Introduction

Extreme precipitation from Tropical Cyclones (TC) can cause significant inland flooding in both coastal and interior areas in the continental U.S. (CONUS) and lead to substantial economic damage. For example, Hurricane Harvey (2017) caused $136.7 billion in losses, mostly from rain-induced inland flooding1. It ranks as the second most expensive disaster after Hurricane Katrina. Similar flooding disasters also occurred in Hurricane Irene (2011), Hurricane Florence (2018), and Hurricane Ida (2021). Tropical Cyclone Precipitation (TCP) will likely increase in the future because a warming climate will increase TC rain rate and reduce TC translation speed2,3,4, causing more frequent stalling5. These changes in TCP patterns will elevate inland flooding risk6. In addition, local factors such as the urban heat island effect7 and aerosols8 can further exacerbate risk by intensifying TCP.

While the physical risk of TCP is changing, there have also been substantial shifts in regional demographics and economic development. These changes are likely to continue in the future9. Therefore, inflation-adjusted GDP is an essential criterion to quantify historical10,11,12,13 and future trends14 in TC damages. Two studies10,15 have concluded that there is no trend in destruction because of the lack of trends in landfalling TC frequency and intensity13 in CONUS. In contrast, Klotzbach et al.11 used a more extended period of record and found positive trends in hurricane-related economic losses in CONUS. In addition, there are also observed increases in socio-economic exposure globally and regionally16 from TC wind damages11,14,17. While many discussions have focused on damages caused from TC winds, few10,18 have focused on impacts from TCP and inland flooding. In a changing climate, local mitigation, adaptation, and risk management are crucial for reducing the effects of TCs19. Here we use a long-term, high-resolution TCP record (at 0.25°) derived from daily in-situ precipitation observations20, high-resolution population data (highest resolution of 1 km)21 and Gross Domestic Product (GDP, 1 km)22 to quantify spatial and temporal changes in physical TCP risk and socio-economic exposure to TCP in CONUS over 70 years (1949 to 2018). We summarize changes at the local, state, and national levels to inform mitigation and adaptation strategies.

Results

Temporal trends in TCP

About 3.7 million km2 (30%) of CONUS has consistent TCP records from 1949 to 2018 (Fig. 1a) and 52,000 km2 (0.4% of CONUS) has a statistically significant positive trend in annual TCP. There is an even larger area (144,000 km2, 1.2% of CONUS) that has a statistically significant positive trend in annual maximum event TCP (Fig. 1b). There is also small area with negative trend (~28,000 km2 for annual TCP and 1540 km2 for annual maximum event TCP), although not statistically significant. Most locations with increasing annual TCP trends are located in the mid-Atlantic and south-eastern Atlantic coast (Fig. 1a). In comparison, extreme TCP (Fig. 1b) has a more expansive spatial pattern with significant positive trends in the northeastern Atlantic coast, Gulf coast, and some inland locations. There are different spatial patterns for increases in daily rain rate (Supplementary Fig. 2), annual TCP days (Supplementary Fig. 3) and durations of TCP events (Supplementary Fig. 4) and they all contribute to increases in total and extreme TC in Fig. 1. Previous studies have found either positive trends in TCP23,24,25,26 or no trend in TCP27,28 over CONUS. Differences in the length of the record, the definition of the TCP (radius to the TC center), or the source of rainfall data are possible causes of the lack of agreement in past studies. Our novel dataset shows that there is substantial spatial variability in annual total and extreme TCP trends. The mainly positive trends are caused by increased TC rain rate and event duration.

Fig. 1: Trends in tropical cyclone precipitation from 1949 to 2018 in CONUS.
figure 1

The trend at each location is calculated using Sen’s slope (mm year-1) for: a annual total Tropical Cyclone Precipitation (TCP), and b annual maximum event TCP from 1949 to 2018 in CONUS. Black dot indicates that the trend at that location is statistically significant based on the Mann–Kendal test at the 95% significance level.

Full size image

Changes in TCP Probability

We have chosen 100 and 200 mm as the thresholds to quantify changes in the Return Periods (RP) of extreme TCP events across the CONUS because 100 mm is a general precipitation threshold for generating flooding29 and 200 mm indicates an extreme precipitation event30,31. The RPs are calculated for two overlapping 50-year time windows (Fig. 2) to evaluate changes between the early period (1949 to 1998) and the late period (1969–2018). The overlapping period of record was selected so that we are using a robust and conservative estimate of changes in return periods for extreme events. Please see the “Methods” and Supplementary Fig. 10 for results of the sensitivity analysis using non-overlapping periods. Only 529,760 km2 of areas in CONUS have <25 years RP for >100 mm event TCP in the early period (Fig. 2a), and it increases to 904,750 km2 in the late period (Fig. 2b) with a relative increase of 70%. The >200 mm TCP (Fig. 2c, d) events are rarer but more destructive. We observe a ten-fold increase in the areas with <25 years RP from 11,550 km2 in the early period to 117,810 km2 in the late period. Geographically, locations with a higher risk of extreme TCP are mostly in coastal areas during the early period, but these higher-risk regions extend much further inland during the later period (Fig. 2b, d).

