Storyline attribution of human influence on a record-breaking spatially compounding flood-heat event

Abstract

Attribution of compound events informs preparedness for emerging hazards with disproportionate impacts. However, the task remains challenging because space-time interactions among extremes and uncertain dynamic changes are not satisfactorily addressed in the well-established attribution framework. For attributing the 2020 record-breaking spatially compounding flood-heat event in China, we conduct a storyline attribution analysis by designing simulation experiments via a weather forecast model, quantifying component-based attributable changes, and comparing with historical flow analogs. We quantify that given the large-scale circulation, anthropogenic influence to date has exacerbated the extreme Mei-yu rainfall in the mid-lower reaches of the Yangtze River during June–July 2020 by ~6.5% and warmed the co-occurring seasonal extreme heat in South China by ~1°C. Our projections show a further intensification of the compound event by the end of this century, with moderate emissions making the rainfall totals ~14% larger and the season ~2.1°C warmer in South China than the 2020 status.

Anthropogenic climate change intensified the record-breaking spatially compounding flood-heat event in monsoonal China.

INTRODUCTION

The summer of 2020 was marked by a number of disastrous weather and climate extremes to eastern China, with the mid-lower reaches of the Yangtze River (MLRYR) and South China (SC) suffering the most. A quasi-stationary monsoonal front hosted a parade of convective storms along the MLRYR during June and July that produced a huge amount of precipitation beyond the previous record by a wide margin (Fig. 1A and fig. S1A) and caused a death toll of more than 100 along with economic losses of about $18 billion (1). In tandem with the monsoonal flood, an unprecedented seasonal heat event griped SC, expressed as more than 30 extremely hot days (i.e., daily maximum temperature exceeds 35°C) over much of the region, posing considerable pressure on agriculture, infrastructure, energy, and health sectors (2) (Fig. 1B). Both events were listed among the top-10 impactful extreme events in 2020 in China. Compared to occurrences in isolation, their concurrence may amplify economic losses, raise disaster fatalities, cause multi-breadbasket harvest failures, and stretch the regional or even national emergency response system thin (e.g., simultaneous allocation of flood and drought relief resources).

Fig. 1. Anthropogenic intensification of the compound event.

(A and B) Patterns for observed precipitation and surface air temperature anomalies (relative to 1961–1990) during June–July 2020 in eastern China. (C and D) Ensemble mean differences in precipitation and surface air temperature during June–July 2020 in eastern China between A2020 and N2020. Black dots in (C) and (D) mark significant differences between the ensemble means at the 0.05 level based on a two-sample t test. Dashed blue and red rectangles display the geographic boundaries of the MLRYR and SC, respectively. (E) Violin plots for the kernel density distributions of the simulated changes in precipitation averaged in the MLRYR (in percentage; blue; left y axis) and co-occurring surface air temperatures averaged in SC (in °C; orange; right y axis) due to historical and future anthropogenic warming and wetting. The inserted box whisker plots in (E) present the 5th, 25th, median (white dot), 75th, and 95th percentiles of member-specific simulated changes. (F) Daily evolutions of the anthropogenically forced changes in area-weighted mean MLRYR precipitation (blue; left y axis) and SC temperature (orange; right y axis) during June–July 2020, with the curve and shading presenting the ensemble mean and the 5 to 95% range of the changes among individual ensemble members. The Pearson correlation coefficient (r) between the detrended ensemble mean daily-scale precipitation and temperature changes, alongside its P value (P), are also shown in the upper right.

The two extremes are physically interconnected via a meridionally distributed Mei-yu trough-Western Pacific Subtropical High dipole pattern intrinsic to the Asian summer monsoon system on an intra-seasonal timescale (3, 4). Given the physical linkage between the two, spatially compounding flood-heat events should be common to eastern China. The 2020 case, however, is exceptional because both components were far above the 95th percentile, which is a first over the past 60 years (see fig. S1). Considerable efforts were soon devoted to the drivers, mechanisms, and predictability for the 2020 monsoonal rainfall, from the perspective of natural climate variability of various scales (5–9). Considering the huge impacts in proportion to the event intensity, a more concerning yet still elusive question pertains to the role of anthropogenic climate change in the spatially compounding event.

