Sequential Landfall of Tropical Cyclones in the United States: From Historical Records to Climate Projections

In this study, we examine sequential landfalling tropical cyclones (TCs) along U.S. East and Gulf Coasts. We find that Florida and Louisiana are most prone to sequential landfall risk. The minimal time between sequential landfalling TC has decreased for most regions since 1979, although the trend is not statistically significant given limited data. A climate projection indicates a significant increase in sequential landfalls over the 21st century under the SSP5 8.5 scenario, with the chance of a location experiencing a less‐than‐10‐day break between two TC impacts being doubled for most regions. The increases in sequential landfalls in the historical period and projected future climate are both related to increased landfall frequency, even though the storm season has been slightly expanding and may continue to expand. This study highlights a new type of TC hazard resulting from the temporal compounding of landfalls and urges the improvement of coastal resilience.

It is interesting from both scientific and social perspectives to ask if SLTs have been increasing and will increase in the future. Previous research has examined TC climatology features that are related to SLTs. Webster et al. (2005) found global increases of storm number since 1970s. Wang and Toumi (2021) found the fraction of storms entering coastal regions has an increased trend. Kossin (2018) found the TC translation speed has decreased globally, and Hall and Kossin (2019) showed the stalling times of US landfalling TCs have increased. These results indicate increasing risks from TCs but did not directly address the trend of SLTs. For future projections, previous studies have focused on global to basin-scale changes of storm intensity and frequency. Most models project increased storm intensity and decreased frequency (Knutson et al., 2020) and some models project increased storm frequency (Bhatia et al., 2018;Emanuel, 2013). Some models also project a slowdown of TC movement (Emanuel, 2021) and increased TC landfall frequencies (Emanuel 2013(Emanuel , 2021. Fewer studies have projected landfall hazards in the future (Emanuel, 2017;Marsooli et al., 2019;Xu et al., 2020), and none of the climate projection studies has focused on SLTs.
In this study, we investigate SLTs in the coastal United States. We examine the time intervals between storm landfalls and impacts, as metrics describing SLTs, and their spatial and temporal variations. We also investigate how the climatology of these time intervals may change in the future, as simulated by a synthetic TC model downscaling six CMIP6 climate models. To explore the observed and simulated spatial and temporal variations of SLTs, we apply a Poisson-Gaussian model to connect TC landfall intervals to TC landfall frequency and seasonality.

Data and Method
To study the historical climatology of SLTs in the coastal United States, we used TC observations from the International Best Track Archive for Climate Stewardship (IBTrACS). IBTrACS provides six-hourly TC locations and intensities. We used data from 1979 (the satellite era) to avoid the influence of missing records in earlier decades (Moon et al., 2019).
To study climate change effect, we applied synthetic TCs generated with a statistical-deterministic TC model (Emanuel 2006(Emanuel , 2021. The synthetic storms were generated under the environments of the NCAR/NCEP reanalysis and each of six CMIP6 climate models (CanESM5, CNRM-CM6-1, GFDL-CM-4, EC-Earth3, IP-SL-CM6A-LR, MIROC6) for control  and Shared Socioeconomic Pathway 5 8.5 scenarios (SSP5 8.5;2070-2100. 4400 and 6200 storms were generated from each climate model for control and SSP5 8.5 scenarios, respectively, and 5018 storms were generated from the reanalysis in the control period. We divided the U.S. East and Gulf coastlines into eight regions, namely Texas, Louisiana, Mississippi-Alabama, West Florida, East Florida, Georgia, South Carolina, and North Carolina (regions in higher latitudes have been seldomly affected by SLTs and thus are not investigated in this study). We also specifically examine the East Louisiana region (90.59°W to 89.29°W) as it was hit by SLTs in 2020. To focus on SLTs that are spatially close enough to cause compound impacts, we divided the coastlines into 186 mileposts with 100-km spacing in Mexico and 50-km spacing in the US (detailed in Jing & Lin, 2020) and analyzed SLTs for each milepost. We selected all storms in IBTrACS and synthetic storm simulations that approached within 250 km of a milepost (following Jing & Lin, 2020) and defined the landfall time as when the storm first entered the 250-km circle centered on the milepost. We computed the landfall intervals between storms for each milepost and defined the minimum interval over a year and across all mileposts in each defined coastal region to be the minimal landfall interval (MLI) in the year for the region. We further defined the storm impact duration (SID) for a milepost as the duration when the storm was within 250 km of the milepost, and we calculated the minimal impact interval (MII), the minimum over a year and across the mileposts for the region of the time interval between the end of the impact of the first storm and the start of the impact of the second storm. Annual landfall frequencies for each region are also calculated.
