This paper explores simulated changes to the cool-season (November–March) storm-surge and coastal-flooding events at the Battery in New York City, New York (NYC), during most of the twenty-first century using several climate models and a previously developed multilinear regression model. The surface wind and pressure forcing for the surge predictions are obtained from an ensemble of 6 coupled global climate models (GCM) and 30 members from the Community Earth System Model. Using the “RCP8.5” emission scenario, both the single-model and multimodel ensemble means yielded insignificant (significance level p > 0.05) simulated changes to the median surge event (>0.61 m above astronomical tide) between a historical period (1979–2004) and the mid-to-late twenty-first century (2054–79). There is also little change in the return interval for the moderate-to-high surge events. By the mid-to-late twenty-first century, there is a poleward shift of the mean surface cyclone track in many of the models and most GCMs demonstrate an intensification of the average cyclone. There is little effect on the future surge events at the Battery because most of these storm changes are not in a region that favors more or higher-amplitude surges at NYC. Rather, projected sea level rise dominates the future simulated change in the number of flooding events by the mid-to-late twenty-first century. For example, the projections show about 23 times as many coastal-flooding events (tide + surge ≥ 2.44 m above mean lower low water; 1983–2001) in 2079 when compared with 1979, and the return intervals for some major coastal floods (e.g., the December 1992 northeaster) decrease by a factor of 3–4.
Extratropical storms during the cool-season months (November–March) often create coastal flooding along the east coast of the United States, which results in hazardous conditions, coastal erosion, and property damage. The change in water height beyond the tide that is forced by surface winds and pressure is referred to as the storm surge, and the total water level (astronomical tide + surge) determines the depth of flooding.
Like much of the East Coast, the coastline from near Atlantic City, New Jersey, to Montauk, New York, is vulnerable to storm surges from coastal storms because its position relative to the storm track favors frequent onshore (easterly) winds from nearby cyclones along the East Coast (Fig. 1). In addition, the shallow bathymetry of the coastal shelf promotes the transport of the water by surface wind stresses (Fig. 1; Resio and Westerink 2008). At the Battery in New York City, New York (NYC), between October and March 1959–2007, weak extratropical-cyclone surges (≥0.61 and < 1.0 m) have occurred between 1 and 14 times per season, whereas stronger (≥1.0 m) surges have occurred between 0 and 2 times per season (Colle et al. 2010). Some of these events have caused significant coastal flooding for NYC, such as the surge during the December 1992 northeaster (Colle et al. 2008).
The variations in frequency, track, and intensity of nearby extratropical storms that influence coastal flooding are strongly related to larger-scale atmospheric patterns of variability, such as the North Atlantic Oscillation (NAO; Talke et al. 2014) and El Niño–Southern Oscillation (ENSO) (Hirsch et al. 2001; DeGaetano et al. 2002). For example, Talke et al. (2014) demonstrated that the annual maximum water level at the Battery was strongly anticorrelated (correlation coefficient R = −0.92; p = 0.076) to the NAO between 1856 and 2013. Some studies have also found a qualitative relationship between the frequency of East Coast storms and ENSO events (Hirsch et al. 2001; DeGaetano et al. 2002). For example, Hirsch et al. (2001) showed that that there were about 44% more East Coast storms during El Niño periods as compared with neutral or La Niña periods. An increase in cyclone frequency and/or intensity could potentially force an increase in annual surge frequency since coastal cyclones generate surge events (Colle et al. 2008).
To study long-term storm-surge variability from extratropical cyclones, multidecadal simulations of surface wind and pressure are required. These data are currently available from global climate models (GCMs). Most GCMs typically feature coarse resolutions (horizontal resolution of 1°–2°) and tend to underpredict both the frequency and intensity of extratropical cyclones (Colle et al. 2013). To simulate coastal flooding with GCM data, statistical relationships (von Storch and Reichardt 1997; Roberts et al. 2015) can represent surge at a point, whereas other approaches create two-dimensional (2D) fields of surge with numerical hydrodynamic models (Flather et al. 1998; Lowe and Gregory 2005; Woth et al. 2006). Von Storch and Woth (2008) provide a comprehensive summary of twenty-first-century surge variations over Europe using numerical approaches. Numerical approaches for storm surge often use 2D barotropic numerical models (Flather et al. 1998; WASA Group 1998; Lowe and Gregory 2005; Woth et al. 2006; Butler et al. 2007), but this approach can be prohibitively expensive for multiple decades using a relatively large ensemble. As a result, Roberts et al. (2015) developed a multilinear regression model for storm surge along the New York–New Jersey coastline, building upon the work of Salmun et al. (2011) and Salmun and Molod (2015). Roberts et al. (2015) showed that the regression model has a level of skill in predicting storm surge at the Battery that is similar to that of a three-dimensional hydrodynamical model run in the region [Stevens Institute of Technology New York Harbor Observing and Prediction System (SIT-NYHOPS); Georgas 2010]. The regression developed in Roberts et al. (2015) makes it more feasible to predict decades of 3-hourly surge time series with ensembles using GCM data and explains more than ~60% of the variance in the observed surge time series.
