Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/228578451
CITATIONS
READS
38
161
3 authors, including:
Jason Goetz
Alexander Brenning
11 PUBLICATIONS 75 CITATIONS
SEE PROFILE
SEE PROFILE
Geomorphology
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / g e o m o r p h
University of Waterloo, Department of Geography and Environmental Management, 200 University Avenue West, Waterloo, Ontario, Canada N2L 3G1
Hemmera, Suite 250 - 1380 Burrard Street, Vancouver, British Columbia, Canada V6Z 2H3
a r t i c l e
i n f o
Article history:
Received 2 March 2010
Received in revised form 20 January 2011
Accepted 4 March 2011
Available online 10 March 2011
Keywords:
Landslide susceptibility modeling
Generalized additive model
Logistic regression
SHALSTAB
Factor of Safety
Digital terrain analysis
a b s t r a c t
Physically based models are commonly used as an integral step in landslide hazard assessment. Geomorphic
principles can be applied to a broad area, resulting in rst order assessment of landslide susceptibility. New
techniques are now available that may result in the increased accuracy of such models. We investigate the
possibility to enhance landslide susceptibility modeling by integrating two physically-based landslide models,
the Factor of Safety (FS) and the Shallow Stability model (SHALSTAB), with traditional empiricalstatistical
methods that utilize terrain attribute information derived from a digital elevation model and land use
characteristics related to forest harvesting. The model performance is measured by the area under the receiver
operating characteristic curve (AUROC) and sensitivity at 90% and 80% specicity both estimated by bootstrap
resampling. Our study examines 278 landslide initiation points in the Klanawa Watershed located on
Vancouver Island, British Columbia, Canada. We use a generalized additive model (GAM) and a logistic
regression model (GLM) combining physical landslide models, terrain attributes and land use data, and GAMs
and GLMs using only subsets of these variables. In this study, all empirical and combined physicalempirical
models outperform the physically-based models, with GAMs often performing signicantly better than GLMs.
The strongest predictive performance is achieved by the GAMs using terrain attributes in combination with
land use data. Variables representing physically-based models do not signicantly improve the empirical
models, but they may allow for a better physical interpretation of empirical models. Also, based on bootstrap
variable-selection frequencies, land use data, FS, slope and plan/prole curvature are relatively the most
important predictor variables.
2011 Elsevier B.V. All rights reserved.
1. Introduction
Despite considerable advances in landslide hazard analysis and
risk management, landslides continue to present an acute threat to
lives and property in mountainous regions worldwide. It is estimated,
for example, that China, Japan and Nepal average over 150 deaths
annually from direct and indirect landslide hazards. Average annual
cumulative economic costs of landslides in Canada, the United States,
Japan and India are over four billion dollars (Sidle and Ochiai, 2006).
Changing climatic patterns, continuous deforestation in mountainous
regions and increased urbanization and development in landslide
susceptible regions have contributed to a global increase in landslide
activity and risk (Schuster, 1996). Improved landslide susceptibility
models and maps are required to reduce the impacts of landslides on
infrastructure and human populations.
It is our objective to improve landslide susceptibility models by
combining physically-based approaches with statistical and machine-
Corresponding author. Tel.: + 1 519 888 4567x35783; fax: + 1 519 746 0658.
E-mail address: jgoetz@uwaterloo.ca (J.N. Goetz).
0169-555X/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.geomorph.2011.03.001
377
2. Study area
Our study area is located in the Klanawa River watershed on the
southwestern coast of Vancouver Island, British Columbia, Canada
(Fig. 1). This site encompasses a total area of 610 km2 with 960 m of
relief. The lithology is composed of grano-dioritic rocks and calcalkaline volcanic rocks. Climate effects coupled with its rugged terrain
formed by Pleistocene glaciations makes the study area generally
prone to landsliding. The Klanawa River is located in a temperate
maritime climate with annual precipitation typically greater than
3000 mm (Guthrie et al., 2008). The forest cover is generally
comprised of western hemlock and western redcedar trees. There
have been extensive human activities in this area related to the forest
industry and clear-cutting practices. As of 2001 approximately 46% of
the study area has been logged (Guthrie et al., 2008). Landslides in
this area have been noticed to frequently occur in deforested areas,
which are commonly adjacent to logging roads. In general, landslides
occurring on Vancouver Island have been observed to occur in greater
spatial densities when adjacent to roads. In addition, there has been a
considerable increase in the number of landslides after three decades
of forestry activities (Guthrie, 2002).To exclude the low-lying valley
oor that is not susceptible to landsliding, we only consider terrain
above 150 m elevation, which is an area of 394 km2, in model
construction and assessment.
