Você está na página 1de 10

HYDROLOGICAL PROCESSES Hydrol. Process. 25, 37113720 (2011) Published online 15 April 2011 in Wiley Online Library (wileyonlinelibrary.

com) DOI: 10.1002/hyp.8096

Performance evaluation of interpolation methods for incorporating rain gauge measurements into NEXRAD precipitation data: a case study in the Upper Guadalupe River Basin
Hongjie Xie,1 * Xuesong Zhang,2 Beibei Yu1 and Hatim Sharif3
1

Laboratory for Remote Sensing and Geoinformatics, Department of Geological Sciences, University of Texas at San Antonio, San Antonio, TX 78249, USA 2 Joint Global Change Research Institute, Pacic Northwest National Laboratory, College Park, MD 20740, USA 3 Department of Civil and Environmental Engineering, University of Texas at San Antonio, San Antonio, TX 78249, USA

Abstract:
High spatial and temporal resolution of precipitation data is critical input for hydrological budget estimation and ash ood modelling. This study evaluated four methods [Bias Adjustment (BA), Simple Kriging with varying Local Means (SKlm), Kriging with External Drift (KED), and Regression Kriging (RK)] for their performances in incorporating gauge rainfall measurements into Next Generation Weather Radar (NEXRAD) multi-sensor precipitation estimator (MPE; hourly and 4 4 km2 ). Measurements from a network of 50 gauges at the Upper Guadalupe River Basin, central Texas and MPE data for the year 2004 were used in the study. We used three evaluation coefcients percentage bias (PB), coefcient of determination (R2 ), and NashSutcliffe efciency (NSE) to examine the performance of the four methods for preserving regional- and local-scale characteristics of observed precipitation data. The results show that the two Kriging-based methods (SKlm and RK) are in general better than BA and KED and that the PB and NSE criteria are better than the R2 criterion in assessing the performance of the four methods. It is also worth noting that the performance of one method at regional scale may be different from its performance at local scale. Critical evaluation of the performance of different methods at local or regional scale should be conducted according to the different purposes. The results obtained in this study are expected to contribute to the development of more accurate spatial rainfall products for hydrologic budget and ash ood modelling. Copyright 2011 John Wiley & Sons, Ltd.
KEY WORDS

NEXRAD; MPE; rain gauge; precipitation; geostatistics

Received 8 December 2009; Accepted 8 March 2011

INTRODUCTION Precipitation, characterized by high spatial and temporal variability, is one of the primary controlling factors for many physical processes such as the hydrological cycle, land surface and air interaction, and forest and agriculture crops growth. Previous studies have shown that accurate estimation of precipitation is critical for simulating runoff, time shift of hydrographs, sediment delivery, and nutrient yield (Arnaud et al., 2002; Hossain et al., 2004; Chaplot et al., 2005; Kalinga and Gan, 2006; Younger et al., 2009). Traditionally, precipitation measurements from rain gauges are usually too sparsely distributed to capture the spatial variability. Weather radar measurements of precipitation, which provide precipitation data with much higher spatial resolution in comparison to rain gauges, have served meteorology for over 40 years and hydrology for around 20 years (Krajewski and Smith, 2002). The U.S. Next Generation

* Correspondence to: Hongjie Xie, Laboratory for Remote Sensing and Geoinformatics, Department of Geological Sciences, University of Texas at San Antonio, San Antonio, TX 78249, USA. E-mail: hongjie.xie@utsa.edu Copyright 2011 John Wiley & Sons, Ltd.

Weather Radar (NEXRAD), which can provide precipitation products with high spatial resolution (4 4 km2 ) at hourly temporal scale, has been used in a wide range of hydrological, meteorological, and climatological studies (Bedient et al., 2000; Bradley et al., 2002; Knebl et al., 2005; Kalin and Hantush, 2006). Many studies have been performed to validate the NEXRDA rainfall estimates (Young et al., 2000; Jayakrishnan et al., 2004; Xie et al., 2005, 2006; Wang et al., 2008; Habib et al., 2009). Many efforts attempted to incorporate rain gauge measurements into NEXRAD precipitation products using spatial interpolation techniques (Seo, 1998; Steiner et al., 1999; Seo and Breilenbach, 2002; Haberlandt, 2007; Li et al., 2008). Different methods, ranging from simple multiplicative correction methods to complex geostatistical methods, have been applied in previous studies on improving NEXRAD products. For example, Steiner et al. (1999) applied a simple Bias Adjustment (BA) method to correct NEXRAD in Goodwin Creek, a small research watershed in northern Mississippi. Li et al. (2008) developed a linear regression based Kriging method to improve daily NEXRAD precipitation using rain gauge data and applied it in Texas to estimate daily spatial precipitation in 2003. These results have shown

