Document Actions

IJEGE-11_BS-Bellugi-et-alii

background image
Italian Journal of Engineering Geology and Environment - Book www.ijege.uniroma1.it © 2011 Casa Editrice Università La Sapienza
399
DOI: 10.4408/IJEGE.2011-03.B-045
SPATIALLY EXPLICIT SHALLOW LANDSLIDE SUSCEPTIBILITY
MAPPING OVER LARGE AREAS
d
ino
BELLUGI
(*, **)
, w
illiam
E. DIETRICH
(*)
, J
onatHan
STOCK
(***)
, J
im
m
C
kean
(****)
,
b
Rian
KAZIAN
(*****)
& P
aul
HARGROVE
(******)
(*)
Department of Earth & Planetary Science, University of California, Berkeley
(**)
Department of Computational Science and Engineering, University of California, Berkeley
(***)
United States Geological Survey, Menlo Park, California
(****)
United States Forest Service, Rocky Mountain Research Station, Boise, Idaho
(*****)
Intel Corporation, Santa Clara, California
(******)
Lawrence Berkeley National Laboratory, Berkeley, California
likely extent of hazard for a given storm. This suggests
that campaigns to collect local precipitation data and
detailed shallow landslide location maps after major
storms could be used to calibrate models and improve
their use in hazard assessment for individual storms.
K
ey
words
: shallow landslides, drainage area, slope stabili-
ty, Shalstab, parallel computing
INTRODUCTION
Downscaling of climate model precipitation pre-
dictions can now generate spatially detailed (i.e. at
a scale of kilometres to tens of kilometres) maps of
storm rainfall that include orographic and wind effects
(e.g. G
Rell
et alii, 1995; G
ioRGi
et alii, 2001, m
iCHa
-
lakes
et alii, 2001; w
idmann
et alii, 2003; s
alatHé
,
2003; R
oe
, 2005; a
ndeRs
et alii, 2007; d
ettinGeR
et
alii, in press). Such climate models typically operate
over large areas (regional to continental to global).
This allows an exploration of the effects of extreme
storm events on hazard generation (through flooding
and landsliding) and emergency response over these
large areas. Landslides are always local, that is they
are spatially discrete events, and from a management
perspective the more spatially explicit the landslide
susceptibility can be delineated, the more useful it is
for planning. Gross maps based on threshold slopes
derived from coarse grained digital elevation data (e.g.
30 m grid), for example, provide limited guidance. The
increasing availability of higher resolution topographic
ABSTRACT
Recent advances in downscaling climate model
precipitation predictions now yield spatially explicit
patterns of rainfall that could be used to estimate shal-
low landslide susceptibility over large areas. In Cali-
fornia, the United States Geological Survey is explor-
ing community emergency response to the possible
effects of a very large simulated storm event and to
do so it has generated downscaled precipitation maps
for the storm. To predict the corresponding pattern
of shallow landslide susceptibility across the state,
we have used the model Shalstab (a coupled steady
state runoff and infinite slope stability model) which
susceptibility spatially explicit estimates of relative
potential instability. Such slope stability models that
include the effects of subsurface runoff on potentially
destabilizing pore pressure evolution require water
routing and hence the definition of upslope drainage
area to each potential cell. To calculate drainage area
efficiently over a large area we developed a parallel
framework to scale-up Shalstab and specifically intro-
duce a new efficient parallel drainage area algorithm
which produces seamless results. The single seamless
shallow landslide susceptibility map for all of Califor-
nia was accomplished in a short run time, and indicates
that much larger areas can be efficiently modelled. As
landslide maps generally over predict the extent of in-
stability for any given storm. Local empirical data on
the fraction of predicted unstable cells that failed for
observed rainfall intensity can be used to specify the
background image
D. BELLUGI ,w. E. DIETRICH, J. STOCk, J. MckEAN, B. kAZIAN & P. HARGROVE
400
5th International Conference on Debris-Flow Hazards Mitigation: Mechanics, Prediction and Assessment Padua, Italy - 14-17 June 2011
cap/arkstorm.html for details). They provided hourly
rainfall simulation for all of California with resolution
ranging from 2 to 6 km. The challenge here is can we
predict shallow landslides across the entire state us-
ing the highest resolution state-wide topographic data
(10m cells) and assess not only the potential location
by the also the magnitude (number of landslides) for a
given area given the predicted rainfall patterns.
