Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

GEODERk.

MA
ELSEVIER Geoderma 69 (1996) 157-173

Testing the performance of a


one-dimensional solute transport model (LEACHC)
using response surface methodology
Nicholas M. Holden a,*, Andrew J. Rook b, David Scholefield b
Department of Geographical Sciences, University of Plymouth, Drake Circus, Plymouth, PL4 8AA, UK
b Institute of Grassland and Environmental Research, North Wyke, Okehampton, Devon, EX20 2SB, UK

Received 8 May 1995; accepted 14 September 1995

Abstract

The independent comparison of a three-dimensional data set from a large soil block (5 m X 3 m
X 1 m) with output from a one-dimensional model of soil water and solute transport (LEACHC) is
considered. The lack of work comparing site-specific solute breakthrough with model predictions is
outlined. A statistical method of limiting the number of runs required to provide a good representation
of model output based on Response Surface Methodology is presented. This method is both a
sensitivity analysis, and a means of identifying how a model behaves. As an example the model
LEACHC was tested using response surface methodology and was found to be insensitive to input
variable alteration. A lack of compatibility between sampling methods and the assumptions of the
model may account for it being unable to predict solute movement similar to that measured in a large
soil block. It is concluded that the technique presented is well suited to testing one-dimensional
models.

1. Introduction

A number of one-dimensional soil water models have been published over the last decade
for predicting the movement o f solutes through the soil profile (e.g. SLIM, Addiscott and
Whitmore, 1991; LEACHM, Hutson and Wagenet, 1992; CRACK, Jarvis, 1989; MACRO, Jarvis,
1991 ). These models each describe solute transport differently: (i) SLIM is a capacity layer
model using m o b i l e / i m m o b i l e water segmentation, (ii) LEACHM, a numerical solution of
the Richards' and the Convection-Dispersion equations, (iii) CRACK is a model based on

* Correspondingauthor. Present address: Peat TechnologyCentre, UniversityCollege Dublin,EarlsfortTerrace,


Dublin 2, Ireland.

0016-7061/96/$15.00 © 1996 Elsevier Science B.V. All rights reserved


SSDI0016-7061 (95)00066-6
158 N.M Holden et al. / Geoderma 69 (1996) 15~173

the entry of water into soil peds and (iv) MACROwhich divides water flow into macropore
and micropore domains and applies gravity flow to macropore water and Richards' equation
to micropore water. Addiscott and Wagenet (1985) presented a critical summary of
approaches used in soil water modelling in recent times and suggested that few models are
independently tested.
There is a sparse literature on independent model testing (as opposed to testing by the
developer), but the objective assessment of the goodness of prediction has been addressed
by the authors of papers on objective evaluation of models and was summarized by Willmott
et al. (1985) and Loague and Green ( 1991 ). Initially, a subjective, graphical comparison
is always used because if there is an obvious visual difference between observed and
predicted data then there is a weakness in either the formulation or the operation of the
model being tested.
Cameron and Wild (1982) tested three models using total profile -~6C1- extracted tYom
augered samples. They used 10 replicate cores to derive a mean datum for each depth
segment in the soil profile. They did not attempt to assess the prediction of three-dimensional
variability.
Vinten and Redman (1990) reported a test of a capacity-type model (Addiscott et al.,
1986) which works in a similar manner to SLIM ( Addiscott and Whitmore~ 1991 ). The data
from 4 field plots were used and particular attention focused on the estimation of co, the
permeability factor in the model. In assessing the model, individual plot data were utilized
to quantify a and best-fit curves were generated by manipulation of c~. It was found that
best-fit values were very different from values derived. Vinten and Redman concluded that
parameter estimation is crucial to successful prediction, but did not consider the limited
spatial variation from their field plots to any great extent.
Pennell et al. (1990) evaluated 5 pesticide leaching models against a data set derived
IYom 16 samples taken from each of 4 quadrants and at 8 depths on 7 days. The 16 samples
were used to calculate "average solute concentrations" which were in turn converted to a
single "field-average" concentration. The field average concentration was used for all
comparisons between modelled and measured data. In evaluating the models there was no
attempt to assess the potential range of prediction that each model could achieve. Costa et
al. (1994) evaluated 4 models in a similar manner to Pennell et al. (1990) with no attempt
to predict the spatial variability measured. Jabbro et al. (1994) used data from 18 plots
equipped with zero tension pan lysimeters to estimate an average breakthrough for model
evaluation with comparison of the measured mean against a single best prediction curve.
Roth et al. ( 1991 ) suggested that a shortage of field data and the complexity of the sytem
accounts for the lack of understanding of field scale solute transport. Holden et al. (1995a)
and Scholefield et al. (1995) presented results of chloride and nitrate breakthrough respec-
tively, in a large undisturbed soil block (5 m × 3 m × l m) (the North Wyke Soil Block)
resulting from tracer pulse experiments made under steady-state, near saturated conditions.
The experiments yielded fine spatial and temporal resolution solute breakthrough curves at,
on average, 4 points at 9 depths within the block. The data were obtained with the objective
of testing existing solute transport models. Arising directly from this objective was how to
best undertake the task of assigning values to variables in a one-dimensional model which
required comparison with a three dimensional dataset.
N.M. Holden et al. /Geoderma 69 (1996) 157-173 159

lOO

90 -

80 ;" i
70

60 ,I' . . . . :, i"
5(? )
j ,-
40 ~ :
r

20 " ' '~

10 ~ ' J ..........

-0.5 0 0.5 1 15 2 2.5 3


time (days)
Fig. 1. The range in measured chloride breakthrough at 0.40 m within the North Wyke Soil Block at a surface
water flux of 1 mm h ~ showing great variation in peak concentration and time-to-peak concentration. Each line
represents the breakthrough measured at a single sampler.

