Forecasting Rainfall-Induced Landslides
Faraz S. Tehrani, Ph.D.
Deltares
Delft University of Technology
Data Science Symposium, Deltares, Delft, November 201815 november 2018
Acknowledgement
• Team
Giorgio Santinelli, Gennadii Donchyts, Meylin Herrera
• Programs
• Impact of Extreme Weather
Ferdinand Diermanse & Robert McCall
• Urban Engineering
Jelle Buma & Mandy Korff
15 november 2018 2
Outline
 Background
 Data
 Prediction
 Ongoing activities
15 november 2018 3
Landslide research plan
15 november 2018 4
Forecasting
Post-hazard
Analysis
Risk
Assessment
Mitigation
Background Data Prediction Current activities
Landslide research plan
15 november 2018 5
Forecasting
Post-hazard
Analysis
Risk
Assessment
Mitigation
Background Data Prediction Current activities
• Global
• Regional
• Local
• Site-specific
Landslide Forecasting
15 november 2018 6
Background Data Prediction Current activities
• Improving awareness and hazard understanding
• Global landslide susceptibility maps
• Early warning
• Emergency response
Target susceptibility map
Datasets
Global landslide inventory
Rainfall
Digital Elevation Models
Soil
Vegetation Index
Temperature
Soil moisture
Lithology
Land cover
Drainage network
Road network
Tehrani & Santinelli (2019)
15 november 2018 7
Database
Herrera (2018)
• PostgreSQL / PostGIS database with all relevant data
Landslide
eventsLandslide_id
Landcover
Lat, Lon
Landslide_id
SoilMoisture
Lat, Lon
Landslide_id
Rainfall
Lat, Lon
Landslide_id
VegetationIndex
Lat, Lon
Landslide_id
Topography
Lat, Lon
Landslide_id
Temperature
Lat, Lon
Landslide_id Lat, Lon
Join Join
Join
Join Join
Background Data Prediction Current activities
Global Landslide Catalogue
15 november 2018 8
Background Data Prediction Current activities
• NASA
• 10,988 landslides
• 2007 – 2018
• Based on media
Herrera (2018)
Global Landslide Catalogue
15 november 2018 9
Background Data Prediction Current activities
4542 landslides to be used
• NASA
• 10,988 landslides
• 2007 – 2018
• Based on media
PERSIANN CDR
• 1983-Present
• 0.25° x 0.25°
TRMM 3B42 (Daily)
• 0.25° x 0.25°
• 1998-Present
TRMM 3B43 (Monthly)
• 0.25° x 0.25°
• 1998-Present
15 november 2018 10
day 0day -1day -2day -3…day -10
short-termlong-term
Background Data Prediction Current activities
Precipitation
Tehrani et al. (2019)
Shuttle Radar Topography Mission (SRTM1)
• 2000
• 1ʺ× 1ʺ (approximately 30 m × 30 m)
Advanced Land Observing Satellite (ALOS)
• 2011
• 1ʺ× 1ʺ (approximately 30 m × 30 m)
Multi-Error-Removed Improved-Terrain (MERIT)
• 2017
• 3ʺ× 3ʺ (approximately 90 m × 90 m)
15 november 2018 11
Digital Elevation Models
Elevation relief = Elevationmax – Elevationmin
Tehrani et al. (2019)
Background Data Prediction Current activities
15 november 2018 12
Soil & Bedrock
SoilGrids
• 2017
• 250 m x 250 m
• Depth to the bedrock
• Sand fraction
• Silt fraction
• Clay fraction
Hengl et al. (2017)
Tehrani et al. (2019)
Background Data Prediction Current activities
[cm]
15 november 2018 13
Vegetation Index
Tehrani et al. (2019)
Normalized Difference Vegetation Index ( -1<NDVI<+1)
• Distinct colors (wavelengths) of visible and near-infrared sunlight
reflected by the plants
• Green leaves strongly absorb visible light and reflect near-infrared light
NDVI = (NIR — VIS)/(NIR + VIS)
MOD13Q1 v.6
• 2000-Present
• 250 m x 250 m
Background Data Prediction Current activities
15 november 2018 14
Machine Learning
Background Data Prediction Current activities
Logistic Regression
• a classification algorithm
• returns a probability value
• maps to two or more discrete classes
1
( )
1 z
p z
e


0 0 1 1 2 2 3 4 n nz w x w x w x w x w x     
15 november 2018 15
Datasets
Background Data Prediction Current activities
x0 x1 x2 x3 x4 x5
E0
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10
