The objective is to assess the use of the Climate Prediction Center morphing method (CMORPH) (~0.073° latitude–longitude, 30 min resolution) rainfall product as input to the physics-based fully distributed Gridded Surface–Subsurface Hydrologic Analysis (GSSHA) model for streamflow simulation in the small (21.4 km2) Hortonian watershed of the Goodwin Creek experimental watershed located in northern Mississippi. Calibration is performed in two different ways: using rainfall data from a dense network of 30 gauges as input, and using CMORPH rainfall data as input. The study period covers 4 years, during which there were 24 events, each with peak flow rate higher than 0.5 m3 s−1. Streamflow simulations using CMORPH rainfall are compared against observed streamflows and streamflow simulations using rainfall from a dense rain gauge network. Results show that the CMORPH simulations captured all 24 events. The CMORPH simulations have comparable performance with gauge simulations, which is striking given the significant differences in the spatial scale between the rain gauge network and CMORPH. This study concludes that CMORPH rainfall products have potential value for streamflow simulation in such small watersheds. Overall, the performance of CMORPH-driven simulations increases when the model is calibrated with CMORPH data than when the model is calibrated with rain gauge data.
Satellite precipitation products are increasingly becoming available at high resolutions that are appropriate for hydrological applications. However, given the large errors in high-resolution satellite precipitation estimates (e.g., Hong et al. 2007; Villarini and Krajewski 2007; Ebert et al. 2007; Tian et al. 2007; Habib et al. 2009; Zeweldi and Gebremichael 2009a,b), questions remain with regard to the capabilities and limitations of satellite precipitation estimates for hydrological applications. Previous studies have shown that, for very large basins (>50,000 km2), hydrologic simulations driven by satellite precipitation input perform well in reproducing the occurrence of daily flood events and magnitude of low flows, although they might contain biases in peak flows (Yilmaz et al. 2005; Su et al. 2008; Collischonn et al. 2008; Artan et al. 2007; Hughes et al. 2006; Hong et al. 2006; Wilk et al. 2006). These studies have also shown that at such large scales, hydrologic model simulations driven by the existing rain gauge network, albeit sparse in some regions, perform well. However, studies on the capabilities and limitations of satellite precipitation products for hydrological applications in small watersheds are still lacking.
The objectives of this study were to assess the use of high-resolution satellite precipitation product as input to a hydrologic model for streamflow simulation and to assess the impact of model calibration strategy (model calibrated with rain gauge input data versus model calibrated with satellite precipitation input data). The study region is the Goodwin Creek experimental watershed (GCEW), with an area of 21.4 km2, located in northern Mississippi (Fig. 1). This watershed was selected for this study because of its dense instrumentation network: a rain gauge network (1.4 stations per km2), a stream gauge network (0.65 flumes per km2), one meteorological station, and two soil climate analysis networks. The satellite precipitation product selected is the Climate Prediction Center morphing method (CMORPH; Joyce et al. 2004), which is available every 30 min at a spatial resolution of 0.073° across the globe (60°N–60°S) since December 2002. The hydrologic model selected was the physically based, fully distributed Gridded Surface–Subsurface Hydrologic Analysis (GSSHA) model, which was shown to have adequate performance in simulating streamflow for this watershed (Downer and Ogden 2004).
2. Data and method
a. Study area
The GCEW is an agricultural watershed located in northern Mississippi (Fig. 1). Its watershed area is 21.4 km2, with an outlet located at latitude 34°13′55″N and longitude 89°54′55″W. A detailed description of the watershed is provided in Alonso and Binger (2000). Land use consists of pasture (44%), forest (27%), gullied land (15%), and active cultivation (14%) (Blackmarr 1995). Soil texture consists of silt loam (80%), clay loam (19%), and sand (1%). Elevation ranges from 71 to 127 m. Channels are incised 2–3 m, and the groundwater table is several meters below the ground. Climate is characterized by long, hot, humid summers, and short, mild winters, with a mean annual temperature of 18°C and a mean annual precipitation of 1450 mm (Alonso and Binger 2000). The GCEW has been operated as an experimental watershed by the U.S. Department of Agriculture–Agricultural Research Service (USDA–ARS) National Sedimentation Laboratory (NSL), and is well instrumented for hydrological studies. A layout of the network of 30 rain gauges in the watershed is shown in Fig. 1.