The data set obtained by Holden et al. (1995a) comprised chloride breakthrough curves
from a number of samplers throughout a large soil block. The spatial variability observed
in "time-to-peak" concentration and magnitude of peak concentration (Fig. 1 ) was such
that it was not rational to derive "average" breakthrough curves at each depth; the average
curve being unrepresentative of any mechanism occurring in the soil. Furthermore, for the
purposes of evaluation, it is unreasonable to assume that the model output generated from
mean input values will be comparable in any meaningful way to mean breakthrough curves
unless the distribution of all variables is of the same type (e.g. normal) and there is an
overall linear response in the model. It is however, reasonable to consider the possibility of
a purely mechanistic model being able to predict breakthrough at specific locations in the
soil if it is initialized with accurately determined input variables that correspond to the soil
influencing the location of interest. The problem is that determination of the influential soil
variables is normally very difficult. The soil block breakthrough data being utilized here
are not supported by field data ( water release, hydraulic conductivity) specific to each point
of interest. This paper presents a method of model evaluation based on Response Surface
Methodology that gives the model the "best chance" of predicting the measured break-
through at each sampling location in the soil block using the data available.

2. Methods

2.1. Selection of model

To assess the usefulness of Response Surface Methodology (RSM) it was necessary to


select a model with which to work. Any of the models mentioned in the introduction to this
paper would have been acceptable but LEACHMwas chosen, specifically the version LEACHC.
Costa et al. (1994) concluded that LEACHM best represented their data under steady-state
160 N.M. Holden et al. / Geoderma 69 (1996) 157-173

conditions while Pennell et al. (1990) are less enthusiastic about LEACHP but still recom-
mend it. It is an assumption of the model that the soil layers are homogeneous and thus it
would appear unsuitable for modelling a soil block with preferential flow. Hutson and
Wagenet (1992) state however, that "...it has been shown that use of such a model in an
'ad hoc stochastic' manner is possible, in which repeated executions of the model with
different input values of possible field conditions can provide accurate descriptions of the
variability of water and chemical in the field...". If the locations of preferential flow (as
indicated by breakthrough measured from individual samplers) are assumed to be due to
variations in flow between Darcian flow fields of suitable dimension, then the model should
be able to predict the range of breakthrough measured in the field. Hutson and Wagenet
(1992) do not suggest in their introduction to LEACHM that its use should be restricted to
" h o m o g e n e o u s " soil, rather that its use should be pushed towards the development of
strategies for modelling field conditions. The LEACHC version was used as it is designed to
predict C1 movement (Hutson and Wagenet, 1992).

2.2. Model testing

Initially, the model was run using mean input values for comparison with mean measured
breakthrough in a similar manner to previously published work (Pennell et al., 1990; Costa
et al., 1994; Jabbro et al., 1994).
The method of model evaluation used is based on the assumption that if a model is capable
of predicting the correct result, it should be given every reasonable chance to do so. If all
input variables were measured and the model run with all possible combinations of variables,
it would be possible to assess whether the model could predict a range of outputs similar to
those measured in the field. Unfortunately, a model with for example, just 5 input variables
in each of 9 layers, with 10 levels per variable would require 450 runs to cover all combi-
nations. A highly developed mechanistic model such as LEACHM has many more variables
per layer each with many possible levels and thus it is obviously not possible to execute the
model for all combinations of variables. To overcome this a statistically derived subsample
can be taken to achieve a reasonable number of combinations of variables so as to obtain a
representative expression of the model' s response.

2.3. Response surface methodology

Response surface methodology is used to study systems with a large number of variables
all of which influence a particular feature of the system (Myers, 1971 ). In the case presented
here, model output is influenced by a large number of model inputs. The model response
(~/) can be expressed as a function of the input variables ( ~'i):

r / = f ( ~',, sr2. . . . . ~',,) (1)


where the form o f f is complex (and in this case represents the interactions of equations in
the model). RSM depends on the approximation of f by a simple low order polynomial
equation, e.g. if the response was approximately linear, then Eq. ( 1 ) could be expressed as:
n=/3o +/3~x~ +/3~x~ + ... +/3.x~ ~2)
N.M. Holden et al. / Geoderma 69 (1996) 157-173 161

where xi is a coded form of the natural variable ~'i and/3,, the equation coefficient (Myers,
1971 ). In using RSM the following stages are considered: (i) experimental design, i.e., the
selection of the combinations and levels of variables to be considered, (ii) fitting a regression
equation to the observation and (iii) analysis of the regression equation to identify the most
sensitive variables (Kleijnen, 1974). The current paper presents an application of RSM to
model evaluation. Further details of the technique can be found in Davies (1956), Hill and
Hunter (1966), Herzberg and Cox (1969), Myers ( 1971 ) and Kleijnen ( 1974, 1975).

2.4. The simplex designfor generating the response surface

Where many variables are required to initialize a model (which is typical for mechanistic
solute transport models) it is necessary to assume a first-order (linear) response surface in
order to keep the number of combinations manageable (e.g. for a central composite design
assuming a second-order response surface, the number of combinations would be 2 k factorial
points plus 2K+ n2 additional points. In the example used in this paper, k = 45, thus the
number of combinations would be ca. 3.5 X 1013). A complex response surface would
require thousands of combinations for reliable prediction when using a model with a large
number of inputs. The complicated relationships encompassed in a mechanistic model such
as LEACnC are likely to be best approximated by a high-order response surface but the first-
order surface is useful in order to generate a range of responses from the model. It is not
the intention of this work to use the response surface to simplify prediction by eliminating
the need for the model, but to produce a statistical sample of combinations of input para-
meters to encompass the range of possible combinations. Further, more detailed investiga-
tions of regions of the response surface can be undertaken if required. The first-order
response surface can be determined using a manageable number of model runs. The linear
response could be used to predict best-fit parameters at a later stage. A simplex design (see
Myers, 1971 for a derivation of the terminology) was used. The steps used to construct the
design are illustrated using an example with 4 variables:
(i) an orthogonal matrix of size N × N ( where N = K + 1, K = number of variables) was
formed. All values in the first column (Xi.l) were set to 1. In the second column, the first
element (X2.1) was set to K and all other elements to - 1. For the remaining columns (J),
the first J - 2 elements were set to 0, the next to K - ( J - 1) and all the remainder to - 1:
K---4 .'. N = 5 1 4 0 0 0
1 -1 3 0 0
1 -1 -1 2 0 (3)
1 -1 -1 -1 1
1 -1 -1 -1 -I
(ii) each element of the matrix was multiplied by ( 1/m °5) where m was the sum of the
absolute values of all the elements in the column:
1/75 4/78 0 0 0
1/75 -1/78 3/76 0 0
1/75 -1/78 -1/76 2/74 0 (4)
1/75 -1/78 -1176 -1174 1172
1/75 -1178 -1176 -1174 -1/72
162 N.M. Holden et al. / Geoderma 69 (1996) 15~173