There are many approaches for mapping shallow
landslide potential across a landscape, ranging from
purely empirical statistical approaches to process
based models (e.g. review in C
asadei
et alii, [2003]).
There are insufficient studies in California to map the
shallow landslide risk solely on statistical relation-
ships. We elect as a first step in using the new drain-
age area algorithm to use the Shalstab model [m
ont
-
GomeRy
& d
ietRiCH
, 1994; d
ietRiCH
et alii, 1995;
d
ietRiCH
et alii, 2001] which couples a steady state
shallow subsurface flow model (to predict pore pres-
sure distribution) with an infinite slope model. This
model does not allow us to relate storm magnitude to
landslide frequency, but its simplicity allows it to be
applied across the entire state of California in the ab-
sence of spatially explicit data on such controlling fac-
tors as soil depth, root strength, soil saturated conduc-
tivity, and soil friction angle. Landslide susceptibility
maps produced using the Shalstab model (or other
hydrologically dynamic models), which successfully
delineate areas of failure tend to also greatly over pre-
dict the extent of landsliding for a given storm event
(e.g. C
asadei
et aii., [2003]). Following Stock et al.
[this volume] we suggest that this problem can be ad-
dressed empirically by relating the proportion of un-
stable cells that fail in a given storm to some measure
of that storm magnitude
This first application of the slope stability model
in a parallelized framework points to a need for sys-
tematic data collection on precipitation and landslide
locations for a wide range of storms and across the
diverse landscape found in a large area such as Cali-
fornia. It is not clear to us at this point whether more
mechanistic landslide models will be able to overcome
the problem of over prediction and without consider-
able local calibration successfully delineate specific
areas of instability for a given storm.
METHODS
TOPOGRAPHIC DATA
data (10m grid to 1 m (LiDAR-based)), the grid size
of which approaches the scale of common shallow
landslides (typically involving just the soil mantle),
invites more mechanistic landslide models which have
the potential to be broadly applicable. Such models,
however, typically use hydrologic models to predict
potential destabilizing pore pressure, and this means
that drainage area to every cell must be determined.
For small areas, this is readily accomplished, but for
large areas (e.g. > 100 km
2
for 1 m data or 10,000 km
2
for 10 m data) conventional algorithms will not work
efficiently because they cannot address the necessary
amount of memory. This is because the computation of
drainage area is a global operation (i.e. the information
needed may theoretically come from any part of the
landscape) and thus particularly challenging to imple-
ment seamlessly and efficiently.
Here we report the successful development of a
parallel framework for large-scale spatially explicit
landslide susceptibility assessment, and the implemen-
tation of an efficient seamless parallel drainage area
algorithm. Our drainage area algorithm, available on-
line at the National Center for Earth-surface Dynamics
(http://www.nced.umn.edu/), is not dependent on the
choice of landslide model and stands separately from
its application to this problem.
This algorithm development was motivated by a
challenge. As described in s
toCk
et alii (this volume),
the United States Geological Survey (USGS) is con-
ducting a study of the emergency preparedness in the
state of California (414,000 km
2
) for the effects of ex-
treme storms. Ar
k
Storm, their emergency-preparedness
scenario, is intended to represent the most extreme
storm events that have struck California (analogous
those that devastated the state in 1861–62), as it is
assumed such events will now increase in likelihood
with global warming effects on climate. Recent work
by d
ettinGeR
et alii [in press] and R
alPH
et alii [2006]
has illustrated that some of California’s most damaging
storms are atmospheric rivers of moisture that originate
in the tropics and convey vast amounts of water vapour
towards California in a narrow jet (e.g., 100-200 km
wide) of moisture. When they strike California, these
jets can deliver record rainfalls over the course of 1-3
days. Combining data from California’s largest recent
storms (1969 and 1985) d
ettinGeR
et alii [in press]
simulated an atmospheric river storm in a dynamic
meteorological model (see http://meteora.ucsd.edu/
background image
SPATIALLY EXPLICIT SHALLOW LANDSLIDE SUSCEPTIBILITY MAPPING OVER LARGE AREAS
Italian Journal of Engineering Geology and Environment - Book www.ijege.uniroma1.it © 2011 Casa Editrice Università La Sapienza
401
soil friction angle. Combining the two above equations
we obtain the Shalstab equation in its simplest form:
In essence, this model captures the topographic
(i.e. area and slope) control on the spatial variability
of pore pressures, and expresses the relative potential
for shallow landsliding in terms of the ratio of the ef-
fective (steady-state) precipitation and the capability
of the soil to conduct water, in a spatially explicit fash-
ion. In this simple model, the only parameters are the
friction angleφ , and the (saturated) soil bulk density
ρ
s
, which we set to values of 45 º and 1.7 g/cm
3
, fairly
typical of unconsolidated cohesionless soils.
