the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Jun 2023
Windstorm damage relations – Assessment of storm damage functions in complex terrain
Abstract. Extreme winds are by far the largest contributor to Norway’s insurance claims related to natural hazards. The predictive skills of four different damage functions are assessed for Norway at the municipality and national levels on daily and annual temporal scales using municipality level insurance data and the high-resolution Norwegian reanalysis (NORA3) wind speed data for the period 1985–2020. Special attention is given to extreme damaging events and occurrence probabilities of wind speed induced damages. Because of the complex topography of Norway and the resulting high heterogeneity of the population density, the wind speed is weighted with population. The largest per-capita losses and severe damages occur most frequently in the western municipalities of Norway whilst there are seldom any large losses further inland. The good agreement between the observed and estimated losses at municipality and national levels suggests that the damage functions used in this study are well suited for estimating severe wind storm-induced damages. Furthermore, the damage functions are able to successfully reconstruct the geographical pattern of losses caused by extreme windstorms with a high degree of correlation. From event occurrence probabilities, the present study devises a damage classifier that distinguishes between daily damaging and non-damaging events at the municipality level. While large loss events are well captured, the skewness and zero-inflation of the loss data greatly reduces the quality of both the damage functions and the classifier for moderate and weak loss events.
- Preprint
(14138 KB) - Metadata XML
-
Supplement
(13241 KB) - BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on nhess-2023-90', Anonymous Referee #1, 28 Jul 2023
The paper investigates the skill of four storm damage functions to reproduce loss from windstorm events for the complex topography of Norway at both municipality and national level. The damage functions use insurance data and the high-resolution reanalysis dataset NORA3 for the period 1985-2020. The authors show that all damage functions are able to reproduce extreme loss events.
The paper covers an interesting and relevant topic that could be of interest to NHESS readers. However, major concerns regarding the data and the adopted methods (see main points below) need to be addressed before considering the paper suitable for publication. Since these concerns are rather substantial, I suggest rejecting the manuscript at this stage, but would encourage the authors to resubmit it once it has been revised.
Main points
- GENERAL
- Are your results representative/transferable to other regions with complex orography? If not, you should mention Norway in the title.
- The study needs to be motivated more clearly in the introduction.
- The review paper by Gliksman et al. (2023) provides a good overview on the topic and should therefore be included.
- DATA
- Why did you choose the year 2015 to adjust the insurance loss for inflation?
- Is the population data gridded (line 123) or at municipality level (line 95)?
- You have adjusted the loss for inflation and then talk about the zero inflation of the loss time series. This is quite misleading.
- It is unclear which wind speed and gust data you extracted from NORA3:
- Line 113: time steps 4-9h
- Line 113: hourly wind speed
- Line 114: daily maximum near-surface wind speed.
- METHODS
- It is not clear how the estimation of municipality level wind speed works.
- Why did you not calculate the storm damage by grid point and then aggregate it by region (municipality or national level), as Pinto et al. (2012) or Karremann et al. (2014a) did, for example? These studies also include a weighting with population density.
- Exponential model: Why did you chose the 95th percentile and not a similar threshold as in the other models? It seems too low to assess extreme events. Why is it necessary/useful to bin the loss with respect to wind speed?
- Klawa model: Why do they use the 98th percentile? Please explain. Have you checked whether this threshold is suitable for Norway (see Karremann et al., 2014b)?
- Why are you interested in different loss classes? Losses are primarily caused by gusts above a certain threshold.
- Modified Prahl model: Why did you choose this particular modification?
- The concept of the damage classifier is not clear. What added value does it offer compared to the exponential or Klawa model, which already classify events/non-events based on wind speed thresholds?
- If you focus on daily losses (line 238), how do you account for longer lasting events? Which day of the storm event is selected as the ‘loss day’?
- FIGURES / TABLES
- In total, 14 figures and 3 tables are too many to include in the main manuscript. You should select the most relevant ones to convey your key findings and move the rest to the Supplementary Material.
- In your spatial maps, you use the same colours for loss, wind speed, errors, thresholds, etc. Consider using different colormaps to depict the spatial patterns. This could help the reader better understand what is shown and clearly differentiate the different type of plots.
