Influence of Weather and Climate on Multidecadal Trends in Atlantic Hurricane Genesis and Tracks

Grace Kortum aDepartment of Geosciences, Princeton University, Princeton, New Jersey
bHigh Meadows Environmental Institute, Princeton University, Princeton, New Jersey

Search for other papers by Grace Kortum in
Current site
Google Scholar
PubMed
Close
,
Gabriel A. Vecchi aDepartment of Geosciences, Princeton University, Princeton, New Jersey
bHigh Meadows Environmental Institute, Princeton University, Princeton, New Jersey

Search for other papers by Gabriel A. Vecchi in
Current site
Google Scholar
PubMed
Close
,
Tsung-Lin Hsieh aDepartment of Geosciences, Princeton University, Princeton, New Jersey
bHigh Meadows Environmental Institute, Princeton University, Princeton, New Jersey

Search for other papers by Tsung-Lin Hsieh in
Current site
Google Scholar
PubMed
Close
, and
Wenchang Yang aDepartment of Geosciences, Princeton University, Princeton, New Jersey

Search for other papers by Wenchang Yang in
Current site
Google Scholar
PubMed
Close
Open access

Abstract

This study investigates the relative roles of sea surface temperature–forced climate changes and weather variability in driving the observed eastward shift of Atlantic hurricane tracks over the period from 1970 to 2021. A 10-member initial condition ensemble with a ∼25-km horizontal resolution tropical cyclone permitting atmospheric model (GFDL AM2.5-C360) with identical sea surface temperature and radiative forcing time series was analyzed in conjunction with historical hurricane track observations. While a frequency increase was recovered by all the simulations, the observed multidecadal eastward shift in tracks was not robust across the ensemble members, indicating that it included a substantial contribution from weather-scale variability. A statistical model was developed to simulate expected storm tracks based on genesis location and steering flow, and it was used to conduct experiments testing the roles of changing genesis location and changing steering flow in producing the multidecadal weather-driven shifts in storm tracks. These experiments indicated that shifts in genesis location were a substantially larger driver of these multidecadal track changes than changes in steering flow. The substantial impact of weather on tracks indicates that there may be limited predictability for multidecadal track changes like those observed, although basinwide frequency has greater potential for prediction. Additionally, understanding changes in genesis location appears essential to understanding changes in track location.

Significance Statement

From the 1970s to the present, there has been an increase in the frequency of North Atlantic hurricanes, but they have also shifted in location to the east, away from land. We explore whether this shift in hurricanes’ locations was caused by climatic factors or randomness to understand if and how these trends will persist. We also consider whether the shift was due to a change in where hurricanes started or how they moved over their lifespan. Analyzing data from observed and simulated hurricanes, we find that the shift was made more likely by climate factors, but ultimately occurred due to random variability in the hurricanes’ starting locations. These results suggest a higher uncertainty in the future location and impact of hurricanes and highlight the importance of studying why hurricanes originate where they do.

© 2024 American Meteorological Society. This published article is licensed under the terms of the default AMS reuse license. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Grace Kortum, gkortum@alumni.princeton.edu

Abstract

This study investigates the relative roles of sea surface temperature–forced climate changes and weather variability in driving the observed eastward shift of Atlantic hurricane tracks over the period from 1970 to 2021. A 10-member initial condition ensemble with a ∼25-km horizontal resolution tropical cyclone permitting atmospheric model (GFDL AM2.5-C360) with identical sea surface temperature and radiative forcing time series was analyzed in conjunction with historical hurricane track observations. While a frequency increase was recovered by all the simulations, the observed multidecadal eastward shift in tracks was not robust across the ensemble members, indicating that it included a substantial contribution from weather-scale variability. A statistical model was developed to simulate expected storm tracks based on genesis location and steering flow, and it was used to conduct experiments testing the roles of changing genesis location and changing steering flow in producing the multidecadal weather-driven shifts in storm tracks. These experiments indicated that shifts in genesis location were a substantially larger driver of these multidecadal track changes than changes in steering flow. The substantial impact of weather on tracks indicates that there may be limited predictability for multidecadal track changes like those observed, although basinwide frequency has greater potential for prediction. Additionally, understanding changes in genesis location appears essential to understanding changes in track location.

Significance Statement

From the 1970s to the present, there has been an increase in the frequency of North Atlantic hurricanes, but they have also shifted in location to the east, away from land. We explore whether this shift in hurricanes’ locations was caused by climatic factors or randomness to understand if and how these trends will persist. We also consider whether the shift was due to a change in where hurricanes started or how they moved over their lifespan. Analyzing data from observed and simulated hurricanes, we find that the shift was made more likely by climate factors, but ultimately occurred due to random variability in the hurricanes’ starting locations. These results suggest a higher uncertainty in the future location and impact of hurricanes and highlight the importance of studying why hurricanes originate where they do.

© 2024 American Meteorological Society. This published article is licensed under the terms of the default AMS reuse license. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Grace Kortum, gkortum@alumni.princeton.edu

1. Introduction

Tropical cyclones in the North Atlantic (hurricanes) pose a substantial risk to people’s safety and well-being in the United States, Caribbean, and Central America as one of the principal natural threats to the region (Pielke et al. 2008; Smith and Katz 2013). As a result, it is important to investigate what has driven the considerable variability in the frequency, intensity, and location of North American tropical cyclones over time, as well as how they are expected to behave in the future.

Over the last ∼50 years from 1970 to 2021, there has been a generally increasing trend in tropical cyclone frequency in the North Atlantic (e.g., Fig. 1). Over the same period, there has been a southward and eastward shift in the location of tropical cyclone tracks (e.g., Kossin et al. 2010; Fig. 1).

Fig. 1.
Fig. 1.

(top) The recorded frequency of hurricanes and (middle) the average longitude of hurricanes have shown an increasing trend since 1970, while (bottom) the average latitude of hurricanes has been generally decreasing since 1970. We observe that frequency rises above the mean for the time period and latitude falls below the mean for the time period at around the year 1995. Series are shown with a 5-yr rolling average applied. Data are from the IBTrACS dataset (Knapp et al. 2010).

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

This observed increase in tropical cyclone frequency is thought to be caused by some combination of radiative forcing changes (particularly aerosols) and internal climate variability, in particular, the Atlantic meridional mode (AMM) and the Atlantic multidecadal oscillation (AMO) (Delworth and Mann 2000; Mann and Emanuel 2006; Kossin et al. 2010; Booth et al. 2012; Villarini and Vecchi 2013; Walsh et al. 2015; Vecchi et al. 2017; Murakami 2022). Meanwhile, an analysis of multicentennial proxy-based hurricane reconstruction in the Bahamas suggested that random fluctuation can dominate track location in the Atlantic (Wallace et al. 2020). The impact of internal climate variability and radiative forcing on North Atlantic tropical cyclone tracks has also been studied. For example, a northeastward extension of the North Atlantic subtropical high (Elsner et al. 2000) and similarly a weakening of the western North Atlantic subtropical high in negative North Atlantic Oscillation (NAO) phases (Kossin et al. 2010) have been associated with more recurving tropical cyclone tracks. Kossin et al. (2010) also observed that positive AMM phases are correlated with more landfalling storms, while positive El Niño–Southern Oscillation (ENSO) phases have been tied to the opposite effect (Xie et al. 2005; Kossin et al. 2010). Colbert and Soden (2012) attributed the decrease in landfalling tropical cyclones during El Niño to a change in steering flow but posited that the changes driven by the AMM were caused by a change in tropical cyclone genesis locations. In addition, Kossin et al. (2010) find that while there has been an increase in the frequency of deep tropical storms (which are located more in the south and east of the Atlantic) since the 1980s, these storms do not make up a substantial portion of the storms that hit land, and the frequency of the Gulf of Mexico storms that comprise the majority of storms making landfall in North America has not been changing.

Anthropogenic climate change has also been projected to affect the trajectories of North Atlantic tropical cyclones (Colbert et al. 2013). Vecchi and Soden (2007) projected using model simulations that increasing CO2 may be associated with an eastward movement of tropical cyclone genesis locations, while Murakami and Wang (2010) projected that climate change will yield an increase in tropical cyclones making landfall in the northeast of the United States and a decrease in the southeast of the United States. Liu et al. (2017) discuss eastward shifts in North Atlantic hurricanes in response to greenhouse induced warming. Vecchi and Knutson (2008, 2011) discuss a century-scale eastward shift of North Atlantic tropical storms and hurricanes since 1871. It is worth noting that tropical cyclone tracks have shifted westward in recent decades in most ocean basins (Wang and Toumi 2021) except the North Atlantic, where our analysis is focused.

We aim to investigate the relative role of randomness (or weather-scale variability) and SST-forced climate in driving the observed recent multidecadal changes in the spatial distribution and frequency of hurricanes. A primary limitation to identifying the statistical significance of the observed signals in tropical cyclone activity attributed to changes in climate is limited data availability due to the infrequency of tropical cyclones. In addition, the attribution of trends due to forced or internal climate change from weather variability is difficult due to the short time period over which accurate observations have been collected. As a result, model generated tropical cyclone tracks are a critical tool for assessing the extent to which observed signals are due to weather or climate factors, because models can produce ensemble members with varying weather conditions but identical climate trends over time. We use a combination of observed and model simulated hurricane tracks to bring new evidence to the issue of weather versus climate modulation of hurricane frequency and track distribution.

This study seeks to answer the following questions:

  1. Were the observed multidecadal changes in tropical cyclone frequency and track distribution from 1970 to the present driven by SST-forced changes in climate or weather variability?

  2. Are multidecadal trends in the spatial distribution of tropical cyclone tracks produced primarily by year-to-year variability in steering flow, year-to-year variability in genesis locations, or a combination of the two?

To address these questions, we first analyze historically observed hurricane tracks in the context of an ensemble of model generated tracks with identical climate conditions over time but variable weather between ensemble members in order to isolate climate-driven and weather-driven trends. Next, we use a statistical model to simulate tracks with different genesis locations and steering flow conditions and thus identify the roles of genesis location changes and steering flow changes in driving shifts in track location.

2. Data and methods

a. Datasets

Data for observed tropical cyclone tracks were taken from IBTrACS, the International Best Track Archive for Climate Stewardship from the NOAA National Climatic Data Center (Knapp et al. 2010).

Model tropical cyclone tracks were generated from the Geophysical Fluid Dynamics Laboratory (GFDL) Atmospheric Model version 2.5 with C360 resolution (AM2.5-C360). This model is the atmospheric component of the GFDL Climate Model version 2.5 (CM2.5) (Delworth et al. 2012) using a higher-resolution cubed-sphere finite-volume dynamical core of Putman and Lin (2007). This is a similar model to the atmospheric component of the GFDL Hi-Resolution Forecast-Oriented Low Ocean Resolution (HiFLOR) model (Murakami et al. 2015) and is inspired by the results of Chen and Lin (2011, 2013). The horizontal grid spacing is approximately 25 km for the atmospheric simulation. The prescribed observational SST time series is developed by Chan et al. (2021). The greenhouse gas and aerosol concentrations follow the Coupled Model Intercomparison Project phase 5 (CMIP5) protocol (Taylor et al. 2012). Before year 2005, the concentrations follow the CMIP5 historical experiments, and after year 2005 the concentrations follow the RCP4.5 scenario.