We considered storms that reach tropical storm intensity (>17.5 m/s) at landfall in the historical analysis. For synthetic analysis with larger samples, we examined MLI and MII also for hurricanes (>32.5 m/s at landfall). We estimated the probability distribution and return period for MLI and MII in the control simulations and climate-model projections. The climate projections of MLI and MII were bias-corrected. We bias-corrected the future projection of storm frequency and cumulative density functions (CDFs) of MLI and MII in each model by comparing the climate-model and reanalysis-based estimations for the control period. The storm frequency was bias-corrected through a multiplicative factor, and CDFs of MLI and MII were bias-corrected through quantile-quantile-mapping. Besides the MLI and MII distributions for each of the six climate models, we also calculated the weighted average of the six models, by assigning weights to the models based on the mean square error of the model-simulated CDFs of MLI and MII for the control period compared to reanalysis-based simulations. Marsooli et al. (2019) applied similar methods to perform bias correction and model combination for storm surge projections.
To understand the connections between MLI and landfall frequency and seasonality, we propose a Poisson-Gaussian landfall model. The model assumes storm arrivals in a region to form a nonstationary Poisson process. The Poisson rate of storm landfall is modulated by interannual and seasonal variations, which can be described as the interannual variation, is the landfall frequency in year t, and   E S s , the seasonal variation, is the likelihood of landfall occurrence on day s relative to the likelihood of occurrence during the season. We model the seasonality with a Gaussian function, historical observations and climate simulations. To study how these parameters have changed in the historical period, we divided the historical period (1979-2019) into two periods (1979-1999 and 2000-2019) and fitted the model for each period. To study how these parameters may change from the control to the future climate, we fitted the model for the control and future periods for each climate model. We used the method of maximum likelihood for model fitting. We performed Monte Carlo simulations of TC landfall using the model fitted by the historical observation to validate the model.
The Poisson-Gaussian model assumes storms are conditionally independent (given the environmental condition), and so does the synthetic storm modeling, as feedbacks of storms to the environment is not captured in the one-way coupling system. Physical interactions between storms are possible but rare (Brand, 1970;Hoover, 1961;Schenkel, 2016;Xu et al., 2013). Also, previous research has examined the independence assumption on storm genesis and landfalls and found that the Poisson distribution generally fits the observations well (e.g., Lin et al., 2012;Rumpf et al., 2009;Wahiduzzaman et al., 2021).

Historical Record of SLTs
We first use East Louisiana as an example to study the historical variation of the annual landfall frequency, mean SID, and MLI. Figures 1a-1d shows that, for East Louisiana, the landfall frequency has increased by 0.0167 ± 0.0156 storms/year since 1979. SID has increased by 0.016 ± 0.011 days/year since 1979. Also, MLI has decreased (−0.909 ± 0.508 days/year) since 1979, implying that the chance of SLTs has increased. The increasing trends of the landfall frequency and mean SID pass the Mann-Kendall significance test (95% confidence level), but the trend of MLI cannot pass the test, possibly due to data limitations (only data in years with at least two landfalling storms at a milepost are used to calculate MLI). The probability distribution of yearly MLI is fitted with a Generalized Pareto Distribution (GPD) (Figure 1d). The shape parameter of the fitted GPD is negative, implying that the GPD has a light tail and the yearly MLI is unlikely to take values much larger than the mean.
We extended the above analysis to the other coastal regions (Figures 1e-1h . Although the trend cannot pass the Mann-Kendall test due to data limitation, the increased landfall frequency indicates increased potential for short MLI (see discussion with the theoretical modeling below). The decrease in MLI and increase in SID implies a decrease of MII (not shown) and thus an increase of impact from SLTs for most coastal areas. We also fit GPD to MLI for each region; Figure 1h shows the scale of the fitted GPD. Small value of the scale indicates the region is more likely to experience short MLI. The lowest values of scale are found in East Florida (24.41), West Florida (32.02), and Louisiana (37.79), meaning that these locations are most prone to the threat of SLTs. Shape parameters (not shown) for all regions are negative except for Florida, implying that MLI is unlikely to take values much larger than the mean in most regions.