b. CMORPH precipitation estimates
A number of satellite precipitation products result from different approaches that combine the more accurate (but infrequent) microwave (MW) with the more frequent (but indirect) infrared (IR) to take advantage of the complementary strengths. For detailed information on different kinds of high-resolution satellite precipitation algorithms, we refer interested readers to a comprehensive book edited by Gebremichael and Hossain (2009). In this study, we used the widely used high-resolution CMORPH satellite rainfall product. CMORPH (Joyce et al. 2004) employs an approach in which IR data are used only to derive a cloud motion field that is subsequently used to propagate raining pixels; thus, only rainfall estimates that have been derived from microwave data are used in the procedure. CMORPH is a global (in longitude and 60°N–60°S) rainfall product with resolutions of 0.073° and 30 min.
c. GSSHA hydrologic model
GSSHA is a physically based, distributed-parameter, structured-grid, hydrologic model that simulates the hydrologic response of a watershed subject to given hydrometeorological inputs (Downer et al. 2005). The watershed is divided into cells that constitute a uniform finite difference grid. Processes that occur before, during, and after a rainfall event are calculated for each grid cell and then the responses from individual grid cells are integrated to produce the watershed response. Major components of the model include precipitation distribution, precipitation interception, infiltration, evapotranspiration, surface water retention, surface runoff routing, channel flow routing, unsaturated zone modeling, and saturated groundwater flow.
During an event, rainfall is spatially and temporally distributed over the watershed. Rainfall may be intercepted by vegetation before reaching the land surface. Once an initial interception demand is reached, a fraction of the precipitation will reach the land surface. Upon reaching the land surface, precipitation may infiltrate as a result of gravity and capillary forces. Water remaining on the land surface may run off as two-dimensional (2D) overland flow after a specified retention depth representing microtopography has been reached. This water may eventually enter a stream and be routed to the watershed outlet as one-dimensional (1D) channelized flow. Between precipitation events, soil moisture accounting, evapotranspiration, and 2D lateral groundwater flow may be occurring.
In this study, the current version of GSSHA (5.0) was applied at the GCEW. This watershed had been previously simulated with an earlier version of GSSHA (Downer and Ogden 2004). Both versions showed considerable skill in predicting streamflow. In this study, the GCEW model consisted of an overland flow grid with 125-m resolution and 1357 cells. The model was implemented with a time step of one minute. Streamflow was routed in a series of channel segments, nonorthogonal to the stream. Infiltration was simulated using the Green and Ampt with Redistribution (GAR) method (Ogden and Saghafian 1997) and soil moistures were simulated using a two-layer soil moisture accounting method (Downer 2007). The Penman–Montieth method was used to calculate potential evapotranspiration (ET), which was used to calculate actual ET based on soil moisture and root depth. Distributed parameters for overland flow, infiltration, ET, and soil moisture accounting were developed from index maps derived from land use/cover and soil type information described by Blackmarr (1995); soil maps were obtained from the USDA–NRCS soil surveys, and land use and cover maps were obtained from ground surveys and satellite imagery. Initial values of parameters were developed from observed data and literature values. Sensitive parameters (overland roughness, channel roughness, soil hydraulic conductivity, surface retention depth, soil moisture depth from which ET occurs, and top soil layer depth for the GAR method) were calibrated using the Shuffled Complex Evolution (SCE) method (Duan et al. 1992), and the objective function used was minimizing the weighted sum of the relative mean absolute errors in peak discharge and total volume (see Downer and Ogden 2004 for further information). The model was calibrated to outlet streamflow data for the year 2003. We performed model calibration using rainfall input data from rain gauges and CMORPH separately, resulting in two sets of calibration parameter values, which allowed us to assess the impacts of rainfall data source on model calibration and validation.