Two important issues arise in the application of
this model. The first is that mostly due to the hydro-
logical simplifications, the model can not be linked to
specific storm events The second is that as mentioned
above, this model tends to over-predict the landsliding
potential [d
ietRiCH
et alii, 2001]. These issues point
to the need for a calibration procedure discussed in a
following section.
With respect to parallelization, it is important to
note that computing the right hand side of equation (3)
is a purely local operation: every grid cell has a local
slope θ, and a drainage area a, both already assigned.
This is what is referred to as an “embarrassingly par-
allel operation” [f
osteR
, 1995], as no communication
is required between any two cells. The computation
of slope is performed on a 3 by 3 neighborhood, us-
ing a moving window. Communication in this case
is required only when operating at the boundary of a
processor’s data domain. Thus, the real challenge for
the parallelization of Shalstab lies entirely in the area
computation, which is made possible by our new par-
allel algorithm.
PARALLEL FRAMEwORk
We utilize the Unified Parallel C (UPC) language
[C
aRlson
, et alii, 1999], an extension of the C lan-
guage based on the Partitioned Global Address Space
(PGAS) parallel programming model. UPC offers
programming abstractions similar to shared memory
(where any process can read/write data allocated by
another process), while allowing control over data
layout that is critical to high performance and scal-
ability [y
eliCk
et alii, 2007]. UPC offers the program-
Basic topographic data for California were ob-
tained from USGS’s National Elevation Dataset
(http://ned.usgs.gov/). The data are gridded at a reso-
lution of 1/3 arc-second (approximately 10 m). Pre-
processing consisted in assembling the data into larger
tiles (1000 tiles, 972 km
2
each) and removing depres-
sions using the standard ArcGis Fill tool. Depressions
greater than 1.5 m were ignored, effectively prevent-
ing closed basins (e.g. Death Valley) from being filled.
During the pre-processing steps the data were resam-
pled at exactly 10m resolution, and saved in standard
binary IEEE 32-bit floating point format for compat-
ibility and efficiency.
SHALLOw LANDSLIDE SUSCEPTIBILITY
MODEL
We adopted the widely used shallow landslide
susceptibility model Shalstab [Montgomery and Di-
etrich, 1994; d
ietRiCH
et alii, 1995; d
ietRiCH
et alii,
2001] to estimate relative potential of shallow land-
slides initiated by storm rainfall. Shalstab couples a
hydrological model to a limit-equilibrium slope stabil-
ity model to calculate the critical steady-state rainfall
necessary to trigger slope instability at any point in a
landscape. Rainfall is assumed to infiltrate to a lower
conductivity layer and flow along topographically de-
termined paths above an impermeable layer. Under
the simplifying assumption that soil transmissivity
does not vary with depth, the degree of saturation of
the soil profile can be written as
where h (m) is the height of the saturated soil
above the impermeable layer, z (m) is the total height
of the soil, q (m/day) is the effective steady-state pre-
cipitation, a (m2)is the upslope drainage area, b (m)is
the grid cell width, T (m
2
/day) is the depth-integrated
soil transmissivity and
θ
(degrees) is the local
slope. For the general (and conservative) case of co-
hesionless soils, the one-dimensional infinite slope
stability model can be expressed as:
where ρ
s
and ρ
w
are the bulk densities of soil and
water, z is the soil thickness,
g
is gravity and φ is the
(1)
(2)
(3)
background image
D. BELLUGI ,w. E. DIETRICH, J. STOCk, J. MckEAN, B. kAZIAN & P. HARGROVE
402
5th International Conference on Debris-Flow Hazards Mitigation: Mechanics, Prediction and Assessment Padua, Italy - 14-17 June 2011
upstream tiles which may not be identifiable a priori.
This presents a significant obstacle for parallelization,
as it is not possible to determine a partitioning scheme
for the grid such that communication and cross-thread
dependencies are minimized during this computation.
For this reason we develop a new efficient parallel
drainage area algorithm.
