This paper was peer-reviewed for scientific content.

Pages 919-923. In: D.E. Stott, R.H. Mohtar and G.C. Steinhardt (eds). 2001. Sustaining the Global Farm. Selected papers from the 10th International Soil
Conservation Organization Meeting held May 24-29, 1999 at Purdue University and the USDA-ARS National Soil Erosion Research Laboratory.
Using Fuzzy Logic-Based Modeling to Improve the Performance of the
Revised Universal Soil Loss Equation
L.T. Tran*, M.A. Ridgley, M.A. Nearing, L. Duckstein and R. Sutherland

*L.T. Tran, M.A. Ridgley, and R. Sutherland, University of Hawaii, Geography Department, 2424 Maile Way, Honolulu, HI 96822,
USA; M.A. Nearing, National Soil Erosion Laboratory, USDA-ARS, West Lafayette, IN 47907-1196, USA; L. Duckstein, Ecole Nationale
du Genie Rural, des Eaux et des Forets, 19 avenue du Maine, 75732 Paris Cedex 15, France. *Corresponding author: [email protected].
Current address: Center for Integrated Regional Assessment, Penn. State University; 2217 Earth & Engineering Science Building, University
Park, PA 16802, USA.
ABSTRACT
This paper reports the application of fuzzy logic-
based modeling (FLBM) to improve the performance of
the Revised Universal Soil Loss Equation (RUSLE). The
FLBM approach was to make the RUSLEs structure
more flexible in describing the relationship between soil
erosion and RUSLE factors and in dealing with data and
model uncertainties while not requiring any further
information. The approach used in this study consists of
two techniques: multiobjective fuzzy regression (MOFR)
and fuzzy rule-based modeling (FRBM). First, MOFR
was used to derive the relationship between soil loss and
a combination of RUSLE factors. These MOFR models
were in turn linked together in a FRBM framework.
Then these fuzzy rules were applied to adjust the RUSLE
prediction corresponding to each combination of RUSLE
factors.
The Nash-Sutcliffe model efficiency of the fuzzy
model on a yearly basis was 0.70 while RUSLE's was
0.58. On an average annual basis, the efficiency was 0.90
and 0.72 for the fuzzy model and RUSLE, respectively.
With several good characteristics, the FLBM approach
can be used to improve the performance of RUSLE with
little effort and modification to the existing RUSLE
model.
INTRODUCTION
This paper reports the application of fuzzy logic-based
modeling (FLBM) to improve the performance of the
Revised Universal Soil Loss Equation (RUSLE). For the
purpose of conservation planning, the prediction accuracy of
RUSLE is very important in making sound decisions on how
soil should best be protected from erosion. However, an
analysis of over 1700 plot-years of data, taken from more
than 200 plots at 21 sites in the U.S., showed that soil
erosion was not adequately described merely by the
multiplication of six RUSLE factors in all cases. For
instance, data indicated that the relationship between rainfall
erosivity factor (EI) and soil loss, when other RUSLE
factors were held constant, was not always linear. The aim
of the FLBM approach was to make the RUSLEs structure
more flexible in describing the relationship between soil
erosion and RUSLE factors and in dealing with data and
model uncertainties while not requiring any further data. The
paper is organized as follows RUSLE and its limitations
are discussed in the next section. Data and methodology are
presented in the section 3. Section 4 is devoted to results and
discussion followed by conclusion in the last section.
The Revised Universal Soil Loss Equation
RUSLE is an empirical equation derived from a large
amount of field data and has been widely used as an erosion
prediction and conservation planning tool in the U.S. as well
as worldwide. The model computes soil erosion using values
representing major factors influencing erosion, including
climate erosivity, soil erodibility, topography, and land use
and management. Keeping the same format of the Universal
Soil Loss Equation (USLE), RUSLE is expressed as follows:
A = R K L S C P
Where A is the mean annual soil loss (t ha
-1
); R is the
rainfall and runoff erosivity (MJ mm ha
-1
h
-1
); K is the soil
erodibility factor (t ha h (ha MJ mm)
-1
); LS is the combined
dimensionless slope length and slope steepness factor; C is
the dimensionless cover-management factor, and P is the
dimensionless supporting practices factor.
The rainfall and runoff factor (R) is the average annual
total of the storm EI values, which equal the total storm
energy (E) times the maximum 30-min intensity (I
30
). The
relation between soil losses to the EI parameter was assumed
to be linear (Renard et al., 1997). Compared to USLE,
RUSLE included more precise values of R for the western
half of the United States; and more corrections, more refined
smoothing, and the filling of data gaps for the eastern United
States (Renard et al., 1997).
Generally K in RUSLE was computed in a similar
manner as in USLE. The difference of K between the two
models is an adjustment added in RUSLE to take into
account seasonal changes, such as freezing and thawing, soil
moisture and soil consolidation (Renard et al., 1994a). On
the other hand, compared to USLE, LS in the RUSLE was
refined by assigning new equations based on the ratio of rill
to interrill erosion and accommodating complex slopes
(Renard et al., 1994b).
The C factor in USLE was computed based on cropping
sequence, surface residue, surface roughness, and canopy

