Abstract
Tropical cyclones (TCs) are among the most devastating natural phenomena, causing substantial economic damage and severe impacts on human life and infrastructure. Prolonged extreme ocean temperature events, known as marine heat waves (MHWs), affect 52% of landfalling TCs globally and provide favorable conditions for TC rapid intensification (RI). Here, we use four decades of global data to demonstrate that TCs experiencing RI during MHWs resulted in 60% more billion-dollar disasters compared to those without heat wave influence. Our analysis reveals these costlier storms produced consistently higher maximum wind speeds, storm surges, and precipitation rates. We found that, for billion-dollar TCs, even controlling for similar levels of coastal development, MHW TCs cause significantly higher damages than their counterparts, confirming that their enhanced intensity characteristics amplify economic impacts beyond what would be expected from exposure alone. These findings underscore the critical need for improved disaster preparedness as MHWs increasingly supercharge TCs’ intensity, leading to more frequent billion-dollar disasters.
Marine heat waves supercharge tropical cyclones, leading to unprecedented strengthening and greater economic damages worldwide.
INTRODUCTION
Human activities have led to unprecedented global warming, making tropical cyclones (TCs) stronger over the past four decades (1, 2). Many of these intense storms undergo rapid intensification (RI), where wind speed increases by at least 30 knots within 24 hours, before making landfall (3, 4). This phenomenon is particularly problematic, as it leaves little time for evacuation and preparation for the event, leading to greater damage and loss of life. For example, in October 2023, Hurricane Otis rapidly intensified from a tropical storm to a Category 5 hurricane within a single day and made landfall near Acapulco, Mexico. Otis was the strongest storm ever to make landfall in the eastern Pacific with catastrophic 143 knots winds (5). The hurricane caused about $16 billion worth of damage and at least 52 deaths (5). Such abrupt changes in TC wind speed make it difficult, if not impossible, for models to reliably predict these events. Otis, for example, was one of the most poorly predicted RI events in recent decades, despite all the improvements of RI forecast tools (6). Hurricane Ian (2022) intensified to a Category 4 hurricane and made landfall on the Gulf coast of Florida with winds of 135 knots (7). Similarly, Category 5 super typhoon Goni (2020), which had winds of 170 knots at landfall, the strongest ever at landfall in recorded history (8), underwent RI while passing through the Philippine Sea. A common factor for all these cyclones is that they traversed a patch of ocean that exceeded 30°C along their trajectories before making landfall.
These observations bring us to a critical question: Do TCs traversing exceptionally warm ocean waters associated with marine heat waves (MHWs) preferentially undergo RI and become stronger, leading to more severe impacts upon landfall?
There is a pressing need to understand this phenomenon further as the proportion of TCs undergoing RI has shown an upward trend since the 1980s (9–12). Moreover, RI is likely to become more frequent in the future according to projections (13–17). Global climate models also suggest that anthropogenic warming has increased TC precipitation rates due to enhanced atmospheric moisture (18–20). However, the relationships between MHWs, TC RI, and precipitation patterns require further clarification. When TCs traverse regions experiencing MHWs, the extreme ocean temperature anomalies can significantly alter air-sea interactions. MHWs create a stronger moisture disequilibrium between the ocean surface and atmosphere, resulting in latent heat flux that is up to three times greater than normal conditions (18). This enhanced latent heat flux fuels more vigorous convection near the TC center, which manifests as heavier precipitation. Furthermore, TCs undergoing RI demonstrate distinctly different precipitation characteristics compared to those in slow intensification. RI TCs uptake approximately three times more moisture (up to 2500 km from the cyclone center) than non-RI TCs (21). This enhanced moisture uptake contributes to higher extreme rainfall totals over land (22).
To address the abovementioned research question, this study makes several key contributions to the understanding of TC behavior and impact. We provide the first global statistical analysis investigating the combined effects of MHWs and RI on landfalling TCs and their economic impacts. Through this analysis, we quantify how the synergistic interaction between MHWs and RI potentially influences TC destructive characteristics, including wind speeds, precipitation rates, and storm surges. This comprehensive examination of hazard characteristics explores how TC intensity characteristics amplify economic impacts when controlling for similar levels of coastal development. We hypothesize that TCs experiencing RI during MHWs will exhibit enhanced destructive characteristics compared to those without MHW influence and that these enhanced characteristics, rather than simply their landfall locations, are primarily responsible for their disproportionate contribution to economic damages. After establishing these relationships, we address how this understanding can lead to better coastal hazard preparedness by identifying the key drivers of compounding MHW and RI events. This analysis could, in principle, inform early-warning systems about which variables to monitor for improved RI forecasting during MHWs, presenting a causal chain from fundamental understanding to actionable insights for disaster risk reduction.
RESULTS
Global MHW characteristics and their implications for RI
An environment conducive to TC formation includes warm ocean waters of at least 26.5°C (23), the presence of the Coriolis force (requiring a minimum distance of at least 500 km from the equator), cyclonic low-level circulation, low vertical wind shear (less than about 37 km/hour), abundant humidity near the mid-level of the troposphere (about 5000 m), and instability between the surface and the mid-level (24, 25). Hence, not all ocean waters are conducive to the spin-up of TCs. Consequently, active TC basins are geographically constrained to regions where these conditions are met. These active TC basins are shown in Fig. 1E.
Fig. 1. Global MHW characteristics.
Forty-one-year (September 1981 to October 2023) average properties of MHWs based on high-resolution (0.25° × 0.25°) gridded daily SST data obtained from the National Oceanic and Atmospheric Administration (NOAA) OI SST v2.1 dataset. (A) Mean SST anomaly (°C), (B) mean annual frequency (events per year), (C) mean annual duration (days per year), and (D) mean annual maximum intensity relative to 90th percentile (degrees Celsius per year). Note: This analysis covers only the MHWs occurring during TC activity seasons, i.e., May through November in the Northern Hemisphere and November through April in the Southern Hemisphere. (E) The seven active TC basins.
Building on these foundational conditions for TC formation, MHWs—defined as at least 5 days of sea surface temperatures (SSTs) above the 90th percentile of the linearly detrended 1991–2020 climatological mean—may provide more favorable conditions for tropical cyclogenesis (26–29). MHWs are associated with elevated SSTs that can extend through the mixed layer (30, 31) and sometimes deeper, potentially surpassing the critical SST threshold of 26.5°C necessary for TC genesis (23). These elevated temperatures may be accompanied by processes that drive increased heat flux into the ocean, such as enhanced insolation or reduced latent heat loss, which can strengthen upper ocean stratification (32–34). This increased stratification reduces the depth of storm-induced vertical mixing (35–37). Therefore, the upper ocean maintains a higher heat content, with warm water anomalies concentrated closer to the surface (18). The increased thermal energy availability in the upper ocean supports and potentially intensifies cyclonic activity. Such conditions not only facilitate the initial formation of TCs but also create an environment highly conducive to their RI.
Understanding the global patterns of MHWs and their potential to influence TC behavior requires examining their fundamental characteristics across different ocean basins. Using 41 years of high-resolution (0.25° × 0.25°) gridded daily SST data from the National Oceanic and Atmospheric Administration (NOAA) OI SST v2.1 dataset, we analyze MHW frequency, duration, and intensity patterns in relation to historical TC activity (Fig. 1). The mean detrended SST anomaly of MHW events shows positive anomalies across all active TC basins, ranging from weak anomalies (≈1° to 2°C above the detrended baseline) to moderate anomalies (≈2° to 4°C). The strongest positive detrended anomalies (>4°C, locally >7°C) appear along the mid-latitude western boundary current extensions and other extratropical regions (Fig. 1A). In regions where SST often exceeds 26.5°C during the warm season (i.e., the tropical sectors of the active basins; see fig. S1), MHW anomalies of ≈1° to 3°C further raise the likelihood of RI by combining high SSTs with abundant atmospheric moisture, thereby creating favorable conditions for TC intensification (18). For instance, in the Gulf of Mexico and the northwestern Caribbean Sea, RI is on average 50% more likely during MHWs (29).
The mean MHW frequency in the active TC basins is typically between ~1.0 and 1.5 events per year (Fig. 1B). Localized hot spots exceed two events per year, most prominently in the Bay of Bengal and Arabian Sea of the North Indian Ocean, along the Philippine and South China Sea sectors of the Northwest Pacific, and in parts of the Australian basin (especially the Timor and Arafura seas). Extended MHW durations further amplify the potential for TC RI: within the active basins, average durations cluster around 10 to 20 days per year, with pockets surpassing 25 days per year (Fig. 1C), providing a prolonged heat reservoir that can sustain cyclogenesis and storm strengthening.
The largest MHW intensities appear along the subtropical western boundary current regions (e.g., the Kuroshio or Gulf Stream) (Fig. 1D); however, these higher latitude hot spots rarely contribute to landfalling TCs that undergo RI closer to the tropics, where MHW anomalies coincide with already high SSTs (38, 39). Consequently, the subtropical boundary currents typically mark transition zones for extratropical cyclone phases rather than providing prolonged, low-shear conditions essential for RI (40, 41). In active TC basins, areas that persistently register high relative MHW intensity values (Fig. 1A) represent an abnormal condition in which SSTs conducive to TC RI are not only achieved but exceeded. Across these core TC basins, mean i_max_rel (representing the maximum intensity above the SST threshold) is more modest, generally ~0.3° to 0.8°C above the threshold (Fig. 1D). This suggests that, in these basins, even moderate increases in mean MHW intensities can further enhance already favorable SST conditions conducive to TC RI. MHWs during warm seasons represent extreme ocean temperatures, in which the baseline SST exceeds the minimum threshold required for TC formation; therefore, even small increases above these already high temperatures can boost the potential for TC RI.
To further illustrate how MHWs affect TC intensification beyond simply raising SSTs, we examined the potential intensity (PI) of TCs; a theoretical upper limit of TC intensity based on thermodynamic conditions (42). Over half (51.7%) of globally landfalling TCs between September 1981 and October 2023 occurred during MHW conditions. PI analysis globally during this period shows that, on average, MHW TCs maintain higher PI values than non-MHW TCs during the prelandfall period (Fig. 2A), indicating a more favorable thermodynamic environment for intensification. The composites aligned to landfall show that before landfall (−5 to −1 days), MHW storms have mean PI values about 3 to 5.5 m/s (~4.5 to 8%) higher than non-MHW storms, with the largest difference (~5.5 m/s or 8%) occurring around a day before landfall (Fig. 2A). The daily box plots (Fig. 2, B and C) confirm these temporal patterns, with median values demonstrating the sustained higher PI of MHW TCs prelandfall. Additionally, individual TC PI analysis for Cyclones Idai and Kenneth that made landfall in Mozambique within ~1.5 months in 2019 (fig. S2) shows that Idai, which moved through waters without MHWs and experienced two RI events, showed variable PI values that dropped to ~50 to 55 m/s midway through its life cycle before increasing again, while Kenneth, which passed over MHW regions, maintained steady high PI values around 80 m/s throughout its life cycle while experiencing three RI events. This is because PI is closely tied to SST through an exponential relationship, where even small increases in already-warm waters (as in MHWs) can lead to disproportionately large gains in maximum PI (43). This thermodynamic framework explains our observational findings that MHWs can provide a more conducive environment for maintaining TC strength.
