Advanced risk-based event attribution for heavy regional rainfall events

The study addresses how anthropogenic climate change alters the probability of occurrence of specific extreme rainfall events (risk-based event attribution) and complements assessments of how event severity changes (storyline attribution). While there is broad agreement on attributing temperature extremes to human influence, risk-based attribution of extreme precipitation remains challenging because local rainfall is controlled by moisture availability, atmospheric circulation, vertical stability, mesoscale systems, and complex topography that typical GCMs cannot adequately resolve. High-resolution regional models are required, yet previous risk-based EA studies seldom evaluated both background synoptic circulation and mesoscale systems together. This work aims to combine large-ensemble GCM simulations with high-resolution RCM downscaling to quantify anthropogenic influences on the likelihood of major regional heavy rainfall events in Japan (2017, 2018, and 1993), explicitly conditioning on background circulation to reduce dynamical uncertainty.

Prior attribution research has successfully linked temperature extremes to anthropogenic warming, but precipitation attribution shows less consensus due to complex regional dynamics. Previous risk-based EA for heavy rainfall often relied on GCMs with coarse resolution that poorly represent convective systems, and while some studies used RCMs, evaluations focused mainly on rainfall statistics without fully resolving synoptic patterns and mesoscale processes. Storyline EA using high-resolution RCMs has been effective for case studies (e.g., July 2018 Japan rainfall, tropical cyclones), but cannot address event probabilities as it conditions on event occurrence. Studies of tropical cyclone impacts reveal substantial anthropogenic effects on rainfall and storm characteristics; however, cyclone genesis rates and frequency changes remain uncertain. This paper builds on and integrates these approaches by employing paired large-ensemble GCM and high-resolution RCM simulations to connect regional rainfall processes, background circulation types, and anthropogenic warming effects in a risk-based framework.

• Modeling framework: Used paired large-ensemble simulations from MRI-AGCM3.2 (~60 km) and NHRCM (20 km) over East Asia as part of the d4PDF project. Constructed 100-member ensembles for historical (HIST) and counterfactual non-warming (NonW) climates, spanning 1981 to present (over 3000 years per ensemble for 1981–2010). The 20-km RCM resolves meso-α/β-scale systems (fronts, tropical depressions, squall lines) and Japanese island topography, though not meso-γ convective lines.
• Forcing and scenarios: HIST uses observed SST/sea ice (COBE-SST2), historical anthropogenic and natural forcings (RCP4.5 after 2006). NonW fixes GHGs, anthropogenic and volcanic aerosols at 1850 levels and removes observed warming trends from SST/sea ice while retaining natural forcings.
• Events examined: Three Japanese heavy rainfall episodes: Case2018 (Japan’s Inland Sea; late June–early July, long-duration stationary rainband with strong southerly moisture advection), Case2017 (West Kyushu; orographic enhancement on the windward side), and Case1993 (East Kyushu; typhoon-induced, two similar-track TCs making landfall two days apart).
• Observational/reanalysis datasets: Daily precipitation from AMeDAS (~1300 stations at ~17-km spacing) for 1981–present; JRA-55 reanalysis for geopotential heights and column-integrated water vapor flux; TC best tracks from RSMC Tokyo Typhoon Centre.
• Validation and diagnostics: Computed regression coefficients linking local heavy rainfall frequency (days >100 mm/day over defined regions) from NHRCM to large-scale fields from MRI-AGCM3.2 (850-hPa geopotential height, column moisture flux, TC density) using 3000-year samples (100 members × 30 years). Compared leading interannual circulation modes (Rossby wave patterns) with JRA55.
• Indices and thresholds: Defined regional domains for the three cases. For Case2018 used maximum 72-h rainfall (28 June–8 July); for Case2017 maximum daily rainfall (1–10 July); for Case1993 maximum daily rainfall (1–31 July). Due to RCM underestimation of very intense daily rainfall, used the 50-year return level from the 1981–2010 HIST simulations as the event threshold for each case to assess exceedance probabilities.
• Probability estimation: Constructed return period curves for HIST vs NonW ensembles and computed exceedance probabilities for case-like thresholds. Estimated sampling uncertainty via Monte Carlo bootstrap (5000 resamples) to produce 2nd–98th percentile confidence intervals.
• Background circulation conditioning: Analyzed synoptic anomalies (Z850 and moisture flux) from JRA55 and ensemble-mean MRI-AGCM3.2 for 2017 and 2018. Defined a meridional dipole index (difference in 850-hPa geopotential height between [10°–30°N, 130°–160°E) and [35°–50°N, 150°E–180°E)) as a circulation metric for Case2017, calculated for July to assess dynamical conditioning.
• Long-term frequency assessment: Mapped HIST–NonW differences in July heavy rainfall frequency (days >100 mm/day) over 1981–2010 using RCM outputs to contextualize regional tendencies.