cover, weighted by the percentage of erosive rainfall during
the six crop stages. Compared to USLE, RUSLE included
some more factors in determining C, such as prior land use,
canopy cover, surface cover, surface roughness, and soil
moisture. However, the key difference in computing C
between the two models is the use of time-varying
computation with a 15-day interval in RUSLE (Renard et al.,
1997), which was considered a major improvement in
estimating soil loss (Renard et al., 1994b).
P in RUSLE, which was also considered more advanced
than those in USLE, was computed with a combination of
empirical and process-based erosion technology, based on
hydrologic soil group, slope, row grade, ridge height, and the
10-year single storm erosion index value (Renard et al.,
1997). A complete description of all RUSLE factors can be
found in USDA Agricultural Handbook Number 703
(Renard et al., 1997).
Risse et al. (1993) carried out a comprehensive analysis
to assess the error associated with the USLE. Results from
this study showed that the overall Nash-Sutcliffe model
efficiency was 0.75 on an average annual basis and 0.58
when compared on a yearly basis. These values are
considered reasonable as the same data set showed an annual
erosion variability of 35% between replicated plots (Yoder
et al., 1998). However, looking at more detail, USLE over
predicted soil loss on plots at low erosion rates while
underpredicted for plots at high erosion rates.
Relying heavily on Risse et al.s work and using data
from the same sites and the same periods, Rapp (1994)
found a similar result for RUSLE with a model efficiency of
0.73 on an average annual basis and 0.58 on a yearly basis.
Similar to USLE, RUSLE tended to overpredict on plots
with low erosion rates and under predict on plots with high
erosion rates. For this similarity in performance of USLE
and RUSLE, Yoder et al. (1998) explained that RUSLE was
developed from the basic USLE for the purpose of extending
its application rather than increasing its predictive accuracy
for normal cropping situations. As a consequence, they
would be expected to provide a similar degree of accuracy.
With more data and many improvements, RUSLE was
considered scientifically superior to USLE (Renard et al.,
1994b). Consequently, it was expected that RUSLE would
perform better than USLE. However, this was not the case as
presented above. Besides the issue of data limitation, the
relatively moderate performance of RUSLE can be attributed
to several theoretical problems. First of all, to make the
model more practicable, runoff, an important factor to which
soil loss is closely related, was not explicitly dealt with but
incorporated within the R factor (Morgan, 1995). On the
other hand, although soil losses are directly proportional to
the EI parameter, this relation is generally nonlinear.
Scattergram of soil losses versus EI values at barley sites in


Figure 1. Measured and RUSLE-predicted soil loss versus EI values for barley sites.
0
10
20
30
40
50
60
70
80
90
0 50 100 150 200 250 300
EI(100.ft.tonf.in/acre/h)
S
o
i
l

l
o
s
s
(
t
o
n
/
a
c
r
e
)
Transition zone between
rules 1 & 2
Fuzzy rule 1
(MOFR 1)
Fuzzy rule 2
(MOFR 2)
RUSLE
MOF
R 1
MOF
R 2
Combination of MOFR
1 & MOFR 2
Upper bound of
MOFR 2
Lower bound of
MOFR 2
Measured soil loss

