Abstract

Significant concern has been expressed regarding the ability of satellite-based precipitation products such as the National Aeronautics and Space Administration (NASA) Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA) 3B42 products (version 6) and the U.S. National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center’s (CPC) morphing technique (CMORPH) to accurately capture rainfall values over land. Problems exist in terms of bias, false-alarm rate (FAR), and probability of detection (POD), which vary greatly worldwide and over the conterminous United States (CONUS). This paper directly addresses these concerns by developing a methodology that adjusts existing TMPA products utilizing ground-based precipitation data. The approach is not a simple bias adjustment but a three-step process that transforms a satellite precipitation product. Ground-based precipitation is used to develop a filter eliminating FAR in the authors’ adjusted product. The probability distribution function (PDF) of the satellite-based product is adjusted to the PDF of the ground-based product, minimizing bias. Failure of precipitation detection (POD) is addressed by utilizing a ground-based product during these periods in their adjusted product. This methodology has been successfully applied in the hydrological modeling of the San Pedro basin in Arizona for a 3-yr time series, yielding excellent streamflow simulations at a daily time scale. The approach can be applied to any satellite precipitation product (i.e., TRMM 3B42 version 7) and will provide a useful approach to quantifying precipitation in regions with limited ground-based precipitation monitoring.

1. Introduction

Over the last decade, there has been a growing acceptance of remotely sensed products in terms of both monitoring precipitation and for supporting hydrological modeling applications (Artan et al. 2007; Beighley et al. 2009; Collischonn et al. 2008; Hughes et al. 2006; Su et al.2008; Yilmaz et al. 2005). There is an increasing realization that the selection of the type of input precipitation is more important than the choice of the hydrologic model in terms of producing robust hydrologic simulations (Wilk et al. 2006). There are now at least 12 distinct satellite precipitation products available (Ebert et al. 2007). Some of the more commonly utilized products that cover large swaths of the planet include the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN; Hsu et al. 1997), the National Aeronautics and Space Administration (NASA) Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA; Huffman et al. 2007), and the U.S. National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center’s (CPC) morphing technique (CMORPH; Joyce et al. 2004). The primary trend in the development of satellite precipitation products is to move toward merged datasets that take advantage of the best available data, regardless of source. For example, the TMPA 3B42, version 6 product is merged based on the combination of passive microwave (PMW) and infrared (IR) precipitation estimates from all available satellites with the utilization of PMW-calibrated IR-based rainfall estimates to fill PMW coverage gaps. There are two TMPA 3B42, version 6 products, including a research version (TMPA-Research) in which monthly ground rain gauge data are applied to correct for bias and a real-time version (TMPA-RT) without gauge bias correction. The pure intermediate IR-based TRMM 3B41RT product provides a minimal latency (several hours) supporting near-real-time applications, whereas there is up to a 6-week latency in the availability of the TMPA 3B42 research product. The CMORPH is similar to the TRMM products in that CMORPH merges satellite precipitation data from both infrared and microwave sensors. CMORPH is unlike TMPA in that CMORPH data are based on high-resolution IR that is examined to estimate the motion of rainfall patterns to obtain a smooth morphing of PMW rain patterns between PMW snapshots. Additionally, CMORPH is a pure satellite product with no rain gauge correction. Latency for the standard CMORPH product is 20–44 h.

While the promise of remotely sensed satellite products to monitor precipitation in regions of the world with limited ground-based observations is great, this promise has yet to be realized. Satellite products are plagued by problems associated with a lack of precipitation detection, false detection of precipitation, and bias. There has been some work that has focused on addressing the deficiencies of satellite precipitation products, particularly focusing on the adjustment of bias (Boushani et al. 2009; Vila et al. 2009; Wilk et al. 2006). The goal of this paper is to validate a methodology that allows users to adjust satellite products such as TMPA-RT that initially yield poor streamflow simulations and transform these products into a platform that can support robust hydrological modeling. The approach is based on applying ground data [U.S. National Weather Service (NWS) rain gauge; NWS multisensor precipitation estimator (MPE)] to minimize bias and to eliminate the lack of detection and false alarms. This methodology is validated with streamflow simulations based on the Soil and Water Assessment Tool (SWAT) of the San Pedro River basin in Arizona. This study demonstrates that satellite products can be adjusted to produce acceptable streamflow simulations even with sparse ground data when dealing with satellite products at a 0.25° spatial resolution and daily time scale. Consequently, the developed methodology can support near-real-time hydrologic modeling for moderately large (1000s km2) watersheds in regions with limited ground-based precipitation data.

2. Study area

The examined watershed is the San Pedro River basin, located between Tombstone and Benson, Arizona (Fig. 1), and has been previously examined by Serrat-Capdevila et al. (2007). The studied reach of the San Pedro River basin has U.S. Geological Survey (USGS) stream gauges located at the inlet (Tombstone, USGS gauge 09471550) and outlet (Benson, USGS gauge 09471800) of the basin (Fig. 1). This basin is a moderately large watershed—1971 km2 in area. Landscape within the watershed consists of mountains that rim the eastern and western sides of the basin, typical of basin and range topography, with lower elevations in the central valley associated with the San Pedro River (Fig. 1). Land cover consists mostly of rangeland (RNGB; 93%), which is covered by scrub and brush. Minor areas with forest (FRSD; 3%), rangeland–grass (RNGE; 2%), agriculture (AGRL; 1%), urban (URBS; <1%), and wetland (<1%) land covers also exist within the basin. Precipitation in the basin has a bimodal seasonality, with high-intensity summer rain and lower-intensity winter precipitation. Summer rainfall is attributable to monsoonal circulation and moisture produced from dissipating tropical cyclones (Webb and Betancourt 1992). Wintertime precipitation is associated with frontal storm systems originating from the Pacific Ocean. Streamflow response is significant only during the warm season (July–September) and is insignificant during the cool season (October–June) because of the sporadic occurrence of low-intensity precipitation events during this period.

Fig. 1.