The grid over which upslope area information
must travel to each cell can be viewed as a con-
nected directed graph [d
asGuPta
et alii, 2006], with
the nodes representing the grid cells and with edges
representing flow from uphill to downhill neighbours
(figure 1). The graph construction procedure comes at
no cost, as it can be inserted in the flow direction and
partitioning procedure that requires a complete pass
over each grid cell. Pointers to nodes which have no
incoming edges (corresponding to local maxima in the
mers a seamless view of the data layer and a trans-
parent communication layer (i.e. no explicit message
passing is required), but data can be defined as local
(i.e. near and reachable without communication costs)
or as global (i.e. potentially far and more expensive
to reach for improved efficiency. In particular, UPC
allows for operations such as the slope computation
discussed above to be simple to implement as well
as efficient: the global address space allows for the
transparent reference of an address outside of the local
domain (in this case the boundary in the neighboring
processor’s domain), while the partitioning ensures
that all memory access within a processor’s domain
is local. These characteristics make the language well
suited for our application, where most computations
are local but there is the need to obtain global infor-
mation, as in the case of drainage area. The Lawrence
Berkeley National Laboratory UPC compiler, used for
this application, is now supported under many operat-
ing systems and architectures, and is freely available
at http://upc.lbl.gov.
The parallel system used for testing and devel-
opment is named Franklin, the National Energy
Research Scientific Computing Center (NERSC)’s
38,288-CPU 2.3 GHz Opteron Cray XT-4 running
a Linux-based operating system. It consists of 9572
quadprocessor nodes, each with 8GB of memory, in-
terconnected with a high-speed SeaStar-2 network,
and sharing a Lustre Parallel File System (LPFS) with
436Tb of user disk space. Theoretical peak perform-
ance is 9.2 GFlops/second per core, or 352TFlop/sec-
ond for the whole system.
DRAINAGE AREA ALGORITHM
Generally, slope stability calculations, when im-
plemented on a grid or a mesh, operate on informa-
tion (such as topographic attributes) that is assigned
to each cell, and thus are trivial to parallelize by using
spatial domain decomposition of the (tiled) datasets
[w
ilkinson
& a
llen
, 1999]. Most of the attributes
that may be assigned to each grid cell (for example
topographic slope) are also local in nature, requiring
information only from pre-defined neighbourhoods of
the target cell. Drainage area to a point, a fundamental
landscape attribute, is defined as the total basin area
above a specific point from which flow can reach such
a point. The computation of drainage area is a glo-
bal operation, as information is needed from all other
Fig.1 - Idealized topography in 3-D view (a), and corre-
sponding directed graph in plan view (b). Nodes
correspond to grid cells and directed edges cor-
respond to flow to downhill neighbours. Edge
weights (illustrated here as edge thickness) corre-
spond to the proportion of flow from outgoing cell.
The upper-left node has no incoming edges and is
a starting point for the algorithm. The lower-right
node will be the last node to be processed.
Fig. 2 - Flow chart of initialization phase of the drainage
area algorithm
background image
SPATIALLY EXPLICIT SHALLOW LANDSLIDE SUSCEPTIBILITY MAPPING OVER LARGE AREAS
Italian Journal of Engineering Geology and Environment - Book www.ijege.uniroma1.it © 2011 Casa Editrice Università La Sapienza
403
tionship between a metric of rainfall and the fraction of
unstable grid cells predicted by Shalstab that actually
failed in historic storms (equation 1, Stock et al., this
volume). This relationship accounts for the effect of
rainfall intensity and duration on landslide abundance.
They used digital landslide inventories mapped from
air photographs taken after storms to construct storm-
specific landslide catalogs for two southern Califor-
nia sites (Sunland and Santa Paula) and one northern
California site (Montara). In the Santa Paula area they
isolated landslides triggered by storms that occurred in
1969, 1998, 2001, and 2005; in the Sunland area they
isolated landslides which occurred after storms in 1998
and 2005; in the Montara area they isolated landslides
which occurred in 1955, 1982, and 2005.