Short-term rain: Accumulated rain for the day prior to and the day of landslide
Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide
Elevation relief: The difference between maximum and minimum elevation
Slope: Average slope of at the location of landslide
NDVI before landslide
Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
Non-Landslide Cases
Rain: Noise was only applied to rainfall data
DEM: Noise was only applied to topography data
All: Noise was applied to both rainfall and topography features
9084 Landslide/Non-Landslide cases
15 november 2018 16
Datasets
Background Data Prediction Current activities
x0 x1 x2 x3 x4 x5
E0
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10
Short-term rain: Accumulated rain for the day prior to and the day of landslide
Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide
Elevation relief: The difference between maximum and minimum elevation
Slope: Average slope of at the location of landslide
NDVI before landslide
Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
15 november 2018 17
Datasets
Background Data Prediction Current activities
x0 x1 x2 x3 x4 x5
E0
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10
Short-term rain: Accumulated rain for the day prior to and the day of landslide
Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide
Elevation relief: The difference between maximum and minimum elevation
Slope: Average slope of at the location of landslide
NDVI before landslide
Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
15 november 2018 18
Datasets
Background Data Prediction Current activities
x0 x1 x2 x3 x4 x5
E0
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10
Short-term rain: Accumulated rain for the day prior to and the day of landslide
Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide
Elevation relief: The difference between maximum and minimum elevation
Slope: Average slope of at the location of landslide
NDVI before landslide
Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
15 november 2018 19
Landslide detection
Background Data Prediction Current activities
MSc Thesis Project (Meylin Herrera):
Landslide detection using Machine Learning with application in landslide
susceptibility mapping
Mud Creek, California, 20/05/2017
15 november 2018 20
Landslide detection
Background Data Prediction Current activities
MSc Thesis Project (Meylin Herrera):
Landslide detection using Machine Learning with application in landslide
susceptibility mapping
Mud Creek, California, 20/05/2017
Optical Radar
15 november 2018 21
Landslide detection
Background Data Prediction Current activities
Pixel-based or OBIA Training
Yes
Adjust
parameters
• Satellite Imagery:
Sentinel-2 /Landsat-
8
• DEM
• NDVI
• Landslide Inventory
(Franeltalia, 2018)
Triggering factor:
• Rainfall data
Controlling factors:
• Topography from DEM
(Slope, height, aspect)
• NDVI
• Soil moisture
• Soil Composition
• Landcover
• Temperature
Test
INPUT ML
Accuracy
Assessment
Testing
Good overall
accuracy ?
Areas with data
scarcity
No
generateLandslide
Inventory maps
OUTPUT
(LSM)
Input Data
Input Data
Susceptibility Mapping (LSM)
Landslide Detection
(LD)
Validation
Herrera (2019?)
Some remarks
 A database was created for global rainfall-induced landslides
 Preliminary analysis showed that rainfall-induced landslides can be
predicted with a reasonable accuracy
 More controlling features need to be added
 Accuracy and resolution of the data is important and must be
improved
 Landslide detection algorithms will be developed
 Global model is for “awareness” and as a first step towards
regional and local predictions and planning
 Climate scenarios can be applied to the model for global prediction
of landslides in future
15 november 2018 22

DSD-INT 2018 Global Landslides analysis and forecasting - Tehrani

  • 1.
