je XXXIX-B8, 2012
\spect, Elevation, Soil
acture; the ANN output
nts of the occurrence of
"e Qutput
i’
Analysis result
rt of LS analysis
les
and the equation (1),
nput factors (quantitative
0
0
0
NN trainings would be
mentioned, the qualitative
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
values were selected from 1,200 landslide records, and the
summary of different frequencies every sample were
considered as the output data. The number of neurons of hidden
layer is more than twice the input data. The structure of Neural
network is “9-18-17. The BP learning algorithm, with network
training error of 10-6, and 500 times of training (Figure 2), was
implemented to train various ANN architectures.
(4.30 +
£1.25
0.20
vi
0. Sant
t5
0 77736 "ioo 150 200 250 300 350 400 450 500
Training times
Error
Figure 2. Error and training times of ANN
Some landslide test data were from worldwide authoritative
website (United States Geological Survey (USGS),
http://landslides.usgs.gov/recent; | EM-DAT International
Disaster Database, http://www.em-dat.net/; International
Consortium on Landslides Website (ICL), http://iclhq.org/).
Gathering 100 landslide test data, the accuracy of the network
is about 82.31%. The output data were considered to express
the LS index values of pixels. The higher the value of LS, the
more susceptible is that pixel to the occurrence frequencies of
landslide. Based on the 1kmx1km grid, the LS map of China is
then produced by ANN learning and multiplying the weighted
factor layers (Figure 3).
Figure 3. LS map of China based on multi-factors
From the figure can clearly see that the southwest of China has
higher susceptibility index than other areas, which also could
be consistent with landslide reports from news. Meanwhile, the
rainfall data from TRMM satellite can be used to predict the
landslide area by overlaying the LS map.
49
5. CONCLUSIONS
The most prominent features of landslide include information
diversity, fuzziness, uncertainty and randomness, which bring
great difficult for information processing. The processing
ability of traditional ways and methods is limited, and may
cause a lot of useful information losses, result in the degree of
reliability reducing. With the strong nonlinear mapping
capability, can simulate the possibility of the landslide and the
relationship of complex factors (Chacón et al, 2006).
A landslide susceptibility model has been developed for the
whole country applying a scoring system with a set of relevant
factors based on BP ANN, merging nonlinear elements by
qualitative and quantitative indices. The result has been tested
by the data of worldwide authoritative website, and the
accuracy of the network is about 82.31% to meet landslide
study requirements. The high susceptibility parts mainly in the
southwest of China, characterized by the presence of
landslide-prone sedimentary rocks, high seismicity, frequent
severe earthquake and rainfall events and significant human
activities in this area.
The landslide susceptibility evaluation and mapping of China
generated in this study constitutes a preliminary step for further
more detailed susceptibility and hazard research, as well as a
useful method for risk assessment, and provides theoretical
basis for prediction and forecast of landslide disasters
throughout China.
Acknowledgements
The work described in this paper was supported by National
Basic Research Program of China (2012CB957702) and
Innovation Program of Shanghai Municipal Education
Commission(Project ID:10ZZ25) , and also supported by the
Centre of Spatial Information Science and Sustainable
Development, Tongji University.
References
Henderson L J. Emergency and disaster: pervasive risk and
pubic bureaucracy in developing nations [J]. Pubic
Organization Review: A Global Journal, 2004, 4: 103-119.
Kirschbaum D B, Adler R, Hong Y, et al. A global landslide
catalog for hazard applications: method, results, and limitations
[7]. Nature Hazards, 2010, 52: 561-575.
