Professional Documents
Culture Documents
www.elsevier.com/locate/advwatres
Abstract
In this study, we derive analytical solutions of the rst two moments (mean and variance) of pressure head for one-dimensional
steady state unsaturated ow in a randomly heterogeneous layered soil column under random boundary conditions. We rst lin-
earize the steady state unsaturated ow equations by Kirchho transformation and solve the moments of the transformed variable
up to second order in terms of rY and rb , the standard deviations of log hydraulic conductivity Y lnKs and of the log pore size
distribution parameter b lna. In addition, we also give solutions for the mean and variance of the unsaturated hydraulic
conductivity. The analytical solutions of moment equations are validated via Monte Carlo simulations.
2004 Elsevier Ltd. All rights reserved.
and
CU z; njza Ua ag hHa i 1hb0 U1 ni ag
dU2 z eag z hqiag z hHa0 U1 ni; 20
1 ag zb02 hqiag zb0 Y 0 z
dz Kg z 2
which involves the cross-covariance functions
hqi 0 2 hb0 U1 ni, hq0 U1 ni, hHa0 U1 ni, and hY 0 zU1 ni.
Y z ag zb0 q0 q0 Y 0 z ; 12
2 By writing (11) and (14) in terms of n, and multi-
plying the derived equations by b0 , taking the ensemble
subject to the boundary conditions
mean, and assuming that b0 , Y 0 , and q0 are independent,
eag hHa i we obtain the equation for the covariance hb0 U1 ni
U0 a , Ua ; 13
ag dhb0 U1 ni hqiag r2b ag n
ne ; 21
dn Kg
1 0
U a Ua ag hHa i 1b ag Ha0 ; 14
subject to the following boundary condition:
and
hb0 U1 nijna Ua ag hHa i 1r2b : 22
1 2
U2 a Ua ag b Ha0 a2g hHa i ag hHa i 1b02
0
Here we have utilized the fact that b and Ha are un-
2 correlated, i.e., hb0 Ha0 i 0, at the particular boundary
1 2 conditions in our problem.
a2g Ha0 ; 15
2 Similarly, the equations and their corresponding
boundary conditions for covariance hq0 U1 ni,
where Kg and ag are the geometric means of the satu-
hHa0 U1 ni, and hY 0 zU1 ni are given as:
rated hydraulic conductivity Ks and the pore-size dis-
tribution parameter a, respectively. In the following dhq0 U1 ni r2q
eag n ; 23
derivation, both Kg and ag are considered as constants dn Kg
within each layer. By taking the ensemble mean of these
equations and their corresponding boundary conditions, hq0 U1 nijna Ua ag hq0 Ha0 i; 24
and solving these mean equations, one has
hqi ag z dhHa0 U1 ni hq0 Ha0 i ag n
0
hU zi Ua e eag a ; 16 e ; 25
ag Kg dn Kg
symmetric with respect to z and n, to nd the variance of where U0 hU0 i. Expanding the logarithm in the last
U, we only need to solve hY 0 zU1 ni from (27) and equation and collecting terms at separate order (up to
(28) for the case of z 6 n: second order) yields the equations for the total head up
" to second order
0 1 hqikr2Y 2eag z eag aza=k
hY zU ni
Kg 1 a2g k 2
ag k 1 ag h0 z lnag U0 z; 35
#
eag nnz=k
: 32 ag h1 z ag b0 h0 z b0 U1 z=U0 z; 36
ag k 1
or
By substituting (29)(32) into (19) and (20), solving
for CU z; n, and setting n z, we obtain the variance 1 U1 z
r2U z
1
h z h z b0
0
; 37
a ag U0 z
2Ua 0 0
r2U z U2a a2g r2Ha U2a ag hHa i 12 r2b hq Ha i and
Kg
2hqiUa r2b 1
eag z eag a ag hHa i 1 ag h2 z b0 h1 z b02 h0 z
ag Kg 2
ag z 1eag z ag a 1eag a
" #2
U2 z 1 U1 z
hqi2 r2b 0 : 38
ag z 1eag z ag a 1eag a 2 U z 2 U0 z
a2g Kg2
r2q hqi2 kr2Y e2ag a
eag z eag a 2 2.3.1. First moment of head
ag Kg2
2
ag Kg2 1 a2g k2
By taking the mean of (35), (37), and (38) and solve
2ag keag 1=kza 1 ag ke2ag za 1 ag k for hh0 i, hh1 i, and hh2 i, we have
33
1 1 hqi ag z
The rst four terms on the right hand side of (33) are the h0 z lnag U0 z ln eag hHa i e eag a ;
ag ag Kg
contributions of the input variabilities through the lower 39
boundary, while the remaining three terms on the right
hand side are the contributions of the respective b, q, hh1 zi 0, and
and Y variabilities. Note that both terms with r2q and r2Y
are in the order of exp2ag z for large values of z, while r2b 1 2 0 hb0 U1 zi
2 hh2 zi rb h z
the term with r2b is in the order of ag z 1 exp2ag z. ag 2 ag hU0 zi
Because a large r2U corresponds to a large head variance hU2 zi r2U z
(see next subsection), this explains why as the increase of : 40
elevation z, the contribution of r2b to head variance is ag hU zi 2ag hU0 zi2
0