    Forecasting Rainfall-Induced Landslides FarazS. Tehrani, Ph.D. Deltares Delft University of Technology Data Science Symposium, Deltares, Delft, November 201815 november 2018
  • 2.
    Acknowledgement • Team Giorgio Santinelli,Gennadii Donchyts, Meylin Herrera • Programs • Impact of Extreme Weather Ferdinand Diermanse & Robert McCall • Urban Engineering Jelle Buma & Mandy Korff 15 november 2018 2
  • 3.
    Outline  Background  Data Prediction  Ongoing activities 15 november 2018 3
  • 4.
    Landslide research plan 15november 2018 4 Forecasting Post-hazard Analysis Risk Assessment Mitigation Background Data Prediction Current activities
  • 5.
    Landslide research plan 15november 2018 5 Forecasting Post-hazard Analysis Risk Assessment Mitigation Background Data Prediction Current activities • Global • Regional • Local • Site-specific
  • 6.
    Landslide Forecasting 15 november2018 6 Background Data Prediction Current activities • Improving awareness and hazard understanding • Global landslide susceptibility maps • Early warning • Emergency response Target susceptibility map Datasets Global landslide inventory Rainfall Digital Elevation Models Soil Vegetation Index Temperature Soil moisture Lithology Land cover Drainage network Road network Tehrani & Santinelli (2019)
  • 7.
    15 november 20187 Database Herrera (2018) • PostgreSQL / PostGIS database with all relevant data Landslide eventsLandslide_id Landcover Lat, Lon Landslide_id SoilMoisture Lat, Lon Landslide_id Rainfall Lat, Lon Landslide_id VegetationIndex Lat, Lon Landslide_id Topography Lat, Lon Landslide_id Temperature Lat, Lon Landslide_id Lat, Lon Join Join Join Join Join Background Data Prediction Current activities
  • 8.
    Global Landslide Catalogue 15november 2018 8 Background Data Prediction Current activities • NASA • 10,988 landslides • 2007 – 2018 • Based on media Herrera (2018)
  • 9.
    Global Landslide Catalogue 15november 2018 9 Background Data Prediction Current activities 4542 landslides to be used • NASA • 10,988 landslides • 2007 – 2018 • Based on media
  • 10.
    PERSIANN CDR • 1983-Present •0.25° x 0.25° TRMM 3B42 (Daily) • 0.25° x 0.25° • 1998-Present TRMM 3B43 (Monthly) • 0.25° x 0.25° • 1998-Present 15 november 2018 10 day 0day -1day -2day -3…day -10 short-termlong-term Background Data Prediction Current activities Precipitation Tehrani et al. (2019)
  • 11.
    Shuttle Radar TopographyMission (SRTM1) • 2000 • 1ʺ× 1ʺ (approximately 30 m × 30 m) Advanced Land Observing Satellite (ALOS) • 2011 • 1ʺ× 1ʺ (approximately 30 m × 30 m) Multi-Error-Removed Improved-Terrain (MERIT) • 2017 • 3ʺ× 3ʺ (approximately 90 m × 90 m) 15 november 2018 11 Digital Elevation Models Elevation relief = Elevationmax – Elevationmin Tehrani et al. (2019) Background Data Prediction Current activities
  • 12.
    15 november 201812 Soil & Bedrock SoilGrids • 2017 • 250 m x 250 m • Depth to the bedrock • Sand fraction • Silt fraction • Clay fraction Hengl et al. (2017) Tehrani et al. (2019) Background Data Prediction Current activities [cm]
  • 13.
    15 november 201813 Vegetation Index Tehrani et al. (2019) Normalized Difference Vegetation Index ( -1<NDVI<+1) • Distinct colors (wavelengths) of visible and near-infrared sunlight reflected by the plants • Green leaves strongly absorb visible light and reflect near-infrared light NDVI = (NIR — VIS)/(NIR + VIS) MOD13Q1 v.6 • 2000-Present • 250 m x 250 m Background Data Prediction Current activities
  • 14.