Modeling studies on twenty-first-century cyclone-track changes for North America and the North Atlantic Ocean have demonstrated a poleward shift and a reduction in the overall frequency of the simulated surface extratropical-cyclone tracks by 2079–2100 (Yin 2005; Colle et al. 2013; Chang et al. 2012; Chang 2013). These mean cyclone shifts are partly the result of weaker lower-tropospheric meridional temperature gradients in a warmer global climate (Pinto et al. 2007; Mizuta et al. 2011; Colle et al. 2013; Chang et al. 2012; Chang 2013; Zappa et al. 2013). Some studies, such as Mizuta et al. (2011), Colle et al. (2013), and Marciano et al. (2015), show evidence that there may be an increase in the frequency of relatively deep cyclones during the Northern Hemispheric winter by the late twenty-first century. The effects of a reduction in the number of extratropical cyclones, an increased frequency of deeper extratropical cyclones, and a shift of the mean cyclone track could alter future coastal-flooding impacts for the NYC region.
An increase in sea level will also increase the occurrence of coastal flooding. At the Battery, the local mean sea level (LMSL, 1.78 m above station datum; 1983–2001) has risen at approximately 3 mm yr−1 since 1900 (Center for Operational Oceanographic Products and Services 2016; Fig. 2), which is almost 2 times the rate of global sea level rise (SLR) since 1900 (~1.50 mm yr−1; IPCC 2013, pp. 1137–1216). Over the twentieth century, a 0.43-m increase in the LMSL since 1856 has contributed to an increase in the 10-yr flood level by 0.72 ± 0.25 m (Talke et al. 2014), and, throughout the twenty-first century, the regional SLR is expected to accelerate. A recent study for the New York–New Jersey coastline projected a regional SLR of 0.59–1.14 m (95th-percentile range) by 2100 above a reference level (LMSL; Zhang et al. 2014; Fig. 2) following the Representative Concentration Pathway 8.5 (RCP8.5) emissions scenario from the Intergovernmental Panel on Climate Change Fifth Assessment Report (AR5; IPCC 2013). In a similar way, the New York City Panel on Climate Change (NPCC2; Horton et al. 2015) projected an SLR between 0.39 and 1.91 m (90th-percentile range) above LMSL by 2100 using a carbon emission scenario that is similar to RCP8.5.
To determine whether simulated changes of twenty-first-century coastal flooding are reliable, the differences as compared with a historical period must be shown not to originate from random variations (i.e., they are statistically significant). The approach often used to characterize projections of future climate is to analyze solutions from multiple models [i.e., multimodel ensembles (MMEs)] and compare them with a historical period. In contrast, ensembles obtained from one model using perturbed initial conditions [single-model ensembles (SMEs)] are also used in much the same way as MMEs (Kay et al. 2015). A comparison between the MMEs and SMEs may provide more confidence in any projected changes to the twenty-first-century surge since each ensemble format has its own benefits and drawbacks (Tebaldi and Knutti 2007); if both ensembles yield similar results, confidence in the projections is increased.
This study will answer the following questions:
Can the historical storm-surge climatological behavior between 1979 and 2004 for November–March at the Battery be accurately represented by a multilinear regression forced with atmospheric data from both a single-model and a multimodel ensemble of climate models?
How does the cool-season (November–March) surge climate vary during a future period (2006–79) given the RCP8.5 emissions scenario using a single-model ensemble and multimodel ensemble?
How does regional sea level rise, consistent with the same global-warming emissions scenario used for surge predictions, affect twenty-first-century coastal flooding?