3712

H. XIE ET AL.

Figure 1. Location of the study area and gauge locations with identication numbers (ID). The three rain gauges used as examples in the Performance Evaluation of Different Interpolation Methods at Local Scale Section are gauge IDs 4 and 18 at Kerr County (kr4, kr18) and 1 at Guadalupe County (gp1)

the potential of further improving NEXRAD data using rain gauge data to provide more accurate spatial precipitation estimation. The objective of this study is to evaluate four spatial interpolation methods [Bias Adjustment (BA), Simple Kriging with varying Local Means (SKlm), Kriging with External Drift (KED), and Regression Kriging (RK)] for their performances in incorporating gauge measurements into NEXRAD precipitation products. We selected the Upper Guadalupe River Basin (UGRB) as the study area because of the availability of high quality precipitation records from 50 rain gauges across this area. In previous research, it was common to evaluate the performance of different methods to capture the rainfall characteristics at regional scale which is important for hydrologic budget estimation at watershed scale. However, minimal research has been devoted to evaluate the accuracy in localized precipitation that is critical for accurate ash ood forecasting. Therefore, results from this study will provide insight into the performances of different methods in terms of characterizing precipitation characteristics at both regional and local scales, which are valuable for developing accurate precipitation data for ecological and hydrologic modelling. METHODOLOGY Study area The study area lies in the Hill Country of central Texas, covering the UGRB and partly intersecting the
Copyright 2011 John Wiley & Sons, Ltd.

recharge zone of the Edwards Aquifer (Figure 1). The region belongs to a semi-arid climate regime and has annual mean rainfall of approximately 800 mm. The area is highly vulnerable to hydrologic extremes: intense precipitation events, severe ooding, and droughts (Scanlon et al., 2004). Data source The NEXRAD multi-sensor precipitation estimator (MPE), developed by the National Weather Service (NWS) Ofce of Hydrology since March 2000, is a product that merges rainfall measurements from rain gauges, and rainfall estimates from NEXRAD and Geostationary Operational Environmental Satellite (GOES) products. The NWS West Gulf River Forecast Center (WGRFC) switched from Stage III to MPE as the preferred precipitation estimation programme in October 2003, and ended Stage III in December 2004 (Wang et al., 2008). The MPE precipitation products are hourly accumulation with a spatial resolution of about 4 4 km2 . There are 50 rain gauges covering four counties: Kerr, Kendall, Comal, and Guadalupe in south-central Texas (Figure 1). All rain gauge data have a 6-min accumulation time-step. Year 2004 data of MPE and 50 rain gauges are used in the study. Spatial interpolation methods Four methods, applied for incorporating rain gauge data into the NEXRAD estimation, are BA, SKlm (Goovaerts, 2000), KED (Haberlandt, 2007), and RK
Hydrol. Process. 25, 37113720 (2011)

INTERPORATION METHODS FOR RAINFALL MAPPING

3713

(Hengl et al., 2004). BA is a simple and widely used method to adjust the system error in the Radar estimates. In BA method, it rst calculates the ratio of the mean gauge rainfall measurements over the mean radar rainfall estimates and then multiplies this ratio with radar data to estimate spatial rainfall distribution. Kriging methods have been shown as effective multivariate geostatistical techniques (Lloyd, 2005; Cheng et al., 2007; Haberlandt, 2007; Zhang and Srinivasan, 2009) for rainfall distribution estimation. In order to efciently implement the geostatistical techniques automatically for multiple time-steps, we revised the geographic information system (GIS)-based precipitation interpolation programme developed by Zhang and Srinivasan (2009) to read in NEXRAD data as auxiliary variable and extended this GIS programme to include RK and BA methods. Kriging is a group of advanced geostatistical techniques that provides the best linear unbiased estimate. In Kriging methods, rainfall value at location x is estimated as z O x D m x C x , where m is the expected mean rainfall and is the residual. In SKlm, KED, and RK methods, we calculate the expected mean rainfall as a linear equation of NEXRAD rainfall value m x D 0 C 1 R x and estimate the residual as a weighted sum of residuals that are calculated as the difference between rain gauge recorded and expected mean rainfall values at surrounding rain gauges x D n iD1 xi [ i ] . R x is the NEXRAD rainfall value at location x. 0 and 1 are linear coefcients. i is the residual at rain gauge i. n is the number of rain gauges. xi is the linear weight of residual at rain gauge i. Different Kriging methods use different assumptions and equations to calculate m and . The algorithms used by the three Kriging methods are introduced in the following paragraphs. More detailed information on SKlm, KED, and RK was provided by Goovaerts (1997) and Hengl et al. (2004). In RK method, it is assumed that the expected mean rainfall is a known constant. The concurrent pairs of NEXRAD and rain gauge observed rainfall values are used to estimate the linear coefcient 0 and 1 by generalized least squares (GLS). The optimal weights ( xi ) are solved using the Ordinary Kriging system, n x D hxi i D 1, . . . , n xj hij jD1 1 n xj D 1
jD1