Fig. 2: Comparison of Return Period (RP, in years) for 100 and 200 mm Tropical Cyclone Precipitation (TCP) events between the early period (1949–1998) and the late period (1960 to 2018).
figure 2

Each panel is a RP for 100 mm TCP events in the early period, b RP for 100 mm TCP events in the late period, c RP for 200 mm TCP events in the early period and, d RP for 100 mm TCP events in the late period. Only locations with robust estimation of the empirical exceedance probability density curve are displayed.

Full size image

Cities have concentrations of people and economic activities, so they usually have greater exposure to extreme weather hazards32. Rapid urbanization and elevated risk of severe weather jointly exacerbate the damage risk33,34. Therefore, we chose nine large cities (locations shown as Supplementary Fig. 5) along the coast for analysis, and many have increased TCP risk (Fig. 3). Houston has the most significant increase, with the 20-year event TCP increased from <200 mm in the early period to >400 mm in the late period. The tail of the distribution during the late period increases to > 600 mm as a result of extreme events like Tropical Storm Alison (2001), Hurricanes Ike (2008), and Harvey (2017) in the late period. Other large cities, including New Orleans, Mobile, Tampa, Jacksonville, and Raleigh, have also experienced substantial increases in TCP risk (Fig. 3b–f). Cities further to the north generally have lower TCP risk and demonstrate different patterns of change in TCP risk. For example, TCP risk increased by 50% in New York City and decreased slightly in Boston, while it did not change in Washington, D.C. This is partly a function of the uncertainties in determining TCP risk in regions that experience fewer TCs.

Fig. 3: Return Period (RP) for event Tropical Cyclone Precipitation (TCP) calculated for U.S. cities by the empirical kernel density function of exceedance probability.
figure 3

Each panel represents: a Houston, TX. b New Orleans, LA. c Mobile, AL. d Tampa, FL. e Jacksonville, FL. f Raleigh, NC. g District of Columbia, D.C. h New York City, N.Y.C, NY. i Boston, MA. Confidence Intervals (CI) are estimated by the upper (95%) and lower (5%) boundaries of the Poisson distribution.

Full size image

TC rain rate generally follows the Clausius–Clapeyron equation, and many previous studies have demonstrated that higher global mean atmospheric temperature will lead to more precipitation2,35,36. Our analysis shows substantial increases in TC rain rates over CONUS during the last 70 years (Supplementary Figs. 2, 6) and increased probabilities of higher rain rates in many major cities (Supplementary Fig. 7). At the same time, the mean translation speed of TCs decreased from 10.55 kt in the early period to 10.37 kt (Supplementary Fig. 8) in the late period (~1.7%). This reduction is most pronounced in the distribution’s 0 to 20 kt range. We also find high spatial variability in changes in TC translation speeds, and clustered grids with slowing downs are in the southern part of CONUS (Supplementary Fig. 9). There is still debate on the global trend in TC translation speed37,38,39. Nevertheless, our records indicate that the recent increase in TCP risk in CONUS is associated with a substantial rise in TC rain rate and a slight decrease in regional TC translation speed over land.

Changes in socio-economic exposure to TCP

In CONUS, substantial demographic and economic changes have occurred in the last 70 years. For example, the southern “Sunbelt” population has increased since 1950, driven by economic growth, demand for amenities, and housing supply40. Meanwhile, these locations are prone to TC-related disasters because of their proximity to the ocean. We created a TCP socio-economic exposure index (TCPEI, see data and methods for details) by combining demographic and economic data with physical TCP risk. The TCPEI quantifies how socio-economic exposure to extreme TCP events has changed between the early period (1949–1998) and the late period (1969–2018). There is a general increase in TCPEI from the early period (Fig. 4a, c) to the late period (Fig. 4b, d). About 55 million people were exposed to > 100 mm TCP in the early period, and it increased to 107 million people (92% relative increase) in the late period (Fig. 4c, d). The population with > 1 TCPEI (the highest category for TCP > 200 mm, Fig. 4g) has increased from zero to 4 million. Meanwhile, the population with > 0 TCPEI (lowest category for TCP > 200 mm) has risen ~6 folds from 4.9 million to 29 million. This pattern indicates that the densely populated areas are experiencing more relative increases in exposure to more extreme TCP events.