The San Pedro basin between Tombstone and Benson, AZ. Watershed map illustrates 29 subbasins with related hydrographic features superimposed on a DEM. Indicated is the spatial distribution of centroids of TRMM grid cells (squares) and 0.25° TRMM grid cells (dashed lines). MPE/TRMM grid cells are labeled as follows: northernmost grid cell 1; middle, west grid cell 2; middle, east grid cell 3; southernmost, west grid cell 4; and southernmost, east grid cell 5.

Fig. 1.

The San Pedro basin between Tombstone and Benson, AZ. Watershed map illustrates 29 subbasins with related hydrographic features superimposed on a DEM. Indicated is the spatial distribution of centroids of TRMM grid cells (squares) and 0.25° TRMM grid cells (dashed lines). MPE/TRMM grid cells are labeled as follows: northernmost grid cell 1; middle, west grid cell 2; middle, east grid cell 3; southernmost, west grid cell 4; and southernmost, east grid cell 5.

3. Data sources

The data sources utilized in this study include digital data for the San Pedro River basin [Hydrologic Unit Code (HUC) 15050202], including a digital elevation model (DEM; USGS 90-m resolution, 3-arc-second elevation model) for the continental United States (CONUS), a hydrography layer (National Hydrography Dataset, high resolution, 1:100 000; USGS 2009a), a soil layer [State Soil Geographic database (STATSGO), 1:250 000; U.S. Department of Agriculture; US EPA 2010], a land use layer [land use/land cover Geographic Information Retrieval and Analysis System (GIRAS) dataset, 1:250 000; USGS 1986], and a USGS stream gauge location with USGS surface runoff data (USGS 2009b) as tab-separated text files.

Daily temperature and rain gauge precipitation data (2003–09) were downloaded from the U.S. National Climatic Data Center (NCDC 2009). The closest NWS station to the examined watershed is the Bisbee–Douglas International Airport, Arizona [Cooperative Observer Program (COOP) ID 022664], which is henceforth referred to as the Douglas rain gauge and has a complete record for 2003–09. Four satellite precipitation products were used in this study. Three TRMM datasets (version 6) were analyzed, including two TRMM-RT products—3B41 and 3B42—and the TMPA-Research version of 3B42 (Huffman et al. 2007). The fourth satellite product examined was CMORPH (Joyce et al. 2004). These products have fine spatial (0.25° × 0.25°) and temporal (3 h; 1 hourly for TRMM 3B41RT) resolutions. TRMM 3B41RT, TMPA-RT, and CMORPH are pure satellite products, whereas TMPA-Research incorporates monthly rain gauge data to correct for bias. TMPA-Research, TMPA-RT, and TRMM 3B41RT data were obtained online through the Goddard Earth Sciences Data and Information Services Center (GES DISC) with the GES DISC Interactive Online Visualization and Analysis Infrastructure (Giovanni; NASA 2010). CMORPH was downloaded from the Climate Prediction Center (CPC 2010). Subdaily data for each TRMM and CMORPH grid cell (Fig. 1) were aggregated into a daily estimate of precipitation (2003–09) with Visual Basic.

A fifth type of precipitation data utilized in this study was the MPE. In the western United States, this product is based on Parameter-elevation Regressions on Independent Slopes Model (PRISM) climatology (Daly et al. 1994), which is used to interpolate between rain gauges and to adjust for orography. This methodology forms the basis for the NWS Mountain Mapper product, which has a 4 km2 and 6-hourly resolution. MPE data were obtained from the NWS Colorado River Forecast Center in an XMRG format, which is a binary format developed by the NWS to store gridded data. XMRG files were processed using C++ (see Xie et al. 2005), producing American Standard Code for Information Interchange (ASCII) text files. Programming in Visual Basic aggregated MPE data into a daily estimate of precipitation at a 0.25° spatial resolution directly comparable to the spatial extent of the TRMM and CMORPH products.

4. Methodology

a. Precipitation data analysis

To better understand how input precipitation can affect hydrologic model results, precipitation data were analyzed. The analysis focused on an intercomparison of precipitation values and quantitative estimates of satellite product performance within the examined study area (Fig. 1). The quantitative estimate of satellite product performance was based with reference to MPE data aggregated to a 0.25° resolution or rain gauge data within TRMM and CMORPH grid cells, and the product evaluation is based on the criteria established by Ebert et al. (2007), which include the probability of detection (POD), false-alarm rate (FAR), and equitable threat score (ETS).

b. Precipitation product adjustment

The primary goal of this paper was to adjust satellite products to eliminate false alarms and missed precipitation events and to correct for bias-facilitating hydrologic modeling supported by satellite precipitation products. To accomplish this goal, a filter was needed that could act as a mechanism to screen for false alarms and missed precipitation events. We developed the filter based on two approaches: 1) utilizing actual ground precipitation data (MPE or rain gauge) and 2) developing the filter by autoregression (AR) analysis of the ground precipitation record based on a Markov chain approach, which has been commonly used to develop synthetic precipitation datasets (Wilks 2006). For both approaches we omitted dates with missing data, 1–2 days during the entire 6-yr time series, and considered values only above the detection limit (MPE and rain gauge, 0.254 mm) as positive for precipitation on a given date. Douglas rain gauge data, which is actually located outside of the basin (Fig. 1), demonstrates the utility of our approach in terms of the spatial transfer of the adjustment methodology to satellite grids that do not correspond with the location of ground-based precipitation monitoring. Consequently, a total of four filters were developed to facilitate satellite product adjustment (MPE-actual, MPE-AR, rain gauge-actual, rain gauge-AR).

Developed filters were utilized to remove false alarms F from satellite products; in the adjusted dataset, the precipitation value is set to zero. Dates in which pure satellite products do not record precipitation were adjusted through substitution of ground precipitation data into the adjusted satellite product. The adjusted TMPA-RT product from the San Pedro basin consisted of 27%–35% substituted ground precipitation data. These two adjustments transformed the satellite product into a product with no false alarms (FAR = 0) or missing dates (POD = 1) and with a perfect ETS = 1 compared with the ground precipitation data.

