Sie sind auf Seite 1von 9

HYDROLOGICAL PROCESSES Hydrol. Process. 19, 507 515 (2005) Published online 8 December 2004 in Wiley InterScience (www.interscience.wiley.com).

DOI: 10.1002/hyp.5675

How to construct recursive digital lters for baseow separation


K. Eckhardt*
METCON Umweltmeteorologische Beratung, Jappopweg 9h, 25421 Pinneberg, Germany

Abstract:
Recursive digital ltering of hydrographs is a baseow separation method that can easily be automated and has been recommended for providing reproducible results. In the past, different formulations of the most simple lter type, the so-called one-parameter lter, have been proposed. In this paper, a theoretical framework is developed for lter algorithms that were constructed under the assumption that the outow from an aquifer is linearly proportional to its storage. It is shown that these one-parameter lters describing an exponential baseow recession are all special cases of a two-parameter lter whose equation is specied. Its parameters are the recession constantwhich can be objectively determined by a recession analysisand BFImax , the maximum value of the baseow index that can be modelled by the algorithm. This introduces a subjective element into the baseow calculation, since BFI max is not measurable. A preliminary analysis based on the results of conventional separation techniques shows that it might be possible to nd typical BFImax values for classes of catchments that can be unequivocally distinguished by their hydrological and hydrogeological characteristics. Copyright 2004 John Wiley & Sons, Ltd.
KEY WORDS

hydrograph; baseow separation; recursive digital ltering; model validation

INTRODUCTION There is widespread agreement that good correspondence of measured and calculated streamow at the catchment outlet is not a sufcient criterion for the validity of a physically based hydrologic model, but that additional knowledge concerning catchment-internal processes is needed (e.g. Beven, 1989). Yet, for many catchments such information is not available and the models performance can only be assessed by comparing calculated and measured streamow. However, the information provided by a hydrograph is not limited to the absolute height of the runoff. For example, a separation of the hydrograph allows for identication of runoff components originating from different compartments of the considered system, and responding differently delayed and smoothed to rainfall events. An example of the potential benet of this information for model verication has been given by Eckhardt et al. (2002). In this paper, a special separation technique is considered, the recursive digital ltering of hydrographs. It serves to partition the streamow into two components, direct runoff and baseow: yk D fk C bk with y f b k D D D D total streamow direct runoff baseow time step number 1

* Correspondence to: Dr K. Eckhardt, METCON Umweltmeteorologische Beratung, Jappopweg 9h, 25421 Pinneberg, Germany. E-mail: eckhardt@metcon-umb.de

Copyright 2004 John Wiley & Sons, Ltd.

Received 27 March 2003 Accepted 8 January 2004

508

K. ECKHARDT

Baseow is usually associated with water discharged from ground water storage. This process provides considerable smoothing. Hence, in the frequency spectrum of a hydrograph, long waves will be more likely to be associated with baseow while the high frequency variability of the streamow will primarily be caused by direct runoff. It should therefore be possible to identify the baseow by low-pass ltering the hydrograph. In this paper, the focus lies on the most simple lter type, the so-called one-parameter lters. One-parameter lters have been presented by Lyne and Hollick (1979), Chapman (1991) and Chapman and Maxwell (1996), for example. A theoretical framework is developed for lter algorithms that were constructed under the assumption that the outow from an aquifer is linearly proportional to its storage. A generally valid formulation of these lters is provided that facilitates their construction, yet also reveals some fundamental problems in their application. PREVIOUSLY PUBLISHED ONE-PARAMETER FILTERS Lyne and Hollick (1979) proposed the lter equation fk D afk subject to fk 0 or, in terms of the baseow, yk C yk 1 3 2 subject to bk yk , where a is the lter parameter. The LyneHollick algorithm has been used by Nathan and McMahon (1990), Arnold and Allen (1999) and Arnold et al. (2000), for instance. Yet, Chapman (1991) pointed out that the LyneHollick algorithm incorrectly provides a constant streamow y or baseow b, respectively when direct runoff has ceased (f D 0) and therefore developed the new algorithm bk D abk
1 1

1Ca yk 2 1 a

yk

bk D

3a 1 bk 3 a

1 3 a

a yk C yk a

subject to bk yk . Equivalent to Equation (4) is the formulation fk C fk 1 5 2 When the quick runoff component has ceased, the lter parameter now gets the hydrological signicance of the recession constant a in the equation 6 bk D abk 1 bk D abk
1

most used for describing the baseow recession during dry-weather periods without groundwater recharge. Therefore, the lter parameter a can now be objectively derived by a recession analysis. Chapman and Maxwell (1996) simplied Equation (4) to bk D a 2 a bk
1