the system of equations below


n xj C jD1

hij D C hxi

i D 1, . . . , n

(2)

where C h is the spatial covariance between two points separated by distance h. In SKlm and RK, the expected mean rainfall is assumed to be known, while KED assumes the expected mean rainfall is unknown. The major difference between KED and RK lies in that KED incorporates the estimation of the linear regression coefcients into the Ordinary Kriging equation system, while RK conducts the linear regression and Ordinary Kriging separately. The optimal weights in KED are solved using the following equations
n xj hij jD1 D hxi iD1 n
iD1 n xi

u0

u1 R i 3

i D 1, 2, . . . , n,

D 1, i DR x

xi R

Evaluating precipitation estimated by different methods To be consistent with previous research on spatial rainfall estimation using NEXRAD, we use rain gauge observed rainfall as ground truth because NEXRAD measures the atmospheric moisture above ground while a rain gauge collects precipitation falling on the ground. It is worth noting that measurement of small rainfall events is challenging for both rain gauges and NEXRAD (http://weather.noaa.gov/radar/radinfo/radinfo.html# clear; Habib et al., 2009), which may lead to signicant sampling errors. However, small amount of precipitation has relatively insignicant effect on hydrologic modelling (Westcott et al. 2008; Zhang and Srinivasan, 2009). Therefore, we focussed on those hours with mean spatial precipitation rate larger than 01 mm h 1 . A total of 737 h was selected. The accumulated areal mean precipitation depth of these 737 h accounts for about 95% of annual mean precipitation (10174 mm) for the year 2004. Regional-scale evaluation. The performance of different methods to estimate spatial rainfall at regional scale (the entire UGRB) is evaluated using cross-validation (Isaaks and Srivastava, 1989) for each of the selected 737 h. In the procedure, each of the rain gauge measurements is temporarily removed at a time and the remaining data are used to estimate a rainfall value for the location where the temporarily removed rain gauge is located at. The difference between the gauge observed and estimated value is then used to evaluate the accuracy of an interpolation method. Let zil , i D 1, . . . n, l D 1 . . . , L , denotes the i-th station measurement on the l-th hour, where L is the number of total hours being used in the comparison. Let z Oil denote the predicted value for the same gauge.
Hydrol. Process. 25, 37113720 (2011)

where xi is the weight assigned to the residual at rain gauge i, x is the Lagrange parameter accounting for the constraint on the weights, hij denotes the separation distance between rain gauges i and j. The h is semivariagram function. The general procedures in SKlm method are similar to those of RK, except that SKlm uses ordinary least squares (OLS) to t the linear regression between co-located rain gauge and NEXRAD rainfall values and employ spatial covariance function to interpolate the residuals. The optimal weights in SKlm are obtained by solving
Copyright 2011 John Wiley & Sons, Ltd.

3714

H. XIE ET AL.

Table I. Evaluation coefcients of precipitation maps produced by different methods for the three hours examined Date Hour 8, 24 April 2004 Evaluation coefcients Property Mean Min Max CV PB R2 NSE Mean Min Max CV PB R2 NSE Mean Min Max CV PB R2 NSE MPE 597 0 20 083 2146 086 0732 385 0 36 199 4243 094 07 716 0 32 133 2241 087 081 BA 774 0 25 082 028 085 083 678 0 64 198 018 094 093 932 0 41 133 004 086 086 SKlm 787 0 32 100 061 085 085 662 0 62 203 124 093 093 915 0 42 136 050 092 092 KED 767 0 32 102 189 085 085 666 0 55 194 021 094 094 912 0 46 138 079 092 092 RK 786 0 33 102 052 085 085 662 0 63 203 071 093 092 908 0 44 137 043 092 092

Accuracy

Hour 8, 22 June 2004

Property

Accuracy

Hour 9, 22 June 2004

Property

Accuracy

Units for mean, min, and max are mm h

and for PB is %.

