Journal of Hydrology (2008) 352, 168 173

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/jhydrol

A comparison of baseflow indices, which were

calculated with seven different baseflow separation
methods
K. Eckhardt *

University of Hohenheim, Institute of Physics and Meteorology, Garbenstrasse 30, 70593 Stuttgart, Germany

Received 4 June 2007; received in revised form 10 January 2008; accepted 15 January 2008

KEYWORDS Summary Baseflow indices for 65 North American catchments are compared, which were
Baseflow separation; calculated with seven different baseflow separation methods (HYSEP1, HYSEP2, HYSEP3,
Baseflow index; PART, BFLOW, UKIH, and Eckhardt). Since the true values of the baseflow index are
Recession analysis unknown, it cannot be said, which one of the methods gives the best estimates. However,
the results of the Eckhardt method appear to be hydrologically more plausible than those
of the other algorithms. Furthermore, this algorithm can be applied to hydrographs of any
time step length. A simple technique for assessing the recession constant one of the
parameters of the algorithm is presented.
2008 Elsevier B.V. All rights reserved.

Introduction
Baseflow is a streamflow component which reacts slowly to
rainfall and is usually associated with water discharged from
groundwater storage. Knowledge about baseflow is useful in
the assessment of water quality and low-flow conditions.
It can also be used to calibrate or validate hydrological
models.
Neff et al. (2005) presented a study, which aimed at
producing estimates of baseflow in watersheds tributary to
the Great Lakes in North America. In order to develop
regression models describing baseflow in ungaged areas,
they calculated baseflow indices (the long-term ratio of
baseflow to total streamflow) for 959 gages with six differ-
ent separation methods: the fixed-interval, sliding-interval,
and local minimum HYSEP methods (Sloto and Crouse,
1996), PART (Rutledge, 1998), BFLOW (Arnold and Allen,
1999), and UKIH (Piggott et al., 2005).
In the present study, a selection of these results is com-
pared to those provided by a separation method developed
by Eckhardt (2005). Furthermore, a simple technique for
assessing the recession constant one of the parameters
of the algorithm is presented.
Data

* Tel.: +49 711 45922152; fax: +49 711 45922461. From the 959 gages whose streamflow was analysed by Neff
E-mail address: eckhardt@uni-hohenheim.de et al. (2005), 65 USGS gages were randomly selected for the