Fig. 2. PI evolution for MHW and non-MHW TCs (827 versus 773 storms) globally during September 1981 to October 2023.
(A) Time series of mean PI from 5 days before to landfall. (B and C) Daily box plots of PI for MHW and non-MHW TCs, respectively, showing the distribution of values for each indicated day relative to landfall. PI data are extracted from the Commonwealth Scientific and Industrial Research Organization global gridded ERA5-derived PI dataset (104). d, days.
To investigate how these MHW patterns relate to TC behavior over time, we analyze the historical record of global landfalling TCs using the International Best Track Archive for Climate Stewardship (IBTrACS) dataset (Fig. 3). An examination of global TC trends (Fig. 3A) indicates a statistically significant increase (P < 0.05) in non-MHW TCs, whereas the MHW TCs rise more modestly and the trend remains statistically insignificant. This pattern reflects the use of detrended SST in the MHW definition: As the baseline SST rises over time, only short-term, anomalous excursions above this moving reference are counted as MHWs, thereby reducing apparent trends in MHW frequency despite ongoing ocean warming. This methodological effect has been noted in recent studies, which show that applying a detrended or moving baseline substantially dampens or even eliminates upward trends in MHW frequency compared to using a fixed climatological baseline (44, 45).
Fig. 3. Historical analysis of TCs based on the occurrence of RI and the presence of MHWs from 1981 to 2023.
This figure depicts the temporal evolution of TC frequencies over a period of 41 years worldwide and for the seven active TC basins: (A) Global, (B) East Pacific, (C) North Atlantic, (D) Northwest Pacific, (E) North Indian, (F) Southwest Indian, (G) Australian, and (H) Southwest Pacific. Each subplot contains the total number of TCs in each category. The dashed lines show various trends in the events. The red line represents TCs that underwent RI and were influenced by MHWs, the green line represents MHW TCs without RI, the pink line represents non-MHW TCs that underwent RI, and the blue line represents non-MHW TCs without RI. The shaded area indicates a 95% confidence interval. Double asterisk indicates that the trend is significant (P ≤ 0.05).
The strongest acceleration in rapidly intensified MHW TCs is instead confined to individual Northern Hemisphere basins, most notably the East Pacific and North Atlantic (Fig. 3, B and C). Upward trends in MHW-related TCs (red and pink lines) are clear in the East Pacific, North Atlantic, and North Indian basins, while the Southwest Indian and Southwest Pacific basins show little or insignificant change. Conversely, non-MHW TCs (green and blue lines) now either hold steady or increase in most basins. The Southwest Pacific basin (Fig. 3H) stands out for its low TC activity, with only three events occurring over this period and no clear trends discernible. Only in the North Indian basin (Fig. 3E), the non-MHW TCs exhibit a significant upward trend, coinciding with a marked increase in overall TC activity since 2005. This unique pattern is tightly linked to the rising SSTs and increased moisture availability in both the Arabian Sea and the Bay of Bengal, a direct consequence of global warming (46). The Northern Hemisphere basins, in general, demonstrate an upward shift in the aggregated number of TCs, particularly since 2005, that aligns with observations of enhanced TC PI colocated with increasing SSTs (12). This can have far-reaching implications, especially for coastal communities.
It should be noted that, while elevated SSTs due to MHWs are an important factor for TC formation and RI, they are not the only contributors. Other conditions, such as low vertical wind shear, high mid-tropospheric humidity, atmospheric instability, preexisting disturbances, high ocean heat content, and adequate Coriolis force, can also play important roles in creating environments favorable for TC formation and intensification. To further contextualize the basin-specific MHW versus non-MHW trends, we have examined all global TCs and their RI behavior (see fig. S3). Overall, the global total of RI TCs exhibits a slight but statistically significant increasing trend; however, the trends within most individual basins are muted or not statistically significant, pointing to compensating effects. This phenomenon likely reflects two competing factors: (i) a decline in favorable large-scale atmospheric conditions in many basins (47, 48), which reduces RI events; and (ii) more frequent localized SST extremes from MHWs that can help fuel intensification for TCs passing over them, provided that other necessary atmospheric conditions are met (15, 21). The net outcome is an overall increase in global RI TCs, reflecting how a more hostile background environment can be partially offset by intensifying MHW conditions.
Recent studies further demonstrate that exceptionally warm ocean states associated with MHWs can compensate for moderately unfavorable atmospheric conditions and still promote RI. For example, Le Hénaff et al. (49) documented that Hurricane Michael intensified rapidly over the Gulf of Mexico despite relatively strong vertical wind shear, due to the presence of a Loop Current eddy and anomalously high ocean heat content. Similarly, Rathore et al. (27) showed that Super Cyclone Amphan underwent RI even as vertical wind shear increased, primarily as a result of an underlying MHW. A recent basin-scale analysis (18), over the Northwest Pacific and North Atlantic basins, found that storms encountering MHWs intensify much more rapidly than those that do not, even after accounting for distributions of shear and humidity. Notably, their results indicate that large-scale vertical wind shear during MHW-associated TCs was higher compared to non-MHW cases. Despite this, storms over MHWs exhibited enhanced RI, a behavior attributed to increased moisture convergence and higher mid-level relative humidity resulting from greater ocean-atmosphere moisture disequilibrium and enhanced latent heat fluxes.
Prelandfall MHW and RI impacts on TC wind speed and precipitation evolution
To understand the influence of TC characteristics on the extent of damage caused by landfalling TCs, we examine the evolution of two TC hazard indicators (i.e., wind speed and precipitation) at different lead times pre- and postlandfall (i.e., 5 days before landfall to 2 days after). Wind speed and precipitation severity are good proxies for the economic losses associated with landfalling TCs in coastal regions (50, 51). Based on the occurrence of RI and the presence of close MHWs (see relevant section in Materials and Methods), we can consider four scenarios as shown in Figs. 4 and 5: RI with MHW (red lines), RI without MHW (purple lines), no RI with MHW (orange lines), and no RI without MHW (blue lines). TCs that undergo RI reach nearly double the peak wind speed of non-RI storms at landfall (~80 knots versus 40 knots; Fig. 4), and the divergence between the two groups is already apparent 5 days before landfall. The presence of a MHW provides an additional and sustained increase in wind speed exclusively for RI storms: RI TCs influenced by MHWs maintain wind speeds 5 to 10 knots higher than RI TCs without MHWs throughout most of the prelandfall period. This difference is statistically significant at all lead times (P < 0.05; fig. S4, top) and peaks at ≈10 knots (15 to 20%) ~3 to 5 days before landfall. The effect of MHWs on non-RI TCs appears to be minimal, indicating that MHWs primarily supercharge the wind speeds of more intense TCs. Overall, non-RI TCs demonstrate a more gradual increase in wind speed, peaking at around 40 knots at landfall.
Fig. 4. Evolution of TC wind speed at different lead times relative to landfall.
(A) Wind speed (knots). Wind speeds represent the maximum wind speed of the TC track from IBTrACS data at the time step of interest. (B to E) Wind speed box plots showing distribution at daily intervals for each category: (B) RI with MHW, (C) RI without MHW, (D) no RI with MHW, and (E) no RI without MHW. The red line represents TCs that underwent RI and were influenced by MHWs, the purple line represents MHW TCs without RI, the orange line represents non-MHW TCs that underwent RI, and the blue line represents non-MHW TCs without RI. The color shading shows a 95% confidence interval. d, days.
Fig. 5. Evolution of TC precipitation at different lead times relative to landfall.
(A) Mean total precipitation (millimeters per hour). Total precipitation is averaged for all grids within a 5° distance from ERA5 reanalysis with a spatial resolution of 0.25°. (B to E) Precipitation box plots showing distribution at daily intervals for each category: (B) RI with MHW, (C) RI without MHW, (D) no RI with MHW, and (E) no RI without MHW. The red line represents TCs that underwent RI and were influenced by MHWs, the purple line represents MHW TCs without RI, the orange line represents non-MHW TCs that underwent RI, and the blue line represents non-MHW TCs without RI. The color shading shows a 95% confidence interval. hr, hour; d, days.
After landfall, all TC categories show a rapid decrease in wind speed within the first day, with the strongest storms (i.e., RI TCs) experiencing the sharpest decline. RI TCs lose ~40% of their wind speeds 1 day after landfall, compared to about 15% for non-RI TCs (60% lower weakening rate). The rapid weakening of the RI TCs postlandfall is primarily due to the sudden loss of their primary energy source (i.e., warm ocean waters) and increased surface friction over land. Studies show that hurricanes undergoing RI tend to reach higher peak intensities, making them more prone to rapid decay after landfall, as stronger storms experience greater absolute weakening due to enhanced surface roughness and reduced heat flux (52). Additionally, landfalling TCs often encounter dry air intrusion and increased vertical wind shear, which disrupts the storm’s core and accelerates weakening (53). This pattern has been observed across multiple basins, including the North Atlantic, Eastern and Western Pacific, and the Indian Ocean, where intense storms exhibited rapid postlandfall decay (54, 55).
TC precipitation typically ramps up beginning 2 days before landfall, peaking a few hours before the TC center reaching land (Fig. 5A). The persistent difference between RI TCs with MHW influence and those without is evident from about 2.5 days before landfall to about half a day after landfall. The peak difference is observed about 3 to 6 hours before landfall. RI TCs influenced by MHWs show the largest rise with mean rates accelerating from ≈0.8 mm/hour on day −3 to ≈3.0 mm/hour just before landfall. When a nearby MHW is present, RI storms maintain a persistent precipitation premium of 0.2 to 0.3 mm/hour (≈10 to 12%) over their non-MHW counterparts from day −3 through day +1, and the difference is statistically significant within ±24 hours of landfall (P < 0.05; fig. S4, bottom). This sustained enhancement in precipitation for RI TCs with MHWs highlights the role of MHWs in providing extended periods of high PI (fig. S2), which support stronger convection, increased moisture convergence, and higher latent heat flux near the storm center (18).
