Changes in Atlantic major hurricane frequency since the late-19th century

The study investigates whether there has been a century-scale change in the number of the most intense hurricanes (major hurricanes, Saffir-Simpson Categories 3–5) in the North Atlantic. While the Atlantic has long records of hurricane activity (HURDAT2 back to 1851), these records suffer from severe inhomogeneities due to evolving observing practices and capabilities, complicating trend detection. Recent decades show increased basin-wide hurricane and major hurricane counts, but these may reflect improved detection rather than true climate trends. The purpose of this work is to develop a homogenized record of basin-wide North Atlantic hurricane and major hurricane frequency since 1851 to reassess long-term changes and place recent increases in a century-scale context. This question is important because major hurricanes account for a disproportionate share of U.S. hurricane damages and because understanding long-term variability versus forced trends informs projections and risk management.

Theory and models generally project that, with warming, tropical cyclone peak intensity and the fraction reaching major hurricane status should increase, though projected changes in major hurricane frequency in individual basins like the North Atlantic are mixed. Homogenized satellite-era observations since the early 1980s show an increase in the fraction of major hurricanes among all tropical cyclones globally and in the Atlantic, and an increase in rapid intensification events since the 1980s. However, recent multidecadal changes likely include substantial contributions from internal climate variability (e.g., Atlantic Multidecadal Variability) and non-greenhouse gas forcing, especially aerosols, which can modulate Atlantic hurricane activity. Prior studies have identified undercounts in pre-satellite hurricane records and developed methods to estimate missing storms using ship track density and other proxies, indicating that recorded increases in basin-wide hurricane counts since the late 19th century are consistent with observing practice changes. Differences in basin-wide versus U.S.-strike statistics and concerns about track shifts and landfall proxies have also been noted in the literature.

The authors extend and adapt an observing system emulation to homogenize basin-wide North Atlantic hurricane (HU, Vmax ≥ 33 m/s) and major hurricane (MH, Vmax ≥ 50 m/s) counts from 1851 to 2019. The core approach uses satellite-era storm tracks (assumed fully sampled; primary period 1972–2019) combined with pre-satellite-era ship track density from ICOADS (1851–1971) to probabilistically estimate the number of storms likely missed each pre-satellite year. For each satellite-era storm, an ensemble is constructed by shifting storm occurrence dates (0 to ±50 days in 5-day increments; 21 shifts) and sampling storm wind radii from parameterized distributions (100 realizations), yielding 2100 realizations per pre-satellite year per storm. Detection criteria: a storm is deemed detected if either (a) one land observation lies within the modeled hurricane-wind radius (R33) or (b) at least two ship observations lie within the tropical storm-wind radius (R17) with at least one within R33. For MH detection, a single land or ship observation within the 50 m/s wind radius (R50) suffices, with no latitude restriction. For HU detection, first detection must be equatorward of 40°N. Radii parameterizations are based on observational studies: R17 and R33 follow lognormal formulations tied to Vmax and scaled to mean extents, and R50 is derived from HWIND wind radii (1998–2013), fit with a lognormal distribution for MH radii. The expected number of missed storms per pre-satellite year is the sum over satellite-era storms of (1 − detection probability) for that year. Uncertainty is quantified via bootstrap resampling: for each pre-satellite year, 10,000 bootstrap samples are drawn from the combined ensemble (2100 realizations × 48 satellite years). The estimated missing counts are added to recorded HURDAT2 counts to form adjusted series for HU and MH. Temporal smoothing (15-year centered running means) is used for low-frequency assessment, with sensitivity checks to 9–25 year windows. To quantify secular trends, Poisson regression models with a log link are fit to annual counts (USA strikes and basin-wide HU/MH), with time as a covariate (years/100). For ratio metrics (e.g., USA strikes as a fraction of basin-wide counts, and MH/HU ratio), Binomial regression with a logistic link and time as a covariate is used. Trends are reported as the time-dependent coefficients and associated p-values. Multiple start dates are analyzed (1851, 1878, 1900, and for context, 1980) to account for key observing inhomogeneities (pre-1878 pre-Signal Corps era; 1878–1900 transitional period; satellite-era intensity records since 1980). Key assumptions include: perfect detection by ships and land points; representativeness of ICOADS ship tracks; inclusion of all detectable storms in HURDAT2; radial symmetry of storms; accurate wind measurements to threshold; ships not avoiding storms; representativeness of modern storm tracks for pre-satellite eras; sufficient ancillary information to identify storms upon threshold wind observation; and accurate event counting in HURDAT2. Sensitivity tests, including alternative fully sampled start (1966) and leave-one-out analyses of satellite years, indicate robustness of the adjustments.