After nearly 20 years of development (10), a well-established probabilistic attribution method has enabled quantification of anthropogenic influence on the likelihood of specific (or classes of) weather and climate extremes (11–15). However, it faces some special challenges when attributing the 2020-like compound events (16). For record-breaking events, observations of mere decades are too brief to obtain meaningful statistics (e.g., return period) to guide event definition (17, 18), thus introducing large uncertainties in the remaining attribution steps (19, 20). The issue is more prominent for low-likelihood compound events because factoring in intra-event dependence requires a substantially enlarged sample size for the statistic estimates (16, 21). Moreover, applying a univariate extreme-tailored attribution method to multivariate extreme events is never straightforward. It requires a systematic transformation of all steps, including event definition, model evaluation, bias correction, and quantification of likelihood changes (22, 23). Nonetheless, there is a general lack of well-accepted theories, metrics, and techniques to underpin such an intermethod transformation.

Attribution of the compound event also encounters similar obstacles prevalent to the simulation and attribution of monsoonal precipitation extremes, including insufficient model resolution for convective processes and limited model skills for complex interactions and configurations among circulation agents (24, 25). The dependence between variables and interacting processes in space add dimensionality and complexity to the dynamic aspects, further discounting the suitability of conventional tools and methods in attributing spatially compounding events (26, 27). Even trickier is the poorly understood response of regional atmospheric circulations to anthropogenic forcings, which has been identified as a major source of uncertainties in the probabilistic attribution and projection of dynamically driven extremes (28).

Here, we adopt a “storyline” perspective (29, 30) by applying the pseudo-global warming approach to examine human influence on the compound event. Specifically, through constraining the underlying large-scale atmospheric circulations, we answer to what extent the well-understood anthropogenic disturbances to the climate system’s thermodynamic aspects have altered the severity of the unprecedented monsoonal compound flood-heat event. We achieve this by (i) running a weather forecast model to create and compare large ensembles of factual and counterfactual simulations, (ii) dissecting the estimated overall human influence into component-based attributable changes to enhance attribution credibility, and (iii) performing an observation-based flow analog analysis as a cross-validation. In addition, we take advantage of a collection of seven single-model initial-condition large ensembles (31) to display uncertainties in the response of regional dynamics to anthropogenic climate change and discuss the implications on the attribution of univariate and compound extreme events. Although the attribution conclusions are case specific, our analysis opens a promising avenue for the attribution of dynamically complex, low-likelihood, and high-impact compound events.

RESULTS

Event-based storyline attribution

We perform a suite of high-resolution (15 km) simulations using the Weather Research and Forecasting [WRF; (32)] model with large-scale (100° to 130°E; 15° to 40°N) atmospheric dynamics (three-dimensional geopotential heights and winds) constrained and initial and boundary thermodynamic conditions modified [air temperature, specific humidity, sea surface temperature (SST), and greenhouse gas concentrations]. In short, although details are available in Methods, we start with an “actual conditions” hindcast experiment (referred to as “A2020” hereafter) that consists of 48-member simulations driven by realistic initial and boundary conditions. Under the assumption of near-constant relative humidity, we use the “pseudo-global warming” approach (33–35) to estimate anthropogenically induced changes in the climate system’s thermodynamic states (36, 37), which are either removed from the current climate to create a “naturalized conditions” ensemble (“N2020”) or added to the current climate based on certain shared socioeconomic pathways [SSPs; (38)]. The large-scale dynamics are constrained in the same manner (see below and Methods) in factual and counterfactual simulations without nudging or data assimilation applied.

Initialized by the observed weather patterns (from the fifth-generation European Centre for Medium-Range Weather Forecasts atmospheric reanalysis of the global climate) several days (3, 5, and 7 days) ahead of the 2020 monsoon season, the A2020 simulations well reproduce the regional–to–large-scale circulation anomalies, both horizontally (figure omitted) and vertically (fig. S2). The surface weather extremes in A2020 are in good agreement with observations in terms of their spatial patterns, magnitudes, and intraseasonal evolutions, albeit with an overestimate in precipitation intensity (fig. S3) due to stronger vertical velocity in the simulations (fig. S2). The consistency between the A2020 simulations and observations indicates the decisive role of the dynamics in the compound event. The four experiments share highly similar meteorological patterns (figure omitted), consistent with previous studies [e.g., (35–37)] that reported essentially unchanged large-scale dynamic patterns among pseudo-global warming experiments. The model verification confirms that all three counterfactual experiments produce dynamically driven events similar to those in A2020, thus allowing for a like-for-like comparison to quantify the effect of anthropogenic thermodynamic forcings.