378
Fig. 2. Typical landslides on Vancouver Island, British Columbia. (A) Stereopair of landslides in a natural setting. Landslides usually begin as shallow translational failures and break
up, losing cohesion, as they move down-slope. (B) Landslides initiating from a concave slope into a gully system. (C) Landslides in a forested setting. (D) Landslide showing both the
distinct planar failure surface, and the complete disintegration of material downslope.
379
Fig. 3. Map of forest-harvesting related land use. The logged areas represent a mosaic of forest cuts from the years 1995 to 2002.
1
tan
sin 1
a
tan
380
density of water (s/w). Essentially, the lower the Q/T value, the more
susceptible to landsliding a location is.
Unconditionally unstable slopes are dened as the slope gradient
being equal to or greater than the friction angle:
tan tan
a
C + cos1 min TR sin
; 1 = r tan
sin
Fig. 4. Parameter optimization of r and in the SHALSTAB model using the entire study
area for illustration. The median optimal values, based on the 100 bootstrap replications
are = 40.6 (std. dev. 3.1) and r = 1.78 (std. dev. 0.17).
Fig. 5. Parameter optimization of T/R using the entire study area for illustration. (A) An
illustration of the effects of a range of T/R values on model performance of FS and PT-GAM
at different cohesions. The optimal T/R value, based on the entire study area, is 214 m with
an FS performance of 73.6% AUROC. The median optimal T/R value, based on the 100
bootstrap replication, is 92 m with a median FS performance of 71.6% AUROC. (B) An
illustration of the effects of a range of T/R values on the correlations between slope and log
catchment area to log FS at different cohesions.
381
382
Table 1
Descriptive statistics of the morphometric and physical model predictor variables used for modeling landslide susceptibility.
Predictor variable
Non-landslide points:
Median (std. dev)
Landslide points:
Median (std. dev)
AUROC (%)
Study area
AUROC (%)
Bootstrap test set (std. dev)
log FSa
Slope (degrees)
log(Q/T)a
Catchment slope
Elevation (m)
Distance to road (m)
TWI
Plan curvature
Prole curvature
log catchment area
0.06
23.0
2.5
23.9
371
354
6.0
0.001
0.000
3.6
0.22
31.5
3.0
27.7
472
112
5.7
0.000
0.001
3.7
73.6
73.0
71.1
67.0
61.7
62.7
58.5
57.6
54.6
52.7
71.9
72.1
68.9
65.0
62.9
61.9
60.1
57.4
52.4
50.2
(0.31)
(10.3)
(0.8)
(7.7)
(183)
(840)
(1.5)
(0.006)
(0.007)
(0.5)
(0.15)
(7.3)
(1.0)
(5.5)
(152)
(346)
(0.9)
(0.011)
(0.008)
(0.4)
(2.2)
(2.1)
(2.6)
(2.4)
(2.2)
(2.0)
(2.3)
(2.4)
(3.6)
(2.4)
5. Discussion
5.1. Model interpretation
Landslides are typically more prone to occur in steep convergent
areas (Montgomery and Dietrich, 1994). These curvature conditions
force soil water to converge at the soilbedrock contact or where the
soil meets an underlying impermeable layer (Wilson and Dietrich,
1987).After heavy rainstorms or long periods of rain, upwardly
concave slopes can hold more water for a longer period of time (Lee
and Min, 2001). A combination of antecedent rainfall conditions and
rainstorm or rapid snowmelt can result in an increase of pore water
pressure, which can lead to hillslopes becoming more susceptible to
failure (Talebi et al., 2008).
The empirical model results indicate that hillslopes with concave
prole and convergent plan curvature tend to have increased
Table 2
Model performance of GAM, GLM, and physically-based models (SHALSTAB and FS) estimated using the bootstrap and the training set (median value and standard deviation). The
median variable frequency represents the average number of variables included in each model for the bootstrap.