Our study period covers the entire years of 2003, and 2005 through 2007 (we excluded 2004 because of a lack of streamflow observations). During this period, there were 24 storm events that produced peak streamflow rates greater than 0.5 m3 s−1, which Senarath et al. (2000) estimated as the smallest value that allows accurate estimation of event volumes at GCEW. We used the events in 2003 for calibration, and the events in 2005 through 2007 for validation. Our approach had the following main steps: 1) calibration of GSSHA using rain gauge data, simulation of streamflow from GSSHA using rain gauge measurements and CMORPH rainfall estimates (separately) as inputs, and comparison of the simulations with observed streamflow; 2) recalibration of GSSHA using CMORPH rainfall data, simulation of streamflow from GSSHA using CMORPH rainfall estimates as inputs, and comparison of the simulations with observed streamflow; and 3) comparison of the performance of CMORPH streamflow simulations obtained from the two calibration approaches.
e. Performance statistics
We conducted model performance evaluations based on a visual inspection of simulated and observed streamflow hydrographs and statistical comparisons of the three hydrograph parameters (time to peak, peak flow rate, and event runoff depth). For the statistical comparisons between observed (OBS) and simulated (SIM) hydrograph parameters, we used correlation coefficient (R), bias ratio, and normalized root-mean-square error (NRMSE), as defined below:
where E(·), V(·), and Cov(·) are the expectation, variance, and covariance operators, respectively. In practical applications, sample estimates are used. The correlation coefficient R varies from −1 (perfectly negative correlation) to 1 (perfectly positive correlation), with 0 indicating no correlation, and R is not affected by bias. The ideal bias ratio is 1. NRMSE measures the average degree of agreement between observed and simulated hydrograph parameters, and the ideal NRMSE is zero.
3. Results and discussions
a. Streamflow simulations set A: GSSHA calibrated with rain gauge data
The purpose of these simulations was to assess the effects of CMORPH rainfall estimates on simulated streamflow hydrographs when GSSHA model was calibrated with rain gauge data. We calibrated GSSHA model parameters with rain gauge data using nine events in 2003. Model calibration parameter estimates are given in Table 1. Comparisons between observed and simulated (when the model was forced by rain gauge rainfall) are shown in Fig. 2. The simulations reproduced well the shapes of observed hydrographs.
We forced the rain gauge–calibrated GSSHA with rainfall inputs from gauges and CMORPH, separately, and simulated streamflow for 15 events that took place between 2005 and 2007. Comparisons of observed and simulated streamflow are given in Fig. 3. Also shown are corresponding watershed-average rainfall values derived from gauge measurements and CMORPH estimates. The steep gradients in the rising and receding limbs of the event hydrographs indicate a quick response of the watershed, suggesting that the runoff in the GCEW was mainly due to direct surface runoff contribution. This indicates that the surface runoff was dominated by a Hortonian mechanism as reported in Senarath et al. (2000). CMORPH simulations captured all the observed events. To measure the performance of the simulations, we compared the hydrograph parameters of simulated and observed streamflow: time to peak, peak flow rate, and event runoff depth. Figure 4 presents the comparisons in the form of scatterplots and performance statistics, Fig. 5 provides the frequency at which each hydrograph parameter was correctly or incorrectly estimated, and Table 2 gives the hydrograph parameter values and other relevant information. The results are discussed for each hydrograph parameter as follows.