By comparing A2020 and N2020, we find that anthropogenic climate change has intensified the 2020 Mei-yu rainfall (Fig. 1C) and warmed SC significantly during the same season (Fig. 1D). For the domain average (rectangles in Fig. 1, C and D), the 2020 Mei-yu rainfall totals increased by nearly 6.5% (5 to 95% range: 3.7 to 10.1%) due to historical human influence (Fig. 1E). Simultaneously, anthropogenic influence has led to a rise in SC-averaged temperature by around 1°C (0.9° to 1.1°C) in June–July 2020 (Fig. 1E), with the most remarkable warming found in the immediate south of heavy rainfall centers (Fig. 1D). Future warming and wetting (the 2090s versus the 2010s) are projected to further intensify the compound event, with the extreme Mei-yu rainfall amount experiencing extra increases by 6.4% (2.7 to 10%) and 14% (9.7 to 21.1%) under SSP1-2.6 and SSP2-4.5, respectively, and SC extreme temperatures further rising by 0.8°C (0.7° to 1.0°C) and 2.1°C (1.8° to 2.7°C) in tandem (Fig. 1E).

Through examining the daily evolution, it can be observed that the best estimates for attributable precipitation changes in the MLRYR and temperature changes in SC are generally positive (Fig. 1F), with daily attributable rainfall and temperature changes positively correlated (Pearson’s correlation coefficients around 0.5) (Fig. 1F). This suggests an anthropogenic intensification of the compound event occurring on a daily basis (i.e., heavier rainfall in the north and hotter weather in the south). Compared to the lower-warming scenarios (i.e., A2020 versus N2020 and SSP126 versus A2020), the intensity responses of both extremes under SSP2-4.5 exhibit markedly larger day-to-day variability and wider intramember spread, implying a nonlinear influence of climate change on these variables and potentially nonlinear intensification of the compound event as a result. For wet extremes, this may result from amplified local dynamical responses to increased latent heating in a wetter environment (39); while for hot extremes, the ever-increasing magnitude of temperature responses intraseasonally could suggest progressively weakened constraint from evaporative cooling that likely results from early-season intense evapotranspiration followed by accelerated soil moisture depletion (40, 41).

Composition of attributable changes

In addition to providing the overall quantitative statements on human influence, the storyline attribution further allows for quantification of component-based attributable changes. For the record Mei-yu rainfall, we quantify the influences of anthropogenic climate change on regional moisture budget. Theoretically, the precipitation change within the domain (P′) is mostly contributed by the forced changes in column-scale vertical (−〈ω∂_pq〉′) plus horizontal moisture advection (−〈V_h · ∇_hq〉′) and surface evapotranspiration (E′), given that changes in atmospheric moisture storage over the season (−∂_t〈q〉′) and the residual term δ are several orders of magnitude smaller (see Eq. 2 in Methods). Anthropogenic influence on the vertical moisture advection can be further decomposed into forced changes in moisture profile (pure thermodynamic term, −〈ω_ref∂_pq′〉), vertical velocity (pure dynamic term, −〈ω′∂_pq_ref〉), and the strength of their nonlinear interactions [−〈ω′∂_pq′〉; (42); see Eq. 3 in Methods]. Regarding the source of atmospheric moisture, anthropogenically caused warming of the neighboring oceans, especially in the SC Sea and the northeastern Indian Ocean, boosts the level of water vapor evaporated into the overlying atmosphere, which is readily advected to the target region by the East Asian summer monsoon and converges there (fig. S4). Anthropogenic intensification of the 2020 extreme Mei-yu rainfall is then largely mediated through surges in vertical moisture advection, in which stronger dynamic-thermodynamic interactions (−〈ω′∂_pq′〉) and disproportionate moistening of the lower troposphere (−〈ω_ref∂_pq′〉) play a primary and secondary role respectively (Fig. 2A and fig. S5, A and B). Here, the enhanced nonlinear interaction refers to the invigoration of ascending motions, particularly at the mid-upper troposphere due to more latent heat released within wetter storms (figs. S5B and S6A), dubbed as the “dynamic amplification” effect (39, 43, 44). The contribution from responses of local-to-regional dynamics (−〈ω′∂_pq_ref〉) is smaller (ensemble-mean) but uncertain in sign (error bars in Fig. 2A) and largely accounts for the intramember spread of the simulated precipitation responses (Fig. 2C), even with large-scale dynamics equally constrained among ensemble members. The component-based attribution also holds for projected responses of the extreme Mei-yu rainfall to future warming and wetting (Fig. 2, A and C, and fig. S5).