To better understand the variation of MLI, we investigate the connection between MLI and TC climatology features. Figure 2a shows the relationship between MLI and the annual landfall frequency in the historical record. Smaller MLI is associated with larger landfall frequency. Linear fitting of mean MLI to annual landfall frequency indicates that one more landfall on average causes a 7.15 ± 2.4 day shortening of MLI. Figure 2b shows the relationship between the fitted E λ (annual frequency averaged over each period) and E σ (seasonality variation averaged over each period) in the historical record, and how they jointly influence MLI. We find that E λ and E σ are positively correlated, implying that as the storm frequency increases (Figure 1e), storm season has also been slightly expanding (i.e., seasonal distribution being slightly flattened). Figure 2b also shows that lower MLI is associated with smaller E σ and larger E λ although the pattern appears unclearly due to data limitations. Thus, we perform idealized simulations using the Poisson-Gaussian model; Figure 2c shows the expectation of MLI (E(T)) in the  E λ σ space. The theoretical results show that given simulations. The small discrepancies of simulations and observations are related to the model assumption of storm independency and statistical uncertainties. The result indicates that MLI has a decreasing potential over the historical period mainly because the storm landfall frequency has increased. Although the storm season has been slightly expanding (i.e., TC activity in the off-peak season has increased relatively more than in the peak season), the increase of TC activity in the peak season is still significant and likely responsible for the decrease of MLI.

Climate Projection of SLTs
In this section we use the synthetic storm data set to examine SLTs under climate change. Figure 3 examines the CDF of MLI. We find that the probability for short TC-MLI (MLI between TCs; Figures 3a-3h) increases from the control to the future climate (e.g., in Louisiana the probability of MLI <20 days increases from 0.49 to 0.65), while the predicted degree of increase varies among climate models. The probability for short hurricane-MLI (MLI between hurricanes; Figures 3i-3p) also increases (e.g., in Louisiana the probability of hurricane-MLI <20 days increases from 0.52 to 0.61), and the changes in the Gulf Coast regions and Florida are larger than those in East Coast regions. The overall chance of the hurricane-MLI being smaller than 20 days for the Gulf Coast (Texas to West Florida) increases from 0.46 to 0.59 and for East Coast (East Florida to North Carolina) increases from 0.46 to 0.56 (not shown). All changes pass the two-sample Kolmogorov-Smirnov test on differences in distributions (with 5% significance level) and are hence statistically significant.
To assess overall sequential landfall risk, we combine MLI, SID, and multi-landfall frequency to compute return period of MII (Figure 4). The change in MII is significant, which is dominated by the change in MLI and landfall frequency rather than SID (not shown). For example, the return period of a 10-day TC-MII The CDF is calculated based on data in years when at least two TCs (or hurricanes) make landfall at a milepost in the region. Blue dots show observations (for TCs only), black curve shows the reanalysis-based simulation for the historical period, red curve shows the weighted average projection for the future climate, and the spread of the six climate models is shown as error bars. Bias-correction based on the reanalysis simulation is applied for the climate projections. (For North Carolina, three of six climate models (CanESM5, CNRM-CM6-1, IPSL-CM6A-LR) cannot predict SLTs for hurricanes in the historical period due to limited number of intense storms in historical simulations, so we omit the SSP5 8.5 simulations from these climate models for this location.).
would decrease by about half for most regions from the control to future climate (e.g., in Louisiana the return period decreases from 15 to 8 years). We found more significant increase of the chance of experiencing a short hurricane-MII (break between hurricane impacts). In Louisiana the return period of a 20-day hurricane-MII would decrease from 35 to 4 years. For the Gulf Coast (Texas to West Florida) the return period of a 20-day hurricane-MII would decrease from 12 to one year, and for the East Coast (East Florida to North Carolina) it will decrease from 25 to two years (not shown).