2. Data and methods
a. Water-level and storm-surge data
Given the recent work performed on the Battery’s past and future coastal-flooding frequency (Talke et al. 2014; Horton et al. 2015) and the prominent coastal infrastructure nearby, the focus of this study is the Battery in NYC. Raw water-level data for the Battery were obtained from the NOAA Center for Operational Oceanographic Products and Services Tides and Currents website (http://tidesandcurrents.noaa.gov) in hourly increments from 20 November 1926 (earliest year of hourly verified water level on record) to 31 December 2015. The water-level time series was detrended by removing a linear best-fit line of 3.0 mm yr−1, and then the detrended monthly mean water-level anomaly, calculated from the Permanent Service for Mean Sea Level (http://www.psmsl.org/; accessed 30 March 2016), was removed. This was done to remove the combined effects of thermal expansion/contraction, intraseasonal variability in winds, and local anthropogenic effects to the gauge site, such that the water level is primarily the sum of astronomical tide and storm surge. Storm-surge data were then calculated by subtracting the predicted astronomical tides obtained through NOAA from the filtered water-level data. Storm surge defined in this manner assumes tide–surge interactions are negligibly small relative to the surge height. The tide–surge interaction was found to be reasonably small by studying “Q–Q” plots (e.g., Bernier and Thompson 2007) of the tidal residual (water level − astronomical tide) at different tidal stages. Time series of storm surge were coarsened to 3 h (0000, 0300, 0600 UTC, etc.) to work with the regression described below.
A storm maximum surge (peak with at least 24 h of separation) of 0.61 m or greater is referred to as a surge event as in Colle et al. (2010) and Roberts et al. (2015), which represents the 95th percentile of storm maximum surges at the Battery (Roberts et al. 2015). When the water level (astronomical tide + surge) exceeds 2.44 m (peak with 24 h of separation) above mean lower low water (MLLW, 1.0 m above station datum; 1983–2001), it is considered to be a coastal-flood event. Both the surge event and coastal-flood event are based on the flooding potential at the Battery around high tide and are thresholds used by the National Weather Service to issue coastal-flood advisories and warnings, respectively. Our definition of a surge event does not represent the actual coastal-flood events at the Battery because many surges have occurred below high tide. Our coastal-flood-event definition does represent actual flooding events at the Battery, however, since it is defined with respect to MLLW.
b. GCM data
The study uses 6-h data from seven GCMs as part of phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2012). Six GCMs compose the MME; the Community Earth System Model (CESM) and its 30 ensemble members represent an SME. The MME membership represents a compromise between some commonly used models in recent East Coast cyclone studies (e.g., Colle et al. 2013) and 6-h data availability when this project was collecting data. The SME data are from the Large Ensemble Project (Kay et al. 2015). Both CMIP5 and Large Ensemble Project data used here are forced with the RCP8.5, which predicts approximately 8.5 W m−2 of global radiative forcing by 2100. The considered GCMs are listed in Table 1 together with their relevant attributes. The data from five of the GCMs used in this work were also used by Colle et al. (2013) and were found to realistically simulate the surface cyclone climatological behavior over the western North Atlantic Ocean.
The GCM data employed for surge predictions are the lowest-model-level (LML) u and υ wind components (U) and mean sea level pressure data (MSLP). All data are separated into a historical period (from 1 January 1979 to 31 March 2004) and a future period (from 1 November 2006 to 31 March 2079) for the cool-season months of November–March. For the CESM, only one 14-yr time slice and two 10-yr time slices (1990–2004, 2026–35, 2071–80) were available with 6-h data; therefore, in the context of the SME the historical period (P1) refers to the years between 1990 and 2004 and the early and mid-to-late future periods correspond to 2026–35 (P2) and 2071–80 (P3), respectively.
To discuss mechanisms that could lead to future changes in storm surge, surface cyclones in the MME GCMs throughout the 100-yr study period were tracked and analyzed. Surface cyclones were automatically tracked using the MSLP field in all of the MME GCMs (Table 1) following the methods in Colle et al. (2013). In brief, the Hodges (1994, 1995) cyclone-tracking algorithm was used to obtain the cyclone track using 6-h spectral bandpassed (wavenumbers equal to or less than 5 and larger than 70 are removed) MSLP data. Cyclones that were included had at least a 24-h lifetime and a track distance of more than 1000-km. Colle et al. (2013) determined that there is a 5%–10% uncertainty associated with using this approach to identify surface cyclones.