0022-1694/$ - see front matter 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.jhydrol.2008.01.005
A comparison of baseflow indices, which were calculated with seven different baseflow separation methods
present study, and the data (mean daily discharge) re-
trieved from the USGS National Water Information System.
The baseflow index (BFI) values provided by the three
HYSEP methods, PART and UKIH, which are presented in this
study, are taken from the report of Neff et al. (2005). Neff
et al. (2005) also showed results from BFLOW, which were
obtained after three passes of the filter over the streamflow
time series (see below for technical details). For this study,
the BFI values from BFLOW were newly calculated by pass-
ing the filter only once over the time series, as it is more
common for the filter technique. This results in higher BFI
values for this method.
Methods
Introduction
The methods whose results are discussed in this paper serve
to partitioning the streamflow in two components, direct
runoff and baseflow
y k fk b k 1 for perennial streams with hard rock aquifers. In the present
where y is the total streamflow; f is the direct runoff; b is
the baseflow; and k is the time step number.
The assumption that the outflow from the aquifer is lin-
early proportional to its storage leads to the model of an
exponential baseflow recession during periods without
groundwater recharge
bk1 bk eDt=s 2 Recession analysis
where Dt is the time step length and s is the characteristic
time constant, or
bk1 abk 3
respectively, with the recession constant
a eDt=s 4
Baseflow separation
HYSEP is a computer program, which mimics three manual
separation methods (Sloto and Crouse, 1996). In the fixed-
interval method, the streamflow hydrograph is partitioned
into non-intersecting intervals. The minimum discharge in
each interval is assumed to describe the baseflow over the
entire interval. The sliding-interval method differs from
the fixed-interval method in that the interval is continously
moved over the hydrograph. The HYSEP local minimum,
PART (Rutledge, 1998), and UKIH (Piggott et al., 2005)
methods consist of connecting local minima of the hydro-
graph with straight lines. They differ in how the local min-
ima are identified.
The program BFLOW (Arnold and Allen, 1999) uses a
method, which was apparently first suggested by Lyne and
Hollick (1979). Baseflow is usually associated with water dis-
charged from groundwater storage. This process provides
considerable smoothing. Hence, in the frequency spectrum
of a hydrograph, long waves will be more likely to be asso-
ciated with baseflow while the high frequency variability of
the streamflow will primarily be caused by direct runoff. It
169
should therefore be possible to identify the baseflow by low-
pass filtering the hydrograph.
Lyne and Hollick (1979) proposed the filter equation
1a
bk a  bk1 y k y k1 5
2
subject to bk 6 yk. This equation, with a value of 0.925 for
the filter parameter a, is implemented in BFLOW.
Chapman (1991) criticized the LyneHollick algorithm as
providing theoretically incorrect results and presented an-
other filter equation. Eckhardt (2005) showed that the
Chapman filter is a special case of the following algorithm:
1  BFImax abk1 1  aBFImax y k
bk 6
1  aBFImax
subject to bk 6 yk, where a is the recession constant of Eq.
(3) and BFImax is the maximum value of the baseflow index
that can be modelled by the algorithm, that is the BFI,
which would be obtained, if a constant time series were fil-
tered. Eckhardt (2005) suggested to set BFImax = 0.80 for
perennial streams with porous aquifers, BFImax = 0.50 for
ephemeral streams with porous aquifers, and BFImax = 0.25
study, a stream is considered perennial, if it is waterless
during less than 10% of the time.
The filter parameter a can be determined by a recession
analysis. For the present study, the recession analysis was
carried out according to the correlation method (Langbein,
1938) as explained in the following section.
The time series, which are analysed in this study, consist of
daily mean values of streamflow. For the calculation of the
recession constant, every streamflow value yk is taken into
consideration, which is part of a recession period of at least
five days, i.e. it must be
4000
3000
streamflow at time step k+1
2000
1000
0
0 1000 2000 3000 4000
streamflow at time step k
Figure 1 Scatter plot of streamflow yk+1 against yk during
recession periods, Flint River, Michigan, USA (USGS 04148000,
unit: 1 ft3/s). Dashed line: line through origin with slope
a = 0.975, the value, which is considered the recession constant.
170 K. Eckhardt

