Typhoon Dujuan rainfall near Tokyo in September 2026 likely intensified by human-driven climate change
Contact Authors
Davide Faranda, IPSL-CNRS, France, davide.faranda@lsce.ipsl.fr
Haosu Tang, University of Sheffield, UK, haosu.tang@sheffield.ac.uk
Marco Zanchi, CNRS, France, marco.zanchi@lsce.ipsl.fr
Neven S. Fučkar, University of Oxford, UK, neven.fuckar@ouce.ox.ac.uk
Suzana Camargo, Columbia University, USA, suzana.camargo@columbia.edu
Citation
Faranda, D., Tang, H., Zanchi, M., Fučkar, N. S., & Camargo, S. J. (2026). Typhoon Dujuan rainfall near Tokyo in September 2026 likely intensified by human-driven climate change. ClimaMeter, Institut Pierre Simon Laplace, CNRS. https://doi.org/10.5281/zenodo.23020225
Press Summary
Weather situations like Typhoon Dujuan now bring about 5-15% heavier rainfall over the Tokyo region than they would have in the past (1950-1987); the winds show no statistically significant change. When similar systems form today, the rainfall they produce is heavier than it would be in a world without climate change.
This heavier rainfall is accompanied by a warmer, more moisture-laden atmosphere over the western Pacific. The most intense rainfall was focused on the Kanto region around Tokyo, Yokohama and Chiba, and the adjacent Pacific coast of Honshu.
The event was linked to a rare weather pattern, a typhoon tracking north along the Pacific coast of Japan and making landfall on the Boso Peninsula, just east of Tokyo.
Approximately 37 million people and 1.7 trillion USD (equivalent to about €1.6 trillion) in economic activity were exposed within the rainfall-hazard area. Of the resulting rain-linked damages, an estimated 6% is attributable to human-driven climate change.
We attribute the intensification of the rainfall mainly to human-driven climate change, with natural variability playing only a secondary role.
Figure 1. ClimaMeter analysis of the August 2026 European drought (3-month window, June to August 2026). (a, b) circulation (Z500) and temperature anomalies; (c) precipitation during the event and (d) the three-month SPEI (brown = drier, green = wetter); (e, f) present-minus-past changes in circulation and temperature; (g) precipitation changes and (h) SPEI-3 changes; (i) the months in which similar past events occurred; (j) changes at the selected cities. The two dials summarise the attribution (left) and how rare the pattern is (right).
Typhoon Dujuan made landfall on the Boso Peninsula in Chiba Prefecture, just east of Tokyo, on 21 September 2026, after tracking north along the Pacific coast of Japan. More specifically, before reaching the Tokyo region, Dujuan was a well-developed tropical cyclone. On September 18, while south-southeast of Chichijima, Japan Meterological Agency (JMA) reported a central pressure around 975 hPa, and maximum sustained winds around 108 km/h. It brought violent winds, with gusts reported reaching near 180 km/h, and torrential rain to the greater Tokyo area and the Kanto region, prompting authorities to urge millions of residents to evacuate.
The storm left seven people dead and five missing after triggering landslides and flooding around Tokyo; a woman in her sixties was killed when mud flowed into her home in Yokosuka. Power to some 58,000 households in Chiba Prefecture was temporarily cut, and more than 460 houses were damaged. Rainfall was extreme in places: Oshima, an island about 120 km south of central Tokyo, recorded 832 mm of rain over 48 hours, more than double its average for the whole of September. The disruption forced the early closure of the Tokyo Game Show on its final day.
An important thermodynamic element for why the rainfall became so severe is the extraordinarily moist lower atmosphere. At the Tokyo JMA station during September 21, temperatures were around 21–22°C with dew points very close to the air temperature, and relative humidity near 99–100% during part of the storm. In other words, the near-surface atmosphere was essentially saturated. Tokyo was not actually experiencing the strongest part of Dujuan's winds, but it was exposed to an enormous amount of moisture and rainfall associated with the cyclone.This analysis focuses on a tight inner-core domain centred on Tokyo and the Boso Peninsula, so the attribution isolates the changes in the storm core rather than averaging them over the wider ocean. Following the ClimaMeter default for tropical cyclones, it uses the single landfall day (21 September 2026). The Circulation Anomalies panel (Fig. 1a), the anomaly of the sea-level pressure pattern, shows the deep low of the typhoon off the Pacific coast; the Precipitation Data (Fig. 1c) panel shows the heavy rain band over the Kanto region and the coast at landfall.
