precipitation data as input to the SWAT model failed to capture heavy rainfall events and underestimated the peaks during monsoon (Himanshu et al. 2018; Kumar & Lakshmi 2018). In general, the performance of satellite-based precipitation products is more reliable in flat terrain than mountainous regions (Derin & Yilmaz 2014; Zhu et al. 2016). Several satellite products performed much better in the plain areas but failed to capture precipitation events successfully in the mountainous counterparts (Kumar & Lakshmi 2018; Prakash et al., 2018; Jena et al. 2020). Similarly, the performance of satellite-based precipitation products in the mountainous region of CRB was poor mainly due to the underestimation of peaks. The peaks in streamflow occur during monsoon when moisture-laden air is lifted by striking Himalayan mountains and results in heavy cloud formation in the CRB. These orographic effects cause high rainfall events during the monsoon season, since satellite products use passive microwave or infrared sensors which mostly fail to detect the orographic change in rainfall over complex topography (Shige et al. 2013). This might be one reason for the underestimation of flow in the mountainous areas using satellite-based precipitation products (Figure 5). Similarly, the TRMM and CHIRPS data underestimated the flow in the Western Ghats, India (Sharannya et al. 2020). In another study, satellite precipitation products underestimated the heavy precipitation events in the Hindukush Himalayan region (Sharannya et al. 2020). According to Jena et al. (2020), CHIRPS and PERSIANN-CDR could only capture 22 and 16.67% of total cloudburst events over northwest Himalaya, respectively. In the CRB, very high flows occur during the monsoon, while low flows occur over the rest of the year. In this study, the hydrologic performance of PERSIANN-CDR and MERRA2 was unsatisfactory, and the performance of CHIRPS was the worst at a daily scale (Figures 4 and 5). The hydrologic performance of CHIRPS over India and other regions is a blend of overestimation and underestimation of flow (Prakash 2019), with a poor performance on a daily basis over West Africa (Dembele & Zwart 2016). The contrasting results of CHIRPS precipitation data over Pakistan were observed during the post-monsoon season (Nawaz et al. 2021). Our results show the poor performance of CHIRPS data as input to the SWAT model at a daily scale. In contrast, Prakash's (2019) study shows a good performance of CHIRPS in monsoon-dominated catchments.

The hydrologic performance of satellite precipitation products was satisfactory based on R² (>60) and NSE (>50) but it was unsatisfactory as indicated by PBIAS (>30%). These results show that although the performance of satellite-based products was good, the underestimation of peak flows makes them unsuitable for streamflow simulation in this region. Our study confirms that PBIAS is a major indicator that determines the accuracy in simulating streamflow. For instance, higher PBIAS values of the SWAT model using the MERRA2, PERSIANN-CDR and CHIRPS data are translated into poor performance of these datasets. The higher biases in satellite precipitation data might be translated into larger PBIAS in streamflow simulation. According to Su et al. (2008), any bias in the input data can be transformed into simulated streamflow. Similarly, previous studies interpreted larger PBIAS as a result of biases in precipitation data in streamflow simulation (Zhu et al. 2016; Musie et al. 2019). The hydrologic performance of satellite precipitation data can be improved if bias is removed from precipitation data (Himanshu et al. 2018).

Several studies have been conducted using satellite and reanalysis precipitation products worldwide (Tuo et al. 2016; Li et al. 2018; Beck et al. 2019; Tarek et al. 2020) but fewer in this region or study basin using the SWAT model (Himanshu et al. 2018; Kumar & Lakshmi 2018; Sharannya et al. 2020). To the best of our knowledge, the study conducted by Ahmed et al. (2020) was perhaps the only study that evaluated the performance of satellite precipitation using the SWAT model in the CRB. The results achieved in this study were more reliable compared to a recent study by Ahmed et al. (2020) in the same basin using satellite precipitation products (e.g. IMERG-F v6 and TRMM 3B42 v7) as input data to the SWAT model. They obtained the best-simulated results in terms of R² (0.86 and 0.89) and NSE (0.77 and 0.82) during calibration and validation, respectively, at a monthly scale. Our study obtained much improved results of R² (0.94 and 0.91 during calibration and validation, respectively) and NSE (0.94 and 0.89 during calibration and validation, respectively) using ERA5 data. These results show that the hydrologic performance using reanalysis data in the studied basin was better than using satellite precipitation data by Ahmed et al. (2020) in the CRB. The calibration results show a higher correlation between observed and simulated streamflow at monthly than daily scales. The evaluation indices showed a higher performance of IMERG-F than 3B42 for simulating streamflow. However, the performance of IMERG-F for simulating peaks was poor at a daily scale. Our study showed a better performance using satellite products at monthly than daily scales, while the ERA5 reanalysis precipitation product showed promising results both at daily and monthly scales. In addition, ERA5 captured peaks perfectly both at daily and monthly scales in the study basin (Figures 4 and 5). The simulated flows using IMERG-F and 3B42 data showed moderate underestimation (22.8% in calibration and 21.5% in validation) and extreme underestimation (30.9% in calibration and 31.1% in validation), respectively. In our study, simulated flows using satellite precipitation data showed higher PBIAS for PERSIAN-CDR (33.5%), CHIRPS (42.4%) and MERRA2 (38.2%) during calibration at a daily scale (Figures 4 and 5). However, biases were quite low in simulated flows using ERA5 (−6%) and CFSR (−3.9%) reanalysis precipitation data. This comparison shows that the hydrologic performance of the reanalysis precipitation products was slightly better than satellite-based precipitation in the CRB.