Nearby rain gages recorded hourly rainfall data
for all the storms of interest, with the exception of the
1955 and 1982 Montara storms (which were thus ex-
cluded from their study). Using the Santa Paula data as
a training dataset, they found that the fraction of unsta-
ble cells that actually failed increased as a power law
with the maximum 6-hour averaged rainfall intensity
(unstable fraction = 0.00001 * I
6-hr
2.7
). The 1998 and
2005 Sunland events fell nicely on this curve (figure
6, Stock et al., this volume), suggesting that such a
metric could be applied regionally. However, the 2005
Montara event in northern California had more than
an order of magnitude lower fraction of unstable cells
for similar 6-hour intensities. This points to the need
for further region-specific calibration is needed if one
wishes to apply such a method state wide.
STORM SIMULATION
Dettinger et al. [in press] simulated an atmospher-
ic river storm (Ar
k
Storm) combining data from two
of California’s most intense storms on record (Janu-
ary 1969 in southern California and February 1986 in
northern California) as initial conditions in a spatially
explicit dynamic meteorological model to describe a
rapid sequence of several major storms over the state,
yielding precipitation totals that go well beyond what
actually occurred during the two separate events. They
used a General Circulation Model (GCM) that depicts
the world’s climate over time at a coarse scale (hun-
dreds of kilometres), coupling it over California with
the state-of-the-art Weather Research and Forecast
(WRF) model. WRF down-scales weather in a nested
fashion down to a 2-km grid size, thus capturing the
elevation data), and the information they contain, are
pushed into queues belonging to the thread that owns
the downhill node (figure 2). The algorithm routes the
drainage area information across this graph, operating
only on nodes from the queues (i.e. nodes which have
received all upslope information).
Individual threads poll the queues belonging to
them, and remove the first item in the queue. The area
information contained is then distributed to all the re-
ceiving node’s downhill neighbours on the graph, ac-
cording to the chosen flow weighting scheme. When
information is sent across an edge of the graph, the
edge is deleted. If, as a consequence of an edge de-
letion, a node no longer has any incoming edges, it
becomes available for processing and it is placed onto
the appropriate queue (figure 3). The algorithm termi-
nates when the graph is fully disconnected, and all the
queues are empty.
This algorithm is general, in other words it does
not depend on the choice of flow partitioning. In our
application we implement the multi-direction slope-
dependent flow partitioning scheme, in which flow is
distributed to all neighbouring downhill cells propor-
tionately to the local slope [Q
uinn
, et alii, 1991].
STORM INFLUENCE ON LANDSLIDES
Stock et al. [this volume] found an empirical rela-
Fig. 3 - Flow chart of processing phase of the drainage
area algorithm
background image
D. BELLUGI ,w. E. DIETRICH, J. STOCk, J. MckEAN, B. kAZIAN & P. HARGROVE
404
5th International Conference on Debris-Flow Hazards Mitigation: Mechanics, Prediction and Assessment Padua, Italy - 14-17 June 2011
relevant orographic and wind controls on precipitation.
They provided hourly rainfall totals on a 2-km grid in
southern California, and a 6-km grid over the whole
state, effectively simulating the “perfect storm” sce-
nario for the state of California (figure 4).
An example of the seamless drainage area com-
putation resulting from the application of our parallel
drainage area algorithm is shown in figure 5. The re-
sults of the application of the parallel Shalstab to the
State of California are shown in figure 6. They are
classified based on their q/T value (eq. 3), the ratio of
effective steady-state precipitation and soil transmissiv-
ity required for instability. Two additional classes are
also shown, representing areas having too low
a gradient (unconditionally stable), and those
with too high a gradient (unconditionally un-
stable). Table 1 illustrates the number of square
kilometres and the percent area belonging to
each class, for the entire state, reflecting the un-
derlying distributions of steep and convergent
areas. The results indicate that landslide sus-
ceptibility is generally focused in these steep
convergent areas, and that while landslides
pose a significant hazard in the state of Califor-
nia, the susceptible areas are a relatively small
fraction of the entire landscape. For example,
using a log(q/T) threshold of -2.8, as suggested
by Dietrich et al., 2001 (for 10m data), 5.42%
of the landscape would be considered to be
susceptible to shallow landsliding. Assuming a
soil transmissivity value of 65 m
2
/day [Mont-
gomery and Dietrich, 1994], a log(q/T) thresh-
old of -2.8 threshold would be equivalent to a
critical rainfall rate of 103 mm/day.
The regression derived by Stock et al.
[this volume], relating the fraction of unstable
cells (calculated using our parallel Shalstab
run), which may actually fail under the ArkStorm sce-
nario, show that within the areas having similar lithol-
ogy as the Santa Paula training site, the abundance
of landslides under the ArkStorm scenario reflect the
abundance of failures observed after the extreme win-
ter storms of 1969.
