Flood loss estimation using 3D city models and remote sensing data

Kai Schröter, Stefan Lüdtke, Richard Redweik, Jessica Meier, Mathias Bochow, Lutz Ross, Claus Nagel, Heidi KreibichEnviron. Model. Softw 2018 · DOI

TL;DR

This paper presents a novel approach to estimating flood losses in urban areas by integrating 3D city models and remote sensing data. By leveraging standardized data sources, the study demonstrates comparable flood loss prediction accuracy to models based on detailed empirical surveys, paving the way for more accessible and transferable flood risk assessments.

What Problem Does This Paper Address?

Accurate flood risk assessment is crucial for optimizing flood risk management strategies. However, a significant challenge lies in obtaining detailed, object-related data needed for spatially explicit risk information. While traditional flood loss models often rely on inundation depth and generic damage functions, this paper addresses the need for more sophisticated, multi-variable models that can better capture building vulnerability. The authors aim to bridge the gap between the availability of detailed data and the requirements of advanced flood loss modeling by exploring the potential of readily available standardized data sources like 3D city models and remote sensing data.

Methods

The core of this research lies in developing and evaluating multi-variable flood loss models using Random Forests (RF). RF is chosen for its ability to handle complex, non-linear relationships and its robustness against overfitting. The study employs a systematic approach:

  • Variable Selection: RF's variable importance feature is used to identify the most relevant predictors from various data sources, aiming for parsimonious models that balance predictive performance with data requirements.
  • Model Development: Flood loss models are trained using different combinations of data, including empirical survey data, 3D city model attributes, and numerical spatial measures derived from remote sensing.
  • Model Evaluation: Performance is assessed using metrics like Mean Squared Error (MSE), Mean Absolute Error (MAE), Mean Bias Error (MBE), and quantile-based measures (QR90, HR) to evaluate accuracy, precision, and prediction uncertainty. A leave-one-out cross-validation procedure is used to test model transferability.

Datasets

The study utilizes several datasets for training and evaluation:

  • Empirical Survey Data (CATI): Computer-aided telephone interviews conducted after floods in Dresden, providing information on flood situation, building characteristics, and relative flood loss. This dataset serves as a benchmark.
  • 3D City Model Data (CityGML): Geometric and thematic attributes of buildings in Dresden, including building area, height, roof type, and construction year. This data is available at Levels of Detail (LOD) 1 and 2.
  • Remote Sensing Data (Numerical Spatial Measures): Derived from IKONOS satellite imagery, these measures capture urban structure characteristics like land cover proportions, shape, and spatial arrangement within building blocks.
  • Inundation Depth Maps: Generated from hydro-numeric simulations for three flood scenarios in Dresden.
  • ICPR Model: A standard stage-damage function used for flood risk management in Dresden, serving as a lower benchmark.

Key Insights and Ideas

  • Standardized Data Sufficiency: The study demonstrates that readily available standardized data from 3D city models and remote sensing can yield flood loss prediction performance comparable to models built on detailed, empirical survey data. This is a significant finding for making flood risk assessment more accessible.
  • Geometric Properties Matter: Building geometric properties (area, height, shape, density) derived from 3D city models and remote sensing are identified as strong predictors of flood loss, directly or indirectly linked to building types and their associated vulnerabilities.
  • Random Forests for Complexity: RF proves to be a powerful tool for handling the heterogeneity and complexity of urban flood loss modeling, effectively identifying important variables and providing insights into prediction uncertainty.
  • Transferability Potential: The success of models built on standardized data suggests a greater potential for spatial transferability of flood loss models, a critical aspect for broader application.

Key Results

  • The citydb_sm model, which combines 3D city model data and spatial measures, achieved the best performance in terms of MAE and MSE among models using standardized data.
  • The citydb model (using only 3D city model data) showed the best scores for QR and HR, indicating lower prediction uncertainty and more reliable predictions.
  • Models based on standardized data sources (e.g., citydb_sm) performed comparably to models based on empirical survey data (cati complete and cati reduced).
  • The sdf model (using only water depth) showed inferior performance, highlighting the benefit of incorporating additional variables.
  • In a leave-one-out cross-validation, the citydb_sm and citydb models demonstrated comparable performance to the empirical data-based models, with citydb_sm yielding better MAE, MBE, and MSE.
  • The icpr model consistently performed worse than the models based on local data, underscoring the advantage of using region-specific data.
  • The citydb model reproduced the mean of empirical relative loss values and captured some of the variability in relative loss for comparable inundation depths.
  • The expected annual damage (EAD) calculated using the citydb model was 3.5 times higher than that calculated using the icpr model, demonstrating the substantial impact of model choice on investment decisions for flood mitigation.

Limitations and Future Work

  • Data Availability (LOD): The study acknowledges that current 3D city models primarily offer LOD1 or LOD2 information, which is mostly geometric. Future work could benefit from higher LODs (LOD3/LOD4) that provide details on building openings, sill heights, and internal facilities.
  • Sample Size: The relatively small sample size (n=80) of empirical data might influence variable importance and model performance analyses. While steps were taken to mitigate this (bootstrapping, cross-validation), larger datasets would further strengthen findings.
  • Spatial Transferability Testing: While the study demonstrates comparable performance to empirical models, the "acid test" for the derived flood loss models would be their application and performance testing in a different city, investigating the assimilation of local data.
  • Uncertainty Visualization: The paper suggests that further research could explore how to effectively visualize the predictive uncertainty provided by RF models to aid decision-making and improve risk communication.
  • Broader Data Integration: The study highlights the potential for integrating more sophisticated and complex modeling approaches by exploiting available data sources, suggesting that 3D city model-based technology offers significant opportunities in this regard.

Summary updated

BibTeX
@article{schroter2018flood,
  author = {Kai Schröter and Stefan Lüdtke and Richard Redweik and Jessica Meier and Mathias Bochow and Lutz Ross and Claus Nagel and Heidi Kreibich},
  title = {{Flood loss estimation using 3D city models and remote sensing data}},
  journal = {Environ. Model. Softw.},
  year = {2018},
  doi = {10.1016/j.envsoft.2018.03.032},
}
Figure