Introduction

As the scientific evidence linking anthropogenic warming to increases in extreme weather events mounts, there is considerable interest in how severe convective storms may respond. The most violent types of convective storms are those capable of producing tornadoes, which garner particular attention due to their destructive and unpredictable nature. However, many challenges exist when examining tornadic responses to climate change. Aside from the many knowledge gaps on tornadoes, they also include the inhomogeneity in reporting and damage assessments, spatially and temporally sparse radiosonde data, and incomplete or limited archives of historical events. These problems are further exacerbated when examined globally due to the lack of uniformity in the standards of tornado reporting between countries. For Japan, in addition to the lack of dense observational stations, sounding reports are only available (aside from a few sporadic reports) at 0900 and 2100 Japan Standard Time (JST; 0000 and 1200 UTC). This potentially limits the number of events commonly considered environments associated with tornadoes (i.e., within 3 h), typically occurring between 1100 JST and 1800 JST (Supplementary Fig. 1a).

To alleviate the problems posed by non-homogeneous and scarce radiosonde data, pseudo-proximity (grid point) soundings are commonly used in their place. Gridded soundings can provide better insight into severe convective environments, including tornadoes, as they offer higher spatiotemporal resolution than observations. While the majority of studies utilizing pseudo-proximity soundings for convective environments are focused in the United States1,2,3,4,5, the quantitative understanding obtained from those studies has led to their use, primarily by global reanalysis, in examining tornadic and other severe convective environments throughout the world6,7,8,9,10,11.

As we develop a deeper understanding of severe convective environments, studies on their climatology, recent trends, and anticipated changes in response to a warming climate have been performed. In terms of recent trends in severe convective environments, the general understanding is that as the climate has warmed, there is greater potential for atmospheric instability and convective potential. By the same notion, deep layer vertical wind shear, an important component for storm organization, has weakened in the mid-latitudes due to decreased meridional temperature gradient12. However, changes vary significantly both seasonally and regionally due to complex differences in regional and atmospheric characteristics13,14,15, so it is important to examine how severe weather events, in the case of this study, tornadic environments, change on a regional level. Studies examining future changes commonly use relatively coarse resolution models, acknowledging that while convective thunderstorms may not be explicitly resolved, which encounter limitations due to high computational demand, an implicit understanding of future changes in convective environments can be achieved while allowing us to evaluate their uncertainties16,17,18,19,20,21. To date, only Muramatsu et al. 22 have investigated future changes in tornado environments over Japan. However, their results were based only on one model ensemble for the end-of-century climate, and indices for tornado environments were determined based on tornadoes in the United States at coarser time increments, which may not represent tornadoes over Japan

This study aims to expand upon previous studies to provide a comprehensive perspective of tornadic events in Japan using environmental proxies. We first examine the ability of several commonly used convective parameters that govern severe weather development through reanalysis datasets. These include common measures for strong convective potential [convective available potential energy (CAPE), convective inhibition (CIN), and the K-index (KI)] and the bulk wind difference (BWD). In addition, the lifting condensation level (LCL) and storm-relative helicity (SRH), which represents low-level moisture and updraft rotation indicative of tornadic potential, are also investigated. Multivariate parameters that combine these indices [Energy Helicity Index (EHI), K-Helicity index (KHI), and the Significant Tornado Parameter (STP)] are then examined for a clearer understanding of environments that distinguish between weak and significant tornadoes. Lastly, we evaluate how their frequency has and may change in the near future using reanalysis and a high-resolution large ensemble regional climate experiment.

