Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

GEOPHYSICALRESEARCHLETTERS,VOL. 28, NO.

6, PAGES 1091-1094,MARCH 15,2001

The characterization of landslide size distributions

Colin P. Stark
Lamont-Doherty Earth Observatory of Columbia University, Palisades, New York

Niels Hovius
Department of Earth Sciences, University of Cambridge, Cambridge, UK

Abstract. Landslide size distributions generally exhibit imposed on another fixed by the geomorphic system itself,
power-law scaling over a limited scale range. The range is through properties suchas regolith depth, root strength, and
set by the mapping resolution, by the number of observed mean relief. One persistent question has been whether we
events, and by the slope failure processitself. This property can distinguish these natural length scales from those im-
of self-similarity is an important insight into the physics of posed by the mapping methods. Another uncertainty has
hillslope failure. Typically, however, a large proportion of been whether the limitations of mapping schemes bias the
the landslide data does not fit a simple power law. These estimates of the power-law scaling exponents. These ques-
data are always ignored in order to characterize the scaling.
tions must be answeredif we are to be able to (i) predict
We show that landslide data sets from New Zealand and
extremeevents[Castillo,1988]and assesslandslidehazard
Taiwan exhibit two scalingregimes, separated by a crossover with any accuracy;(ii) comparedata sets;(iii) estimaterates
scalethat is purely an artefact of mapping resolution. Below of processessuch as erosion and montane forest disturbance;
this scale the landslide data are undersampled. We propose (iv) run simulationsof systemsdominated by landslides.
a general model for the size distribution of observedland- Our study addressesthese issues.
slideswhich can account for the whole population of mapped
slope failures. The model quantifies the undersampling of Characterizing landslide size
smaller landslides and provides an improved estimation of distributions
the power-law scaling of larger landslides. Estimates of this
scaling suggestthat the area disturbed by landsliding, and There is now a wealth of evidence that landslide popu-
perhaps the landslide sediment yield, are essentially depen- lations exhibit some power-law scaling across a portion of
dent on the frequency of smaller landslides. Higher reso- their size distributions. This notion is now so well estab-

lution landslide maps will be required in order to quantify lished that the non-scaling portions of the observed data
these fluxes. Our results also indicate that the probability of are generally ignored. In most cases, the bulk of an ob-
extreme landslide events is lessthan previous studies would served distribution must be disregarded. For example, fig-
predict. ure I showstypical landslidesizedata, from Taiwan [Hovius
et al., 2000], plotted as probabilitydensityp(x) for sizex in
Introduction both log-linear and log-log forms. The region of the distri-
bution consideredto exhibit power-law scaling, and to which
Landsliding is the most important geomorphic processin a power-law model can be fit with adequate precision, de-
humid mountains; it can also be the most dangerous. Since scribesonly 25% of the observations,as the log-linear graph
the size-frequencydistribution of landslides determines the shows. Three quarters of the observed data are not fit by a
hazard posed by such slope failure, it is vitally important power-law model in this case, and this caseis typical: log-log
that we quantify this distribution accurately. The size dis- plots hide the inadequacy of a pure power-law model distri-
tribution is also of fundamental importance if we are to bution. The problem we wish to address is therefore: what
estimate integrated effects such as the extent of landslide simple model can we use to describeas much of the observed
disturbance and the yield of eroded material. data as possible? We will then establish whether existing es-
Many recent studieshave shownthat landslide size distri- timates of power-law scaling need to be revised. Before we
butions exhibit power-law scaling. By this we mean that the begin, some important points need to be addressed:
number of observed landslides is a negative power function
of the landslide size, across some size range. This infer- (i) A distributionn(x)dx of landslidesof area x may be
ence holds true whether we define landslide size as the scar expressedusing a probability density function p(x)
area [Hoviuset al., 2000] or the total area disturbed[Pel- and a total landslide population N where,
letier et al., 1997],and whetherthe landslidesare triggered n(x) dx - Np(x) dx. (1)
over a long period of time [Hoviuset al., 1997]or almost
instantaneously [Harp and Jibson,1996;Keefer,1994]. The (ii) If the landslide sizes were perfectly power-law dis-
range of observedpower-law scaling, however, is always lim- tributed, their probability density function would be a
ited by the extent of the landslide data. This scale range is Pareto distribution(equation2). This study assesses
the effect on a Pareto distribution of undersampling
the smaller landslides, which is a process known as
Copyright2001 by the AmericanGeophysicalUnion.
censoring.
Papernumber2000GL008527. (iii) For a sizedistribution to be expressedas a probability
0094-8276/01/2000GL008527505.00 density, that density must integrate to unity. There-

