• Arlot, S., , and Celisse A. , 2010: A survey of cross-validation procedures for model selection. Stat. Surv.,4, 40–79, doi:10.1214/09-SS054.

  • Chang, C.-P., , Yang Y.-T. , , and Kuo H.-C. , 2013: Large increasing trend of tropical cyclone rainfall in Taiwan and the roles of terrain. J. Climate, 26, 41384147, doi:10.1175/JCLI-D-12-00463.1.

    • Search Google Scholar
    • Export Citation
  • Czajkowski, J., , Simmons K. , , and Sutter D. , 2011: An analysis of coastal and inland fatalities in landfalling US hurricanes. Nat. Hazards, 59, 15131531, doi:10.1007/s11069-011-9849-x.

    • Search Google Scholar
    • Export Citation
  • Hawkins, D. M., , Basak S. C. , , and Mills D. , 2003: Assessing model fit by cross-validation. J. Chem. Inf. Comput. Sci., 43, 579586, doi:10.1021/ci025626i.

    • Search Google Scholar
    • Export Citation
  • Hebert, C. G., , Weinzapfel R. A. , , and Chambers M. A. , 2010: Hurricane severity index: A new way of estimating a tropical cyclone’s destructive potential. Impactweather, Inc., 5 pp.

  • Irish, J. L., , and Resio D. T. , 2010: A hydrodynamics-based surge scale for hurricanes. Ocean Eng., 37, 6981, doi:10.1016/j.oceaneng.2009.07.012.

    • Search Google Scholar
    • Export Citation
  • Jordan, M. R., , and Clayson C. A. , 2008: Evaluating the usefulness of a new set of hurricane classification indices. Mon. Wea. Rev., 136, 52345238, doi:10.1175/2008MWR2449.1.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2006: Time to replace the Saffir–Simpson hurricane scale? Eos, Trans. Amer. Geophys., 87, 36, doi:10.1029/2006EO010003.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2008: Tropical cyclone destructive potential by integrated kinetic energy. Bull. Amer. Meteor. Soc., 89, 219221, doi:10.1175/BAMS-89-2-219.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2013: Classification of hurricanes: Lessons from Katrina, Ike, Irene, Isaac and Sandy. Ocean Eng., 70, 124128, doi:10.1016/j.oceaneng.2013.06.007.

    • Search Google Scholar
    • Export Citation
  • Matyas, C. J., , and Cartaya M. , 2009: Comparing the rainfall patterns produced by Hurricanes Frances (2004) and Jeanne (2004) over Florida. Southeast. Geogr., 49, 132156, doi:10.1353/sgo.0.0046.

    • Search Google Scholar
    • Export Citation
  • Peduzzi, P., , Chatenoux B. , , Dao H. , , De Bono A. , , Herold C. , , Kossin J. , , Mouton F. , , and Nordbeck O. , 2012: Global trends in tropical cyclone risk. Nat. Climate Change, 2, 289294, doi:10.1038/nclimate1410.

    • Search Google Scholar
    • Export Citation
  • Pielke, R. A., Sr., , and Pielke R. A. Jr., 1997: Hurricanes: Their Nature and Impacts on Society. John Wiley and Sons, 298 pp.

  • Powell, M. D., , and Reinhold T. A. , 2007: Tropical cyclone destructive potential by integrated kinetic energy. Bull. Amer. Meteor. Soc., 88, 513526, doi:10.1175/BAMS-88-4-513.

    • Search Google Scholar
    • Export Citation
  • Rappaport, E. N., 2000: Loss of life in the United States associated with recent Atlantic tropical cyclones. Bull. Amer. Meteor. Soc., 81, 20652073, doi:10.1175/1520-0477(2000)081<2065:LOLITU>2.3.CO;2.

    • Search Google Scholar
    • Export Citation
  • Rappaport, E. N., 2014: Fatalities in the United States from Atlantic tropical cyclones: new data and interpretation. Bull. Amer. Meteor. Soc., 95, 341346, doi:10.1175/BAMS-D-12-00074.1.

    • Search Google Scholar
    • Export Citation
  • Saffir, H. S., 1977. Design and construction requirements for hurricane resistant construction. American Society of Civil Engineers Preprint 2830, 20 pp.

  • View in gallery

    Tracks of Hurricane Jeanne and Hurricane Frances (both in 2004; http://www.csc.noaa.gov/hurricanes).

  • View in gallery

    A comparison of the wind field (kt) from (left) category 3 Hurricane Jeanne and (right) category 2 Hurricane Frances. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

  • View in gallery

    The correlation of four existing indices vs damage (2013 billion USD; solid dot) and death toll (solid triangle).