Results

Identifying tornadic environments and their recent trends

Based on 469 tornadic events (67 F2+ on the Fujita scale, 215 F1, and 187 F0; Fig. 1), Fig. 2 shows boxplots of each parameter as a function of tornado intensity using the ERA5m. Consistent with previous studies, stronger tornadoes exhibit higher instability (CAPE and KI), low-level moisture (LCL), and stronger SRH. The slightly stronger CIN for F2+ tornadoes may indicate that stronger buoyant updrafts can overcome moderate levels of CIN or assist in capping weak convective initiation until environments are more favorable or built up for intense convection. The BWD does not differ significantly between intensities, consistent with Thompson et al. 1 and others, indicating that other environmental contributions distinguish tornadic intensity once BWD is sufficiently strong for convective development. The EHI, KHI, and STP indices discriminate between tornadic intensities reasonably well, though there are considerable overlaps in distribution. While there are limitations to these indices, as evidenced by the overlaps, we determine that values of 0.7 (EHI), 1.3 (KHI), and 1 (STP) are reasonable indicators of F2+ environments for Japan using the ERA5m. We note that median EHI and STP values are roughly two to three times lower than F2+ environments in the United States1,3. This is due to higher CAPE environments over the United States, where median values for F2+ in Thompson et al. 1 are more than twice what is shown in Fig. 2. This also likely explains the lack of recorded F4+ tornadoes over Japan. Additionally, thresholds for EHI, KHI, and STP using the ERA5p are 0.5, 1, and 0.5, respectively (Fig. 2 and Supplementary Fig. 2). This is needed for our climate change analysis, as the simulations for this study are also on pressure-level coordinates.

Fig. 1: Tornado reports from 1979 – 2021 used in this study.
figure 1

Red triangles indicate F2+ tornadoes and blue triangles indicate F0 and F1 tornadoes. Black boxes indicate domains used for trend analysis.

Full size image
Fig. 2: Box and whisker plots from pseudo-proximity soundings from the ERA5m.
figure 2

a CAPE, (b) CIN, (c) LCL, (d) KI, (e) BWD, (f) SRH, (g) EHI, (h) KHI, and (i) STP in the proximity of tornadoes in Fig. 1, separated by the Fujita scale rating. The boxes extend to the 25th and 75th percentiles, and the horizontal line within the box represents the median. Whiskers represent the 10th and 90th percentiles. The dotted blue line represents the 25th and 75th percentiles, and the solid red line represents the median for F2+ tornadoes from the ERA5p.

Full size image

Long-term trends (1979–2021) for F2+ environments in several regions with a high frequency of F2+ events (Fig. 1) are shown in Table 1 and Fig. 3. As expected, areas further south exhibit more hours with F2+ environments, mainly due to the influence of higher atmospheric instability (CAPE and KI). Long-term trends are mixed between regions and tornadic indices. Significant trends (p-value < 0.05) are seen for EHI and STP in Kanto and EHI in Kyushu and Ryukyu. The KHI exhibits the lowest interannual variability magnitude, likely culminating in a muted trend signal. The overall increasing trends can largely be attributed to the increasing trends in CAPE and KI, as BWD and SRH show insignificant trends (Table 1). CIN and LCL have shown a gradual but significant increase in all regions due to decreasing low-level relative humidity23, which is unfavorable for tornadic environments. It is noted that in addition to increasing trends in instability, there are significantly increasing trends in tropical cyclone frequency (Supplementary Fig. 3), which accounts for roughly 20% of all tornado events in Japan9, which may have influenced some long-term trends of F2+ environments, particularly in Ryukyu, where significantly increasing trends in EHI is not in supported by trends in the individual parameters used to compute EHI. The lone region with no significant trends in tornado parameters is the Chubu-Kinki region, despite significant trends in CAPE and KI. Note, however, that the highest frequency of tornado environments has occurred in recent years.

Table 1 Forty-three year trend statistics (per decade) for each parameter.
Full size table
Fig. 3: Forty-three year (1979–2021) timeseries of F2+ tornado environments.
figure 3

Annual 3-h frequency using the ERA5m for EHI > 0.7, KHI > 1.3, and STP > 1 over the (a) Kanto, (b) Chubu-Kinki, (c) Kyushu, and (d) Ryukyu regions shown in Fig. 1. Dashed lines represent the 43-year trend using the Theil-Sen slope estimate.

Full size image

Model validation and future changes in tornadic environments

The HIST is compared with the ERA5p for thermodynamic and kinematic parameters, as the d4PDF provides model output at pressure levels. The HIST (Fig. 4) reproduces the spatial distribution of the ERA5p (Supplementary Fig. 3). Prominent features include stronger atmospheric instability (CAPE, KI) in the southern portion of our domain due to high lower tropospheric moisture. Higher values are also depicted on the Pacific Ocean side compared to the Sea of Japan side, likely for the same reason. For CIN, high values are seen over China and portions of the Korean Peninsula, typically generated from warming at high elevations and subsidence of dry air from continental westerlies. Similar characteristics to CIN are also seen for LCL. Differences between HIST and ERA5p CIN are most pronounced over water east of the landmass, which may indicate that warm low-level airmass off land that commonly produce high CIN over water are weaker in simulations. BWD follows the climatological location of the mid-latitude jet, which is highest near central Japan. Lastly, SRH is greater at higher latitudes and over land. This region frequently experiences the passage of extratropical cyclones24,25, which can produce strong near-surface winds. The resulting low-level shear induced by surface friction is likely depicted here11.

