Introduction

Tropical cyclone (TC) rainfall has become a leading cause of coastal damage in recent decades1,2,3,4, surpassing wind impacts and storm surges and accounting for more than 80% of TC-related socioeconomic losses in some regions5. While debates remain regarding the translation speed and intensity decay of landfalling TCs under global warming6,7,8,9, observations have largely shown a migration of global TCs10 and their associated rainfall11 toward coasts and an increase in TC intensity across global coastal regions12,13. TC-related disasters increasingly threaten inland areas far beyond coastlines, with impacts reaching into East Asian hinterlands and the continental United States14,15,16,17. However, the spatial shifts in TC rainfall lack robust quantification at the global scale, as current evidence largely stems from regional studies17,18,19,20,21. Advancing the understanding of spatial shifts in global TC rainfall, particularly over land22,23, is crucial for disaster mitigation.

Heavy rainfall from TCs poses heightened risks of catastrophic flash floods over land24,25,26,27,28,29,30, and worryingly, most land areas located slightly inland from coastlines (>100 km) remain underprepared14. While changes in multiple TC rainfall metrics, such as rate3,31,32, accumulation33,34,35, and frequency36,37,38 under a warming climate are fairly well studied, little is known about changes in the landward extent of global TC rainfall with rising sea-surface temperature (SST). The observed global radial expansion of TC rainfall footprints relative to TC centers39 highlights the need to detect and attribute directional shifts. In particular, changes in the landward extent of TC heavy rainfall warrant greater attention.

Previous studies have reached a consensus that urbanization intensifies both TC rainfall via urban surface roughness40,41,42,43,44 and flood severity through high impervious area ratios45,46,47, and amplifies exposure risk owing to high population density48. This study identifies the landward extension of global terrestrial TC rainfall under climate warming, highlighting the growing population exposure risk associated with coastal urbanization.

Here, we use satellite-based rainfall observations, including the Multi-Source Weighted-Ensemble Precipitation (MSWEP) and Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG), to quantify changes in the landward extent of terrestrial TC rainfall along continental coasts over the past 44 years. Reanalysis, Coupled Model Intercomparison Project Phase 6 (CMIP6) simulations, and idealized Weather Research and Forecasting (WRF) model simulations are employed to identify potential physical drivers, including nearshore SST warming and coastal urbanization. Finally, satellite imagery reveals an increase in inland population exposure to TC-induced floods, providing independent evidence consistent with our findings.

Results

TC rainfall extends inland

Globally, terrestrial TC heavy rainfall (≥30 mm per 3 h) occurs primarily within 200 km of continental coastlines (Fig. 1). The landward distance of global TC rainfall from coastlines (\(\overline{{{D}_{{{\rm{land}}}}}_{30}^{500}}\); Methods) shows a statistically significant (P < 0.05) increasing trend of 3.3 ± 1.7 km per decade over 1980–2023 (Fig. 1a), suggesting that global TC rainfall is extending inland. The trend is more pronounced for higher-intensity rainfall (Supplementary Fig. 1a), and it remains evident under various spatial constraints (e.g. varying either the inland boundary of the TC center or the definition of the TC rainfall radius; Supplementary Fig. 1b, c), after masking out grids where rainfall uncertainties tend to be amplified (e.g. extreme convective rainfall grids49, coastline-adjacent rainfall grids50, and rapid population-growth grids51; Supplementary Fig. 1d), after removing signals associated with the main climate oscillations (e.g. the El Niño–Southern Oscillation (ENSO) and the Atlantic Multidecadal Oscillation (AMO); Supplementary Fig. 1d), or after excluding tropical depressions (to eliminate the potential influence of improved satellite detection in inland tropical depression records; Supplementary Table 1). Although the trend for Category 3–5 TCs (the group with the smallest sample size) shows no discernible trend, all other TC intensity levels show clear increasing trends, and grouping by translation speed does not affect these trends or their significance (P < 0.1; Supplementary Table 1). It is noted that the landward extension of global TC rainfall, based on the synthetic distance (\(\overline{{D}_{30}^{500}}\); Methods), is statistically significant (P < 0.05) only within the coastal zone (±100 km from coastlines; Supplementary Fig. 2a), even though the increasing trend in the frequency of TC-center records is higher on the ocean side than on the land side within this coastal zone (Supplementary Fig. 2b). This finding remains robust under smaller landmass thresholds and is independent of TC trajectories (Supplementary Fig. 2c).