RI TCs also exhibit a markedly steeper increase in precipitation rates. These TCs show an increase of 0.6 to 0.8 mm/day compared to the more modest 0.4 to 0.5 mm/day for non-RI TCs. This divergence is particularly pronounced in the period from 1 day before landfall to several hours after landfall, during which RI TCs consistently have higher precipitation rates than their non-RI counterparts. The number of offshore RI events within 400 km of the coastline has tripled between 1980 and 2020 (15). This trend suggests that landfalling TCs are increasingly likely to have undergone RI just before making landfall, explaining the systematic peak in rainfall rates during this time window. RI TCs typically experience enhanced moisture convergence and deep convective organization near landfall, which provide substantial latent heat energy to fuel intense convection (14, 56). These storms frequently develop a well-organized inner core, characterized by compact eyewalls and deep convective towers, which maximizes precipitation efficiency (22, 57, 58). Additionally, near-surface frictional convergence increases as the storm nears land, temporarily enhancing rainfall rates at landfall (59). However, this effect is transient. After landfall, increasing terrain interaction and disruption of the storm’s core structure weaken the convection, leading to a rapid decrease in precipitation rates (60). The observed precipitation pattern is further influenced by large-scale environmental conditions. Rapidly intensifying storms often occur in regions of low vertical wind shear and strong upper-level divergence, both of which sustain vigorous convection (61). The combination of these factors results in a short-lived peak in precipitation around landfall, with lower precipitation rates before RI onset and after the storm moves inland.
For non-RI TCs, the impact of MHWs on precipitation is much smaller (typically less than 5%) and not statistically significant. Although it may appear counterintuitive that non-RI storms exhibit slightly higher disk-mean rainfall than RI storms during the −5- to −2-day window, this pattern is consistent with global precipitation climatologies and the known structural evolution of TCs. Multiple satellite-based studies have shown that slowly intensifying and weaker storms tend to maintain broader stratiform rain shields and persistent outer rainbands, which enhance area-averaged rainfall at larger radii. As storm intensity increases, the radius of maximum rainfall contracts inward, resulting in more concentrated rainfall near the core and reduced rain coverage at larger distances (62). Spatial variability in TC rainfall is further shaped by storm structure and environmental shear, with weaker or slowly intensifying vortices retaining greater rainfall asymmetry and broader regions of moderate rain outside the core (63).
Long-term composites using Tropical Rainfall Measuring Mission Precipitation Radar reveal that, in the hours leading up to RI, the primary change is a sharp increase in inner-core rainfall coverage, especially stratiform, while the area-mean rain rate remains nearly constant until after RI has begun. In contrast, slowly intensifying storms retain a loosely organized stratiform shield that sustains elevated area-averaged rainfall over large disks, even as maximum winds remain modest. The contribution of stratiform rain is maximized during the build-up to RI, but mean rain intensity and volumetric totals only rise sharply once the core has fully contracted and intensification is underway (64). That said, our IBTrACS-based intensity analysis (table S1) reveals that 85 to 89% of non-RI storms (compared to ~50% for RI storms) are undergoing slow intensification, with sustained 24-hour wind increases of 10 to 30 knots/day. This prevalence of slowly intensifying TCs with their broader stratiform rain shields explains the slightly higher 5°-disk mean precipitation rates compared with the more compact, rapidly contracting RI storms.
Economic impacts of MHW and RI events on landfalling TCs
To assess how the combined effects of MHWs and RI translate into economic consequences, we examine relationships between TC intensity measures and resulting damage costs. Focusing exclusively on 789 landfalling TCs with documented economic impact data, we compare physical characteristics (wind speed, precipitation, and storm surge) against damage outcomes across different scenarios. By categorizing events into damage thresholds (<$1 billion and ≥$1 billion) and TC types (based on RI occurrence and MHW influence), we can determine whether higher economic impacts stem primarily from enhanced TC intensity or from coastal development patterns in affected regions. Comparing the four aforementioned cases, TCs with RI (red and purple in Fig. 6A) demonstrate noticeably higher wind speeds than those without RI (orange and blue in Fig. 6A) in every cost range above $1 million. The presence of MHWs (red and orange in Fig. 6A) appears to increase wind speeds compared to their non-MHW counterparts (purple and blue in Fig. 6A), with the largest separation (≈10 to 15 knots in the median) occurring in the $100 million to $1 billion and $1 billion to $10 billion bins. Consistent with Fig. 4, TCs with RI and MHWs exhibit the strongest winds in the three highest cost brackets, although RI storms without MHWs reach comparable speeds in the >$10 billion classes.
Fig. 6. Relationship between TC intensities and economic loss.
(A) Mean overland wind speeds (Ws) versus cost range. (B) Mean overland total precipitation rates (Tp) versus cost range. Box plots in (A) and (B) show distributions for different scenarios: RI with MHW, RI without MHW, no RI with MHW, and no RI without MHW. Notches in the box plots indicate the 95% confidence interval around the median; sharp wedges occur when this interval is larger than the interquartile range due to small sample sizes in some categories. (C to F) Joint distributions of mean overland wind speed and precipitation for different scenarios: RI with MHW, RI without MHW, no RI with MHW, and no RI without MHW. Background shading represents joint exceedance probabilities from 0.01 to 1.00. Dots indicate individual TCs, colored by damage costs. Text insets show the percentage of TCs in each damage category. The cost unit is 2023 US dollars. B, billion; M, million.
Median precipitation (Fig. 6B) generally rises with increasing economic loss, and storms that traverse MHW regions (red and orange) retain the advantage across most cost brackets. In the four sub-billion ranges (<$1 million, $1 million to $10 million, $10 million to $100 million, and $100 million to $1 billion), the MHW categories record the highest or co-highest medians, consistent with the finding from Fig. 5 that MHWs enhance TC rainfall, particularly up to 12 hours after landfall. Furthermore, in the $10 billion to $100 billion bracket, rapidly intensified MHW events (red) show a substantially higher median precipitation than any non-RI group, and their median closely matches the levels observed for these storms in the $1 billion to $10 billion range.
The joint relationship between wind speed and precipitation rate (Fig. 6, C to F) shows a similar pattern. In RI scenarios, ≥$1 billion events cluster toward higher wind speeds and moderate-to-high precipitation rates. This clustering is more pronounced when MHWs are present. Non-RI scenarios show a more dispersed distribution, with ≥$1 billion events less clearly separated from <$1 billion events in terms of wind and precipitation rates. TCs that experienced RI while traversing MHWs have the highest economic losses: 15 (of 71) TCs caused damages exceeding $10 billion ($1 billion). In contrast, TCs that experienced RI without the influence of MHWs have fewer points in the high-intensity zone, suggesting that, while RI alone can result in strong storms, the absence of MHWs mitigates the extreme wind and precipitation hazards. This results in reduced economic losses, with only 7 (of 45) TCs with RI and without MHW causing over $10 billion ($1 billion) in damage. The non-RI with MHW cases (Fig. 6E) also display a similar number of billion-dollar TCs to non-MHW TCs (≈11%). The 50% joint exceedance curves for marginal TC intensities in this figure (i.e., half the points lie within the light blue area and half outside) well represent the implications of the occurrence of RI and the presence of MHWs on the economic impacts of landfalling TCs. The frequency/severity of TC damage outside this light-blue–shaded region clearly shows how RI and MHWs influence not only the physical predictors of TCs but also their potential to wreak economic damage. Figures S5 to S7 illustrate similar joint relationships for other paired intensities [i.e., maximum surge versus wind speed (Ws) and total precipitation rate (Tp)].
Population density and infrastructure in affected areas are critical in translating physical impacts into economic damage. A high storm surge in a sparsely populated area may cause less economic damage than a smaller surge in a densely populated, economically important region. The increasing trend in built-up volume suggests growing infrastructure development in coastal regions, potentially increasing exposure to TC impacts over time.
To isolate the specific influence of MHWs and RI on damage costs while controlling for exposure differences, we used propensity score matching (PSM) (65) using built-up volume from the Global Human Settlement Layer (GHSL) dataset (66) as a proxy for coastal development and exposure (see Materials and Methods for more details on the PSM analysis). This statistical technique helps create comparable groups of TCs by matching those with similar built-up volumes but different characteristics (i.e., with/without RI and MHW), allowing a more reliable assessment of how TC intensity influences damage costs. Our analysis shows that, even after controlling for exposure, TCs with RI and MHW influence caused significantly higher damages compared to their matched counterparts (Table 1). For TCs with similar built-up volumes, RI TCs with MHW influence produced 93% higher damages than RI TCs without MHW influence (t test, P < 0.05) and 876% higher damages than non-RI TCs without MHW (Mann-Whitney, P < 0.05). This statistically significant pattern also holds consistently for all events (Table 1) and across different continental regions (fig. S8). This pattern confirms that, while high exposure is a necessary precondition for extreme damages, the enhanced intensity characteristics of TCs experiencing both RI and MHWs significantly amplify the resulting economic impacts beyond what would be expected from exposure alone.
Table 1. PSM analysis of TC damage costs.
This table presents the results of PSM analysis examining the effect of MHWs and RI on TC damage while controlling for built-up volume. The analysis compares the treatment group “RI with MHW” against different control groups. Mean damage values are in billions of 2023 US dollars. Statistically significant P values (P < 0.05) are indicated with asterisks. Note: Absolute standardized mean difference (|SMD|) values of 0.2 to 0.5 indicate small, 0.5 to 0.8 indicate medium, and >0.8 indicate large differences between two groups (105, 106). MW refers to Mann-Whitney U test, a nonparametric test that is more robust for nonnormally distributed data like damage costs. Asterisks indicate significance levels based on statistical tests: ** denotes cases where both MW and t test support significance at the 5% level; * denotes cases where one of the tests indicates significance.
| Comparison | Matched pairs | Treatment mean (B$) | Control mean (B$) | Difference (%) | |SMD| (before) | |SMD| (after) |
|---|---|---|---|---|---|---|
| All events | ||||||
| RI with MHW versus RI without MHW | 218 | 4.31 | 2.24 | 92.9* | 0.04 | 0.08 |
| RI with MHW versus no RI with MHW | 171 | 2.79 | 0.58 | 384.0** | 0.01 | 0.14 |
| RI with MHW versus no RI without MHW | 226 | 4.18 | 0.43 | 876.3** | 0.18 | 0.26 |
| Billion-dollar events | ||||||
| RI with MHW versus RI without MHW | 114 | 15.08 | 8.34 | 80.8* | 0.09 | 0.04 |
| RI with MHW versus no RI with MHW | 60 | 11.18 | 3.65 | 206.6* | 0.01 | 0.15 |
| RI with MHW versus no RI without MHW | 86 | 14.66 | 2.82 | 419.5** | 0.27 | 0.38 |
Key variables in driving compound MHW and RI events
Several characteristics of MHWs and TCs play important roles in initiating RI events. We examine which TC and MHW properties provide the greatest value in predicting whether RI will occur over the subsequent 24 hours. Using the XGBoost machine learning (ML) algorithm, a gradient-boosted decision tree algorithm, in combination with the Shapley additive explanations (SHAP) (67) framework, we evaluate the contribution of various TC and MHW predictors (listed in table S3) to understand which factors most strongly influence the occurrence of RI in the presence of MHWs. All MHW metrics in our analysis are area-averaged values calculated from MHWs ranging from ~7580 km2 to ~1.9 million km2 in size. XGBoost was selected for its superior performance with interpretable ML using SHAP (68) and its demonstrated effectiveness in our previous work (69) for RI prediction in the presence of MHWs. SHAP values quantify how much each predictor contributes to the model’s prediction, allowing us to identify the key drivers of MHW-induced RI events and their relative importance in the prediction process. The optimized hyperparameters for the XGBoost model are provided in the Supplementary Materials (table S4). For detailed methodology, please refer to the “Predictor importance analysis” section.
