Geology Reference
In-Depth Information
different measures, and variable trends, of soil
erosion for all conditions being equal.
Nonetheless, a survey of the literature will
show that when scientists attempt to test a new
soil erosion model or the application of a model in
a new environment, they almost invariably rely
upon measured data for comparison with model
output results. Also, invariably, the documented
fact that the data have enormous natural variabil-
ity is ignored: the measured data are assumed to
be correct. If one were to model, for example, the
40 replicated plots from the Wendt
et al
. (1981)
study, with essentially the same soil, cover and
rainfall conditions, the input parameters for the
model would be nearly or exactly the same for all
the plots. Given that the models are essentially
all deterministic in nature, the output of the
model would be a single value. In that case one
could see where the modelled value falls within
the distribution of the 40 measured values.
However, if one only has a single (or two at best)
measured erosion values with which to compare,
one has no idea where that measured value lies
within the distribution associated with the natu-
ral measurement variability. In most cases that
variability would be much larger than recognized.
Nearing
et al
. (1999) provided a more universal
scheme for characterizing replicated plot variabil-
ity, and Nearing (2000) attempted to develop a
procedure for using that information in model
validation studies, but those concepts have been
neither widely recognized nor implemented.
A major limitation that erosion modellers face
in quantifying and comparing soil erosion rates is
the lack of long-term data. The paradigm the
world over for funding scientific research is the
two- to five-year grant, which is a serious problem
in terms of collecting long-term data. A study by
Edwards and Owens (1991) found that soil erosion
measurements on nine small watersheds in Ohio
over 28 years were dominated by a few large
storms. The five largest erosion-producing events
out of more than 4000 accounted for 66% of the
total erosion. On one watershed, one storm caused
more than half of the 28-year total. Nearing
et al
.
(2007) looked at 11 years of data from six small
watersheds in the Walnut Gulch Experimental
Watershed in southeastern Arizona. In each case
the single largest storm on the record contributed
between 9% and 11% of the total sediment yield
for the 11-year period of record, and approximately
50% of the sediment yield came from between six
and ten events during the 11 years. Lane and
Kidwell (2003) looked at data from four small
watersheds in the Santa Rita Experimental Range
in southern Arizona measured over 16 years. They
found that the year with the largest erosion event
accounted for between 18% and 26% of the total
measured sediment. This temporal variability is
one reason why we need soil erosion models.
Appropriately constructed, a process-based model
may have the ability to extrapolate a short record
of measured erosion to a longer time frame.
Nonetheless, the problem is that models devel-
oped and parameterized from short records that
do not contain the extreme event probably will
not effectively represent the extreme event. The
most likely scenario will be that the impact of the
extreme event will be under-predicted. This is
one area obviously ripe for further research.
Jetten
et al
. (2003) published a review on the
application of models in terms of spatial distribu-
tions of erosion rates within watersheds. Not sur-
prisingly they found that the models were able to
characterize sediment yields from watershed out-
lets only moderately well, a result they attributed
to 'the high spatial and temporal variation of ero-
sion and sediment transport and our inability to
assess and/or describe this variability in terms of
the input parameters normally used in erosion
models.' The models performed even more poorly
in terms of characterizing the spatial erosional
patterns within the watersheds: 'The application
of the LISEM tested here shows that accurate pre-
dictions at the grid-cell resolution at which the
model is run are impossible.' They found that
the finer and more detailed the resolution for the
model inputs and grid, the worse were the spatial
predictions. Obtaining good spatial predictions of
measured erosion requires extensive and detailed
spatial datasets (van Oost
et al
., 2004). The
number of scientific papers in the literature