M
e
a
s
u
r
e
d

S
o
i

L
o
s
s

&

(
P
r
e
d
i
c
t
e
d

-

M
e
a
s
u
r
e
d
)
0
2
4
6
Measured RUSLE-Measured FRUSLE-Measured
(
t
o
n
/
a
c
r
e
)
(a)
0
2
4
6
8
10
Measured RUSLE-Measured Fuzzy model-Measured
(
t
o
n
/
a
c
r
e
)
(b)
-2
0
2
4
6
8
Measured RUSLE-Measured Fuzzy model-Measured
(
t
o
n
/
a
c
r
e
)
(c)
-4
0
4
8
12
16
Measured RUSLE-Measured Fuzzy model-Measured
(
t
o
n
/
a
c
r
e
)
(d)
-20
0
20
40
60
Measured RUSLE-Measured Fuzzy model-Measured
(
t
o
n
/
a
c
r
e
)
(e)
Mean Median Standard deviation

Figure 2. Expected accuracy of the RUSLE and the fuzzy model for different categories of
measured soil loss (ton/acre) on a yearly basis; (a), <1; (b), 1-5; (c), 5-10; (d), 10-20; and (e), >20.

Figure 1 illustrates this point. Perhaps the linear assumption
in the RUSLE is only applicable within a certain range but
not the whole scope of all EI values. With the structure of a
simple multiplication of several factors, RUSLE, however,
cannot accommodate such a nonlinear relationship.
Another theoretical problem is that the interdependence
between different factors is quite considerable. For instance,
the relationships between L and S in USLE were derived
from soil-loss measurement from mostly medium-textured,
poorly aggregated surface soils (Wischmeier and Smith,
1978). These relationships, in turn, were used to determine K
values. As a consequence, errors and shortcomings from
these relationships would carry over into K (Renard et al.,
1997). There is also a similar problem for the rainfall
erosivity factor R (Wischmeier and Smith, 1978).
Furthermore, interactions between K and C caused
problems in delineating values of these two factors (Renard
et al., 1997).
DATA AND METHODOLOGY
Data
The data set used in this study was obtained from Rapp
(1994). It was originally supplied by the USDA-ARS
Southwestern Watershed Research Center, containing year-