The final adjustment involved a correction for bias present in the pure satellite products. Our approach to correct for bias was based on the transformation of the probability distribution function (PDF) of the satellite product into the PDF of the ground-based precipitation data. The PDF of both satellite products and ground-based precipitation data were determined through curve fitting with a mixed exponential function, as suggested by Foufoula-Georgiou and Lettenmaier (1987), Wilks (1999), and Woolhier and Roldán (1982). The PDFs of ground (rain gauge, MPE) and satellite (TRMM 3B41RT, TMPA-RT, CMORPH) datasets were determined by curve fitting, with two curves defining an upper and lower domain. Adjustment of the satellite product value (known) into a ground-based value (unknown) was accomplished by simultaneously solving satellite product and ground-based precipitation curve fit equations for a given probability. This process is illustrated in Fig. 2, which plots the PDFs for satellite and ground precipitation data from a single grid cell. As shown in Fig. 2, there were three possible combinations of satellite/ground-based curves that could be used in simultaneous equations to solve for the adjusted value, which included 1) upper-domain satellite/upper-domain ground (light gray); 2) upper satellite/lower ground (medium gray); and 3) lower satellite/lower ground (dark gray). Detection limit (on a daily basis) used to develop lower curves was 0.254 mm for ground-based precipitation data and 1.0 mm for satellite products.

Fig. 2.

Illustration of PDFs adjustment process based on two exponential curves with an upper curve (satellite) and lower curve (ground product). Three combinations of satellite/ground-based curves were used to calculate an adjusted value, which included: 1) upper-domain satellite/upper-domain ground (light gray), 2) upper satellite/lower ground (medium gray), and 3) lower satellite/lower ground (dark gray). Example of how the process can produce an adjusted TMPA-RT value is provided.

Fig. 2.

Illustration of PDFs adjustment process based on two exponential curves with an upper curve (satellite) and lower curve (ground product). Three combinations of satellite/ground-based curves were used to calculate an adjusted value, which included: 1) upper-domain satellite/upper-domain ground (light gray), 2) upper satellite/lower ground (medium gray), and 3) lower satellite/lower ground (dark gray). Example of how the process can produce an adjusted TMPA-RT value is provided.

c. Hydrological model selection and setup

To provide validation for the performance of unadjusted and adjusted satellite precipitation products, simulated streamflow from a hydrologic model was evaluated. The semidistributed hydrologic model selected for this study was the Soil and Water Assessment Tool (SWAT), which is a physically based model with demonstrated global applications, and it has been validated at the watershed scale through the publication of hundreds of referred papers (see Gassman et al. 2007). SWAT is a widely available modeling platform that supports the incorporation of diverse GIS and weather data sources (as described in section 3), which are incorporated into the model, as illustrated in Fig. 3. The major data preparation steps involve watershed delineation, Hydrologic Response Unit (HRU) definition, and the importation of weather and streamflow data (Fig. 3).

Fig. 3.

Conceptual drawing of SWAT model structure.

Fig. 3.

Conceptual drawing of SWAT model structure.

There are six major databases that define the SWAT model (management, basin, HRU, subbasin, routing, groundwater; Table 1; Fig. 3). No ad hoc adjustments were made to model parameters to force a model calibration with observed streamflow data, and all adjustments were made to more realistically represent the values of the parameters in the study area. No changes were made to the management database, which has the most sensitivity parameter—the curve number (CN). Within the basin database, the surface runoff lag coefficient (SURLAG) was adjusted to yield a better temporal fit between simulated and observed streamflow at the watershed outlet, as suggested in the SWAT User’s Manual (Neitsch et al. 2002). Within the HRU, subbasin, and routing databases, the Manning values for overland flow, tributary channels, and the main channel were changed to values more consistent with recommendations for rangeland settings (Neitsch et al. 2002). Additionally, the hydraulic conductivity of tributary and main channel alluvium (CH_1, CH_2) was selected based on the average hydraulic conductivity of soils in the STATSGO soil database that underlie both the tributaries and the main channel in the basin. Lastly, there are a several parameters that influence modeled evapotranspiration values. Most significantly from the HRU database is the soil evaporation compensation factor, which allows the user to modify how soil evaporative demand varies as a function of depth within the soil. From the groundwater database, the parameters include the threshold depth for water in shallow aquifer required for return flow to occur (GWQMN), groundwater reevaporation coefficient (GW_REVAP), threshold depth for water in the shallow aquifer for “reevap” or percolation to the deep aquifer to occur (REEVAPMN), and baseflow alpha factor (ALPHA_BF). Evapotranspiration is extreme in the San Pedro basin, and all simulations have annual evapotranspiration values in excess of 90% of the annual precipitation. Parameters that affect evapotranspiration were set to values (Table 1) at the extreme edge of permissible ranges to minimize modeled evapotranspiration, producing the best possible fit between actual and simulated streamflow for both dry and wet periods, and suboptimum model performance was recorded with the selection of any other permissible value.

Table 1.

SWAT model parameters.

SWAT model parameters.
SWAT model parameters.

A series of SWAT simulations, identical except for daily precipitation input, were executed for five precipitation types (rain gauge, MPE, TMPA-Research, TMPA-RT, CMORPH) for the San Pedro basin between Tombstone and Benson. Simulations spanned a time series of 1 January 2003–31 December 2008. Precipitation data from 1 January 2003 to 30 September 2005 were used to initialize the model (warm-up period), with the simulation period spanning 1 October 2005–31 December 2008 corresponding with the record of streamflow present at the watershed outlet (Benson). Simulations were not extended into 2009 because this was a particularly dry year and the SWAT model tends to overestimate streamflow during dry periods (Feyereisen et al. 2007; Van Liew et al. 2005, 2007).

d. Quantification of streamflow results

Simulated streamflow was evaluated with three measures of goodness of fit relative to observed streamflow: mass balance error (MBE), Nash–Sutcliffe efficiency coefficients (NS; Nash and Sutcliffe 1970), and root-mean-square error standard deviation ratio (RMR). The three quantitative goodness-of-fit measures were defined as follows:

 
formula
 
formula
 
formula

where Qobs,a was the average observed streamflow. Additionally, Qsim,i and Qobs,i were the simulated and observed streamflow at the ith observation, respectively; and n was the number of observations. Acceptable simulations had simulated and observed streamflow within 25% (MBE) and had NS values >0.50 with RMR values <0.70 (Moriasi et al. 2007). Negative NS values reflected a simulation that performed worse than if the average observed streamflow was utilized for correlation. Legate and McCabe (1999) have documented that the NS value is a superior measure of goodness of fit for hydrological time series datasets. Efficiency measures were determined at monthly and daily (time averaged over a 3-day period) time scales.