To better understand the change of SLTs in the future, we examine the change of annual landfall frequency E λ and seasonality E σ in the control and future climate simulations from the six climate models. We find that the averages of both E λ and E σ increase in the future. The increase in average E λ from the control to the future climate ranges from 2.38 to 6.31 storms across the regions while the increase in average E σ ranges from 3.69 to   (Emanuel, 2021). The projected change in E λ over the 21st century is beyond the historical variation from the period of 1979-1999 to the period of 2000-2019 (increase by up to 0.9 storms) while the increase in E σ is within the historical variation (increase by up to 9.6 days). According to the Poisson-Gaussian model (Figure 2c), such changes in storm frequency and seasonality lead to the significant decrease of MLI (e.g., in Louisiana expected MLI decreases from around 17-6 days: Figure 3) and MII (Figure 4), resulting from increased storm activity in the peak season.

Conclusions
This study examines various climatological features of SLTs using observation and climate model projections for U.S. East and Gulf Coasts. We describe SLTs with storm landfall interval (MLI), impact duration (SID), and impact interval (MII). We find East Florida, West Florida, and Louisiana are most prone to sequential SLTs risk. Except for Texas, the landfall frequency and annual mean SID have significantly increased since 1979, with the largest increase in landfall frequency found in Mississippi-Alabama (+0.04 ± 0.0182 storms/ year) and the largest increase in SID found in Louisiana (+0.0286 ± 0.0105 days/year). MLI decreases for most regions, except West Florida and Mississippi-Alabama, by up to 0.94 ± 0.75 days/year (South Carolina). Although the trend is not statistically significant due to data limitation, the increase of landfalling storms indicate increased potential for short MLI for most regions in the U.S. Coasts. The decreasing potential of MLI and increases in landfall frequency and SID indicate decreased MII and thus an increased impact from SLTs for most coastal areas. Applying a Poisson-Gaussian model, we found the decrease of MLI is consistent with the increase of annual landfall frequency, although the storm season has also been slightly expanding. Previous findings of more storms (Wang & Toumi, 2021;Webster et al., 2005) and longer storm stalling time (Hall & Kossin, 2019) and our results of increased likelihood of temporally compounding storms consistently indicate an increased TC threat along the U.S. coastlines over the past decades.
Climate projections using a synthetic storm model indicate SLTs will significantly increase under the SSP5 8.5 scenario, especially for the Gulf Coast. The return period of a 10-day MII between TCs would decrease by about half for most coastal regions over the 21st century (e.g., in Louisiana from 15 to 8 years). The chance of experiencing a short MII between hurricanes would increase more significantly (e.g., in Louisiana the return period of a 20-day hurricane-MII would decrease from 35 to 4 years). This large increase in SLTs is consistent with the projected significant increase of annual landfall frequency in the model (increased by 2.38-6.31 storms on average over the 21st century for US coast) although the storm season is projected to be slightly expanding. Previous research has projected a global slowdown of storm translation speed (Emanuel, 2021) and more landfalling storms in the future (Emanuel 2013(Emanuel , 2021. Consistent with the previous findings, this study shows explicitly that the time intervals between storm impacts may significantly decrease in the future. This study highlights a new type of TC hazard resulting from temporally compounding landfalls while previous studies focus on single-storm hazards such as surge (Marsooli et al., 2019), rainfall (Emanuel, 2017), and compound flooding (Gori et al., 2020). Previous findings on intensified single-storm hazards calls for improvement of coastal reliability while this study on increased SLTs urges the improvement of coastal resilience. For example, the power system in Texas requires around 17 days to fully recover after hurricane landfall (Ouyang et al., 2012). For Texas, hurricane-MII being less than 17 days is a 40-year event in the current climate but would be a less-than-10-year event in the future (Figure 4i). To be prepared for potentially more frequent SLTs in the future, the recovery time of power grids and other infrastructure systems may need to be shortened along U.S. coastlines.
This study used a specific synthetic storm model to investigate SLTs in the future climate. This storm model predicts increased TC frequencies (Emanuel 2013(Emanuel , 2021 while large uncertainties still exist on how TC frequency will change in the future (Knutson et al., 2020). This study underlines again the importance of accurately projecting TC frequency, which greatly influences the estimation of SLTs and other TC hazards (Marsooli et al., 2019). It is important for future studies to use other methods, including regional (Wright et al., 2015) and global climate models that can explicitly simulate intense TCs (e.g., Bhatia et al., 2018), to validate the change of SLTs in the future found in this study.