(iii) all elements of the matrix were multiplied by N°5;


(iv) the first column of the matrix was removed to leave the design matrix:
3.16 0 0 0
- 0.79 2.74 0 0
-0.79 -0.91 2.24 0 (5)
-0.79 -0.91 -1.12 1.58
-0.79 -0.91 -1.12 -1.58
The input variables of the model were coded such that 0 = mean of the real variable value
and the maximum value in the design = the maximum measured value with other values
determined based on linear scaling between these points. For example, if the first variable
had a min., mean and max. of 2, 4 and 6, then the codes would translate to: code 0 = 4, code
3.16 = 6 and 1 coded unit = 0.6.

2.5. Application to input variables f o r LEACHC

The only modification to the code of the model was to add a few lines permitting the soil
water chloride values to be written to a separate file for simplifying the task of data analysis.
There are about 250 variables that are used by the model with 9 soil layers, depending on
the method of initialization. For the soil block data (Holden et al., 1995a, b), this would
require about 460 input variable combinations (depending on how the model was to be
run) each of three levels to satisfy the simplex design. For the purposes of the model testing
presented here this was simplified by excluding variation in the crop growth, ground cover,
and the initial chemical conditions in the soil. Attention was focused on just five variables
in each of 9 layers (%clay; %silt, %organic carbon, bulk density and saturated hydraulic
conductivity). These variables were chosen because they control the nature of the pore-size
distribution, total porosity and the movement of water in the soil. A total of 46 model runs
were required by the simplex design to determine the response surface. These variables
were selected because data were available for clay, silt and organic carbon from 54 locations
within the North Wyke Soil Block (samples having been obtained while installing the
suction samplers) which permitted LEACHC to be run using particle size-based regression
models to estimate water retention. As data for bulk density and hydraulic conductivity
were not available without damaging the block, these were estimated respectively from
values obtained from adjacent soil and by estimation based on the rates of breakthrough
observed in the experiments. The initial input variables used are presented in Table 1.
(Methods after Avery and Bascomb, 1982). Due to the uncertainty of the suitability of the
bulk density and saturated hydraulic conductivity data, a second set of variables was used
for an additional 46 runs of the model with different mean values and a slightly larger range
respectively.

3. Results

3.1. M e a n input variables vs mean breakthrough

Mean measured breakthrough was derived from the soil block data by performing a linear
interpolation between measured data points and calculating mean values for hourly intervals.
N.M. Holden et al. / Geoderma 69 (1996) 157-173 163

Table 1
Initial and modified input parameters for the RSM test of the model LEACHC based on measured values of percent
clay, silt and organic carbon and estimates of bulk density (g cm 3) and saturated hydraulic conductivity (mm
d t)

Parameter Depth (m) Mean Maximum Modified mean Modified maximum

Clay (c) 0.1 14.48 18.36


Silt (z) 0.t 24.15 26.77
Organic carbon (oc) 0.1 0.929 1.267
Bulk density (db) 0.1 1.3 1.4 l 1.4
Sat. Hydr. Cond, (ks) 0.1 10584 12528 10584 15876
Clay (c) 0.2 15.78 20.09
Silt (z) 0.2 22.65 26.65
Organic carbon ( oc ) 0.2 0.605 1.055
Bulk density (db) 0.2 1.3 1.4 1 1.4
Sat. Hydr. Cond. (ks) 0.2 10584 12528 10584 15876
Clay (c) 0.3 9.78 14.11
Silt (z) 0.3 20.29 24
Organic carbon (oc) 0.3 0.15 0.2
Bulk density (db) 0.3 1.35 1.45 1.05 1.45
Sat. Hydr. Cond. (ks) 0.3 10584 12528 10584 15876
Clay (c) 0.4 9.17 13.47
Silt (z) 0.4 20.99 30.28
Organic carbon (oc) 0.4 0.14 0.19
Bulk density (db) 0.4 1.35 1.45 1.05 1.45
Sat. Hydr. Cond. (ks) 0.4 10584 12528 10584 15876
Clay (c) 0.5 7.51 9.45
Silt (z) 0.5 20.84 23.74
Organic carbon (oc) 0.5 0.061 0.09
Bulk density (db) 0.5 1.35 1.45 1.05 1.45
Sat. Hydr. Cond. (ks) 0.5 10584 12528 10584 15876
Clay (c) 0.6 7.8 11.81
Silt (z) 0.6 18.86 22.9
Organic carbon (oc) 0.6 0.037 0.07
Bulk density (db) 0.6 1.4 1.5 1. l 1.5
Sat. Hydr. Cond. (ks) 0.6 10584 12528 10584 15876
Clay (c) 0.7 4.87 7.11
Silt (z) 0.7 19.84 22.87
Organic carbon (oc) 0.7 0.025 0.04
Bulk density (db) 0.7 1.4 1.5 1.1 1.5
Sat. Hydr. Cond. (ks) 0.7 10584 12528 10584 15876
Clay (c) 0.8 5.88 9.34
Silt (z) 0.8 19.63 22.99
Organic carbon ( oc ) 0.8 0.03 0.05
Bulk density (db) 0.8 1.4 1.5 1.1
Sat. Hydr. Cond. (ks) 0.8 10584 12528 10584 15876
Clay (c) 0.9 5.08 8.06
Silt (z) 0.9 21.5 24.56
Organic carbon (oc) 0.9 0.02 0.05
Bulk density (db) 0.9 1.4 1.5 1.1 1.5
Sat. Hydr. Cond. (ks) 0.9 10584 12528 10584 15876
164 N.M. Holden et al. / Geoderma 69 (1996) 157-173