5. Results

a. Precipitation data

As a reference we compared the TMPA-Research product from the grid that corresponds with Douglas rain gauge data. For the period of 2004–08, the TMPA-Research (1310 mm) overestimated precipitation by 13% compared to the Douglas rain gauge data (1160 mm). The primary intercomparison noted for the precipitation data within the San Pedro basin was that the variation between precipitation products was significantly greater than the internal variation of precipitation values for an individual product. Additionally, there was no significant correlation (r2 < 0.4) between either warm season (July–September) or total precipitation values (2004–08) for 0.25° aggregated MPE, TMPA-Research, or TMPA-RT products with mean grid elevation within the study area.

For the same 2004–08 period, both TMPA-Research [average (avg) = 1089 mm, standard deviation (std dev) = 63 mm] and MPE (avg = 871 mm, std dev = 95 mm) underestimated precipitation in comparison to the Douglas rain gauge by 6% and 25%, respectively. Conversely, during the same period, CMORPH (avg = 1789 mm, std dev = 241 mm) and TMPA-RT (avg = 2247 mm, std dev = 226 mm) overestimate precipitation in comparison with the Douglas rain gauge by 52% and 94%. CMORPH and TMPA-RT had significantly higher averages compared with all ground-based precipitation products (p value from t-test statistic <0.0005). R. F. Alder (2009, personal communication) suggested that TRMM products have been recently enhanced and consequently an additional intercomparison for the water year 2008 (October 2008–September 2009) was completed. Some improvements in bias are noted for the TMPA-RT product; however, overall, similar trends existed between ground and satellite products. MPE (avg = 126 mm, std dev = 16 mm) underestimates precipitation by 14% compared with the Douglas rain gauge (146 mm). CMORPH (avg = 357 mm, std dev = 28 mm) and TMPA-RT (avg = 208 mm, std dev = 13 mm) overestimate precipitation compared with the Douglas rain gauge data by 65% and 144%, respectively. Additionally, we examined the near-real-time TRMM 3B41RT product that vastly overestimates precipitation (avg = 1529 mm, std dev = 187 mm, 946%). The average of CMORPH, TMPA-RT, and TRMM 3B41RT were significantly higher than MPE values from the corresponding grid cells (p < 0.0005).

The reference for quantitative estimate of satellite product performance was determined by a comparison of TMPA-Research from the grid that corresponds with Douglas rain gauge data (Table 2). Interestingly, both the TMPA-Research and TMPA-RT products produced fewer robust results when compared to 0.25° aggregated MPE data from corresponding TRMM grid cells within the San Pedro basin (Table 2). MPE data aggregated to correspond with CMORPH performed better in all metrics (Table 2) than TRMM products (p < 0.0005), consistent with previous observations in CONUS (Tian et al. 2007).

Table 2.

Quantitative estimate of satellite product performance.

Quantitative estimate of satellite product performance.
Quantitative estimate of satellite product performance.

A PDF was fitted for the daily precipitation values for all products examined. A mixed exponential function was then used to fit the data with two exponential functions fitting high and low values. The boundary between these domains varied between 7 and 13 mm and was selected based on the most significant natural break in the datasets as determined by visual inspection. The fitted curves based on this approach are presented in Fig. 4 for rain gauge, MPE, TMPA-Research, and TMPA-RT data for the representative grid cell centered at 32.125°N, 110.375°W (grid cell 2). Regression values for all products were high with r2 values in excess of 0.9. The curve fit was generally better for the lower domain in all products. Note that TMPA-RT exhibited consistently high precipitation values at all exceedance probabilities compared with rain gauge, MPE, and TMPA-Research products (Fig. 4).

Fig. 4.

Daily exceedance probability (January 2004–December 2008) of NWS rain gauge (rain gauge), MPE, TMPA-Research, and TMPA-RT from grid cell 2, which has a centroid of 32.125°N, 110.375°W.

Fig. 4.

Daily exceedance probability (January 2004–December 2008) of NWS rain gauge (rain gauge), MPE, TMPA-Research, and TMPA-RT from grid cell 2, which has a centroid of 32.125°N, 110.375°W.

b. Streamflow based on original precipitation products

For the San Pedro basin, simulated streamflow can be divided based on whether precipitation data were pure satellite products (TMPA-RT, CMORPH) or had a ground-based component (rain gauge, MPE, TMPA-Research). Excellent matches between observed and simulated streamflow were obtained for rain gauge, MPE, and TMPA-Research products (Table 3; Fig. 5a). Simulations from these products had consistently positive, and yet acceptable, MBE (8%–12%) as well as excellent monthly and daily NS values (≥0.89). Positive MBE associated with the earlier-mentioned simulations can be accounted for by the overestimation of streamflow during dry periods, mainly the cooler season (October–June; Table 3). Previous studies have documented that the SWAT modeling has problems with accurately accounting for soil moisture storage and water loss through infiltration and evapotranspiration during dry climatic periods and performs better in more humid—as opposed to arid—climate regimes (Feyereisen et al. 2007; Van Liew et al. 2005, 2007). CMORPH streamflow simulation did not have acceptable MBE (63%) even when accounting for the overestimation of streamflow during dry periods, with marginal efficiency measures (monthly NS = 0.51; Table 3). TMPA-RT has unacceptable MSE (165%) and negative NS values, indicative of a poor simulation (Table 3).

Table 3.

SWAT-simulated streamflow results from unadjusted products.

SWAT-simulated streamflow results from unadjusted products.
SWAT-simulated streamflow results from unadjusted products.
Fig. 5.

Observed daily streamflow and simulated streamflow based on SWAT modeling for the San Pedro basin watershed using (a) actual and (b) adjusted precipitation data.