For each hour, the following statistics were computed Percentage Bias, PB D Ol Z Zl Zl 100
2
2 05

Correlation Coefcient % , R2 n Ol zil Zl z Oil Z iD1 D 05 n n 2 Ol zil Zl z Oil Z


iD1 iD1

NashSutcliffe Model efciency coefcient, which indicates how well the plot of the rain gauge observed value versus the predicted value ts the 1 : 1 line, and ranges from 1 to 1 (Nash and Sutcliffe, 1970). Larger R2 and NSE value indicate better agreement between interpolations and observations. Local-scale evaluation. The method that provides better results at regional scale may not outperform the other methods for localized rainfall estimation. In addition to regional-scale evaluation, we evaluate the performance of different rainfall estimation methods at each rain gauge location. The same evaluation statistics (i.e. PB, R2 , and NSE) were used to evaluate the agreement between the estimated and rain gauge observed rainfall at each rain gauge. The form of equations used for local-scale evaluation are the same as those listed in the previous section, except that the spatially observed and estimated rain fall values are replaced with the observed and estimated rainfall at a specic rain gauge for a specied time period. The time series of observed and estimated rainfall are then used to evaluate the performance of different methods at each rain gauge location.

NashSutcliffe efciency, NSE


n

z Oil D 10
iD1 n

zil zl

6 zil
2

iD1

where Zl is the mean of gauge measured precipitation O l is the mean of predicted precipiof the l-th hour, Z tation at rain gauge locations of l-th hour. Var zil is the variance of gauge measured precipitation of the l-th hour, Var z Oil is the variance of interpolated rainfall values at rain gauges of the l-th hour. PB is calculated to represent the systematic bias of spatial rainfall distribution estimated by different methods. Negative PB value indicates underestimation and positive PB value indicates overestimation. The smaller the absolute value of PB, the better the performance of the predicting method. R2 is the square of the Pearsons product-moment correlation coefcient and describes the proportion of the total variance in the observed data that can be explained by the model. R2 is an indicator of the strength of the relationship between the observed and simulated values. NSE is also called
Copyright 2011 John Wiley & Sons, Ltd.

RESULTS Performance evaluation of different interpolation methods for regional-scale rainfall estimation Comparison of spatial precipitation maps interpolated by different methods. Here, we select three hours with signicant rainfall accumulation (the 8th hour, 24 April 2004; 8th and 9th hours, 22 June 2004) as examples to show the difference between maps produced by different methods as well as the MPE maps. Table I
Hydrol. Process. 25, 37113720 (2011)

INTERPORATION METHODS FOR RAINFALL MAPPING

3715
SKlm

60

No Correction

60

BA

60

Observed

40

40

40

20

20

20

20 40 Predicted KED

60

20 40 Predicted RK

60

20 40 Predicted

60

60

60

Observed

40

40

Data Points Trend line

20

20

1:1 line

20 40 Predicted

60

20 40 Predicted

60

Figure 2. Scatter plots of rain gauges observed rainfall versus estimated precipitation values from the MPE (no correction) and the four interpolation methods, at hour 8 on 24 April 2004

lists the four performance evaluation coefcients of ve precipitation maps for all the three hours. The coefcients for the MPE maps were directly calculated by comparing the radar cell values with co-located and concurrent gauge measurements. In addition, other properties [mean, maximum, minimum precipitation values, and coefcient of variation (CV)] of the ve maps for each hour are also included in Table I. The PB values show that the MPE underestimates 21, 42, and 22% of areal mean precipitation, respectively, for the 8th hour on 24 April 2004, the 8th hour on 22 June 2004, and the 9th hour on 22 June 2004, as compared with the gauge measurements. These PB values are much larger than the 7% underestimation for the same time period but only using uniform rainfall events reported by Wang et al. (2008). This suggests that for non-uniform rainfall events, the difference between radar estimates and gauge observations should be much larger than that of the uniform rainfall events. The PB values for four interpolation methods are between 004% (the BA method) and 189% (the KED method). The BA method shows the lowest PB values among the four methods, which suggest that the simple bias correction is effective in reducing the overall bias. However, in most cases, NSE of the BA interpolated map is relatively lower than those of the three Kriging methods. This indicates that, although the simple BA method effectively removed the overall bias, its performance in terms of reducing difference between observed and estimated rainfall values is not as good as the Kriging methods. R2 values of rainfall maps derived by the four methods as well as the MPE product are high and close to each other, except that MPE and BA obtained lower R2 values than three Kriging methods (085086 vs 092) for the
Copyright 2011 John Wiley & Sons, Ltd.

