Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

DSD-INT 2018 Data-driven impact modelling for the estimation of disaster damage - Wagenaar


Published on

Presentation by Dennis Wagenaar (Deltares) at the Data Science Symposium 2018, during Delft Software Days - Edition 2018. Thursday 15 November 2018, Delft.

Published in: Software
  • Be the first to comment

DSD-INT 2018 Data-driven impact modelling for the estimation of disaster damage - Wagenaar

  1. 1. Data-driven impact modelling Data-driven impact modelling for the estimation of disaster damage Dennis Wagenaar
  2. 2. • Decision making for disaster prevention, response and recovery. • Disaster impacts are typically what determines how to respond and what measures are feasible . • Possible disaster impacts: • Flood damage • Flood casualties • People affected • Required assistance • Required building materials Why model disaster impacts
  3. 3. Application of impact modelling • CBA infrastructure investments • Risk screening studies • Adaptive delta management • Setting insurance premiums • Climate change impact • How much money is needed for recovery? • Allocation of recovery funds Impact Based Forecasting: Warning information and preparation event Prioritization of rescue and aid actions
  4. 4. Current approach: Types of disaster damage Category Tangible Intangible Direct • Capital (buildings incl. contents, crops, cars) • Production losses, income losses • Casualties, injuries, monuments • Social disruption, emotional damage Indirect • Production losses / loss of utility services outside flooded area; • Unemployment, migration. • Cutting of infrastructure lines • Loss of potential for attracting investors • Reputation damage Delft-FIAT Multiplication factor 5-100% of capital damage
  5. 5. Current approach: Delft-FIAT Hazard Vulnerability Exposure
  6. 6. Vulnerability: The damage function Expert/synthetic approach • Expert or group of experts come together and estimate a damage function. • Elements of object of interest can be assessed individually. • Weakness: experts typically have one setting in mind. Data-driven approach (empirical) • Regression analysis on available data points of past flood damage. • Weakness: Data availability and bad fits.
  7. 7. Things not to do with current approach: • Copy damage functions from the literature (transfer) • CBA for measures that influence anything but water depth and probability (e.g. resilience measures) • Allocation of recovery funds • Detailed prioritization of emergency help • Uncertainty estimation (for insurance premiums) Gaps in current approach • Large differences in damage functions (hidden assumptions)
  8. 8. The multi-variable data-driven approach From: Damage fraction = f(water depth) To: Damage fraction = f(water depth, warning time, wave height, …..) DF = f(water depth) DF = f(water depth, warning time, waves height, …..) Multi-variable damage models can be build from data with Machine learning methods!
  9. 9. • Multi-variable linear regression • Tree methods • Regression trees • Bagging trees • Random Forest • Artificial Neural Networks • Bayesian Networks • Support Vector Machines Supervised Machine Learning methods Example with limited number of dimensions
  10. 10. Required data and supervised Machine Learning depth velocity duration warning population …. damage 1.3 0.25 190 3 4…. €32,000 2.1 0.15 250 5 2… €21,000 …. ….. ….. ….. …. … …. 1.8 0.2 160 4 3… ? Target/ dependent variableFeatures / independent variables / characteristics Training data Application data
  11. 11. Model transfer ? • We found out that in some cases multi-variable flood damage models can be transferred. • Transfer from Germany to Netherlands worked well • Transfer from Netherlands to Germany went wrong • We are working on techniques to make this transfer work better
  12. 12. We need more data! Training data • Purpose to find relationships between features and damage. • We only need samples! • Good heterogeneous samples better than large amount. • Definitions and collection guidelines. • Collection probably manual labor? • Global organization to coordinate collection efforts Application data • Purpose data required for model application. • Needs to be complete for study area. • Collection using remote sensing (street view?) • Need good hazard models • Can sometimes be estimated Building feature recognition