The AM2.5-C360 model generates realistic tropical cyclone annual cycles (Yang et al. 2021) and response to ENSO (Hsieh et al. 2022). A set of 10 ensemble members that was generated under the climate and forcing conditions of the last 52 years, but with varying initial conditions resulting in different weather, was used to represent different potential scenarios that could have occurred in the North Atlantic between 1970 and 2021. Initial conditions for the 10 ensemble members are set as 10 different restart files at the beginning of each year. These restart files are from an AM2.5-C360 control simulation where radiative forcings are fixed and monthly SST climatology is specified. Climatic forcing from greenhouse gases and aerosols is included through the sea surface temperature forcing. Important to this analysis is the fact that all 10 ensemble members were generated with identical sea surface temperature forcing, and thus any interensemble-member differences are considered to be weather-driven. For more information about the model design and specifications of AM2.5-C360, refer to Chan et al. (2021), Hsieh et al. (2022), and Yang et al. (2021).

b. Data processing

Tropical cyclone tracks were computed from the model output using the method of Harris et al. (2016). Corresponding sea level pressure fields for the model tropical cyclone tracks were also used. For the model and observed tracks, data points were selected according to the definition of genesis as the first point where an intensity of 17.5 m s−1 was sustained for the next 24 consecutive hours and lysis as the last point where an intensity of 17.5 m s−1 was sustained for the previous 24 consecutive hours. In addition, only warm core storms that lasted for at least 2 days according to these genesis and lysis conditions were included in the dataset. We considered tropical storms that reached a maximum intensity of at least 33 m s−1 to be hurricanes.

For model sea level pressure data, both the monthly and 6-hourly frequencies were used, and pressure gradients were computed along the zonal and meridional directions then averaged along a 5° running window in both longitude and latitude. The time frame of 1970–2021 was used for all analysis.

c. Spatial clustering

A clustering method was applied to the observed and model tracks together in order to identify and isolate specific types of hurricanes based on their position and movement. Following Elsner (2003), Ramsay et al. (2012), and Ng and Vecchi (2020), a k-means clustering algorithm (MacQueen 1967) was used; this method seeks to minimize the differences between vectors in each cluster. As in Kossin et al. (2010), we clustered all of the tropical storms, not only hurricanes.

In the application here, clusters were based off simplified vectors representing the movement of each storm. For each track, a vector of 20 dimensions was constructed that included the first recorded location (genesis-point latitude and longitude), last recorded location (lysis-point latitude and longitude), and eight evenly spaced latitude–longitude pairs in between based on an interpolation of the track. The k-means clustering was run once on the vectors representing the tropical storm tracks across all years and all 10 ensemble members in addition to the historical tracks to construct consistently defined cluster assignments.

Average silhouette scores (which are a metric for describing how well each of the n samples is partitioned into its cluster) were computed for clustering with k = 2 to k = 9 clusters. The silhouette scores indicated that having four clusters yielded a similar coherence level to three clusters, and a much higher coherence level than five clusters. As a result, we employed four clusters because it provided more descriptive power than two clusters and was consistent with past studies of North Atlantic tropical cyclone tracks, which have used between three and five distinct groups (Kossin et al. 2010; Colbert and Soden 2012; Ng and Vecchi 2020). The clusters were labeled from 1 to 4, with 1 being the cluster that was located the farthest to the west on average and 4 being the cluster that was located the farthest to the east on average.

d. Bootstrap resampling

The storm tracks dataset contained 10 ensemble members, which is too small of a sample size to estimate sample statistics without making assumptions about the underlying probability distribution of the variables being assessed. As a result, a bootstrap resampling method (e.g., Efron and Tibshirani 1991) was used to generate additional bootstrap ensemble members. These new bootstrap ensemble members were generated by selecting data for each year from the 10 existing ensemble members and combining them back together to form a full time series from 1970 to 2021. For example, in a bootstrap ensemble member, the data for year 1970 might be taken from ensemble member 5, while the data for year 1971 were taken from ensemble member 2. This was randomized in that for each year, a number was randomly drawn with replacement from 1 through 10, and this number would correspond to the ensemble member from which the data for that year would be taken. This method assumes that the noise in each Atlantic hurricane season is statistically independent between years, conditioned on the prescribed SST time series. Therefore, these bootstrap ensemble members would be equally likely to have occurred as the 10 original ensemble members. One million bootstrap ensemble members were generated out of a possible 102021 − 1970+1 = 1052 options. A limitation of these bootstrap ensemble members as opposed to having a million new “real” ensemble members is that they were all limited to having only the exact same tropical cyclone tracks that had been produced already for the 10 original ensemble members. As a result, the clustering algorithm was not rerun for the bootstrap ensemble members, and each tropical cyclone track maintained the same cluster label as it had originally been assigned.

We sought to analyze the bootstrap ensemble members which exhibited large spatial shifts in track distribution from the period before 1995 to the period after 1995. These shifts were modeled in terms of the proportion of tracks in each cluster. Specifically, a proportional decrease in the fraction of tracks in cluster 2 relative to cluster 3 between the two time periods was considered an “eastward” shift, while a proportional increase in the fraction of tracks in cluster 2 relative to cluster 3 between the two time periods was considered a “westward” shift. The reason we define these shifts in terms of clusters 2 and 3 is because these are the two clusters that drive the multidecadal shift in the observed tracks; this will be shown later in the results section. Throughout the rest of the paper, we discuss the eastward shift and westward shift in terms of proportional changes between clusters 2 and 3 only. However, we note that defining the eastward shift and westward shift in terms of changes between clusters 1 and 3 yields qualitatively consistent results.

e. Regression model specification

In addition to diagnosing explicit tropical cyclones in the AM2.5-C360 global model, we develop a regression model to generate synthetic tropical cyclone tracks in order to investigate the relative importance of background steering flow and genesis location. The regression model is useful because it allows the generation of synthetic tracks based on factors including steering flow and genesis location. Then, scenarios where only steering flow changes year-to-year while genesis locations are held constant over time can be generated and compared against other scenarios where genesis locations change over time while steering flow is held constant year-to-year. The synthetic tropical cyclone model has three components: 1) genesis location, 2) track movement, and 3) intensity evolution, similar to the methods of Emanuel et al. (2006) and Lee et al. (2018). In contrast to these existing methods that used only large-scale environmental conditions as an input, the genesis and intensity components of our synthetic tropical cyclone model used information diagnosed from AM2.5-C360. The track movement component is a regression model.

Many past studies have used statistical models to simulate tropical cyclone tracks, and they bring the benefit of computational ease and more direct interpretability compared to dynamical models. Both Emanuel et al. (2006) and Lee et al. (2018) apply a Beta and Advection Model (BAM) to predict track movement. These BAMs use a combination of a beta-drift effect and an advection effect based off wind velocities at 850 and 250 hPa to predict track movement in each time step. Hall et al. (2021) and Hall and Jewson (2007) apply an autoregressive model for zonal and meridional track movement.

We build a multiple linear regression model that incorporates elements of the approaches of Emanuel et al. (2006), Hall and Jewson (2007), and Lee et al. (2018). Separate statistical models were generated to predict the average zonal translation velocity (u) and meridional translation velocity (υ) between times t and t − 1 (6 h earlier) based off of beta drift and advection from large-scale steering flow. These average velocities were then used to update the position of the tropical cyclone from its location at time t − 1 to its location at time t. The parameters of our BAM fit are summarized in Table 1 along with variable definitions.

Table 1.

Description of regression variables.

Table 1.

Three variables were included in the regression to represent beta drift: β, β(intensity), and β(intensity2). The best way to parameterize beta drifts is an open question. For example, Lee et al. (2018) use four cases with beta drift parameters of 1, 2.5, 1 m s−1 × cos(ϕ), or 2.5 m s−1 × cos(ϕ). In this model, we remove any assumptions on the size of the beta drift parameter by empirically deriving a coefficient on cos(ϕ). Merlis et al. (2016) describe that higher tropical cyclone intensity is correlated with a stronger beta drift effect, and we incorporate this interaction with the β(intensity) and β(intensity2) terms.

Last, large-scale steering flow is modeled as the geostrophic winds corresponding to the sea level pressure gradients in the zonal and meridional directions, following the geostrophic balance equations:
fυg=1ρpx,fug=1ρpy,
where p is sea level pressure and f is Earth’s vorticity and equal to 2ω sin(ϕ).

Sea level pressure gradients were the rolling mean of the sea level pressure gradient across 5° of latitude and longitude. This component of the regression model follows the advection components of Emanuel et al. (2006) and Lee et al. (2018). We explore two settings, one using sea level pressure data at a monthly resolution and one using sea level pressure data at a 6-hourly resolution.

Because we assume velocity to be driven only by beta drift and large-scale steering flow, the constant term in the regression was assumed to be zero.

The regression model is summarized in the following equations:
ut=b0β+b1β(intensity)+b2β(intensity2)+b31ρfpy+ϵ,υt=b0β+b1β(intensity)+b2β(intensity2)+b31ρfpx+ϵ,
where β is df/dy=[2ωcos(ϕ)]/a, and ω, ϕ, a, and ρ are as defined in Table 1. The regression coefficients are bn, and ϵ represents the error term.

f. Track simulation using the regression model

The regression model was used to produce simulated tropical cyclone tracks. Because the regression model did not predict genesis location, intensity evolution, or length, these factors had to be accounted for in the simulated tracks. For the choice of genesis location, different studies have taken different approaches, such as sampling genesis locations from a probability distribution (Emanuel et al. 2006; Hall and Jewson 2007) or sampling historical genesis locations (Vickery et al. 2000). For our simulation experiments, we use the full set of genesis locations from the 52 years in each of the 10 ensemble members to increase the sample size.

Simulations were run in two ways to account for length and intensity. In the first, track length and intensity evolution were prescribed directly from the corresponding dynamical model storm track with the same genesis location. Under this method, for each track being simulated using the statistical model, the dynamical model storm whose genesis location was being used for the simulation was identified. Then, in each time step of the simulation, the intensity inputted to the regression model was the intensity of the corresponding dynamical model track after the same number of hours, and the simulation was stopped at the same length (in hours) as the corresponding dynamical model track. In the second method, we sought to remove any variability in track length and intensity that was not due to variability in genesis location, so that the simulation experiments could be better used to isolate the effects of only genesis location. A simple model for track length was developed using a linear regression of length on latitude and longitude of genesis. In the model data, this regression had a correlation coefficient of 0.59. For each track simulated with the statistical model, length was predicted based on the genesis location. An average intensity evolution vector was calculated by interpolating all tracks in the dynamical model output to have the same length from genesis to lysis and then calculating the average intensity at each point. In the simulation experiments, this average intensity vector was interpolated again to match the predicted storm length and then was inputted to the regression model at each time step.

Simulations were produced using monthly mean sea level pressure data for the regression model trained with monthly data and 6-hourly instantaneous sea level pressure for the regression model trained with 6-hourly data. The tracks in the 10 ensemble members were simulated using the regression models to test their ability to capture the ensemble mean shift in track location over time.