9th hour on 22 June 2004. Overall, the high R2 values indicate that all methods/maps have similar performance in capturing the precipitation distribution pattern; in other words, the R2 criterion does not assess the four methods for different performances. The scatter plots in Figure 2 show that, although the NEXRAD MPE product obtained high R2 value, its trend line is substantially deviating from the 1 : 1 line, as a result of the underestimation of MPE as compared with gauge data. After incorporating rain gauge observed values, the trend lines of the four methods are very close to the 1 : 1 line. Scatter plots for the two other hours (not shown) have similar results. In Figure 3, the general patterns of precipitation distribution obtained by different methods are similar, while the maximum rainfall rates from the four interpolation maps are distinguished larger than those from the MPE map (Table I). This is because that the relatively high rain gauge rainfall rates as compared with the MPE (i.e. negative PB values of the MPE maps) are transferred into the four interpolated maps. Other properties (mean, minimum, maximum, CV) of the four interpolated rainfall maps are similar. Overall performances of different interpolation methods at regional scale. We used box-and-whisker plots to show the evaluation coefcients of different methods for 737 h. The explanation of the symbols in Figure 4 is described as follows. The lower and upper lines of the box are the 25th and 75th percentiles of the samples. The line in the middle of the box is the sample median (or 50th percentile). The whiskers are lines extending above and below the box. The plus signs are values that are more than 15 times the interquartile range (difference between the 25th and 75th percentiles) away from the top
Hydrol. Process. 25, 37113720 (2011)

3716

H. XIE ET AL.

(1)

(2)

(3)
Figure 3. Spatial precipitation estimated by different methods at the 8th hour on 24 April 2004 (1), the 8th hour on 22 June 2004 (2), and the 9th hour on 22 June 2004 (3). (a, b, c, d, and e are the original MPE, BA, SKlm, KED, and RK maps, respectively)

or bottom of the box. The notches in the box represent a robust estimate of the uncertainty about the medians for box-to-box comparison. Boxes whose notches do not overlap indicate that the medians of the two groups differ at a signicance level of 005 (Matlab, 2007). In the plot of PB values of 737 rainfall hours (upper panel of Figure 4), we used absolute values, because the positive and negative may compensate with each other and lead to a false impression of small median and mean bias values. Similarly as shown in Table I, the MPE of 737 h has negative bias while all the other methods have positive bias. It is evident that median PB values of the MPE are much higher than those of the other four methods that incorporated with rain gauge data. In terms of PB, the BA method has the smallest median than RK and SKlm, whereas the PB values obtained by SKlm and RK are more compacted compared with those of BA and KED methods. The largest PB value of SKlm is 514%, which is much less than 50968%, the largest PB value of BA (not shown). In terms of NSE (Figure 4, lower panel) values, the SKlm and RK methods performed similar to each other and signicantly better than the other two methods. The MPE has the smallest NSE values. Plot of R2 (Figure 4, middle panel) indicates that the MPE
Copyright 2011 John Wiley & Sons, Ltd.

performed the best for capturing precipitation distribution pattern as compared with the rain gauge observations, followed by SKlm and RK. Overall, large bias (PB values) of MPE map as compared with the rain gauge observations indicate that it is preferred to incorporate high quality rain gauge data into MPE map to provide more accurate rainfall map at regional scale. Compared with MPE products, the simple BA method obtained much smaller PB, but cannot improve R2 and NSE. Among the four interpolations approaches, the overall performance of SKlm and RK are similar to or better than the others. Performance evaluation of different interpolation methods at local scale Comparison of rainfall values estimated by different methods at three rain gauges. In addition to accuracy evaluation at regional scale, we also conducted evaluation of the predicted rainfall at each rain gauge. The difference between the spatial precipitation maps, as exemplied in Figure 3, indicates that different methods predict varied rainfall values at local scale (e.g. at rain gauge point). For illustration purpose, we plot (Figures 5 and 6) the accumulated hourly rainfall of areal mean of the UGRB (a) and at three rain gauges (which are randomly chosen)
Hydrol. Process. 25, 37113720 (2011)