Table 1 Baseflow indices for 65 North American catchments, calculated with seven different methods of baseflow separation
USGS gage HYSEP1* HYSEP2* HYSEP3* PART* BFLOW UKIH* Eckhardt
number
04015475 0.77 0.76 0.53 0.77 0.67 0.48 0.68
04017000 0.77 0.77 0.57 0.81 0.69 0.53 0.67
04018000 0.87 0.87 0.67 0.86 0.78 0.69 0.74
04018900 0.75 0.75 0.56 0.81 0.69 0.50 0.69
04019300 0.70 0.71 0.45 0.54 0.53 0.27 0.58
04020000 0.61 0.61 0.49 0.69 0.65 0.47 0.67
04024098 0.61 0.62 0.53 0.55 0.54 0.35 0.57
04032500 0.52 0.52 0.41 0.48 0.52 0.31 0.57
04057820 0.81 0.82 0.70 0.85 0.76 0.62 0.73
04057900 0.78 0.77 0.68 0.83 0.71 0.61 0.71
04058000 0.79 0.79 0.68 0.80 0.73 0.59 0.70
04058130 0.93 0.93 0.88 0.93 0.83 0.76 0.78
04058400 0.85 0.85 0.75 0.91 0.78 0.66 0.75
04059400 0.74 0.74 0.66 0.79 0.66 0.43 0.65
04062230 0.88 0.88 0.68 0.89 0.79 0.66 0.75
04062270 0.88 0.88 0.76 0.87 0.79 0.66 0.75
04062300 0.84 0.84 0.65 0.83 0.75 0.59 0.71
04064000 0.83 0.84 0.74 0.90 0.82 0.74 0.77
04065300 0.74 0.74 0.67 0.77 0.70 0.56 0.70
04065393 0.78 0.78 0.67 0.82 0.72 0.58 0.69
04065397 0.83 0.83 0.69 0.85 0.75 0.63 0.72
04065600 0.87 0.86 0.75 0.86 0.73 0.59 0.72
04067958 0.83 0.83 0.78 0.87 0.83 0.75 0.76
04069416 0.79 0.79 0.76 0.76 0.80 0.70 0.74
04072150 0.47 0.47 0.35 0.44 0.49 0.26 0.54
04073050 0.73 0.74 0.65 0.76 0.72 0.57 0.70
04077630 0.88 0.87 0.85 0.90 0.86 0.82 0.78
04079602 0.87 0.87 0.80 0.85 0.76 0.65 0.74
04085281 0.65 0.65 0.57 0.65 0.64 0.46 0.66
04085395 0.78 0.77 0.66 0.82 0.72 0.57 0.72
04086150 0.75 0.75 0.60 0.79 0.70 0.52 0.69
04087040 0.58 0.58 0.51 0.61 0.59 0.41 0.63
04087159 0.39 0.39 0.36 0.36 0.41 0.28 0.45
04095050 0.66 0.65 0.48 0.57 0.54 0.32 0.59
04096272 0.75 0.74 0.64 0.80 0.70 0.57 0.71
04097200 0.92 0.92 0.89 0.94 0.79 0.76 0.76
04097970 0.94 0.94 0.89 0.94 0.82 0.77 0.75
04145500 0.44 0.45 0.33 0.40 0.46 0.17 0.52
04147990 0.66 0.67 0.59 0.74 0.66 0.47 0.68
04148000 0.71 0.70 0.57 0.74 0.70 0.53 0.69
04148160 0.49 0.49 0.37 0.40 0.42 0.20 0.48
04148200 0.86 0.87 0.72 0.88 0.71 0.54 0.72
04148720 0.61 0.62 0.53 0.55 0.56 0.41 0.60
04151000 0.55 0.53 0.46 0.53 0.59 0.38 0.61
04160000 0.62 0.61 0.50 0.58 0.58 0.35 0.61
04163500 0.60 0.60 0.49 0.52 0.53 0.32 0.59
04164010 0.67 0.65 0.55 0.59 0.57 0.37 0.64
04164150 0.60 0.60 0.53 0.60 0.61 0.42 0.65
04164350 0.45 0.44 0.25 0.29 0.35 0.13 0.48
04164360 0.35 0.37 0.30 0.35 0.43 0.22 0.51
04164450 0.50 0.49 0.34 0.38 0.42 0.18 0.32
04164600 0.64 0.65 0.53 0.58 0.58 0.41 0.63
04165200 0.63 0.63 0.53 0.55 0.54 0.33 0.62
04171500 0.87 0.87 0.82 0.92 0.80 0.70 0.76
04205700 0.71 0.71 0.68 0.74 0.73 0.60 0.72
04206208 0.70 0.70 0.64 0.68 0.54 0.51 0.68
04206212 0.51 0.51 0.42 0.45 0.48 0.28 0.54
04216418 0.55 0.55 0.51 0.56 0.59 0.41 0.62
04220500 0.53 0.54 0.48 0.55 0.57 0.37 0.64
04232034 0.71 0.70 0.67 0.74 0.72 0.61 0.72
04232046 0.77 0.77 0.65 0.70 0.64 0.46 0.67
04232630 0.56 0.56 0.47 0.49 0.46 0.22 0.35
04235276 0.63 0.64 0.45 0.50 0.51 0.29 0.57
04235300 0.61 0.60 0.55 0.61 0.62 0.45 0.64
04244000 0.74 0.74 0.71 0.77 0.74 0.60 0.73
*
Values taken from Neff et al. (2005).
A comparison of baseflow indices, which were calculated with seven different baseflow separation methods 171