200

+ 2 S.D., measured
150
A

t00
E mean, measured
predicted
50 ~
"E

8 ~ 00~ ~
-5 5 10 15 20 25
-50 V - 2 S.D., measured time (days)

-100 -

Fig. 2. Mean breakthrough curves ( + 2 × standard deviation) calculated from the data in Fig. 1. and the LEACHC
predictions using mean input variables.

A graph of mean measured breakthrough and modelled breakthrough (Fig. 2) shows that
there was no similarity between the predicted and measured values. It is clear that the
difference between measured and predicted values of both the magnitude of peak concen-
tration and the time-to-peak concentration were very large. The prediction of time-to-peak
concentration was markedly poor, being 4 days later than that measured.

3.2. RSM results using initial estimates of bulk density and K,~,

There was little difference in peak concentration predicted by LEACHCfor the 46 runs of
the model undertaken using the initial estimates of input variables (Fig. 3). Predicted time-
to-peak concentration using the same 46 model runs was more variable with a range of 19
hours. Predictions were more variable in the middle and deeper soil layers than at the
surface.
Response surfaces for the peak concentration and time-to-peak concentration were deter-
mined for each soil layer by performing a multiple linear regression (Ryan et al., 1985) of
the predicted output on the five input variables varied at each depth. In order to identify
which variables were most influential, the calculated coefficients were multiplied by a
scaling factor related to the magnitude of the range of the variable. Table 2 presents two
examples of the calculated coefficients for predicting peak concentration scaled by the
variable range. In many cases (not all data shown) the coefficients were 0 indicating no
relationship could be established. The mean values obtained for all 46 runs of the model
suggested that bulk density in the surface horizons was of some importance when predicting
peak concentration and time-to-peak concentration (Table 3) but otherwise no single
variable was particularly important or unimportant. It should be noted that the 0 values for
K~ coefficients occurred because all coefficients were rounded to 4 decimal places, and the
magnitude of Ks made the coefficient values relatively small, thus the rounding made the
coefficients appear less important than they perhaps were.
N.M. Holden et al. / Geoderma 69 (1996) 157-173 165
High conductivity High conductivity and or clay
52 - content

47 .~ "7-~ ~/ ,"'i ~ 0.1m


I / /' ',
A ",
42 i ' , ' "

oc32 i .,. . . . . . . . ' I 0.3m


o _ _ - - _ _ / ..... r ...... i ........................... 0.4m
~ , , , * - - ~ ' ~ - I¢'/~-~--'~- - , , - S - ~ - - j . . Z . ~ - - / . ~ 0.5 m 0.6 m
27 ! ~ ~ ) ~ \ _ ~ .~,,X~..._,X_~.---~_ " = 0.7 m
•----~ --". . . . -------" - --.--~" . . . . 0.8 m O . g m

22 ~ -- ~ ~ . . . . . . . .
0 10 20 30 40 50 A
run number
High clay content
20 - High bulk density ,;~ ~

"O /, ,~,.., .,~, ..=, .,~ . . . . . ~. . . . . . . . ~ ....... u.um


18 - ,I ,.,"",J , f ~ , l ,j~'J ,/,~__...~,~ -~'-- J ~ _ ~-.~- 0.8 m

i/ ~ /-,,U/-.-~,~/ -,,~/f\ ~'/'\ r"~\ /.._._. A_.~ /"--/~--/06m

13 '-."" ....... "-'--'--' ",...... " " .............................. ..... -............. 0.3m

1 2 , "~-'~ ""~"~' -~" . . . . . . "- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 0.2m


" ~ v ~ ~ 0 . 1 m
E 11
10~- - ~ --- ~ . . . . . . . . . .
0 10 20 30 40 50
run number B
Fig. 3. ( A ) variation in peak concentration, and (B) variation in time-to-peak concentrations as predicted by
L E A C H C using initial measured and estimated input variables obtained from the North W y k e Soil Block and
implemented using RSM.

3.3. RSM results using extended ranges of bulk density and Ksa,

The extended range of bulk density and saturated hydraulic conductivity that was used
tended to smooth out the variation in peak concentration (Fig. 4A) but greater peak con-
centrations tended to be associated with larger values of bulk density. Variation in simulated
"time-to-peak" concentration (Fig. 4B) was much less near the soil surface than at depth.
It is interesting to note that with the extended values, at 0.90 m, the peak concentration
could be simulated as occurring sooner than at 0.80 m. This could have been due to the
numerical rounding up procedures used in the model code. When LEACI-ICwas run with the
values of the water release characteristic estimated from the particle size distribution, it
seemed to be most sensitive to changes in bulk density, but greatly reducing the value of
the bulk density appeared to have reduced the responsiveness of the model. The relative
importance of the scaled regression coefficients (data not shown) indicated that with high
166 N.M. Holden et al. / Geoderma 69 (19961 157~173

Table 2
Regression coefficients for predicting peak concentration from: clay, silt, organic carbon, bulk density and saturated
hydraulic conductivity ( weighed by [ max-mean ] )

Parameter Depth Range Peak concentration Depth Range Peak Depth Range Peak
concentration concentration