Fig. 4: Spatial distribution of convective parameters and their future changes.
figure 4

(left) HIST ensemble mean, and (right) absolute change between the 2-K ensemble mean and HIST ensemble mean for (a) 95th percentile CAPE, (b) 95th percentile CIN, (c) median LCL, (d) 95th percentile KI, (e) median BWD, and (f) median SRH. Hatching indicates agreement in the response across all 2-K members.

Full size image

For +2-K changes, we focus on areas where all ensembles agree on the sign of the change (Fig. 4). Atmospheric instability parameters exhibit an increase across the analysis domain, primarily due to a projected increase in low-level specific humidity (not shown). For CIN, there is an increase over Japan and areas to the west. The spatial distribution is comparable to Chen et al. 26, who examined end-of-century changes to CIN using the CCSM4, which is one of the ΔSST patterns for d4PDF. LCL exhibits minimal change in 2-K, with only the Kanto region and parts of China showing slightly lowering LCL heights consistent across ΔSST patterns. This lowering of LCL is due to a subtle increase in projected relative humidity (not shown), which, interestingly, differs from the observed trends. However, we note that while low-level relative humidity from the ERA5 also showed decreasing trends throughout much of Japan from 1979–2010, the HIST ensemble showed no significant trend (not shown), which may explain the differences between observed and climate simulations in some areas. For BWD, decreases in meridional temperature gradients due to warming temperatures to the north act to weaken mid-latitude vertical shear, as reported in previous studies21. Lastly, small negative differences in SRH are depicted over Japan, which may be attributed to a projected delay in the Baiu front onset27,28 or a decrease in cyclone passage in future climates25, acting to reduce low-level wind shear.

The spatial distribution of EHI, KHI, and STP frequencies is shown in Fig. 5. A higher frequency of F2+ tornado environments is depicted in the southern portion of mainland Japan, such as Kanto and Kyushu, consistent with high CAPE and KI. High EHI and STP values are exhibited near the coast, while the KHI is distributed uniformly over land. This is due to the sensitivity of surface elevation and the lifted parcel in computing CAPE compared to KI, which uses specific pressure levels. Over the Ryukyu region, very few F2+ events are depicted in KHI due to low SRH values. The +2-K responses are mostly consistent between tornado parameters in depicting northern Japan to have more hours with F2+ environments, with more pronounced changes in June–August (Supplementary Fig. 5). This also signals that increases in F2+ environments occur primarily during the first half of the peak observed tornado season (Supplementary Fig. 1b) and less so in the second half (September–November). Greater changes are exhibited on the Sea of Japan side, synonymous with areas with larger changes in atmospheric instability compared to the Pacific Ocean side. The clear exception is in the Kanto region for EHI, which also sees high magnitude change in CAPE comparable to the Sea of Japan side. These features are muted for STP because of the introduction of LCL and BWD, which are projected to decrease slightly in +2-K. In contrast, the KHI exhibits smaller changes compared to EHI and STP. This is due to the difference in projected change between CAPE and KI, as CAPE in some areas can double from HIST, while KI increases only by 1–2 K. Lastly, the spatial distribution of EHI, KHI, and STP separated by ΔSST patterns are shown in Supplementary Fig. 6, confirming the +2-K changes over Japan are in agreement that more frequent F2+ environments should be expected, regardless of the tornado parameter. The clear difference to the +2 K changes is near the Ryukyu islands, where the future ensemble does not show agreement across members in any tornado parameter (Fig. 5), especially in SON (Supplementary Fig. 5) and across ΔSST patterns (Supplementary Fig. 6). It is possible that projected decreases in tropical cyclone activity in future climates in this region, as highlighted in many studies25,29,30, and differences in the magnitude changes, may have contributed to the uncertainty in this region’s projected changes.

Fig. 5: Spatial distribution of F2+ tornado environment frequency and their future changes.
figure 5

a EHI > 0.5, (b) KHI > 1, and (c) STP > 0.5.