Fig. 4: Spatial Changes in the population-based Tropical Cyclone Precipitation Exposure Index (TCPEI), comparison of population and area for different ranges of TCPEI, relative changes in TCP probabilities, and relative changes in population and Gross Domestic Product (GDP).
figure 4

Each panel represents: a >100 mm daily TCPEI in the early period, b >100 mm daily TCPEI in the late period, c population comparison between the early period and the late period for different ranges of >100 mm daily TCPEI, d comparison of the impacted area between the early period and the late period for different ranges of >100 mm daily TCPEI, e >200 mm daily TCPEI in the early period, f >200 mm daily TCPEI in the late period, g population comparison between the early period and the late period for different ranges of >200 mm daily TCPEI, h the comparison of the impacted area between the early period and the late period for different ranges >200 mm daily TCPEI, i relative changes in >100 mm daily TCP probability, j relative changes in >200 mm daily TCP probability, k relative changes in the logarithm of the population, l relative changes in the logarithm of GDP. The 1970 Population and the 1949–1998 TCP are used to calculate the early period TCPEI. The 2020 population and 1969–2018 TCP are used to calculate the late period TCPEI. In panels c, d, g, h, “m” stands for million and “k” stands for thousands.

Full size image

We also observe significant changes in TCPEI regionally. By averaging the TCPEI within state boundaries, we demonstrate that most states have experienced increases in the median TCPEI, but the magnitude varies from state to state (Supplementary Fig. 11). Louisiana, Florida, and North Carolina have the largest absolute increases (0.49 to 0.70 in median) and relative increases (97 to 100%) in population exposure for > 100 mm TCP (Supplementary Fig. 11a, b). Louisiana, Florida, and Texas have the largest increases in population exposure to >200 mm TCP (0.19 to 0.29 increases in the median with 39 to 58% relative increases, Supplementary Fig. 11c, d). This pattern indicates that states with larger populations are experiencing significantly elevated exposures to extreme TCP events. Florida, Texas, and North Carolina rank high in the relative population growth from 1990 to 2020 (Supplementary Fig. 12). Louisiana has slower population growth, so its significant increase in population exposure is primarily due to elevated physical TCP risk.

Economic activities are closely associated with population changes and can be substantially disturbed by extreme weather events like TCs. Similar to the population TCPEI, we have found substantial increases in the exposure of economic activities to extreme TCP events (Supplementary Fig. 13). There are expansions of areas with high TCPEI in the late period (Supplementary Fig. 13b, f) compared with the early period (Supplementary Fig. 13a, e). The > 200 mm TCPEI has more significant relative increases than the >100 mm TCPEI. For example, the GDP exposed to > 100 mm TCP increased by 2%, from $5.2 trillion to $5.3 trillion (Supplementary Fig. 13c). Meanwhile, the GDP exposed to >200 mm TCP increased by 124%, from $581 billion to $1.3 trillion (Supplementary Fig. 13g). Most states had increases in median GDP TCPEIs (Supplementary Fig. 14), and the southern states had more pronounced relative increases. For example, Florida, Mississippi, Texas, Alabama, and Louisiana rank as the top 5 states highest relative increase (32–131%) in the 200 mm TCPEI (Supplementary Fig. 14c, d).

The TCPEI combines physical TCP risk and socioeconomic information. The TCP risk has larger relative increase (> 60% at many locations, Fig. 4i, j) than the population (<20% at many locations in the sourthern U.S., Fig. 4k) and GDP (<10% at most locations, Fig. 4l). Therefore, the increase of physical TCP is the major driver of increasing TCPEI over the CONUS. On the other hand, the changes in population/GDP intensified (attenuated) the regional TCPEI changes (e.g., in the southern U.S., Fig. 4a, b), with major influences from the changing physical TCP risk.

Discussion

Large areas in the southern U.S. and coastal mid-Atlantic have significant positive trends in annual total TCP and maximum event TCP during the last 70 years. Widespread increases in TCP probability occur in coastal regions and they are most pronounced in the southern U.S. Elevated TCP risk is also observed in most coastal cities. Those changes in physical TCP risk are associated with increased TCP rain rates and decreased TC translation speed6,41. Our records also show that the most significant decreases in TC translation speed are in the slowest moving TCs (<20 knots, Supplementary Fig. 8a) and in TCs tracking towards the northwest (Supplementary Fig. 8b). Large spatial variabilities also exist in the temporal trend in TCP. These variabilities are a result of multiple factors, including anthropogenic climate change3,41,42, natural variability (e.g., El Niño-Southern Oscillation and Atlantic Multidecadal Variability) and forcing from volcanic activity28,43,44, and inconsistencies and limitations of historical observations20. Therefore, we need more high-quality and long-term historical TCP data and carefully controlled numerical experiments for a better understanding of changes in TCP and its relationship with climate/environment.