Fig. 5.

Observed daily streamflow and simulated streamflow based on SWAT modeling for the San Pedro basin watershed using (a) actual and (b) adjusted precipitation data.

Examination of streamflow simulations during the summer monsoonal season (July–September) yields additional insights, as illustrated in Table 4. All products exhibit poorer performance during the summer of 2008 as opposed to the previous two summers. MPE and TMPA-Research products support excellent streamflow simulations during all summers. Rain-gauge-based simulation has excessive MSE (37%) during the summer of 2007. CMORPH simulation yielded an acceptable simulation only during the summer of 2006. Lastly, TMPA-RT had unacceptable results during all three summers.

Table 4.

SWAT-simulated streamflow results from summer season.

SWAT-simulated streamflow results from summer season.
SWAT-simulated streamflow results from summer season.

To check the sensitivity of varying date definition on the performance of 3-hourly TRMM products in our daily streamflow simulations, we executed two additional simulations (2004–08) based on the TMPA-Research product. The baseline simulation presented earlier is defined with a day beginning with TRMM 0600 UTC data, which produced a daily efficiency comparable to rain gauge and MPE-based simulations (daily NS = 0.90; Table 3). Additional simulations based on beginning the day with 0300 and 0900 UTC data yielded similar MBE and monthly/daily NS values, as reported in Table 3.

c. Streamflow based on adjusted precipitation products

Adjustments to TMPA-RT and CMORPH datasets are based on two filters—actual and AR—applied with either MPE or rain gauge (RG) data. Maximum variation in streamflow simulations resulting from the application of the actual versus AR filter to a single type of adjusted precipitation product is minimal (MSE ≤ 4%, NS ≤ 0.07). Streamflow simulations based on the MPE-adjusted TMPA-RT product yielded acceptable results (Table 5; Fig. 5b). These results were comparable to streamflow simulations based on rain gauge, MPE, and TMPA-Research products, with MBE attributable to the overestimation of streamflow during dry periods (Table 5).

Table 5.

Simulated streamflow from adjusted satellite products. Adjustments based on the AR filter.

Simulated streamflow from adjusted satellite products. Adjustments based on the AR filter.
Simulated streamflow from adjusted satellite products. Adjustments based on the AR filter.

To demonstrate the utility of the approach to regions with limited ground-based precipitation monitoring, we applied the MPE values and filter based on a single grid to the adjustment of TRMM data in all five grids that define the basin (grid filters 1–5; Fig. 1; Table 5). This series of model runs based on the spatial transfer (Fig. 6) of a single dataset/filter throughout the entire basin documents the potential of our approach for watersheds with limited ground-based precipitation data. All five MPE spatial transfers yielded acceptable streamflow simulations with positive MBE (7%–33%) when accounting for the overestimation of streamflow during dry periods (11%–14%) and excellent daily (NS = 0.75–0.90; Table 5) efficiency measures. Additionally, to evaluate uncertainty of using spatially transferred MPE data in the adjusted TRMM product, an ensemble of 50 simulations were executed where inserted MPE values to account for missed precipitation were randomly selected within a range between 0% and 200% of the original MPE value. All ensemble simulation results were quite close, with a maximum ensemble range between low and high values of MSE confined to a range that was less than 10% and NS varied by less than 0.05.

Fig. 6.

Representation of the spatial transfer process.

Fig. 6.

Representation of the spatial transfer process.

Streamflow simulations based on Douglas rain-gauge-adjusted TMPA-RT data did not perform as well as models based on MPE-adjusted products but yielded acceptable results (Table 5; MSE = 35%, NS = 0.75). Discounting the 10% dry period overestimation, rain-gauge-adjusted TMPA-RT simulations had a streamflow MSE that was barely within the 25% limit for acceptable simulations (Moriasi et al. 2007). Lastly, the adjustment methodology, based on both MPE and RG data, also improved CMORPH-based simulations (Table 5; MSE = 34%–41%, NS = 0.77–0.79); discounting the 24%–25% overestimation during dry periods.

Examination of streamflow simulations based on adjusted precipitation products during monsoonal season (July–September) yielded additional insights, as illustrated in Table 6. MPE-adjusted TMPA-RT and CMORPH data produced acceptable simulations during all three summers. However, rain-gauge-adjusted TMPA-RT produced an acceptable simulation for only the summer of 2006 and rain-gauge-adjusted CMORPH was acceptable for the summers of 2006 and 2007 (Table 6). The single rain gauge used to adjust satellite data, located in Douglas, indicated that the spatial transfer of ground-based precipitation over significant distances (59–101 km) may work only during more humid periods, such as summer 2006.

Table 6.

SWAT-simulated streamflow results from summer season for adjusted satellite products. MPE-based adjustments based on all MPE grids and the AR filter.

SWAT-simulated streamflow results from summer season for adjusted satellite products. MPE-based adjustments based on all MPE grids and the AR filter.
SWAT-simulated streamflow results from summer season for adjusted satellite products. MPE-based adjustments based on all MPE grids and the AR filter.

6. Discussion and significance

a. Overview of results

Significant inconsistencies existed in terms of bias, FAR, and POD with satellite precipitation products such as TRMM, PERSAINN, and CMORPH (Ebert et al. 2007). The quality of satellite precipitation products varied greatly worldwide and even over CONUS (Ebert et al. 2007; Tian et al. 2007). Bias compared with ground-based methods was a significant issue. Positive bias was present throughout central and western CONUS, especially during the warm season (Ebert et al. 2007; Tian et al. 2007). A major problem in warm arid regions is evaporation of hydrometeors below cloud base, which is difficult to quantify in satellite products (Ferraro et al. 1998; McCollum et al. 2002; Rosenfeld and Mintz 1988; Scofield 1987; Spencer et al. 1989).