Full size image

Discussion

In this study, we investigate the climatology of F2+ tornadic environments in Japan, how they have changed over the last few decades, and how they may continue to change in response to global warming. The development of the high spatial and temporal resolution ERA5 reanalysis allows for a better understanding of tornadic events in Japan, which has yet to be possible with observational soundings or coarser resolution reanalysis. From this, we were able to confirm what has been concluded in many studies, which is that F2+ tornadoes can be discriminated from weak tornadoes based on commonly used thermodynamic and kinematic parameters, and much better using multivariate parameters.

Trends using multivariate parameters suggest several regions with statistically significant increasing trends in F2+ environments. While the long-term increasing trend can partly be attributed to long-term increases in atmospheric instability, which is strongly related to regional warming trends, it does not fully explain these recent features. It would be interesting to monitor these environmental characteristics to determine whether they can be attributed to climate change or are more likely a response to internal climate variability. Additionally, further insight may be gained by using the recently extended ERA5, which now dates to 1940 but was not openly available at the time of this study. Compared to HIST, the +2-K climate shows much of Japan with increasing F2+ environments. Northern regions of Japan are expected to see nearly a doubling of F2+ environments based on EHI, KHI, and STP (Fig. 5). Eastern and southern Japan near Kanto, Chubu-Kinki, and Kyushu regions also showed increases in most indices. However, we note that some ΔSST patterns showed similar or even slight decreases in frequency. The primary driver is increased atmospheric instability, which is closely linked to increases in near-surface temperatures.

The F2+ tornado environments are projected to increase in some of Japan’s most densely populated regions, some of which may already be observed, as evidenced by trends in atmospheric instability and some multivariate tornado parameters. Therefore, the general public should be increasingly conscious of the dangers of tornadoes, and appropriate preparation should be undertaken to mitigate their societal impacts. For the forecast and research community, tornado-related indices should be continually re-examined and refined as we develop a more robust understanding of tornadogenesis through enhancements in numerical simulations, observational and reanalysis datasets, and larger tornado observation sample sizes.

In future studies, it is desirable to evaluate climate models at convective permitting (<4 km) resolutions, preferably with model-level vertical coordinates, as the ERA5m exhibits clear advantages compared to ERA5p in the representing convective environments. Many have shown the benefits of convective permitting models by simulating convective thunderstorms explicitly or for greater spatial and temporal representation of tornado environments that implicit methods may fail to capture, particularly when determining if adequate triggering mechanisms exist to initial convection, justifying the heavy computation burden necessary to perform such simulations31,32,33,34,35. Additionally, the representation of F2+ environments is highly sensitive to the indices applied, as our study suggests. Muramatsu et al. 22 also examined changes in F2+ tornadic environments using a threshold of CAPE ≥ 500 J kg−1 and surface to 3 km SRH ≥ 150 m2 s−2, which was determined for tornadoes in the U.S., using 6-h reanalysis data. In their 1993–2003 JJA model results, F2+ frequency near Kanto is ~4.5, and ~11 for Kyushu (approximated from Fig. 7 in Muramatsu et al. 22), all lower than our JJA results using ERA5p (not shown). It is also important to note that regardless of which tornado indices are used, they serve as guidance for operational forecasters to determine the potential for tornadoes. As such, there are likely a decent number of missed forecasts and even more false alarms. Finally, as Tochimoto36 comprehensively reviews, tornadoes can develop under various synoptic and mesoscale environments. Therefore, further studies, preferably with much larger sample sizes, should be performed in the future. All these factors contribute to recognizing that future increases in environments favorable for strong tornadoes should not be assumed as more tornadoes are on the ground. Continued research is required in many aspects of tornado environments.

Methods

CAPE and KI quantify the potential for strong convective thunderstorm development. CIN suppresses moist convection development by producing negative buoyant energy in the lower troposphere. LCL heights signify abundant low-level moisture and the cloud base, which can be useful in distinguishing between violent and non-violent tornados37,38. BWD promotes convective organization, and SRH represents the potential of updraft rotation of convective storms. The formulas for each parameter are detailed in Supplementary Table 1.