Fig. 2. Anthropogenic influence on key components to precipitation and temperature extremes.

(A) Simulated changes [forced signal as expressed by ensemble mean (shaded bars) and 5 to 95% range of individual ensemble member values (error bars)] in the area-weighted seasonal precipitation and atmospheric moisture budget terms in the MLRYR during June–July 2020 due to historical and end-of-century (SSP1-2.6 and SSP2-4.5) anthropogenic warming and wetting. (B) Simulated changes in the area-weighted land surface energy budget terms in SC during June–July 2020 due to historical and future anthropogenic warming and wetting. The meanings of the items labeled along the x axis in (A) and (B) are introduced in the main text. (C) Scatterplots showing the relationship between member-specific simulated changes in the dynamic term of vertical moisture advection (x axis) and precipitation in the MLRYR under differing scenarios. (D) Scatterplots between member-specific simulated changes in the area-weighted mean downward shortwave radiation (x axis) and surface air temperature in SC under differing scenarios. The slope of the linear regression equation (b), the Pearson correlation coefficient (r), and its P value (P) are indicated.

For the colocated record seasonal heat in SC, most anthropogenic contributions come from the background man-made warming (fig. S7A). At the local scale, descending branches of the secondary cell in between the two extremes also play an indispensable role in modulating the magnitude and spatial pattern for the air temperature response (fig. S7B). On one hand, it directly elevates near-surface air temperature through enhancing sinking motion and adiabatic heating (figs. S5C and S6A); on the other hand, it reduces cloud covers (fig. S5D) and increases downward shortwave radiation (SW_d) reaching the ground (Fig. 2B). With the upward reflected shortwave radiation (SW_u) changing little, the net surface radiation (R_n) is positive in response, which is then partitioned into land-surface sensible heat (H) directly warming the air atop and evapotranspiration latent heating (λE) (Fig. 2B). As for the longwave radiation, although warmer troposphere emits more downward energy to the ground (LW_d), its implication on surface air temperature is mostly offset by the enhanced radiative cooling (upward longwave radiation from surface, LW_u) caused by higher land skin temperatures. The intramember spread in temperature changes (T′) is sourced primarily from responses of downward shortwave radiation (${{\mathrm{SW}}_{\mathrm{d}}}^{\mathrm{\prime }}$) (Fig. 2D). Elucidating the human influence on enhanced descending motion atop SC, i.e., the source of incremental incoming shortwave radiation, is fundamental to understand the SC heat intensity in the compound event. The increased atmospheric stability owing to greater warming of upper tropospheric temperature and equivalent potential temperature seems the direct reason (fig. S6B). The anthropogenically strengthened dynamic amplification in the MLRYR is also at play by scaling up the low-level convergence locally and thus indirectly facilitating divergent outflows from SC (fig. S6A). Further comparison of human influence on adiabatic heating, land-air energy exchange, and background large-scale warming warrant better-designed factorial experiments (45, 46).