All values of q/T above the “unconditionally sta-
ble” value were used by Stock et al. [this volume] to
Fig. 4 - Total accumulated precipitation during wRF
simulation of ArkStorm scenario, in 6-km state-
wide wRF nest (a), and 2-km southern Califor-
nia wRF nest (b). Notice that colour bars are not
all the same; the domain of panel (b) is indicated
by white rectangle in panel (a). Figure from Det-
tinger et al. [in press]
Fig. 5 - Detail showing the intersection of four drainage area data tiles
from a 1m LiDAR survey of the Eel River, CA (National Center
for Airborne Laser Mapping). Panel (a) shows sequentially
processed data tiles. The red arrows point to discontinuities
in drainage area results. Panel (b) shows the same tiles proc-
essed by our parallel algorithm with a seamless grid
Tab 1 - Distribution of Shalstab grid cells per stability
class
background image
SPATIALLY EXPLICIT SHALLOW LANDSLIDE SUSCEPTIBILITY MAPPING OVER LARGE AREAS
Italian Journal of Engineering Geology and Environment - Book www.ijege.uniroma1.it © 2011 Casa Editrice Università La Sapienza
405
The s
toCk
et alii [this volume] findings suggest
that a systematic effort be made across the state to
collect local rainfall data and map landslide scars for
specific storm events and then compute the proportion
of failed to predicted cells for a given area. Differ-
ent slope stability models could be used in such re-
gressions. This approach has the advantage over the
now widespread use of intensity-duration precipita-
tion threshold plots for landsliding (e.g. R
ossi
et alii,
2006; G
uzzetti
et alii, 2008; b
aum
& G
odt
, 2009) in
providing estimates of the number of landslides in an
area. Given the complex legacy of land use effects,
uncertain fire history, and the presently unknown in-
fluence of bedrock type (and its tectonic history) there
are good reasons to question whether this approach
can be successful in general. Nonetheless, such data
would prove invaluable in evaluating models and,
hopefully, improving hazard prediction. Some simple
steps can be taken to explore if improvements can be
made in shallow landslide prediction at the state level
using more mechanistic models. Data on soil thick-
ness, vegetation cover, soil material properties, and
geology (glaciated versus unglaciated) could be used
to parameterize models that include root strength.
Models that predict parameters such as soil depth (e.g.
d
ietRiCH
et alii, 1995) could exploit the parallelized
framework and possibly improve estimates of the spa-
tial structure of soil depth. Hydrologically dynamic
delineate areas likely to fail in the storm. However, ex-
perience with Shalstab suggests that, when using 10m
data, a log(q/T) threshold value of -2.8 will capture the
vast majority of shallow landslide scars [d
ietRiCH
et alii,
2001]. Figure 7 shows in red all the area of California
that would be potentially unstable using this threshold.
DISCUSSION
Despite its simplicity, the application of Shal-
stab over the entire state of California at 10 m cell
size required the development of a parallelized form
of the drainage area determination for each cell. As
topographic resolution increases further (from 10 m
to 1 m data spacing) through airborne laser mapping,
our ability to map landslides will greatly improve, but
to make spatially explicit landslide predictions over
large areas, parallelization of computation algorithms
will be necessary. The largest area that a modern desk-
top computer can load into memory ranges from 100
km
2
for 1 m data to 10,000 km
2
for 10 m data. This im-
plies that larger areas must be processed on a distrib-
uted memory architecture. Thus for seamless results
over areas such as California, the only alternatives to
parallelization are databases which are not practicable
in terms of speed. Nevertheless, our scaling tests sug-
gest that if we had 1 m data for all of California, even
using all of the 9572 nodes available on Franklin (one
of the world’s most powerful computers) would still
require several hours for Shalstab to run.
Fig. 7 - Map showing potentially unstable areas for the
state of California, as determined by a threshold
of log(q/T) -2.8. The inset map shows the Sunland
test area southern California
Fig. 6 - Map showing Shalstab for the state of California.