much more important than that of r2Y [9,23]. For unsaturated ow, up to rst-order, (39) implies
0 < aU0 z < expaz. This requires that hqi 6 Kg 1
2.3. Conversion from U to h eag hHa iz =1 eag az for the case of inltration, or
hqi 6 Kg eag hHa i =eag z eag a for evapotranspiration (i.e.,
As the variable U is only an intermediate variable, hqi > 0). The latter can be interplated as the maximum
once the rst and second moments of transformed var- water ux at location z and could be used to calculate
iable U are solved, we must transform them back to the the maximum evapotranspiration rate at surface.
originalP variable, the total head h. By writing
hz 1 n
n1 h z, recalling the expansions for a and U,
and substituting these into relationship ahz lnaUz, 2.3.2. Second moments of head
we have The cross-covariance between total head h and other
independent variables can be derived from (37), for
example
ag 1 b0 h0 h1
" #
X1
Un hq0 U1 zi
b0 0
ln ag e U hq0 h1 zi ; 41
0 ag hU0 zi
n0 U
!
X1
Un and the variance of the pressure head, which is the same
0 0
lnag U b ln 1 0
; 34 as the variance of the total head, can be derived from
n0 U (37)
Z. Lu, D. Zhang / Advances in Water Resources 27 (2004) 775784 779
r2b and
0 2 21 ag h0 z
r2w z r2h z 1 a g h z
a2g a2g hU0 zi K 2 z ag Kg eag z U2 z Y 0 zU1 z
conductivity is k 1:0 m for all layers. The statistics of solution is simply removed. The sample statistics of the
the logarithm of the pore size distribution parameter are ow eld, i.e., the mean prediction of head and its
given as hbi 0:5, 1.0, and 0.5 (in the unit of ln[1/m]), associated uncertainty (variance) as well as the mean
respectively, with CVa 10% for all layers. and variance of unsaturated hydraulic conductivity, are
For the purpose of comparison, we conducted Monte then computed from the rest of realizations. These sta-
Carlo simulations. For the three layers, we generate tistics are considered the true solutions that are used
three sets of realizations, each of which includes 50,000 to compare against the derived analytical solutions of
one-dimensional unconditional realizations. Each set of the moment equations.
these realizations has been tested separately by com- Fig. 1a compares the mean pressure head derived
paring their sample statistics (the mean, variance, and from Monte Carlo simulations (the solid line) and ze-
correlation length) against the input statistics. The roth- and second-order analytical solutions (dashed line
comparisons show that the generated random elds and dashed-dotted line). It is seen from the gure that
reproduce the specied mean and covariance structure while the zeroth-order solution slightly deviates from
very well. Realizations of the log hydraulic conductivity Monte Carlo results, the second-order solution is almost
elds for the whole column are then composed by three identical to the latter. A comparison of the standard
realizations chosen from each set of realizations. deviations of pressure head computed from Monte
The steady state unsaturated ow equation, i.e. (1), is Carlo simulations and analytical solutions is illustrated
solved, using Yehs algorithm [20], for each realization in Fig. 1b. It shows that the two solutions are very close.
of the log hydraulic conductivity eld together with Fig. 2 compares the unsaturated hydraulic conductivity
three independently-generated random numbers repre- statistics resulted from Monte Carlo simulations and
senting the logarithm of the pore size distribution analytical solutions to the moment equations. Again,
parameter for the three layers. If a solution of pressure these results are in excellent agreement, though the
head contains any positive values (i.e., the column is analytical results are systematically underestimated. The
partially saturated), the realization corresponding to this reason for such underestimation is still not clear and
20 20
Zeroth order Analytical
Second order Monte Carlo
15 15
Monte Carlo
Elevation (m)
Elevation (m)
Case 1
10 10 2Y = 0.693, 2 = 9.95E-3
q = -0.002 m/d, 2q = 4.0E-8
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.50 1.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 1. (a) Mean and (b) standard deviation of pressure head for the base case (Case 1).