Model
LPT-GAM
LPT-GLM
LT-GAM
LT-GLM
PT-GAM
PT-GLM
T-GAM
T-GLM
FS
SHALSTAB
Bootstrap
Study area
AUROC
(%)
Sensitivity (%) at
90% specicity
Sensitivity (%) at
80% specicity
Mean variable
frequency
AUROC
(%)
Sensitivity (%) at
90% specicity
Sensitivity (%) at
80% specicity
80.8
80.3
80.8
80.3
74.9
73.7
74.9
73.7
71.9
68.9
47.7
45.6
48.4
46.2
31.6
29.1
31.0
28.2
19.5
19.0
63.4
62.7
63.8
62.4
48.3
45.6
47.7
46.7
41.8
38.1
6.1
6.0
5.9
5.8
4.3
4.4
3.9
4.1
83.4
83.3
83.8
83.2
77.7
77.1
77.4
76.6
73.6
71.1
56.4
54.4
56.4
56.8
35.8
35.8
36.7
34.5
24.0
24.4
69.7
69.7
70.4
67.2
54.4
51.2
54.0
51.2
47.0
42.2
(2.0)
(2.1)
(2.0)
(2.0)
(2.2)
(2.2)
(2.2)
(2.2)
(2.2)
(2.6)
(5.4)
(5.8)
(5.4)
(5.7)
(5.4)
(5.5)
(5.0)
(5.4)
(5.0)
(10.7)
(4.6)
(5.1)
(4.5)
(5.0)
(5.4)
(5.1)
(5.5)
(5.3)
(6.7)
(6.9)
(0.9)
(0.9)
(1.0)
(1.1)
(1.0)
(1.1)
(1.1)
(1.2)
383
Fig. 6. Diagram of model performance and signicance. The performance of each empirical model has been mapped in a diagram where the arrow points to a model that had a lesser
performance; a solid line indicates that there is no difference, statistical or otherwise, between model performance. The numbers adjacent to the arrows indicate the percent
difference in performance between comparing models. Additionally, the statistical signicance of model differences are indicated using signicance codes following the percent
difference.
or deposition. Prole curvature characterizes the subsurface acceleration or deceleration of ow down a slope, which in turn is related
to potential erosion or deposition rates and consequently spatially
Fig. 7. Landslide susceptibility maps for physically-based (FS),and combined models (LPT-GAM and PT-GAM) applied to the subarea shown in Fig. 1. The models are trained on the
entire study area. The area designated as Not applicable is a portion of the study area below the models' 150 m threshold.
384
Fig. 8. Transformation of predictor variables in the generalized additive models, for PT-GAM (A) and T-GAM (B), that utilize the entire study area as training sample. A spline function
for non-parametric smoothing of the variables, s(variable), indicates a nonlinear transformation. The dotted lines represent condence bands. In the plan curvature plot, a negative
value indicates a convergent surface, a positive value indicates a divergent surface, and value of 0 indicates the plan is straight. In the prole curvature plot, a negative value indicates
a convex surface, a positive value indicates a concave surface, and a value of 0 indicates no surface curvature.
R a
1
T sin
In this case, T/R and a will have no inuence on FS, and only
constants such as C, r and and the spatially variable slope angle will,
resulting in a monotonically decreasing nonlinear function of .
The reverse situation occurs when the specic catchment area (or,
on planar slopes, slope length) is smaller than some threshold
T
sin
R
Table 3
Variable-selection frequencies and percentage of nonlinear occurrence for GAM models
on 100 bootstrap training samples.
Variable
LPT-GAM
LT-GAM
PT-GAM
T-GAM
Distance to road
Logging
Plan curvature
logFS
Prole curvature
Elevation
Slope
log(Q/T)
TWI
log catchment area
Catchment slope
100 (44%)
100
99 (56%)
99 (52%)
93 (75%)
37 (100%)
22 (32%)
21 (33%)
19 (32%)
10 (30%)
9 (33%)
100
100
100
94
39
100
27
16
12
100
98
81
27
32
26
21
23
12
99 (66%)
93 (87%)
25 (100%)
99 (87%)
38 (37%)
21 (14%)
17 (29%)
(44%)
(50%)
(77%)
(91%)
(90%)
(26%)
(25%)
(33%)
(69%)
(52%)
(100%)
(100%)
(31%)
(38%)
(52%)
(22%)
(25%)
385
386
for different plan curvatures in the upslope contributing area indirectly through the specic catchment factor, however no information
on the slope prole is incorporated into these models. Therefore, the
inclusion of an empirical variable for prole curvature is important to
represent the potential range of ow and inltration characteristics
for different prole curvatures that the physically-based landslides
susceptibility models may not account for.