0. I m 0.8 m 0.1 m (I.8 m 0.1 m 0.8 m

Constant - 33.3437 7.4768


Clay 0.1 3.88 0,8218 0.0695 0.4 4.30 0.2670 0.0344 (t.7 2.24 04621 0.00
Silt 0.1 2.62 0.8387 0.0707 0.4 9.29 0.2768 0.0353 0.7 3.03 (I.4869 (I.0(I
Org. carbon (1.1 (I.34 0.1427 0.0724 0.4 (I.(15 (I.2878 0.0370 (I.7 (I.(/2 (t.6855 0.0
Bulk density 0.1 0.10 1.1469 0.0746 0.4 0.10 (I.2993 (I.9615 (1.7 0.10 0.5484 0./10
Hydr. Cond. 0.1 1944 0.1944 0.0000 0.4 1944 0.3888 0.00 0.7 1944 (I,5832 0.00

Clay 0.2 4.31 0.1871 0.0793 0.5 1.94 0.3331 0.00 0.8 3.46 0.3823 0,00
Silt [I.2 4.00 0.1932 0.0820 0.5 2.90 (I.3468 0.00 0.8 3.36 0.4076 0.00
Org. carbon 0.2 0.45 0.2018 0.0855 0.5 0.03 0.3736 0.00 0.8 0.02 (I.4355 0.00
Bulk density 0.2 0.10 0.2085 0.9117 0.5 0.10 0.3736 0.(X) (1.8 0.10 (I.5427 0.00
Hydr. Cond. 0.2 1944 0.1944 0.0000 0.5 1944 (I.5832 0.00 (1.8 1944 0.5832 0.00

Clay 0.3 4.33 0.2221 0.0632 0.6 4.01 0.3609 0.00 (I.9 2.98 (I.6359 0.00
Silt 0.3 3.71 0.2300 0.0653 0.6 4.04 0.3777 0.00 0.9 3.06 (I.3045 0.110
Org. carbon 0.3 0.05 0.2380 0.0676 0.6 0.03 0.3392 0.00 0.9 0.03 0.6638 0.00
Bulk density (1.3 0.10 0.2476 0.9297 0.6 0.10 (I.4183 0.00 0.9 0.10 0.7524 0.00
Hydr. Cond. 0.3 1944 0.1944 0.0000 0.6 1944 0.3888 0.00 0.9 1944 (I.7776 0.00

Table 3
Mean regression coefficients (scaled using variable range) for predicting peak concentration and time to peak
based on the RSM data for L E A C H C using data in Table 1 ( the parameter coding corresponds to that in Table I }

Parameter Coefficients to predict: Parameter Coefficients to predict: Parameter Coefficients to predict:

Peak Time-to- Peak Time- Peak Time-


concentration peak concentration to-peak concentration to-peak

cl0 0.3904 0.2408 c40 0.1241 0.1525 c70 0.2012 0.2045


zl0 0.2246 0.2750 z40 0.2856 0.2220 z7(I 0.1472 (I.0871/
ocl0 0.1916 0.0940 oc40 0.1251 0.0832 oc70 0.2072 0.1226
dbl0 0.3965 0.2750 db40 0.4378 0.2255 db70 0.1658 0.0981
ksl0 0.1728 0.0648 ks40 0.1296 0.0432 ks70 0.1944 0.0864
c20 0.1434 0.1414 c50 0.2768 0.1808 c80 0.2577 0.3670
z20 0.1480 0.1460 z50 0.1222 0.1062 z80 0.2748 0.2030
oc20 0.2113 0.0849 oc50 0.1316 0.1144 oc80 0.2936 0.2168
db20 0.4013 0.1661 db50 0.1316 0.1144 db80 0.1361 0.1065
ks20 0.2160 0.1080 ks50 0.1944 0.1728 ks80 0.1512 0.1296
c30 0.1380 0.1799 c60 0.1958 0.2626 c90 0.1595 0.1984
z30 0.1428 0.1499 z60 0.1235 0.0796 z90 0.0477 0.0274
oc30 0.1300 0.0902 oc60 0.1108 0.0715 oc90 0.0891 0.0449
db30 0.4217 0.2363 db60 0.1367 0.0881 db90 0.1010 0.0509
ks30 0.1080 0.0432 ks60 0.1512 0.0216 ks90 0.1728 0.1296
N.M. Holden et al. / Geoderma 69 (1996) 157-173 167

conductivity and lower bulk density, the model became insensitive, with many coefficients
being 0 when predicting peak concentration, and very close to 0 when predicting time to
peak concentration. When compared to the mean scaled coefficients presented in Table 3,
it can be seen that the bulk density in the simulated surface horizons was important, and
that the other variables were of lesser importance when predicting peak concentration
(Table 4). Greatly altering the conductivity had little effect on the relative importance of
the variables (Table 4).
A measure of the influence of altering bulk density and saturated hydraulic conductivity
was obtained by subtracting peak concentration values obtained with initial variables from
those obtained with extended variables (Fig. 5A) and by using the same procedure for
obtaining "time-to-peak" concentrations differences (Fig. 5B). (These calculations were
performed for each depth but limited data is presented for clarity). All peak concentrations
were reduced by extending the range of conductivity and reducing the bulk density, most
effect being seen in the simulated surface horizons. Increasing the conductivity and lowering
bulk density values resulted in faster times to peak concentration, most importantly in the
deeper horizons.

4. Discussion

4.1. Mean measured breakthrough and prediction using mean variable values

It is the authors' opinion that the mean breakthrough curve at any particular depth is not
of great value because it represents no known mechanism operating in the soil. Furthermore,
there was no reason to believe that the combination of mean input variables would predict
the estimated mean breakthrough. The data (Fig. 2) support this belief. This contrasts with
the findings of Costa et al. (1994) who found mean values to provide good inputs for
predicting mean breakthrough.