INTERPORATION METHODS FOR RAINFALL MAPPING

3717

methods. For example, at kr18, KED estimated rainfall depth is about 5 mm lower than that of SKlm and RK on 22 June 2004 (Figure 6c). Table II lists the evaluation coefcients for estimated rainfall time series at rain gauges kr4, kr18, and gp1 by different methods for 737 h in the year 2004. At some localized region, there may be substantial difference between the performances of difference interpolation methods. For example, at rain gauge gp1, PB values of KED and BA are 2851 and 5012%, respectively, which has signicant implication for local-scale hydrologic modelling and analysis. Similar as in Table I, the differences of R2 values between four methods are small and the SKlm and RK methods are in general better than BA and KED methods, whereas KED and BA perform similar. Overall performances of different interpolation methods at local scale. We used box-and-whisker plots to show the evaluation coefcients of different methods at the 50 rain gauges. The plot of PB values at 50 rain gauges (upper panel of Figure 7) shows that all methods have similar median values. The largest absolute PB values of BA and MPE are 5012 and 4434%, respectively, which are larger than those of SKlm, KED, and RK (ranging between 3484 and 3640). For R2 (Figure 7, middle panel), MPE and KED obtained slightly higher median than the other methods. In terms of NSE (Figure 7, bottom panel), the SKlm and RK methods performed similar to each other and signicantly better than the other methods. In contrast to regional-scale evaluation, BA does not perform the best for removing the bias at local scale.

Figure 4. Box-and-whisker plots of the PB (PBIAS) (%), R2 , and NSE values of the selected 737 h in 2004 for the MPE (no correction) and the four interpolation methods (BA, SKlm, KED, and RK)

(b, c, d) from six different methods (gauges only, no correction or MPE only, BA, SKlm, KED, and RK) on 2 days (9 June 2004 and 22 June 2004). It is found, on 9 June 2004 (Figure 5a), that the accumulated areal mean precipitation amounts estimated by the four interpolation methods are very close to the rain gauge observed values. The MPE estimates similar rainfall as the four methods before the 12th hour and more rainfall then after. On 22 June 2004 (Figure 6a), the four methods and rain gauge get similar areal mean rainfall accumulation values, whereas MPE shows less accumulated rainfall (by 27%). Although the four interpolation methods perform similar to each other for areal mean precipitation estimation, substantial difference between the rainfall values estimated at local scale (rain gauge locations: kr4, kr18, and gp1) are shown in parts bd of Figures 5 and 6. For example, MPE and BA show higher rainfall than rain gauge observed, whereas SKlm, KED, and RK show lower rainfall at rain gauge kr4 on 9 June 2004 (Figure 5b). On 22 June 2009, at rain gauge gp1, BA estimated much lower rainfall than SKlm, KED, and RK. For the three Kriging methods, there is appreciable difference between rainfall values estimated by the three
Copyright 2011 John Wiley & Sons, Ltd.

DISCUSSION AND CONCLUSIONS In Sections Performance Evaluation of Different Interpolation Methods for Regional-Scale Rainfall Estimation and Performance Evaluation of Different Interpolation Methods at Local Scale, we compared four spatial rainfall interpolation methods as well as the MPE maps for capturing the observed rainfall characteristics at regional and local scales. For regional-scale assessment, the high R2 values indicate that the NEXRAD MPE products are able to capture the spatial precipitation distribution patterns well, but with large bias as compared with rain gauge observed values, the median and mean absolute PB values of 22 and 42%, respectively. With respect to selecting methods for regional-scale spatial rainfall estimation, the inconsistent performance of different methods for each time-step and each evaluation coefcient makes it difcult to simply choose one optimal method. The number of hours that different methods performed the best among 737 h is listed in Table III. In terms of the number of hours that a method performs the best, BA, MPE, and SKlm perform the best for PB, R2 , and NSE, respectively. It is also worth noting that the number of hours that one method performed the best for one evaluation coefcient is less than 62% of 737 h. By
Hydrol. Process. 25, 37113720 (2011)

3718
(a)

H. XIE ET AL.

(b)

(c)

(d)

Figure 5. Accumulated hourly precipitation of areal mean (a) and at three rain gauges (b, c, d) estimated by different methods on 9 June 2004. No correction means the original MPE data. (b, c, d denotes rain gauge kr4, kr18, gp1, respectively)
(a) (b)

(c)

(d)

Figure 6. Accumulated hourly precipitation of areal mean (a) and at three rain gauges (b, c, d) estimated by different methods on 22 June 2004. No correction means the original MPE data. (b, c, d denotes rain gauge kr4, kr18, gp1, respectively)

selecting one method, we may not obtain optimal results for other portion of the hours (more than 38%). Local-scale evaluation results indicate that rainfall values estimated by different methods at a specic rain gauge can be substantially different from each other. This, to some extent, reveals the importance of evaluating the performance of different methods at local scale. The number of rain gauges that different methods performed the best is listed in Table IV. Although, SKlm performed the best at the all 50 rain gauges in terms of NSE, it only
Copyright 2011 John Wiley & Sons, Ltd.