Fig. 1: Observed changes in the global spatial distribution of terrestrial tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h) over 1980–2023.
Fig. 1: Observed changes in the global spatial distribution of terrestrial tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h) over 1980–2023.The alternative text for this image may have been generated using AI.
Full size image

a Time series of the landward distance of global TC rainfall (solid lines) and its trend (dashed lines), derived from the Multi-Source Weighted-Ensemble Precipitation (MSWEP; Methods) dataset; shaded areas indicate two-sided 95% confidence intervals derived from linear regression (P values are derived from the Mann–Kendall test; Methods). b Linear trends in the landward distance of TC rainfall from the MSWEP (1980–2023; blue bars) and Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG; 1998–2023; red bars) datasets in the Northern Hemisphere (NH) and Southern Hemisphere (SH). Single and double asterisks indicate significance at the 90% and 95% confidence levels, respectively (Mann–Kendall test; Methods); error bars denote 95% confidence intervals. c Global distribution of TC tracks (gray lines) and rainfall frequency (count of occurrences, within a specified grid, where rainfall rate ≥30 mm per 3 h; shading) based on the best-track (Methods) and MSWEP datasets. d–f compare TC rainfall frequency distributions between pre-2001 (blue shading) and post-2001 (red shading) along the continental coasts of the western North Atlantic (d), Bay of Bengal (e), and western North Pacific (f); gray dots denote TC-center records, and the curved black solid and dashed lines denote the 500-km offshore and 100-km inland boundaries, respectively. The straight black solid line denotes the along-coast boundary. TC rainfall frequency is calculated for each 0.1° × 0.1° latitude–longitude grid over land. TC centers are restricted to a coastal zone spanning from 500 km offshore to 100 km inland. A TC rainfall radius threshold of 500 km and a landmass size threshold of 790,000 km2 are applied throughout.

Hemispherically, \(\overline{{{D}_{{{\rm{land}}}}}_{30}^{500}}\) in the Northern Hemisphere (NH) shows a statistically significant (P < 0.05) increase of 3.8 ± 1.8 km per decade, and this result is broadly reproduced with the IMERG dataset (Fig. 1b; Methods). In contrast, no significant trend is observed in the Southern Hemisphere (SH) (Fig. 1b). A similar hemispheric contrast is also evident in \(\overline{{D}_{30}^{500}}\) for TCs within the coastal zone (Supplementary Fig. 2d), which may partly reflect the asymmetric hemispheric distribution of TC activity (~70% of TCs occur in the NH52) and differences in coastal anthropogenic influences53 (e.g. comparatively lower socioeconomic development in northern Australia (NAU)54).

Regionally, statistically significant landward extensions of TC rainfall in the NH are observed along the continental coasts of the western North Atlantic (WNA), Bay of Bengal (BOB), and western North Pacific (WNP); eastern North Pacific TCs are excluded because their trajectories are predominantly oceanward55 (Supplementary Fig. 2d). Specifically, prominent signals are found in the southeastern United States, southeastern Mexico, eastern and southeastern India, Bangladesh, eastern and southeastern China, and the eastern Indo-China Peninsula (Fig. 1c–f). In contrast, the trends in \(\overline{{D}_{30}^{500}}\) are not statistically significant in the two SH regions: eastern Mozambique (EMZ) and NAU (Supplementary Fig. 2d).

Nearshore SST warming

Globally, only in the WNA, BOB, and WNP do both the synthetic distance of TC heavy rainfall (Supplementary Fig. 2d) and nearshore (from 500 km offshore to the coastline) SST (Supplementary Fig. 3) show statistically significant increasing trends over 1980–2023. Heavy rainfall is primarily triggered by strong lower-tropospheric moisture convergence56 and sustained by high specific humidity57 under climate warming. The fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis for the global climate and weather (ERA5; 1980–2023; Methods) during the TC peak season shows statistically significant increases in lower-tropospheric (1000–700 hPa) specific humidity over most of these three basins (from 500 km offshore to 200 km inland), with stronger increases on the ocean side than on the land side (Supplementary Fig. 3a–c). These increases in specific humidity, however, show no systematic landward extension in their spatial pattern. Likewise, trends in monthly lower-tropospheric moisture flux and its convergence over the continental coasts (≤200 km from the coastline) of these three basins are generally insignificant (Supplementary Fig. 4), despite scattered localized significant increases in specific months, such as August in the WNA (Supplementary Fig. 4a), October in the BOB (Supplementary Fig. 4e), and July in the WNP (Supplementary Fig. 4g).

Notably, after removing the nearshore SST trend signal, the landward-trending signal in the synthetic distance of TC heavy rainfall in the WNA, BOB, and WNP largely disappears (Fig. 2a–c). This indicates that nearshore SST warming is closely associated with the observed landward extension of TC heavy rainfall in these basins. The model results of the Detection and Attribution Model Intercomparison Project (DAMIP) in the CMIP6 (1980–2014; Methods) indicate an increasing trend in the nearshore SST in these three basins in the multi-model mean of the historical (HIST) simulations (the medians are close to the corresponding Extended Reconstructed SST (ERSST) trends; Fig. 2d–f). Under greenhouse gas (GHG) forcing, there is a comparable increase in these basins. In contrast, no significant change is found under anthropogenic aerosol (AER) and natural (NAT) forcings. These results suggest that the nearshore SST increase is mainly associated with GHG forcing, whereas AER and NAT contribute little or even negatively (BOB and WNP; Fig. 2e, f).