• Model capability: The 60-km GCM reproduced background synoptic circulation patterns (e.g., zonal dipole, enhanced NPSH and associated moisture transport) and the 20-km RCM captured regional rainfall gradients linked to topography. Regression analyses confirmed physically consistent relationships between heavy rainfall frequencies and large-scale drivers: West Kyushu rainfall increases with active NPSH and southwesterly moisture convergence; East Kyushu rainfall correlates with approaching typhoons; Inland Sea rainfall relates to a western trough–eastern ridge dipole and southerly moisture flow.
• Long-term frequency differences (1981–2010, July): HIST–NonW differences show stronger increases on the west side of Kyushu due to thermodynamic moisture increases enhancing southwesterly convergence; differences are smaller and less detectable over East Kyushu (due to TC uncertainty) and the Inland Sea (due to diverse event types and variability).
• Event-specific risk-based attribution: • Case2018 (Japan’s Inland Sea, 72-h rainfall): The 2018-like threshold had a 20.7-year return period in HIST vs 68.0 years in NonW, corresponding to a probability increase from 1.47% (NonW; 1.22–1.71%) to 4.82% (HIST; 4.38–5.27%). • Case2017 (West Kyushu, daily rainfall): Return period changed from 53.5 years (NonW; 1.87%, 1.61–2.12%) to 36.0 years (HIST; 2.78%, 2.48–3.08%). • Case1993 (East Kyushu, typhoon-induced daily rainfall): No significant change in probability detectable between HIST and NonW.
• Background conditioning: Ensemble-mean circulation anomalies indicate 2017 and 2018 events were dynamically conditioned (enhanced NPSH and southerly moisture advection; zonal dipole), reducing dynamical uncertainty and enabling detection of anthropogenic thermodynamic contributions. TC-related risks remain uncertain due to natural variability, GCM TC reproducibility limitations, and aerosol influences on TC potential intensity.

The study demonstrates that reliable risk-based attribution for regional heavy rainfall requires explicit conditioning on background circulation states that favor event development. Thermodynamic increases in atmospheric moisture under warming generally elevate rainfall potential, but detection of changes in dynamic circulation remains difficult due to large natural variability. By using large ensembles, the authors isolate signals of favorable synoptic patterns (e.g., enhanced NPSH, zonal dipole) in ensemble-mean fields for 2017 and 2018, which constrain dynamics sufficiently to reveal increased event probabilities attributable to anthropogenic warming. In contrast, typhoon-induced events depend strongly on cyclone approach frequency and structure; current GCMs outside the RCM domain and aerosol-related effects introduce substantial uncertainty, precluding robust attribution of probability changes for Case1993. The results reconcile why long-term regional frequency maps may show weak signals in some areas (e.g., Inland Sea), yet event-specific, circulation-conditioned analyses can detect significant anthropogenic influence. Integrating storyline EA (process-level, conditioned magnitudes) with risk-based EA (probabilities) yields comprehensive assessments of extreme rainfall under climate change.

By pairing large-ensemble GCM simulations with high-resolution RCM downscaling and conditioning on synoptic circulation, the study attributes increased risks of two record-breaking heavy rainfall events in western Japan (2017 and 2018) to anthropogenic warming, while finding no detectable change for a typhoon-induced 1993 event. The approach highlights that successful risk-based EA for regional rainfall depends on reducing dynamical uncertainty via background conditioning and leveraging large ensembles to separate signals from variability. The work extends high-resolution RCM studies toward robust probabilistic attribution and underscores the value of combining storyline and risk-based frameworks. Future research should improve representation of mesoscale convective systems (e.g., finer than 20 km), enhance tropical cyclone simulation reliability, quantify aerosol–TC interactions, and develop circulation-based conditioning metrics applicable across regions to generalize attribution findings.

• Resolution constraints: The 20-km RCM cannot fully resolve meso-γ scale line-shaped convective systems and fine-scale orographic precipitation, potentially underestimating very intense daily rainfall (addressed by using model-based return thresholds).
• Tropical cyclone uncertainty: GCM limitations in simulating TC genesis, tracks, and structure outside the RCM boundary, combined with large natural variability, hinder detection of TC-related probability changes.
• Aerosol effects: Cooling by aerosols can reduce TC potential intensity and counteract greenhouse gas warming effects, adding uncertainty to TC-induced rainfall attribution.
• Detection challenges: Dynamic circulation changes are difficult to detect at the current warming stage due to large variability; long-term frequency maps may mask event-specific signals without circulation conditioning.
• Regional focus: Results are specific to western Japan cases and may not directly generalize without adapting conditioning metrics and validating in other regions.