• Recorded HURDAT2 shows strong century-scale increases in basin-wide hurricane and major hurricane counts, with about a threefold increase in recorded major hurricanes from the mid-19th century to recent decades. In contrast, U.S. strike counts do not exhibit such increases and may even show nominal decreases.
• After adjusting for likely missing storms, the apparent long-term increases in basin-wide hurricane and major hurricane counts largely disappear for 1878–2019. The adjusted series exhibit substantial multidecadal variability with peaks in the late 19th, mid-20th, and early 21st centuries, and minima in the early 20th and especially the 1960s–1980s.
• The increases in basin-wide hurricane and major hurricane activity since the 1970s are not a continuation of a century-scale trend but represent a rebound from a deep minimum in the 1960s–1980s.
• The century-scale decrease in the fraction of basin-wide storms striking the USA seen in raw HURDAT2 disappears after adjustment; the USA-strike fraction for major hurricanes remains roughly stationary (about 20–30% over 15-year periods) since the mid-19th century.
• The basin-wide MH/HU ratio in raw HURDAT2 increases over the century, but after adjustment it shows no significant long-term trend for 1878–2019, instead exhibiting multidecadal fluctuations: minima around 25–30% in the mid-1850s and around the 1980s, and maxima around 40–50% in the early-to-mid 20th century and early 21st century.
• Poisson and Binomial regressions: Raw HURDAT2 basin-wide major hurricane counts show significant positive trends (p < 0.01), whereas adjusted major hurricane trends are not significant after 1878 and only marginally significant over 1851–2019. Adjusted basin-wide hurricane counts show positive trends, but the post-1980 increases reflect recovery from prior minima rather than secular change.
• The adjustments indicate roughly three hurricanes per year were missed in the 1860–1880 period and around one major hurricane per year missed through much of the pre-satellite era, with larger undercounting during the World Wars due to sparse ship reports.
• The adjusted records align qualitatively with documentary and proxy reconstructions indicating that the late 20th century was an unusually inactive period for Atlantic hurricane activity.

The central research question—whether the most intense North Atlantic hurricanes have increased over the past century—is addressed by removing observing system biases from historical records. Findings indicate that recorded century-scale increases in basin-wide hurricane and major hurricane frequencies and the MH/HU ratio are largely artifacts of changing observing practices. After homogenization, there is no significant century-scale increase in major hurricane frequency or MH/HU ratio over 1878–2019. The recent (post-1980) rise in major hurricane activity and MH/HU ratio is better interpreted as a rebound from a marked mid-to-late 20th century minimum. The results reconcile discrepancies between basin-wide counts and U.S. strike counts by demonstrating that improved detection inflated basin-wide records relative to more consistently monitored U.S. strikes. These outcomes are significant for attribution and projection: they suggest that internal multidecadal variability and aerosol forcing likely exerted strong influences on 20th century Atlantic major hurricane activity, potentially masking greenhouse-gas-induced intensification signals in historical records. While theory and projections point toward increased intensity in a warming climate, the adjusted historical data do not show a clear century-scale intensification trend, emphasizing caution when attributing recent increases solely to greenhouse warming. The adjusted datasets provide critical targets for evaluating high-resolution dynamical and statistical models and for disentangling the roles of internal variability, aerosols (anthropogenic, dust, volcanic), and greenhouse gases in past and future hurricane activity. A nominal decrease in the fraction of basin-wide hurricanes striking the USA as hurricanes may hint at track shifts projected under warming, but trends for major hurricane strikes are not significant.

The study develops a homogenized record of North Atlantic basin-wide hurricane and major hurricane frequencies from 1851–2019 by emulating historical observing capabilities using satellite-era storm tracks and ship track densities. After adjustment, century-scale increases in major hurricane frequency and the MH/HU ratio evident in raw records are not significant over 1878–2019, and recent increases since the 1970s reflect recovery from a mid-to-late 20th century minimum rather than a secular rise. These results underscore that observing system changes strongly influenced recorded trends and that internal variability and aerosol forcing likely masked greenhouse-gas-related intensification signals in the historical period. Future research should: (1) use the homogenized records as benchmarks for century-scale simulations with high-resolution models; (2) quantify the relative contributions of internal multidecadal variability versus different aerosol species (sulfate, dust, volcanic) to 20th century changes; (3) refine detection emulation by incorporating non-ship observational sources and improved intensity proxies; and (4) assess potential shifts in storm tracks and landfall fractions under continued greenhouse warming.

Key limitations stem from methodological assumptions and data inhomogeneities: (1) assuming ships and land points are perfect detectors likely underestimates the number of missed storms, particularly in the 1800s; (2) ICOADS ship tracks may not fully represent all observations, potentially biasing adjustments; (3) reliance on HURDAT2 completeness can bias adjustments low if detectable storms were omitted; (4) assuming radial symmetry of wind fields introduces random errors; (5) assuming accurate historical wind estimates at detection thresholds may bias counts if measurements were systematically off; (6) assuming ships did not avoid storms underestimates missed counts; (7) using modern storm tracks to represent pre-satellite climatology may damp real historical variations; (8) identification upon threshold wind observation presumes sufficient ancillary information; (9) potential misclassification or splitting/merging of events in HURDAT2 could bias counts. Uncertainties are larger for individual years and especially pre-1878 and 1878–1900 transitional periods. Some residual inhomogeneities may remain, as suggested by marginal trends over 1851–2019.