y k3 > y k2 > y k1 > y k > y k1 > y k2 7 instead. In other words: deviations of 2% from the theoret-
ical value of a must be accepted, because they rely on ran-
If (a) such a recession were long enough that the streamflow
dom errors. If the recession constant is determined as the
yk and yk+1 consisted entirely of baseflow, (b) there were no
slope of a straight line, which is fitted to the upper bound
groundwater recharge during the time steps k and k + 1, and
of the scatter plot, then points (yk, yk+1) are allowed to
(c) the assumption, that the aquifer is a linear reservoir,
lie above the line, as long as this deviation is less than or
were correct, the following relation would hold:
equal to 2% of the value of the point, which lies perpendic-
y k1 ay k 8 ularly below on the line. This criterion was applied in the
present study.
with the recession constant a, see Eq. (3). Hence, in a plot
of yk+1 against yk, all points should lie on a line through ori-
gin with slope a. In reality, however, this is not the case as
Results
the aforementioned conditions are oftentimes not perfectly
Several workers have observed variations in the recession
fulfilled. One example is shown in Fig. 1. Here, as in Fig. 3,
behaviour with seasonal changes in evapotranspiration,
data from the Flint River (Michigan, USA, USGS 04148000,
see e.g. Szilagyi et al. (2007). According to their analysis,
gauge located at 4306 0 2500 N, 8337 0 0300 W, drainage area:
streamflow recesses faster during the warm season. This
1536 km2) is used. The Flint River was arbitrarily selected
finding is not supported by the present study. First, the
from the 65 catchments in order to exemplify how the re-
recession analysis, which is described in the preceding sec-
sults of the analysis look like.
tion, was exclusively applied to the streamflow records of
Fig. 1 shows that the recession between yk and yk+1 can
the summer months June, July, and August. Then the
assume different speeds. There are several potential rea-
streamflow of December, January, and February was ana-
sons for a faster streamflow recession (i.e. smaller values
lysed. The mean recession constant of all 65 rivers for sum-
yk+1) than expected: the streamflow could still contain di-
mer, which results from these calculations, is 0.960. This
rect runoff, so that it recesses faster between the time
value is only marginally smaller than the mean recession
steps k and k + 1 than if it consisted entirely of baseflow.
constant for winter, which amounts to 0.961. Probably, sig-
A decreasing groundwater recharge would contribute to a
nificant seasonal changes in the streamflow recession only
faster recession as well. Inversely, it is less probable that
become visible in catchments with special characteristics,
the streamflow recesses more slowly than supposed. A
namely shallow groundwater tables and an abundant deeply
growing fraction of direct runoff in the streamflow or an
rooting vegetation.
increasing groundwater recharge would be preceded by
Table 1 contains the BFI values, which were calculated
rainfall, leading to an overall increasing discharge and thus
with the different baseflow separation methods. Mean value
preventing the respective streamflow values from being
and standard deviation are specified in Table 2.
processed. Actually, those points in Fig. 1, which represent
the slowest recessions and form the upper bound of the
scatter plot, lie almost on a line through origin. The reces-
sion constant is found as the slope of such a line, which is Table 2 Mean value and standard deviation of the BFI
fitted to the upper bound of the scatter plot. values
In the following, it is assumed that only measurement er-
rors lead to computed recessions, which are slower than the Mean value Standard deviation
expected ones. Hirsch and Costa (2004) estimate the accu- HYSEP1 0.70 0.14
racy of the USGS streamflow data to 510% of the actual HYSEP2 0.70 0.14
flow. Part of this percentage, however, are systematic er- HYSEP3 0.59 0.15
rors, which will not influence the ratio of consecutive PART 0.69 0.18
streamflow values. If the random error is supposed to be BFLOW 0.65 0.13
1%, then this error could cause the ratio of yk+1 to yk not UKIH 0.49 0.17
to assume the value of a, but the value of Eckhardt 0.65 0.10
1:01  y k1
9
0:99  y k