References
Gliksman D et al. (2023) Review Article: A European Perspective on wind and storm damage – from the meteorological background to index-based approaches to assess impacts. NHESS 23, https://doi.org/10.5194/nhess-23-2171-2023
Karremann MK, Pinto JG, von Bomhard PJ and Klawa M (2014a) On the clustering of winter storm loss events over Germany. NHESS 14, https://doi.org/10.5194/nhess-14-2041-2014
Karremann MK, Pinto JG, Reyers M and Klawa M (2014b) Return periods of losses associated with European windstorm series in a changing climate. ERL 9, https://doi.org/10.1088/1748-9326/9/12/124016
Pinto JG, Karremann MK, Born K, Della-Marta PM and Klawa M (2012) Loss potentials associated with European windstorms under future climate conditions. Clim Res 54, https://doi.org/10.3354/cr01111
Citation: https://doi.org/10.5194/nhess-2023-90-RC1 - AC1: 'Reply on RC1', Ashbin Jaison, 15 Sep 2023
- GENERAL
-
RC2: 'Comment on nhess-2023-90', Anonymous Referee #2, 04 Aug 2023
This paper compares the performance of four different damage functions in estimating windstorm losses over Norway. The damage functions are assessed at municipality and national spatial scales, and daily and annual time scales. All damage functions can reproduce the spatial loss patterns of the most extreme storms, and when aggregated nationally have high temporal correlations with observations on daily timescales. Time series of national aggregate losses on annual timescales correlate well with the deterministic damage models, but the probabilistic models produce very large errors for some years. Using the probability of damage function from the probabilistic models, a damage classifier is developed to distinguish between damaging and non-damaging wind speeds at the municipality level. The classifier has a low hit rate when considering all events, but does predict the most extreme events.
I think this is a worthwhile study and the paper has some interesting results. However, I have a few major concerns that need addressing before I can recommend publication. Detailed comments are listed below.
Major comments:
- I think it needs to be emphasised that the damage functions were developed to work on different spatial and temporal scales. For example, the Klawa and Ulbrich (2003) cubic model was originally used for annually and nationally aggregated data, whereas the Prahl et al (2012) model was applied to smaller scale daily data (hence the need to be stochastic). It is surprising that the cubic model works so well on the municipality scale.
- As you mention there are zeros in the daily municipality data, so for the linear regression step for the exponential and cubic models (L(ν) = β0 + β1d(ν)) the residuals will not be Gaussian. It’s not clear to me how you dealt with the zeros – were they included in the fit? Was the data binned for the cubic model as well as the exponential one?
- For the cubic model, in this paper each municipality is fitted separately, whereas in the Klawa & Ulbrich (2003) they aggregate the data nationally (and annually) then apply the linear regression fit. How much does the regression fit vary between municipalities? i.e. did you prove it’s necessary to fit each region separately?
- Klawa & Ulbrich (2003) only fitted above the 98th percentile assuming that no damage occurred below this. How good an approximation is this for the Norwegian data? From fig 1 there is clearly damage below the 98th percentile but hard to tell how significant this is because I assume there is a very high proportion of days with zero loss at lower wind speeds.
- L214: It’s confusing to define the evaluation of damage classifier here, before you’ve defined the damage classifier. Maybe put section 2.6 before 2.5?
- L256 and L326: “The choice of wind data has the potential to influence the performance of the damage functions…” and “The predictive performances of the damage function and the damage classifier confirms the importance of weighting wind speed with population for better performance of the damage functions.” In the paper you show that the population weighting gives different maximum wind speeds for the municipalities (Fig 7b), but you don’t actually show that it makes the damage models perform better. Does it?
- In the abstract and conclusions (L324) it’s stated that all models perform well on national scales, but from Fig S4 it looks like the probabilistic models perform very poorly on this scale.
Minor comments:
- L103 “Figure 1 highlights a record high number of claims in years 1992, 2011 and 2015. This can be attributed to the New Year Storm in 1992, storm Dagmar in 2011 and storms Nina and Ole in 2015 (Table 1).” In the figure it looks like there’s high loss in 1994, not 1992.
- Fig 2: How do you have zero on a log scale (y-axis)? It looks like zero losses are not actually plotted so this should be stated.
- Section 2.5, L204 “For robust storm-damage relations, extreme care should be taken in the parameters estimation of damage functions. To ensure robustness of the damage functions, we bin the loss data with respect to wind speeds to eliminate the sensitivity of damage functions to extreme events.” I’m not sure what you mean by this. Are you talking about the binning done when fitting the parameters, or do you bin the data when evaluating the errors as well?