Fig. 2: Trends in the synthetic distance of tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h) and nearshore sea-surface temperature (SST) in the Northern Hemisphere basins during the TC peak season.
Fig. 2: Trends in the synthetic distance of tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h) and nearshore sea-surface temperature (SST) in the Northern Hemisphere basins during the TC peak season.The alternative text for this image may have been generated using AI.
Full size image

a–c Comparison of trends in the synthetic distance of TC heavy rainfall from the coastline before and after removing the SST trend signal (shown in Supplementary Fig. 3a–c) for basins: western North Atlantic (August–October; a), the Bay of Bengal (May, October, and November; b), and western North Pacific (July–October; c), based on the Multi-Source Weighted-Ensemble Precipitation and Extended Reconstructed SST (ERSST; Methods) datasets (1980–2023). Single and double asterisks indicate significance at the 90% and 95% confidence levels, respectively (Mann–Kendall test; Methods); error bars denote 95% confidence intervals. Trends of the nearshore SST in basins: western North Atlantic (d), the Bay of Bengal (e), and western North Pacific (f), based on the ERSST dataset (1980–2023) and Coupled Model Intercomparison Project Phase 6 (CMIP6) simulations (1980–2014). Simulations are forced by historical (HIST), greenhouse gas (GHG), anthropogenic aerosol (AER), and natural (NAT) forcings, respectively (Methods). Red horizontal line shows the trend from the ERSST dataset. Boxes show interquartile ranges (25th–75th percentiles), center lines indicate medians, and whiskers extend to the 5th and 95th percentiles.

Coastal urbanization effects under SST warming

To examine whether SST warming is associated with the landward extension of TC rainfall and to explore possible mechanisms, we conduct a suite of WRF experiments (Supplementary Fig. 5; Methods). In the control (CTRL; Methods) scenario, the synthetic distances of TC heavy rainfall in the warmer (SST = 302 K) case are significantly (P < 0.05) more landward than those in the cooler (SST = 300 K) case, whereas the distances of TC centers relative to the coastline show no significant difference between the two cases (Fig. 3a). Consistent with this rainfall shift, the warmer case exhibits enhanced lower-tropospheric convergence (below 800 hPa) and a stronger convergent updraft specifically over the baseline urban area, while no comparable enhancement is evident in the ocean-side band symmetric about the TC center (Fig. 3b and Supplementary Fig. 5d, e). Together, these patterns suggest that SST warming is closely associated with the landward extension of TC rainfall, potentially by further amplifying the land–ocean contrast in the intensification of the convergent updraft.

Fig. 3: Numerical experiments and a schematic diagram showing how sea-surface temperature (SST) warming and coastal urbanization can contribute to the landward extension of tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h).
Fig. 3: Numerical experiments and a schematic diagram showing how sea-surface temperature (SST) warming and coastal urbanization can contribute to the landward extension of tropical cyclone (TC) heavy rainfall (≥30 mm per 3 h).The alternative text for this image may have been generated using AI.
Full size image

a SST-warming effects on the synthetic distance of TC rainfall and the distance of TC centers relative to the coastline for the control (CTRL) scenario. b SST-warming-induced changes in meridional-mean profiles (Methods) of horizontal wind convergence (shading) and zonal–vertical wind vectors (vertical wind speed multiplied by 100; arrows) for the CTRL scenario (Supplementary Fig. 5e minus d). c Effects of SST warming and coastal urbanization on the landward distance of TC rainfall relative to the coastline. d Urbanization-induced changes in meridional-mean profiles (Methods) of horizontal wind convergence (shading) and zonal–vertical wind vectors (vertical wind speed multiplied by 100; arrows) under SST warming (Supplementary Fig. 5g minus e). Each boxplot in panels (a) and (c) summarizes 128 ensemble members (8 sub-experiments × 16 3-h intervals; Methods). Single and double asterisks indicate that the difference between the two paired samples is significant at the 90% and 95% confidence levels, respectively (paired-sample t-test; Methods). Thin blue lines at the bottom indicate the zonal extent of baseline urban land in the CTRL scenario, and thick red lines indicate the zonal extent of developed urban land in the urbanization (URBAN) scenario. The red dashed box marks lower-tropospheric anomalous ascent over the urban area. e Schematic of TC rainband enhancement under coastal urbanization with SST warming.

Notably, both the global 30-m land-cover dynamics monitoring product (GLC_FCS30D; 1985–2022; Methods) and European Space Agency Climate Change Initiative Land Cover project (CCI_LC; 1992–2022; Methods) show statistically significant (P < 0.05) landward extension of coastal cities along the continental coasts of the WNA, BOB, and WNP, regardless of the impervious-area-ratio threshold (Supplementary Fig. 6a, b), whereas the urbanization levels in the EMZ and NAU remain substantially lower (Supplementary Fig. 6f, g). Among the three NH basins, this extension of coastal cities is most pronounced in the BOB, reaching 4.6 ± 1.2 km per decade in GLC_FCS30D under an impervious-area-ratio threshold of 0.2 (Supplementary Fig. 6a). This observational evidence motivates our urbanization sensitivity experiments (Methods) to assess whether coastal urbanization can further enhance the landward extension of TC heavy rainfall.