An emerging approach to facilitate the utility of satellite precipitation products in supporting hydrological modeling is to correct the bias present in the satellite datasets. This approach was able to generate precipitation datasets that produced acceptable hydrological simulations (i.e., Boushani et al. 2009; Vila et al. 2009; Wilk et al. 2006). Harris et al. (2007) cautions that bias corrections can add artifacts to a time series of streamflow data that results in gross inaccuracies in hydrologic simulations. Bias-correction approaches are typically applied over longer monthly time scales (Huffman et al. 2007); however, recent research has applied bias-correction methodologies at an hourly time scale (Boushani et al. 2009). The studies listed earlier do not generally address issues associated with a lack of detection or false alarms.

The goal of this paper was to validate a methodology that will allow users to adjust satellite products based on ground data, removing bias, lack of detection, and false alarms. Results in Table 5 demonstrate the promise of the developed methodology. There was just one major discrepancy noted in this approach during June 2007 (Fig. 5b), where the extreme positive bias present in TMPA-RT data could not be corrected. Significantly, this methodology can be applied to any satellite precipitation product, including the new version of TRMM products (version 7) and data from the planned Global Precipitation Mission. Another significant result of this study was the successful application of the methodology using just a single ground-based dataset to adjust all five 0.25° TRMM grids within the watershed. Our study demonstrates that one time series of ground-based data can be successfully utilized to adjust all grids of satellite data in a moderately large watershed (Table 5). Even if the ground dataset is located outside the examined watershed, as was the case with the Douglas rain gauge data, acceptable streamflow simulations were still realized, especially during more humid periods, such as the summer of 2006 (Table 6). This conclusion is highly applicable to watersheds in poorly gauged regions of the world where there may be only one rain gauge in geographic proximity to a basin. Other bias-reduction approaches (i.e., Boushani et al. 2009) only produce robust results if numerous gauges are available to facilitate satellite product adjustment. The significance of the developed method is the retention of the spatial variability of the satellite product capturing interbasinal variability in precipitation. Therefore, in sparsely gauged regions, models can be driven at a 0.25° resolution instead of using limited gauge data as input, allowing the modeler to more realistically capture the variability of basin hydrologic processes.

b. Designing an operational product: Issues

A major goal in the hydrologic community is to make satellite products work for hydrologic applications on land in a near-real-time environment. Our approach has been validated at a daily time scale and could conceivably support hydrological modeling at a subdaily temporal resolution. Therefore, the developed approach could support the utilization of satellite precipitation data for flood prediction and nowcasting in an operational environment, especially with on-the-fly adjustment of products like TRMM 3B41RT that have a minimal latency. Comparison of adjusted basin precipitation values from the water year 2009 is instructive with adjusted TMPA-RT (avg = 98 mm, std dev = 17 mm) and adjusted TRMM 3B41RT (avg = 146 mm, std dev = 70 mm) yielding comparable total precipitation values to the Douglas rain gauge (146 mm) and 0.25° aggregated MPE (avg = 126 mm, std dev = 16 mm). These results are noteworthy in that the unadjusted products have extreme positive biases (TMPA-RT avg = 208 mm, std dev = 173 mm; TRMM 3B41RT avg = 1529 mm, std dev = 187 mm).

The initial application of the developed methodology was based on modeling with SWAT. The results were deemed particularly encouraging because robust simulations at a daily time scale were obtained in an arid region. The SWAT model was not designed for daily but rather for monthly modeling, and it has been documented to perform less robustly in dry climatic settings (Gassman et al. 2007; Van Liew et al. 2005, 2007). Consequently, further validation could be obtained with simulations based on other hydrologic models [e.g., Gridded Surface Subsurface Hydrologic Analysis, Variable Infiltration Capacity (VIC) Macroscale Hydrology] that could support subdaily modeling and applications such as flood forecasting. However, the results of this study directly support the premise that this approach could be easily applied to longer-term hydrological applications, such as drought monitoring and water and agricultural management.

Operationally, application of the approach would begin with the ingestion of ground-based (e.g., rain gauge) and satellite precipitation data (e.g., TMPA-RT). Ground data are used to develop a filter to address the lack of detection or false alarms in the satellite product. Most ideally, the PDFs of two precipitation datasets would be updated on a daily (or subdaily) basis, providing a means for addressing the issue of bias continuously. To test the sensitivity of our proposed method to the frequency of PDF update, we ran a simulation of adjusted TMPA-RT, where the PDFs were adjusted on an annual basis at the beginning of the water year, simulating a setting where minimal resources are available to implement this procedure. In this situation, only three annual adjustments (1 October 2006–08) to the PDF were made as opposed to the baseline simulation that reflects continuous daily adjustment over the entire period of record. With only annual adjustment, simulated streamflow was not significantly different from the baseline TMPA-RT simulations presented in Table 5.

This methodology can also be applied in hybrid operational situations in regions that frequently have missing ground-based observations, which is common throughout the developing world (Su et al. 2008). The approach in this situation would involve the construction of the PDFs based on dates that have both satellite and ground precipitation data available. Uncertainty in POD and FAR for dates with missing ground precipitation data would be addressed through Markov Chain analysis to determine the likelihood of missed or incorrectly identified precipitation events.

c. Potential downscaling applications

The developed methodology can be applied to the downscaling of satellite products. The demonstrated robustness of the spatial transfer of ground-based precipitation over significant distances indicates that the adjustment approach could easily be applied to the adjustment of satellite data within a 0.25° grid. The basis of a downscaling methodology would involve the development of a higher resolution—conceivable down to 4 km2—downscaled satellite product based on ordinary kriging or a more complex nonlinear interpolation approach (Chowdhury et al. 2009). Each downscaled grid could be adjusted based on the methodology delineated in this paper. Future work is needed to validate the minimal space–time resolution and basin size to which the adjustment methodology can support acceptable hydrological simulations through downscaled satellite precipitation data. The major advantage of the development of a downscaled satellite product through the approach articulated in this paper lies with the ability of this approach to better capture, in detail, interbasinal variability of precipitation that will provide a more detailed characterization of the hydrological processes at the watershed scale.