A series of experiments were run to investigate the effect of changing steering flow while genesis location is held constant across years and the effect of changing genesis location while steering flow is held constant across years. For these experiments, we tested how well the regression models could simulate an extreme track shifting scenario given that it had information about solely genesis locations from 1970 to 2021 or solely sea level pressure from 1970 to 2021. One experiment would be run as follows:

  1. A target bootstrap ensemble member with extreme track shift is selected.

  2. Constant sea level pressure annual cycle simulations are run:

    1. The genesis locations and times from each of the 52 years of the extreme bootstrap ensemble member were selected.

    2. Annual sea level pressure data from each of the 10 ensemble members and their 52 years was selected [for a total of 520 (=10 × 52) annual cycles of sea level pressure data].

    3. A set of 520 scenarios would be generated using this one set of annually changing genesis locations paired with each of the 520 sea level pressure annual cycles that are repeated 52 times.

  3. Constant genesis location annual cycle simulations are run analogously:

    1. The sea level pressures from each of the 52 years of the extreme bootstrap ensemble member were selected.

    2. Annual genesis location data from each of the 10 ensemble members and their 52 years was selected.

    3. A set of 520 scenarios would be generated using this one set of annually changing sea level pressures paired with each of the 520 genesis location annual cycles held constant annually.

  4. As a result, two sets (corresponding to constant genesis and constant sea level pressure runs) of 520 scenarios that each contain 52 years’ worth of simulated tracks would be produced. Each of the two sets then contained on the order of (520 scenarios) × (52 years) × (∼5 tracks yr−1) ≈ 160 000 simulated tracks.

  5. These two sets were compared to the target extreme bootstrap ensemble member to see how well they each captured its shift pattern given only knowledge about its genesis locations or only knowledge about its sea level pressures.

This procedure is summarized in Fig. 2. Of note, these experiments, despite holding genesis location or sea level pressure constant across years, still allowed for variability within the year, always pairing genesis points with the sea level pressure associated from the correct time of year. Throughout the paper, “constant genesis” or “constant sea level pressure” then always just refers to the variables being held constant interannually but never spatially or intra-annually.

Fig. 2.
Fig. 2.

Schematic describing the constant genesis and constant sea level pressure simulation procedure for each target extreme bootstrap ensemble member. Each of these rectangles comprises a blue half and green half representing 52-yr-long sets of sea level pressure data and genesis location data that would be used to produce a 52-yr-long dataset of simulated tracks. Note that “ensemble member” is abbreviated to “ens.”

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

We used this method to simulate hurricane tracks from the top 100 eastward-shifting bootstrap ensemble members and the top 100 westward-shifting bootstrap ensemble members using both the monthly sea level pressure regression model and the 6-hourly sea level pressure regression model. In total, 2 693 600 unique tracks were simulated using predictions from the regression models.

An analogous method was followed to simulate the tracks in the original 10 ensemble members under constant sea level pressure and constant genesis location conditions.

3. Results

a. Spatial distribution of tracks

The k-means clustering method effectively identified four types of storms characterized by different track pathways. The distributions of the locations of tracks in each cluster are presented in Fig. 3 and the distributions of genesis locations for each cluster are shown in Fig. 4. Storms in cluster 1 travel the farthest west and comprise almost all storms that reach Central America. Storms in cluster 2 tend to take a more northward trajectory, often approaching the East Coast of the United States. Clusters 3 and 4 are located the farthest to the east, with storms in cluster 3 starting farther to the south on average than cluster 4, and storms in cluster 4 taking more of an eastward trajectory, having a mean lysis point in the northeast of the Atlantic approaching Europe. Table 2 summarizes key statistics about the numbers of storms in each cluster. In both observations and the model ensembles, more of the storms were in clusters 1 and 2 than clusters 3 and 4 (Table 2). However, there were some differences in the overall model distribution of tracks relative to observations, with the model having a lower share of tracks in cluster 1 by 8 percentage points and higher share of tracks in cluster 2 by 8 percentage points (Table 2).

Fig. 3.
Fig. 3.

(top),(middle) Track density of hurricanes in each cluster. (bottom left) All model and observed hurricane tracks with their cluster assignments. (bottom right) Mean points of genesis, midpoint, and lysis for each cluster, with error bars representing standard deviation.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

Fig. 4.
Fig. 4.

(top),(middle) Density of genesis locations for each of the four clusters. (bottom left) All genesis points with their cluster assignments. (bottom right) For each gridded square, the color and opacity demonstrate the likelihood that a track with genesis in that location would be assigned to each cluster. For example, a square that is halfway between red and green indicates that 50% of storms with genesis in that location were assigned to cluster 3 and 50% were assigned to cluster 4, while a white square indicates a 25% chance of being assigned to each cluster (complete uncertainty).

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

Table 2.

Summary of cluster statistics for model and observed hurricanes.

Table 2.

b. Changes in the frequency and distribution of tracks over time

In addition, changes over time in the frequency and distribution of tracks were investigated. For each of the 10 ensemble members, along with the ensemble mean and historical tracks, the percentage of tracks in each cluster before 1995 and the percentage of tracks in each cluster from 1995 onward was computed. The percentage point change in these values is shown for each cluster in Fig. 5 and Table 3. Throughout this paper, “pre-1995” will refer to the period from 1970 through 1995, while “post-1995” will refer to the period from 1996 through 2021. Additionally, the percentage changes in the average annual number of storms between the pre-1995 and post-1995 periods were computed, with the results shown on the right of Fig. 5. Across the ensemble members and observed data, there was a consistent increase in the mean annual number of storms between the two periods (Fig. 5 and Table 2). In the observed data, there was an increase in cluster 3 (by 3.72 percentage points) and a decrease in cluster 2 (by 4.50 percentage points) between the pre- and post-1995 periods, along with a small increase in cluster 1 and small decrease in cluster 4. Among the model ensemble members, there was not a lot of consistency in which clusters showed increases or decreases between the time periods, apart from cluster 3, which had a proportional increase in frequency in 8 of the 10 ensemble members (Fig. 5).

Fig. 5.
Fig. 5.

(left),(center) Summary of the change in the proportion of total storms assigned to each cluster from the pre-1995 period to the post-1995 period for each of the 10 ensemble members, ensemble mean, and observed tracks. (right) Percent change in total number of storms from the pre-1995 period to the post-1995 period for each of the 10 ensemble members, ensemble mean, and observed tracks.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

Table 3.

Changes in the fraction of storms in each cluster between 1970–95 and 1996–2021 for model and observed hurricanes.

Table 3.

The cluster distribution and storm frequency shift between the pre-1995 and post-1995 periods for a set of 100 000 bootstrap ensemble members are shown as a histogram in Fig. 6 and as a mean and standard deviation in Table 3. Table 3 also shows p values indicating the likelihood that a given bootstrap ensemble would have a change fraction of storms in a cluster that is consistent in sign with the observations. Most (77%) of the 100 000 bootstrap ensemble members had an increase in cluster 3 between the pre- and post-1995 periods (on average by 3.03 percentage points) and most (63%) also had a decrease in cluster 2 (on average by 1.95 percentage points) (Table 3). However, this pattern was much less consistent among bootstrap ensemble members than what was observed for the change in frequency, where all 100 000 of the bootstrap ensemble members had an increase in hurricane frequency between the two time periods (Fig. 6).

Fig. 6.
Fig. 6.

(left),(center) Probability density estimation representing the distribution of the percentage point change in the proportion of total storms assigned to each cluster between the pre-1995 period and the post-1995 period for 100 000 bootstrap ensemble members. The gray vertical line is at 0. (right) Probability density estimation representing of the percent change in annual number of storms between the pre-1995 and post-1995 periods for the same 100 000 bootstrap ensemble members.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

The change in observed and model ensemble mean tropical cyclone genesis density and track density between the pre-1995 and post-1995 periods was also investigated (Fig. 7). The change in model ensemble genesis and track densities (Fig. 7, bottom row) followed similar spatial patterns to the observed changes in genesis and track densities (Fig. 7, top row), but the magnitudes of these changes were smaller by about 50%. One difference between the observed and model spatial patterns of changes in track density is that the model ensemble mean shows negative values along the east coast of North America, while the observational data show more positive values in the same region. We have conceptualized the observations as a single realization from the observed distribution and the model ensemble mean as the expected value from the model distribution and note that at smaller spatial scales we expect to see more noise in track changes between ensemble members. For example, although the model ensemble mean shows a pattern of negative values (a decrease in track density) along the East Coast, 3 of the 10 ensemble members show positive values more consistent with observations.

Fig. 7.
Fig. 7.

(top) Spatial histogram of difference in the location of (left) observed hurricane track and (right) genesis density between the periods 1970–95 and 1996–2021. (bottom) Spatial histogram of difference in the location of (left) modeled hurricane track and (right) genesis density between the periods 1970–95 and 1996–2021. The model histograms show the change combined across all 10 ensemble members. Observed data are from IBTrACS and model data are from the GFDL AM2.5-C360 model; more detail on data sources is provided in section 2.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

The following sections seek to isolate the causes behind the variance in shifts in cluster partitioning from pre-1995 to post-1995. Two factors are considered in depth: genesis location and steering flow.

c. Linking genesis location and track distribution

First, the role of genesis location in driving differences in track location was considered. There are clear qualitative differences between the general genesis regions for the storms in each cluster, as was expected given that genesis location was one of the factors used to determine cluster assignment (Fig. 4). Genesis for cluster 1 occurs primarily around the Caribbean and Gulf of Mexico, genesis for cluster 2 occurs a bit to the northeast of cluster 1, genesis for cluster 3 is highly concentrated around the southeast of the region off the coast of Africa, and genesis for cluster 4 is very spread out throughout the basin. To better visualize the overlap between these four histograms and the information about cluster partitioning that can be gained from knowing genesis location, Fig. 4 shows the likelihood of being assigned to each cluster given a specific genesis location. Around the periphery of the region where genesis occurs, there is generally a high level of certainty about cluster partitioning from the genesis point, but in particular around 10°–20°N, 70°–40°W, the genesis location does not seem to provide enough information to predict cluster partitioning, with many locations having under 50% certainty of cluster assignment based on genesis location.

For the observed and model ensemble tracks, the shift in track distribution was coupled with a shift in genesis distribution to the southeast between the pre-1995 and post-1995 periods (Fig. 7).

We also considered the 100 bootstrap ensemble members with the largest eastward shift (increase in cluster 3 relative to cluster 2) and the largest westward shift (increase in cluster 2 relative to cluster 3) between the pre-1995 and post-1995 periods. For these 100 eastward-shifting and 100 westward-shifting bootstrap ensemble members, we show the average change in distribution of genesis points from the pre-1995 period to post-1995 period in Fig. 8. This plot shows how the changes in track location observed in the extreme eastward- and westward-shifting bootstrap ensemble members were also associated with a significant westward and eastward movement in the distribution of genesis location. This shift in genesis location is on the order of 5%–10% of the total counts for genesis points per square (Fig. 8).

Fig. 8.
Fig. 8.

