NG A
A
u.tw
ons and
for the
for the
ognized
blished.
datasets
ner is a
yretation
e image,
3. Then,
result is
in assist
methods,
\, 1991).
forrential
ce of the
à all the
standable
ness of a
actice to
in expert
| purpose
tifies the
»pagation
'orldwide
;e serious
portation,
ses, over-
itting and
landslide
an areas
ny urban
>s of the
ghroy &
aracterize
rticularly
h spatial
ventional
approach
’AIWAN.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Satellite imagery can also be used to collect data on the relevant
parameters involved such as soils, geology, slope,
geomorphology, land use, hydrology, rainfall, faults, etc.
Multispectral images are used for the classification of lithology,
vegetation, and land use, Stereo SPOT imagery is used in geo-
morphological mapping or terrain classification (Liu, 1987; Wu
et al., 1989; Liang, 1997; Liu, 1999; Hsu, 2002).
For landslide inventory mapping the size of the landslide
features in relation to the ground resolution of the remote
sensing data is very important. A typical landslide of 40000 m?,
for example, corresponds with 20x20 pixels on a SPOT Pan
image and 10*10 pixels on SPOT multi-spectral images. This
would be sufficient to identify a landslide that has a high
contrast, with respect to its surroundings e.g. bare scarps within
vegetated terrain, but it is insufficient for a proper analysis of
the elements pertaining to the failure to establish characteristics
and type of landslide. Imagery with sufficient spatial resolution
and stereo capability such as SPOT or IRS can be used to make
a general inventory of the past landslides. However, they are
mostly not sufficiently detailed to map out all landslides (Hsiao
et al, 2003). It is expected that in the future the Very High
Resolution (VHR) imagery, such as from IKONOS-2, might be
used successfully for landslide inventory (Westen, 2000). By
using the criteria for visual interpretation, artificial intelligent of
expert system and automatic procedures can be developed to
improve the efficiency and accuracy of landslide mapping
(Kojima et al., 2000, Liu et al., 2001).
Artificial Neural Networks (ANNs) have been used successfully
in many applications such as pattern recognition, function
approximation, optimization, forecasting, data retrieval, and
automatic control (Robert, 1990, Zurada, 1992). ANNs have
been found to be powerful and versatile computational tools for
orgagizing and correlating information in ways that have
proved useful for solving certain types of problems too complex,
too poorly understood, or too resource-intensive to tackle using
more traditional computational methods.
2. METHODOLOGY
2.1 Traditional landslide interpretation methods
Individual landslides are generally small and located in certain
locations of a slope. Landslides occur in a large variety,
depending on the type of movement such as (slide, topple, flow,
fall, spread), the speed of movement (mm/year-m/sec), the
material involved (rock, debris, soil), and the triggering
mechanism (earthquake, rainfall, human interaction). Survey
methods usually include ground survey, aerial or space-borne
survey, or a combination.
Ground survey can be high accurate, but slow. When hazards
take places, accessibility is low. Therefore, it is impossible to
make the survey in near real-time or in a complete coverage
after a torrential rainfall.
Photographic or image interpretation approach can be adopted
and implemented manually, automatically, or semi-
automatically. Manual interpretation requires well-trained
geologist to delineate the landslides under a stereoscopic
environment. The advantage of this approach is that individual
landslide can be defined very clearly. However, the subject
Judgement is the disadvantage. Automatic classification of
landslides is based on certain criteria and computing algorithms.
The advantage for image classification is the objectiveness of
the approach. In a real case, limitations are due to the spatial
and spectral resolutions of the images. More than 50% of the
rainfall-induced landslides in Taiwan are less than 50 m in
length. Landslides of this scale are not readily identifiable using
images of a pixel-size larger than 10 m. By pixel-wise
classification, landslides can occupy only individual or just a
few pixels without forming an outer shape of landslides.
Moreover, commission and omission errors can further
complicate the situation.
2.2 Interpretation Signatures
Key rules for this study are summarized from literatures, case
studies, and expert experiences, as shown in Table 1.
Key Rule Contents
Colour Tone|Brown, deep brown, bright. brown, green
Criterion brown
Location In the vicinity of ridge lines, road sides, and
Criterion the cut-off side of a river channel
Shape Lenticular-shaped or spoon-shaped, or
Criterion cumulated as tree-shaped in river basins, or a
triangular or rectangular-shape if located near
river banks
Direction The longitudinal axis is in the direction of|
Criterion gravity or perpendicular to flow-lines
Shadow Shadows are applied to assist the interpreter to
Criterion percept river bottoms and ridges in 2D images
Table 1 rules of interpretation for landslides
The rules of interpretation for landslides in Table 1 are to be
implemented as computing algorithms for automatic
identification. For example, the colour tone of a new landslide
is usually an expression of bare lands with unique spectral
signature. NDVI (Normalized Vegetation Index) is one of the
20 vegetation indices, useful for this purpose. Equation of
NDVI is as follows:
appt «MER (1)
NIR+R
This index is derived from the reflectance of red band and NIR
band. It is also an indicator of biomass. The value of NDVI is in
the range of —1 and +1. À negative value designates a bare land.
The location criterion of a landslide can be realized by using
DTM (Digital Terrain Model) for generating a ridgeline and by
digitising roads from the 1:5000 orthophoto maps, which are
the most common maps in Taiwan. Subsequently, a vicinity
analysis can be implemented.
The direction criterion is implemented by intersection
operation of the ridgelines and buffer zones generated by river-
lines.
The shape criterion and shadow criterion are not implemented
in this study. However, slope criterion is added. Statistics shows
that highest possibility of landslides take place on slopes of
15°~30°, and then on slopes of 30°~45° (Hsiao et. al., 2003).
A synergy of satellite images, DTM, existing roads, and
drainage lines is better implemented in a neural network system
as adopted in this study. A scoring scheme is used to transform
the above-mentioned criteria into the neurons of input layer of
the artificial neural network as shown in Table 2.
Colour Criterion Direction Location Criterion
‘Criterion (Ridge line)
NDVI Score | Buffer Score Buffer Score
Value size size
<0.0 1.0 «50m 1.0 «50m 1.0
0.0~0.25 0.8 50~100 0.8 50~100 0.8
0.25~0.5 0.6 100~150 | 0.6 100~150 0.6
0.5~0.75 0.4 150~200 | 0.4 150~200 0.4
575