c. Multilinear regression
A multilinear regression (MLR) developed in Roberts et al. (2015) was trained and validated to predict 3-h storm surge at the Battery using reanalysis data over 1979–2012 for the October–March period. In this statistical model, two components of summed (15 and 18 h, respectively) surface wind stress (i.e., Aτx and Aτy) and one component of MSLP forcing (or P) were used to generate predictions of storm surge [Eq. (1)] for a point that represents the location of the Battery:
The data for these predictors were collected over the boxed region (hereinafter referred to as the predictor box) illustrated in Fig. 3. The predictors were derived by following the method of Roberts et al. (2015) and were standardized (z score) on the basis of the 1979–2004 November–March Climate Forecast System Reanalysis (CFSR; Saha et al. 2010) predictor climatological statistics. Here P represents the predictor box’s minimum MSLP, and the zonal wind stress predictor Aτx and meridional wind stress predictor Aτy were derived by converting to wind stresses with the method of Garratt (1977), were spatially averaged over the predictor box, and then were summed in time (Roberts et al. 2015). Note that the βs represent the coefficients of the MLR, which are shown in Table 2. The MLR underpredicted relatively large (≥95th percentile) storm maximum surge heights by 6%–38%; therefore, predictions of surge were multiplied by the reciprocal of the training multiple correlation coefficient R (Roberts et al. 2015). The MLR [Eq. (1)] forced with CFSR atmospheric data explained 64% of the variance in the validation-period observed 3-h storm surges at the Battery. Surge predictions were associated with a mean absolute error of 0.29 m and a mean error of −0.01 using Eq. (1).
d. Data treatment
All GCM and CFSR data were linearly interpolated from a 6-h time step to a 3-h time step to work with the MLR’s data format, as described in Roberts et al. (2015). All data were then spatially interpolated bilinearly to a grid of 0.5° × 0.5° (Fig. 3). Gridpoint U wind-component data were reduced from their LML height (indicated by subscript HLML) (Table 1) to 10 m using a power-law transform [Eq. (2)], as has been done in many other studies (Archer and Jacobson 2005; Pryor et al. 2005; Devis et al. 2014):
The term ε was assumed to be constant with a value of 0.11 corresponding to a neutral stability stratification (Hsu et al. 1994) since storms are often associated with strong low-level winds that can effectively mix the boundary layer to a neutral stratification. The height of the LML zHLML was determined from the hypsometric equation with the pressure at the LML sigma level and the surface pressure (Devis et al. 2014). The lowered and interpolated U every 3 h were then converted to wind stress components using the Garratt (1977) drag formulation as in Roberts et al. (2015).
Biases inherent to wind and pressure data from GCMs resulted in biases in the MLR’s predictors (Figs. 4 and 5), which contributed to biases in surge predictions. Therefore, the mean and standard deviation of the MLR’s three predictors were bias corrected (BC) using the CFSR and MLR predictor statistics for the November–March 1979–2004 period.
To bias correct, the MLR predictors were turned into means (indicated by overbars) and anomalies (indicated by primes), and then these data were separated into training (subscript T) and validating (subscript V) datasets using an alternating-year configuration for cross validation. For example, the odd years were used for training and the even years were used for validation. The training and validation periods were also switched (i.e., even years for training and odd years for validation), and the BC approach was repeated. In all validation period data, the mean bias during the training period was subtracted out and the anomaly was multiplied by the ratio of standard deviations S [Eq. (3); Xu and Yang 2012]:
Note that the acronyms GCM and CFSR in Eq. (3) are placeholders for the atmospheric variables (U and MSLP) and that the superscript asterisk denotes BC data. The mean bias is the middle term with parentheses on the right-hand side of Eq. (3).
To form a continuous time series of BC data for the historical period, the validation-period data from each cross-validation configuration were merged. For the future period, the MLR predictors were bias corrected using both cross-validation configurations and then the 3-h surge predictions were averaged to form one time series of 3-h surge for each GCM. The BC approach decreased the percent difference [(GCM − CFSR)/CFSR] for two of the three MLR predictors (Fig. 4), the exception being Aτy. More important, the BC approach resulted in storm maximum surge percentiles that had a smaller spread and converged more onto observations of storm maximum surge percentiles during both ensembles’ historical periods (Fig. 5).
Time series of surge represent only one realization of Earth’s climate state; therefore, we use the range of an ensemble of predictions as an estimate of the variation of the statistic. To simulate an estimate of variability in the observed surge data, observed time series of surge events per season were resampled 1000 times to create a sampling distribution. For predictions of water level (astronomical tide + surge), the tidal uncertainty was taken into account since a surge event can occur at any tidal stage at the Battery. To quantify tidal uncertainty in water-level predictions, a 3-h tidal time series of the Battery Park tide from 1979 to 2012 was added to the 3-h surge predictions 1000 times. The tidal time series were shifted one 3-h time step forward iteratively 1000 times to quantify the flooding potential at different tidal stages.
a. Historical surge period
The regression model using MME and SME input from the historical period predicts more 3-h intervals that meet the surge-event threshold than were observed in the climatic record (Fig. 6). For instance, the cumulative probability of observed 3-h intervals that meet the surge-event threshold is 0.75% (Fig. 6b), whereas the MME and SME have cumulative probabilities that range from 0.9% to 1.8%. Three-hour surge intervals of 0.61–0.81 m are predicted to occur 0.5%–1.2% of the time (Fig. 6a), whereas only 0.6% of 3-h observed data are in this range. For more intense surge thresholds (i.e., ≥0.81 m), both modeled and observed probabilities diminish to less than 0.25% and the differences are smaller between predictions and observed data.