- Section 3.3: Since storms last more than one day, how did you estimate the losses for a single storm? e.g. did you sum the losses over a few days? Choose the maximum loss day?
- Fig 12 caption – what does it mean ‘Annual time series of observed and estimated national losses using the extreme loss class'? Are these not just annually aggregated losses (i.e. the sum of all days?)
- Fig 11 – panels aren’t labelled.
Citation: https://doi.org/10.5194/nhess-2023-90-RC2 - AC2: 'Reply on RC2', Ashbin Jaison, 15 Sep 2023
Status: closed
-
RC1: 'Comment on nhess-2023-90', Anonymous Referee #1, 28 Jul 2023
The paper investigates the skill of four storm damage functions to reproduce loss from windstorm events for the complex topography of Norway at both municipality and national level. The damage functions use insurance data and the high-resolution reanalysis dataset NORA3 for the period 1985-2020. The authors show that all damage functions are able to reproduce extreme loss events.
The paper covers an interesting and relevant topic that could be of interest to NHESS readers. However, major concerns regarding the data and the adopted methods (see main points below) need to be addressed before considering the paper suitable for publication. Since these concerns are rather substantial, I suggest rejecting the manuscript at this stage, but would encourage the authors to resubmit it once it has been revised.
Main points
- GENERAL
- Are your results representative/transferable to other regions with complex orography? If not, you should mention Norway in the title.
- The study needs to be motivated more clearly in the introduction.
- The review paper by Gliksman et al. (2023) provides a good overview on the topic and should therefore be included.
- DATA
- Why did you choose the year 2015 to adjust the insurance loss for inflation?
- Is the population data gridded (line 123) or at municipality level (line 95)?
- You have adjusted the loss for inflation and then talk about the zero inflation of the loss time series. This is quite misleading.
- It is unclear which wind speed and gust data you extracted from NORA3:
- Line 113: time steps 4-9h
- Line 113: hourly wind speed
- Line 114: daily maximum near-surface wind speed.
- METHODS
- It is not clear how the estimation of municipality level wind speed works.
- Why did you not calculate the storm damage by grid point and then aggregate it by region (municipality or national level), as Pinto et al. (2012) or Karremann et al. (2014a) did, for example? These studies also include a weighting with population density.
- Exponential model: Why did you chose the 95th percentile and not a similar threshold as in the other models? It seems too low to assess extreme events. Why is it necessary/useful to bin the loss with respect to wind speed?
- Klawa model: Why do they use the 98th percentile? Please explain. Have you checked whether this threshold is suitable for Norway (see Karremann et al., 2014b)?
- Why are you interested in different loss classes? Losses are primarily caused by gusts above a certain threshold.
- Modified Prahl model: Why did you choose this particular modification?
- The concept of the damage classifier is not clear. What added value does it offer compared to the exponential or Klawa model, which already classify events/non-events based on wind speed thresholds?
- If you focus on daily losses (line 238), how do you account for longer lasting events? Which day of the storm event is selected as the ‘loss day’?
- FIGURES / TABLES
- In total, 14 figures and 3 tables are too many to include in the main manuscript. You should select the most relevant ones to convey your key findings and move the rest to the Supplementary Material.
- In your spatial maps, you use the same colours for loss, wind speed, errors, thresholds, etc. Consider using different colormaps to depict the spatial patterns. This could help the reader better understand what is shown and clearly differentiate the different type of plots.
References
Gliksman D et al. (2023) Review Article: A European Perspective on wind and storm damage – from the meteorological background to index-based approaches to assess impacts. NHESS 23, https://doi.org/10.5194/nhess-23-2171-2023
Karremann MK, Pinto JG, von Bomhard PJ and Klawa M (2014a) On the clustering of winter storm loss events over Germany. NHESS 14, https://doi.org/10.5194/nhess-14-2041-2014
Karremann MK, Pinto JG, Reyers M and Klawa M (2014b) Return periods of losses associated with European windstorm series in a changing climate. ERL 9, https://doi.org/10.1088/1748-9326/9/12/124016
Pinto JG, Karremann MK, Born K, Della-Marta PM and Klawa M (2012) Loss potentials associated with European windstorms under future climate conditions. Clim Res 54, https://doi.org/10.3354/cr01111
Citation: https://doi.org/10.5194/nhess-2023-90-RC1 - AC1: 'Reply on RC1', Ashbin Jaison, 15 Sep 2023
- GENERAL
-
RC2: 'Comment on nhess-2023-90', Anonymous Referee #2, 04 Aug 2023
This paper compares the performance of four different damage functions in estimating windstorm losses over Norway. The damage functions are assessed at municipality and national spatial scales, and daily and annual time scales. All damage functions can reproduce the spatial loss patterns of the most extreme storms, and when aggregated nationally have high temporal correlations with observations on daily timescales. Time series of national aggregate losses on annual timescales correlate well with the deterministic damage models, but the probabilistic models produce very large errors for some years. Using the probability of damage function from the probabilistic models, a damage classifier is developed to distinguish between damaging and non-damaging wind speeds at the municipality level. The classifier has a low hit rate when considering all events, but does predict the most extreme events.