Previous studies have shown that the multivariate relationship of several thermodynamic and kinematic indices better distinguishes tornado events from non-tornadic convection and tornado intensity. Therefore, we used three co-variate tornado parameters to examine environments favorable for F2+ tornadoes: the EHI2:

$$\text{EHI}=\frac{\text{CAPE}\times \text{SRH}}{\mathrm{160,000}}$$
(1)

the slightly modified KHI39:

$$\text{KHI}=\frac{{\text{KI}}^{2}\times \sqrt{\text{SRH}}}{\mathrm{20,000}}$$
(2)

And the STP1:

$$\text{STP}=\frac{\text{CAPE}}{1000\,\text{J}\,{\text{kg}}^{-1}}\times \frac{\text{SRH}}{100\,{\text{m}}^{2}{\text{s}}^{-2}}\times \frac{\text{BWD}}{20\,\text{m}\,{\text{s}}^{-1}}\times \frac{2000\,\text{m}-\text{LCL}}{1500\,\text{m}}$$
(3)

For each parameter, an environment where CIN > 125 J kg−1 (approximately the 90th percentile of F2+ tornadoes in Fig. 2) is omitted.

Tornado database

Damaging wind gust events from 1961 onward are archived by the Japan Meteorological Agency’s tornadoes and severe winds database. The dataset provides the type of event (tornado, downburst, gust front, funnel cloud, dust devil), intensity, time, location, and prevailing meteorological conditions such as a passage of a cold front or a presence of an upper-level pressure trough. For this study, we focus on tornadic events between 1979 and 2021. Any tornadic reported over water or without intensity, location, or occurrence time is omitted. To avoid duplicate reports, we assume all reports that occurred within 30 min and 0.3° to be from the same parent thunderstorm, and the highest intensity reports are used. Several changes have been implemented since the start of this archive, most notably to the rating system to estimate tornado intensities. The Japanese Enhanced Fujita scale was implemented in April 2016 to replace the Fujita scale based on differences in damage assessment between Japan and the United States. For simplicity, however, we assume these changes to be negligible.

Reanalysis

Pseudo-proximity sounding utilizes the European Centre for Medium-Range Weather Forecasts high-resolution reanalysis (ERA540). At the time of this study, the ERA5 provides the highest spatial and temporal resolution among reanalysis datasets available (~31 km grid spacing). For three-dimensional computations, we use hourly outputs at 29 pressure levels from 1000 to 100 hPa for ERA5p and 78 hybrid/sigma model levels from near-surface to ~100 hPa from the ERA5m. The ERA5m dataset has been shown to better reproduce convective and kinematic parameters from radiosondes observations than coarser resolution reanalysis41. Similar conclusions are seen when we compare ERA5m soundings to coarser resolution reanalyzes such as the Japanese 55-year reanalysis (JRA-5542) or pressure level data from the ERA5p43 (Supplementary Table 2 and Supplementary Fig. 7). This is particularly true for parameters requiring high vertical resolution at lower levels (CAPE, CIN, SRH). These results demonstrate that higher horizontal and vertical resolution in the ERA5m is better suited for convective environment evaluation.

To evaluate tornadic environments, we extract the maximum values for each parameter from the closest hourly reanalysis soundings and within a 0.75° grid box for each tornado report.

Ensemble climate simulations

Dynamically downscaled regional simulation from the database for Policy Decision-Making for Future Climate Change (d4PDF44,45) large ensemble dataset is used for historical and future climate simulations. Downscaling experiments from a 60-km atmospheric general circulation (AGCM) to 20 km were performed using the non-hydrostatic regional climate model (NHRCM46). Each ensemble simulation contains 60 years of historical (1951–2010; HIST) and a 2-K warming scenario. Fifty HIST AGCM ensemble members were created using different initial conditions, and small sea surface temperature (SST) perturbations were downscaled using the NHRCM. The 2-K scenario represents a climate that is 2-K warmer relative to pre-industrial levels, corresponding to a climate around 2040 under the Coupled Model Intercomparison Project phase 5 Representative Concentration Pathway 8.5 scenario. Fifty-four ensembles were created by adding SST warming patterns (ΔSST) from six CMIP5 simulations. Due to computational limitations in storing and computing environmental parameters with three-dimensional data, we examine 320 years for HIST (10 members × years 1979–2010) and 1800 years for 2-K (5 members × 6 ΔSST × 60 years). We used 6-h output at the same 29 pressure levels as the reanalysis.