EGU2020-7334
https://doi.org/10.5194/egusphere-egu2020-7334
EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Probabilistic Flood Loss Models for Companies

Lukas Schoppa1,2, Tobias Sieg1,2, Kristin Vogel2, Gert Zöller3, and Heidi Kreibich1
Lukas Schoppa et al.
  • 1GFZ German Research Centre for Geosciences, Section 4.4 Hydrology, Potsdam, Germany (lukas.schoppa@gfz-potsdam.de)
  • 2Institute of Environmental Science and Geography, University of Potsdam, Potsdam, Germany
  • 3Institute of Mathematics, University of Potsdam, Potsdam, Germany

Flood risk assessment strongly relies on accurate and reliable estimation of monetary flood loss. Conventionally, this involves univariable deterministic stage-damage functions. Recent advancements in the field promote the use of multivariable probabilistic loss estimation models which consider damage controlling variables beyond inundation depth. Although companies contribute significantly to total loss figures, multivariable probabilistic modeling approaches for companies are lacking. Scarce data and heterogeneity among companies impedes the development of novel company flood loss models.

We present three multivariable flood loss estimation models for companies that intrinsically quantify prediction uncertainty. Based on object-level loss data (n=1306), we comparatively evaluate the predictive performance of Bayesian networks, Bayesian regression and random forest in relation to established stage-damage functions. The company loss data stems from four post-event surveys after major floods in Germany between 2002 and 2013 and comprises information on flood intensity, company characteristics and private precaution. We examine the performance of the candidate models separately for losses to building, equipment, and goods and stock. Plausibility checks show that the multivariable models are able to identify and reproduce essential relationships of the flood damage processes from the data. The comparison of the prediction capacity reveals that the proposed models outperform stage-damage functions clearly while differences among the multivariable models are small. Even though the presented models improve the accuracy of loss predictions, wide predictive distributions underline the necessity for the quantification of predictive uncertainty. This applies particularly to companies, for which the heterogeneity and variation in the loss data are more pronounced than for private households. Due to their probabilistic nature, the presented multivariable models contribute towards a transparent treatment of uncertainties in flood risk assessment.

How to cite: Schoppa, L., Sieg, T., Vogel, K., Zöller, G., and Kreibich, H.: Probabilistic Flood Loss Models for Companies, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-7334, https://doi.org/10.5194/egusphere-egu2020-7334, 2020

Comments on the presentation

AC: Author Comment | CC: Community Comment | Report abuse

Presentation version 2 – uploaded on 04 May 2020
Included affiliations on title slide
  • CC1: Comment on EGU2020-7334, John Maskell, 04 May 2020

    When you say "precaution" do you mean privately installed flood protection measures? If so, what were the most common types of measures you recorded and how did you define the precaution ratio?

    • AC1: Reply to CC1, Lukas Schoppa, 04 May 2020

      Thank you for the question. You are correct; in this context, "precaution" refers to private precautionary measures. During the survey, companies were asked whether they implemented a selection of private precautionary measures prior to the damaging flood or not. Based on previous analyses, we extracted a subset from the queried precautionary measures and computed a ratio to assess the flood precaution of the respective company. The eight measures, from which we computed the precaution ratio, can be classified into adaption, mitigation, and emergency measures.

      Classification Precautionary measure
      Adaption Adapted use of flood-prone area
      Adaption Relocation of susceptible equipment
      Mitigation Improve flood resilience of building; e.g. basement waterproofing
      Mitigation Installation of water barriers
      Emergency Saving equipment / saving goods and stock
      Emergency Use of water pumps
      Emergency Shut-down of machinery and power
      Emergency Preventing contamination

      The precaution ratio is the number of measures a company implemented prior to the flood divided by the number of measures that this company could have potentially implemented (some of the measures are not relevant for all companies). Hence, the predictor variable “precaution” ranges from zero (i.e., no relevant measure implemented) to one (i.e., all relevant measures implemented).

       

  • CC2: Comment on EGU2020-7334, Antara Dasgupta, 06 May 2020

    Hi Lukas, thanks for responding my query in the chat. The chat is a bit fast paced and hard to keep up with for answering all the questions, you did really well nonetheless. :)

    I leave my last question here for you to respond to as per your own convenience.

    Can you please elaborate a bit on what is meant by flood experience and how it can be quantified?  

    • AC2: Reply to CC2, Lukas Schoppa, 06 May 2020

      Hello Antara,

      indeed, it was hard to keep up with the chat. Thank you for posing your question again as a comment.

      I would like to add one more point why we did not include flow velocity as a predictor (In addition to constraints due to the number and accuracy of the survey answers on flow velocity). We developed the flood loss models for fluvial flooding and, hence, assume that omitting flow velocity is not as decisive as it would be for flash floods.

      As for the flood experience, we used the number of flood events that a company experienced prior to the damaging/surveyed event as a proxy, that is:
      no, one, two, three, four, and five or more previous floods
      Thus, flood experience is a ordinal predictor in our study.

       

       

       

       

      • CC5: Reply to AC2, Antara Dasgupta, 06 May 2020

        Right, all of this makes sense now. This is a very interesting study in my opinion (I work mostly with flood forecasting and remote sensing data assimilation), I look forward to seeing your paper on this! :)

         

         

        • AC5: Reply to CC5, Lukas Schoppa, 06 May 2020

          Great. Thanks a lot for your interest and the positive feedback!

          The manuscript to this work is currently under review in Water Resources Research and hopefully will be available soon.

  • CC3: Question about predictors, Yue Zhang, 06 May 2020

    Dear Lukas,

    This is an interesting and valuable study. I have one question about these predictors. Are there any correlations between them? If so, will these correlations have negative impacts on the robustness of three models? 


    These statistical models (BN, BR and RF) are kind of like black boxes which give us results but the robustness of models depends a lot on inputs. So I am curious about the sensitivity of models to predictors.

    Thank you very much.

    Yue Zhang 
    Beijing Normal University

    • AC3: Reply to CC3, Lukas Schoppa, 06 May 2020

      Hello Yue,

      Thank you for your question!

      Yes, some of the predictors are correlated; yet, rather weakly. The strongest correlation exists between the flood experience and the precaution (~0.54).

      I think that the effect of this collinearity differs across the models. For instance, Bayesian networks require some correlation among the variables in order to establish the network structure.

      I assume that the correlations affect measures of predictor importance, which provide some insights on the model functionalities, rather than the predictive performance itself. However, we did not investigate this with a systematic sensitivity analysis yet. From my experience in developing the models, I got the impression that varying the predictor set indeed influences model performance. However, only for variables with high predictor importance.

      This definitely is an aspect worth studying in further work on the models.

      • CC6: Reply to AC3, Yue Zhang, 06 May 2020

        Hello Lukas,

        Thank you for your reply. Again, this study is really interesting.

  • CC4: Comment on EGU2020-7334, Chiara Arrighi, 06 May 2020

    Hi Lukas,

    how do you estimate exposure values for companies ?

    • AC4: Reply to CC4, Lukas Schoppa, 06 May 2020

      Hello Chiara,

      Thank you for your question!

      The exposure values of the company assets and the monetary damage were queried in the surveys on the object-level. That is, the respondents (mostly the company owners) provided this information.

      For modelling, we used relative loss instead of absolute values.

Presentation version 1 – uploaded on 02 May 2020 , no comments