7. Summary

This study outlines a methodology to adjust satellite precipitation products in terms of bias, false alarms, and lack of detection compared with ground-based precipitation data. The satellite precipitation products examined included TMPA-Research, TMPA-RT, and TRMM 3B41RT, and CMORPH, which have a 0.25° spatial resolution. The developed approach was not a simple bias adjustment but involved three steps that adjusted a satellite product based on a ground-based precipitation product. This methodology was successfully applied in the hydrological modeling of the San Pedro basin in Arizona for a 3-yr time series, yielding excellent streamflow simulations at a daily time scale. Simulated streamflow based on unadjusted TMPA-RT data had unacceptable mass balance error (MSE = 165%; Table 2). Simulations based on adjusted TMPA 3B42 data had acceptable MSE and efficiency coefficients at a daily time scale. The developed approach could be applied to any satellite precipitation product (i.e., TRMM 3B42 version 7). The utility of the developed methodology is its ability to support real-time hydrologic modeling, at a daily time scale, for moderately large (1000s km2) basins with limited ground-based, precipitation monitoring.

Acknowledgments

We acknowledge the support of the NASA Precipitation Science program through Award NA17AE2924 (Dr. Ramesh Kakar). The assistance of research assistants Adriana Torres and Arturo Diaz was greatly appreciated. The data used in this study were acquired using the GES-DISC Interactive Online Visualization and Analysis Infrastructure (Giovanni) as part of NASA’s Goddard Earth Science (GES) Data and Information Services Center (DISC). Additionally, we gratefully acknowledge the assistance of the Colorado Basin River Forecast Center in providing MPE data for this project.

