Vávra, A: Phenological Observation Treatment in the Landscape Mapping of the ...
Marjanović, M: Advanced Landslide Assessment of the Halenkovice Experimental Site
1. Advanced Landslide Assessment of the
Halenkovice Experimental Site
Miloš Marjanović
This presentation is co-financed by the
European Social Fund and the state
budget of the Czech Republic
2. Introduction
Motifs:
raising awareness
need for diverse case studies at different
scales, using different methods
applicability (decision making for land use planning and civil protection)
Objectives:
reliability and coherency of inputs (specially landslide inventory)
performing advanced modeling (many different methods)
evaluating models in the best fashion
providing maps/models as final outputs to be used in
practical/scientific manner
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
3. Introduction
Landslides – mass movements of the ground
Landslide susceptibility – spatial probability
of landslide occurrence (relation to hazard, risk…)
Setting definition:
Classification after Varnes 1978 (defining the mechanism and typology)
Scale/resolution (mid-scale, after Fell et all 2008)
Raster format data structure, pixel resolution 10 m
Definition of geometry (size, depth, area, frequency of landslides)
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
4. Introduction
Problems & perspectives in landslide assessment
lack of data, lack of possibility to relate events with triggers, non-
linearity of the problem…
piling investigations, promising capacities for monitoring (ground
sensors and Remote Sensing) in the future
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
5. Methodology
Methods for data pre-processing and selection: q n (ϕ oi , j − ϕ ei , j )2
Χ =∑ ∑
Chi-square
2
i=1 j =1 ϕ ei , j
Entropy
k ni n
E ( Sin ) = −∑
N
log 2 i
N
i =1
Landslide modeling methods ADVANCED!
Deterministic, Heuristic, Statistical, Fuzzy, Machine Learning
Methods for data evaluation
ROC plot
Kappa-index
n n n
κ =( ∑
i =1
xii − ∑
i =1
yii ) /(1 − ∑y )
i =1
ii
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
6. Methodology
Machine learning - Support Vector Machines
(SVM)
Classification task
Optimization (only two parameters)
Training over sampling splits
Testing the rest of the dataset with trained classifier
Kernels
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
7. Methodology
support vectors
landslide
e.g. aspect
e.g. aspect
stable
e.g. slope e.g. slope
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
8. Methodology
Experiment design
SAGA
SAGA
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
9. Methodology
Experiment design
Testing
Cross-Validation
Training
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
10. Case Study Dataset
Study Area
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
11. Case Study Dataset
Landslide Inventory
CGS survey (1:10 000)
http://mapy.geology.cz/svahove_nestability/
Field investigation
Independent field survey
Continuation from previous studies at the department
(Křivka, Marek, Bíl)
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
12. Case Study Dataset
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
13. Case Study Dataset
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
14. Case Study Dataset
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
15. Case Study Dataset
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
16. Case Study Dataset
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
17. Case Study Dataset
Thematic attributes # attribute source
Morphometric attributes 1 DEM Topo-maps
2 Slope DEM
Hydrological attributes 3 Slope length DEM
Environmental attributes 4 Aspect DEM
Geological attributes 5 Plan/profile curvature DEM
6 Convergence index DEM
7 Drainage elevations DEM
8 Elevation above drainage DEM
9 Drainage buffer DEM
10 LS factor DEM
11 TWI DEM
12 Catchment area DEM
13 Land cover units Orthophoto
nominal
14 Lithological units Geo-maps
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
18. Case Study Dataset
Attribute layers
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
19. Case Study Results
Model accuracy
=== Summary ===
Correctly Classified Instances 304080 = 88.16 %
Incorrectly Classified Instances 40814 = 11.83 %
Kappa statistic 0.1025
Mean absolute error 0.1183
Root mean squared error 0.344
Relative absolute error 75.3045 %
Root relative squared error 136.5789 %
Coverage of cases (0.95 level) 88.1662 %
Mean rel. region size (0.95 level) 50 %
Total Number of Instances 344894
=== Detailed Accuracy By Class ===
TP Rate FP Rate Precision Recall F-Measure MCC ROC Area PRC Area Class
0.932 0.823 0.941 0.932 0.936 0.103 0.555 0.94 0
0.177 0.068 0.156 0.177 0.166 0.103 0.555 0.082 1
Avg.0.882 0.773 0.889 0.882 0.885 0.103 0.555 0.883
=== Confusion Matrix ===
a b <-- classified as: a=non-landslide
300020 21980 | a=0 b=landslide
18834 4060 | b=1
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
20. Case Study Results
Comparison with an earlier, non-predictive model based
on multivariate regression
First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
21. Conclusions
Overall:
model seems promising, but there is room for improvements
the study is in its beginning and it might be interesting to extend
it methodologically and to compare the results
Drawbacks
bad communication between GIS and Machine Learning platform
time consumption
For further notice:
it is necessary to increase the number of folds in optimization
it would be interesting to challenge the algorithm with multi-
class (multinomial) scenario
post-procesing might be good refinement for the overall
accuracy First InDOG Doctoral Conference, 29th October - 1st November 2012, Olomouc
22. Advanced Landslide Assessment of the
Thank You For Your Attention!
Halenkovice Experimental Site
Miloš Marjanović
milos.marjanovic01@upol.cz
This presentation is co-financed by the
European Social Fund and the state
budget of the Czech Republic