2186
Proceedings of the 18
th
International Conference on Soil Mechanics and Geotechnical Engineering, Paris 2013
Highway One, which could be potentially dangerous to
motorists and cause significant disruption to the Auckland road
network.
The soil at the site consists of weathered soil from the
Northland Allochthon formation, which is renowned for its
montmorillonite content (Power, 2005). This formation is
susceptible to landsliding due to seasonal pore-pressure changes
(Lentfer, 2007; O'Sullivan, 2009). The site consists of 3 strata;
the underlying parent rock, a transition zone and a completely
weathered residual soil. The transition zone consists of
unweathered rock fragments in a silty clayey matrix. This
transition zone has many slickensided shear surfaces present,
and is thought to be one of the underlying factors that give rise
to the susceptibility of the formation to landslides. The residual
soil is a silty clay, susceptible to shrink swell movement. In
general, sites in the Northland Allochthon have high ground
water tables even in dry periods (O'Sullivan, 2009).
For a more detailed description of the site and soil
properties, the reader is referred to Harris et al. (2012).
2
METHODOLOGY
A total of 13 VWC sensors were installed along the same cross
section of the slope; at the toe, mid-height and crest. The
sensors consisted of MP406s and ECH
2
O probes (ICT
International Pty Ltd, 2012), which were installed at
approximately 0.25m depth intervals. A tipping bucket rain-
gauge was used to record rainfall events. Recordings were made
via a data logger at an hourly interval.
SEEP/W
(GEO-SLOPE International Ltd, 2009a) was used
for the FEM. The hourly rainfall captured at the site was input
as an influx into the slope. A general evaporation pattern was
applied to the model as a negative influx between rainfall
events. This generalised evaporation pattern was based on a trial
and error method to get the best agreement between the FEM
results and the field monitoring results. The level of evaporation
applied following a rainfall event was determined by the
cumulative rainfall amount of the event. This FEM was coupled
with the limit equilibrium analysis (LEA) program
SLOPE/W
(GEO-SLOPE International Ltd, 2009b). Thus at each hourly
time step in the FEM, the FOS was obtained.
The soil properties used in these models are given in Table 1.
The soil water characteristic curve was described using the Van
Genuchten (1980) method, the parameters of which were
obtained using the pressure plate apparatus. The permeability
was determined using the falling head method. A variety of
triaxial tests, including constant shear drained tests, were used
to determine the shear strength parameters. The shear strength
values used for the top soil layer were reasonably high to force
the slip surface obtained in the LEA to a reasonable depth. φ
b
represents the angle of shearing resistance due to matric
suctions, as described by Fredlund et al. (1978)
Table 1. Soil parameters used in the models.
Van Genuchten (1980) Parameter Shear Strength
k
a
n
m
θ
r
φ
φ
b
c�
m/hr 10
-3
kPa
-1
%
°
°
kPa
Top Soil
36
608 3.27 0.69 38.5 40 20 10
Residual Soil
0.36
608 3.27 0.69 38.5 36 20 0
Transition
Zone
0.036 297 5.23 0.81 37.2 21 20 3
Underlying
Rock
0.0036 29 5.23 0.81 37.2 35 20 5
The ANN was developed using the software
Matlab
(The
MathWorks Inc, 2012). For more information regarding ANNs
and their use in geotechnical engineering, the reader is referred
to Khanlari et al. (2012). The ANN was trained to predict the
LEA-obtained FOS using the sensor readings from the field
monitoring as inputs. The ANN was developed as a closed-loop
recurrent dynamic network, where the FOS predicted by the
ANN for the previous time-step was used as an input for the
prediction of the FOS for the current time-step. The Levenberg-
Marquardt method (Mathworks, 2010) was used to optimize the
ANN, which had 10 hidden layers. The accuracy of the ANN
improved when cumulative rainfall amounts were included as
inputs into the ANN. Thus, cumulative rainfall amounts ranging
from 2 to 200 hours were included as inputs into the ANN.
A second ANN was developed which predicts the LEA
obtained FOS based solely on rainfall data. Thus the future FOS
could be predicted at the site using rainfall forecasts obtained
from the Meteorological Service of New Zealand (2012).
3
RESULTS
A reasonable agreement was obtained between the field
measured and FEM obtained VWC. The permeability of the top
soil layer had to be increased in the FEM in comparison to the
underlying soil layers to obtain the required infiltration amount.
Presumably this reflects the discontinuities such as surface
cracks and vegetation of the soil. In some locations the
agreement was very good, in others the agreement quite poor
The reason for this is thought to be due to natural variability
within the soil, as described by Dai et al. (2002).
To confirm this modelling process, the rainfall record obtained
from the Meteorological Service of New Zealand (2012) leading
up to the 2008 landslide was input into the models. As a FOS of
just above unity was obtained at approximately the same time as
the landslide occurred, it is assumed that the models used in the
development of this EWS were reasonably accurate.
Because few extreme rainfall events occurred during the field
monitoring period, artificial rainfall events were input into the
rainfall record. Such artificial rainfall events can be seen in the
upper graph of Figure 1, at an elapsed time of 1500 hours and
2200 hours. The comparison between the FOS obtained from
the LEA, that obtained from the ANN using sensor data, and
that obtained from the ANN using just rainfall data is shown in
the lower graph of Figure 1. As observed, at each significant
rainfall event there is a large decrease in the FOS. This FOS
recovers rapidly following the rainfall event.
The ANNs are reasonably accurate at predicting the LEA
obtained FOS. The mean squared error of the ANN using sensor
was 0.41. Using just rainfall data, the mean squared error
increased to 1.16. The FOS predicted by the ANNs is
susceptible to large fluctuations, particularly during times of
evaporation. This is seen at elapsed times of approximately
1600 hours and 2600 hours. Because these fluctuations occur
during times of evaporation, they are not critical to the accuracy
of the EWS; however they do indicate that some discrepancies
occur due to the generalised evaporation pattern which was
used. If a deterministic approach was used to measure
evaporation, such as that described by Penman (1948), it is
thought that such discrepancies will be minimised. The
improvement in accuracy from the ANN which uses just rainfall
data as an input, compared to the ANN which uses sensor data
also, indicates that the use of the sensors provides an indication
as to the actual amount of rainfall infiltration in the slope.
To provide an example of the EWS in use, the data
corresponding to the point shown in Figure 1 was input into the
EWS. The resulting plot is shown in Figure 2. Elapsed time = 0
corresponds to the point in time when the data was obtained
from the site. The upper graph in Figure 2 shows the rainfall
record obtained from the site (from an elapsed time of -24 to an
elapsed time of 0). The rainfall is constant as it is obtained
during an artificial rainfall event, as shown in Figure 1 (a). The
rainfall record in Figure 2 from an elapsed time of 0 to and
elapsed time of 5 hours is that obtained from the forecast.
The solid line in the lower graph of Figure 2 is the ANN –
predicted FOS of the last 24 hours, using the sensor data as
inputs. The dotted line is the predicted FOS over the next 5