The frequency in storm-surge events (a peak with at least a 24-h separation) is underpredicted using the SME and MME. Between 1979 and 2004 for November–March there are 98 (55–146 when including uncertainty from natural variability) observed surge events at the Battery, with an average of 3.78 events per season and a range between 0 and 10 events per season. The MME mean has an average of 59.2 (44–95 range) events (Fig. 7a). Most MME members underpredict the frequency of events during the MME’s historical period by 15–30 events, whereas the GFDL underpredicts by only one event. The SME also underpredicts its observed historical surge period with 40.2 (31–49 range) surge events predicted between November and March for 1990–2004, whereas there were 52 (30–86 range) surge events observed in this period. The MME has an average of 2.4 (range of 0–8.5 events per season) during its historical period (Fig. 7a), and the SME has an average of 2.9 (0–10.2 events per season) during its historical period. Overall, the member-to-member variability in the SME is smaller and the historical surge-event frequency is more accurate than that of the MME, since some of the MME members have a large underprediction in surge-event frequency. The overall overprediction in duration of surge events and underprediction in frequency of events is partly a result of the interpolation of atmospheric forcing data from 6 h down to an interval of 3 h.
Between 1926 and 2015 for November–March there were an average of 0.23 observed coastal floods per season, with a range from 0 to 3 floods per season. Between 1979 and 2004, the MME mean predicts an average of 0.28 floods per season (0.02–0.87 floods per season when including uncertainty from tides), with a range from 0 to 5 floods per season. Between 1990 and 2004, the SME mean predicts an average of 0.36 floods per season (0.03–0.52 floods per season with tidal uncertainty) with a range from 0 to 6 floods per season.
b. Twenty-first-century surge predictions
The MME mean does not predict a significant trend (F statistic = 0.21; p > 0.05) in the frequency of surge events over much of the twenty-first century (2006–79; Fig. 8). Here we use an F test to determine whether the slope of the line is different from zero (Lomax 2007, p.10). Over the twentieth century, there was no significant observed trend (F = 0.46; p > 0.05) in surge events per season. Four of the six members of the MME and the MME mean have nonsignificant trends; the BCC and CanESM demonstrate significant trends of 0.0144 (F = 9.38; p = 0.0029) and 0.0177 (F = 5.99; p = 0.0162) events per season between 1979 and 2079. Overall, MME predictions over the twenty-first century exhibit variations that are similar to those of the modeled data during the twentieth century. For example, the future period (2006–79) MME mean, average surge-event frequency was 2.6 events per season, with a range from 0 to 8.7 events per season. This result is similar to the historical period’s MME mean, average surge-event frequency of 2.4 events per season and range from 0 to 8.5 events per season.
During the three SME time slices (i.e., P1, P2, and P3) throughout the twentieth–twenty-first century, the SME mean also does not show a significant trend in surge events per season (FP1 = 1.25 with pP1 > 0.05; FP2 = 1.14 with pP2 > 0.05; FP3 = 1.09 with pP3 > 0.05). The SME mean surge-event frequency varies from 2.87 events per season over 1990–2004 to 2.96 events per season over 2026–35 and to 3.10 events per season over 2071–80 (Fig. 9). Further, modeled surge-event frequencies exhibit variations that are similar to that of estimated variability (range of 0–10 surge events per season). For example, the SME’s range during each period is 0–10.16 events per season from 1990 to 2004, 0–11.87 events per season from 2026 to 2035, and 0–9.87 events per season from 2071 to 2080.