4.2. The response o f LEACHC to variation in input variables

The insensitivity of LEACHC to change in input variables was disappointing. It had been
hoped that it could be forced to predict a range of output values given a range of input
values. Figs. 3 and 4 show that, while response varies, its range is small. The extremes of
the predicted breakthrough can be used to generate a graphical envelope of response ( Fig. 6)
which when plotted alongside the measured data shows no similarity. Holden et al. (1995b)
presented a simple test of LEACHCperformed using water release data as input to the model.
With this they managed to predict a greater range of both peak concentration and time to
peak. The use of water release values obtained either from direct measurement or from
particle size analysis data is legitimate because the model can be run using either type of
data as inputs. It is acknowledged that calculation from particle size data is less accurate
(Hutson and Wagenet, 1992) but there is no indication that it results in less sensitivity.
Wagenet and Rao (1983) tested a similar type of model to LEACHCfor predicting pesticide
movement and stated that variation between model runs was "...not as great as would be
expected..." which agrees with the evaluation of LEACHCpresented here.
168 N.M. Holden et al. / Geoderma 69 (1996) 157-173
50

45

0.1m
4o

3~
-- '----. --/" . . . . . . . . . . . . . . . . . . . . . . 0.2m
30
..... , . . . . . /- ,. . . . . . r . . . . . . ,,', . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 03 m
8
__/,,___/~. ...... /~, ....................... 0.4m
2 5 ~ ~ . . ~ : . ...... O.~r~ 0.6m
'., "- -, - .... -. ~~."X2~-..."E,Z~...-;'XZ-:'..-'~.~ 0.7 m 0 . 8 m 0 . 9 m
f
20 L - ~ ~
0 10 20 30 40 50

run n u m b e r A

20 r ~ t~ ~ o

,.~ .... ~ ..... . ...~,...., ...~ ..... 09m


' ' "/ "/ 'V' V ,/ ,/ j ,/ j ' 0.8m

Ix /\ / x / /x
17 : j ~j 'k.-.-J \ ._.~-x_./x.J ~x.___/\___./-~__ o.6m

~ 1 6 1 ~ , , ~ .0.5m

14 / ~" " . ~ " - .... ......... ........ • - 03m

.e 1 3 ~
i - . ....... -~- . - - ... ....... ,,. . . . . . . . . 0.2 m
12, "

=e 11 i ~ ' ~ ~ _ _ 0 . 1 m

10 # . . . . . =
0 10 20 30 40 50
run n u m b e r B

Fig. 4. variation in peak concentration, and (B) variation in time-to-peak concentrations as predicted by
(A)
LEACHC using measured and second estimates of input variables obtained from the North Wyke Soil Block and
implemented using R S M .

4.3. The effectiveness of the RSM approach to model testing

Due to the poor sensitivity of LEACHCin this case, the effectiveness of the RSM approach
to model testing is less clear than would have been the case had the output been markedly
varied. The RSM approach has two favourable characteristics: (i) the model is forced to
predict with a range of input variables, thus information is obtained with respect to sensitivity
and range of response, and (ii) where measured variable values are missing (as in this case,
where the site could not be disturbed in order to measure the values) a range of reasonable
estimate values can be tested with the minimum of difficulty. Results presented here indicate
that the model has been given a reasonable set of data and it is unable to predict breakthrough
as rapid as that measured in the field, thus it might be argued that the model is flawed in
some respect. With the particular model used in this study, the magnitude of an estimated
input variable has been shown to be relevant in altering the output, but when constrained to
realistic variable ranges, the model cannot predict the measured breakthrough from the
field.
N.M. Holden et al. / Geoderma 69 (1996) 157-173 169

Table 4
Mean regressioncoefficients (scaled using variablerange) for predictingpeak concentrationand time to peak
based on the secondestimates of hydraulicconductivityand bulk densityRSM data for LEACHC using data in
Table 1 (the parametercodingcorrespondsto that in Table 1)

Parameter Coefficientsto predict: Parameter Coefficientsto predict: Parameter Coefficientsto predict:

Peak Time-to- Peak Time- Peak Time-


concentration peak concentration to-peak concentration to-peak

clO 0.1272 0.1523 c40 0.0569 0.1235 c70 0.0838 0.1662


zlO 0.1299 0.1362 z40 0.1179 0.1757 z70 0.0229 0.0437
oclO 0.2538 0.2237 oc40 0.0569 0.0693 oc70 0.0322 0.0616
dbl0 1.6822 0.3245 db40 0.6546 0.1896 db70 0.0258 0.0493
ksl0 0.3528 0.2940 ks40 0.1176 0.1176 ks70 0.0000 0.0588
c20 0.0829 0.1204 c50 0.0854 0.1672 c80 0.0809 0.1009
z20 0.0856 0.1243 z50 0.0222 0.0436 z80 0.0248 0.0507
oc20 0.0894 0.0815 oc50 0.0239 0.0466 oc80 0.0265 0.0541
db20 1.1702 0.2512 db50 0.0239 0.0469 db80 0.0278 0.0346
ks20 0.1176 0.1764 ks50 0.0000 0.0000 ks80 0.0588 0.0000
c30 0.0628 0.1166 c60 0.0849 0.2287 c90 0.0746 0.1246
z30 0.0651 0.1225 z60 0.0222 0.0434 z90 0.0295 0.0532
oc30 0.0673 0.0733 oc60 0.0199 0.0389 oc90 0.0326 0.0588
db30 0.6618 0.3196 db60 0.0061 0.0120 db90 0.0370 0.0666
ks30 0.0000 0.0000 ks60 0.0000 0.0000 ks90 0.0588 0.0588

One possible limitation of the RSM as presented concerns the nature of the relationships
between the input variables. There are significant correlations between clay content, silt
content, organic carbon, and bulk density (conductivity being more difficult to predict from
the more easily measured variables). Moreover, certain combinations of these variables
would be very unlikely to occur at any location within the soil type. In the present experiment
this is not a problem as the unlikely combinations generated a potentially broader range of
response which in fact had little effect. If the methodology was being used with a model
where unlikely combinations were having significant impact, then the operator could take
suitable action such as modifying the values according to a pre-determined plan, or omitting
the model run from the analysis.
The possible existence of a higher order response surface was not investigated becasue
of the number of model runs required to reliably estimate the relationship. This limitation
is largely due to the computer time required to run the model and the time required to
prepare input files. Both these problems are surmountable; computers are constantly becom-
ing faster and the model code could be modified to facilitate improved batch operation. A
better approximation of the response surface would allow more specific conclusions about
the significance of input variables within a model.