ClimaMeter attributes an event by comparing it with its closest historical analogues, that is, past days whose large-scale weather pattern most closely resembles the event. The analogues are split into a past period (1950 to 1987) and a present period (1988 to 2025); the difference between the two sets shows how weather situations like this one have changed as the climate has warmed. All fields are detrended and deseasonalised before the search, and the significance of the changes is assessed with a 1000-member bootstrap. To prevent several consecutive days of the same episode from being counted as separate analogues, the retained analogues are required to be at least seven days apart. The two dials summarise, on the left, how much of the change is due to human-driven climate change versus natural variability, and, on the right, how rare the meteorological pattern is in the historical record. Finally, we find that sources of natural climate variability, notably the Atlantic Multidecadal Oscillation, may have only partly influenced the event. This means that the changes we see in the event compared to the past may be mostly due to human-driven climate change. Similar Past Events (Fig. 1i) cluster in the typhoon season.
A warmer atmosphere holds more moisture (about 7% more per degree of warming), so the physically expected fingerprint of climate change on a tropical cyclone is heavier rainfall, which is the change the analogues below bring out most clearly.
The IPCC Sixth Assessment Report (AR6, Working Group I, Chapter 11) concludes that there is high confidence that, for a given tropical cyclone, human-driven warming increases the associated extreme rainfall, because a warmer atmosphere holds more moisture, and that the proportion of intense (Category 4 to 5) tropical cyclones and the rate of rapid intensification have likely increased. Working Group I, Chapter 12, projects further increases in tropical-cyclone rainfall rates and peak intensities with additional warming.
Working Group II identifies the densely populated coasts of East Asia, including the Tokyo metropolitan area, as highly exposed to tropical-cyclone rainfall, wind and storm surge, with very large concentrated assets. For Dujuan, ClimaMeter finds a robust intensification of the rainfall over the storm core, while the winds show no significant analogue-based change; because the pattern is rare (but not unprecedented) in the 1950 to 2025 record, the analogue-based attribution carries medium-to-high confidence.
We analyse how weather situations similar to Typhoon Dujuan have changed between the past (1950 to 1987) and the present (1988 to 2025) over the Tokyo region and the adjacent Pacific, identifying analogues from the sea-level pressure pattern. Comparing present with past analogues, the Precipitation Changes (Fig. 1g) show that comparable systems now deliver about 5-15% heavier rainfall over the storm core, consistent with the moisture increase expected in a warmer atmosphere. The Windspeed Changes (Fig. 1h) are not statistically significant (about 0%), and the temperature change over the core is small. Finally, we find that sources of natural climate variability, notably the Atlantic Multidecadal Oscillation, may have only partly influenced the event. This means that the changes we see in the event compared to the past may be mostly due to human-driven climate change. Similar Past Events (Fig. 1i) cluster in the typhoon season.
To estimate the population and economic assets exposed to climate change-influenced conditions, we overlay the ClimaMeter rainfall hazard map onto gridded datasets of population density and economic activity using the methodology developed in Faranda et al. 2026 (ERL). The hazard map retains only the grid cells where present-day conditions during events like the studied one are significantly more extreme than in the historical baseline (1950 to 1987). Onto this area, we overlay gridded population and gridded gross domestic product both at a spatial resolution of about 50 km.
We then aggregate the population and economic value within the hazard area and categorize them into three severity classes. The categories are based on how unusual the local rainfall was during the event relative to the 1950-2025 record:
Moderate: falling between the local 98th and 99th percentiles (representing the most extreme one to two per cent of days);
Severe: falling between the 99th and 99.5th percentiles (the most extreme half to one per cent); and
Extreme: exceeding the 99.5th percentile (the rarest and most intense conditions).
Importantly, these figures measure exposure, the people and assets in harm’s way potentially impacted, not the realized losses.