The MME mean does not demonstrate a significant trend (F = 2.01; p > 0.05) in the intensity of extreme (i.e., the seasonal maximum surge for each ensemble member) modeled surge events throughout much of the twenty-first century; neither does the SME mean (FP1 = 0.78 with pP1 > 0.05; FP2 = 1.61 with pP2 > 0.05; FP3 = 1.23 with pP3 > 0.05). Throughout much of the twentieth century (1926–2015), the observed seasonal maximum surge (SMS) was on average 1.02 m and ranged between 0.61 and 2.83 m (Fig. 10a). In twenty-first-century MME predictions, the SMS ranged between 0.61 and 2.37 m, which was similar to the MME’s historical-period data (i.e., SMS = 1.01 m, with a range from 0.61 to 2.50 m). The SME’s predictions of SMS were 0.92 m (0.61–3.10 m) between 1990 and 2004, 0.92 m (0.61–3. 04 m) between 2026 and 2035, and 0.94 (0.64–2.42 m) between 2071 and 2090.
From a Mann–Whitney U test (Mann and Whitney 1947), neither the MME members nor the MME mean demonstrate a statistically significant change in the median surge-event data for 2054–79 when compared with 1979–2004 (Figs. 11a–g). The only significant change occurs for the CanESM’s moderate surge events (>1.0 m), with a median difference of 0.11 m (U statistic = 43; p = 0.003). As a result, the MME mean exhibits nonsignificant changes in the median surge between periods. In a similar way, the Mann–Whitney U test shows that the SME exhibits no significant change in the median surge height between 1991–2000 and 2071–80 (not shown).
c. Twenty-first-century coastal floods
During the twenty-first-century period, the MME mean frequency of coastal floods was 0.30 floods per season (from 0.01 to 0.97 floods per season given tidal uncertainty). This result is similar to the MME mean during the historical period [0.28 floods per season (0.02–0.87 floods per season)]. This prediction does not take into account the effects of SLR. Along the New York–New Jersey coastline, it is very likely that mean sea level will continue to rise at an accelerating pace throughout the twenty-first century (IPCC 2013; Zhang et al. 2014; Horton et al. 2015). To take into account the effects of SLR on coastal flooding, a range that encompassed the projections of SLR (0–1.0 m) (Zhang et al. 2014; Horton et al. 2015) was added to predictions of water level. This procedure results in a plot that shows the relationship between SLR and the number of flooding events using the MME mean (Fig. 12). On this surface, the x, y coordinate illustrates the MME mean number of coastal floods in a particular season given an SLR.
With the addition of the regional sea level projection (Zhang et al. 2014), the frequency of coastal floods increased exponentially throughout the twenty-first century (Fig. 12). The MME mean predicts 0.80 floods per season (0.67–5 floods per season with tidal uncertainty) by 2050 and ~6.6 floods per season (3.7–10 floods per season) by 2079. Uncertainties in modeling the components that affect SLR lead to considerable uncertainty in this prediction. For instance, Zhang et al. (2014) report 95th-percentile uncertainty bounds between 0.59 and 1.14 m around the expected SLR of 0.86 m by 2100. Figure 12 also illustrates how the number of flooding events varies for any value of SLR up to 1.0 m from 2010 to 2079 so that, if future estimates of SLR change, one can get an estimate in the number of flooding events. In addition, in Fig. 12 one can also see that the wiggles around each dashed line are smaller in magnitude than the curve induced from an SLR projection. This situation indicates that the effects on coastal-flooding frequency from cyclone variability are smaller than the effects from SLR.
The MME and SLR results can be used to obtain return (or recurrence) intervals (RI) for various flooding levels. RIs are calculated with and without SLR for observed historical and predicted future (2006–79) water levels:
where n is the number of years on record and m is the relative rank of the event (with the largest event being rank 1).The RI for observed data for November–March between 1926 and 2015 is calculated with surge predictions from the MME and bootstrapped with the tidal data, as described in the methods. As a result, a sample of 1000 historical time series is available to determine the RI for each MME member.
Figure 13a shows the RI using the 1926–2015 observed water levels and bootstrapped tides at the Battery, as well as the historical (1979–2004) and future (2006–79) water levels using the MME mean. Two observed RI estimates are provided, one with Hurricane Sandy (2012) and another without it. Because Sandy occurred in late October, it fell very close to our definition of a cool season (November–March), and Galarneau et al. (2013) noted that this storm had a structure more similar to an extratropical cyclone as it approached landfall. Sandy illustrates the RI sensitivity of the results to including a major storm. Including Sandy reduces the RI for a water level 3.1 m above MLLW from 70 years down to a value of 35 years. This 3.1-m MLLW water level is the peak that occurred during the December 1992 northeaster that flooded part of the NYC subway system (Colle et al. 2008). The future RI estimates from the MME mean are more in line with the observed non-Sandy RIs, since these models did not predict a storm of such magnitude.