20 20
Elevation (m)
10 10
5 5
0 -5 0 -5
10 10-4 10-3 10-2 10-1 100 101 10 10-4 10-3 10-2 10-1 100 101
(a) Mean Unsaturated Hydraulic Conductivity K (m/d) (b) Standard Deviation of Unsaturated K (m/d)
Fig. 2. (a) Mean and (b) standard deviation of unsaturated hydraulic conductivity for the base case (Case 1).
Z. Lu, D. Zhang / Advances in Water Resources 27 (2004) 775784 781
further investigation may be needed. Here we would like the pressure head variance. Second, due to the large
to mention that although the variability of Y and b in variability on Y , the saturated hydraulic conductivity Ks
each layer are not very large, the total variability of ei- in some points of realizations is so low that the medium
ther Y or b for the whole column is still relatively large becomes partially saturated and thus these realizations
because of the contrast between layers [10]. are removed in the Monte Carlo simulations. Overall,
2590 ( 12.7%) out of 20,000 realizations, have been
3.2. Large variability of Y and b removed for this case.
Fig. 4 compares the Monte Carlo results and ana-
Now we would like to investigate the validity of our lytical solutions for a large b variability r2b 0:087
solutions at very large variabilities of Y and b. Fig. 3 (CVa 30%). A few observations can be made from this
depicts the comparisons of Monte Carlo results and gure. First, the analytical solutions are very close to the
analytical solutions at r2Y 4:0 (the coecient of vari- Monte Carlo results even at such a large variability of b.
ation CVKs 732%). Note that at such a large vari- In addition, if we compare this gure with Fig. 3, we nd
ability, the zeroth-order analytical solution of the mean that the head variance due to CVa 30% is much larger
pressure head greatly deviates from the Monte Carlo that that due to CVKs 732%. This nding is consistent
results. However, after including the second-order cor- with our early conclusion [23] that the contribution of b
rections the solution is almost identical to the Monte variability to the head variance is much more important
Carlo results. The head variance from our analytical than is the contribution of Y variability.
solution is reasonably close to Monte Carlo results.
There are two possible reasons that contribute to the 3.3. Random constant a approximation
discrepancy between the head variances computed from
Monte Carlo simulations and analytical solutions. First, Another interesting point we would like to explore is
2
the head variance r2w z hw1 z i from the analytical the inuence of our assumption (or approximation) of
solutions represents the lowest-order approximation of random constant a. We do so with a new set of Monte
20 20
Zeroth order
Case 2
Y = 4.0, = 0.0,
2 2
Second order
q = -0.002 m/d, q = 4.0E-8
2
15 15
Monte Carlo
Elevation (m)
Elevation (m)
10 10
Analytical
Monte Carlo
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.20 0.40 0.60 0.80 1.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 3. (a) Mean and (b) standard deviation of pressure head for Case 2: r2Y 4:0 (CVKs 732%) for each layer.
20 20
Zeroth order Analytical
Second order Monte Carlo
15 15
Monte Carlo
Elevation (m)
Elevation (m)
Case 3
10 10 = 0, = 0.086,
2 2
Y
q = -0.002 m/d, q = 4.0E-8
2
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.50 1.00 1.50 2.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 4. (a) Mean and (b) standard deviation of pressure head for Case 3: r2b 0:086 (CVa 30%) for each layer.
782 Z. Lu, D. Zhang / Advances in Water Resources 27 (2004) 775784
Carlo simulations. In Case 4, instead of generating a of Case 4. Again, in Monte Carlo simulations, a in each
random number as the a value for each layer, we now layer is a correlated random function (correlation length
compose the correlated random elds of a for the whole k 1:0), while in our analytical solutions, a is a random
column in the same way as we did for the realizations of constant. The comparison is illustrated in Figs. 6 and 7.