    15 november 201814 Machine Learning Background Data Prediction Current activities Logistic Regression • a classification algorithm • returns a probability value • maps to two or more discrete classes 1 ( ) 1 z p z e   0 0 1 1 2 2 3 4 n nz w x w x w x w x w x     
  • 15.
    15 november 201815 Datasets Background Data Prediction Current activities x0 x1 x2 x3 x4 x5 E0 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 Short-term rain: Accumulated rain for the day prior to and the day of landslide Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide Elevation relief: The difference between maximum and minimum elevation Slope: Average slope of at the location of landslide NDVI before landslide Sand fraction, Silt fraction, Clay fraction and depth ro bedrock Non-Landslide Cases Rain: Noise was only applied to rainfall data DEM: Noise was only applied to topography data All: Noise was applied to both rainfall and topography features 9084 Landslide/Non-Landslide cases
  • 16.
    15 november 201816 Datasets Background Data Prediction Current activities x0 x1 x2 x3 x4 x5 E0 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 Short-term rain: Accumulated rain for the day prior to and the day of landslide Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide Elevation relief: The difference between maximum and minimum elevation Slope: Average slope of at the location of landslide NDVI before landslide Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
  • 17.
    15 november 201817 Datasets Background Data Prediction Current activities x0 x1 x2 x3 x4 x5 E0 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 Short-term rain: Accumulated rain for the day prior to and the day of landslide Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide Elevation relief: The difference between maximum and minimum elevation Slope: Average slope of at the location of landslide NDVI before landslide Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
  • 18.
    15 november 201818 Datasets Background Data Prediction Current activities x0 x1 x2 x3 x4 x5 E0 E1 E2 E3 E4 E5 E6 E7 E8 E9 E10 Short-term rain: Accumulated rain for the day prior to and the day of landslide Long-term rain: Accumulated rain for days ranging from 10 days to 2 days before landslide Elevation relief: The difference between maximum and minimum elevation Slope: Average slope of at the location of landslide NDVI before landslide Sand fraction, Silt fraction, Clay fraction and depth ro bedrock
  • 19.
    15 november 201819 Landslide detection Background Data Prediction Current activities MSc Thesis Project (Meylin Herrera): Landslide detection using Machine Learning with application in landslide susceptibility mapping Mud Creek, California, 20/05/2017
  • 20.
    15 november 201820 Landslide detection Background Data Prediction Current activities MSc Thesis Project (Meylin Herrera): Landslide detection using Machine Learning with application in landslide susceptibility mapping Mud Creek, California, 20/05/2017 Optical Radar
  • 21.
    15 november 201821 Landslide detection Background Data Prediction Current activities Pixel-based or OBIA Training Yes Adjust parameters • Satellite Imagery: Sentinel-2 /Landsat- 8 • DEM • NDVI • Landslide Inventory (Franeltalia, 2018) Triggering factor: • Rainfall data Controlling factors: • Topography from DEM (Slope, height, aspect) • NDVI • Soil moisture • Soil Composition • Landcover • Temperature Test INPUT ML Accuracy Assessment Testing Good overall accuracy ? Areas with data scarcity No generateLandslide Inventory maps OUTPUT (LSM) Input Data Input Data Susceptibility Mapping (LSM) Landslide Detection (LD) Validation Herrera (2019?)
  • 22.
    Some remarks  Adatabase was created for global rainfall-induced landslides  Preliminary analysis showed that rainfall-induced landslides can be predicted with a reasonable accuracy  More controlling features need to be added  Accuracy and resolution of the data is important and must be improved  Landslide detection algorithms will be developed  Global model is for “awareness” and as a first step towards regional and local predictions and planning  Climate scenarios can be applied to the model for global prediction of landslides in future 15 november 2018 22