1 2

a yk a

and showed that this simpler formulation provides almost identical results to the earlier form. GENERAL FORMULATION OF THE ONE-PARAMETER FILTER In the following, Equation (7) serves as a copy for the general form of a recursive one-parameter lter. It is assumed that bk can be calculated as bk D Abk 1 C Byk 8
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

RECURSIVE DIGITAL FILTERS FOR BASEFLOW SEPARATION

509

subject to bk yk . A and B are functions of the lter parameter a with A > 0 and B > 0. They shall be expressed in a universally valid, hydrologically meaningful way. As yk D bk C fk , Equation (8) can be written bk D Abk C Bfk 1 B
1

It is claimed that, as in the case of the algorithms of Chapman (1991) and Chapman and Maxwell (1996), the lter parameter equals the groundwater recession constant a in Equation (6). If fk D 0, Equation (9) assumes the form A bk D 10 bk 1 1 B Therefore, 1 A B Da 11 n time steps before. We get

As 0 < a < 1, A > 0 and B > 0, it follows that A < 1. Now, the baseow bk at the time step k is traced back to the baseow bk bk D Abk
1

C Byk
2

D A Abk

C Byk
3

C Byk
2

D A[A Abk D ...

C Byk
n 1

C Byk 1 ] C Byk

D An bk

CB
iD0

Ai y k

12
i

Assuming the hypothetical case of a constant streamow (yk be formulated as bk D An bk


n C Byk

D yk for all i D 1, . . . , k ), Equation (12) can Ai 13

n 1 iD0

If we lter a time series of innite length (n approximates 1) we get


n 1

bk D lim An bk
n!1

C Byk
iD0

Ai 14

B 1 A

yk

because A < 1. Application of the recursive lter to a constant time series corresponds to low-pass ltering of a wave of innite length. This wave should pass the lter with the least attenuation. Therefore, this calculation shows the maximum value BFImax of the baseow index (the long-term ratio of baseow to total streamow) that can be modelled by the algorithm: B BFImax D 15 1 A Assuming that in dry-weather periods without groundwater recharge baseow recesses exponentially, Equation (11) was obtained which gives A D 1 B a. On the other hand, if BFImax were equal to 1 it would follow from Equation (15) that B D 1 A and hence A D [1 1 A ] a D Aa. This equation is
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

510

K. ECKHARDT

incorrect because a < 1. Hence, BFImax must be less than 1 as well. There is no wave that passes the lter without attenuation. A and B can now be expressed as functions of the lter parameter a and of BFI max . In Equation (15), A is replaced by 1 B a: B 16 BFImax D 1 1 Ba It follows that BD and AD 1 B a 1 BFImax a D 1 aBFImax 1 1 a BFImax aBFImax 17

18

Hence, the general form of a one-parameter lter whose parameter a corresponds to the groundwater recession constant is 1 BFImax abk 1 C 1 a BFImax yk bk D 19 1 aBFImax subject to bk yk . An equivalent algorithm bk D a bk 1CC
1

C yk 1CC

20

with parameters a and C is used in the hydrologic model AWBM (Boughton, 1993). Equation (19) is obtained from Equation (20) by setting 1 a BFImax CD 21 1 BFImax DISCUSSION A key assumption in the derivation of Equation (19) is that the outow from the aquifer is linearly proportional to its storage. This assumption leads to the model of an exponential baseow recession when there is no groundwater recharge. The linear storage model has been questioned (Wittenberg, 1999). However, Chapman (1999) showed that for the commonly occurring case of recession of duration up to about 10 days, the linear model remains a very good approximation. Baseow separation by recursive digital ltering corresponds to a low-pass ltering of hydrographs. Yet, it has been shown in the previous section that the assumption of an exponential baseow recession implies that no wave passes the lter without attenuation or that BFImax < 1, respectively. At rst glance, this result may seem inacceptable because one tends to attribute long waves completely to the baseow. However, Spongberg (2000) has demonstrated that the frequency spectrum of direct runoff has a broad bandwidth with a non-negligible low-frequency part. So the characteristic of the lter (19) to relate some of the low-frequency variability of the streamow to the direct runoff could be benecial. As BFImax 6D 1 the user of the lter technique now faces the problem that he has not only to specify the parameter a, but also a value for BFImax < 1. If, for example, we set BFImax D 05 in Equation (19) we obtain Equation (7), the lter algorithm of Chapman and Maxwell (1996). If, however, we choose a different BFI max then we will get another lter that yields a different baseow. The one-parameter lter has to be seen as a two-parameter lter.
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