  • View in gallery

    Leading causes of hurricane deaths in the United States during 1970–99. [Data from Rappaport (2000).]

  • View in gallery

    Hurricane Ike (2008) wind field (kt) at landfall, generated by H*Wind. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

  • View in gallery

    Hurricane Sandy wind field (kt) at landfall, generated by H*Wind. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

  • View in gallery

    Correlation and regression coefficients for all three hazard indices against (left) damage (billion USD) and (right) death toll considering all 19 hurricanes. Equation (7) is calculated with a = b = c = d = 1.

  • View in gallery

    Correlation and regression coefficients for all three hazards indices against (left) damage (billion USD) and (right) death toll excluding Hurricane Katrina and Hurricane Sandy.

  • View in gallery

    Hurricanes ranked according to their (left) actual damage (billion USD) and (right) death toll, and the respective SSHS and TCHI rankings. Black bars indicate Hurricanes Frances, Jeanne, and Ike.

  • View in gallery

    Hurricanes ranked (including Sandy and Katrina) according to their actual damage (billion USD) and death toll, and the respective SSHS and TCHI rankings.

  • View in gallery

    Hurricanes ranked (including Sandy and Katrina) using the individual hazard subindices.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 339 339 46
PDF Downloads 256 256 39

Classification of Hurricane Hazards: The Importance of Rainfall

View More View Less
  • 1 School of Civil Engineering, University of Queensland, Brisbane, Queensland, Australia
© Get Permissions
Full access

Abstract

A new hazard index is presented to estimate and rank hurricane severity according to a storm’s damage and death toll after landfall on the continental United States. The index uses three characteristic meteorological aspects of hurricanes: wind, torrential rainfall, and storm surge, each with an individual subindex. Rainfall is identified as an important and frequently dominant hazard in terms of damage and death toll, but is not included in any current hazard scales or indices. The new rainfall subindex adopts rainfall intensity, storm rainfall area, and the forward speed of the system to estimate the rainfall hazard. The new hazard index, applied to recent U.S. hurricanes (2003–12), has better skill than existing scales in terms of ranking the severity of the events by both damage and death toll. Further, the index can provide good quantitative estimates of dollar values for damage and death toll, whereas previous models provide only a scale or ranking. The index provides a basis for improved hazard planning and emergency response, and may also be useful for insurance and risk management processes.

Corresponding author address: Mehdi Rezapour, School of Civil Engineering, University of Queensland, St Lucia, QLD 4067, Australia. E-mail: m.rezapour@uq.edu.au

Abstract

A new hazard index is presented to estimate and rank hurricane severity according to a storm’s damage and death toll after landfall on the continental United States. The index uses three characteristic meteorological aspects of hurricanes: wind, torrential rainfall, and storm surge, each with an individual subindex. Rainfall is identified as an important and frequently dominant hazard in terms of damage and death toll, but is not included in any current hazard scales or indices. The new rainfall subindex adopts rainfall intensity, storm rainfall area, and the forward speed of the system to estimate the rainfall hazard. The new hazard index, applied to recent U.S. hurricanes (2003–12), has better skill than existing scales in terms of ranking the severity of the events by both damage and death toll. Further, the index can provide good quantitative estimates of dollar values for damage and death toll, whereas previous models provide only a scale or ranking. The index provides a basis for improved hazard planning and emergency response, and may also be useful for insurance and risk management processes.

Corresponding author address: Mehdi Rezapour, School of Civil Engineering, University of Queensland, St Lucia, QLD 4067, Australia. E-mail: m.rezapour@uq.edu.au

1. Introduction

Appropriate warning of tropical cyclone severity and timely evacuation from hazards zones are the key to reducing damage and loss of life as these systems make landfall. Here, we use the term tropical cyclone (TC) as a generic term for hurricanes, typhoons, or severe tropical storms. The risk of living and the cost of damage in TC-prone areas are increasing due to growing population and wealth (Peduzzi et al. 2012), and it is therefore crucial to identify, understand, and forecast TC systems to reduce vulnerability to the local population and infrastructure.