outperformed the other methods at 5 and 1 rain gauges in terms of PB and R2 , respectively. Spatial precipitation data are critical for accurate ash ood forecasting and hydrologic budget estimation. In the study, we incorporate the rain gauge data into NEXRDA MPE products, using four methods (i.e. BA, SKlm, KED, and RK) and evaluate their performances using PB, R2 , and NSE at regional and local scales. The results show that the NEXRAD MPE data can capture the patterns of spatial precipitation distribution well, but
Hydrol. Process. 25, 37113720 (2011)

INTERPORATION METHODS FOR RAINFALL MAPPING

3719

Table II. Evaluation parameters of predicted areal mean precipitation and precipitation at three rain gauges for the year of 2004 Evaluation coefcients MPE kr4 PB R2 NSE PB R2 NSE PB R2 NSE 433 077 079 654 065 065 3641 080 069 BA 392 076 084 015 065 074 5012 081 073 SKlm 021 076 088 113 065 080 3198 080 079 KED 555 077 083 052 064 073 2851 081 071 RK 025 076 088 057 065 079 3224 080 079

Table IV. The numbers of rain gauges (total 50) that different methods performed the best for different evaluation coefcients Methods Evaluation coefcient PB R2 NSE MPE 14 14 0 BA 13 7 0 SKlm 5 1 50 KED 13 25 0 RK 5 3 0

kr18

gp1

PB is expressed in %; others have no units.

there is substantial discrepancy between the rain gauge and NEXRAD estimated precipitation. The comparison between different spatial interpolation methods indicates that, although no one method can consistently outperform the other methods in terms of all evaluation coefcients for all time-steps and at all rain gauges, the two Krigingbased methods (SKlm and RK) are in general better than BA and KED, and the PB and NSE criteria are better than the R2 criterion. It is worth noting that the performance of one method for regional-scale evaluation may be different from its performance for local-scale evaluation. This indicates the importance of careful evaluation of the performance of different methods at local or regional scale for different purposes. The results obtained in this study are expected to contribute to the development of spatial rainfall products for hydrologic budget and ash ood modelling.

ACKNOWLEDGEMENTS

This study was in part supported by the U.S. NASA grant (#NNX07AL79G). Provisions of NEXRAD data through Greg Story at the National Weather Service West Gulf River Forecast Center and rain gauge data from Guadalupe Blanco River Authority are sincerely acknowledged. Critical reviews and constructional comments from three anonymous reviewers to improve the quality of this manuscript are greatly appreciated.

REFERENCES Figure 7. Box-and-whisker plots of the PB, R2 , and NSE values at the 50 rain gauges for the MPE (no correction) and the four interpolation methods (BA, SKlm, KED, and RK) Arnaud P, Bouvier C, Cisneros L, Dominguez R. 2002. Inuence of rainfall spatial variability on ood prediction. Journal of Hydrology 260: 216230. Bedient PB, Hoblit BC, Gladwell DC, Vieux BE. 2000. NEXRAD radar for ood prediction in Houston. Journal of Hydrologic Engineering 5: 269277. Bradley AA, Peters-Lidard C, Nelson BR, Smith JA, Yong CB. 2002. Raingauge network design using NEXRAD precipitation estimate. Journal of the American Water Resources Association 38: 13931407. Chaplot V, Saleh A, Jaynes DB. 2005. Effect of the accuracy of spatial rainfall information on the modeling of water, sediment, and NO3 -N loads at the watershed level. Journal of Hydrology 312: 223234. Cheng S, Hsiedh H, Wang Y. 2007. Geostatistical interpolation of spacetime rainfall on Tamshui river basin, Taiwan. Hydrological Processes 21: 31363145. Goovaerts P. 1997. Geostatistics for Natural Resources Evaluation . Oxford University Press: New York. Goovaerts P. 2000. Geostatistical approaches for incorporating elevation into the spatial interpolation of rainfall. Journal of Hydrology 228: 113129. Haberlandt U. 2007. Geostatistical interpolation of hourly precipitation from rain gauges and radar for a large-scale extreme rainfall event. Journal of Hydrology 332: 144157. Hydrol. Process. 25, 37113720 (2011)

Table III. The number of hours that different methods performed the best for different evaluation coefcients Methods Evaluation coefcient PB R2 NSE MPE 27 422 202 BA 450 5 54 SKlm 106 162 320 KED 98 102 108 RK 121 108 104

The sum of the numbers in each row may be larger than 737 because for some hours there are several methods performing the same for a specic evaluation coefcient.

Copyright 2011 John Wiley & Sons, Ltd.