The most important factor is the wind speed at the start of the RI process (Fig. 7, A and B). Note that the positive SHAP values on Fig. 7A indicate weighting toward the prediction of RI occurrence, while negative values indicate weighting toward the non-RI prediction. The color scale indicates the magnitude of the predictor across the test dataset (e.g., for wind speed, red indicates high values and blue indicates low values). Furthermore, Fig. 7B shows the average contribution of each predictor to the model’s prediction, ranked from highest to lowest on the basis of their influence, as measured by the mean absolute SHAP value. The results indicate that average initial wind speeds are more attributable to the positive impact on the RI likelihood. Among the TC predictors, TC latitude ranks second. Storms at lower latitudes are more likely to lead to RI, likely due to higher SSTs and atmospheric moisture content closer to the equator, as well as a weaker Coriolis effect enhancing boundary layer processes (70). TC distance to land also plays an important role, for which SHAP values show a rightward skew for larger distances. This predictor represents the distance to the nearest landmass larger than 1400 km2 (approximately the size of Kauai, Hawaii). The peak of the distance is mostly seen around 400 km offshore. The number of RI events in this offshore area has shown an increasing trend since 1980 due to more favorable thermodynamic environments (higher relative humidity and maximum PI), reduced vertical wind shear, global ocean warming, the influence of the Pacific Decadal Oscillation, anthropogenic forcing, and landward migration of TC activity (15).
Fig. 7. Most influential variables on the prediction of MHW-induced RI events.
(A) For each observation, the associated SHAP value of the selected predictor is calculated and represented on a horizontal axis crossing that predictor. Positive SHAP values on the x axis indicate weighting toward the prediction of RI occurrence, while negative values indicate weighting toward the prediction of no RI occurrence. Asymmetry of the individual dot plot of each predictor indicates a skew in the impact, either more positive or more negative, depending on which side the curve leans toward. The color scale represents the value of the predictor across the test dataset (red, high values; and blue, low values). (B) The average impact of each feature on the model output, measured by the mean absolute SHAP value. All MHW metrics in our analysis are area-averaged values.
TC longitude has a modest overall effect, with SHAP values clustered near zero and a slightly broader tail on the negative side. Colors are mixed across signs, indicating no monotonic east-west relationship. Storm direction ranks lower in importance, but higher directional values, which here indicate a more northward or eastward track, tend to appear on the positive SHAP side, suggesting a slight association with increased RI probability. Storm speed is also lower-ranked, with SHAP values narrowly concentrated around zero, indicating minimal overall influence. Physically, the distribution shows that slower-moving storms tend to nudge the model slightly toward predicting RI, due to longer residence times over warm waters.
Among MHW predictors, the most influential feature is the relative mean intensity of MHW (intensity_mean in table S3), with lower values associated with more frequent RI events. The reason is that moderate MHWs occur more frequently near TC activity areas (Fig. 1D). MHW mean and peak intensity are other major contributors. For MHW peak intensity, the distribution is skewed negative, with lower anomalies on the negative side and higher anomalies extending toward positive SHAP. This means weaker peaks are linked to reduced RI likelihood, while stronger peaks have more positive influence, although less frequent. For MHW mean intensity, the distribution has a long negative SHAP tail dominated by lower anomalies, indicating that weaker mean anomalies can substantially reduce RI likelihood in some cases. There is also a cluster of higher anomalies close to zero on the negative side, suggesting that even stronger anomalies sometimes correspond to reduced RI likelihood, due to other limiting environmental factors. On the positive side, the distribution is skewed right with mixed colors, showing that both moderate and high anomalies can push the prediction toward RI, but the effect varies case by case. MHW surface area shows less impact among the displayed factors in Fig. 7, with its distribution tightly clustered near zero SHAP value. However, it is important to note that MHW surface area ranks 11th among all 27 predictors listed in table S3, not the lowest overall. The SHAP values represent relative importance compared to other predictors rather than absolute importance, meaning features with lower SHAP values still contribute to the prediction but with less statistical influence than higher-ranked predictors.
DISCUSSION
MHWs most effectively enhance TC strengthening in tropical basins where moderate temperature anomalies coincide with already favorable conditions for intensification, rather than in regions like mid-latitude western boundary current extensions (e.g., Gulf Stream and Kuroshio) which typically lack the prolonged, low-shear conditions essential for TC RI (Fig. 1). This enhanced effectiveness translates into observable trends, with individual Northern Hemisphere basins, particularly the East Pacific, the North Atlantic, and the North Indian basins, showing clear increases in rapidly intensified MHW TCs, while global patterns rise more modestly due to basin-specific variability (Fig. 3). The niche of MHW impact is not from an increased number of RI TCs, but the increase in their level of strength and destructiveness. Our findings (Fig. 6) show that the occurrence of billion-dollar TCs is ~1.6 times as frequent when MHWs are present (71 MHW TCs compared to 45 non-MHW TCs). This amplification effect becomes particularly concerning in the context of ongoing warming, where TCs appear to be moving more slowly over land, stalling more frequently (71), and possibly decaying less rapidly after landfall (72), thereby increasing the exposure and vulnerability of coastal communities to storm-related hazards.
TCs undergoing RI exhibit different characteristics when associated with proximate MHWs compared to those without MHW influence. This is primarily because MHWs enhance the maximum sustained wind speeds (up to 20% higher speeds in MHW TCs versus non-MHW TCs in Fig. 4) and promote heavy precipitation (up to 12% higher rates in MHW TCs versus non-MHW TCs in Fig. 5) before the storms make landfall in RI TCs. This behavior highlights the compounding effects of MHWs, which not only enhance the intensity of RI events but also sustain higher wind speeds and precipitation rates both before and after landfall. The relationship between maximum storm surge and damage costs (figs. S5 to S7) shows complex, sometimes counterintuitive patterns across different scenarios. This complexity likely stems from the interaction of multiple factors: local coastal topography, timing relative to tidal cycles, presence of coastal defenses, and variations in population density and infrastructure placement.
In terms of MHW detection criteria, we selected the widely used definition from Hobday et al. (73) with a linearly detrended baseline approach to account for long-term warming trends. Removing the long-term warming signal in SST is required because the PI scales with relative SST anomalies rather than absolute SST values (74), making storms more sensitive to detrended SST anomalies than absolute temperatures, which contrasts with biological applications where absolute temperature thresholds are typically more relevant. Previous studies have demonstrated that modifying MHW definitional parameters affects the frequency and characteristics of detected MHWs (29, 69, 75). To evaluate the robustness of our findings against various MHW thresholding scenarios, we conducted a sensitivity analysis in the Gulf of Mexico and Northwest Caribbean Sea region (latitude 15°N to 31°N, longitude 100°W to 78°W) using three different scenarios: the common definition (90th percentile, 5-day duration), reduced duration threshold (90th percentile, 4-day duration), and reduced intensity threshold (80th percentile, 5-day duration). As expected, less stringent criteria identified more MHWs, resulting in more TCs classified as MHW influenced. This reclassification affected uncertainty bands (figs. S8 and S9, A to C), with MHW TC categories showing narrower confidence intervals due to larger sample sizes and non-MHW categories exhibiting wider bands from reduced sample counts. Despite these changes in classification distribution and uncertainty representation, our key findings remained consistent across all definition scenarios, with variations in mean wind speed and precipitation across different lead times changing by less than 10% between definition scenarios (figs. S8 and S9D).
The increasing coastal development in TC-prone areas may exacerbate the relationship between MHW-influenced RI TCs and the vulnerability of coastal communities. The PSM analysis indicates that TCs experiencing both RI and MHWs produced significantly higher damages than other TCs (Table 1), even when controlling for similar levels of coastal development. Building on this understanding and other evidence that MHWs are becoming more frequent and lasting longer under global warming (76), our results offer timely insights into the evolving dynamics of strong and destructive TCs under the influence of MHWs. They also emphasize that the importance of MHW influences should be given greater consideration in future strategies for climate adaptation and coastal hazard preparedness. For example, operational centers could incorporate alert thresholds for local SST anomalies or MHW intensities to enhance the detection and forecasting of RI events (69).
It is important to note that our analysis does not argue that MHWs alone are sufficient to cause RI in all circumstances. Rather, we examine aggregate behavior across hundreds of TCs to identify systematic patterns in how MHWs influence TC characteristics compared to non-MHW conditions. Our findings demonstrate that MHWs generally make TCs more intense across multiple dimensions—whether through enhanced maximum wind speeds, increased precipitation rates, or elevated storm surge—with particularly pronounced effects on the most severe storms (i.e., RI TCs; Figs. 4 and 5). However, the presence of MHWs cannot overcome all unfavorable atmospheric conditions to guarantee RI in every case. Individual TCs are influenced by many competing environmental factors, and determining whether MHW effects dominate over other potentially unfavorable conditions requires detailed case-by-case analysis, as demonstrated in (27). Our contribution lies in establishing that, on average and at the infrastructure exposure level, MHWs systematically enhance TC intensity and substantially increase the likelihood of billion-dollar disasters, providing critical insights for risk assessment and disaster preparedness in a warming climate.