RECURSIVE DIGITAL FILTERS FOR BASEFLOW SEPARATION

511

SENSITIVITY ANALYSIS Equation (19) has two parameters: a and BFImax . In the following, a sensitivity analysis is carried out in order to see which one of these parameters inuences the ltering result the most. To this purpose, one parameter, respectively, is changed while keeping the other one constant. A sensitivity index I is calculated as ID relative change of the calculated mean baseow relative change of the parameter 22

Streamow records from two catchments are used: a hydrograph measured at the outlet of the Beaverdam Creek catchment in Maryland, USA (BFI 072, see later) and a hydrograph of the Aar, a small river in the southeast of the Rhenish Massif in Germany (BFI 023, see later). a is varied by 0025 and BFImax by 005. The results of the sensitivity analysis are shown in Table I. The lter parameter a exerts a denitely weaker inuence on the calculated mean baseow than BFI max . This is problematic inasmuch as a can be determined by a recession analysis while BFImax is non-measurable. So the result of the recursive digital ltering strongly depends on a quantity that seems to be merely a tting parameter. We can only circumvent this problem if we succeed in empirically nding characteristic BFImax values for classes of catchments that can be unequivocally distinguished by their hydrological and/or hydrogeological characteristics. This is the objective of the next section.

FILTERING HYDROGRAPHS OF CATCHMENTS WITH DIFFERENT CHARACTERISTICS The necessity to select a value of BFImax introduces a subjective element into the baseow calculation. The objective of this section is to minimize this subjective inuence by derivation of characteristic BFImax values for different hydrological and hydrogeological situations. To this purpose, results of conventional separation methods are compared with those of recursive digital lters as dened by Equation (19). The lters are passed once over the hydrograph. For each considered catchment (Table II), the recession constant or the lter parameter a, respectively, is determined by constructing a master recession curve using the matching strip method (see e.g. Nathan and McMahon, 1990). For the Hadley Creek catchment, where the resolution of the streamow time series is not sufcient to construct a master recession curve, the same recession constant is assumed as for the Goose Creek catchment. Perennial streams, porous aquifers In the following, data collected in two Northern American catchments is used, the Brandywine Creek catchment in Pennsylvania and the Beaverdam Creek catchment in Maryland. They were, among others, used by Arnold and Allen (1999) to validate the LyneHollick algorithm. Observed streamow is provided by the US Geological Survey. The eld-based estimates of baseowexpressed by the BFI (Table III, last
Table I. Results of the sensitivity analysis Recession constant a 0925 0925 0925 0900 0950 BFImax Beaverdam Creek calculated base ow (m3 /s) 0439 0411 0467 0444 0431 Sensitivity index I BFImax Aar calculated base ow (m3 /s) 0362 0291 0433 0362 0362 Sensitivity index I

075 070 080 075 075

096 096 042 067

025 020 030 025 025

098 098 000 000

Copyright 2004 John Wiley & Sons, Ltd.

Hydrol. Process. 19, 507 515 (2005)

512

K. ECKHARDT

Table II. Characteristics of the considered catchments Catchment Brandywine Creek, PA, USA Beaverdam Creek, MD, USA Goose Creek, IL, USA Hadley Creek, IL, USA Upper Dill, Germany Dietzh olze, Germany Aar, Germany Flow regime perennial perennial ephemeral ephemeral perennial perennial perennial Area (km2 ) 743 50 122 188 63 81 134 Mean annual precipitation (mm) 1120 1075 940 914 950 950 950

Table III. Baseow indices of Northern American catchments with porous aquifers Catchment Filter parameter a Recursive digital ltering Predened BFImax Brandywine Creek January 1928December 1931 Brandywine Creek January 1952September 1953 Beaverdam Creek April 1950March 1952 Goose Creek January 1955December 1958 Hadley Creek April 1956September 1958 0975 0980 0980 0963 0963 080 080 080 050 050 Calculated BFI 072 058 071 042 020 070 065 072 049 015 Independently estimated BFI

column)are adopted from Arnold and Allen (1999). The maximum of these eld-based estimates is 072. With BFImax D 08 the recursive digital lter bk D abk
1