Overall, there is little change in the future RIs using the MMEs, but these estimates do not include SLR. Figure 13b shows the impact of SLR on the RIs using the MME. SLR dramatically decreases the RIs for all water levels. Given the SLR projection from Zhang et al. (2014), the RI for coastal-flood events decreases from 3.33 years to less than 1 year. Coastal flooding as observed during the 1992 northeaster (peak water level of 3.02 m MLLW) changes from every ~45 years down to a value of ~10 years. This once again emphasizes that SLR will likely be the major determining factor in how the number of flooding events will change in the future.
4. Impact of cyclone changes
The storm-surge results can be put in the context of some recent studies that investigated changes in simulated extratropical cyclones along the East Coast during the twenty-first century as compared with the twentieth century (Colle et al. 2013). Colle et al. (2013) used the “best 7” CMIP5 models to show that during the late twenty-first century (2079–2100 for November–March) the number of cyclones decreases over the western Atlantic for the RCP8.5 scenario when compared with 1979–2004 for November–March while there is a slight (5%–10%) increase in the number of storms along the coastal plain of the East Coast. We unfortunately could not use the same best-7 models in our study since the 6-hourly LML wind data were not available for several of these models (except CCSM4 and CNRM).
We note behaviors in the GCMs used in this study in our cyclone-track data that are similar to those found in Colle et al. (2013). For instance, Fig. 14 shows the percent change in cyclone density (number of cyclones per November–March per 50 000 km2) for the various members of the MME between 2054–79 and 1979–2004. Similar to some cyclone-track-change results found in Colle et al. (2013), the MME mean has a 2%–8% increase in the number of cyclone tracks, focused in the mid-Atlantic region. The BCC and NorESM both show a 10%–15% increase in cyclone-track density over the northeastern United States, whereas the CCSM4 projects a pronounced ~15% increase in cyclone-track density over the southeastern United States. In addition, the MME mean projects a deepening of the median cyclone MSLP < 1000 hPa by 2–3 hPa (U= 18 482; p = 0.09) for 2054–79 that passes through a region around NYC (Fig. 15). Colle et al. (2013) and similar studies showed that deeper (<990 hPa) cyclones occurred more frequently over the Northeast despite an overall reduction of cyclones over the western Atlantic between 1979–2006 and 2079–2100.
The changes in the median surge event between periods may be related to a combination of intensity, location, and frequency changes in the mean cyclone track in the region surrounding NYC. For example, a shift in the mean cyclone track could decrease the intensity of surge events by lessening the probability of strong 10-m winds. Meanwhile, a deepening of the average cyclone MSLP could offset the effect on the average surge event from some of these track changes. As mentioned earlier, the MME mean shows a 2–3-hPa intensification in the median surface cyclone (<1000 hPa) in the region around NYC (Fig. 15). An increase in more intense storms (i.e., <990 hPa) occurs mainly to the north of the Battery (not shown) in simulated data, however. This is likely one of the reasons why the RIs are similar between the historical and future periods.
To study the effect of mean cyclone-track variations on the intensity of surge events at the Battery, surge predictions were matched in time to cyclone tracks that were in the region covered by 90°–60°W and 30°–50°N. If multiple cyclone tracks were present in this region during the same surge event, only the cyclone track that was closest to the Battery was associated with the surge event. To demonstrate the effect that changes in the mean cyclone track have on surge, the maximum in the track-density difference over the mid-Atlantic region was reduced by randomly removing 20% of cyclones that pass through a polygonal region (the magenta region in Fig. 16) that encompassed the maximum track-difference anomaly in the MME mean. This process was repeated 1000 times to take into account the random nature of the removal of storm tracks.
By comparing the difference in the change ΔMD in median surge between 2054–79 and 1979–2004 and the change in median surge event between the same periods calculated with cyclone tracks removed [Eq. (5)], we estimate the effect that the removal of tracks has on the median surge event:
In Eq. (5), ς represents the statistic distributed in Fig. 17. To calculate p values for ς, we sum the number of times that the median is greater than zero (less than zero) to obtain the right-tailed (left tailed) test and then double the most significant test.
After the random removal of 20% of the cyclone tracks (Fig. 16b) 1000 times, a two-tailed bootstrap test (Diaconis and Efron 1983) failed to reject the hypothesis that the change in the MME mean, median surge event (<0.001 m) was different from zero (p > 0.05) (Fig. 17). Only the CCSM4 demonstrated a significant (p = 0.04) change of −0.01 m, but this change is trivial in comparison with the typical surge event (i.e., 0.61 m) and projected SLR. When one considers the region in which the MME mean experiences the largest increase in track density in 2054–79 relative to 1979–2004, it is apparent that there is little overlap with the track density associated with cyclones that generate surge events at the Battery (Fig. 16a). For example, after the removal of tracks associated with the poleward shift in track density, there is little change in track density associated with surge events (Fig. 16b).