Consistent with this motivation, our idealized WRF experiments show that the landward distances of TC heavy rainfall in the urbanization (URBAN; Methods) scenario significantly (P < 0.05) exceed those in the CTRL scenario under SST warming, whereas no significant difference is found between the two scenarios under the cooler SST condition (Fig. 3c). In the warmer case, enhanced lower-tropospheric convergence (below 600 hPa), convergent updraft, and upper-tropospheric divergence (above 400 hPa) are found over the developed urban area (including the newly expanded area) of the URBAN scenario relative to the CTRL scenario, while no comparable change is evident in the ocean-side band symmetric about the TC center (Fig. 3d and Supplementary Fig. 6e, g). These results suggest that the frictional dynamical effect, arising from a higher surface roughness length40 and a larger urban area in the URBAN scenario under SST warming, favors a stronger and broader in situ secondary circulation that supports TC heavy rainfall and its landward extension (Fig. 3e), consistent with the observations.

Discussion

We identify a landward extension of global terrestrial TC rainfall from 1980 to 2023. The landward distance of TC heavy rainfall (≥30 mm per 3 h) along the continental coasts of the NH shows a statistically significant (P < 0.05) increase at a rate of 3.8 ± 1.8 km per decade. Our results are insensitive to the definition of the spatial extent of terrestrial TC rainfall and broadly consistent across the observational datasets examined. Reanalysis datasets and the CMIP6 multi-model ensemble indicate that nearshore SST warming, which is mainly associated with GHG forcing, is closely linked to the landward extension of TC rainfall. WRF experiments suggest that under nearshore SST warming, the landward extension of coastal urban areas can further enhance the landward extension of TC rainfall.

Interestingly, the landward extension of global TC rainfall is statistically significant (P < 0.05) only within the coastal zone (Supplementary Fig. 2a). This suggests that our result is independent of the global-scale radial expansion of TC rainfall footprints relative to TC centers, as reported by Qin et al.39 in an analysis based primarily on oceanic cases. This inference is further supported by the fact that the landward extension remains significant when we restrict the TC rainfall radius to ≤300 km (Supplementary Fig. 1c), indicating that our result is not explained by the outer-rainband enhancement at 300–500 km highlighted for landfalling TCs over the United States by Wright et al.17. Across the major ocean basins, azimuthally averaged TC size and azimuthal asymmetry in TC structure show no significant change over the past four decades58; therefore, they do not appear to account for this landward extension. This extension is also independent of coastal TC trajectories (no landward trend in TC tracks is detected within ±100 km of the coastline, and no increasing trend is found in TC-center records beyond 100 km inland; Supplementary Figs. 1b, 2c), distinguishing it from the global coastward trend in TC tracks reported by Wang et al.10 and TC rainfall reported by Wei et al.11, both of which are based on analyses that include open-ocean cases.

We further identify a concurrent landward extension of coastal urban areas along the continental coasts (≤200 km from the coastline) of the WNA, BOB, and WNP (Supplementary Fig. 6). Although the landward extension rate of TC heavy rainfall (3.8 km per decade) may appear small relative to a 200-km continental coastal zone width, the superposition of TC heavy rainfall and coastal urbanization is associated with a disproportionate increase in inland population exposure (>100 km from the coastline) to TC rainfall, with the growth rate in the WNP reaching ~2.6 million people per decade (Supplementary Fig. 7). Although TC-induced flood risk is also influenced by factors beyond rainfall (e.g. surface impervious area ratio46,47, topography59, and runoff generation25), this pattern suggests that the superposition may shift TC-induced flood risk further inland (Supplementary Fig. 8). Coastal urban infrastructure planning should therefore account for TC-induced heavy rainfall and flooding in historically less-affected inland regions. This consideration should extend beyond the three NH basins examined here; under a warming climate, regions in the SH with high future urbanization potential (e.g. EMZ and NAU) also warrant close attention.

We also recognize that satellite-based precipitation estimates, including MSWEP and IMERG, share common input data and retrieval methods, which limits their independence. While our main findings are consistently observed across both datasets, this interdependence precludes a truly independent validation. Nevertheless, the robustness of our results is supported by extensive sensitivity experiments, such as varying rainfall-radius thresholds, TC center constraints, and masking grids with potential data quality issues, that address multiple sources of uncertainty.

Note, however, that existing global reanalyses and models may not explicitly represent the physical processes underlying the potential influence of coastal urban extension on coastal TC rainfall, owing to their insufficient resolution. Specifically, the dynamic processes by which surface friction enhances lower-tropospheric convergence, anomalous ascent, and upper-tropospheric divergence—evident only in idealized high-resolution WRF experiments—are not captured by these products. Reanalyses may, in principle, embed these frictional effects via their manifestation in assimilated observations, though global models without data assimilation would miss this entirely. Although relatively weak, anthropogenic heat40,42,44 and aerosols60 may also modulate the influence of urbanization on TC rainfall. Future global convection-permitting/downscaling frameworks with urban-canopy schemes and aerosol-cloud-precipitation interactions are expected to improve understanding of the mechanisms linking coastal urbanization to the landward extension of global TC rainfall.