This is the fourth decade in which the Saffir–Simpson hurricane scale (SSHS) has been used as a convenient method for meteorological scientists and decision makers to categorize hurricanes based on surface wind speed. However, the 2004 Atlantic hurricane season showed that the SSHS does not consistently estimate the true destructive potential of hurricanes (Hebert et al. 2010; Jordan and Clayson 2008; Kantha 2006). For example, category 2 Hurricane Frances (2004) caused 12 billion U.S. dollars (USD) worth of damage and led to 49 fatalities, compared with category 3 Hurricane Jeanne (2004), with 7 billion USD worth of damage and 5 fatalities, in the United States, even though both made landfall at almost the same location, just 2 miles apart in Florida (Matyas and Cartaya 2009; Table 1, Figs. 1 and 2). Further, other existing hurricane scales also do not classify accurately the impact of these hurricanes (Table 1). This suggests an opportunity to develop a new tropical cyclone hazard index to address these deficiencies.

Table 1.

A comparison of the characteristics and existing hazard scales for Hurricane Frances (2004) and Hurricane Jeanne (2004). Radius of max wind is given in nautical miles (n mi; 1 n mi = 1.852 km).

Table 1.
Fig. 1.
Fig. 1.

Tracks of Hurricane Jeanne and Hurricane Frances (both in 2004; http://www.csc.noaa.gov/hurricanes).

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

Fig. 2.
Fig. 2.

A comparison of the wind field (kt) from (left) category 3 Hurricane Jeanne and (right) category 2 Hurricane Frances. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

Injury, property damage, and the loss of life and economic activity due to a particular tropical cyclone result from the interaction between a tropical cyclone and the land impacted by the tropical cyclone. Consequently, in recent decades, a range of different scales and indices have been proposed. Kantha (2006, 2008, 2013) suggested that a continuous scale with a dynamical basis would allow emergency response agencies to make better evacuation decisions than are possible with the discrete and rather arbitrary SSHS. Further, Kantha (2006, 2013) asserted that storm size, which is neglected in the SSHS, is an important factor governing cyclone hazards. Three new indices were developed to address perceived deficiencies in the SSHS—the hurricane intensity index (HII), the hurricane hazard index (HHI), and the hurricane surge index (HSI):
eq1
eq2
e1
where is the maximum sustained wind velocity, is a reference value (74 mi h−1), R is the radius of hurricane-force winds [defined as the radius within which the wind speed exceeds 34 knots (kt; 1 kt = 0.51 m s−1)], and is a reference value, equal to 60 mi.

The HHI included storm size for the first time by considering the radius of hurricane-force winds. The HSI was developed to categorize hurricanes on the basis of the resulting storm surge. To test these relationships, Jordan and Clayson (2008) developed a database of past hurricanes making landfall in the United States and calculated the Kantha (2006) indices corresponding to each hurricane’s individual parameters. Jordan and Clayson (2008) concluded that the HII is not significantly better than the SSHS in terms of accuracy for end users and that the HHI and HSI needed to take into account further parameters (e.g., bathymetry and topography) to provide reliable damage and surge forecasts.

Hebert et al. (2010) developed an alternative methodology that considers the size of storms for calculating a hurricane severity index (HSI*, not to be confused with the hurricane surge index). The HSI* is a 50-point scale, with 25 points contributed by the tropical cyclone intensity and 25 points contributed by the size of the wind field. However, as illustrated in Table 1, HSI* does not classify hurricanes accurately in some cases [e.g., Hurricane Frances and Hurricane Jeanne (2004)] because rainfall is not considered in this model (Table 1). A different approach for assessing the destructive potential of tropical cyclones was proposed by Powell and Reinhold (2007), through an integrated kinetic energy model (IKE). Powell and Reinhold (2007) asserted that the IKE is more relevant to damage by wind, storm surge, and waves. The IKE is computed from the surface wind field by integrating the 10-m-level kinetic energy per unit volume over the portions of the storm domain volume Vol with sustained surface wind speed u:
e2
where is air density and are volume elements. Powell and Reinhold (2007) suggested ranking hurricanes with two scales: one for the destructive potential of the wind and the other for the destructive potential of the storm surge and waves . This approach was questioned by Kantha (2008), who argued that structural damage is proportional to the rate of work done by the wind and is not proportional to the energy.

The performance of existing scales for selected U.S. hurricanes is illustrated in Fig. 3, which plots the hazard scale versus the actual damage and death toll. All damage values are corrected for inflation and converted to USD in 2013. It is clear that there is not a well-defined trend or correlation between the existing scales and damage and death toll.

Fig. 3.
Fig. 3.

The correlation of four existing indices vs damage (2013 billion USD; solid dot) and death toll (solid triangle).