Using both GCM data and projections of sea level rise, this is one of the first papers to look at simulated twenty-first-century changes in coastal flooding and storm surge from extratropical storms for New York City. Coastal-flooding and storm-surge events at the Battery in NYC were predicted for most of the twenty-first century by using a multilinear regression model that was forced with 6-hourly surface wind and pressure data from an ensemble of seven coupled GCMs. Six GCMs that were members of CMIP5 composed the multimodel ensemble (Taylor et al. 2012). An additional single-model ensemble that was composed of 30 members using the CESM was compared with the 6-member multimodel approach for one 15-yr and two 10-yr periods (1990–2004, 2026–35, and 2071–80) throughout the twenty-first century.
The comparison approach employed here between the SME and MME demonstrates the relatively large effect that different models and resolutions can have on the surge projections on the climatic time scale (i.e., 20–30 yr). When the RCP8.5 emission scenario is used, however, both the SME and MME means yielded insignificant changes (p > 0.05) on the order of 0.01 m to the median surge event and on the order of approximately 0.001 events per season to the frequency of surge events per season between a 25-yr historical period (1979–2004 for the MME and 1990–2004 for the SME) and a mid-to-late-twenty-first-century period (2054–79 for the MME and 2071–80 for the SME).
The changes in the mean surface cyclone track in the western North Atlantic were randomly removed to quantify the effect that they could have on the median storm-surge event at the Battery. A two-tailed bootstrap test failed to reject the null hypothesis that the modeled effect (<0.001 m) from the removal of storm track in 2054–79 on the median surge event was different from zero. The region in the MME mean that experiences the largest increase in track density in 2054–79 relative to 1979–2004 has little overlap with the track density associated with cyclones that generate surge events at the Battery. Further, while some models produced more intense cyclones in the later twenty-first century, these changes occurred north of the NYC region and thus have little effect on the surge intensity. Therefore, given the magnitude of changes found here and the typical order of an impactful surge event (i.e., ~0.61 m) we conclude that we cannot detect an impactful change in surge event due to changes in cyclones into the twenty-first century.
Regardless of the effect that storm-track changes may have on surge, the height of the sea level largely determines the magnitude of coastal flooding. Projected regional SLR will dominate any twenty-first-century surge changes that may result from extratropical-cyclone changes at NYC. The MME mean combined with an expected regional SLR scenario that projected a sea level rise of 0.86 m by 2100 yielded approximately 23 times as many coastal-flooding events (tide + surge > 2.44 m above MLLW) in 2079 as in 1979, and the flooding return intervals may decrease by a factor of 3–4.
At this time, the study of regional climate change is in its youth, and the conclusions from analyzing these projections carry large uncertainty. It is still necessary to characterize the ability of global climate models to replicate regional climates that are used to produce things like storm-surge distributions. Such characterization will allow us to determine progress and regression in our ability to simulate regional climate as we further refine our simulations. There are many uncertainties associated with this work. One challenge of using the CMIP5 GCM data is that it was necessary to employ a bias correction to improve historical (1979–2004) predictions of surge events, which brings into question the inherent realism of the GCM forcing. Because there was no 10-m CMIP5 wind product available at the time of the study, one was derived using the lowest model level (typically 30–40 m), thus introducing uncertainties. Furthermore, the coarse (1°–2°) horizontal resolutions and 6-h time step in the GCM output likely contributed to the overall underprediction of surge events and the overprediction of the surge-event duration. The authors are currently performing another study on twenty-first-century changes to storm surges in and around NYC that uses similar 6-h CMIP5 data but with a dynamical downscaling approach that predicts storm surges with a barotropic two-dimensional finite-element model.
We thank the three anonymous reviewers for their constructive comments and suggestions to improve this paper. We also thank the CESM group at the National Center for Atmospheric Research for providing those ensemble data and Dr. Kevin Reed for help in accessing the data. This work is supported by the U.S. Department of Commerce National Oceanic and Atmospheric Administration (NOAA) Climate Program Office Modeling, Analysis, Predictions and Projections (MAPP) Program under Grant NA11OAR4310104 and by NOAA through project R/CCP-18 to the Research Foundation of the State University of New York on behalf of New York Sea Grant. The statements, findings, conclusions, views, and recommendations are those of the author(s) and do not necessarily reflect the views of any of these organizations.