Methods

TC tracks

TC best-track observations are obtained from the International Best Track Archive for Climate Stewardship v4r01 (1980–2023)61. Only TCs with a lifetime maximum intensity exceeding 34 kt are considered in this study. The non-TC systems (e.g. extratropical cyclones, tropical waves, disturbances, or others) are excluded. Information for each TC observation, including location and maximum sustained wind speed, is provided at 3-h intervals (00:00, 03:00, 06:00, 09:00, 12:00, 15:00, 18:00, and 21:00 UTC). The distance between the TC center and the nearest coastline is computed in ArcGIS using coastline data from the Global Self-consistent, Hierarchical, High-resolution Geography (GSHHG) database62. A positive distance indicates the TC center is over land, while a negative distance indicates it is over the ocean. All TC centers included in the calculations are restricted to within 500 km offshore of the nearest coastline. The synthetic distance for TC centers refers to their average distance from the coastline when they are located within the specified zone (±100 km from coastlines; see the red overlines in Supplementary Fig. 2b). We set 790,000 km² (approximately the size of New Guinea) as the baseline landmass size threshold. Distances are computed only along the coastlines of continents and landmasses larger than this size. To assess sensitivity to this choice, we repeat the analysis with smaller thresholds (10,000, 5000, 2000, and 1400 km²; Supplementary Fig. 2b).

TC rainfall data

TC-related rainfall data are drawn from the MSWEP v2 dataset63, which combines rain gauge, satellite, and reanalysis data to provide reliable global rainfall estimates. MSWEP provides continuous rainfall coverage over both ocean and land since 1980, with a 0.1° spatial resolution and a 3-h temporal resolution, which matches the temporal span of satellite-based TC records. An additional satellite-based rainfall dataset, IMERG v07 (available since 1998)64, is also used, which provides data at a 0.1° spatial resolution and a 0.5-h temporal resolution. However, it should be noted that MSWEP and IMERG are not fully independent, as MSWEP integrates multiple sources including IMERG. For consistency, all rainfall datasets are temporally aggregated to a 3-h resolution (unit: mm 3 h−1). TC rainfall statistics are based on rainfall grids within a 500-km radius39 of the TC center. To assess the sensitivity to the fixed 500-km TC rainfall radius, we repeat the analysis using alternative radii of 200, 300, and 400 km, and further compare the results with a physics-based dynamic tracking method of the TC rainfall (Supplementary Fig. 1c). The rainfall rate associated with TCs is computed by matching each TC center position at the start of each 3-h interval to the corresponding rainfall field, while preserving the original spatial resolution of each dataset. For each rainfall dataset, the distance from the nearest coastline is computed for each grid point using coastlines from the GSHHG database. Based on the MSWEP v2 dataset, we sequentially mask terrestrial rainfall grids with elevated uncertainties49,50,51 (e.g. extreme rainfall grids (≥50 mm 3 h−1), coastline-adjacent grids (≤25 km from coastline), and rapid population-growth grids) to assess the reliability of terrestrial rainfall estimates (Supplementary Fig. 1d). Rapid population-growth grids are defined using the LandScan Global (LSG) dataset65 as those where the 2024 population exceeds twice the 2000 population and the 2000 population density is ≥300 km−2 66.

Distance from coastlines to TC rainfall

For each TC, we use all 3-h observation times during its lifetime at which the center lies within the specified zone to compute a single per-TC landward distance of TC rainfall from the coastline \({{D}_{{{\rm{land}}}}}_{{rr}}^{r}\) (using Eqs. (1) and (3); terrestrial grids only) and a single per-TC synthetic distance of TC rainfall from the coastline \({D}_{{rr}}^{r}\) (using Eqs. (2), (3) and (4); both oceanic and terrestrial grids).

$${{D}_{{{\rm{land}}}}}_{{rr}}^{r}=\frac{{\sum }_{i=1}^{n}{R}_{{{\rm{l}}}i}\times {d}_{{{\rm{l}}}i}}{{\sum }_{i=1}^{n}{R}_{{{\rm{l}}}i}}$$
(1)
$${D}_{{rr}}^{r}=\frac{{\sum }_{i=1}^{n}{R}_{{{\rm{l}}}i}\times {d}_{{{\rm{l}}}i}+{\sum }_{j=1}^{m}{R}_{{{\rm{o}}}j}\times {d}_{{{\rm{o}}}j}}{{\sum }_{i=1}^{n}{R}_{{{\rm{l}}}i}+{\sum }_{j=1}^{m}{R}_{{{\rm{o}}}j}}$$
(2)

where:

$${R}_{{{\rm{l}}}i}=\left\{\begin{array}{c}{{rain}}_{{{\rm{l}}}i},\,{{rain}}_{{{\rm{l}}}i}\ge {rr}\\ 0,\,\hfill {{rain}}_{{{\rm{l}}}i} < {rr}\end{array}\right.$$
(3)
$${R}_{{{\rm{o}}}j}=\left\{\begin{array}{c}{{rain}}_{{{\rm{o}}}j},\,{{rain}}_{{{\rm{o}}}j}\ge {rr}\\ 0,\,\hfill {{rain}}_{{{\rm{o}}}j} < {rr}\end{array}\right.$$
(4)

where \(r\) is the TC rainfall radius threshold (e.g. 200, 300, 400, 500; unit: km), and \({rr}\) is the rainfall rate threshold (e.g. 1.5, 10, 20, 30, 40, 50; unit: mm 3 h−1); \(n\) is the count of terrestrial rainfall grids, and \(m\) is the count of oceanic rainfall grids; \({{rain}}_{{{\rm{l}}}i}\) is the rainfall rate (unit: mm 3 h−1) of the i-th terrestrial rainfall grid, and \({{rain}}_{{{\rm{o}}}j}\) is the rainfall rate (unit: mm 3 h−1) of the j-th oceanic rainfall grid; \({d}_{{{\rm{l}}}i}\) is the landward distance (≥0, unit: km) from the nearest coastline for the i-th terrestrial rainfall grid, and \({d}_{{{\rm{o}}}j}\) is the offshore distance (<0, unit: km) from the nearest coastline for the j-th oceanic rainfall grid (Supplementary Fig. 9).