The peaks and base flows simulated with input data from three satellite-based precipitation products (PERSIANN-CDR, MERRA2 and CHIRPS) were much lower than observed streamflow (Figures 4 and 5). However, peaks and base flows were perfectly captured using reanalysis data (ERA5 and CFSR) during the calibration and validation periods. These results show that high precipitation events during the monsoon period were not accurately simulated using satellite-based products. The curve number technique in the SWAT model fails to accurately predict several storms during 1 day and defines rainfall events as the sum of all-day rainfall (Kim & Lee 2008). The underestimation of the high-flow period might be partly because of the SCS-CN method in the SWAT model. In another study, the hydrologic performance of ERA5-Land data as input data to a precipitation-runoff model was much better in this region (Nazeer et al. 2021). The hydrologic performance of the SWAT model using reanalysis precipitation products (ERA5 and CFSR) can be used for hydrological modeling, not in the data-scarce region but also where the distribution of rainfall stations is not uniform. This study supports the use of reanalysis precipitation products in hydrological modeling, while numerous studies are in favor of using satellite-based rainfall products (Ahmed et al. 2020; Khatakho et al. 2021).

This study evaluated and assessed the hydrologic performance of the five satellite-based and reanalysis precipitation products using the SWAT model in the CRB. The uneven or sparse distribution of ground-based precipitation network is inadequate to project spatial distribution of precipitation which is required for better assessment and planning of water resources through a hydrological model. Therefore, the high spatial and temporal resolution of satellite and reanalysis precipitation products provides a good alternative to a ground-based network. In this study, the SWAT model is forced with selected satellite-based and reanalysis precipitation products in the CRB to examine their performance at daily and monthly scales.

The performance of the ERA5 precipitation product was the best followed by CFSR, PERSIANN-CDR, MERRA2 and CHIRPS for simulating streamflow at daily and monthly scales in the CRB. The hydrologic performance of these precipitation products was ranked higher at monthly than daily scales. The ERA5 can be the best substitute to rain gauge precipitation data as input data to the hydrologic model in sparsely gauged catchments. The performance of CFSR data was the second-best to simulate streamflow. The results of the SWAT model suggest that ERA5 together with CFSR can be a promising alternative to rain gauge data for streamflow simulation in the CRB. The hydrologic performance of the PERSIANN-CDR and MERRA2 was satisfactory based on R² and NSE but unsatisfactory based on PBIAS values. The values of PBIAS were higher from simulated flows with satellite-based precipitation products, while lower PBIAS using the reanalysis precipitation products. The performance of CHIRPS as input data to the hydrologic model was the worst at daily and monthly scales. It was observed that the performance of precipitation data widely differs based on statistical indices. For instance, all precipitation datasets show very good to satisfactory performance based on R², whereas only ERA5 and CFSR data qualify for good performance based on all three evaluation indices (e.g. R², NSE and PBIAS).

This research can be useful in making a selection of precipitation data from a number of open-source precipitation products to simulate streamflow with a hydrological model in similar regions. These results prove that open-source precipitation products are a good source of hydrological simulation in ungauged catchments. A better hydrologic performance using reanalysis precipitation products than satellite-based would encourage water managers to use these precipitation products in this or nearby ungauged catchments for future runoff modeling. This research evaluated only five open-source precipitation products but there are many other freely available precipitation products that should be assessed and evaluated as a potential source of hydrological modeling. We recommend that results can be promising if further studies are conducted based on satellite precipitation products from different parts of the Himalayan region. However, a longer simulation period could add more confidence in the model performance here.

The limitations of this study are included data availability/data gaps, data quality and model limitations. Global precipitation estimates used in the study area have certain limitations in terms of spatial resolution, uncertainty and ability to capture precipitation extremes in complex terrain. Here, CFSR weather data is used and a weather generator in SWAT provides an option for estimating or extending weather variables to other simulation periods with fragmented information. The model can potentially be improved using finer scale LULC and applying parameters determined from field surveys which would take a long time and considerable human resources.

Data cannot be made publicly available; readers should contact the corresponding author for details.