About 37 million people and 1,742 billion USD of economic activity (roughly €1.6 trillion) are located within the rainfall-hazard area, out of a domain total of about 71 million people and 3.2 trillion USD. By severity class the exposed population splits into roughly 2.9 million (moderate), 15.8 million (severe) and 18.8 million (extreme), so the exposure is dominated by the most intense, extreme-severity rainfall over the Kanto plain around Tokyo, Yokohama and Chiba, one of the most densely populated and economically concentrated regions on Earth (Fig. S4). The wind hazard produced no statistically significant, spatially connected footprint for this event, so no wind exposure is reported.
The methodology for calculating the climate-linked damage share is adapted from Faranda et al. 2026 (QJRMS). In this case, each grid cell represents the geographical spread of the event. We first calculate how much more intense the hazard is today compared to historical analogues, defined as the present-minus-past difference of the analogue composites. Afterwards, we restrict this estimate to the hazard core, representing the cells where the event was most severe and where the change is statistically significant. Applying this relative intensification to the exposed population and assets splits the local impact into two components that sum to 100 per cent in every affected cell: (1) a baseline share, which the weather extreme would have caused anyway, and (2) a climate-change share, the part intensified by the human-driven warming. Because these results are fractions independent of absolute monetary losses, we report percentages only, while the values below are estimates.
We quantify the uncertainty on this attributable share from the analogue bootstrap. For every grid cell, the present-minus-past change (the delta) carries a distribution whose spread is given by the 1000-member bootstrap. We take the 10th and 90th percentiles of that per-cell delta distribution and recompute the climate-attributable damage for each, bracketing the central estimate. Because the low and high percentiles are applied to every cell at once, the resulting range (reported in brackets) is a conservative envelope of the uncertainty in the attributed hazard change rather than a strict confidence interval.
Aggregated over the exposed assets, about 6% (10th-90th percentile 3 to 9%) of the rain-linked damage in the affected area is attributable to human-driven climate change (Fig. S5). Crucially, this percentage represents the climate-driven fraction of realized losses, rather than a proportion of the total exposed assets. The wind hazard produced no significant footprint, so no wind-damage fraction is reported (Fig. S6).
Weather situations similar to Typhoon Dujuan now tend to deliver heavier rainfall over the Tokyo region than in the past (1950 to 1987), by about 5-15%, while the winds show no significant analogue-based change inland. About 37 million people and 1.7 trillion USD of economic activity lie within the rainfall-hazard area, and an estimated 6% (10th-90th percentile 3 to 9%) of the rain-linked damage is attributable to human-driven climate change. Finally, we find that sources of natural climate variability may have only partly influenced the event. This means that the changes we see in the event compared to the past may be mostly due to human-driven climate change.
NB1: The following output is specifically intended for scientists and contain details that are fully understandable only by reading the methodology described in https://www.climameter.org/peer-reviewed-research
Figure S1. Counterfactual analogues. Each small panel is one retained past-period (1950 to 1987) analogue day, plotted as a Z500 anomaly on a common diverging scale.
Figure S2. Factual analogues. As S1 but for the present period (1987 to 2025).
Figure S3. Hazard fields, changes and analogue diagnostics. Rows are Z500, 2 m temperature, precipitation and wind speed; columns are the event field, the past composite, the present composite, and the present-minus-past change (coloured only where statistically significant). Lower rows give analogue-quality and natural-variability index diagnostics.
S4. Rainfall exposure. The ClimaMeter rainfall-hazard area (grid cells where present-day rainfall during events like this one is significantly more extreme than in the past) is overlaid on gridded population and gross domestic product, both at about 50 km resolution, split into moderate (98th to 99th percentile), severe (99th to 99.5th) and extreme (above the 99.5th) classes.
S5. Rainfall vulnerability diagnostic. (a) hazard severity as the local quantile of the event rainfall; (b) baseline share of the local damage due to the rainfall itself; (c) the share added by human-driven climate change; the two sum to 100% in every affected cell. The panel (c) title gives the aggregated climate-driven fraction and its 10th-90th percentile range.
S6. Wind vulnerability diagnostic. As S5 but for the wind hazard; no significant wind footprint was isolated for this event.
S7. Role of natural variability. For each climate-variability index (ENSO, NAO, PDO, AMO and others), the bar is a p-value testing whether that index differs between past and present analogues; a bar below the 0.05 or 0.10 line flags an index whose typical phase has shifted between the two periods.