The inset map shows the Sunland test area south-
ern California
background image
D. BELLUGI ,w. E. DIETRICH, J. STOCk, J. MckEAN, B. kAZIAN & P. HARGROVE
406
5th International Conference on Debris-Flow Hazards Mitigation: Mechanics, Prediction and Assessment Padua, Italy - 14-17 June 2011
cells that failed for a given storm are needed to esti-
mate the magnitude of the landslide response. We an-
ticipate that higher resolution topographic data, more
mechanistic models, and increased computational effi-
ciency through parallelization algorithms will progres-
sively improve shallow landslide hazard predictions
ACKNOWLEDGEMENTS
This research was financially supported by a grant
from the United States Forest Service. This research
used resources of the National Energy Research Sci-
entific Computing Center, which is supported by the
Office of Science of the U.S. Department of Energy
under Contract No. DE-AC02-05CH11231. Further
technical support was offered by the Berkeley UPC
group. A special thanks to Collin Bode of the National
Center for Earth-surface Dynamics who navigated un-
documented ArcGis waters to help automate the ren-
dering of the many thousands of tiles. The authors are
grateful for a thorough review by Kevin Schmidt of
the United States Geological Survey, and for helpful
comments by two anonymous readers.
models which can route water and include the effects
of vegetation at the spatial scale of DEM’s (e.g. w
iG
-
mosta
et alii 2002) could similarly exploit the paral-
lelized framework and improve the spatial structure of
the pore pressure field.
CONCLUSION
Detailed precipitation predictions now available
at the large scale allow and invite the assessment of
the effects of extreme storm parallel version of Shal-
stab allow for fast seamless results even when applied
to large areas. We applied the algorithms to the state
of California (414,000 km
2
) using 10 m resolution
topographic data, but tests show that our application
could process in a few hours all of the United States
(9,830,000 km
2
) or all of Europe (10,180,000 km
2
)
with similar data resolution.
To contribute to the USGS ArkStorm emergency
preparedness study we ran Shalstab for the entire state.
The extent of landsliding for any given storm will be
greatly over-predicted by this model. Empirical rela-
tionships between the percent of predicted unstable
REFERENCES
a
ndeRs
a.m., R
oe
G.H., d
uRRan
d.R. & m
indeR
J.R. (2007) - Small- Scale Spatial Gradients in Climatological Precipitation
on the Olympic Peninsula. J. Hydrometeor, 8: 1068–1081.
b
aum
R.l. & G
odt
J.w. (2009) - Early warning of rainfall-induced shallow landslides and debris flows in the USA. Landslides.
doi:10.1007/s10346-009-0177-0
C
aRlson
w., d
RaPeR
J.m., C
ulleR
d.e., y
eliCk
k., b
Rooks
e. & w
aRRen
k. (1999) - Introduction to UPC and Language
Specification
C
asadei
m., d
ietRiCH
w. e. & m
illeR
n.l. (2003) - Testing a model for predicting the timing and location of shallow landslide
initiation in soil mantled landscapes. Earth Surface Processes and Landforms, 28: 925-950.
d
asGuPta
s., P
aPadimitRiou
C.H.,& v
aziRani
U.V. (2006) - Algorithms. McGraw-Hill Higher Education.
d
ettinGeR
m.d., R
alPH
f.m., H
uGHes
m., d
as
t., n
eiman
P., C
ox
d., e
stes
G., R
eynolds
d., H
aRtman
R., C
ayan
d. & J
ones
l. (in
press). Requirements and designs for a winter storm scenario for emergency preparedness and planning exercises in California.
d
ietRiCH
w .e., R
eiss
R., H
su
m. & m
ontGomeRy
d.R. (1995) - A process-based model for colluvial soil depth and shallow
landsliding using digital elevation data. Hydrological Processes, 9: 383-400.
d
ietRiCH
w.e., b
elluGi
d., & R
eal
de
a
sua
R. (2001) - Validation of the shallow landslide model, SHALSTAB, for forest
management, in M..S. wigmosta, and S. J. Burges, editors, Land Use and watersheds: Human influence on hydrology and
geomorphology in urban and forest areas, Amer. Geoph. Union, Water Science and Application, 2: 195-227.
f
osteR
I. (1995) - Designing and building parallel programs, Addison-Wesley.
G
ioRGi
f., H
ewitson
b., C
HRistensen
J., f
u
C., H
ulme
m., m
eaRns
l.,
von
s
toRCH
H., w
Hetton
P. et alii (2001) - Regional cli-
mate simulation - evaluation and projections. In Climate Change 2001: The Scientific Basis, H
ouGHton
J.t., d
inG
y., G
RiGGs
d.J., n
iGmeR
m.,
van
deR
l
inslen
P.J., d
ai
x. (
eds
). Cambridge University Press.