The study acknowledges several limitations related to data, methodology, and generalizability. Emergency Events Database (EM-DAT), as a widely used disaster loss database, suffers from missing and incomplete records due to reporting biases and data collection inconsistencies, which may affect the accuracy of disaster impact assessments (77, 78). Similarly, European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5) data, while valuable for providing global coverage of atmospheric conditions, have recognized limitations in capturing extreme precipitation events, particularly in complex terrains and regions with limited ground observations (79). This could introduce uncertainties in assessing rainfall contributions to TC intensification. Future research could focus on reducing these uncertainties by incorporating alternative precipitation datasets with finer spatial resolutions, such as North American Land Data Assimilation System (NLDAS), Multi-Radar Multi-Sensor (MRMS), and Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG), which might provide additional insights into the precipitation patterns of landfalling TCs. Regarding built-up volume as a proxy for exposure, the GHSL data provide a standardized global measure of urbanization; however, studies have noted its potential underestimation in low-density areas and overestimation in high-density urban centers (80, 81). These limitations, while inherent to the datasets, do not undermine the study’s key findings but should be considered when interpreting the results.
MATERIALS AND METHODS
Data sources
We obtain global TC best-track data during the period of 1981–2023 from the IBTrACS database (82). This dataset has a temporal resolution of 3 hours and a spatial resolution of 0.1° (~10 km). In our analysis, we use the maximum sustained wind speed (knots), the storm center of circulation (degrees latitude/longitude), the storm translation speed (knots), the storm translation direction (in degrees east of north), the current distance to land from current position (i.e., distance to land in km), and the minimum distance to land over next 3 hours (i.e., landfalling distance in kilometers). To identify TCs with direct landfall, we analyze the IBTrACS dataset and detect all locations along the TCs’ tracks whose distance to land and landfalling distance are equal to zero. In total, we found 1600 TCs with direct landfall among 3464 global TCs during the analysis period.
The precipitation data are acquired from the European Centre for Medium–Range Weather Forecasts ERA5 reanalysis (83) with a temporal resolution of 3 hours and a spatial resolution of 0.25°. We use ERA5 precipitation data at different lead hours pre- and postlandfall to understand the impact of MHW and RI events on increasing the TC power. The mean precipitation is calculated by averaging the ERA5 precipitation rates over the grid points within a 5° radii of the TC center. This region encompasses the inner TC rainband, which is the main area experiencing heavy precipitation (84–87). As such, it is a good representation of the potential threat that TC poses to coastal communities (88). Additionally, this region has an important impact on how TCs develop and intensify (89–91).
All analyses in this study focus exclusively on landfalling TCs identified from the IBTrACS dataset during the period 1981–2023, representing ~46% of all global TCs during this timeframe. For the economic loss analysis of TCs with direct landfall, we primarily rely on the EM-DAT (92). It is widely recognized that the availability and quality of damage data are not homogeneous worldwide, as some countries have well-established systems for collecting and disseminating information on the costs incurred, while others lack the resources or infrastructure to accurately assess and report the economic impacts (77, 93). Despite this limitation, we have done our best to complete and validate our TC damage database. The first challenge in using EM-DAT is that the event names in this database have discrepancies with the storm names reported by IBTrACS. Therefore, we reviewed all event names to ensure that the damage considered is accurately reflected. After this control step, we retrieve 644 damage data associated with 522 TCs and 46 countries worldwide. We searched for available TC reports that include the total damage for the remaining 1078 TCs. To ensure a consistent approach, we added damage data only when the reported economic loss reflects the total damage for the affected country. Using this online search, we added 284 additional damage data from 267 TCs to our database. The resulted global dataset contains the economic damage of 789 TCs with direct landfall from 1981 to 2023.
Definition of RI TC and non-RI TC
An RI event is defined as an increase in the maximum sustained winds of a TC of at least 30 knots over 24 hours (94). Therefore, TCs with at least one RI during their lifetime are referred to as RI TCs; otherwise, they are called non-RI TCs. In total, 1450 of the 3464 TCs (~42%) underwent RI during their lifetime.
Definition of MHW TC and non-MHW TC
Given the objective of this study, the TCs should be classified according to their association with MHWs: When a TC has a close spatiotemporal relationship with MHWs, it is called an MHW-TC, and if a TC is not affected by the MHW region during its lifetime, then it is referred to as a non-MHW-TC. To this end, we first detected global MHWs that occurred between 1981 and 2023 by adapting Hobday’s definition with a linearly detrended baseline approach (73). According to their definition, an MHW event is defined as a period in which SSTs are above the local and seasonally evolving 90th percentile threshold for at least five consecutive days, relative to a climatological mean (average temperature for each day of the year evaluated during the baseline period). While marine biological studies often use time-invariant thresholds due to organisms’ responses to absolute temperature thresholds, TC studies require consideration of relative SST anomalies as TC PI scales with relative rather than absolute SST (74). Therefore, we apply a linearly detrended baseline approach that removes the long-term warming signal, ensuring that only short-term anomalous temperature excursions above the evolving background state are classified as MHWs (45, 95). In our analysis, the baseline period is 30 years and extends from 1991 to 2020. The identified MHW events are also used for global MHW analysis presented in Fig. 1. For comparison with previous studies that used fixed baselines, we provide the equivalent of Fig. 1 using a fixed baseline approach in fig. S1 and the equivalent of Table 1 using fixed baseline methodology in table S2, demonstrating that our key findings remain robust across different MHW detection methodologies.
To investigate how an MHW affects a TC, we identified MHW TCs as TCs whose centers remained within 1° of an MHW event that had been present 10 days or less before the TC’s record time. The spatial threshold can encompass the inner and outer core areas of the TC (96) because air-sea enthalpy fluxes in these regions play an important role in maintaining and enhancing the TC strength (49, 97, 98). The temporal threshold is also assumed to cover the period when MHWs have a better impact on TC dynamics (99). For each TC event, MHW features, including duration, intensity, rate of onset, and spatial extent, are derived based on MHW events that meet the defined spatial and temporal thresholds. Additionally, for the area covered by MHWs, the cumulative area and the minimum distance to the TC center are computed to better quantify the spatial influence of the MHW on TC dynamics. To ensure the MHW’s influence on intensification near landfall, we exclude TCs that encountered MHW regions less than 24 hours before landfall. The resulting feature set is used for subsequent analysis of MHW-TC interactions (see table S3 for the full list of features).
Tropical cyclone potential intensity
To assess the maximum theoretical intensity that TCs could achieve in the given thermodynamic environment, we calculated PI. PI represents the thermodynamic upper bound on TC intensity (42). The maximum PI, , is given by (100)
| (1) |
where and are the surface exchange coefficients for enthalpy and momentum, respectively; is the SST; is the outflow temperature at the level of neutral buoyancy; is the saturation moist static energy at the ocean surface; and is the saturation moist static energy of the overlying atmosphere. Following the methodology of Gilford (101), we used environmental inputs (SST, atmospheric temperature, and specific humidity profiles across multiple pressure levels, and mean sea level pressure) required to compute convective available potential energy, which serves as a proxy for moist static energy differences in the PI formulation.
PSM for exposure control
To isolate the effect of MHWs on RI TCs while controlling for exposure, we used PSM. This statistical technique creates balanced comparison groups by matching treatment observations with control observations that have similar characteristics, allowing for causal inference in observational studies (65). The propensity score is defined as the conditional probability of receiving the treatment given the observed covariates
| (2) |
where is the treatment indicator (1 = RI with MHW, 0 = comparison group) and represents the covariate (i.e., exposure). Built-up volume is used as a proxy for the level of development and exposure of the TC-affected areas. It represents the three-dimensional structure of the built environment, capturing the extent and density of buildings and infrastructure within a given region. The built-up volume data, available in 5-year intervals from the GHSL dataset (66), are matched to the TC season by selecting the closest available year as the reference year for each event. To quantify the built environment affected by TC intensities, the built-up volumes of all grid cells within a 5° radius of the TC eye over land are aggregated. This aggregated value served as an exposure indicator for the specific damage data.
In our study, we defined three separate comparisons: group 1: treatment (RI with MHW) versus control 1 (RI without MHW); group 2: treatment (RI with MHW) versus control 2 (no RI with MHW); and group 3: treatment (RI with MHW) versus control 3 (no RI without MHW).
For each comparison, we estimated propensity scores using logistic regression
| (3) |
The model estimates the probability of a TC being in the treatment group (RI with MHW) based solely on its associated built-up volume.
We implemented nearest-neighbor matching with replacement, which selects for each treatment observation the control observation with the closest propensity score. The matching process can be formalized as
| (4) |
where denotes the set of control units matched to treated unit i and and are the propensity scores for units i and j, respectively.
Then, we assessed the quality of matching by comparing the distribution of built-up volume between treatment and control groups before and after matching. The balance was evaluated using percentage difference, calculated as , where and are the means of the treatment and control groups, respectively. Before matching, there was a 25% absolute standardized mean difference in built-up volume between treatment and control groups, indicating substantial imbalance. After matching, this difference was reduced to 0.91%, demonstrating the effectiveness of PSM in creating comparable groups.
After matching, we calculated the percentage difference in the average treatment effect (ATE) as the mean difference in damage between the matched treatment and control groups
| (5) |
where is the number of treatment units, is the damage for treatment unit i, is the damage for the control unit matched to treatment unit i, and and are the mean damages for the treatment and control groups, respectively.
To assess the statistical significance of the treatment effect, we applied: (i) two-sample t test, which tests the null hypothesis that the means of the two independent groups are equal; and (ii) Mann-Whitney U test, which is a nonparametric test that assesses whether the distribution of damages in the treatment group is stochastically greater than in the control group, which is more robust for nonnormally distributed data like damage costs.
Predictor importance analysis
The application of interpretable ML techniques is particularly valuable in meteorological domains where physical understanding is critical for model acceptance and operational implementation (102). To analyze the importance of each TC and MHW predictors on the formation of RI TCs, we use the XGBoost classifier (a gradient-boosted decision tree algorithm) on our dataset and then use the SHAP model explainer (67) to interpret the model’s predictions. We use the complete set of TC and MHW predictors listed in table S1. We perform hyperparameter tuning using grid search to find the best parameter set, with the final parameter values provided in table S4.
After training the optimized model, we use SHAP to provide interpretable insights into the model’s behavior. SHAP enumerates predictor impact through contribution scores known as Shapley values. SHAP values are calculated by quantifying the change in output’s predictions (here, RI event) with and without the predictor of interest when all possible combinations of predictors are considered. In the Supplementary Materials, the rationale behind the SHAP method’s selection is described.