I think this is a worthwhile study and the paper has some interesting results. However, I have a few major concerns that need addressing before I can recommend publication. Detailed comments are listed below.
Major comments:
- I think it needs to be emphasised that the damage functions were developed to work on different spatial and temporal scales. For example, the Klawa and Ulbrich (2003) cubic model was originally used for annually and nationally aggregated data, whereas the Prahl et al (2012) model was applied to smaller scale daily data (hence the need to be stochastic). It is surprising that the cubic model works so well on the municipality scale.
- As you mention there are zeros in the daily municipality data, so for the linear regression step for the exponential and cubic models (L(ν) = β0 + β1d(ν)) the residuals will not be Gaussian. It’s not clear to me how you dealt with the zeros – were they included in the fit? Was the data binned for the cubic model as well as the exponential one?
- For the cubic model, in this paper each municipality is fitted separately, whereas in the Klawa & Ulbrich (2003) they aggregate the data nationally (and annually) then apply the linear regression fit. How much does the regression fit vary between municipalities? i.e. did you prove it’s necessary to fit each region separately?
- Klawa & Ulbrich (2003) only fitted above the 98th percentile assuming that no damage occurred below this. How good an approximation is this for the Norwegian data? From fig 1 there is clearly damage below the 98th percentile but hard to tell how significant this is because I assume there is a very high proportion of days with zero loss at lower wind speeds.
- L214: It’s confusing to define the evaluation of damage classifier here, before you’ve defined the damage classifier. Maybe put section 2.6 before 2.5?
- L256 and L326: “The choice of wind data has the potential to influence the performance of the damage functions…” and “The predictive performances of the damage function and the damage classifier confirms the importance of weighting wind speed with population for better performance of the damage functions.” In the paper you show that the population weighting gives different maximum wind speeds for the municipalities (Fig 7b), but you don’t actually show that it makes the damage models perform better. Does it?
- In the abstract and conclusions (L324) it’s stated that all models perform well on national scales, but from Fig S4 it looks like the probabilistic models perform very poorly on this scale.
Minor comments:
- L103 “Figure 1 highlights a record high number of claims in years 1992, 2011 and 2015. This can be attributed to the New Year Storm in 1992, storm Dagmar in 2011 and storms Nina and Ole in 2015 (Table 1).” In the figure it looks like there’s high loss in 1994, not 1992.
- Fig 2: How do you have zero on a log scale (y-axis)? It looks like zero losses are not actually plotted so this should be stated.
- Section 2.5, L204 “For robust storm-damage relations, extreme care should be taken in the parameters estimation of damage functions. To ensure robustness of the damage functions, we bin the loss data with respect to wind speeds to eliminate the sensitivity of damage functions to extreme events.” I’m not sure what you mean by this. Are you talking about the binning done when fitting the parameters, or do you bin the data when evaluating the errors as well?
- Section 3.3: Since storms last more than one day, how did you estimate the losses for a single storm? e.g. did you sum the losses over a few days? Choose the maximum loss day?
- Fig 12 caption – what does it mean ‘Annual time series of observed and estimated national losses using the extreme loss class'? Are these not just annually aggregated losses (i.e. the sum of all days?)
- Fig 11 – panels aren’t labelled.
Citation: https://doi.org/10.5194/nhess-2023-90-RC2 - AC2: 'Reply on RC2', Ashbin Jaison, 15 Sep 2023
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 550 | 197 | 40 | 787 | 65 | 37 | 37 |
- HTML: 550
- PDF: 197
- XML: 40
- Total: 787
- Supplement: 65
- BibTeX: 37
- EndNote: 37
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1