Table 3 Correlation between the different sets of BFI values

HYSEP1 HYSEP2 HYSEP3 PART BFLOW UKIH Eckhardt
HYSEP1 1.00
HYSEP2 1.00 1.00
HYSEP3 0.93 0.93 1.00
PART 0.96 0.96 0.94 1.00
BFLOW 0.92 0.92 0.93 0.97 1.00
UKIH 0.90 0.90 0.95 0.95 0.98 1.00
Eckhardt 0.85 0.85 0.87 0.91 0.93 0.92 1.00
172 K. Eckhardt

1.00 1.00

0.80 0.80

Eckhardt

Eckhardt
0.60 0.60

0.40 0.40

0.20 0.20

0.00 0.00
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
HYSEP1 HYSEP2

1.00 1.00

0.80 0.80
Eckhardt

Eckhardt
0.60 0.60

0.40 0.40

0.20 0.20

0.00 0.00
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
HYSEP3 PART

1.00 1.00

0.80 0.80
Eckhardt

Eckhardt

0.60 0.60

0.40 0.40

0.20 0.20

0.00 0.00
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
BFLOW UKIH

Figure 2 Comparison of the BFI values, which were calculated by means of the Eckhardt filter, with those of the other methods.

Table 3 shows the correlation between the different sets The HYSEP1 and HYSEP2 methods give almost identical
of BFI values. In Fig. 2 the results, which were obtained with results. By and large, they correspond to those of PART.
filter Eq. (6), are compared to those of the other six However, in some cases there can be important differences
methods. between HYSEP1 and HYSEP2 on the one hand, and PART on

2500

2000

1500
ft 3/s

1000

500

0
1931- 12- 01

1932- 02- 01

1932- 04- 01

1932- 06- 01

1932- 08- 01

1932- 10- 01

1932- 12- 01

1933- 02- 01

1933- 04- 01

1933- 06- 01

1933- 08- 01

1933- 10- 01

1933- 12- 01

1934- 02- 01

1934- 04- 01

1934- 06- 01

1934- 08- 01

date( yyyy- mm- dd)