The SHAP methodology is built on cooperative game theory, specifically Shapley values, which distribute the “payout” (prediction) among the “players” (features). For each predictor, SHAP calculates its contribution by evaluating the model’s prediction with every possible combination of features with and without the target predictor. The mathematical representation of Shapley values is (103)
| (6) |
where j is the given feature, S is any subset of features that does not include the jth feature and is its size, N is the number of all features, represents all combinations with adding jth feature, is all combinations without j’s presence, function defines the prediction output of the model, and denotes summation over all possible combinations without j.
The SHAP beeswarm plot (Fig. 7A) illustrates the global impact of each predictor. This plot displays detailed explanations of the effects of each predictor on the outcome of interest. The y axis lists predictors, ordered by their overall importance. Each dot represents a single MHW-TC observation. For each observation, the associated SHAP value of the selected predictor is calculated and represented on a horizontal axis crossing that predictor. This value indicates how much that predictor influenced the prediction for that specific MHW-TC event. Positive SHAP values on the x axis indicate weighting toward the prediction of RI occurrence while negative values indicate weighting toward the prediction of no RI occurrence. Asymmetry of the individual dot plot of each predictor indicates a skew in the impact, either more positive or more negative, depending on which side the curve leans toward. The color scale represents the value of the predictor across the test dataset (red, high values; and blue, low values).
The SHAP bar plot (Fig. 7B) provides the average contributions of different predictors on the model’s prediction. This represents the overall importance of each predictor in the model, regardless of whether it tends to increase or decrease RI likelihood. Basically, the more absolute values of these individual SHAP values increase, the more the predictor is likely to impact the prediction. We ranked all variables by their overall impact as determined by the mean absolute SHAP values. A higher mean absolute SHAP value indicates the predictor has a larger impact on the model’s predictions overall.
Acknowledgments
Funding:
Funding was awarded to Hamed Moftakhari through the Cooperative Institute for Research to Operations in Hydrology (CIROH) under the NOAA Cooperative Agreement with The University of Alabama (NA22NWS4320003). Partial support was also awarded to Hamed Moftakhari through NSF award no. 2223893.
Author contributions:
Conceptualization: S.R., H.Mof., H.Mor., and E.F. Methodology: S.R., A.S.G., G.R.F., and H.Mor. Investigation: S.R. Data curation: S.R. Formal analysis: S.R. and H.Mor. Validation: S.R. Software: S.R. Visualization: S.R. Supervision: H.Mof., H.Mor., and G.R.F. Funding acquisition: H.Mof. and H.Mor. Writing—original draft: S.R. Writing—review and editing: A.S.G., G.R.F., H.Mof., E.F., and H.Mor.
Competing interests:
The authors declare that they have no competing interests.
Data, code, and materials availability:
All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. All codes to reproduce the results of this paper are available at https://doi.org/10.5281/zenodo.15242051 (107) and the GitHub repository: https://github.com/sradfar/BillionTCs. TC best track data can be obtained from the International Best Track Archive for Climate Stewardship (IBTrACS) dataset (82), which is freely accessible in CSV format via the National Centers for Environmental Information (NCEI) website at www.ncei.noaa.gov/data/international-best-track-archive-for-climate-stewardship-ibtracs/v04r00/access/csv/. SST data can be obtained from the NOAA Optimum Interpolation Sea Surface Temperature version 2.1 (NOAA OISST v2.1) dataset (108), available in NetCDF format and can be accessed through the ERDDAP data server at www.ncei.noaa.gov/erddap/info/index.html?page=1&itemsPerPage=1000. Precipitation data for this study are from the ECMWF Reanalysis version 5 available at https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=overview. The primary source for the economic loss information of landfalling TCs is the Emergency Events Database (EM-DAT) that can be found at www.emdat.be/. The built-up volume data used in this analysis are downloaded from the GHSL dataset, covering the period from 1980 to 2025. This dataset is publicly available through the Copernicus Emergency Management Service at https://human-settlement.emergency.copernicus.eu/download.php. For this study, the data at 30–arc sec resolution are used. This study did not generate new materials.
Supplementary Materials
This PDF file includes:
Figs. S1 to S10
Tables S1 to S4
REFERENCES
- 1.Knutson T., Camargo S. J., Chan J. C. L., Emanuel K., Ho C.-H., Kossin J., Mohapatra M., Satoh M., Sugi M., Walsh K., Wu L., Tropical cyclones and climate change assessment: Part I: Detection and attribution. Bull. Am. Meteorol. Soc. 100, 1987–2007 (2019). [Google Scholar]
- 2.Kossin J. P., Knapp K. R., Olander T. L., Velden C. S., Global increase in major tropical cyclone exceedance probability over the past four decades. Proc. Natl. Acad. Sci. U.S.A. 117, 11975–11980 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Jewson S., Tropical cyclones and climate change: Global landfall frequency projections derived from Knutson et al. Bull. Am. Meteorol. Soc. 104, E1085–E1104 (2023). [Google Scholar]
- 4.Knutson T. R., Sirutis J. J., Bender M. A., Tuleya R. E., Schenkel B. A., Dynamical downscaling projections of late twenty-first-century US landfalling hurricane activity. Clim. Change 171, 28 (2022). [Google Scholar]
- 5.B. J. Reinhart, A. Reinhart, “Hurricane Otis report” (EP182023, National Hurricane Center, 2024).
- 6.Wang W., Zhang Z., Cangialosi J. P., Brennan M., Cowan L., Clegg P., Hosomi T., Masaaki I., Das A. K., Mohapatra M., Sharma M., Knaff J. A., Kaplan J., Birchard T., Doyle J. D., Heming J., Moskaitis J., Komaromi W. A., Ma S., Sampson C., Wu L., Blake E., A review of recent advances (2018-2021) on tropical cyclone intensity change from operational perspectives, Part 2: Forecasts by Operational Centers. Trop. Cyclone Res. Rev. 12, 50–63 (2023). [Google Scholar]
- 7.L. Bucci, L. Alaka, A. Hagen, S. Delgado, J. Beven, “Hurricane Ian report” (AL092022, National Hurricane Center, 2023).
- 8.A. Francis, B. Strahl, “Annual tropical cyclone report 2020” (Joint Typhoon Warning Center, 2020); www.metoc.navy.mil/jtwc/products/atcr/2020atcr.pdf.
- 9.Bhatia K. T., Vecchi G. A., Knutson T. R., Murakami H., Kossin J., Dixon K. W., Whitlock C. E., Recent increases in tropical cyclone intensification rates. Nat. Commun. 10, 635 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Bhatia K. T., Baker A., Yang W., Vecchi G., Knutson T., Murakami H., Kossin J., Hodges K., Dixon K., Bronselaer B., Whitlock C., A potential explanation for the global increase in tropical cyclone rapid intensification. Nat. Commun. 13, 6626 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Balaguru K., Chang C.-C., Leung L. R., Foltz G. R., Hagos S. M., Wehner M. F., Kossin J. P., Ting M., Xu W., A global increase in nearshore tropical cyclone intensification. Earth’s Future 12, e2023EF004230 (2024). [Google Scholar]
- 12.Klotzbach P. J., Wood K. M., Schreck C. J. III, Bowen S. G., Patricola C. M., Bell M. M., Trends in global tropical cyclone activity: 1990–2021. Geophys. Res. Lett. 49, e2021GL095774 (2022). [Google Scholar]
- 13.Bhatia K., Vecchi G., Murakami H., Underwood S., Kossin J., Projected response of tropical cyclone intensity and intensification in a global climate model. J. Clim. 31, 8281–8303 (2018). [Google Scholar]
- 14.Emanuel K., Will global warming make hurricane forecasting more difficult? Bull. Am. Meteorol. Soc. 98, 495–501 (2017). [Google Scholar]
- 15.Li Y., Tang Y., Wang S., Toumi R., Song X., Wang Q., Recent increases in tropical cyclone rapid intensification events in global offshore regions. Nat. Commun. 14, 5167 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Pérez-Alarcón A., Fernández-Alvarez J. C., Coll-Hidalgo P., Global increase of the intensity of tropical cyclones under global warming based on their maximum potential intensity and CMIP6 models. Environ. Process. 10, 36 (2023). [Google Scholar]
- 17.Patricola C. M., Wehner M. F., Anthropogenic influences on major tropical cyclone events. Nature 563, 339–346 (2018). [DOI] [PubMed] [Google Scholar]
- 18.Choi H.-Y., Park M.-S., Kim H.-S., Lee S., Marine heatwave events strengthen the intensity of tropical cyclones. Commun. Earth Environ. 5, 69 (2024). [Google Scholar]
- 19.Knutson T., Camargo S. J., Chan J. C. L., Emanuel K., Ho C.-H., Kossin J., Mohapatra M., Satoh M., Sugi M., Walsh K., Wu L., Tropical cyclones and climate change assessment: Part II: Projected response to anthropogenic warming. Bull. Am. Meteorol. Soc. 101, E303–E322 (2020). [Google Scholar]
- 20.Liu M., Vecchi G. A., Smith J. A., Knutson T. R., Causes of large projected increases in hurricane precipitation rates with global warming. npj Clim. Atmos. Sci. 2, 38 (2019). [Google Scholar]
- 21.Pérez-Alarcón A., Fernández-Alvarez J. C., Trigo R. M., Nieto R., Gimeno L., Insights into the origin of precipitation moisture for tropical cyclones during rapid intensification process. Trop. Cyclone Res. Rev. 13, 72–87 (2024). [Google Scholar]
- 22.Tan X., Liu Y., Wu X., Liu B., Chen X., Examinations on global changes in the total and spatial extent of tropical cyclone precipitation relating to rapid intensification. Sci. Total Environ. 853, 158555 (2022). [DOI] [PubMed] [Google Scholar]
- 23.Tory K. J., Dare R. A., Sea surface temperature thresholds for tropical cyclone formation. J. Climate 28, 8171–8183 (2015). [Google Scholar]