1091
1092 STARK & HOVIUS: THE CHARACTERIZATION OF LANDSLIDE SIZE DISTRIBUTIONS

ii I I i i'!•111• i i i iiiii I i i i iiiii I i i


fore any pure power-law model distribution must have 10
a cutoff scale (either real or artificial). Negative • -

power-law,Pareto distributions(a • 0) require a


lowerboundc (equation2). i,/./,,,,,. \
_
-

(iv) A sizedistributionmay be estimatedby linearor loga-


rithmic binning, or other methods such as kernel den- _ , _

sity estimation. The log binning method, however,


scalesthe estimatedprobabilitydensityp(x) by a fac-
tor x, since it implicitly log transforms the distribu- :: .
tion. i -- Taiwan
data
• Double
Pareto
fit
(v) Representingthe size distribution as a cumulative : .: = Simulated
data
probability
P(X _• x)- f•ø•p(z)dz
ratherthan -

-
,

-- - Newpowerlaw
a probabilitydensityp(x) is not advisable[Pelletier
10
et al., 1997]. This is because:(a) the residualsin es-
timatesof P(x) are strictlyone-sidedand asymmetri- _----
-
':.-. Old
power
law
'I \/•F-
cally distributed, biasing any regressionfit which as-
10 i i i i i i ill , , ,,,,,,I , , ,,,,,,I ,\•
102 103 104 105
Area,x [m2]
5 Str
nog•
Figure 2. A test of the doublePareto model: an observeddistri-
x
tun•lea•a?rn••dj••
resolution? bution, from the Central Range of Taiwan, is compared with both
a model fit and a synthetic data set generated from that model fit.
Note the ambiguityof p(x) for large events,both in the real and
synthetic estimated densities, which induces an underestimate of
the asymptotic,pure power-lawscalingexponent(c• • 0.7). The
õ • / !1• •t Distorted
power
law - double power-law model fits a much greater range of observations

F / •• • (WeT
undersampling?)
la
_ and is therefore less biased by the ambiguity in the tail density.

I•. r"' / ]•
L ' /
k/ P/r

:daJelr-!
/ ///
•1•• z/
Power law tail

/ sampled-
/ data _
The simulated landslide data drew N -- 1000 samples from a
doublePareto fit to the Taiwan data set (N -- 1086), whosepa-
rameterswere: c•- 1.112,• - 3.217,t -- 519,c - 1,m -- 106.

ol--,,/,,,,
........
, ..... ..... ....
;
101 102 103 104 105 106
Area,x [m2]

'-.. •o---o
Observed
data I Fitrange
(Whataroa)
Newfitrange
(WSA)
'". ;.... Pure
powerlaw
•• 102 • • •- •'L"
•4-• -- Double Pareto fit
• • 'K,N,
--'Actual?
distribution
i
.............................................
104
• • ,:

10
s

•Jgure
1.A•example 1 lO
c=1o
]•ds]•de2 lO
size 3 lO
4 lO
5m:1O
6
dis[Hbu[io•,
from[heCe•- 10

versus]o•(•) forms. •hese •phs i]]us[•[e • fu•d•me•[•] p•ob- 10


lem with treating landslide size distributions: the bulk of the 102 103 104 105
observed pdf is at the rollover x - t and does not scale as a
power-law. The portion of the data that does scale as a power- Area,
x [m2]
law is just the tail of the distribution - in this case about a quarter
of the observedlandslides.Graph (b) also illustratesparameters Figure 3. The NZ data and doublePareto modelfits: (i) WSA
used in the double Pareto model (ch •, t, and cutoffsc, m); in ßN -- 5086, c• -- 1.442,• --- 4.863,t -- 1781; (ii) Whataroa: N --
both graphs the best fit double Pareto model is shown as a solid 3986, c•- 1.480,•-- 3.735, t-- 797. In both cases, c-- 1, m--
curve. 106 .
STARK &: HOVIUS' THE CHARACTERIZATION OF LANDSLIDE SIZE DISTRIBUTIONS 1093

sumesnormally distributederrors; (b) any crossover landsliding to other hillslope processes,or it could simply
from a non-power law to a power law is hidden inte- reflect the under-counting of landslides that are too small
gration smoothing. to be mapped accurately(we demonstratebelow that this
First, we address the use of the Pareto probability density
is indeed the casefor our New Zealand data). Whatever
the reason, we need to model the whole observed distribu-
in modeling the tails of landslide distributions. Second, we
tion in order to estimate as robustly as possiblethe asymp-
derive a modified Pareto distribution which models explic-
totic, negative power-law tail. We introduce here the double
itly the undersampling of small landslides, thus providing
Pareto distribution to serve this purpose.
an excellent description of landslide distributions acrossthe
First, we consider a mapping of our uniform sampler •r
whole observed size range.
to a form which switchesbetween scaling regimes. A simple
Pareto probability distribution way to do this is to map •r through a step function,