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

A key impact of tropical cyclones and hurricanes is clearly missing from all these scales: rainfall. This omission is likely to reduce the accuracy and usefulness of such classifications for hazard identification. For example, Rappaport (2000) shows that 59% of hurricane deaths in the United States over the period 1970–90 result from freshwater flooding (Fig. 4). Rappaport (2014) repeated the same assessment over a longer period, from 1963 to 2012. For that time period, the proportion of fatalities caused by excessive rainfall is significant at 27%, but secondary to storm surge, which is responsible for 49% of fatalities. However, the change arises principally because of the large number of fatalities in Hurricane Katrina, which is clearly an outlier in the data, as will be illustrated later. Rappaport (2014) also shows that the most common cause of fatalities is rainfall; in almost half of the deadly TCs there is at least one death from freshwater flooding.

Fig. 4.
Fig. 4.

Leading causes of hurricane deaths in the United States during 1970–99. [Data from Rappaport (2000).]

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

This paper addresses this issue and develops a new methodology to include rainfall in a tropical cyclone hazard index. The proposed new hazard index includes key rainfall characteristics (e.g., the intensity, duration, and distribution), plus the wind and surge characteristics that are incorporated in the existing scales.

2. Methodology and databases

Although the aim is that ultimately the methodology can be developed worldwide to assess hazards from TCs and hurricanes, initially it is necessary to choose a geographical region with sufficient data availability to verify the model approach. For this purpose, the Atlantic hurricane basin has been selected as the geographical test area, and this is further refined to include only systems that made landfall on the continental United States. There are 19 hurricanes that fall into this category within the last decade (2003–12), from Claudette (2003) to Sandy (2012), and for which relevant data exist. The study period is limited to the last decade to ensure that changes in population density and infrastructure are limited as far as possible, such that the data are reasonably stationary. Indeed, the regression model assumes that the underlying physical processes adopted in the regression model do not change over time. If this assumption is not met, the model may fail for future tropical cyclones. The basis for the chosen databases is outlined below. However, for specific applications, alternative databases could be investigated to determine if model accuracy can be improved; the approach would remain the same.

a. Rainfall database

There are a wide range of rainfall databases available from global weather forecasting models [e.g., the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP)], together with some specific tropical cyclone rainfall measuring missions [e.g., Tropical Rainfall Measuring Mission (TRMM)]. However, because a high spatiotemporal resolution is required, plus real-time or near-real-time data combined with global coverage to have a globally useful model, many databases are not suitable. Table 2 summarizes a range of rainfall databases and outlines their scope and limitations. After review of the available databases, the Precipitation Estimation from Remotely Sensed Imagery Using Artificial Neural Network (PERSIANN) database (http://chrs.web.uci.edu), from the Center for Hydrometeorology and Remote Sensing (CHRS), University of California, Irvine, was ultimately selected.

Table 2.

Rainfall databases and their scope and limitations based on coverage, resolution, and availability.

Table 2.

b. Wind field

The H*Wind database (http://www.aoml.noaa.gov/hrd/data_sub/wind.html) from the National Oceanic and Atmospheric Administration/Atlantic Oceanographic and Meteorological Laboratory/Hurricane Research Division (NOAA/AOML/HRD) was adopted for the present study due to its accuracy, high resolution, and availability. The H*WIND swath maps are particularly useful for calculating the area of each wind speed threshold (e.g., 34, 50, 64, and 87 kt). Figure 5 shows H*Wind data for Hurricane Ike (2008) at landfall as an example.

Fig. 5.
Fig. 5.

Hurricane Ike (2008) wind field (kt) at landfall, generated by H*Wind. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

c. Hurricane forward speed

Both the individual wind and rainfall indices discussed below are sensitive to the forward speed of the system. As a result, the accuracy of the calculated velocity is important for the model. The North Atlantic hurricane database (HURDAT; http://www.aoml.noaa.gov/hrd/hurdat/Data_Storm.html) was chosen to calculate the forward speed C of the system.

d. Individual and combined hazard indices

The new combined hazard index includes the three hazards most typical of tropical cyclones: wind, surge, and rainfall. Each is represented by an individual subindex: 1) tropical cyclone wind index (TCWI), 2) tropical cyclone surge index (TCSI), and 3) tropical cyclone rainfall index (TCRI). The final tropical cyclone hazard index (TCHI) is created by combining these three indices to classify tropical cyclones according to their damage rate and death toll.