REFERENCES

REFERENCES
Artan
,
G.
,
H.
Gadin
,
J.
Smith
,
K.
Asante
,
C.
Baedaragoda
, and
J.
Verdin
,
2007
:
Adequacy of satellite-derived rainfall data for streamflow modeling.
Nat. Hazards
,
43
,
167
185
.
Beighley
,
R. E.
,
K. G.
Eggen
,
T.
Dunne
,
Y.
He
,
V.
Gummadi
, and
K. L.
Verdin
,
2009
:
Simulating hydrologic and hydraulic processes throughout the Amazon River Basin.
Hydrol. Processes
,
23
,
1221
1235
.
Boushani
,
F. I.
,
K-L.
Hsu
,
S.
Sorooshian
,
G-H.
Park
,
S.
Mallaini
, and
W.
Shi
,
2009
:
Bias adjustment of satellite precipitation estimation using ground-based measurement: A case study over the southwestern United States.
J. Hydrometeor.
,
10
,
1231
1242
.
Chowdhury
,
M.
,
A.
Alouani
, and
F.
Hossain
,
2009
:
How much does inclusion of non-linearity and multi-point pattern recognition improve the spatial mapping of complex patterns of groundwater contamination?
Nonlinear Processes Geophys.
,
16
,
313
317
.
Collischonn
,
B.
,
W.
Collischonn
, and
C. E.
Morelli Tucci
,
2008
:
Daily hydrological modeling in the Amazon basin using TRMM rainfall estimators.
J. Hydrol.
,
360
,
207
216
.
CPC
,
cited
.
2010
:
NOAA CPC morphing technique (“CMORPH”).
.
Daly
,
C.
,
R. P.
Neilson
, and
D. L.
Phillips
,
1994
:
A statistical–topographic model for mapping climatological precipitation over mountainous terrain.
J. Appl. Meteor.
,
33
,
140
158
.
Ebert
,
E. E.
,
J. E.
Janowisk
, and
C.
Kidd
,
2007
:
Comparison of near-real-time precipitation estimates from satellite observations.
Bull. Amer. Meteor. Soc.
,
88
,
47
64
.
Ferraro
,
R. R.
,
E. A.
Smith
,
W.
Berg
, and
G. J.
Huffman
,
1998
:
A screening methodology for passive microwave precipitation retrieval algorithms.
J. Atmos. Sci.
,
55
,
1583
1600
.
Feyereisen
,
G. W.
,
T. C.
Strickland
,
D. D.
Bosch
, and
D. G.
Sullivan
,
2007
:
Evaluation of SWAT manual calibration and input parameter sensitivity in the Little River Watershed.
Trans. ASABE
,
50
,
843
855
.
Foufoula-Georgiou
,
E.
, and
D. P.
Lettenmaier
,
1987
:
A Markov renewal model for rainfall occurrences.
Water Resour. Res.
,
23
,
875
884
.
Gassman
,
P. W.
,
M. R.
Reyes
,
C. H.
Green
, and
J. G.
Arnold
,
2007
:
The Soil and Water Assessment Tool: Historical development, applications, and future research directions.
Trans. ASABE
,
50
,
1211
1250
.
Harris
,
A.
,
S.
Rahman
,
F.
Hossain
,
L.
Yarborough
,
A. C.
Hagtzogiou
, and
G.
Easson
,
2007
:
Satellite-based flood modeling using TRMM-based rainfall productions.
Sensors
,
7
,
3416
3427
.
Hsu
,
K-l
,
X.
Gao
,
S.
Sorooshian
, and
H. V.
Gupta
,
1997
:
Precipitation estimation from remotely sensed information using artificial neutral networks.
J. Appl. Meteor.
,
36
,
1176
1190
.
Huffman
,
G. J.
, and
Coauthors
,
2007
:
The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales.
J. Hydrometeor.
,
8
,
38
55
.
Hughes
,
D.
,
L.
Andersson
,
J.
Wilk
, and
H.
Savenije
,
2006
:
Regional calibration of the Pitman model for the Okavango River.
J. Hydrol.
,
327
,
399
410
.
Joyce
,
R. L.
,
J. E.
Janowiak
,
P. A.
Arkin
, and
P.
Xie
,
2004
:
CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution.
J. Hydrometeor.
,
5
,
487
503
.
Legate
,
D. R.
, and
G. J.
McCabe
Jr.
,
1999
:
Evaluating the use of “Goodness-of-Fit” measures in hydrologic and hydroclimate model validation.
Water Resour. Res.
,
35
,
233
241
.
McCollum
,
J. R.
,
W. F.
Krajewski
,
R. R.
Ferraro
, and
M. B.
Ba
,
2002
:
Evaluation of biases of satellite rainfall estimation algorithms over the continental United States.
J. Appl. Meteor.
,
41
,
1065
1080
.
Moriasi
,
D. N.
,
J. G.
Arnold
,
M. W.
Van Liew
,
R. L.
Bingner
,
R. D.
Harmel
, and
T. L.
Veith
,
2007
:
Model evaluation guidelines for systematic quantification of accuracy in watershed simulations.
Trans. ASABE
,
50
,
885
900
.
NASA
,
cited
.
2010
:
TRMM Online Visualization and Analysis System (TOVAS).
.
Nash
,
J. E.
, and
J. V.
Sutcliffe
,
1970
:
River flow forecasting through conceptual models. Part 1: Discussion of principles.
J. Hydrol.
,
10
,
282
290
.
NCDC
,
cited
.
2009
:
Climate station locator.
.
Neitsch
,
S. L.
,
J. G.
Arnold
,
J. R.
Kiniry
,
R.
Srinivasan
, and
J. R.
Williams
,
2002
:
Soil and water assessment tool user’s manual.
Texas Water Resource Institute Rep. TR-192, 455 pp
.
Rosenfeld
,
D.
, and
Y.
Mintz
,
1988
:
Evaporation of rain falling from convective clouds as derived from radar measurements.
J. Appl. Meteor.
,
27
,
209
215
.
Scofield
,
R. A.
,
1987
:
The NESDIS operational convective precipitation estimation technique.
Mon. Wea. Rev.
,
115
,
1773
1793
.
Serrat-Capdevila
,
A.
,
J. B.
Valdés
,
J.
González Perez
,
K.
Baird
,
L. J.
Mata
, and
T.
Maddock
,
2007
:
Modeling climate change impacts – and uncertainty – on the hydrology of a riparian system: The San Pedro Basin (Arizona/Sonora).
J. Hydrol.
,
347
,
48
66
.
Spencer
,
R. W.
,
H. M.
Goodman
, and
R. E.
Hood
,
1989
:
Precipitation retrieval over land and ocean with the SSM/I: Identification and characteristics of the scattering signal.
J. Atmos. Oceanic Technol.
,
6
,
254
273
.
Su
,
F.
,
Y.
Hong
, and
D. P.
Lettenmaier
,
2008
:
Evaluation of TRMM Multisatellite Precipitation Analysis (TMPA) and its utility in hydrologic prediction in the La Plata basin.
J. Hydrometeor.
,
9
,
622
640
.
Tian
,
Y.
,
C. D.
Peters-Lidard
,
B. J.
Choudhury
, and
M.
Garcia
,
2007
:
Multitemporal analysis of TRMM-based satellite precipitation products for land data assimilation applications.
J. Hydrometeor.
,
8
,
1165
1183
.
US EPA
,
cited
.
2010
:
State Soil Geographic (STATSGO) Database for CONUS, Alaska, and Hawaii in BASINS.
.
USGS
,
1986
:
Land use and land cover digital data from 1:250,000- and 1:100,000-scale maps.
Data Users Guide 4, 37 pp. [Available online at http://edc2.usgs.gov/geodata/LULC/LULCDataUsersGuide.pdf]
.
USGS
,
cited
.
2009a
:
National hydrography dataset.
[Available online at http://nhd.usgs.gov]
.
USGS
,
cited
.
2009b
:
USGS water data for Arizona.
.
Van Liew
,
M. W.
,
J. G.
Arnold
, and
D. D.
Bosch
,
2005
:
Problems and potential of autocalibrating a hydrologic model.
Trans. ASABE
,
48
,
1025
1040
.
Van Liew
,
M. W.
,
T. L.
Vieth
,
D. D.
Bosch
, and
J. G.
Arnold
,
2007
:
Suitability of SWAT for the conservation effects assessment project: Comparison on USDA Agricultural Research Service Watersheds.
J. Hydrol. Eng.
,
12
,
173
189
.
Vila
,
D. A.
,
L. G. G.
De Goncalves
,
D. L.
Toll
, and
J. R.
Rozante
,
2009
:
Statistical evaluation of combined daily gauge observations and rainfall satellite estimates over continental South America.
J. Hydrometeor.
,
10
,
533
543
.
Webb
,
R. H.
, and
J. L.
Betancourt
,
1992
:
Climatic variability and flood frequency of the Santa Cruz River, Pima County, Arizona.
U.S. Geological Survey Water Supply Paper 2379, 40 pp
.
Wilk
,
J.
,
D.
Kniveton
,
L.
Andersson
,
R.
Layberry
,
M. C.
Todd
,
D.
Hughes
,
S.
Ringrose
, and
C.
Vanderpost
,
2006
:
Estimating rainfall and water balance over the Okavango River Basin for hydrological application.
J. Hydrol.
,
331
,
18
29
.
Wilks
,
D. S.
,
1999
:
Interannual variability and extreme-value characteristics of several stochastic daily precipitation models.
Agric. For. Meteor.
,
93
,
153
169
.
Wilks
,
D. S.
,
2006
:
Statistical Methods in the Atmospheric Sciences.
2nd ed. Elsevier, 627 pp
.
Woolhier
,
D. A.
, and
J.
Roldán
,
1982
:
Stochastic daily precipitation models, 2. A comparison of distribution and amounts.
Water Resour. Res.
,
18
,
1461
1468
.
Xie
,
H.
,
X.
Zhou
,
E. R.
Vivoni
,
J. M. H.
Hendrick
, and
E. E.
Small
,
2005
:
GIS-based NEXRAD Stage III precipitation database: Automated approaches for data processing and visualization.
Comput. Geosci.
,
31
,
65
76
.
Yilmaz
,
K. K.
,
T. S.
Hogue
,
K-l
Hsu
,
S.
Sorooshian
,
H. V.
Gutpa
, and
T.
Wagener
,
2005
:
Intercomparison of rain gauge, radar, and satellite-based precipitation estimates with emphasis on hydrologic forecasting.
J. Hydrometeor.
,
6
,
497
517
.

Footnotes

Corresponding author address: Kenneth J. Tobin, Center for Earth and Environmental Studies, Texas A&M International University, 5201 University Boulevard, Laredo, TX 78041. Email: ktobin@tamiu.edu