Many previous studies showed increasing risk of extreme weather events under anthropogenic climate change3,45,46. Some provide quantitative estimates of social exposure to extreme heat and cold events, floods, sea level rise, etc.47,48,49,50. Using the TCP data, we proposed a new index that combines changes in the physical probability of extreme events and socioeconomic measurements directly affected by the disaster. We demonstrated that while the physical TCP risk is the primary driver determining the socioeconomic exposures, increases in population and GDP are exacerbating these changes. Areas with high TCPEI have expanded substantially due to rapid urbanization and population growth in the southern and mid-Atlantic states of CONUS during the last 70 years (Fig. 4k, l and Supplementary Fig. 15). Future regional population growth51 and urbanization52 may intensify TCP flooding6,33 and its societal exposure. In addition, the spatial heterogeneity in the population growth and economic development cause differences in local exposure to TCP hazards. Potential damage may increase exponentially due to changes in the magnitude and frequency of extreme TC events53,54. Compound events such as storm surge and pluvial floods can intensify those damages55. At the local level, government and citizens need more detailed mitigation strategies such as updated building codes, expanded insurance, and infrastructure hardening11 when faced with increased exposure to extreme TCP. Our research quantifies how physical risk from TCP and societal exposure changed simultaneously across CONUS during the last 70 years. Geiger et al.14 predict an annual global exposure increase of 26% (33 million people) to TC winds for every 1 °C rise in global mean surface temperature. We demonstrated that the number of people exposed to >200 mm TCP has increased by 24 million (489% relative increase, Fig. 4g) and the GDP exposure to >200 mm TCP has increased by $719 billion (124% relative increase, Supplementary Fig. 10g). Future research needs to address how compound flood56 exposure (extreme floods generated by both rainfall and storm surge) changes by integrating and coupling earth system models, hydrological/hydraulic models, and socio-economic/land cover change models. Substantial disparities exist in social exposure57,58 to TC hazards, and our TCPEI does not explicitly address this. Future metrics of social exposure to extreme weather events also need to reflect social disparity so the most vulnerable social groups to nature disasters can be better protected.

Data and methods

Precipitation data are from the Daily Global Historical Climatology Network (GHCN-D), which provides a complete global daily precipitation record. They are freely accessible from National Centers for Environmental Information, NCEI (https://www.ncei.noaa.gov/products/land-based-station/global-historical-climatology-network-daily). We obtained daily precipitation observations for TCP calculation from GHCN-D gauges that are within a moving daily boundary. Four connected circles define this daily boundary with 600 km radius to four TC centers recorded at 6 h intervals from the International Best Track Archive for Climate Stewardship (IBTrACS). Here a larger radius (600 km) is used instead of 500 km to extract rain gauges because we want to avoid missing the rain gauge observations ouside the boundary when spatial interpolation is processed near the 500 km boundary. The IBTrACS data are also freely available from NCEI (https://www.ncdc.noaa.gov/ibtracs/). After a wind under-catch bias correction20, we use Inverse Distance Weighting (IDW) interpolation59 to estimate the precipitation at a set of grid points with even spacing of 0.25° across the CONUS:

$${P}_{g}=\left[\left.\mathop{\sum }\limits_{i=1}^{N}{\left({d}_{i}\right)}^{-u}{z}_{i}\right]\left/\left[\mathop{\sum }\limits_{i=1}^{N}{\left({d}_{i}\right)}^{-u}\right. \right.\right]$$
(1)

Pg is the precipitation estimation at each grid point. N is the number of gauges within a searching distance (R = 30 km) for each grid point. di is the distance from each gauge to a specific grid within R. u is the weighting power and is set to 2 as the default value. After we calculate Pg for all daily TCP, only grids within 500 km of the storm track finally enter our daily TCP database. We choose the 500 km boundary because it provides consistent criteria for our 70-year record to define TCP. It also includes most precipitation systems in TC41,60, including both convective precipitation in the core and stratiform precipitation in outer bands. This routine has been validated and optimized by comparing available TMPA observations20 and is available upon request. To guarantee the consistency of the data, we only chose grids with at least one consistent gauge with 70 years of complete observations within R (30 km). We have chosen 4816 grids (Supplementary Fig. 1) based on that criterion, and most have more than two gauges within 30 km, which guarantees there are no missing observations in the time series. Therefore, we believe that the final grids (4816) provide good spatial coverage of the TCP’s affected areas across the CONUS.

