Geology Reference
In-Depth Information
the so-called 'LS factor'. They range from a direct
slope angle and length calculation to methods
using drainage network accumulation (as in the
WATEM/SEDEM model). Also the availability of
remote sensing images for vegetation properties
has added to the popularity. When GIS gained
popularity in the late 1990s as a main tool to use
spatial data and convert a DEM to an LS map, the
USLE proved a simple method to predict the soil
loss in units of t ha
−1
y
−1
for each grid cell (see Mati
et al
., 2000; Sanjay
et al
., 2001; Bayramin
et al
.,
2007; Pandey
et al
., 2008b). However, it is not
clear what these large-scale soil loss maps depict.
Are the values meant to be the soil loss from a
pixel to the next downstream pixel, or the average
soil loss of a field or a slope inside a pixel? If the
former is meant, the model should have provision
for accumulation of sediment over a network (see
examples below); if the latter is meant, the prop-
erties of the pixel are seen as characterizing the
fields/slopes inside the pixel, which may be ques-
tionable. This makes calibration difficult. At best,
spatial patterns of erosion can and should be veri-
fied against reality. This was done for instance by
Bou Kheir
et al
. (2006) who used USLE type
parameters in a factorial approach to determine
erosion susceptibility in five classes of an area in
Lebanon, and checked these classes successfully
against observed rill density.
An example of a more complex approach was
given by Zhou and Wu (2008) who calculated the
soil loss for several provinces in China using
inverse distance interpolation for the rainfall
data, an NDVI (Normalized Difference Vegetation
Index) based assessment of the cover factor and a
30 m resolution DEM for the LS factor. Moreover,
they introduced a sediment delivery ratio (SDR)
to predict the amount of sediment entering the
stream network. This permitted the authors to
compare the measured and predicted annual sedi-
ment load for two years in six gauging stations,
whereby the SDR was used to calibrate the
results. They then interpreted changes in the
SDR in relation to changes in annual rainfall.
Takeuchi
et al
. (2009) adapted TOPMODEL suc-
cessfully to simulate 23 peak runoff events and
sediment loads from a 4423 km
2
catchment in
China. A different strategy is to model the trans-
port capacity and compare that to the sediment
load. This is done at this scale by the WATEM/
SEDEM model and the PESERA model. Both
models use a more complex relief factor as a proxy
for both flow detachment and transport capacity,
and therefore are capable of showing areas with
net detachment and net deposition. PESERA,
moreover, uses a SDR for the soil loss on the
slopes nearest to the stream.
It is clear that such models are very difficult to
calibrate. The necessary data at this scale do not
exist; possibly sediment trapped in dams of large
areas could serve as a calibration, but this has not
been attempted yet as far as we know. Also, as noted
above, it is not quite clear what pixel-based erosion
values signify at this scale. Nevertheless, maps
based on the output of these models are often used.
On this scale the poor model performance for
small events is also encountered. For example,
Lenzi and Di Luzio (1997) use AGNPS to predict
total runoff and sediment loss from 20 events in a
77 km
2
catchment. They show that the runoff is
predicted with an R
2
of 0.97 and the sediment loss
with an R
2
0.72. However, these values are deter-
mined by two very large rainstorms for which the
runoff is well predicted. If these are not consid-
ered, the explained variance for runoff and sedi-
ment loss drops to 0.31 and 0.28 respectively.
Similar results are given for total runoff, peak
runoff and sediment yield by Rode and Frede
(1999), using AGNPS on two German catchments
of 82 and 129 km
2
(23 and 35 events respectively).
While from the above it seems that homogeneity
is an issue for plots, this is perhaps even more so
for catchments where rainfall variability plays an
important role (Capolongo
et al
., 2008).
The GCTE results, however, also show that
many models have problems with the prediction of
extreme events. This could be due to a number of
reasons: firstly, the system may not behave the
same for moderate and large events. During a heavy
rainstorm connectivity may be dynamic and differ-
ent than under normal circumstances, and no
longer match the fixed connectivity of the model.
Also many variables considered as static constants
in the model are in fact dynamic, in particular those
=