\(\overline{{{D}_{{{\rm{land}}}}}_{{rr}}^{r}}\) (used in Fig. 1a, b, Supplementary Fig. 1, and Supplementary Table 1) is defined as the annual arithmetic mean of \({{D}_{{{\rm{land}}}}}_{{rr}}^{r}\), and \(\overline{{D}_{{rr}}^{r}}\) (used in Fig. 2a–c and Supplementary Fig. 2a, c, d) is defined as the annual arithmetic mean of \({D}_{{rr}}^{r}\).

Environmental variables

We consider four environmental variables related to TC rainfall: hourly relative vorticity, monthly specific humidity, monthly moisture flux, and SST. The first three variables are derived from ERA5 (0.25° × 0.25°)67, whereas SST is obtained from ERSSTv5 (2° × 2°)68, covering the period 1980–2023. The ERA5 hourly relative vorticity is used in a physics-based dynamic tracking method to identify TC rainfall field. Specifically, we compute 3-h means of the 850-hPa relative vorticity field and match each TC center position at the start of each 3-h interval to the corresponding mean vorticity field within a 500-km radius; grids with relative vorticity ≥1 × 10−5 s−1 in the NH and ≤−1 × 10−5 s−1 in the SH are classified as TC rainfall grids69.

To quantify the vertically integrated moisture transport in the lower troposphere, we compute the moisture flux (\({Q}_{u}\), \({Q}_{v}\)) over the 1000–700 hPa layer as follows:

$${Q}_{u}=\frac{1}{g}{\int }_{{\!\!\!\!p}_{{{\rm{upper}}}}}^{{p}_{{{\rm{lower}}}}}q\cdot u\,{dp}$$
(5)
$${Q}_{v}=\frac{1}{g}{\int }_{{\!\!\!\!p}_{{{\rm{upper}}}}}^{{p}_{{{\rm{lower}}}}}q\cdot v\,{dp}$$
(6)

where \(q\) is the specific humidity (unit: kg kg−1); \(u\) and \(v\) are the zonal and meridional wind components (unit: m s−1); \({p}_{{{\rm{lower}}}}\) and \({p}_{{{\rm{upper}}}}\) denote the lower (1000 hPa) and upper (700 hPa) pressure levels, respectively; and \(g\) is the gravitational acceleration (unit: m s−2). The moisture flux convergence (unit: kg m−2 s−1) is then calculated as the negative divergence of the moisture flux vector:

$${{\rm{Moisture}}}\; {{\rm{flux}}}\; {{\rm{convergence}}}=-\left(\frac{\partial {Q}_{u}}{\partial u}+\frac{\partial {Q}_{v}}{\partial v}\right)$$
(7)

Removal of climate factors

Following previous studies13, the influence of selected climate indices is controlled using a linear regression–based removal approach (shown in Fig. 2a–c and Supplementary Fig. 1 d). Let \(Y(n)\) denote the target time series in year \(n\) (\(n\) = l, 2,…, 44 for 1980–2023), and let \(X(n)\) denote the climate index time series, including ENSO, AMO and SST. For each climate index, a linear regression is independently performed as \(Y\left(n\right)=\beta X\left(n\right)+\varepsilon (n)\), where the \(\beta\) is the regression coefficient associated with selected climate index, and the \(\varepsilon (n)\) represents the residual time series.

CMIP6 data

To conduct an attribution analysis of changes in nearshore SST, we use outputs from 15 CMIP6 models70, including its subproject DAMIP71 (Fig. 2d–f). The model outputs analyzed are ACCESS-CM2, ACCESS-ESM1-5, BCC-CSM2-MR, CanESM5, CESM2, CNRM-CM6-1, E3SM-2-0, FGOALS-g3, GFDL-ESM4, GISS-E2-1-G, HadGEM3-GC31-LL, IPSL-CM6A-LR, MIROC6, MRI-ESM2-0, and NorESM2-LM72,73,74,75,76,77,78,79,80,81,82,83,84,85,86. These simulations cover the period from 1980 to 2014. Simulations from the historical all-forcing runs are compared with experiments from DAMIP, which are designed to isolate the effects of individual external forcings on historical climate changes. Specifically, DAMIP provides experiments under three distinct forcings: GHG, which includes only greenhouse gas forcings; AER, which includes only anthropogenic aerosol forcings; and NAT, which includes only natural forcings such as solar variability and volcanic aerosols.