(left) The (top) mean of the difference in the distribution of genesis locations between the pre-1995 and post-1995 periods, (middle) mean of the difference in average sea level pressure between the pre-1995 and post-1995 periods, and (bottom) average sea level pressure in the pre-1995 period for the 100 extreme bootstrap ensemble members with the largest increase in cluster 2 relative to cluster 3 between these periods. (center) As in the left panels, but for the extreme bootstrap ensemble members with the largest decrease in cluster 2 relative to cluster 3 between the pre-1995 and post-1995 periods. (right) The difference between the left and center columns. Flow vectors in the middle and bottom rows illustrate the resultant geostrophic flows from the sea level pressure graphs.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

d. Linking steering flow and track distribution

An analogous investigation into the role of changes in sea level pressure (and thus a simple estimate of steering flow) in driving the changes in track distribution was conducted. Figure 8 also shows the shifts in monthly sea level pressure averaged across June–November between the pre-1995 and post-1995 periods for the mean of the 100 eastward- and westward-shifting bootstrap ensemble members. The similarities between the mean differences in sea level pressure for pre-1995 to post-1995 between the extreme westward- and eastward-shifting bootstrap ensembles demonstrate a temporal trend in large-scale sea level pressure that is likely common to all ensemble members. However, there was still a difference between the change in sea level pressure for the westward-shifting ensembles and the change in sea level pressure for the eastward-shifting ensembles, including a larger increase in sea level pressure in the North Atlantic centered around 20°N, 45°W in the westward-shifting ensembles compared to the eastward-shifting ensembles. Based on this location and sign of this sea level pressure difference, one would expect more westward-steering flows in the westward-shifting ensembles compared to the eastward-shifting ensembles. We additionally show the mean sea level pressure over the pre-1995 period (Fig. 8, bottom row). This comparison highlights the relatively small magnitude of the changes in sea level pressure between the two time periods associated with eastward and westward shifts relative to mean sea level pressures. These changes are on the order of less than 0.1% of the mean sea level pressures over the pre-1995 period (Fig. 8, see left two panels of middle row compared to left two panels of bottom row).

The findings thus far have supported the hypotheses that changes in (i) genesis location, (ii) steering flow, or (iii) both between the pre-1995 and post-1995 periods could have led to the shifts in track distribution between these periods. To isolate the influences of genesis location and steering flow on track distribution more precisely, we performed targeted experiments with the statistical track prediction model.

e. Regression model results

The results of the regression models are presented in Table 4. Spatial R2 values ranged from 0.39 to 0.6 for the four regressions. The 6-hourly sea level pressure data models had higher R2 values than the models run using monthly sea level pressure data. As expected from an approximated geostrophic balance to the large-scale steering flow, the coefficients on (1/ρf)(p/y) were negative and the coefficients on (1/ρf)(p/x) were positive.

Table 4.

Parameters of storm motion model based on a regression. * indicates p < 0.01. Heteroskedasticity robust standard errors shown in parentheses.

Table 4.

f. Statistical model simulation ability and error

The regression model’s track simulation ability was assessed in three ways: first, with a qualitative look at the trajectories of the regression-simulated tracks with respect to the dynamical model tracks; second, with an assessment of the average displacement between the regression-simulated tracks and dynamical model tracks; and third, via a quantification of the ability of the regression-simulated tracks to reproduce shifts in track over decadal time scales, specifically those of the extreme scenario bootstrap ensemble members. The first two methods are discussed in the appendix.

To assess the ability of the regression model simulations to represent spatial shifts in tracks between the pre-1995 and post-1995 periods, track simulations were produced based off the genesis locations and sea level pressures of the 100 westward- and 100 eastward-shifting bootstrap ensemble members. Four different regression models were tested: the 6-hourly model and monthly model with and without hurricane length and intensities either taken from the corresponding dynamical model track or separately modeled. Histograms were then produced binned at a 5° latitude × 5° longitude resolution showing the difference in the normalized distribution of points in the pre-1995 and post-1995 periods for the bootstrap ensemble member tracks and regression simulated tracks. The correlation of this histogram for each of the bootstrap ensemble members with the analogous histogram from the regression simulated tracks produced from each of the four regression models was calculated (200 R values for each model). To account for the spatial heterogeneity produced by the histograms’ discrete binning, a rolling mean across every pair of adjacent grid squares was applied before the correlation was calculated. The distribution of these R values is shown in Fig. 9.

Fig. 9.
Fig. 9.

For the two statistical prediction models—(top) “monthly model” and (bottom) “6-hourly model”—the distribution of correlations of the differences in the normalized distribution of track points between the pre-1995 and post-1995 periods between the regression-simulated tracks and their target bootstrap ensemble member are shown. These are grouped by whether the target ensemble member was a westward- or eastward-shifting scenario, as well as whether the intensities of the target bootstrap ensemble member’s storms were used in the predictions.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

These results indicate that the regression model has skill in reproducing the change in the spatial distribution of tracks, although it performs better for the eastward-shifting bootstrap ensemble members than the westward-shifting ones. This difference in performance was tested with KS tests comparing the set of correlations for the eastward shift to those for the westward shift for each model. The KS tests indicated that these distributions were likely distinct from each other (p < 0.001). We hypothesize that this may be due to the larger signal in the eastward-shifting ensembles, reflecting the combined climate and weather impact. Modeling a storm’s intensity evolution and length instead of prescribing it from the corresponding dynamical model track did not materially affect the regression models’ abilities to capture the spatial track distribution (Fig. 9).

g. Synthetic tropical cyclone simulation results

Constant sea level pressure and constant genesis simulation experiments were run according to the procedure described in the methods section (Fig. 2) using the monthly and 6-hourly regression models. In these experiments, storm length was predicted using the regression model based on genesis location and storm intensity evolution was applied from the average track intensity vector, instead of applying the length and intensity values from the dynamical model tracks. Because it did not have a large effect on the statistical model’s ability to represent the multidecadal track shifting patterns (as represented in Fig. 9), this method was selected for the benefit of allowing a better isolation of the effects of changing genesis location from effects of the storm’s length or intensity evolution.

The regression model’s ability to reproduce the ensemble mean shift in tracks was investigated under the conditions of 1) holding sea level pressure annual cycle constant over all years (Fig. 10, left panel), 2) holding genesis location annual cycle constant over all years (Fig. 10, center panel), and 3) using the actual 52 years of genesis locations and sea level pressure time series (Fig. 10, right panel). These results demonstrate that the ensemble mean shift in track is recovered by the regression model when annual genesis locations are inputted to the regression model but not when only annual sea level pressures are inputted to the regression model. The southeastward shift in tracks produced by simulating tracks with year-to-year changes in sea level pressure and not genesis location (subtly visible in the center panel of Fig. 10) is about an order of magnitude smaller than the actual ensemble mean shift and the shift produced by simulating tracks with year-to-year changes in genesis location.

Fig. 10.
Fig. 10.

(top) Difference in the normalized track density between the pre-1995 and post-1995 periods across all ensemble members. (middle) Difference in the normalized track density between the pre-1995 and post-1995 periods for simulated tracks generated from the (left) genesis locations or (center) sea level pressures of the 10 ensemble members, paired with sea level pressure held constant in the left panel and the genesis location held constant in the center column. (right) Both the genesis locations and sea level pressures of the 10 ensemble members were used for simulated track generation. (bottom) As in the middle panels but with the 6-hourly regression model instead of the monthly regression model.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

The results of a constant genesis and constant sea level pressure simulation of one example extreme bootstrap ensemble member are shown in Fig. 11. In the figure it is clear that constant sea level pressure (changing genesis location) simulations for both the monthly and 6-hourly regression models qualitatively reproduce the spatial patterns of the track shifts of the target scenario. However, we note that the constant sea level pressure simulation using the monthly data model does not capture the decrease in track density around the Gulf of Mexico and Central America. While the constant genesis (changing sea level pressure) simulations appear to have some similar patterns to the target such as an area of decreasing point concentration off the East Coast of the United States and increasing point concentrations off the coast of West Africa, the relationship is less pronounced. This shows that interannually changing the genesis location produced a much larger difference in track positions between the pre-1995 and post-1995 periods than interannually changing sea level pressures with interannually constant genesis positions.

Fig. 11.
Fig. 11.

(top) Difference in the normalized distribution of track points between the pre-1995 and post-1995 periods for one eastward-shifting bootstrap ensemble member. (middle) Difference in the normalized distribution of track points between the pre-1995 and post-1995 periods for simulated tracks generated from the (left) genesis locations or (center) sea level pressures of the target bootstrap ensemble member, paired with sea level pressure held constant in the left panel and the genesis location held constant in the center panel. (right) Both the genesis locations and sea level pressures of the 10 ensemble members were used for simulated track generation. (bottom) As in the middle panels, but with the 6-hourly regression model instead of the monthly regression model. Each panel represents all 520 of the 52-yr simulations of the target scenario, which are then aggregated on one normalized histogram.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

To assess these relationships more quantitatively, and to investigate their robustness across different target bootstrap ensemble members, correlations were computed between 1) the difference histograms of each of the 520 changing sea level pressure simulations and 520 changing genesis location simulations of each target bootstrap ensemble member and 2) the difference histograms of their corresponding target bootstrap ensemble member. Again, a two-bin rolling mean was applied to the histograms before the correlation was calculated. The results are presented as distributions in Fig. 12 and as averages in Table 5. These distributions show that for the 6-hourly data model, the changing genesis location simulations captured the eastward and westward shifts of the bootstrap ensemble members about equally well as the simulations that were run using both the target bootstrap ensemble member’s genesis locations and sea level pressures. Still, the simulations produced with the monthly data model show that the changing genesis location experiments capture the shifts substantially better than the changing sea level pressure experiments. This is evidenced by the mean spatial R values, which are lower for the changing sea level pressure simulations compared to the changing genesis location and changing genesis location and sea level pressure experiments (Table 5).

Fig. 12.
Fig. 12.

For predictions from the two regression models—(top) “monthly model” and (bottom) “6-hourly model”—the distribution of correlations of the differences in the normalized distribution of track points between the pre-1995 and post-1995 periods between the changing genesis and changing sea level pressure experiment simulated tracks and their target bootstrap ensemble member is shown. These are grouped by whether the target ensemble member was a west or east shifting scenario.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

Table 5.

Average spatial correlation of the distributions of dynamical model tracks and regression model generated tracks under different modeling conditions.

Table 5.

Therefore, we conclude that the regression model of synthetic tropical cyclones is able to simulate the multidecadal shifts in tropical cyclone tracks in the ensemble mean and in bootstrap pseudoensemble members. Furthermore, genesis locations are a key annual input to the regression model for producing these multidecadal shifts.

4. Discussion

a. Influence of climate and weather on multidecadal trends in hurricane activity

Tropical cyclone tracks from the GFDL AM2.5-C360 model were used to evaluate the role of weather and climate in driving two observed trends in observed North Atlantic tropical cyclone tracks: an increase in frequency in the period of 1996–2021 compared with the period 1970–95 and an eastward shift in the spatial distribution of the tracks over a similar time period. Because the model’s ensemble members were constructed with identical signals of forced and internal climate variability in SST, any variations between ensemble members were determined to be driven by weather-scale differences, while trends common across ensemble members were attributed to climate-scale drivers (a combination of forcing and climate variability). The robust increase in the frequency of tropical cyclones among the ensemble members as well as a bootstrapped estimation of the ensemble distribution provided evidence that an increase in frequency was driven by a climate effect (Figs. 5 and 6). This finding is in agreement with past studies that have attributed this frequency change to the positive phase of the AMO and the associated generally positive phases of the AMM that have occurred over the period from 1995 to the present (e.g., Goldenberg et al. 2001; Kossin and Vimont 2007; Klotzbach and Gray 2008; Klotzbach 2011), as well as aerosol-driven changes (e.g., Booth et al. 2012). We did not here attempt to isolate the impact of SST signals from radiative forcing and internal variability, but have instead focused on isolating SST-driven (“climate”) effects from unforced (“weather”) effects.