The population for 1970 is from The Global Population Density Grid Time Series v1 and the 2020 population is from the Gridded Population of the World (GPW) v4. They are compiled by the Socioeconomic Data and Application Center (SEDAC) of Columbia University both have 1 km spatial resolution. GDP data at 1 km resolution are obtained for 1990 and 2015 from the Gridded global datasets for Gross Domestic Product and Human Development Index from 1990 to 201522. We re-scaled the population and GDP data for each 0.25° grid by summing all 1 km grid values within each TCP grid box.

We then aggregate daily TCP observations at each grid into event TCP based on the duration of each TC (from the start to the end of each TC event over CONUS) and sum total TCP in each year as the annual total TCP. A 70-year climatology is constructed for annual total TCP, annual maximum event TCP, and annual maximum daily TCP at each grid. The Mann–Kendall non-parametric test61,62 is used to detect trends in those time series and the Sen’s slope63 is calculated for each time series. We use a 95% significance level for both non-parametric methods for all trend detections. We also calculate return periods for daily and event TCP samples at each grid by constructing empirical exceedance probability density64. For any sample of TCP observations (event or daily) at a specific grid, the non-parametric kernel density estimation (KDE) function is defined as:

$${\hat{f}}_{h}\left(p\right)=\frac{1}{{nh}}\mathop{\sum }\limits_{i=1}^{n}K\left(\frac{p-{p}_{i}}{h}\right)$$
(2)

where p1, p2, … pn are observations of TCP at each grid with an unknown distribution; n is the sample size. h is the bandwidth set as 0.02 to provide enough detail for each KDE curve. K is the kernel smoothing function, which we used a Gaussian kernel smoother, expressed as:

$$K\left({p}^{* },{p}_{i}\right)=\exp \left(-\tfrac{{\left({p}^{* }-{p}_{i}\right)}^{2}}{{2\sigma }^{2}}\right)$$
(3)

where σ is the length scale for the input space. After a KDE curve is fitted, we can estimate the cumulative distribution function (CDF) and the Exceedance Probability Function (EPF) (1 − CDF). By using a linear interpolation approach65, we can find the Return Period (RP) estimate for each TCP threshold (e.g, > 100 mm). We estimated the RPs of daily and event TCP thresholds for all 4816 grids and compared the estimates from the early period (1949–1998) to the late period (1969–2018). We did not make extrapolations, so our most extended return period is 50 years. We removed all locations where we could not determine a value from this empirical EPF calculation. We also assume both the event and daily TCP at each location follow a Poisson Distribution6,66 and use it to calculate Confidence Intervals (CI) by the upper 95% and lower 5% boundary for each kernel. Finally, we compare the spatial distribution of RP for different magnitudes of TCP between the early and late periods and demonstrate how the whole TCP probability profile changes at night cities in the CONUS. We also finished a sensitivity analysis that compares two non-overlapping periods of 35 years (Supplementary Fig. 10). Results show that there is little change in the pattern of TCP RP change as compared to the overlapping test in Fig. 3. The long tails of the distribution for extreme values have systematically shifted to a shorter RP. Therefore, the RP estimated with a longer period of record (Fig. 3) will have less uncertainty in the estimates of extreme values.

To quantify society’s exposure to extreme TCP events, we created a TCP exposure index (TCPEI):

$${{{{{\rm{TCPEI}}}}}}=\frac{{{{{{{\rm{log }}}}}}}_{10}{{{{{\rm{SEC}}}}}}}{{{{{{\rm{RP}}}}}}}$$
(4)

Where the SEC stands for Social Economic Criteria, in this study, we utilized two datasets: population and annual GDP in U.S. Dollars ($). We use the logarithm of 10 for the SEC to re-scale the population and GDP for spatial analysis. The RP is the Return Period for a specific magnitude of TCP (e.g., 100 mm). The TCPEI increases as population or GDP increases or RP becomes shorter (higher physical TCP risk). Since there can be spatial and temporal changes in both physical TCP risk and population/socio-economic activities, our TCPEI can quantify the spatial variability of socio-economic exposure and identify hot spots vulnerable to TC floods. Supplementary Fig. 16 shows a flowchart that summarizes all data processing and analysis used for this study.