4.4. Best prediction o f breakthrough using R S M to establish input variable values

Using the response surface regression coefficients, it should be possible to predict the
magnitude of input variables required to obtain a prediction of breakthrough similar to that
measured. From the calculated coefficients, it can be determined that the response surface
170 N.M. Holden et al. / Geoderma 69 (1996) 15~173

0
o 10 20 30 40 50
-1 0.8m

-2
g
-3 .' 0.3m

-4 , 0.2 m

-5

-6 -- 01m
8
-7

-8
run number A

15
O8m

& o.5 02 m

~ 0 •

0 5 ~t0 15 J20 25 30 35 40 45 50

-0.5

-1 run number B
Fig. 5. Differences in (A) peak concentration for 4 depths, and (B) time-to-peak concentration for 3 depths,
based on the two estimates of bulk density and saturated hydraulic conductivity.

of the model is not well approximated by a linear regression because the coefficients were
different when bulk density and conductivity were altered. It is however possible to try and
determine from each set of coefficients the variable values likely to predict near to measured
breakthrough. It is reasonable to conclude that bulk density and conductivity had the greatest
effect. When bulk density was made artificially high (2 g cm 3) the effect was to raise the
peak concentration (Table 5) by up to 30 mg 1- ~ compared to the original runs, and to
alter the time to peak slightly at depth. Raising the conductivity an order of magnitude to
105840 mm d I reduced the peak concentration by 20 mg 1 ] compared to that obtained
with raised bulk density, but this had little effect on the time to peak. The combination of
high bulk density and conductivity produced breakthrough little different from the curves
predicted using the measured and estimated variables indicating that the model is constrained
numerically in the kind of output that it can produce.
4.5. Observation o n t h e p e r f o r m a n c e Of LEACHC

A possible reason for the poor prediction by LEACHC is the individual samplers within
the soil block not conforming to the assumptions of the theory behind the model. The model
N.M. Holden et al. / Geoderma 69 (19961 157-173 171

100 -

80
7O ~,
g 60 i r .-.~ measured
(curves as Fig. 1)
50 i d ih : ':
predicted
40 (envelope of response)

8 30 -,

20 iJ
IO
i

-2 0 2 4 6 8 10 12
time (days)
Fig. 6. The envelope of response of LEACHC based on the variables presented in Table 1 compared with measured
values for 0.40 m within the North Wyke Soil Block at a surface water flux of I mm h ~. The 3 prediction curves
represent the range of response of LEACHC to the range of input data used.

Table 5
Peak concentration and time to peak values for predicted breakthrough using LEACHC with bulk density set high
(2.(I g/cm 2) and conductivity set high ( 105840 m m / d

Depth High bulk Time-to-peak High Time-to-peak Both high peak Time-to-peak
(m) density peak concentration conductivity concentration concentration concentration
concentration (days) peak (days) (rag/l) (days)
(mg/I) concentration
(mg/l)

0.1 70 11.2 51 11.2 71 11.2


0.2 52 11.8 39 12.2 53 11.8
(1.3 44 12.6 33 13 45 12.8
0.4 40 13.4 31 14.6 41 13.4
0.5 39 14 30 15.4 4(1 14
0.6 37 14.6 29 16 38 14.4
0.7 36 15.2 28 16.4 38 15.2
0.8 35 16 28 17.6 37 15.8
0.9 35 17 27 17.8 36 16.2

is based on equations used to describe water movement and chemical transport in a Darcian
flow field of size equivalent to a Representative Elementary Volume (REV - - Bear, 1972).
Each individual sampler is probably not sampling a REV, but this is unclear because the
magnitude ( and shape) of the volume cannot be measured for the soil block. The arithmetiC
mean breakthrough calculated and presented in Fig. 2 is, however, clearly different from
the predicted breakthrough so there is little reason to believe that the breakthrough in an
REV would be significantly slower. It should be noted that the breakthrough predicted by
the mean variable values was little different from those generated during the RSM runs of
the model. It is not possible to assess whether the samplers are working in what could be
termed a Darcian flow field.
172 N.M. Holden et al. / Geoderma 69 (1996) 157-173

Another problem with LEACHC for predicting solute transport in the soil block is the
inability of the model to encompass preferential flow processes. If preferential flow is
thought of as the relative displacement of antecedent soil water (and thus solute) by
incoming water, then where displacement is small (i.e., most incoming water is channeled
through selective lines of flow) there is much preferential flow. LEACHC is not able to
predict such conditions very well because of the macroscale assumptions of Richards'
equation and the Convection-Dispersion equations. The North Wyke soil block is believed
to be dominated by preferential flow processes (Holden et al., 1995a; Scholefield et al.,
1995) which have high spatial variability and are not well simplified in terms of macroscale
assumptions of Darcian flow. A model such as MACRO (Jarvis, 1991) would possibly be
more suitable for predicting solute movement in the soil block because it assumes a two-
phase flow system. Such work on independent evaluation of other one-dimensional models
is being undertaken using RSM.