G
Rell
G.a., d
udHia
J. & s
tauffeR
d.R. (1995) - A description of the fifthgeneration Penn State NCAR Mesoscale Model (MM5).
NCAR Tech. Note TN-398 + STR. 122 pp.
G
uzzetti
f., P
eRuCCaCCi
s., R
ossi
m. & s
taRk
C.P. (2008) - The rainfall intensity–duration control of shallow landslides and
debris flows: an update. Landslides. 5(1): 3-17.
background image
SPATIALLY EXPLICIT SHALLOW LANDSLIDE SUSCEPTIBILITY MAPPING OVER LARGE AREAS
Italian Journal of Engineering Geology and Environment - Book www.ijege.uniroma1.it © 2011 Casa Editrice Università La Sapienza
407
m
iCHalakes
J., C
Hen
s., d
udHia
J., H
aRt
l., k
lemP
J., m
iddleCoff
J. & s
kamaRoCk
w. (2001) - Development of a next genera-
tion regional weather research and forecast model. Developments in Teracomputing: Proceedings of the Ninth ECMWF
Workshop on the Use of High Performance Computing in Meteorology, W. Zwieflhofer and N. Kreitz, Eds., World Scien-
tific, 269–276.
m
ontGomeRy
d.R. & d
ietRiCH
w.e. (1994) - A physically-based model for topographic control on shallow landsliding, Water
Resources Research, 30(4): 1153-1171.
Q
uinn
P., b
eaven
k., C
HevallieR
P. & P
lanCHon
o. (1991) - The prediction of hillslope flow for distributed hydrological mod-
eling using digital terrain models. Hydrological processes, 5: 59-79.
R
alPH
f.m., n
eiman
P.J., w
iCk
G.a., G
utman
s.i., d
ettinGeR
m.d., C
ayan
d.R. & w
Hite
a.b. (2006) - Flooding on California's
Russian River: Role of atmospheric rivers. Geophys. Res. Lett., 33, L13801, doi:10.1029/2006GL026689.
R
oe
G.H. (2005) - Orographic precipitation. Annual Review of Earth and Planetary Sciences, 33: 645–671.
R
ossi
m., P
eRuCCaCCi
s. & G
uzzetti
f. (2006) - A review of rainfall thresholds for the initiation of landslides. Geophys Res
Abstr. 8: 02323.
s
alatHé
e.P. (2003) - Comparison of various precipitation downscaling methods for the simulation of streamflow in a rain-
shadow river basin. International Journal of Climatology, 23: 887-901. doi: 10.1002/joc.922
s
toCk
, J.d. & b
elluGi
d. (this volume) - An empirical method to forecast the effect of storm intensity on shallow landslide
abundance. Proceedings of the 5th International Conference on Debris-Flow Hazards Mitigation: Mechanics, Prediction
and Assessment, Italian Journal of Engineering Geology and Environment.
w
idmann
m., C.s. b
RetHeRton
, e.P. s
alatHé
, 2003, Statistical Precipitation Downscaling over the Northwestern United States
Using Numerically Simulated Precipitation as a Predictor. J. Climate, 16:799–816.
w
iGmosta
m.s., n
iJssen
b., s
toRCk
P. & l
ettenmaieR
d.P. (2002) - The Distributed Hydrology Soil Vegetation Model, In-
Mathematical Models of Small watershed. Hydrology and Applications, s
inGH
v.P. & f
ReveRt
d.k.,
eds
., Water Resource
Publications, Littleton, CO.: 7-42.
w
ilkinson
b. & a
llen
C.m. (1999) - Parallel Programming, Prentice- Hall.
y
eliCk
k., b
onaCHea
d., C
Hen
w., C
olella
P., d
atta
k., d
uell
J., G
RaHam
s.l., H
aRGRove
P., H
ilfinGeR
P., H
usbands
P., i
anCu
C., k
amil
a., n
isHtala
R., s
u
J., w
elCome
m. & w
en
t. (2007) - Productivity and performance using partitioned global
address space languages. In Proceedings of the 2007 international Workshop on Parallel Symbolic Computation (London,
Ontario, Canada, July 27 - 28, 2007). PASCO '07. ACM, New York, NY, 24-32.
Statistics