C 4 1 a yk 5 4a

23

is obtained from Equation (19). Application of this lter (Figure 1) provides a baseow that corresponds very well to the eld-based estimates of the BFI. Ephemeral streams, porous aquifers The data in this subsection stems from the same sources as in the previous section. This time the Goose Creek and Hadley Creek catchment in Illinois, USA are considered. In these catchments groundwater runoff was estimated based on standard hydrograph separation techniques (Arnold and Allen, 1996). This time, BFImax D 05 is selected, i.e. the algorithm of Chapman and Maxwell (1996) (Equation (7)) is used to calculate the baseow. In case of the Goose Creek catchment (Figure 2), the independently estimated BFI is underestimated while the BFI of the Hadley Creek catchment is overestimated by the lter (Table III). Perennial streams, hard rock aquifers In this subsection, three mesoscale low mountain range catchments situated in the southeast of the Rhenish Massif in Germany serve as examples for catchments with mainly hard rock aquifers. Kaviany (1978) analysed
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

RECURSIVE DIGITAL FILTERS FOR BASEFLOW SEPARATION

513

4 mm/d 2 0 1950-11-01 1950-12-01 1951-01-01 1951-02-01 1951-03-01 1951-04-01 1951-05-01 1951-06-01 1951-07-01 1951-08-01 1951-09-01 1951-10-01 1957-07-01 1951-11-01 1957-09-01

measured streamflow

calculated baseflow

Figure 1. Measured streamow and calculated baseow at the outlet of the Beaverdam Creek catchment in Maryland, USA

10 8 6 4 2 0

mm/d

1956-01-01

1956-03-01

1956-05-01

1956-07-01

1956-09-01

1956-11-01

1957-01-01

1957-03-01

measured streamflow

calculated baseflow

Figure 2. Measured streamow and calculated baseow at the outlet of the Goose Creek catchment in Illinois, USA

the streamow at the catchment outlets with two different methods of baseow separation: a simple graphical method proposed by Natermann (1951) and an approach based on an evaluation of the monthly minimum runoff developed by Kille (1970). The resulting baseow indices lie between 020 and 025. Dening BFImax D 025 the lter assumes the form bk D 3abk
1

C 1 4 a

a yk

1957-05-01

24

For all three catchments, BFI D BFImax and hence a good correspondence with the independently estimated BFI values is found (Figure 3, Table IV).
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

514

K. ECKHARDT

12 10 8 mm/d 6 4 2 0 1986-11-01 1987-01-01 1987-03-01 1987-05-01 1987-07-01 1987-09-01 1987-11-01 1988-01-01 1988-03-01 1988-05-01 1988-07-01

measured streamflow

calculated baseflow

Figure 3. Measured streamow and calculated baseow at the outlet of the Aar catchment in Germany

Table IV. Baseow indices of German catchments with hard rock aquifers Catchment Filter parameter a Recursive digital ltering Predened BFImax Upper Dill January 1961December 1975 Dietzh olze January 1972December 1975 Aar January 1961December 1975 0955 0955 0959 025 025 025 Calculated BFI 025 025 025 025 020 023 Independently estimated BFI

DISCUSSION AND CONCLUSIONS The general form of a so-called one-parameter algorithm for baseow separation by recursive digital ltering is given by Equation (19). This lter describes an exponential baseow recession during periods without groundwater recharge. The lter parameter corresponds to the recession constant and thus can be determined by a recession analysis. Yet, Equation (19) shows that it is misleading to speak of a one-parameter lter. There is an innite number of potential lter formulations depending on a second parameter, BFImax , the maximum value of the baseow index that can be modelled by the algorithm. Unfortunately, BFImax on the one hand is a sensitive parameter and on the other hand is a non-measurable quantity. In order to minimize the subjective inuence that a user of the lter technique exerts on the results by his choice of BFImax , attempts have been made to nd typical BFImax values for classes of catchments that can be distinguished by their hydrological and hydrogeological characteristics. The previous section shows that this attempt might be successful. Based on the presented preliminary results it is suggested to select BFImax 080 for perennial streams with porous aquifers, BFImax 050 for ephemeral streams with porous aquifers, and BFImax 025 for perennial streams with hard rock aquifers. However, this analysis is far from being complete and has to be continued. It would not only be favourable if more data was available. One must
Copyright 2004 John Wiley & Sons, Ltd. Hydrol. Process. 19, 507 515 (2005)