3720

H. XIE ET AL. 2004. Unique aspects of proposed San Antonio/Guadalupe Hydrologic Observatory in Texas . http://adsabs.harvard.edu/abs/2004AGUFM. H31C0393S (accessed on 9 January 2008). Seo DJ. 1998. Real-time estimation of rainfall elds using radar rainfall and rain gauge data. Journal of Hydrology 208: 3752. Seo D-J, Breidenbach JP. 2002. Real-time correction of spatially nonuniform bias in radar rainfall data using rain gauge measurements. Journal of Hydrometeorology 3: 93111. Steiner M, Smith JA, Burges SJ, Alonso CV, Darden RW. 1999. Effect of bias adjustment and rain gauge data quality control on radar rainfall estimation. Water Resources Research 35(8): 24872503. Wang X, Xie H, Sharif H, Zeitler J. 2008. Validating NEXRDA MPE and stage III precipitation products for uniform rainfall on the Upper Guadalupe River Basin of the Texas Hill Country. Journal of Hydrology 348(12): 7386. Westcott NE, Knapp HV, Hilberg SD. 2008. Comparison of gage and multi-sensor precipitation estimates over a range of spatial and temporal scales in the Midwestern United States. Journal of Hydrology 351(12): 112. Xie H, Zhou X, Vivoni E, Hendrickx J, Small E. 2005. GIS based NEXRAD precipitation database: automated approaches for data processing and visualization. Computational Geosciences 31(1): 6576. Xie H, Zhou J, Hendrickx E, Vivoni H, Guan H, Tian YQ, Small EE. 2006. comparison of NEXRAD stage III and gauge precipitation estimates in central New Mexico. Journal of the American Water Resources Association 42(1): 237256. Young CB, Nelson BR, Bradley AA, Krajewski WF, Kruger A. 2000. Evaluating NEXRAD multisensor precipitation estimates for operational hydrologic forecasting. Journal of Hydrometeorology 1(3): 241254. Younger PM, Freer JE, Beven KJ. 2009. Detecting the effects of spatial variability of rainfall on hydrological modelling within an uncertainty analysis framework. Hydrological Processes 23: 19882003. Zhang X, Srinivasan R. 2009. A comparison of geostatistical approaches for spatial precipitation estimation. Journal of the American Water Resources Association 45(4): 894906.

Habib E, Larson BF, Graschel J. 2009. Validation of NEXRAD multisensory precipitation estimates using an experimental dense rain gauge network in South Louisiana. Journal of Hydrology 373(34): 463478. Hengl T, Heuvelink GBM, Stein A. 2004. A generic framework for spatial prediction of soil variables based on regression-kriging. Geoderma 120: 7593. Hossain F, Anagnostou EN, Borga M, Dinku T. 2004. Hydrological model sensitivity to parameter and radar rainfall estimation uncertainty. Hydrological Processes 18(17): 32773299. Isaaks EH, Srivastava RM. 1989. Applied Geostatistics . Oxford University Press: New York. Jayakrishnan R, Srinivasan R, Arnold JG. 2004. Comparison of raingauge and WSR-88D stage III precipitation data over the TexasGulf River Basin. Journal of Hydrology 292: 135152. Kalin L, Hantush MM. 2006. Hydrologic modeling of an Eastern Pennsylvania watershed in Middle Tennessee. Journal of Hydrologic Engineering 9(5): 555569. Kalinga AO, Gan TY. 2006. Semi-distributed modelling of basin hydrology with radar and gauged precipitation. Hydrological Processes 20: 37253746. Knebl MR, Yang Z, Hutchison K, Maidment DR. 2005. Regional scale ood modeling using NEXRAD rainfall, GIS, and HEC-HMS/RAS: a case study for the San Antonio River Basin Summer 2002 storm event. Journal of Environmental Management 75(4): 325336. Krajewski WF, Smith JA. 2002. Radar hydrology: rainfall estimation. Advances in Water Resources 25: 13871394. Li B, Eriksson M, Srinivasan R, Sherman M. 2008. A geostatistical method for Texas NexRad data calibration. Environmetrics 19(1): 119. Lloyd CD. 2005. Assessing the effect of integrating elevation data into the estimation of monthly precipitation in Great Britain. Journal of Hydrology 308: 128150. Matlab. 2007. MATLAB Technical Documentation . The MathWorks Inc.: Natick, Massachusetts. Nash JE, Sutcliffe JV. 1970. River ow forecasting through conceptual models: part I. A discussion of principles. Journal of Hydrology 10(3): 282290. Scanlon BR, Musgrove M, Sansom A, Xie H, Wilcox B, Sharp J, Alexander C, Vieux B, Jackson R, Archer S, Keese K, Reedy R.

Copyright 2011 John Wiley & Sons, Ltd.

Hydrol. Process. 25, 37113720 (2011)

Você também pode gostar