1) Time to peak (tp)
As can be seen from Table 2, tp for observed hydrographs ranged from 1.0 to 12.0 h with an average value of 5.15 h; for simulated hydrographs, tp ranged from 1.5 to 17.0 h with an average value of 5.78 h when forced by CMORPH data, and from 1.0 to 11.5 h with an average value of 4.87 h when forced by rain gauge data. From the maximum tp values, one can see that CMORPH gave longer than observed time to peak values for some of the events.
Inspection of Fig. 4a indicates that, on average, CMORPH tp values were higher by 12% than the observations. There was also a significant disagreement in the event-to-event tp fluctuation: correlation between CMORPH simulation and observed tp values was significant but low (R = 0.60), and showed slight improvement (R = 0.75) when the simulations were forced by rain gauges.
Figure 5a presents the number of cases of tp values that were correctly or incorrectly estimated by CMORPH and rain gauge simulations. We assume that if the error in simulated tp is within one hour, the simulated tp values are acceptable. For about one-third of the storms, CMORPH gave acceptable tp values, and for the remaining storms CMORPH gave significantly longer or shorter time to peaks. Rain gauge simulations, on the other hand, gave acceptable tp values for about half of the storms.
2) Peak flow rate (Qp)
As can be seen from Table 2, the event of 5 April 2005 had the highest observed Qp at 62.2 m3 s−1, which was substantially underestimated by both CMORPH (31.8 m3 s−1) and rain gauge (20.6 m3 s−1) simulations. The second, third, and fourth highest observed Qp during the validation period were also underestimated by both CMORPH and gauge simulations. So as far as the extreme events go, both CMORPH and gauge simulations underestimated peak flow rates, with CMORPH showing slightly better performance. The event of 13 April 2007 had a small observed Qp at 3.9 m3 s−1, which was alarmingly overestimated by CMORPH (71.9 m3 s−1) but almost correctly estimated by gauges (3.2 m3 s−1). Similar alarming overestimations by CMORPH (while gauges were giving reasonably accurate values) were reported for the events of 11 October 2006, 23 February 2007, and 7 December 2007. This reveals that the performance of CMORPH varied tremendously from one event to the other while rain gauges had relatively more consistent performance, indicating the presence of large random errors in CMORPH rainfall estimates.
The above statement is further corroborated by Fig. 4b. The correlation between observed and CMORPH-simulated Qp was low at R = 0.47, indicating that only 22% of the variability in observed Qp was explained by CMORPH simulations. The correlation was better (R = 0.74) for observed and rain gauge–simulated Qp.
We assume that if the error in simulated Qp is within 10%, the simulated Qp values are acceptable. Figure 5b reveals that CMORPH simulations overestimated Qp for 47% of the events and underestimated Qp for 53% of the events. The performance is slightly better but still low for rain gauge simulations: overestimation of Qp for 27% of the events, underestimation for 60% of the events, and acceptable estimation for 13% of the events.
3) Event runoff depth (Vd)
The accuracy of CMORPH Vd estimates varied substantially from event to event, overestimating some, underestimating others, and occasionally yielding reasonably accurate values (Table 2). The correlation between CMORPH-simulated and observed Vd was low only at 0.48, while the correlation was higher (R = 0.73) between rain gauge–simulated and observed Vd (Fig. 4c). Assuming that errors within 10% in simulated Vd are acceptable, CMORPH simulations gave acceptable estimations of Vd for only 7% of the events, while rain gauges gave acceptable estimates for 13% of the events.
b. Streamflow simulations set B: GSSHA calibrated with CMORPH data
The purpose of these simulations was to assess the effects of CMORPH rainfall estimates on simulated streamflow hydrographs when the GSSHA model was calibrated with CMORPH data. We recalibrated GSSHA with CMORPH rainfall input for the calibration period 2003. Model parameter estimates are given in Table 1. We simulated streamflow using CMORPH rainfall input for the validation period 2005–07 and show the results in Fig. 6 and Table 2. The simulations captured all the observed events. Furthermore, the CMORPH-calibrated simulations reduce the substantial overestimation of peak flow observed in the gauge-calibrated simulations (e.g., compare the CMORPH simulation hydrographs for 23 February 2007 and 13 April 2007).