WRF experiments

Experimental design and model configuration

We use the Advanced Research WRF (version 4.3) in its idealized configuration to explore the contributions of SST warming and coastal urbanization to TC rainfall distribution. A three-domain nested configuration is used with horizontal resolutions of 9, 3, and 1 km, all centered at the same location. Four scenarios are configured: CTRL with SST 300 K, CTRL with SST 302 K, URBAN with SST 300 K, and URBAN with SST 302 K. The CTRL and URBAN scenarios share the same modified land region, extending from the coastline (defined as 50 km) to 1500 km in the zonal direction (Supplementary Fig. 5a), but differ in urban extent and surface roughness. The CTRL scenario represents a low-intensity setting with narrower urban coverage and moderate surface roughness (baseline), while URBAN features a broader urban extent and higher roughness (developed), reflecting a more densely built environment. A cropland region is prescribed to the west of the urban area to represent non-urban land use. The spatial configurations within innermost domain (d03) are illustrated in Supplementary Fig. 5b, c.

A weak, axisymmetric vortex in gradient-wind and hydrostatic balance is imposed at the center of d03, with maximum tangential winds at the surface level87, and is allowed to develop into a TC over a homogeneous ocean surface with a prescribed SST of 300 K or 302 K6 (depending on the scenario) prior to land-surface modification (defined as 0 h). All scenarios are initialized with identical atmospheric (Jordan’s tropical sounding88) and oceanic conditions except for the prescribed SST. To keep TC intensity at Category 1–2 or weaker after the land-surface modification and minimize its influence on the results, we conduct eight sub-experiments for each scenario by modifying the land-surface setting at eight time points corresponding to maximum 10-m wind speeds (5-point averaged) of 48–55 m s⁻¹ (1 m s⁻¹ increments), and then run each experiment for 60 h. The observation interval is defined as the 48-h period (with a 3-h interval) from 12 to 60 h after the land-surface modification, with the first 12 h discarded to avoid adjustment noise. Model configurations and physical parameterizations89,90,91,92,93,94,95 are summarized in Supplementary Table 2.

Postprocessing and diagnostics

To quantify spatial shifts in TC rainfall, we compute both the synthetic distance (shown in Fig. 3a) and the landward distance (shown in Fig. 3c) of heavy rainfall (≥30 mm per 3 h) grids relative to the coastline in d03 for the four scenarios (the landward distance is evaluated within the 50–130 km zonal range), using 3-h model outputs within the observation interval. The calculation methods are similar to those used in Eqs. (1)–(4). The TC center is diagnosed at each output time (i.e. at the end of each 3-h interval) by locating the minimum sea level pressure. The meridional-mean profiles of horizontal wind convergence and zonal–vertical wind vectors (shown in Fig. 3b, d and Supplementary Fig. 5d–g) are constructed by averaging zonal–vertical slices sampled every 10 km in the meridional direction within the black rectangular box (meridional width: 100 km, fixed zonal range: −150 to 150 km31, always centered meridionally on the diagnosed TC center; see Supplementary Fig. 5b, c), at each 3-h interval within the observation interval. Horizontal wind convergence (unit: s−1) is computed similarly to Eq. (7), with positive values indicating convergence (inflow).

Land cover data

GLC_FCS30D96 is the first global dynamic land cover product at 30-m resolution, covering 1985–2022, with updates every five years before 2000 and annually thereafter. To assess the level of coastal urbanization, we compute the impervious area ratio based on the post-processed and optimized GLC_FCS30D (Version v2)97. To enhance the credibility of the results, we also use the CCI_LC dataset (1992–2022), which provides annual updates at 300-m resolution (v2.0.7cds for 1992–2015 and v2.1.1 for 2016 onward)98. The impervious area ratio is computed using Eq. (8) for each 0.1° × 0.1° latitude–longitude grid. A grid is classified as a city if its impervious area ratio is greater than or equal to a given threshold. To test robustness, we apply two thresholds: 0.240 and 0.3543, separately (Supplementary Fig. 6a, b). The annual average landward distance of city grids within the continental coastal zone (defined as within 200 km of the coastline), denoted as \(\overline{{D}_{{{\rm{city}}}}}\), is computed using Eq. (9).

$${{{\rm{I}}}{{\rm{mpervious}}}\; {{\rm{area}}}\; {{\rm{ratio}}}}_{{{\rm{grid}}}}=\frac{{N}_{{{\rm{impervious}}}\; {{\rm{pixels}}}}}{{N}_{{{\rm{all}}}\; {{\rm{pi}}}{{\rm{x}}}{{\rm{el}}}{{\rm{s}}}}}$$
(8)
$$\overline{{D}_{{{\rm{city}}}}}=\frac{\sum {{\rm{I}}}{{\rm{mpervious}}}\; {{\rm{area}}}\; {{\rm{ratio}}}_{{{\rm{city}}}\; {{\rm{grid}}}} \times {d}_{{{\rm{city}}}\; {{\rm{grid}}}}}{\sum {{\rm{Impervious}}}\; {{\rm{area}}}\; {{\rm{ratio}}}_{{{\rm{city}}}\; {{\rm{grid}}}}}$$
(9)

where the subscript \({{\rm{grid}}}\) denotes each 0.1° × 0.1° latitude–longitude grid; \({N}_{{{\rm{impervious\; p}}}{{\rm{i}}}{{\rm{xels}}}}\) denotes the count of impervious pixels in each grid; \({N}_{{{\rm{all\; pixel}}}{{\rm{s}}}}\) denotes the count of all pixels in each grid; \({d}_{{{\rm{city\; gri}}}{{\rm{d}}}}\) denotes the distance from the nearest coastline to the center of a city grid; and \({{{\rm{I}}}{{\rm{mpervious}}}\; {{\rm{area}}}\; {{\rm{ratio}}}}_{{{\rm{city}}}\; {{\rm{grid}}}}\) denotes the impervious area ratio of the city grid.