e. TCWI

Surface wind speed V (Pielke and Pielke 1997; Kantha 2006, 2008; Saffir 1977; Hebert et al. 2010), wind field size A (Hebert et al. 2010; Kantha 2006), and storm forward speed (Kantha 2006) are chosen as the most important factors for the wind index. For two otherwise similar tropical cyclones, it is clear that the system that has a lower forward speed is more hazardous, because it will have more time to inflict damage to property, trees, infrastructure, etc. Therefore, there is an inverse ratio between TCWI and C. Taking these factors together; Eq. (3) presents the function used to estimate the hazard from tropical cyclones due to their wind field characteristics:
e3
where V = 34, 50, 64, and 87 kt are considered as wind speed thresholds. Items with subscript 0 correspond to reference values [e.g., = 40 kt; Pielke and Pielke (1997)] representing a destructive wind speed (the wind speed that initiates damage to property). As shown in Table 3, varies with V (Hebert et al. 2010), while = 11 kt (Kantha 2006).
Table 3.

Wind speed thresholds of (Hebert et al. 2010).

Table 3.

f. TCSI

Surface wind speed (Kantha 2006) is also selected as the dominant determinant of storm surge and 1 is adopted to take into account the effect of wind field size, instead of radius to maximum wind (Kantha 2006) or (Irish and Resio 2010). As illustrated in Fig. 6, tropical cyclone wind fields are often asymmetric and complicated, particularly at landfall. Therefore, the area is selected as a more accurate indicator than radius for assessing the size of the storm system. The importance of the tropical cyclone approach angle θ is not yet fully clear. Since the aim of this study is to show the severity of mesoscale weather systems, the most critical condition has been chosen (Irish and Resio 2010). Taking these factors together, the TCSI is considered to be a function of the following parameters used to estimate the hazard from storm surge:
e4
where is the maximum surface wind speed, is the area surface with speed greater than 34 kt, and the 0 subscript refers to reference values; that is, = 40 kt and is determined from data derived from the storm wind database (H*WIND), and is the average value of for the selected storms.
Fig. 6.
Fig. 6.

Hurricane Sandy wind field (kt) at landfall, generated by H*Wind. The vertical and horizontal axes are degrees lat and lon, respectively. (Source: NOAA/HRD/AOML.)

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

g. TCRI

Floods caused by heavy rainfall are not considered in existing tropical cyclone scales, despite the major hazard that results (Fig. 2). The most important meteorological parameters for rainfall in terms of the hydro-meteorological severity of the system are intensity I, distribution (which is considered to be an area A above an intensity threshold), and duration. Clearly, a storm system with higher rainfall intensity has a potentially greater probability to generate larger floods (other factors being equal). Furthermore, another important factor is the area of the catchment impacted by rainfall. The area encompassed by the intensity threshold has therefore been taken into account, rather than catchment size itself. This is because rainfall may not be distributed over the whole catchment.

The tropical cyclone forward speed, in addition to the rainfall area, determines the duration of the rainfall. Slow-moving cyclones are likely to result in longer rainfall duration (Chang et al. 2013) and, consequently, greater probability of flood. As a result, unlike V and A, the rainfall severity has an inverse ratio with the cyclone forward speed. Taking the above issues together, the TCRI is formulated as
e5
where the subscript 0 again corresponds to references values used to normalize the parameters. Here, is the average rainfall intensity calculated using all the selected hurricane cases, is the average forward speed calculated from HURDAT (using 3000 cases), and is the average area with more than 2 mm h−1 of rainfall.

h. TCHI

Different model formulations, including linear and nonlinear relationships between TCWI, TCRI, TCSI, and the combined TCHI, have been considered, as proposed in Eq. (6). Each individual hazard subindex (i.e., TCWI, TCRI, and TCSI) is calculated using Eqs. (7).
e6
e7a
e7b
e7c

The TCHI has been modeled twice, for both cyclone damage and death toll. Given the complexity and interrelation of the above three indices with the damage and death toll, different linear and nonlinear regression models were tested within standard statistical software [Statistical Package for the Social Sciences (SPSS)] to determine the correlation coefficients in Eqs. (6) and (7). There are three weighted correlation coefficients (A, B, and C), as well as power correlation coefficients (a, b, …, g), which determine the linearity and nonlinearity of the model regression equations [Eqs. (6) and (7)]. Using all the data, a number of different model scenarios were tested to determine the best physically meaningful coefficients. Table 4 shows four such model scenarios, which includes the final two selected for the model. With all coefficients a, b, …, g equal to 1 (first scenario in Table 4), the model regression is linear; otherwise, it is nonlinear (remaining scenario in Table 4). Furthermore, the meteorological parameters adopted in each individual subindex [e.g., , , etc.] have a physical basis, as described above. Consequently, negative values for the correlation coefficients are not sensible. Therefore, model results yielding negative coefficients are not valid, for example, the coefficient C for the damage index in model scenario 1 in Table 4. Discarding nonphysical results to take into account these physical constraints and taking the models with the highest correlation coefficients, the coefficients for scenarios 3 and 4 (Table 4) were selected for the damage and death toll regression equations, respectively.