Overall, the methods implemented in this study, which combine
empirical and physically-based approaches and include bias-reduced
error estimation, were presented as a general framework to enhance
the analysis of performance for landslide susceptibility models. In
addition, the use of a nonlinear regression technique, such as the
GAM, demonstrated the importance of representing the nonlinear
relationships of predictor variables of landslide occurrence, which
allows for a more exible and interpretable analysis of landslide
susceptibility.
Acknowledgements
This research was funded through an NSERC Discovery Grant
Individual awarded to A. Brenning. We acknowledge constructive
comments provided by the anonymous referees.
References
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression for
landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan.
Geomorphology 65, 1231.
Barnett, P.J., Singhroy, V.H., Shirota, J., Leney, S.J., 2004. Methods for remote engineering
geology and terrain analysis in boreal forest regions of Ontario, Canada. The
Geological Society of America 10, 229241.
Beguera, S., 2006. Validation and evaluation of predictive models in hazard assessment
and risk management. Natural Hazards 37, 315329.
Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of
basin hydrology. Hydrology Science Bulletin 24, 4369.
Bivand, R.S., 2000. Using the R statistical data analysis language on GRASS 5.0 GIS
database les. Computers & Geosciences 26, 10431052.
Brenning, A., 2005. Spatial prediction models for landslide hazards: review, comparison
and evaluation. Natural Hazards and Earth System Sciences 5, 853862.
Brenning, A., 2008. Statistical geocomputing combining R and SAGA: the example of
landslide susceptibility analysis with generalized additive models. In: Bhner, J.,
Blaschke, T., Montanarella, L. (Eds.), SAGA Seconds Out (= Hamburger Beitrge
zur Physischen Geographie und Landschaftskologie, 19), pp. 2332.
Brenning, A., 2009. Benchmarking classiers to optimally integrate terrain analysis and
multispectral remote sensing in automatic rock glacier detection. Remote Sensing
of Environment 113, 239247.
Brenning, A., Grasser, M., Friend, D., 2007. Statistical estimation and generalized
additive modeling of rock glacier distribution in the San Juan Mountains, Colorado,
USA. Journal of Geophysical Research 112. doi:10.1029/2006JF000528.
Chang, K.T., Chiang, S.H., 2009. An integrated model for predicting rainfall-induced
landslides. Geomorphology 105, 366373.
Chung, C.F., Fabbri, A.G., 2003. Validation of spatial prediction models for landslide
hazard mapping. Natural Hazards 30, 451472.
Conrad, O., 2006. SAGA program structure and current state of implementation. In:
Bhner, J., McCloy, K.R., Strobl, J. (Eds.), SAGA Analysis and Modelling Applications
(= GttingerGeographischeAbhandlungen, 115), pp. 3952.
Dai, F.C., Lee, C.F., Ngai, Y.Y., 2002. Landslide risk assessment and management: an
overview. Engineering Geology 64, 6487.
Dietrich, W.E., Bullugi, D., Real de Asua, R., 2001. Validation of shallow landslide model,
SHALSTAB, for forest management. In: Wigmosta, M.S., Burges, S.J. (Eds.), Land Use
and Watersheds, Human Inuences on Hydrology and Geomorphology in Urban
Forest Areas. Water Science and Application 2. American Geophysical Union,
Washington, DC, pp. 195227.
Efron, B., Tibshirani, R., 1993. An Introduction to the Bootstrap. Chapman and Hall.
Frattini, P., Crosta, G., Carrara, A., 2010. Techniques for evaluating the performance of
landslide susceptibility models. Engineering Geology 111, 6272.
Glade, T., Anderson, M., Crozier, M.J., 2005. Landslide Hazard and Risk. John Wiley &
Sons, Chichester.
Gomes, R., Guimares, R.F., Osmar, A., Carvalho Jr., O.A., Fernandes, N.F., Vargas Jr., E.A.,
Martins, E.S., 2008. Identication of the affected areas by mass movement through