For an idealized landslide size distribution to be power


law or Paretodistributedacrossthe rangex E [c,c•) the size ={1[l+tanh/•)l}v (6)

probability density is defined as, where a and -),determine the steepnessof the step. This step
function is a scaled logistic distribution. By then stretching
p(x) • c•c•x-•-• c > 0, c•> 0 (2) y using
wherea is the power-lawscalingexponent.(e.g. figurelb). x = exp(y) (7)
Typical scaling analyses of landslide populations have ei- we obtain a mapping for •r -• x which will yield a probability
ther directly[Hoviuset al., 1997,2000]or indirectly[Keefer, densityof p(x) with a doublescalingregimeand a crossover
1994;Pelletier et al., 1997]set c sothat only the tails of the length scale at x = 1, where
observedsize distributions were characterized as power-law.
= I1+ x (8)
Simulating a sequence of landslides
A more generalform for •r(x) can be written so that'
A useful application of a landslide size distribution ex- (i) x spansthe range[c,m];
pressedas an probability density is in the stochastic model-
(ii) p(x) is negativepower-lawdistributedfor t << x •_ m
ing of geomorphic systems, where the realistic simulation
with exponent
of mass-wasting processesremains a signifcant challenge
(iii) p(x) is positivepower-lawdistributedfor c _• x <• t
[Bendaand Dunne,1997]. In orderto simulatea landslide
with exponent
sequencewhose idealized size distribution is a pure power-
law, we simply draw samples from a Pareto distribution. (iv) p(x) has a crossover
scaleat x - t.
Such samples are difficult to compute directly. In prac- In this case the sampling distribution takes the form,
tice, we generate a sequenceof uniformly distributed ran-
dom numbers and transform these samples in such a way
[Presset al., 1988]that the transformedsequence
distributed.
is Pareto vr(x)
=[l•+
+(m/t)-•
• ]•/•
(x/t)_ (9)
which gives the transformation
The probability•r(x) of an eventX •_ x is definedas,
=t [1+ - (lo)
P{X _• x} = •r(x) -- p(z) dz (3) The probability density of x then takes a censored,or double
Pareto form. This model considers the actual landslide sizes
=• x(•r)= c(1- •r)-t/• (4)
to be negative power-law distributed, but assumesthat the
Notice that •r(x) is the cumulativedistributionof x, and observedlandslide population undersamplesthis distribution
thereforeit spans•r E [0,1]. We sample•r from a uniform in a positive quasi-power-law fashion below a mapping res-
olution scale.
distribution,
p(•r) -0 U[0, 1) (5) undersampling - censoring
and transform •r -• x using equation 4. This yields a se-
quenceof randomsamples{x•} which are suitablyPareto
distributed acrossthe range x • [c,c•) (e.g. figure 1).
We can checkthis by applyingthe transformationp(x) =
p(x) + i ^ i pøwer
= z/ [1d-(m/t)-•']
•/• x xlt)

law
=Pare
W•[P(•r)
ß
Double Pareto probability distribution
wherer/--t(1- 5) and 5- vr(c).
Simulating a sequence of landslide observations
Experiencewith mapped landslidedata sets (figures2
and 3) indicatesthat a observedsize distributionsexhibit The above equations make it straightforward to generate
two scaling regimes: for large landslides we observe power- a sequence
(xj • of simulatedlandslidesizeobservations
that
law scaling with a negative scaling exponent, whereas for are double Pareto distributed. Random samples are drawn
small landslides we observe an apparent positive power-law from the uniformdistributionp(•r),
scaling. A peak in the distribution separates these two scal-
p(•r) •. U[5,1] (12)
ing regimes and therefore reflects a crossoverlength scale.
This crossovercould be set either by the mapping resolution and transformed to landslide size x using equation 10. These
or by the landslide processitself. In other words, the distri- samples(x•) are distributedas required(equation11; e.g.
bution towards small events could be a real crossover from figure 2).
1094 STARK & HOVIUS: THE CHARACTERIZATION OF LANDSLIDE SIZE DISTRIBUTIONS