Y , already described above. It is now seen that the analytical solutions are in excel-
Fig. 5 compares results from our analytical solutions lent agreement with Monte Carlo results. In Case 6, the
in which a in each layer is a random constant against layer conguration is the same as in Case 4, but now we
those from Monte Carlo simulations where a in each increase the correlation length of a in the top layer from
layer is a correlated random function. Comparing it to 1 to 2.5 m, i.e., the top layer is 4 correlation length in
Fig. 1 shows that our analytical solutions are in excellent thickness (10 m). The results are shown in Fig. 8. Cer-
agreement with Monte Carlo results in the rst and tainly, compared to Fig. 5, the agreement between
second layer (counting from the bottom) but there is a Monte Carlo results and our analytical solutions has
discrepancy in the top layer, especially for the head been signicantly improved. Fig. 9 shows such com-
variance. A similar pattern is observed by comparing the parison for unsaturated hydraulic conductivity.
unsaturated conductivity statistics (not shown here). We The results from these two cases imply that when the
suspect that this may be due to the large thickness of this layer thickness is relatively small (in physical length) or
layer (10 correlation length). the correlation length of a is relatively large, the corre-
In order to check this, we analyzed two more cases. In lated random function may be approximated very well
Case 5, the top layer (10 m) is further divided into two by a random constant. These results are consistent with
layers with thickness of 5 m each. The properties of the the nding of Yeh et al. [19] and that of Hopmans [6]. In
third layer are the same as those of the top layer in Case fact, a random constant is a special case of correlated
4, and the fourth layer are the same as the second layer eld with a correlation length of innity.
20 20
Zeroth order Analytical
Second order Monte Carlo
15 15
Monte Carlo
Elevation (m)
Elevation (m)
Case 4
10 10 2Y = 0.693, 2 = 9.95E-3
q = -0.002 m/d, q = 4.0E-8
2
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.50 1.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 5. (a) Mean and (b) standard deviation of pressure head for Case 4. All parameters are similar to the base case, except that b in Monte Carlo
simulations is a spatially correlated random function in each layer rather than a random constant.
20 20
Analytical
Monte Carlo
15 15
Elevation (m)
Elevation (m)
Case 5
10 Zeroth order 10 2Y = 0.693, 2 = 9.95E-3
Second order
q = -0.002 m/d, q = 4.0E-8
2
Monte Carlo
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.50 1.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 6. (a) Mean and (b) standard deviation of pressure head for Case 5.
Z. Lu, D. Zhang / Advances in Water Resources 27 (2004) 775784 783
20 20
Case 5
Zeroth order Analytical
15 Second order 15 Monte Carlo
Monte Carlo
Elevation (m)
Elevation (m)
10 10
5 5
0 -5 0 -5
10 10-4 10-3 10-2 10-1 100 101 10 10-4 10-3 10-2 10-1 100 101
(a) Mean Unsaturated Hydraulic Conductivity K (m/d) (b) Standard Deviation of Unsaturated K (m/d)
Fig. 7. (a) Mean and (b) standard deviation of unsaturated hydraulic conductivity for Case 5.
20 20
Zeroth order Analytical
Second order Monte Carlo
15 15
Monte Carlo
Elevation (m)
Elevation (m)
Case 6
Y = 0.693, = 9.95E-3
2 2
10 10
q = -0.002 m/d, q = 4.0E-8
2
5 5
0 0
-6.0 -5.0 -4.0 -3.0 -2.0 -1.0 0.0 0.00 0.50 1.00
(a) Mean Pressure Head (m) (b) Standard Deviation of Head (m)
Fig. 8. (a) Mean and (b) standard deviation of pressure head for Case 6.
20 20
Case 6
Zeroth order Analytical
15 Second order 15 Monte Carlo
Monte Carlo
Elevation (m)
Elevation (m)
10 10
5 5
0 -5 0 -5
10 10-4 10-3 10-2 10-1 100 101 10 10-4 10-3 10-2 10-1 100 101
(a) Mean Unsaturated Hydraulic Conductivity K (m/d) (b) Standard Deviation of Unsaturated K (m/d)
Fig. 9. (a) Mean and (b) standard deviation of unsaturated hydraulic conductivity for Case 6.
solutions in literature for unsaturated ow in heteroge- [7] Indelman P, Or D, Rubin Y. Stochastic analysis of unsaturated
neous soil column, our solutions are not limited to the steady state ow through bounded heterogeneous formations.