measured streamflow calculated baseflow

Figure 3 Measured streamflow and calculated baseflow (Eckhardt filter) for a section of the Flint River hydrograph.
A comparison of baseflow indices, which were calculated with seven different baseflow separation methods
the other hand, see, for example, gage number 04019300.
On average, the lowest BFI values result from the UKIH
method.
The results from Eq. (6), denoted in the tables and fig-
ures by Eckhardt, most likely resemble those of BFLOW.
Overall, the Eckhardt filter tends to reduce high BFI values
and to raise low ones compared to the other methods.
Therefore, it gives BFI values, which vary slightly less. In
particular, no BFI value greater than 0.8 can be calculated
(because BFImax = 0.8), while the values provided by the
other algorithms sometimes exceed 0.8.
Discussion and outlook
If one looks at the time series of baseflow generated by the
different methods, then HYSEP1, HYSEP2, HYSEP3, PART,
and UKIH show one flaw in that characteristic points of
the hydrograph are connected with straight lines. This
immediately gives the impression of an unrealistic baseflow
progression. The recursive digital filters (BFLOW, Eckhardt)
in contrast provide extensively smooth time series of base-
flow. Fig. 3 shows an example. The results of these two
methods are not very different. However, Eq. (6) has the
advantage of being hydrologically more plausible than Eq.
(5), which is used in BFLOW (Chapman, 1991).
Eq. (6) can also be applied to time series with a time step
length Dt 0 other than one day. Caution must only be paid to
the filter parameter a, the recession constant. If it is as-
sessed from a hydrograph with the same time step length
Dt 0 , then the criterion (7) must possibly be changed, and
care must be taken that the time span between the stream-
flow measurements, which are used to calculate the ratio
between yk+1 and yk, is not too short. It should rather be
greater than one day than less than one day. On the other
hand, if the recession constant was not determined by the
analysis of a hydrograph with the time step length Dt 0 , but
from a hydrograph with another time step length Dt, such
that Dt 0 = cDt (for example: Dt = 1d, c = 1/24 ) Dt 0 = 1h),
then a must be replaced by
0 gations Report 2005-5217, 23p.
a0 eDt =s ecDt=s eDt=s c ac 10 Piggott, A.R., Moin, S., Southam, C., 2005. A revised approach to
compare Eq. (4).
A fundamental problem is that the true values of the BFI
are unknown. Therefore, one cannot say, which one of the
methods approximates reality best. Tracer measurements
provide an independent estimate of the baseflow contribu-
tion to streamflow. An example is the study of Soulsby
et al. (2004). They used geochemical tracer measurements
to assess the BFI in the Feshie catchment/Scotland (Gauge
No. 8013 of the Scottish Environment Protection Agency).
According to their analysis the BFI lies between 50.3% and
56.1%. If one wants to use Eq. (6) for baseflow separation,
one has to take into account that about 28% of the Feshie
catchment are covered by shallow alpine soils or bedrock.
Eckhardt (2005) suggested to set BFImax = 0.25 under these
conditions. Application of the recession analysis presented
in Section Recession analysis and of Eq. (6) with an area
173
weighted mean BFImax of 0.65 (=28% 0.25 + 72% 0.80) to
the daily streamflow data provided by the National River
Flow Archive at CEH Wallingford results in BFI = 0.58.
This example shows that tracer measurements and base-
flow separation by means of filter Eq. (6) may yield similar
results. However, it also makes clear that it would be desir-
able to test more thoroughly how to choose the filter
parameter BFImax depending on the hydrological and hydro-
geological conditions in a catchment and how to combine
the suggested values of BFImax in heterogeneous catch-
ments. Calculation of an area weighted mean BFImax as in
case of the Feshie catchment is a simple, but probably not
the best method. The proposed investigations require inde-
pendently acquired information about the baseflow contri-
bution to streamflow. Tracer measurements will be
helpful to this end.
A free computer program for recession analysis and base-
flow separation by means of the Eckhardt filter is provided
by the author on request.
References
Arnold, J.G., Allen, P.M., 1999. Validation of automated methods
for estimating baseflow and groundwater recharge from stream
flow records. Journal of the American Water Resources Associ-
ation 35, 411424.
Chapman, T.G., 1991. Comment on Evaluation of automated
techniques for base flow and recession analyses by R.J. Nathan
and T.A. McMahon. Water Resources Research 27, 17831784.
Eckhardt, K., 2005. How to construct recursive digital filters for
baseflow separation. Hydrological Processes 19, 507515.
Hirsch, R.M., Costa, J.E., 2004. US stream flow measurement and
data dissimination improve. Eos 85, 197203.
Langbein, W.B., 1938. Some channel storage studies and their
application to the determination of infiltration. Transactions of
the American Geophysical Union 19, 435445.
Lyne, V.D., Hollick, M., 1979. Stochastic time-variable rainfall
runoff modelling. Hydrology and Water Resources Symposium.
Institution of Engineers, Australia, Perth, pp. 8992.
Neff, B.P., Day, S.M., Piggott, A.R., Fuller, L.M., 2005. Base Flow in
the Great Lakes Basin. US Geological Survey Scientific Investi-
the UKIH method for the calculation of baseflow. Hydrological
Sciences Journal 50, 911920.
Rutledge, A.T., 1998. Computer programs for describing the
recession of ground-water discharge and for estimating mean
ground-water recharge and discharge from streamflow data. US
Geological Survey Water-Resources Investigations Report 98-
4148, 43p.
Sloto, R.A., Crouse, M.Y., 1996. HYSEP: A computer program for
streamflow hydrograph separation and analysis. US Geological
Survey Water-Resources Investigations Report 96-4040, 46p.
Soulsby, C., Rodgers, P.J., Petry, J., Hannah, D.M., Malcolm, I.A.,
Dunn, S.M., 2004. Using tracers to upscale flow path under-
standing in mesoscale mountainous catchments: two examples
from Scotland. Journal of Hydrology 291, 174196.
Szilagyi, J., Gribovszki, Z., Kalicz, P., 2007. Estimation of catch-
ment-scale evapotranspiration from baseflow recession data:
numerical model and practical application results. Journal of
Hydrology 336, 206217.