by-year information of over 1700 plot-years from more than
200 individual plots at 21 sites. It included previously
determined USLE factor values, crop types and yields, and
rotation sequences for each year and the plot dimensions.
The number of plots and duration were different from site to
site. The average is nearly eight years per plot for all 21
sites. Many plots had duplicated measurements to take into
account the natural variability of soil loss. With the original
data set, Rapp (1994) used the RUSLE computer program to
identify the RUSLE factor values by providing all necessary
information to the program. This data set is considered
reliable since this work was supervised by Ken Renard, the
team leader of the RUSLE project. The complete data set as
well as a list of sites with individual plot conditions can be
found in Rapp (1994).
Methodology
As analyzed in the previous section, the RUSLE had
several problems related to model structure and parameters.
To make the RUSLEs structure more flexible in describing
the relationship between soil erosion and the RUSLE factors,
and in dealing with uncertainties of parameters, while not
requiring any additional information, a FLBM approach was
applied. This approach includes two techniques: multi-
objective fuzzy regression (MOFR) and fuzzy rule-based
modeling (FRBM).
A fuzzy rule-based model comprises several single fuzzy
rules. Each fuzzy rule generally consists of a set of fuzzy
set(s) as argument(s) A
k
and a consequence B also in the
form of a fuzzy set such that
If A
1
and A
2
and..... and A
K
then B
A detailed technical discussion of FRBM can be found in
Brdossy and Duckstein (1995). In this study, each fuzzy
rule was derived by the means of MOFR (discussed later) to
describe the relationship between soil loss and the EI factor
within a certain range of other of RUSLE factors. There are
different methods for combining fuzzy rule consequences.
The method of additive combination of fuzzy rule responses
was used in this study (Brdossy and Duckstein, 1995).
Often a consequence from applying a fuzzy rule system is a
fuzzy set. The task of transforming a fuzzy consequence into
a crisp number is called defuzzification. For this study, the
maximum-weighted sum defuzzification method developed
by Tran (1999) was utilized.
MOFR is a fuzzy regression model developed by Tran
(1999) and Tran and Duckstein (accepted), which is capable
of combining central tendency and possibilistic properties of
statistical and fuzzy regression, respectively. MOFR
overcomes several shortcomings of fuzzy and statistical
regression approaches (e.g., sensitivity to data outliers of
fuzzy regression, difficulties of verifying distribution
assumptions, insufficient and/or inaccurate input and/or
output data, vagueness of the relationship between input and
output variables in statistical regression). Furthermore,
MOFR is robust with respect to y-direction outlier (often
referred to simply as outlier, as distinct from the x-direction
outlier often referred to as the leverage point). Hence, the
model can be used when only few data are available, while
this is not the case for least-squares regression. This feature
is essential for MOFR to be applied to a small subset of
RUSLE data. A technical description of this MOFR can be
found in Tran (1999), and Tran and Duckstein (accepted).
In fact the FLBM approach used in this study can be
viewed as a fuzzy piecewise linear regression model
(FPLRM), in which MOFR was applied to derive the linear
equation for each segment of data and FRBM was used to
link those segments together. However, in contrast with
conventional piecewise regression models, FPLRM does not
have a single joint point but a nonlinear curve in the
transition zone between two consecutive segments (Figure
1).
To apply the fuzzy rule-based modeling approach to
RUSLE, the following steps were carried out:
1. Divide the data set into several subsets of certain RUSLE
parameter ranges;
2. Apply MOFR to derive the relationship between soil loss
and EI for each certain range of other RUSLE factors;
and
3. Use FRBM to link these MOFR models and apply the
fuzzy rule set to compute predicted soil loss
corresponding to a combination of RUSLE factors.
A detailed description of these steps can be found in
Tran (1999). Figure 1 illustrates the fuzzy logic-based model
applied for barley data.
Predictions from the RUSLE and the fuzzy model were
compared with measured data in several ways, including r
2

and the model efficiency defined by Nash and Sutcliffe
(1970), which is calculated as follows:

=
N
i
m mi
N
i
ci mi
] ) Q Q ( [
] ) Q Q ( [
R
1
2
1
2
2
1

where Q
mi
and Q
ci
are the measured and computed values,
respectively, of event i, and

m
Q
is the mean of measured
values.
RESULTS AND DISCUSSION
Results showed that the fuzzy model performed better
than RUSLE. The Nash-Sutcliffe model efficiency of the
fuzzy model on a yearly basis was 0.70 while RUSLE's was
0.58 (Table 1). On an average annual basis, the efficiency
was 0.90 and 0.72 for the fuzzy model and RUSLE,
respectively, an improvement of >25% with respect to the
performance of RUSLE. Similar picture can be seen if r
2
is
used to compare the performance of RUSLE and the fuzzy
model (Table 1). Furthermore, the problem of over
prediction at low soil loss rates and under prediction at high
soil loss rates was reduced with the fuzzy model. Fig. 2
shows that the over prediction at soil loss rates <10 ton ac
-1

was decreased significantly. On the other hand, the fuzzy
model lessened considerably the over prediction at soil loss
rates >20 ton ac
-1
. It should be mentioned that over
prediction at low soil loss rates is less critical than under
prediction at high soil loss rates.
The FLBM approach applied to RUSLE did not only
make the structure of the model more flexible and more
realistic in describing the relationship between soil loss and
rainstorm parameter, but it also overcame the problems of
uncertainty in the RUSLE parameters. For instance, due to

Table 1: model on a yearly basis and on an average annual basis.
Parameter Subset n RUSLE RUSLE fuzzy model
On a yearly basis
Nash & Sutcliffe Calibration 957 - 0.612 0.666
model efficiency Validation 745 - 0.576 0.753
Whole 1702 0.586 0.599 0.698
r
2
Calibration 957 - 0.617 0.674
Validation 745 - 0.576 0.753
Whole 1702 0.604 0.603 0.704
On an average annual basis
Nash & Sutcliffe
model efficiency
203 0.719 0.721 0.896
r
2
203 0.736 0.740 0.905
No recalibration was made for the RUSLE.
Recalibration was made for the RUSLE using the calibration subset (957 data points).