Table 4.

Example model coefficients. Boldface coefficients were adopted in the final model.

Table 4.

Given the limited number of samples, a cross-validation procedure was adopted to examine the robustness of the model. Consequently, the leave-one-out cross-validation (LOOCV) method was applied repeatedly to the dataset to examine the variance in model parameters, goodness of fit, and errors in model predictions (Hawkins et al. 2003; Arlot and Celisse 2010). The last four rows in Table 4 show the average of the model coefficients and the variance of the coefficients obtained from the LOOCV procedure. These show an acceptable consistency with the optimum damage and death toll coefficient obtained from the total dataset (the boldface scenarios in Table 4), and nearly identical R2 values.

3. Results

First, for all 19 hurricanes, the individual hazard indices are plotted versus damage and death toll using Eq. (7) calculated with a = b = c = d = 1 (Fig. 7). There is no significant correlation, with R2 = 0.0415 and 0.0945 for TCRI versus damage and death toll, respectively. In Fig. 7, dots in the ovals corresponding to Hurricane Katrina (2005) and Hurricane Sandy (2012) are responsible in large part for the poor regression. The damage and fatalities caused by these two events are significantly different from the rest of the population and dependent on the land characteristics. These two hurricanes made landfall in very densely populated areas, including major cities, leading to greater hazards exposure and therefore higher loss of life and greater damage. Consequently, Hurricane Katrina and Hurricane Sandy were removed from the database and modeling proceeded with the remaining 17 hurricanes, yielding the model coefficients in Table 4. Model performance for Hurricane Katrina and Hurricane Sandy is further discussed later.

Fig. 7.
Fig. 7.

Correlation and regression coefficients for all three hazard indices against (left) damage (billion USD) and (right) death toll considering all 19 hurricanes. Equation (7) is calculated with a = b = c = d = 1.

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

Figure 8 shows the resulting regression between the four hazard indices and the damage and death toll. Of the individual subindices, TCRI yields the best correlation for both damage and death toll, with regression values of 0.43 and 0.74, respectively. This is consistent with Rappaport (2000) and Czajkowski et al. (2011), who showed that fatalities caused by freshwater flooding are dominant in comparison with wind or coastal flooding. The combined TCHI has regression coefficients of 0.81 and 0.88 for damage and death toll, respectively. This compares very favorably with the regression coefficients for SSHS and HII, HHI, and HSI* shown in Fig. 3. This suggests the new indices more accurately represent the physical processes causing damage and fatalities.

Fig. 8.
Fig. 8.

Correlation and regression coefficients for all three hazards indices against (left) damage (billion USD) and (right) death toll excluding Hurricane Katrina and Hurricane Sandy.

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

To assess this further, the severity of each hurricane is ranked according to the actual damage and death toll. This is plotted in Fig. 9, together with the rank predicted by the SSHS and the TCHI. As illustrated in Fig. 9, the TCHI ranks the severity of each hurricane significantly more accurately than the SSHS. This demonstrates that the new index can classify hurricanes more consistently than the existing scales, which are based only on wind and/or storm surge. The results show that, in both the damage and death toll models, the rainfall subindex (TCRI) is dominant in comparison with the wind and storm surge indices (i.e., parameter B in Table 4, compared with parameters A and C). Again, this is consistent with the data from Rappaport (2000, 2014) discussed earlier.

Fig. 9.
Fig. 9.

Hurricanes ranked according to their (left) actual damage (billion USD) and (right) death toll, and the respective SSHS and TCHI rankings. Black bars indicate Hurricanes Frances, Jeanne, and Ike.

