Presentation by Dennis Wagenaar, Deltares, at the Data Science Symposium, during Delft Software Days - Edition 2019. Thursday, 14 November 2019, Delft.
DSD-INT 2019 How machine learning will change flood risk and impact assessments - Wagenaar
1. How machine learning will change flood
risk and impact assessments
Dennis Wagenaar
2. • Group vision paper
• Based on brainstorm
• Collaboration between:
• Universities
• World Bank
• Startups
• Deltares
Understanding risk field lab on urban flooding
5. Machine learning in flood risk and impact assessments
Not completely
new:
• Hydroinformatics
• Remote sensing
• Could be applied
more
Much more potential!
7. Where can we apply it?
Predictive Descriptive
Exposure Urban growth
modelling
Identification current
build-up
Hazard Flood modelling Mapping current and
past floods
Impact Flood impact
modelling
Assessing flood
impacts
8. How machine learning works
Indicators Response (variable of interest)
Trainingdata
Historical records of indicators
(e.g. rainfall, wind speed, building data)
Historical records of response
(e.g. damage)
Applicationdata
Indicators data new case
(e.g. rainfall, wind speed, building data)
Response new case
(e.g. damage)
9. A machine learning algorithm you may already know: Linear regression
X: Indicator data (e.g. water level)
Y: Response (e.g. damage)
Blue dots: Historical data
Red line: Model
New predictions made by
looking up y for a given x
10. Example of more advanced machine learning algorithm: Decision trees
Well known algorithms
• Linear regression
• Multi-variable linear regression
• Polynominal regression
• Logistic regression
• Decision/regression tree
• Random Forest
• Artificial Neural Networks (ANN)
• Convolutional Neural Network (CNN)
• Support Vector Machines (SVM)
• Bayesian Networks
11. • Very good physics based
models
• Machine learning can’t deal
with system change
• Useful for forecasting
frequent events
• Sometimes better or cheaper
• Surrogate models
• Google: Automatic
calibration based on remote
sensing data
Predictive hazard: Modelling the water
12. • Social media (e.g. twitter)
• Flood mapping
• Water depths from photos
• Remote sensing
• Optical data – cloud
cover/night
• Synthetic Aperture Radar
(SAR)
• Automatically label floods
Descriptive hazard: Observing floods
13. From the air
• Global building footprints
• Global road information
• Should become available at one
point soon.
From the ground
• 360 degree street view
• Building materials
• Building entrance heights
Descriptive exposure: Automatic detection
360 degree street view images
14. • Predicting impact
• Predict flood damage
• Predict health impacts
• Predict flood casualties
• Predict required resources
• Already done, lack of data
• Exposure data could become
available
Predictive impacts: The final step
15. • Machine learning raises ethics and
bias questions
• Automatic weaponry
• Facial recognition
• Aggravating inequalities through
biased training sets
• Increased complexity
• Misuse
• Lack of uncertainty
communication
• Overhyped
• Working group, guideline
Ethics and bias
16. Model chains and machine learning
Wind speed Surge
model
Overland
flow model
Flood damage
model
Traditional chain
Hybrid chain
Pure machine
learning
Wind speed Surge
model
Overland
flow model
ML flood
damage
model
Wind speed
Distance to
shore
Elevation
ML flood
damage
model
Damage
Damage
Damage
Key is making the
right choice per
component
17. Machine learning methods vs traditional methods (1)
Exact known
relationships
Complex
processes with
many variables
Use formulas
based on physics
Consider data-
driven methods
18. Machine learning methods vs traditional methods (2)
Extrapolation or
system changes
No
extrapolation or
system changes
Use formulas
based on physics
Consider data-
driven methods
19. What will become possible
Better modelsNew applications
• Targeted early actions
• Early harvesting crops
• Strengthening
buildings
• Quick estimates of
recovery needs
• Parametric insurance