Data analysis to the tail of the observed distribution, using only 25-50%
of the data, we can employ 95% of the data and fit the
Southern Alps, New Zealand whole observed distribution. Estimates of the asymptotic
We have applied the double Pareto approach to the anal- power-law scaling are much more robust as a result. This
ysis of landslide data acquired from the western Southern will facilitate more reliable predictions of extreme events.
Alps, New Zealand (figure 3). Two data sets were used; We applied this technique to landslide data from New
both were acquired from the same set of air photographs, Zealand and Taiwan, and found that the power-law decay
obtainedbetween1948and 1985: (i) a map spanningmuch of the size distributions was underestimated in both cases.
of the westernAlps (labelledWSA); (ii) a finer resolution Our revised estimates of power-law scaling suggest that the
map covering Whataroa catchment only. Both data sets total area of landslide disturbance is set by the number of
were obtainedusingthe samemappingtechniques[Hofius smallest landslides. •Ve also determined that these small
et al., 1997] by the sameperson(N.H.); they differedonly landslides are not counted reliably using current mapping
in the self-imposed level of detail. methods, and that we will need to improve mapping res-
Our results show that that the crossover scale is indeed olution if we are to assess bulk effects such as the rate of
a mapping artefact. The modeled threshold scaleswere t = forest disturbance. The bulk sediment flux, however, is less
797m2 and i - 1781m2 for the Whataroa and WSA data dependent on the frequency of small landsliding, and can in
sets respectively,which suggestmapping resolutionsof Lc m some casesbe estimated from existing data.
28 m (Whataroa) and Lc m 42 m (WSA). Landslidecounts
at or below these resolutionswere undersampled, so the true Acknowledgments. This researchis a contribution to
size distribution was censored. the NASA Solid Earth and Natural Hazards Program, and to
Our analyses also indicate that the numbers of mapped the NASA Seasonal-to-Interannual Prediction Project at God-
dard SpaceFlight Center, supported by NASA's Global Modeling
events in the scale range previously used for scaling esti-
and Analysis Program under RTOP 622-24-47. This study was
mation were contaminated by this censoring, even though funded in part by a LDEO research fellowship. We would like
the chosen range was somewhat above the crossoverresolu- to thank Ph. Davy and J. K. Weissel for their advice. LDEO
tion scale. The previous estimate of the scaling exponent contribution no. 6092.

c• m 1.15 (WSA) thereforeneedsto be adjusted. Our dou-


ble Pareto model fits, which account for the censoringbias, References
indicate steeper scaling, with c• • 1.44 - 1.48. Benda, L., and T. Dunne, Stochastic forcing of sediment supply
The scaling exponent of the underlying landslide size dis- to channel networks from landsliding and debris flow, Water
tribution explicitly determines the impact of large versus Resour. Res., 33, 28z19-2863, 1997.
small landslides on integrated measures such as the total Castillo, E., Extreme value theory in engineering, Academic
area disturbed, or the volume of material yielded. Our re- Press, San Diego, 1988.
Harp, E. L., and R. W. Jibson, Landslides triggered by the 1994
vised estimates of around c• m 1.5 suggest that the distur-
Northridge, California, earthquake, Bull. Seisin. $oc. Amer.,
bance area is strongly controlled by the area of the smallest 86, S319-S332, 1996.
landslides, whereas the volume eroded by slope failure is Hovius, N., C. P. Stark, and P. A. Allen, Sediment flux from
influenced uniformly by landslides of all sizes. a mountain belt derived by landslide mapping, Geology, 25,
231-234, 1997.
Central Range, Taiwan Hovius, N., C. P. Stark, H.-T. Chu, and J.-C. Lin, Supply and
removal of sediment in a landslide-dominated mountain belt:
In the Central Range of Taiwan the role of larger land- Central Range, Taiwan, J. Geol., 108, 73-89, 2000.
slidesis more significant(figure 2). Our new scalingesti- Keefer, D. K., The importance of earthquake-induced landslides
mate is c• • 1.11, as compared to a previous estimate of to long-term slope erosion and slope-failure hazards in seismi-
c• • 0.7 [Hoviuset al., 2000]. So, the total area disturbed cally active regions, Geomorphology, 10, 265-284, 1994.
Pelletier, J. D., B. D. Malamud, T. A. Blodgett, and D. L. Tur-
by landsliding is slightly dominated by small scale slope fail-
cotte, Scale-invariance of soil moisture variability and its im-
ure, whereas the total volume flux is set by the largest land- plications for the frequency-sizedistribution of landslides, Eng.
slides. Therefore, both the upper and lower length scalesof Geol., 48, 254-268, 1997.
landsliding must be obtained if we are to assessthe bulk con- Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vet-
sequencesof landsliding in the Central Range. At present terling, Numerical Recipes in C, Cambridge University Press,
these parameters cannot be measured. Cambridge, 1988.

Conclusions C. P. Stark, Lamont-Doherty Earth Observatory of Columbia


University, 61 Route 9W, Palisades, NY 10964, USA. (email:
We have shown that mapped landslide size distributions cstark@ldeo.columbia.edu)
are strongly influenced by the map resolution and by the N. Hovius,Department of EarthSciences, Downing Street,
undersampling of smaller landslides. We have demonstrated University of Cambridge, Cambridge, UK. (email: nhovius
@esc.cam.ac.uk)
that the observeddistributions can be modeled in full using
a double Pareto probability density function, which allows
us to assessthe effective resolution of the landslide mapping
and to treat the distortion of the asymptotic scaling caused (Received March 3, 2000; revised June 15, 2000;
by this artefact. Instead of having to restrict our model fit acceptedJune 23, 2000.)

You might also like