RECURSIVE DIGITAL FILTERS FOR BASEFLOW SEPARATION

515

also be aware of the fact that rather simple separation methods were used to calibrate the algorithm (19) with respect to BFImax . An optimization according to the results of newer approaches, e.g. of tracer experiments, would possibly lead to other recommendations for BFImax . A special problem arises for small BFImax values. The difference between the actual and maximum value of the BFI depends on the extent to which the constraint bk yk takes effect. The smaller BFImax and, hence, the lower the baseow is compared to the total streamow, the smaller is the inuence of this constraint on the ltering result. Therefore, when the streamow of the catchments with hard rock aquifers is ltered, no difference between the calculated BFI values and the predened BFI max D 025 is found, although a certain variability of the baseow index should be expected as indicated by the independently estimated BFI values. It has yet to be tested whether this class of catchments requires a fundamentally different type of lter algorithm. Furthermore, it should be noted that the ltering results will probably be inuenced by the size of the considered catchment as well. The longer the travel time to the catchment outlet, the more runoff peaks are dispersed. This will feign a systematically higher baseow contribution than in smaller catchments. The database available for the present study, however, is not sufcient to test this assumption in practice.

ACKNOWLEDGEMENTS

The author thanks Tom Chapman for his valuable suggestions.


REFERENCES Arnold JG, Allen PM. 1996. Estimating hydrologic budgets for three Illinois watersheds. Journal of Hydrology 176: 57 77. Arnold JG, Allen PM. 1999. Validation of automated methods for estimating baseow and groundwater recharge from stream ow records, Journal of the American Water Resources Association 35: 411 424. Arnold JG, Muttiah RS, Srinivasan R, Allen PM. 2000. Regional estimation of base ow and groundwater recharge in the Upper Mississippi river basin. Journal of Hydrology 227: 2140. Beven KJ. 1989. Changing ideas in hydrology: the case of physically-based models. Journal of Hydrology 105: 157172. Boughton WC. 1993. A hydrograph-based model for estimating the water yield of ungauged catchments. Hydrology and Water Resources Symposium, Institution of Engineers Australia, Newcastle; 317 324. Chapman TG. 1991. Comment on Evaluation of automated techniques for base ow and recession analyses by R. J. Nathan and T. A. McMahon. Water Resources Research 27: 1783 1784. Chapman TG. 1999. A comparison of algorithms for stream ow recession and baseow separation. Hydrological Processes 13: 701714. Chapman TG, Maxwell AI. 1996. Baseow separationcomparison of numerical methods with tracer experiments. Hydrological and Water Resources Symposium, Institution of Engineers Australia, Hobart; 539 545. Eckhardt K, Haverkamp S, Fohrer N, Frede H-G. 2002. SWAT-G, a version of SWAT99.2 modied for application to low mountain range catchments. Physical Chemistry of the Earth 27: 641644. Kaviany E. 1978. Zur Hydrogeologie im Niederschlagsgebiet der Dill (Hessen). Gieener Geologische Schriften 19: 248. Kille K. 1970. Das Verfahren MoMNQein Beitrag zur Berechnung der mittleren langj ahrigen Grundwasserneubildung mit Hilfe der monatlichen Niedrigwasserabu sse. Zeitschrift der Deutschen Geologischen Geseltschaft, Sonderheft Hydrogeologie/Hydrogeochemic : 8995. Lyne VD, Hollick M. 1979. Stochastic time-variable rainfall runoff modelling. Hydrology and Water Resources Symposium, Institution of Engineers Australia, Perth; 89 92. Natermann E. 1951. Die Linie des langfristigen Grundwassers (Au L) und die Trockenwetterabulinie (TWL). Die Wasserwirtschaft, Sonderh. 1950: 1220. Nathan RJ, McMahon TA. 1990. Evaluation of automated techniques for base ow and recession analyses. Water Resources Research 26: 1465 1473. Spongberg ME. 2000. Spectral analysis of base ow separation with digital lters. Water Resources Research 36: 745752. Wittenberg H. 1999. Baseow recession and recharge as nonlinear storage processes. Hydrological Processes 13: 715726.

Copyright 2004 John Wiley & Sons, Ltd.

Hydrol. Process. 19, 507 515 (2005)

Das könnte Ihnen auch gefallen