interdependency between variables, values of a particular
RUSLE parameter used in combination with other
parameters within certain ranges might be incorrect.
Through multiplication of all factors and nothing else to
compute soil loss, this kind of error directly affects the result
of the RUSLE. In contrast, the fuzzy model took into
account the issue of interdependency and provided the best
fit between soil loss and rainstorm energy.
In addition to central values, the fuzzy model provided lower
and upper bounds on the predicted range of soil loss (Fig. 1).
These bounds are valuable information for both scientific
understanding and for management decisions. For instance,
the level of uncertainty associated with a given range of
RUSLE factors can be evaluated from these predicted ranges
of soil loss.
In terms of modeling, FRBM makes the tasks of model
development and updating quite easy. It is because each
fuzzy rule can be developed or updated independently. This
feature is not available in conventional modeling techniques,
as the work must be done over for the entire model.
CONCLUSION
This analysis showed that the FLBM approach has
several good features. For instance, the approach is quite
simple, no other data outside RUSLE are needed, and the
main structure of RUSLE is maintained. Moreover, the
performance of RUSLE was improved significantly with the
use of this approach. Hence it would be worth a try to
employ the whole data set used in RUSLE to develop the
fuzzy RUSLE model.
REFERENCES
Brdossy, A. and L. Duckstein. 1995. Fuzzy rule-based
modeling with applications to geophysical, biological
and engineering systems. CRC Press, Boca Raton.
Morgan, R.P.C. 1995. Soil erosion and conservation (2
nd

Edition). Longman, England.
Nash, J.E. and J.E. Sutcliffe. 1970. River flow forecasting
through conceptual model. Part 1 A discussion of
principles. Journal of Hydrology (Amsterdam) 10: 282-
290.
Rapp, J.F. 1994. Error assessment of the Revised Universal
Soil Loss Equation using natural runoff plot data. Master
Thesis, School of Renewable Natural Resources, The
University of Arizona.
Renard, K.G., G.R. Foster, D.C. Yoder and D.K. McCool.
1994a. RUSLE revisited: Status, questions, answers, and
the future. Journal of Soil and Water Conservation 49
(3): 213-220.
Renard, K.G., J.M. Laflen, G.R. Foster and D.K. McCool.
1994b. The Revised Universal Soil Loss Equation. p.
105-124. In R. Lal (ed.) Soil Erosion, Research and
Methods (2
nd
Edition). Soil and Water Conservation
Society, Ankeny, Iowa.
Renard, K.G., G.R. Foster, G.A. Weesies, D.K. McCool and
D.C. Yoder. 1997. Predicting soil erosion by water: a
guide to conservation planning with the Revised
Universal Soil Loss Equation (RUSLE). USDA
Agricultural Handbook, No. 703. U.S. Gov. Print. Office,
Washington, DC.
Risse, L.M., M.A. Nearing, A.D. Nicks and J.M. Laflen.
1993. Error assessment in the Universal Soil loss
Equation. Soil science Society of America Journal 57:
825-833.
Tran, L. and L. Duckstein. Accepted. Multi-objective fuzzy
regression with central tendency and possibilistic
properties. Fuzzy Sets and Systems.
Tran, L. 1999. Multi-objective Fuzzy Regression and its
Application to Erosion Prediction. Ph.D. Dissertation,
Department of Geography, The University of Hawaii.
Wischmeier, W.H. and D.D. Smith. 1978. Predicting rainfall
erosion losses - A guide to conservation planning. USDA
Agricultural Handbook No. 537. U.S. Gov. Print. Office,
Washington, DC.
Yoder, D.C., G.R. Foster, G.A. Weesies, K.G. Renard, D.K.
McCool and J.B. Lown. 1998. Evaluation of the RUSLE
soil erosion model. Proceedings of the 1998 ASAE
Annual International Meeting, Orlando, Florida, July 11-
16, 1998.