5. Conclusions

The use of response surface methodology for testing one-dimensional soil water and
solute transport models should prove productive. The design matrix is easily computed for
any number of variables, and implementation of the design is particularly suited to models
that can be run in batch mode (i.e. without the operator needing to initiate each run of the
model). The results from RSM permit a measure of variable sensitivity to be obtained, and
allow the operator to obtain a structured picture of how the model behaves.
In the case of LEACHC, the model has been shown to be insensitive to what should be
critical variables (clay, silt and organic carbon - - the variables controlling pore size distri-
bution), and to have only moderate sensitivity to bulk density and conductivity when used
to predict with the conditions of the experiment undertaken using the North Wyke Soil
Block (Holden et al., 1995a), even when using highly unlikely values for input variables.

Acknowledgements

The field data used in this paper were obtained from work funded by The Natural
Environment Research Council, UK. J.F. Dowd and A.G. Williams assisted in field data
collection. NMH was funded by NERC, (UK). DS and AJR were part funded by BBSRC
(UK).

References

Addiscott, T.M. and Wagenet,R.J., 1985.Conceptsof solute leachingin soils: a reviewof modellingapproaches.
J. Soil Sci., 36: 411-424.
Addiscott, T.M. and Whitmore,A.P., 1991. Simulationof solute leaching in soilsof differingpermeabilities.Soil
Use Manage., 7: 94-102.
Addiscott, T.M., Hays, P.J. and Whitmore,A.P., 1986. Applicationof simple leaching models in heterogeneous
soils. Geoderma,38: 185-194.
N.M. Holden et al. / Geode rma 69 (1996) 157-173 173

Avery, B.W. and Bascomb, C.L. (Editors), 1982. Soil Survey Laboratory Methods. Soil Survey Technical
Monograph No. 6. Soil Survey of England and Wales, Harpenden, 83 pp.
Bear, J., 1972. Dynamics of Fluids in Porous Media. Dover, New York, 764 pp.
Cameron, K.C. and Wild, A., 1982. Prediction of solute leaching under field conditions: an appraisal of three
methods. J. Soil Sci., 33: 659-669.
Costa, J.L., Knighton, R.E. and Prunty, L., 1994. Model comparison of unsaturated steady-state solute transport
in a field plot. Soil Sci. Soc. Am. J., 58: 1277-1287.
Davies, O.L., 1956. Design and Analysis of Industrial Experiments. 2nd. ed. Hafner, New York, 637 pp.
Herzberg, A.M. and Cox, D.R., 1969. Recent work on the design of experiments: a bibliography and a review. J.
R. Stat. Soc. Ser. A, 132: 29-67.
Hill, W.J. and Hunter, W.G., 1966. A review of response surface methodology: A literature survey. Technometrics,
8: 571-590.
Holden, N.M., Scholefield, D., Dowd, J.F. and Williams, A.G., 1995a. Fine resolution chloride breakthrough in a
large soil block with preferential flow. Soil Sci. Soc. Am. J., in press.
Holden, N.M., Scholefield, D., Williams, A.G. and Dowd, J.F., 1995b. Interfacing a fine resolution data set and a
one-dimensional model to test quality of output with respect to preferential flow. In: B. Wiezik (Editor),
Hydrological Processes in the Catchment. Cracow University of Technology, Cracow, Poland, pp. 249-259.
Hutson, J.L. and Wagenet, R.J., 1992. Leaching Estimation And Chemistry Model. A process based model of
water and solute movement, transformations plant uptake and chemical reactions in the unsaturated zone.
Water Resources Institute, Comell University, Ithaca.
Jabbro, J.D., Jemison, J.M., Fox, R.H. and Fritton, D.D., 1994. Predicting bromide leaching under field conditions
using SLIM and MACRO. Soil Sci., 157: 215-223.
Jarvis, N.J., 1989. CRACK: a model of water and solute movement in cracking clay soils. Report 159, Swedish
University of Agricultural Sciences, Department of Soil Science, Uppsala, Sweden.
Jarvis, N.J., 1991. MACRO - - a model of water movement and solute transport in macroporous soils. Swedish
University of Agricultural Sciences, Department of Soil Science, Reports and Dissertations No. 9, Uppsala,
Sweden.
Kleijnen, J.P.C., 1974. Statistical Techniques in Simulation. Part 1. Marcel Dekker, New York, 285 pp.
Kleijnen, J.P.C., 1975. Statistical Techniques in Simulation. Part 2. Marcel Dekker, New York, 775 pp.
Loague, K. and Green, R.E., 1991. Statistical and graphical methods for evaluating solute transport models:
overview and application. J. Contam. Hydrol., 7:51-73.
Myers, R.H., 1971. Response Surface Methodology. Allyn and Bacon Inc., Boston, 246 pp.
Pennell, K.D., Hornsby, A.G., Jessup, R.E. and Rao, P.S.C., 1990. Evaluation of five simulation models for
predicting aldicarb and bromide behaviour under field conditions. Water Resour. Res., 26: 2679-2693.
Roth, K., Jury, W.A., Fltihler, H. and Attinger, W., 1991. Transport of chloride through an unsaturated field soil.
Water Resour. Res., 27:2533-254 h
Ryan, B.F., Joiner, B.L. and Ryan, T.A., 1985. Minitab Handbook. 2nd. ed. PWS-KENT Publishing Co., Boston,
386 pp.
Scholefield, D., Holden, N.M., William, A.G. and Dowd, J.F., 1995. Nitrate transport by preferential flow in a
large soil block under pasture. J. Environ. Qual., in press.
Vinten, A.J.A. and Redman, M.H., 1990. Calibration and validation of a model of non-interactive solute leaching
in a clay-loam arable soil. J. Soil Sci., 41: 199-214.
Wagenet, R.J. and Rao, B.K., 1983. Description of nitrate movement in the presence of spatially variable soil
hydraulic properties. Agric. Water Manage., 6: 227-242.
Willmott, C.J., Ackleson, S.G., Davis, R.E., Feddema, J.J., Klink, K.M., Legates, D.R., O'Donnell, J. and Rowe,
C.M., 1985. Statistics for the evaluation and comparison of models. J. Geophys. Res., 90: 8995-9005.

You might also like