Water Resour Res 1993;29:11418.
gravity-dominated regime but vaild for the entire [8] Lu Z, Neuman SP, Guadagnini A, Tartakovsky DM. Conditional
unsaturated zone. Our solutions are second order in moment analysis of steady state unsaturated ow in bounded
terms of the standard deviations of the log hydraulic randomly heterogeneous porous soils. Water Resour Res
conductivity and the pore size distribution parameter. 2002;38(4), doi:10.1029/2001WR000278.
The accuracy of these second order solutions is veried [9] Lu Z, Zhang D. Stochastic analysis of transient ow in hetero-
geneous, variably saturated porous media: the van Genuchten
using Monte Carlo simulations. Numerical examples Mualem constitute model. Vadose Zone J 2002;1:13749.
show that these solutions are valid for relatively large [10] Lu Z, Zhang D. On stochastic modeling of ow in multimodal
variabilities in soil properties. heterogeneous formations. Water Resour Res 2002;38(10):1190,
Our solutions of the rst two moments of the pressure doi:10.1029/2001WR001026.
head are derived based on the assumption (or approxi- [11] Romano N, Brunone B, Santini A. Numerical analysis of one-
dimensional unsaturated ow in layered soils. Adv Water Resour
mation) that the pore size distribution parameter a is a 1998;21:31524.
random constant in each layer. Numerical examples [13] Russo D. Determining soil hydraulic properties by parameter
indicated that such an approximation may be appro- estimation: on the selection of a model for the hydraulic
priate if the ratio of the correlation length of a in any properties. Water Resour Res 1988;24:4539.
layer to the layer thickness is relatively large (e.g., 0.25 [14] Russo D, Bouton M. Statistical analysis of spatial variability in
unsaturated ow parameters. Water Resour Res 1992;28(7):1925
in Case 6). In the limit that this ratio goes to innity, the 91.
random constant treatment becomes exact. [15] Tartakovsky DM, Neuman SP, Lu Z. Conditional stochastic
averaging of steady state unsaturated ow by means of Kirchho
transformation. Water Resour Res 1999;35(3):73145.
[16] Unlu K, Nielsen DR, Biggar JW. Stochastic analysis of unsatu-
References rated ow: one-dimensional Monte Carlo simulations and com-
parisons with spectral perturbation analysis and eld
[1] Andersson J, Shapiro AM. Stochastic analysis of one-dimensional observations. Water Resour Res 1990;26(9):220718.
steady state unsaturated ow: a comparison of Monte Carlo [17] van Genuchten MTh. A closed-form equation for predicting the
and perturbation methods. Water Resour Res 1983;19(1): hydraulic conductivity of unsaturated soils. Soil Sci Soc Am J
12133. 1980;44:8928.
[2] Brooks RH, Corey AT. Hydraulic properties of porous media. [19] Yeh T-C, Gelhar LW, Gutjahr AL. Stochastic analysis of
Hydrol. Pap. 3, Colo. State Univ., Fort Collins, 1964. unsaturated ow in heterogeneous soils: 1. Statistically isotropic
[3] Ferrante M, Yeh JT-C. Head and ux variability in heterogeneous media. Water Resour Res 1985;21:44756.
unsaturated soils under transient ow conditions. Water Resour [20] Yeh JT-C. One-dimensional steady-state inltration in heteroge-
Res 1999;35(4):14719. neous soils. Water Resour Res 1989;25(10):214958.
[4] Foussereau X, Graham WD, Rao PSC. Stochastic analysis of [21] Zhang D, Wallstrom TC, Winter CL. Stochastic analysis of
transient ow in unsaturated heterogeneous soils. Water Resour steady-state unsaturated ow in heterogeneous media: Compar-
Res 2000;36(4):891910. ison of the BrooksCorey and GardnerRusso models. Water
[5] Gardner WR. Some steady state solutions of unsaturated mois- Resour Res 1998;34(6):143749.
ture ow equations with application to evaporation from a water [22] Zhang D, Winter CL. Nonstationary stochastic analysis of steady-
table. Soil Sci 1958;85:22832. state ow through variably saturated, heterogeneous media.
[6] Hopmans JW, Schukking H, Torfs PJJF. Two-dimensional Water Resour Res 1998;34(5):1091100.
steady-state unsaturated water ow in heterogeneous soils with [23] Zhang D, Lu Z. Stochastic analysis of ow in a heterogeneous
autocorrelated soil hydraulic properties. Water Resour Res unsaturated-saturated system. Water Resour Res 2002;38(2),
1988;24(12):200517. 10.1029/2001WR000515.