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

In addition, the new model (TCHI) can also provide estimates of the damage and the number of fatalities likely to be caused by a given hurricane, which is not possible using the SSHS. The model can estimate the damage and death toll with root-mean-square errors of 4.2 billion USD and 13 people, respectively. The RMSE between the hurricane damage for each event and the mean damage value (9.8 billion USD) was also calculated. This demonstrates that the new index (RMSE 4.2 billion USD) has less than half the error in estimating the damage value compared to taking the mean value (9.88 billion USD). This is a significant improvement. Thus, if the TCHI (Table 5) can be developed within a forecasting model, the model has the potential to enable improved hazard planning and emergency response. In hindcast mode with the present data, the model may also be useful for improving insurance and risk management tools since it provides an improved correlation between TC characteristics and impacts by incorporating the physical process that causes the most damage and greatest number of fatalities. Further, Table 5 demonstrates that the new index is able to rank existing hurricanes by damage and death toll with an error of about 2 in rank position, compared with an error in rank position of roughly 6 using the SSHS. Specifically, the model ranks Hurricanes Frances and Jeanne (black boldface bars in Fig. 9) much more accurately than does the SSHS, which ranks the hazard from those systems in reverse compared to their actual impacts. Likewise, the SSHS categorizes Hurricane Ike (Fig. 9) at rank position 7 for both damage and death toll, while in fact Hurricane Ike was the most costly and the second most deadly hurricane in the selected dataset. Consequently, in summary, the TCHI provides far better estimates of severity (rank) than does the SSHS, while also providing good quantitative estimates of the actual damage and death toll.

Table 5.

TCHI model accuracy compared with the SSHS.

Table 5.

4. Discussion

In cases of TC landfalls, the damage and casualties arise from two aspects. First are the meteorological characteristics of the storm such as surface wind speed, storm size, rainfall intensity, storm forward speed, surge, etc. Second are the characteristics of the land impacted, including hydrological factors (e.g., catchment size, catchment slope, and land cover), as well as societal factors (e.g., population and property density, risk awareness, and extent of disaster management planning). The new indices proposed here focus on the meteorological aspects, and consider for the first time all of the main hazards of tropical cyclones: wind, surge, and rainfall.

Although the results show a good performance of the new index compared with existing scales, limitations remain. Because land characteristics are not accounted for, the new subindices cannot accurately estimate the amount of damage or the death toll for events impacting major metropolitan areas (e.g., Hurricanes Sandy and Katrina; see Fig. 7). Consequently, those two events were removed from the regression modeling. Nevertheless, the new indices still rank Hurricanes Sandy and Katrina more accurately than does the SSHS, particularly for damage (Fig. 10 and Table 5). In terms of the individual subindices, the rankings of Katrina and Sandy appear sensible, even though they were excluded from the original regression. For example, the majority of the damage and fatalities for Sandy arose from storm surge, with wind responsible for the rest of the damage. Rainfall was not a significant factor. The individual subindices (Fig. 11) give a ranking for Sandy that is consistent with reality. Moreover, Katrina is ranked highest by all three individual indices. Further work will consider the influence of land and population characteristics in order to further improve the model performance and broaden the applicability of the model.

Fig. 10.
Fig. 10.

Hurricanes ranked (including Sandy and Katrina) according to their actual damage (billion USD) and death toll, and the respective SSHS and TCHI rankings.

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

Fig. 11.
Fig. 11.

Hurricanes ranked (including Sandy and Katrina) using the individual hazard subindices.

Citation: Weather and Forecasting 29, 6; 10.1175/WAF-D-14-00014.1

5. Conclusions

A new tropical cyclone or hurricane hazard index has been presented to estimate the damage and death toll, and to rank the severity, of these damaging storm systems. The key difference compared with previous indices or scales is the inclusion of rainfall. Existing approaches are based on the wind characteristics (e.g., the Saffir–Simpson hurricane scale) or based on storm surge (e.g., HSI; Kantha 2006). The new analysis is based on existing open-access databases—H*WIND, HURDAT, and PERSIANN—to extract the wind field characteristics, forward speed, and rainfall characteristics, respectively.

Three separate subindices were developed, corresponding to the three dominant hazards of wind, rainfall, and surge. In addition, an appropriate model for the combined hazard was developed by merging these three individual indices to estimate the total damage and death toll. The regression coefficients were R2 = 0.81 and 0.88, respectively (excluding Hurricanes Sandy and Katrina), indicating a significantly improved model compared to existing scales. The new tropical cyclone hazard index (TCHI) can also provide quantitative estimates of damage and death toll. The results show that, in both the damage and death toll models, the rainfall subindex (TCRI) is dominant in comparison with the wind and storm surge indices, consistent with data showing that freshwater flooding is the leading cause of death in most hurricanes (Rappaport 2000). Further work is required to account for the significantly greater damage and death toll when systems impact densely populated areas or major cities.

This methodology has the potential to be used in real-time parametric models to better classify approaching hurricanes according to their physical hazards and to estimate the potential damage and death toll, enabling improved hazard planning and emergency response.