- 24.Foltz G. R., Balaguru K., Hagos S., Interbasin differences in the relationship between SST and tropical cyclone intensification. Mon. Weather Rev. 146, 853–870 (2018). [Google Scholar]
- 25.Kanada S., Aiki H., Buffering effect of atmosphere–ocean coupling on intensity changes of tropical cyclones under a changing climate. Geophys. Res. Lett. 51, e2023GL105659 (2024). [Google Scholar]
- 26.Oguejiofor C. N., Wainwright C. E., Rudzin J. E., Richter D. H., Onset of tropical cyclone rapid intensification: Evaluating the response to length scales of sea surface temperature anomalies. J. Atmos. Sci. 80, 1971–1994 (2023). [Google Scholar]
- 27.Rathore S., Goyal R., Jangir B., Ummenhofer C. C., Feng M., Mishra M., Interactions between a marine heatwave and Tropical Cyclone Amphan in the Bay of Bengal in 2020. Front. Clim. 4, 861477 (2022). [Google Scholar]
- 28.Mawren D., Hermes J., Reason C., Marine heat waves and tropical cyclones-Two devastating types of coastal hazard in South-eastern Africa. Estuar. Coast. Shelf Sci. 277, 108056 (2022). [Google Scholar]
- 29.Radfar S., Moftakhari H., Moradkhani H., Rapid intensification of tropical cyclones in the Gulf of Mexico is more likely during marine heatwaves. Commun. Earth Environ. 5, 421 (2024). [Google Scholar]
- 30.Somavilla R., González-Pola C., Fernández-Diaz J., The warmer the ocean surface, the shallower the mixed layer. How much of this is true? J. Geophys. Res. Oceans 122, 7698–7716 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Gupta A. S., Thomsen M., Benthuysen J. A., Hobday A. J., Oliver E., Alexander L. V., Burrows M. T., Donat M. G., Feng M., Holbrook N. J., Perkins-Kirkpatrick S., Moore P. J., Rodrigues R. R., Scannell H. A., Taschetto A. S., Ummenhofer C. C., Wernberg T., Smale D. A., Drivers and impacts of the most extreme marine heatwave events. Sci. Rep. 10, 19359 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Holbrook N. J., Sen Gupta A., Oliver E. C., Hobday A. J., Benthuysen J. A., Scannell H. A., Smale D. A., Wernberg T., Keeping pace with marine heatwaves. Nat. Rev. Earth Environ. 1, 482–493 (2020). [Google Scholar]
- 33.Pak G., Noh J., Park Y.-G., Jin H., Park J.-H., Governing factors of the record-breaking marine heatwave over the mid-latitude western North Pacific in the summer of 2021. Front. Mar. Sci. 9, 946767 (2022). [Google Scholar]
- 34.Oliver E. C., Benthuysen J. A., Darmaraki S., Donat M. G., Hobday A. J., Holbrook N. J., Schlegel R. W., Sen Gupta A., Marine heatwaves. Ann. Rev. Mar. Sci. 13, 313–342 (2021). [DOI] [PubMed] [Google Scholar]
- 35.Balaguru K., Chang P., Saravanan R., Leung L. R., Xu Z., Li M., Hsieh J.-S., Ocean barrier layers’ effect on tropical cyclone intensification. Proc. Natl. Acad. Sci. U.S.A. 109, 14343–14347 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Holbrook N. J., Scannell H. A., Gupta A. S., Benthuysen J. A., Feng M., Oliver E. C. J., Alexander L. V., Burrows M. T., Donat M. G., Hobday A. J., Moore P. J., Perkins-Kirkpatrick S. E., Smale D. A., Straub S. C., Wernberg T., A global assessment of marine heatwaves and their drivers. Nat. Commun. 10, 2624 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Yu J., Zhang H., Wang H., Tian D., Li J., Upper-ocean structure variability in the Northwest Pacific Ocean in response to tropical cyclones. Front Mar. Sci. 10, 1245348 (2023). [Google Scholar]
- 38.Mei W., Xie S.-P., Intensification of landfalling typhoons over the northwest Pacific since the late 1970s. Nat. Geosci. 9, 753–757 (2016). [Google Scholar]
- 39.Kossin J. P., Emanuel K. A., Vecchi G. A., The poleward migration of the location of tropical cyclone maximum intensity. Nature 509, 349–352 (2014). [DOI] [PubMed] [Google Scholar]
- 40.Klein P. M., Harr P. A., Elsberry R. L., Extratropical transition of western North Pacific tropical cyclones: An overview and conceptual model of the transformation stage. Weather Forecast. 15, 373–395 (2000). [Google Scholar]
- 41.Bieli M., Camargo S. J., Sobel A. H., Evans J. L., Hall T., A global climatology of extratropical transition. Part I: Characteristics across basins. J. Clim. 32, 3557–3582 (2019). [Google Scholar]
- 42.Emanuel K. A., The dependence of hurricane intensity on climate. Nature 326, 483–485 (1987). [Google Scholar]
- 43.Demaria M., Kaplan J., Sea surface temperature and the maximum intensity of Atlantic tropical cyclones. J. Clim. 7, 1324–1334 (1994). [Google Scholar]
- 44.Giménez L., Boersma M., Wiltshire K. H., A multiple baseline approach for marine heatwaves. Limnol. Oceanogr. 69, 638–651 (2024). [Google Scholar]
- 45.Smith K. E., Gupta A. S., Amaya D., Benthuysen J. A., Burrows M. T., Capotondi A., Filbee-Dexter K., Frölicher T. L., Hobday A. J., Holbrook N. J., Malan N., Moore P. J., Oliver E. C. J., Richaud B., Salcedo-Castro J., Smale D. A., Thomsen M., Wernberg T., Baseline matters: Challenges and implications of different marine heatwave baselines. Prog. Oceanogr. 231, 103404 (2025). [Google Scholar]
- 46.Deshpande M., Singh V. K., Ganadhi M. K., Roxy M., Emmanuel R., Kumar U., Changing status of tropical cyclones over the north Indian Ocean. Clim. Dyn. 57, 3545–3567 (2021). [Google Scholar]
- 47.Chand S. S., Walsh K. J., Camargo S. J., Kossin J. P., Tory K. J., Wehner M. F., Chan J. C., Klotzbach P. J., Dowdy A. J., Bell S. S., Declining tropical cyclone frequency under global warming. Nat. Clim. Change 12, 655–661 (2022). [Google Scholar]
- 48.P. J. Klotzbach, M. M. Bell, A. J. DesRosiers, L. G. Silvers, “Discussion of 2024 Atlantic hurricane season to date and forecast thoughts on the rest of the season” (Department of Atmospheric Science, Colorado State University, 2024).
- 49.Le Hénaff M., Domingues R., Halliwell G., Zhang J. A., Kim H.-S., Aristizabal M., Miles T., Glenn S., Goni G., The role of the Gulf of Mexico ocean conditions in the intensification of Hurricane Michael (2018). J. Geophys. Res. Oceans 126, e2020JC016969 (2021). [Google Scholar]
- 50.Song J. Y., Alipour A., Moftakhari H. R., Moradkhani H., Toward a more effective hurricane hazard communication. Environ. Res. Lett. 15, 064012 (2020). [Google Scholar]
- 51.Alipour A., Yarveysi F., Moftakhari H., Song J. Y., Moradkhani H., A multivariate scaling system is essential to characterize the tropical cyclones' risk. Earth’s Future 10, e2021EF002635 (2022). [Google Scholar]
- 52.Phillipson L. M., Toumi R., A physical interpretation of recent tropical cyclone post-landfall decay. Geophys. Res. Lett. 48, e2021GL094105 (2021). [Google Scholar]
- 53.Rappaport E. N., Franklin J. L., Schumacher A. B., DeMaria M., Shay L. K., Gibney E. J., Tropical cyclone intensity change before US Gulf Coast landfall. Weather Forecast. 25, 1380–1396 (2010). [Google Scholar]
- 54.Song J., Klotzbach P. J., Duan Y., Guo H., Recent increase in tropical cyclone weakening rates over the western North Pacific. Geophys. Res. Lett. 47, e2020GL090337 (2020). [Google Scholar]
- 55.Colomb A., Kriat T., Leroux M.-D., On the rapid weakening of very intense tropical cyclone Hellen (2014). Mon. Weather Rev. 147, 2717–2737 (2019). [Google Scholar]
- 56.Zagrodnik J. P., Jiang H., Rainfall, convection, and latent heating distributions in rapidly intensifying tropical cyclones. J. Atmos. Sci. 71, 2789–2809 (2014). [Google Scholar]
- 57.Rogers R., Reasor P., Lorsolo S., Airborne Doppler observations of the inner-core structural differences between intensifying and steady-state tropical cyclones. Mon. Weather Rev. 141, 2970–2991 (2013). [Google Scholar]
- 58.Lockwood J. W., Lin N., Gori A., Oppenheimer M., Increasing flood hazard posed by tropical cyclone rapid intensification in a changing climate. Geophys. Res. Lett. 51, e2023GL105624 (2024). [Google Scholar]
- 59.Lu P., Lin N., Emanuel K., Chavas D., Smith J., Assessing hurricane rainfall mechanisms using a physics-based model: Hurricanes Isabel (2003) and Irene (2011). J. Atmos. Sci. 75, 2337–2358 (2018). [Google Scholar]
- 60.Zhai G., Xu W., Su P., Qin L., Liao X., Characteristics of precipitation changes during tropical cyclone processes in China from 1980 to 2019. Sci. Rep. 14, 13654 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 61.Richardson J. C., Torn R. D., Tang B. H., An analog comparison between rapidly and slowly intensifying tropical cyclones. Mon. Weather Rev. 150, 2139–2156 (2022). [Google Scholar]
- 62.Lonfat M., Marks F. D. Jr., Chen S. S., Precipitation distribution in tropical cyclones using the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager: A global perspective. Mon. Weather Rev. 132, 1645–1660 (2004). [Google Scholar]
- 63.Chen S. S., Knaff J. A., Marks F. D. Jr., Effects of vertical wind shear and storm motion on tropical cyclone rainfall asymmetries deduced from TRMM. Mon. Weather Rev. 134, 3190–3208 (2006). [Google Scholar]
- 64.Tao C., Jiang H., Zawislak J., The relative importance of stratiform and convective rainfall in rapidly intensifying tropical cyclones. Mon. Weather Rev. 145, 795–809 (2017). [Google Scholar]
- 65.Rosenbaum P. R., Rubin D. B., The central role of the propensity score in observational studies for causal effects. Biometrika 70, 41–55 (1983). [Google Scholar]
- 66.M. Pesaresi, P. Politis, GHS-BUILT-V R2022A-GHS built-up volume grids derived from joint assessment of Sentinel2, Landsat, and global DEM data, for 1975-2030, European Commission, Joint Research Centre (JRC) (2022).
- 67.Lundberg S. M., Lee S.-I., A unified approach to interpreting model predictions. Adv. Neural Inf. Proces. Syst. 30, 4768–4777 (2017). [Google Scholar]
- 68.E. Dillon, J. LaRiviere, S. Lundberg, J. Roth, V. Syrgkanis, Be careful when interpreting predictive models in search of causal insights (2021); https://medium.com/data-science/be-careful-when-interpreting-predictive-models-in-search-of-causal-insights-e68626e664b6.