Population data

We use the high-resolution LSG dataset65 to represent human population distribution and dynamics. This dataset provides annual population estimates at a spatial resolution of 30 arcsec degree, covering the period from 2000 to 2024. The consistent temporal coverage and fine spatial granularity allow us to identify grids undergoing rapid urbanization over this timeframe. We integrate these population estimates with data on TC‑induced heavy rainfall and flood extents at the grid cell level to quantify population exposure to combined rainfall and flood hazards.

Exposure to TC heavy rainfall

To comprehensively estimate exposure to TC heavy rainfall, we compute the exposure metric in each 0.1° × 0.1° latitude–longitude grid within the exposed area from 2000 to 2023 using Eq. (10), focusing on the WNA, BOB, and WNP basins (Supplementary Fig. 7). This metric combines the annual frequency of TC heavy rainfall (from MSWEP), defined using a 500-km TC rainfall-radius threshold, with the corresponding population amount (from LSG). TC centers are constrained to the coastal domain shown in Fig. 1d–f.

$${{{\rm{E}}}{{\rm{xposure}}}\,{{\rm{to}}}\,{{\rm{rainfall}}}}_{{{\rm{grid}}},\,{year}}={{{\rm{population}}}}_{{{\rm{grid}}},\,{year}}\times {f}_{{{\rm{grid}}},\,{year}}^{{{\rm{rain}}}{{\rm{fall}}}}$$
(10)

where \({{{\rm{population}}}}_{{{\rm{grid}}},{year}}\) is assumed to be constant for each exposed grid within the year; \({f}_{{{\rm{grid}}},{year}}^{{{\rm{rain}}}{{\rm{fall}}}}\) is the annual frequency of TC heavy rainfall (≥30 mm per 3 h) in each exposed grid.

Exposure to TC-induced flooding

We compile recorded TC-induced flood events from 2000 to 2021 using the Dartmouth Flood Observatory (DFO) catalogue99, which aggregates information from news reports, government records, instruments, and remote sensing sources, focusing on the WNA, BOB, and WNP basins. The DFO dataset is widely used in global flood studies for its lack of temporal bias59. To isolate events directly attributable to TC rainfall, we filter the DFO records based on the reported cause of flood, retaining only those explicitly attributed to tropical cyclones, tropical storms, hurricanes, or equivalent terms for the three NH basins. To ensure a spatiotemporal association with TC rainfall, we match each flood event in the DFO dataset to best-track TCs with overlapping timestamps and compute the minimum distance from the event to the corresponding TC center (TC centers are constrained to the coastal domain shown in Fig. 1d–f). Event–TC pairs with distances greater than 500 km are excluded. To fill the data gaps in flood spatial distribution from 2019 to 2021, we use the Moderate Resolution Imaging Spectroradiometer imagery and adopt the water body identification method consistent with Tellman et al.59 for data reconstruction. We filter flood pixels (at a spatial resolution of 250 m) based on the matched TC impact areas, retaining only the flood pixels within 500 km of the TC center. For the period from 2000 to 2021, we compute the exposure to TC-induced floods in each 30 arcsec × 30 arcsec latitude–longitude grid within the exposed area using Eq. (11) (Supplementary Fig. 8).

$${{{\rm{Exposure}}}\; {{\rm{to}}}\; {{\rm{flood}}}}_{{{\rm{grid}}},\,{year}}={{{\rm{population}}}}_{{{\rm{grid}}},\,{year}}\times {f}_{{{\rm{grid}}},\,{year}}^{{{\rm{flood}}}}$$
(11)

where \({{{\rm{population}}}}_{{{\rm{grid}}},\,{year}}\) is assumed to be constant for each exposed grid within the year, and \({f}_{{{\rm{grid}}},{year}}^{{{\rm{flood}}}}\) is the annual frequency of flood occurrence in each exposed grid.

Statistical significance

Linear trends in Figs. 1a, b and 2; Supplementary Figs. 1–4, 6a, b, 7d–f and Supplementary Table 1 are estimated using simple linear regression. For each data series, the slope is computed using the scipy.stats.linregress() function in Python, which performs ordinary least squares regression to fit a linear model. The statistical significance of trends is evaluated using a non-parametric Mann–Kendall test, with P < 0.05 considered statistically significant. To visualize trend uncertainty in line plots, 95% confidence intervals for the slope are computed and displayed as shaded regions around the regression lines. One-sided paired-sample t-test is used to assess statistical significance in Fig. 3a, c (P < 0.05), testing whether the mean paired difference is greater than zero for each comparison.