Acknowledgments

The authors thank the Center for Hydrometeorology and Remote Sensing (CHRS), University of California, Irvine, for providing the PERSIANN data and hosting a visit by MR.

REFERENCES

  • Arlot, S., , and Celisse A. , 2010: A survey of cross-validation procedures for model selection. Stat. Surv.,4, 40–79, doi:10.1214/09-SS054.

  • Chang, C.-P., , Yang Y.-T. , , and Kuo H.-C. , 2013: Large increasing trend of tropical cyclone rainfall in Taiwan and the roles of terrain. J. Climate, 26, 41384147, doi:10.1175/JCLI-D-12-00463.1.

    • Search Google Scholar
    • Export Citation
  • Czajkowski, J., , Simmons K. , , and Sutter D. , 2011: An analysis of coastal and inland fatalities in landfalling US hurricanes. Nat. Hazards, 59, 15131531, doi:10.1007/s11069-011-9849-x.

    • Search Google Scholar
    • Export Citation
  • Hawkins, D. M., , Basak S. C. , , and Mills D. , 2003: Assessing model fit by cross-validation. J. Chem. Inf. Comput. Sci., 43, 579586, doi:10.1021/ci025626i.

    • Search Google Scholar
    • Export Citation
  • Hebert, C. G., , Weinzapfel R. A. , , and Chambers M. A. , 2010: Hurricane severity index: A new way of estimating a tropical cyclone’s destructive potential. Impactweather, Inc., 5 pp.

  • Irish, J. L., , and Resio D. T. , 2010: A hydrodynamics-based surge scale for hurricanes. Ocean Eng., 37, 6981, doi:10.1016/j.oceaneng.2009.07.012.

    • Search Google Scholar
    • Export Citation
  • Jordan, M. R., , and Clayson C. A. , 2008: Evaluating the usefulness of a new set of hurricane classification indices. Mon. Wea. Rev., 136, 52345238, doi:10.1175/2008MWR2449.1.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2006: Time to replace the Saffir–Simpson hurricane scale? Eos, Trans. Amer. Geophys., 87, 36, doi:10.1029/2006EO010003.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2008: Tropical cyclone destructive potential by integrated kinetic energy. Bull. Amer. Meteor. Soc., 89, 219221, doi:10.1175/BAMS-89-2-219.

    • Search Google Scholar
    • Export Citation
  • Kantha, L., 2013: Classification of hurricanes: Lessons from Katrina, Ike, Irene, Isaac and Sandy. Ocean Eng., 70, 124128, doi:10.1016/j.oceaneng.2013.06.007.

    • Search Google Scholar
    • Export Citation
  • Matyas, C. J., , and Cartaya M. , 2009: Comparing the rainfall patterns produced by Hurricanes Frances (2004) and Jeanne (2004) over Florida. Southeast. Geogr., 49, 132156, doi:10.1353/sgo.0.0046.

    • Search Google Scholar
    • Export Citation
  • Peduzzi, P., , Chatenoux B. , , Dao H. , , De Bono A. , , Herold C. , , Kossin J. , , Mouton F. , , and Nordbeck O. , 2012: Global trends in tropical cyclone risk. Nat. Climate Change, 2, 289294, doi:10.1038/nclimate1410.

    • Search Google Scholar
    • Export Citation
  • Pielke, R. A., Sr., , and Pielke R. A. Jr., 1997: Hurricanes: Their Nature and Impacts on Society. John Wiley and Sons, 298 pp.

  • Powell, M. D., , and Reinhold T. A. , 2007: Tropical cyclone destructive potential by integrated kinetic energy. Bull. Amer. Meteor. Soc., 88, 513526, doi:10.1175/BAMS-88-4-513.

    • Search Google Scholar
    • Export Citation
  • Rappaport, E. N., 2000: Loss of life in the United States associated with recent Atlantic tropical cyclones. Bull. Amer. Meteor. Soc., 81, 20652073, doi:10.1175/1520-0477(2000)081<2065:LOLITU>2.3.CO;2.

    • Search Google Scholar
    • Export Citation
  • Rappaport, E. N., 2014: Fatalities in the United States from Atlantic tropical cyclones: new data and interpretation. Bull. Amer. Meteor. Soc., 95, 341346, doi:10.1175/BAMS-D-12-00074.1.

    • Search Google Scholar
    • Export Citation
  • Saffir, H. S., 1977. Design and construction requirements for hurricane resistant construction. American Society of Civil Engineers Preprint 2830, 20 pp.

1

The area (radius) with at least 34-kt surface wind speed.

Save