This document presents preliminary results of a study analyzing the sensitivity of shallow landslide hazards to weather factors. The study developed a statistical model using two techniques - Generalized Linear Model and Random Forest - to predict shallow landslides based on static thematic data at 30m resolution and dynamical weather predictions from the WRF model at 3km resolution. The results found the statistical models had good agreement with observations. Weather predictions were more important for a 2011 rainfall event while static predictors were more important for a 2013 event. The study concludes numerical weather predictions, especially hourly rainfall intensities and soil moisture, provide useful information for shallow landslide prediction and the statistical model bridges micro and meso scales of landslides and weather forecasts.
1. Weather factors sensitivity in shallow landslide hazard:
preliminary results
Perna(1) M., Capecchi(1) V., Crisci(2) A., Corongiu(3) M., Manetti(3) F.
(1) CNR IBIMET - Consorzio LaMMA, Sesto Fiorentino, Italy, (2) CNR IBIMET, Firenze, Italy, (3) Consorzio LaMMA, Sesto Fiorentino, Italy
Data and methodsGoal
Header
[1] Mercogliano et al: A prototype forecasting chain for rainfall induced shallow
landslides, Natural Hazards & Earth System Sciences, 13,2013
[2] Segoni et al: Towards a definition of a real-time forecasting network for rainfall
induced shallow lanslides., Natural Hazards & Earth System Sciences, 9, 2009
[3] Catani et al: Landslide susceptibility estimation by random forests technique:
sensitivity and scaling issues., Natural Hazards & Earth System Sciences, 990 13, 2013.
The study and the poster received the support of the
Results โ Variable importance
Input static thematic predictors (30 m spatial
resolution):
โ
Geomorphological: elevation, slope, altitude above
channel network, etc...
โ
Geological: distance from main tectonic features,
soil permeability, Slope Structural Setting, etc...
โ
Hydrological: Time of concentration, Topographic
Wetness Index, convergence/divergence to overland
flow, etc...
โ
Climate: mm of rainfall for an event with 100 years
as returning period
Input dynamical WRF numerical weather predictions
(3 km spatial resolution)
๎
Rainfall amounts (mm/24h) and intensities (mm/h)
๎
Soil moisture
Results - 25OCT2011
Might rainfall-induced shallow landsliding prediction
benefit from the information provided by the WRF
numerical weather predictions? (see ref. [1],[2])
Two heavy rainfall events occurred on 25OCT2011
(Lunigiana) and on 18MAR2013 (Garfagnana)
triggered a great number of shallow landslides,
causing damages, injuries and human losses.
Conclusions
Data and methods
We developed a
quantitative
indirect statistical
modeling.
Two statistical
techniques are
considered:
Generalized Linear
Model (GLM) and
Breiman's Random
Forest (RF, ref [3]).
Results - 18MAR2013
(I) Results found are in good agreement with
observations (see ROC curves and AUC values)
(II) 18MAR2013: GLM model performs better than RF
(III)25OCT2011: RF performs slightly better than GLM
(IV)WRF data are important predictors for
25OCT2011 event, whereas static predictors prevail
for 18MAR2013
FINDINGS
(I) The statistical model is simple, efficient and
provides reliable results even if NWP data are a not
downscaled towards higher spatial resolution
(II) NWP hourly precipitation intensities and soil
moisture should be considered as input predictors
since they are classified โimportantโ when deep
convection occurs (25OCT2011 rainfall event)
DRAWBACKS/LIMITS
(I) The statistical model is used as a โblack-boxโ, no
tuning performed so far
(II) the statistical model is data-driven. It might
performs differently in different areas or for different
rainfall events
VariableimportanceisprovidedbytheRandomForestalgorithm
Conclusions
Bridge the gap
between micro-ฮณ
scale (โค 20-30 m)
typical scale of
landslide
occurring at basin
scale, with the
meso-ฮณ scale
(2-20 km) which is
the typical scale of
the NWP forecast.
LIFE/12/ENV/IT/001054