Time changes in the spatial distribution of tracks, as modeled by changes in the cluster assignments, generally did not show a trend common to all the ensemble members (Fig. 5). For cluster 3, eight of ten ensemble members had an increase in the relative number of storms over time, and for the two ensemble members that had a relative decrease in cluster 3, the magnitude of the change was small. This indicates that climate factors made increases in cluster 3 more likely but not guaranteed.

Furthermore, although the ensemble mean tropical cyclone track density change between the pre-1995 and post-1995 periods did show an eastward shift, the change was not as pronounced as in the observed tracks (Fig. 7). The shift in the model ensemble member tracks also did not capture much of the southward movement in tracks seen in observations over this time period (Fig. 7).

The bootstrap estimations of the ensemble distributions of the shifts in cluster partitioning also showed a positive shift in the distribution of the proportion of storms in cluster 3, with a mean of 3.02 percentage points; this shift was compensated for by decreases in clusters 1 (by 1.21 percentage points) and 2 (by 1.92 percentage points). Out of the 100 000 bootstrap ensemble members, 63% had a decrease in cluster 2 and 77% had an increase in cluster 3. This shows that although climate factors may have made the observed decrease in cluster 2 and increase in cluster 3 more likely to occur, a change in the opposite direction could have happened as well due to weather variability.

Although the historically observed shifts in clusters 1 and 2 were a few percentage points different from the means of the ensemble distributions, they are overall consistent with the bootstrap ensemble distributions, and there is no statistical indication that the ensemble distribution is not representative of observations (p ≈ 1).

These results highlight the importance of even weather-scale variability on multidecadal hurricane track shifts, which have often been attributed to climate variability. For example, Kossin et al. (2010) explained the observed eastward shift in tracks as being driven by positive phases of the AMM along with negative phases of ENSO based on the use of historical track data. An ensemble of AGCM experiments allows us to isolate the effects of SST from those of internal atmospheric variability. Because the AM2.5-C360 model was forced with historical sea surface temperatures, all of its ensemble members would have recorded the effects of ENSO and the AMM in the same way. The fact that we did not then see a consistent shift in the distribution of track locations across all of the ensemble members implies that ENSO and the AMM were not entirely responsible for the historically observed eastward shift in tracks and that even the sign of the shift that was observed was not determined by the history of SST. These results are consistent with the multicentennial analysis of Bahamian hurricanes of Wallace et al. (2020) and argue that multidecadal changes in regional hurricane activity likely have a substantial random (“weather”) component.

We note four potential sources of error in our model experiments. First, one variable with a large influence on the model’s accuracy of tropical cyclone simulations is the accuracy of the historical sea surface temperatures used to generate the predictions (Chan et al. 2021). Although our analysis covers a relatively recent period, there is likely to be error in the sea surface temperature record for the early part of the study period in the 1970s and 1980s before the use of microwave imaging sensors (Minnett et al. 2019). Second, there could be additional error in the observations of our historical track dataset, but such errors are expected to be small since all the observations came from the satellite observation era (post-1965) (Vecchi and Knutson 2008). Third, it is possible that the 10 ensemble members did not correctly reflect the true distribution of the track predictions of the AM2.5-C360 model. The sample size of 10 could have led to a bias in the bootstrap estimation of the distributions. Finally, the model could be wrongly estimating the amplitude of weather noise. Weather noise in our model emerges from nonlinear terms in the model equations, but the same effect may also be achieved using stochastic physics schemes (Vidale et al. 2021). We suggest that further studies with other high-resolution AGCMs be conducted to test this possibility.

b. Influence of genesis location and sea level pressure variability on weather-driven multidecadal trends in hurricane activity

We next investigated some of the causes behind the observed multidecadal weather-driven variability in the spatial distribution of tropical cyclone tracks. First, the correlations of genesis location and sea level pressure with extreme shifts in model track distribution were considered, and second, a new statistical track prediction model was developed to compare the relative roles of genesis location and steering flow in producing shifts in track distribution over time.

We found that although genesis location was one of ten variables used to define the cluster assignments of tropical cyclone tracks, knowing genesis location was not enough to consistently predict the cluster that a storm would be assigned to (Fig. 4). There was particularly high uncertainty about track cluster partitioning based on genesis location for the storms that formed in the main development region to the south of 20°N. As a result, our results surrounding the importance of genesis location in driving differences in track distribution are unlikely to be simply an artifact of genesis location being used to define cluster partitioning. In addition, we observed that for model scenarios with extreme shifts in track distribution over multidecadal time scales, movement in overall track distribution was correlated with a significant shift of genesis location in the same direction (Fig. 8). The observed and model ensemble mean shift in track distribution between the pre-1995 and post-1995 periods was also associated with a shift in genesis location (Fig. 7). Similarly, extreme shifts in track distribution were associated with different changes in the climatology of sea level pressure (Fig. 8). Although it appeared that all ensemble members showed a consistent pattern of change in sea level pressure over the pre-1995 to post-1995 periods, there was a significant difference between the sea level pressure shift associated with westward-shifting tracks and eastward-shifting tracks (Fig. 8). The sea level pressure shift associated with westward shifts in track had a higher increase in pressure over the Atlantic between 15° and 50°N, corresponding to a geostrophic flow on the order of 0.1 m s−1 primarily to the west, a small but not negligible change. These correlations provide evidence that both steering flow and genesis location could be driving changes in track distribution. To better elucidate the relative importance, we employed a regression model.

The regression model was run with covariates selected for their physical relevance to track movement (beta drift and large-scale steering flow). It produced coefficients that were generally consistent with expectations of the physical roles of beta and advection in driving track movement. The statistical track model was successful at capturing the general recurving structure of tracks and the location of tracks on an annual mean level (Fig. A1), but it had more error for individual tracks compared to more complex models like the National Hurricane Center’s forecast model (Cangialosi et al. 2020; Fig. A2). Directional biases were present as well (Fig. A2), but they had a high level of variance, and the fact that the regression model simulations of track were used to model shifts in tracks rather than absolute positions of tracks likely diminished the negative effects of this bias. With these caveats in mind, we restrict the use of the regression model to simulating multidecadal shifts in track distribution as were present in model bootstrap ensemble members, where it showed appreciable skill (Fig. 9), making it suitable for use in analyzing the causes behind these shifts. This error assessment is consistent with other studies that have considered Beta and Advection Models to be sufficiently accurate to assess the sensitivity of tropical cyclone tracks to factors like genesis location and steering flow but not accurate enough to be reliably used for forecasting on the level of an individual track (Colbert and Soden 2012).

A set of experiments was run to examine the extent to which the weather-driven multidecadal variability in track could have been produced without any year-to-year variability in genesis location or without any year-to-year variability in sea level pressure. We find that extreme scenarios of large weather-driven multidecadal shifts in track distribution were replicated better by regression model simulations with year-on-year changes in genesis location than simulations with year-on-year changes in sea level pressure (Figs. 11 and 12). For the simulations run with the 6-hourly sea level pressure data model, the shifting scenarios were replicated close to equally well by our regression model simulations with or without year-to-year changes in sea level pressure (Figs. 11 and 12). This was also true of regression model simulations of the multidecadal shift in track distribution of the model ensemble mean (Fig. 10). These findings support the hypothesis that genesis location is critical to weather-driven changes in track distribution. While the track distribution changes were shown to be associated with some consistent sea level pressure shifts on multidecadal time scales (Fig. 8), the sea level pressure changes were small in magnitude and on their own were insufficient to drive the shifts in the spatial distribution of tracks.

Colbert and Soden (2012) discussed the relative roles of genesis location and steering flow in driving changes in track distribution due to internal climate variability, specifically ENSO and the AMM. Using simulations from a Beta and Advection Model, they attributed the ENSO effects on track to be due to changes in steering flow but determined that the AMM effects on track were due to changes in genesis location. The differences in mean sea level pressures in the North Atlantic between El Niño and La Niña years discussed by Colbert and Soden (2012) were on the scale of 2 mb, while the scale of the difference in mean sea level pressures from pre-1995 to post-1995 for the westward- and eastward-shifting scenarios in this study were only on the order of less than 0.5 mb (Fig. 8). Therefore, it seems reasonable that these multidecadal shifts in track may have a different cause than ENSO driven shifts in track, as weather variability on multidecadal scales produced much smaller changes in steering flow than ENSO variability.

Another related paper is Sainsbury et al. (2022), which considers the frequency of recurving tropical cyclones in the North Atlantic from 1979 to 2018. Although this study differs in that it focuses more specifically on recurving versus nonrecurving tropical cyclones, they similarly find that interannual variability in steering flow plays only a minor role in driving variability in recurving tropical cyclone track frequency. They instead attribute most of the interannual changes in recurving tropical cyclone frequency to the frequency of genesis in the main development region and the subtropical North Atlantic. Our results further support this finding that interannual variability in genesis location is an important driver of interannual variability in hurricane tracks.

c. Implications

These results suggest that even if future climate impact on SSTs was completely predictable and the climate effects on hurricane frequency were completely understood, a substantial degree of residual uncertainty could remain in predictions of regional hurricane impacts due to the influence of weather on shifts in track location.

This study also highlights the need for future work investigating the role of weather in driving shifts in tropical cyclone tracks. Although we investigate multidecadal shifts in tropical cyclone tracks, a similar framework of analysis could be applied to evaluate the role of weather variability in driving shifts in track distribution on shorter time scales, such as that of the ENSO signal, or in forecasted scenarios.

The finding that weather-driven shifts in the spatial distribution of tropical cyclone tracks can be attributed to changes in genesis location underscores the importance of understanding the factors that influence hurricane genesis location. The ability of models to accurately capture changes in genesis location could have a big impact on their ability to describe shifts in tracks and correctly capture the impacts of tropical cyclones. Our results also highlight the need to develop a better understanding of weather effects on genesis location. Large-scale indices on tropical cyclogenesis (e.g., Emanuel and Nolan 2004; Hsieh et al. 2020; Wang and Murakami 2020) have been developed to quantify the climate signal, but weather-scale variability is also important.

5. Conclusions

A 10-member ensemble of AM2.5-C360 model tropical cyclone tracks from 1970 to 2021 (with consistent forced and internal climate variability between ensemble members) was analyzed to assess the role of weather and climate variability in producing multidecadal changes in hurricane activity.

We find that the historically observed increase in tropical cyclone frequency over this time period is likely to have been modulated by changes in climate, in accordance with past studies that have attributed this factor to positive phases of the AMO and the AMM and radiative forcing, particularly from aerosols (Goldenberg et al. 2001; Kossin and Vimont 2007; Klotzbach and Gray 2008; Klotzbach 2011; Booth et al. 2012). We have not attempted to isolate the impacts of radiatively forced SST changes and those arising from climate variabiliity, both of which plausibly contributed to changes from 1970 to the present (e.g., Vecchi et al. 2017). On the contrary, we find evidence that the historically observed eastward shift in the spatial distribution of tropical cyclone tracks over the period of 1970–2021 was made more likely due to changes in climate, but weather variability was substantial enough that it could have caused a shift to occur in the opposite direction as well.