c. Comparison of set A and set B streamflow simulations
Is it better to use GSSHA calibrated with rain gauge or CMORPH data when using CMORPH rainfall estimates as input for streamflow simulation? Fig. 7 compares the model performance statistics of set A (when the model was calibrated with rain gauge data) and set B (when the model was calibrated with CMORPH data) simulations in terms of NRMSE and bias ratio for each hydrograph parameter. For time to peak estimation, the statistics barely improved as one goes from gauge-calibrated to CMORPH-calibrated simulation. For peak flow estimation, the statistics improved as one goes from gauge-calibrated to CMORPH-calibrated simulation [bias changed from +54% (overestimation) to −32% (underestimation), and NRMSE dropped from 220% to 106%, indicating more consistent performance for CMORPH-calibrated simulation]. For event runoff depth estimation, the bias statistics deteriorated (from +26% to −39%) but NRMSE improved (from 158% to 93%) as one goes from gauge-calibrated to CMORPH-calibrated simulation.
In general, the CMORPH-calibrated simulations had relatively more consistent performance than the gauge-calibrated simulations. However, in terms of average statistics such as bias, there was no a clear winner, as this could depend on the objective function used during the calibration exercise. We also point out that any improvement obtained by using CMORPH-calibrated simulation might be due to compensation of some of the errors in satellite rainfall estimates, which may have negative implications on the accuracy of other water balance component estimations such as the groundwater.
We have assessed the utility of CMORPH rainfall products (0.073° × 0.073°, 30 min) as input into the physics-based fully distributed hydrologic model GSSHA for streamflow simulation in a small (21.4 km2) Hortonian watershed (Goodwin Creek in Mississippi). For comparison, we also used rainfall data from a dense network of 30 rain gauges as input into GSSHA for streamflow simulation. The study period covered 2003, and 2005 through 2007. During this period, there were 24 events (9 in 2003 and 15 in 2005–2007), each with peak flow rate higher than 0.5 m3 s−1. We used the events in 2003 for calibration, and the events in 2005–2007 for validation. Calibration was performed in two ways: using rainfall data from a dense network of 30 gauges as input, and using CMORPH rainfall data as input. Comparison of simulated and observed streamflow, during both calibration and validation periods, provided a measure of the performance of the simulations.
The CMORPH simulations captured all the 24 events. The bias in the time to peak estimation was −6% for gauge simulation versus +12% (+6%) for CMORPH simulations when the model was calibrated with rain gauge (CMORPH) data. The bias in the peak flow estimation was −24% for gauge simulation versus +54% (−32%) for CMORPH simulations when the model was calibrated with rain gauge (CMORPH) data. The bias in the runoff depth estimation was −35% for gauge simulation versus +26% (−39%) for CMORPH simulations when the model was calibrated with rain gauge (CMORPH) data. The CMORPH simulations had comparable performance with gauge simulations, which is striking given the significant differences in the spatial scale between the rain gauge network and CMORPH. The spatial resolution of CMORPH is about 8 km, which is too coarse compared to the spatial resolution implemented in the model (125 m). In addition, the watershed falls within three CMORPH pixels and contributes to less than 50% of the area within each pixel.
As far as the effect of GSSHA calibration (i.e., gauge calibrated versus CMORPH calibrated) on CMORPH simulations is concerned, the CMORPH-calibrated simulations had relatively more consistent performance than the gauge-calibrated simulations. However, in terms of average statistics such as bias, there was no clear winner as the CMORPH-calibrated simulations improved the bias in the peak flow but deteriorated the bias in the event runoff depth, partly because of the objective function used in the calibration exercise.
The study was supported by NASA NIP Grant NNX08AR31G and NASA Grant NNX10AG77G to the University of Connecticut.