- 69.Radfar S., Foroumandi E., Moftakhari H., Moradkhani H., Foltz G. R., Sen Gupta A., Global predictability of marine heatwave induced rapid intensification of tropical cyclones. Earth’s Future 12, e2024EF004935 (2024). [Google Scholar]
- 70.Li T., Ge X., Peng M., Wang W., Dependence of tropical cyclone intensification on the Coriolis parameter. Trop. Cyclone Res. Rev. 1, 242–253 (2012). [Google Scholar]
- 71.Yamaguchi M., Chan J. C. L., Moon I.-J., Yoshida K., Mizuta R., Global warming changes tropical cyclone translation speed. Nat. Commun. 11, 47 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Li L., Chakraborty P., Slower decay of landfalling hurricanes in a warming world. Nature 587, 230–234 (2020). [DOI] [PubMed] [Google Scholar]
- 73.Hobday A. J., Alexander L. V., Perkins S. E., Smale D. A., Straub S. C., Oliver E. C., Benthuysen J. A., Burrows M. T., Donat M. G., Feng M., A hierarchical approach to defining marine heatwaves. Prog. Oceanogr. 141, 227–238 (2016). [Google Scholar]
- 74.Vecchi G. A., Swanson K. L., Soden B. J., Whither hurricane activity? Science 322, 687–689 (2008). [DOI] [PubMed] [Google Scholar]
- 75.Zhang X., Zheng F., Zhu J., Chen X., Observed frequent occurrences of marine heatwaves in most ocean regions during the last two decades. Adv. Atmos. Sci. 39, 1579–1587 (2022). [Google Scholar]
- 76.Frölicher T. L., Fischer E. M., Gruber N., Marine heatwaves under global warming. Nature 560, 360–364 (2018). [DOI] [PubMed] [Google Scholar]
- 77.Jones R. L., Kharb A., Tubeuf S., The untold story of missing data in disaster research: A systematic review of the empirical literature utilising the Emergency Events Database (EM-DAT). Environ. Res. Lett. 18, 103006 (2023). [Google Scholar]
- 78.Mazhin S. A., Farrokhi M., Noroozi M., Roudini J., Hosseini S. A., Motlagh M. E., Kolivand P., Khankeh H., Worldwide disaster loss and damage databases: A systematic review. J. Educ. Health Promot. 10, 329 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 79.Xie W., Yi S., Leng C., Xia D., Li M., Zhong Z., Ye J., The evaluation of IMERG and ERA5-Land daily precipitation over China with considering the influence of gauge data bias. Sci. Rep. 12, 8085 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 80.J. H. Uhl, S. Leyk, Uncertainty prediction of built-up areas from global human settlement data in the United States based on landscape metrics. arXiv:2205.09023 [physics.soc-ph] (2022).
- 81.Melchiorri M., The global human settlement layer sets a new standard for global urban data reporting with the urban centre database. Front. Environ. Sci. 10, 1003862 (2022). [Google Scholar]
- 82.K. R. D. Knapp, J. Howard, J. P. Kossin, M. C. Kruk, C. J. Schreck III, International Best Track Archive for Climate Stewardship (IBTrACS) Project, version 4, NOAA National Centers for Environmental Information (2018); 10.25921/82ty-9e16. [DOI]
- 83.Hersbach H., Bell B., Berrisford P., Hirahara S., Horányi A., Muñoz-Sabater J., Nicolas J., Peubey C., Radu R., Schepers D., The ERA5 global reanalysis. Q. J. Roy. Meteorol. Soc. 146, 1999–2049 (2020). [Google Scholar]
- 84.Emanuel K., 100 years of progress in tropical cyclone research. Meteorol. Monogr. 59, 15.1–15.68 (2018). [Google Scholar]
- 85.Senn H., Hiser H., On the origin of hurricane spiral rain bands. J. Atmos. Sci. 16, 419–426 (1959). [Google Scholar]
- 86.Willoughby H. E., Marks F. D., Feinberg R. J., Stationary and moving convective bands in hurricanes. J. Atmos. Sci. 41, 3189–3211 (1984). [Google Scholar]
- 87.Samsury C. E., Zipser E. J., Secondary wind maxima in hurricanes: Airflow and relationship to rainbands. Mon. Weather Rev. 123, 3502–3517 (1995). [Google Scholar]
- 88.Yu C.-K., Lin C.-Y., Pun C.-H., Origin of outer tropical cyclone rainbands. Nat. Commun. 14, 7061 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 89.May P. T., Holland G. J., The role of potential vorticity generation in tropical cyclone rainbands. J. Atmos. Sci. 56, 1224–1228 (1999). [Google Scholar]
- 90.Houze R. A. Jr., Chen S. S., Lee W.-C., Rogers R. F., Moore J. A., Stossmeister G. J., Bell M. M., Cetrone J., Zhao W., Brodzik S. R., The hurricane rainband and intensity change experiment: Observations and modeling of Hurricanes Katrina, Ophelia, and Rita. Bull. Am. Meteorol. Soc. 87, 1503–1522 (2006). [Google Scholar]
- 91.Wang Y., How do outer spiral rainbands affect tropical cyclone structure and intensity? J. Atmos. Sci. 66, 1250–1273 (2009). [Google Scholar]
- 92.D. Delforge, V. Wathelet, R. Below, C. L. Sofia, M. Tonnelier, J. van Loenhout, N. Speybroeck, EM-DAT: The Emergency Events Database. Research Square (2023). 10.21203/rs.3.rs-3807553/v1. [DOI]
- 93.Jones R. L., Guha-Sapir D., Tubeuf S., Human and economic impacts of natural disasters: Can we trust the global data? Sci. Data 9, 572 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 94.Kaplan J., DeMaria M., Large-scale characteristics of rapidly intensifying tropical cyclones in the North Atlantic basin. Weather Forecast. 18, 1093–1108 (2003). [Google Scholar]
- 95.Capotondi A., Rodrigues R. R., Gupta A. S., Benthuysen J. A., Deser C., Frölicher T. L., Lovenduski N. S., Amaya D. J., Grix N. L., Xu T., Hermes J., Holbrook N. J., Martinez-Villalobos C., Masina S., Roxy M. K., Schaeffer A., Schlegel R. W., Smith K. E., Wang C., A global overview of marine heatwaves in a changing climate. Commun. Earth Environ. 5, 701 (2024). [Google Scholar]
- 96.Weatherford C. L., Gray W. M., Typhoon structure as revealed by aircraft reconnaissance. Part I: Data analysis and climatology. Mon. Weather Rev. 116, 1032–1043 (1988). [Google Scholar]
- 97.Miyamoto Y., Takemi T., An effective radius of the sea surface enthalpy flux for the maintenance of a tropical cyclone. Atmos. Sci. Lett. 11, 278–282 (2010). [Google Scholar]
- 98.Zhang J. A., Cione J. J., Kalina E. A., Uhlhorn E. W., Hock T., Smith J. A., Observations of infrared sea surface temperature and air–sea interaction in Hurricane Edouard (2014) using GPS dropsondes. J. Atmos. Oceanic Tech. 34, 1333–1349 (2017). [Google Scholar]
- 99.Dzwonkowski B., Coogan J., Fournier S., Lockridge G., Park K., Lee T., Compounding impact of severe weather events fuels marine heatwave in the coastal ocean. Nat. Commun. 11, 4623 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 100.Bister M., Emanuel K. A., Dissipative heating and hurricane intensity. Meteorol. Atmos. Phys. 65, 233–240 (1998). [Google Scholar]
- 101.Gilford D. M., pyPI (v1. 3): Tropical cyclone potential intensity calculations in Python. Geosci. Model Dev. 14, 2351–2369 (2021). [Google Scholar]
- 102.McGovern A., Lagerquist R., Gagne D. J., Jergensen G. E., Elmore K. L., Homeyer C. R., Smith T., Making the black box more transparent: Understanding the physical implications of machine learning. Bull. Am. Meteorol. Soc. 100, 2175–2199 (2019). [Google Scholar]
- 103.C. Molnar, Interpretable Machine Learning: A Guide for Making Black Box Models Explainable (ed.2, 2022); https://christophm.github.io/interpretable-ml-book.
- 104.S. Osbrough, H. Ramsay, Tropical cyclone potential intensity (TCPI), version 1, CSIRO Data Collection (2024); 10.25919/x32h-sr47. [DOI]
- 105.J. Cohen, Statistical Power Analysis for the Behavioral Sciences (Routledge, 2013).
- 106.Andrade C., Mean difference, standardized mean difference (SMD), and their use in meta-analysis: As simple as it gets. J. Clin. Psychiatry 81, 20f13681 (2020). [DOI] [PubMed] [Google Scholar]
- 107.S. Radfar, E. Foroumandi, H. Moftakhari, H. Moradkhani, A. Sen Gupta, G. Foltz, Synergistic impact of marine heatwaves and rapid intensification exacerbates tropical cyclone destructive power worldwide (documentation and scripts), version 1.0, Zenodo (2024); 10.5281/zenodo.15242052. [DOI] [PMC free article] [PubMed]
- 108.Huang B., Liu C., Banzon V., Freeman E., Graham G., Hankins B., Smith T., Zhang H.-M., Improvements of the daily optimum interpolation sea surface temperature (DOISST) version 2.1. J. Clim. 34, 2923–2939 (2021). [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Figs. S1 to S10
Tables S1 to S4
Data Availability Statement
All data and code needed to evaluate and reproduce the results in the paper are present in the paper and/or the Supplementary Materials. All codes to reproduce the results of this paper are available at https://doi.org/10.5281/zenodo.15242051 (107) and the GitHub repository: https://github.com/sradfar/BillionTCs. TC best track data can be obtained from the International Best Track Archive for Climate Stewardship (IBTrACS) dataset (82), which is freely accessible in CSV format via the National Centers for Environmental Information (NCEI) website at www.ncei.noaa.gov/data/international-best-track-archive-for-climate-stewardship-ibtracs/v04r00/access/csv/. SST data can be obtained from the NOAA Optimum Interpolation Sea Surface Temperature version 2.1 (NOAA OISST v2.1) dataset (108), available in NetCDF format and can be accessed through the ERDDAP data server at www.ncei.noaa.gov/erddap/info/index.html?page=1&itemsPerPage=1000. Precipitation data for this study are from the ECMWF Reanalysis version 5 available at https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=overview. The primary source for the economic loss information of landfalling TCs is the Emergency Events Database (EM-DAT) that can be found at www.emdat.be/. The built-up volume data used in this analysis are downloaded from the GHSL dataset, covering the period from 1980 to 2025. This dataset is publicly available through the Copernicus Emergency Management Service at https://human-settlement.emergency.copernicus.eu/download.php. For this study, the data at 30–arc sec resolution are used. This study did not generate new materials.