A regression model (structured similarly to a Beta and Advection Model) was devised to predict tropical cyclone track movement from information about genesis location and steering flow. Simulation experiments indicated that year-to-year changes in genesis locations are a major cause of weather-driven shifts in track distribution on multidecadal time scales.

A primary implication of this work is that weather-scale random variability may have a bigger influence on hurricanes than previously believed, because of the role of weather in driving spatial shifts in hurricane tracks. In addition, these results highlight the importance of understanding shifts in genesis location when studying movements in tracks.

Acknowledgments.

We thank Michael Oppenheimer, Jonathan Hanke, and Samuel Kortum for discussion and comments. This work was supported in part by NSF-2202784 (G.A.V. and W.Y.), DOC/NOAA NA18OAR4320123 (T-L.H. and G.A.V.), and the Carbon Mitigation Initiative at Princeton University. The AM2.5-C360 simulations were performed on computational resources managed and supported by Princeton Research Computing, a consortium of groups including the Princeton Institute for Computational Science and Engineering, the Office of Information Technology’s High Performance Computing Center, and the Visualization Laboratory at Princeton University.

Data availability statement.

Source code of the AM2.5 model is available from https://www.gfdl.noaa.gov/cm2-5-and-flor-quickstart/. IBTrACS data are available at https://www.ncei.noaa.gov/data/international-best-track-archive-for-climate-stewardship-ibtracs/v04r00/access/netcdf/. The code used to produce all figures in this paper is available at https://github.com/gkortum/hurricane-tracks-code.

APPENDIX

Statistical Model Validation

a. Visualization of regression model simulated tracks

To provide a qualitative sense of the regression model’s simulation capacity, examples of simulated tracks predicted from the monthly sea level pressure regression model are shown in Fig. A1. In general, the regression model captures the storms’ direction of movement and recurving structure. At an individual track level, the regression model does not fully capture all of the track shapes; for example, the regression model tracks generally do not extend as far north past 45°N (Fig. A1). One reason for these differences at an individual track level might be the model’s definition of steering flow. Generally, there is a close match between the mean trajectories for the real and predicted tracks. This was common across the dataset: despite some variation in the quality of the predictions of individual tracks, the statistical model produced reasonable predictions of the mean track for most ensemble years.

Fig. A1.
Fig. A1.

Each row shows (left) a comparison of the “actual” (AM2.5-C360 model–generated) and regression model–simulated (“predicted”) tracks for a given ensemble member and year and (right) a comparison of the mean trajectories of the real and simulated tracks. The mean trajectories were produced by interpolating all the storms to have length 30, then taking their mean. In this way, the first point in the mean trajectory is located at the mean genesis location, and the last point in the mean trajectory is located at the mean lysis location; however, individual points in the tracks are no longer spaced at a time step of 6 h.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

b. Track simulation prediction errors

The average prediction errors over time in the track simulations were also assessed in terms of the magnitude and direction of the divergence of the simulated tracks from the actual tracks. Starting at each storm’s genesis point, the average distance between the actual and predicted tracks was calculated for 120 h into the future after genesis (Fig. A2). For a sense of scale, prediction errors were compared to the prediction errors of a naive model, which did not move at all, as well as to the National Hurricane Center’s forecast model (Cangialosi et al. 2020), although this model’s prediction errors were computed for hours into the future from any point, not just hours into the future from the genesis point. The standard deviations of the regression prediction models’ errors are approximately bounded by the errors of the naive model on the upper end and the National Hurricane Center model on the lower end. The directional errors indicate that that both models tend to predict track movements slightly to the northeast of the actual track it intends to simulate with the 6-hourly regression model having less directional bias. The standard deviations of these biases were large, spanning nearly 180°.

Fig. A2.
Fig. A2.

(left) Error in the regression model track predictions for time increments into the future starting with 0 km at the given genesis point. Shaded regions represent one standard deviation. (right) Directional bias of the regression model predictions relative to the actual track movement starting at the genesis point. The two lines start at the origin at time t = 0 when the storm is at its genesis point and the error is 0. Then, over time, they move outward along the radial axis, which represents kilometers of distance between the predicted and real track, while the color of the scatter points represent hours elapsed since genesis. The angle represents the direction that the predicted storm is in relative to the “actual” (AM2.5-C360 model–generated) storm at a given time step. Shaded regions represent one standard deviation.

Citation: Journal of Climate 37, 5; 10.1175/JCLI-D-23-0088.1

REFERENCES

  • Booth, B. B. B., N. J. Dunstone, P. R. Halloran, T. Andrews, and N. Bellouin, 2012: Aerosols implicated as a prime driver of twentieth-century North Atlantic climate variability. Nature, 484, 228232, https://doi.org/10.1038/nature10946.

    • Search Google Scholar
    • Export Citation
  • Cangialosi, J. P., E. Blake, M. DeMaria, A. Penny, A. Latto, E. Rappaport, and V. Tallapragada, 2020: Recent progress in tropical cyclone intensity forecasting at the National Hurricane Center. Wea. Forecasting, 35, 19131922, https://doi.org/10.1175/WAF-D-20-0059.1.

    • Search Google Scholar
    • Export Citation
  • Chan, D., G. A. Vecchi, W. Yang, and P. Huybers, 2021: Improved simulation of 19th- and 20th-century North Atlantic hurricane frequency after correcting historical sea surface temperatures. Sci. Adv., 7, eabg6931, https://doi.org/10.1126/sciadv.abg6931.

    • Search Google Scholar
    • Export Citation
  • Chen, J.-H., and S.-J. Lin, 2011: The remarkable predictability of inter-annual variability of Atlantic hurricanes during the past decade. Geophys. Res. Lett., 38, L11804, https://doi.org/10.1029/2011GL047629.

    • Search Google Scholar
    • Export Citation
  • Chen, J.-H., and S.-J. Lin, 2013: Seasonal predictions of tropical cyclones using a 25-km-resolution general circulation model. J. Climate, 26, 380398, https://doi.org/10.1175/JCLI-D-12-00061.1.

    • Search Google Scholar
    • Export Citation
  • Colbert, A. J., and B. J. Soden, 2012: Climatological variations in North Atlantic cyclone tracks. J. Climate, 25, 657673, https://doi.org/10.1175/JCLI-D-11-00034.1.

    • Search Google Scholar
    • Export Citation
  • Colbert, A. J., B. J. Soden, G. A. Vecchi, and B. P. Kirtman, 2013: The impact of anthropogenic climate change on North Atlantic tropical cyclone tracks. J. Climate, 26, 40884095, https://doi.org/10.1175/JCLI-D-12-00342.1.

    • Search Google Scholar
    • Export Citation
  • Delworth, T. L., and M. E. Mann, 2000: Observed and simulated multidecadal variability in the Northern Hemisphere. Climate Dyn., 16, 661676, https://doi.org/10.1007/s003820000075.

    • Search Google Scholar
    • Export Citation
  • Delworth, T. L., and Coauthors, 2012: Simulated climate and climate change in the GFDL CM2.5 high-resolution coupled climate model. J. Climate, 25, 27552781, https://doi.org/10.1175/JCLI-D-11-00316.1.

    • Search Google Scholar
    • Export Citation
  • Efron, B., and R. Tibshirani, 1991: Statistical data analysis in the computer age. Science, 253, 390395, https://doi.org/10.1126/science.253.5018.390.

    • Search Google Scholar
    • Export Citation
  • Elsner, J. B., 2003: Tracking hurricanes. Bull. Amer. Meteor. Soc., 84, 353356, https://doi.org/10.1175/BAMS-84-3-353.

  • Elsner, J. B., K.-B. Liu, and B. Kocher, 2000: Spatial variations in major U.S. hurricane activity: Statistics and a physical mechanism. J. Climate, 13, 22932305, https://doi.org/10.1175/1520-0442(2000)013<2293:SVIMUS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Emanuel, K., and D. S. Nolan, 2004: Tropical cyclone activity and global climate system. 26th Conf. on Hurricanes and Tropical Meteorology, Miami, FL, Amer. Meteor. Soc., 10A.2, https://ams.confex.com/ams/pdfpapers/75463.pdf.

  • Emanuel, K., S. Ravela, E. Vivant, and C. Risi, 2006: A statistical deterministic approach to hurricane risk assessment. Bull. Amer. Meteor. Soc., 87, 299314, https://doi.org/10.1175/BAMS-87-3-299.

    • Search Google Scholar
    • Export Citation
  • Goldenberg, S. B., C. W. Landsea, A. M. Mestas-Nunez, and W. M. Gray, 2001: The recent increase in Atlantic hurricane activity: Causes and implications. Science, 293, 474479, https://doi.org/10.1126/science.1060040.

    • Search Google Scholar
    • Export Citation
  • Hall, T. M., and S. Jewson, 2007: Statistical modeling of North Atlantic tropical cyclone tracks. Tellus, 59A, 486498, https://doi.org/10.1111/j.1600-0870.2007.00240.x.

    • Search Google Scholar
    • Export Citation
  • Hall, T. M., J. P. Kossin, T. Thompson, and J. McMahon, 2021: US tropical cyclone activity in the 2030s based on projected changes in tropical sea surface temperature. J. Climate, 34, 13211335, https://doi.org/10.1175/JCLI-D-20-0342.1.

    • Search Google Scholar
    • Export Citation
  • Harris, L. M., S.-J. Lin, and C. Tu, 2016: High-resolution climate simulations using GFDL HiRAM with a stretched global grid. J. Climate, 29, 42934314, https://doi.org/10.1175/JCLI-D-15-0389.1.

    • Search Google Scholar
    • Export Citation
  • Hsieh, T.-L., G. A. Vecchi, W. Yang, I. M. Held, and S. T. Garner, 2020: Large-scale control on the frequency of tropical cyclones and seeds: A consistent relationship across a hierarchy of global atmospheric models. Climate Dyn., 55, 31773196, https://doi.org/10.1007/s00382-020-05446-5.

    • Search Google Scholar
    • Export Citation
  • Hsieh, T.-L., W. Yang, G. A. Vecchi, and M. Zhao, 2022: Model spread in the tropical cyclone frequency and seed propensity index across global warming and ENSO-like perturbations. Geophys. Res. Lett., 49, e2021GL097157, https://doi.org/10.1029/2021GL097157.

    • Search Google Scholar
    • Export Citation
  • Klotzbach, P. J., 2011: The influence of El Niño–Southern Oscillation and the Atlantic multidecadal oscillation on Caribbean tropical cyclone activity. J. Climate, 24, 721731, https://doi.org/10.1175/2010JCLI3705.1.

    • Search Google Scholar
    • Export Citation
  • Klotzbach, P. J., and W. M. Gray, 2008: Multidecadal variability in North Atlantic tropical cyclone activity. J. Climate, 21, 39293935, https://doi.org/10.1175/2008JCLI2162.1.

    • Search Google Scholar
    • Export Citation
  • Knapp, K. R., S. Applequist, H. J. Diamond, J. P. Kossin, M. Kruk, and C. Schreck, 2010: NCDC International Best Track Archive for Climate Stewardship (IBTrACS) project, version 3. NOAA National Centers for Environmental Information, accessed 10 January 2022, https://doi.org/10.7289/V5NK3BZP.

  • Kossin, J. P., and D. J. Vimont, 2007: A more general framework for understanding Atlantic hurricane variability and trends. Bull. Amer. Meteor. Soc., 88, 17671782, https://doi.org/10.1175/BAMS-88-11-1767.

    • Search Google Scholar
    • Export Citation
  • Kossin, J. P., S. J. Camargo, and M. Sitkowski, 2010: Climate modulation of North Atlantic hurricane tracks. J. Climate, 23, 30573076, https://doi.org/10.1175/2010JCLI3497.1.

    • Search Google Scholar
    • Export Citation
  • Lee, C.-Y., M. K. Tippett, A. H. Sobel, and S. J. Camargo, 2018: An environmentally forced tropical cyclone hazard model. J. Adv. Model. Earth Syst., 10, 223241, https://doi.org/10.1002/2017MS001186.

    • Search Google Scholar
    • Export Citation
  • Liu, M., G. A. Vecchi, J. A. Smith, and H. Murakami, 2017: The present-day simulation and twenty-first-century projection of the climatology of extratropical transition in the North Atlantic. J. Climate, 30, 27392756, https://doi.org/10.1175/JCLI-D-16-0352.1.

    • Search Google Scholar
    • Export Citation
  • MacQueen, J., 1967: Some methods for classification and analysis of multivariate observations. Proc. Fifth Berkeley Symp. on Mathematical Statistics and Probability, Berkeley, CA, University of California, 281–297, https://projecteuclid.org/ebooks/berkeley-symposium-on-mathematical-statistics-and-probability/Some-methods-for-classification-and-analysis-of-multivariate-observations/chapter/Some-methods-for-classification-and-analysis-of-multivariate-observations/bsmsp/1200512992.

  • Mann, M. E., and K. A. Emanuel, 2006: Atlantic hurricane trends linked to climate change. Eos, Trans. Amer. Geophys. Union, 87, 233241, https://doi.org/10.1029/2006EO240001.

    • Search Google Scholar
    • Export Citation
  • Merlis, T. M., W. Zhou, I. M. Held, and M. Zhao, 2016: Surface temperature dependence of tropical cyclone-permitting simulations in a spherical model with uniform thermal forcing. Geophys. Res. Lett., 43, 28592865, https://doi.org/10.1002/2016GL067730.

    • Search Google Scholar
    • Export Citation
  • Minnett, P. J., and Coauthors, 2019: Half a century of satellite remote sensing of sea-surface temperature. Remote Sens. Environ., 233, 111366, https://doi.org/10.1016/j.rse.2019.111366.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., 2022: Substantial global influence of anthropogenic aerosols on tropical cyclones over the past 40 years. Sci. Adv., 8, eabn9493, https://doi.org/10.1126/sciadv.abn9493.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., and B. Wang, 2010: Future change of North Atlantic tropical cyclone tracks: Projection by a 20-km-mesh global atmospheric model. J. Climate, 23, 26992721, https://doi.org/10.1175/2010JCLI3338.1.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., and Coauthors, 2015: Simulation and prediction of category 4 and 5 hurricanes in the high-resolution GFDL HiFLOR coupled climate model. J. Climate, 28, 90589079, https://doi.org/10.1175/JCLI-D-15-0216.1.

    • Search Google Scholar
    • Export Citation
  • Ng, C. H. J., and G. A. Vecchi, 2020: Large-scale environmental controls on the seasonal statistics of rapidly intensifying North Atlantic tropical cyclones. Climate Dyn., 54, 39073925, https://doi.org/10.1007/s00382-020-05207-4.

    • Search Google Scholar
    • Export Citation
  • Pielke, R. A., Jr., J. Gratz, C. W. Landsea, D. Collins, M. A. Saunders, and R. Musulin, 2008: Normalized hurricane damage in the United States: 1900–2005. Nat. Hazards Rev., 9, 2942, https://doi.org/10.1061/(ASCE)1527-6988(2008)9:1(29).

    • Search Google Scholar
    • Export Citation
  • Putman, W. M., and S.-J. Lin, 2007: Finite-volume transport on various cubed-sphere grids. J. Comput. Phys., 227, 5578, https://doi.org/10.1016/j.jcp.2007.07.022.

    • Search Google Scholar
    • Export Citation
  • Ramsay, H. A., S. J. Camargo, and D. Kim, 2012: Cluster analysis of tropical cyclone tracks in the Southern Hemisphere. Climate Dyn., 39, 897917, https://doi.org/10.1007/s00382-011-1225-8.

    • Search Google Scholar
    • Export Citation
  • Sainsbury, E. M., R. K. H. Schiemann, K. I. Hodges, A. J. Baker, L. C. Shaffrey, and K. T. Bhatia, 2022: What governs the interannual variability of recurving North Atlantic tropical cyclones? J. Climate, 35, 36273641, https://doi.org/10.1175/JCLI-D-21-0712.1.

    • Search Google Scholar
    • Export Citation
  • Smith, A. B., and R. W. Katz, 2013: US billion-dollar weather and climate disasters: Data sources, trends, accuracy and biases. Nat. Hazards, 67, 387410, https://doi.org/10.1007/s11069-013-0566-5.

    • Search Google Scholar
    • Export Citation
  • Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of CMIP5 and the experiment design. Bull. Amer. Meteor. Soc., 93, 485498, https://doi.org/10.1175/BAMS-D-11-00094.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and B. J. Soden, 2007: Increased tropical Atlantic wind shear in model projections of global warming. Geophys. Res. Lett., 34, L08702, https://doi.org/10.1029/2006GL028905.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and T. R. Knutson, 2008: On estimates of historical North Atlantic tropical cyclone activity. J. Climate, 21, 35803600, https://doi.org/10.1175/2008JCLI2178.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and T. R. Knutson, 2011: Estimating annual numbers of Atlantic hurricanes missing from the HURDAT database (1878–1965) using ship track density. J. Climate, 24, 17361746, https://doi.org/10.1175/2010JCLI3810.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., T. L. Delworth, and B. Booth, 2017: Origins of Atlantic decadal swings. Nature, 548, 284285, https://doi.org/10.1038/nature23538.

    • Search Google Scholar
    • Export Citation
  • Vickery, P. J., P. F. Skerlj, and L. A. Twisdale, 2000: Simulation of hurricane risk in the US using an empirical track model. J. Struct. Eng., 126, 12221237, https://doi.org/10.1061/(ASCE)0733-9445(2000)126:10(1222).

    • Search Google Scholar
    • Export Citation
  • Vidale, P. L., and Coauthors, 2021: Impact of stochastic physics and model resolution on the simulation of tropical cyclones in climate GCMs. J. Climate, 34, 43154341, https://doi.org/10.1175/JCLI-D-20-0507.1.

    • Search Google Scholar
    • Export Citation
  • Villarini, G., and G. A. Vecchi, 2013: Projected increases in North Atlantic tropical cyclone intensity from CMIP5 models. J. Climate, 26, 32313240, https://doi.org/10.1175/JCLI-D-12-00441.1.

    • Search Google Scholar
    • Export Citation
  • Wallace, E. J., S. Coats, K. Emanuel, and J. P. Donnelly, 2020: Centennial-scale shifts in storm frequency captured in paleohurricane records from the Bahamas arise predominantly from random variability. Geophys. Res. Lett., 48, e2020GL091145, https://doi.org/10.1029/2020GL091145.

    • Search Google Scholar
    • Export Citation
  • Walsh, K. J. E., and Coauthors, 2015: Hurricanes and climate: The U.S. CLIVAR working group on hurricanes. Bull. Amer. Meteor. Soc., 96, 9971017, https://doi.org/10.1175/BAMS-D-13-00242.1.

    • Search Google Scholar
    • Export Citation
  • Wang, B., and H. Murakami, 2020: Dynamic genesis potential index for diagnosing present-day and future global tropical cyclone genesis. Environ. Res. Lett., 15, 114008, https://doi.org/10.1088/1748-9326/abbb01.

    • Search Google Scholar
    • Export Citation
  • Wang, S., and R. Toumi, 2021: Recent migration of tropical cyclones toward coasts. Science, 371, 514517, https://doi.org/10.1126/science.abb9038.

    • Search Google Scholar
    • Export Citation
  • Xie, L., T. Yan, L. J. Pietrafesa, J. M. Morrison, and T. Karl, 2005: Climatology and interannual variability of North Atlantic hurricane tracks. J. Climate, 18, 53705381, https://doi.org/10.1175/JCLI3560.1.

    • Search Google Scholar
    • Export Citation
  • Yang, W., T.-L. Hsieh, and G. A. Vecchi, 2021: Hurricane annual cycle controlled by both seeds and genesis probability. Proc. Natl. Acad. Sci. USA, 118, e2108397118, https://doi.org/10.1073/pnas.2108397118.

    • Search Google Scholar
    • Export Citation
Save
  • Booth, B. B. B., N. J. Dunstone, P. R. Halloran, T. Andrews, and N. Bellouin, 2012: Aerosols implicated as a prime driver of twentieth-century North Atlantic climate variability. Nature, 484, 228232, https://doi.org/10.1038/nature10946.

    • Search Google Scholar
    • Export Citation
  • Cangialosi, J. P., E. Blake, M. DeMaria, A. Penny, A. Latto, E. Rappaport, and V. Tallapragada, 2020: Recent progress in tropical cyclone intensity forecasting at the National Hurricane Center. Wea. Forecasting, 35, 19131922, https://doi.org/10.1175/WAF-D-20-0059.1.

    • Search Google Scholar
    • Export Citation
  • Chan, D., G. A. Vecchi, W. Yang, and P. Huybers, 2021: Improved simulation of 19th- and 20th-century North Atlantic hurricane frequency after correcting historical sea surface temperatures. Sci. Adv., 7, eabg6931, https://doi.org/10.1126/sciadv.abg6931.

    • Search Google Scholar
    • Export Citation
  • Chen, J.-H., and S.-J. Lin, 2011: The remarkable predictability of inter-annual variability of Atlantic hurricanes during the past decade. Geophys. Res. Lett., 38, L11804, https://doi.org/10.1029/2011GL047629.

    • Search Google Scholar
    • Export Citation
  • Chen, J.-H., and S.-J. Lin, 2013: Seasonal predictions of tropical cyclones using a 25-km-resolution general circulation model. J. Climate, 26, 380398, https://doi.org/10.1175/JCLI-D-12-00061.1.

    • Search Google Scholar
    • Export Citation
  • Colbert, A. J., and B. J. Soden, 2012: Climatological variations in North Atlantic cyclone tracks. J. Climate, 25, 657673, https://doi.org/10.1175/JCLI-D-11-00034.1.

    • Search Google Scholar
    • Export Citation
  • Colbert, A. J., B. J. Soden, G. A. Vecchi, and B. P. Kirtman, 2013: The impact of anthropogenic climate change on North Atlantic tropical cyclone tracks. J. Climate, 26, 40884095, https://doi.org/10.1175/JCLI-D-12-00342.1.

    • Search Google Scholar
    • Export Citation
  • Delworth, T. L., and M. E. Mann, 2000: Observed and simulated multidecadal variability in the Northern Hemisphere. Climate Dyn., 16, 661676, https://doi.org/10.1007/s003820000075.

    • Search Google Scholar
    • Export Citation
  • Delworth, T. L., and Coauthors, 2012: Simulated climate and climate change in the GFDL CM2.5 high-resolution coupled climate model. J. Climate, 25, 27552781, https://doi.org/10.1175/JCLI-D-11-00316.1.

    • Search Google Scholar
    • Export Citation
  • Efron, B., and R. Tibshirani, 1991: Statistical data analysis in the computer age. Science, 253, 390395, https://doi.org/10.1126/science.253.5018.390.

    • Search Google Scholar
    • Export Citation
  • Elsner, J. B., 2003: Tracking hurricanes. Bull. Amer. Meteor. Soc., 84, 353356, https://doi.org/10.1175/BAMS-84-3-353.

  • Elsner, J. B., K.-B. Liu, and B. Kocher, 2000: Spatial variations in major U.S. hurricane activity: Statistics and a physical mechanism. J. Climate, 13, 22932305, https://doi.org/10.1175/1520-0442(2000)013<2293:SVIMUS>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Emanuel, K., and D. S. Nolan, 2004: Tropical cyclone activity and global climate system. 26th Conf. on Hurricanes and Tropical Meteorology, Miami, FL, Amer. Meteor. Soc., 10A.2, https://ams.confex.com/ams/pdfpapers/75463.pdf.

  • Emanuel, K., S. Ravela, E. Vivant, and C. Risi, 2006: A statistical deterministic approach to hurricane risk assessment. Bull. Amer. Meteor. Soc., 87, 299314, https://doi.org/10.1175/BAMS-87-3-299.

    • Search Google Scholar
    • Export Citation
  • Goldenberg, S. B., C. W. Landsea, A. M. Mestas-Nunez, and W. M. Gray, 2001: The recent increase in Atlantic hurricane activity: Causes and implications. Science, 293, 474479, https://doi.org/10.1126/science.1060040.

    • Search Google Scholar
    • Export Citation
  • Hall, T. M., and S. Jewson, 2007: Statistical modeling of North Atlantic tropical cyclone tracks. Tellus, 59A, 486498, https://doi.org/10.1111/j.1600-0870.2007.00240.x.

    • Search Google Scholar
    • Export Citation
  • Hall, T. M., J. P. Kossin, T. Thompson, and J. McMahon, 2021: US tropical cyclone activity in the 2030s based on projected changes in tropical sea surface temperature. J. Climate, 34, 13211335, https://doi.org/10.1175/JCLI-D-20-0342.1.

    • Search Google Scholar
    • Export Citation
  • Harris, L. M., S.-J. Lin, and C. Tu, 2016: High-resolution climate simulations using GFDL HiRAM with a stretched global grid. J. Climate, 29, 42934314, https://doi.org/10.1175/JCLI-D-15-0389.1.

    • Search Google Scholar
    • Export Citation
  • Hsieh, T.-L., G. A. Vecchi, W. Yang, I. M. Held, and S. T. Garner, 2020: Large-scale control on the frequency of tropical cyclones and seeds: A consistent relationship across a hierarchy of global atmospheric models. Climate Dyn., 55, 31773196, https://doi.org/10.1007/s00382-020-05446-5.

    • Search Google Scholar
    • Export Citation
  • Hsieh, T.-L., W. Yang, G. A. Vecchi, and M. Zhao, 2022: Model spread in the tropical cyclone frequency and seed propensity index across global warming and ENSO-like perturbations. Geophys. Res. Lett., 49, e2021GL097157, https://doi.org/10.1029/2021GL097157.

    • Search Google Scholar
    • Export Citation
  • Klotzbach, P. J., 2011: The influence of El Niño–Southern Oscillation and the Atlantic multidecadal oscillation on Caribbean tropical cyclone activity. J. Climate, 24, 721731, https://doi.org/10.1175/2010JCLI3705.1.

    • Search Google Scholar
    • Export Citation
  • Klotzbach, P. J., and W. M. Gray, 2008: Multidecadal variability in North Atlantic tropical cyclone activity. J. Climate, 21, 39293935, https://doi.org/10.1175/2008JCLI2162.1.

    • Search Google Scholar
    • Export Citation
  • Knapp, K. R., S. Applequist, H. J. Diamond, J. P. Kossin, M. Kruk, and C. Schreck, 2010: NCDC International Best Track Archive for Climate Stewardship (IBTrACS) project, version 3. NOAA National Centers for Environmental Information, accessed 10 January 2022, https://doi.org/10.7289/V5NK3BZP.

  • Kossin, J. P., and D. J. Vimont, 2007: A more general framework for understanding Atlantic hurricane variability and trends. Bull. Amer. Meteor. Soc., 88, 17671782, https://doi.org/10.1175/BAMS-88-11-1767.

    • Search Google Scholar
    • Export Citation
  • Kossin, J. P., S. J. Camargo, and M. Sitkowski, 2010: Climate modulation of North Atlantic hurricane tracks. J. Climate, 23, 30573076, https://doi.org/10.1175/2010JCLI3497.1.

    • Search Google Scholar
    • Export Citation
  • Lee, C.-Y., M. K. Tippett, A. H. Sobel, and S. J. Camargo, 2018: An environmentally forced tropical cyclone hazard model. J. Adv. Model. Earth Syst., 10, 223241, https://doi.org/10.1002/2017MS001186.

    • Search Google Scholar
    • Export Citation
  • Liu, M., G. A. Vecchi, J. A. Smith, and H. Murakami, 2017: The present-day simulation and twenty-first-century projection of the climatology of extratropical transition in the North Atlantic. J. Climate, 30, 27392756, https://doi.org/10.1175/JCLI-D-16-0352.1.

    • Search Google Scholar
    • Export Citation
  • MacQueen, J., 1967: Some methods for classification and analysis of multivariate observations. Proc. Fifth Berkeley Symp. on Mathematical Statistics and Probability, Berkeley, CA, University of California, 281–297, https://projecteuclid.org/ebooks/berkeley-symposium-on-mathematical-statistics-and-probability/Some-methods-for-classification-and-analysis-of-multivariate-observations/chapter/Some-methods-for-classification-and-analysis-of-multivariate-observations/bsmsp/1200512992.

  • Mann, M. E., and K. A. Emanuel, 2006: Atlantic hurricane trends linked to climate change. Eos, Trans. Amer. Geophys. Union, 87, 233241, https://doi.org/10.1029/2006EO240001.

    • Search Google Scholar
    • Export Citation
  • Merlis, T. M., W. Zhou, I. M. Held, and M. Zhao, 2016: Surface temperature dependence of tropical cyclone-permitting simulations in a spherical model with uniform thermal forcing. Geophys. Res. Lett., 43, 28592865, https://doi.org/10.1002/2016GL067730.

    • Search Google Scholar
    • Export Citation
  • Minnett, P. J., and Coauthors, 2019: Half a century of satellite remote sensing of sea-surface temperature. Remote Sens. Environ., 233, 111366, https://doi.org/10.1016/j.rse.2019.111366.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., 2022: Substantial global influence of anthropogenic aerosols on tropical cyclones over the past 40 years. Sci. Adv., 8, eabn9493, https://doi.org/10.1126/sciadv.abn9493.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., and B. Wang, 2010: Future change of North Atlantic tropical cyclone tracks: Projection by a 20-km-mesh global atmospheric model. J. Climate, 23, 26992721, https://doi.org/10.1175/2010JCLI3338.1.

    • Search Google Scholar
    • Export Citation
  • Murakami, H., and Coauthors, 2015: Simulation and prediction of category 4 and 5 hurricanes in the high-resolution GFDL HiFLOR coupled climate model. J. Climate, 28, 90589079, https://doi.org/10.1175/JCLI-D-15-0216.1.

    • Search Google Scholar
    • Export Citation
  • Ng, C. H. J., and G. A. Vecchi, 2020: Large-scale environmental controls on the seasonal statistics of rapidly intensifying North Atlantic tropical cyclones. Climate Dyn., 54, 39073925, https://doi.org/10.1007/s00382-020-05207-4.

    • Search Google Scholar
    • Export Citation
  • Pielke, R. A., Jr., J. Gratz, C. W. Landsea, D. Collins, M. A. Saunders, and R. Musulin, 2008: Normalized hurricane damage in the United States: 1900–2005. Nat. Hazards Rev., 9, 2942, https://doi.org/10.1061/(ASCE)1527-6988(2008)9:1(29).

    • Search Google Scholar
    • Export Citation
  • Putman, W. M., and S.-J. Lin, 2007: Finite-volume transport on various cubed-sphere grids. J. Comput. Phys., 227, 5578, https://doi.org/10.1016/j.jcp.2007.07.022.

    • Search Google Scholar
    • Export Citation
  • Ramsay, H. A., S. J. Camargo, and D. Kim, 2012: Cluster analysis of tropical cyclone tracks in the Southern Hemisphere. Climate Dyn., 39, 897917, https://doi.org/10.1007/s00382-011-1225-8.

    • Search Google Scholar
    • Export Citation
  • Sainsbury, E. M., R. K. H. Schiemann, K. I. Hodges, A. J. Baker, L. C. Shaffrey, and K. T. Bhatia, 2022: What governs the interannual variability of recurving North Atlantic tropical cyclones? J. Climate, 35, 36273641, https://doi.org/10.1175/JCLI-D-21-0712.1.

    • Search Google Scholar
    • Export Citation
  • Smith, A. B., and R. W. Katz, 2013: US billion-dollar weather and climate disasters: Data sources, trends, accuracy and biases. Nat. Hazards, 67, 387410, https://doi.org/10.1007/s11069-013-0566-5.

    • Search Google Scholar
    • Export Citation
  • Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of CMIP5 and the experiment design. Bull. Amer. Meteor. Soc., 93, 485498, https://doi.org/10.1175/BAMS-D-11-00094.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and B. J. Soden, 2007: Increased tropical Atlantic wind shear in model projections of global warming. Geophys. Res. Lett., 34, L08702, https://doi.org/10.1029/2006GL028905.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and T. R. Knutson, 2008: On estimates of historical North Atlantic tropical cyclone activity. J. Climate, 21, 35803600, https://doi.org/10.1175/2008JCLI2178.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., and T. R. Knutson, 2011: Estimating annual numbers of Atlantic hurricanes missing from the HURDAT database (1878–1965) using ship track density. J. Climate, 24, 17361746, https://doi.org/10.1175/2010JCLI3810.1.

    • Search Google Scholar
    • Export Citation
  • Vecchi, G. A., T. L. Delworth, and B. Booth, 2017: Origins of Atlantic decadal swings. Nature, 548, 284285, https://doi.org/10.1038/nature23538.

    • Search Google Scholar
    • Export Citation
  • Vickery, P. J., P. F. Skerlj, and L. A. Twisdale, 2000: Simulation of hurricane risk in the US using an empirical track model. J. Struct. Eng., 126, 12221237, https