Observation-based flow analog attribution

As a cross-validation of the model-based attribution, we additionally perform an observation-based statistical attribution via a flow analog approach (47–50). In short, we search for a number of historical (since 1961) daily atmospheric circulation patterns that closely resemble the large-scale dynamics of each day in June–July 2020 by the criterion of Teweles-Wobus score [i.e., TW score; (51)] and filter the linear relation among the analogs best matching the target daily meteorological pattern. The historical detrended daily P_ano and T_ano (ano for anomalies) are then combined in the same way to achieve the approximations of dynamically-induced components during June–July 2020 (see Methods). The thermodynamic part is then obtained as a residual. Note that such an artificial separation of dynamic and thermodynamic components of the events acknowledges weak and uncertain responses of dynamic patterns (essentially horizontal pressure gradients) to thermodynamic forcings. To better match the assumption, we use the daily SLP anomaly field (SLP_ano) rather than geopotential heights to search for analogs. Given the satisfactory reconstruction of large-scale dynamic patterns (fig. S8), we estimate dynamically induced P_ano and T_ano around 201 mm (5 to 95% range: 166 to 238 mm) and 0.47°C (0.36° to 0.60°C) respectively, accounting for 72 and 35% in total anomalies (Fig. 3, A and B). The analog-constructed and observed daily P_ano and T_ano time series evolve in phase, with Pearson’s correlation coefficients of 0.74 and 0.70. Thermodynamic processes contributed ~77 mm (40 to 113 mm) in P_ano and 0.89°C (0.77° to 1.00°C) in T_ano accordingly (Fig. 3, A and B).

Fig. 3. Observation-based flow analog attribution.

(A and B) Observed (red curves), dynamically (blue histograms), and thermodynamically-induced (orange histograms) daily anomalies (relative to 1961–1990) of precipitation in the MLRYR and surface air temperature in SC during June–July 2020. The error bars enclose the 5 to 95% ranges for the statistical estimates for dynamic and thermodynamic contributions. The values shown in the parentheses indicate anomalies of seasonal precipitation totals and mean temperature in observations and the median of estimates for the dynamically and thermodynamically induced components. (C and D) Observed (red), dynamically (blue), and thermodynamically induced (orange) anomalies of June–July precipitation totals in the MLRYR and mean surface air temperature in SC over 1961–2020, with the blue and orange shadings representing the 5 to 95% ranges for the dynamically and thermodynamically induced components, respectively. The Pearson correlation coefficients (r) between observations and dynamic reconstructions and the linear trends of the thermodynamically induced components, along with corresponding P values (P) are indicated in each panel.

The thermodynamic residual should involve both anthropogenically forced change and naturally evolving parts, but the latter can hardly drive directional change over the span of six decades. Given this general expectation, we repeat the flow analog analysis with respect to the entire period 1961–2020 and reasonably approximate the forced thermodynamic changes in Mei-yu rainfall and SC air temperatures by linear trends for multidecadal thermodynamic residual series. We report that the thermodynamic forcings have intensified the extreme Mei-yu rainfall by ~50 mm (i.e., 8.4 mm decade⁻¹ × 6 decades) over 1961–2020, equivalent to ~8% of the observed rainfall totals of 613 mm in 2020, and in the meantime elevated the SC surface temperature by ~0.66°C (i.e., 0.11°C decade⁻¹ × 6 decades) during June–July 2020 (Fig. 3, C and D). We confirm that the statistical attribution is robust, both qualitatively and quantitatively, against the choices of reanalysis data (52), field similarity metrics, trend estimators, and periods for analog reconstruction (e.g., 2010–2019 versus 1961–2019). The quantitative differences between observation- and model-based attributions arise mainly from differing anthropogenic warming levels considered (since the 1960s versus the pre-industrial period) and are also linked to the inevitable inclusion of local physical processes related to anthropogenic warming and wetting (e.g., dynamic amplification of moist convection and soil moisture-temperature interactions) into the dynamically driven component by the flow analog approach.

Attribution uncertainties in large-scale dynamics

The use of the storyline method to pursue a conditional attribution answer (thermodynamics only but no overall risk) is based on the dynamic aspects of climate change not being consistent or robust in observations, theories, and models (28–30, 53). Despite the reported inconsistency, very few existing attribution studies have formally quantified the level of uncertainties in forced changes in large-scale dynamics. We use multimodel initial-condition large-ensemble simulations to distinguish systematic responses of atmospheric dynamics to warming, away from noisy internal variability. We select cooler and warmer periods to compare dynamics and the resulting extremes as a first-order approximation for their responses to warming. There are ample samples in the large ensembles showing reasonable resemblance to the 2020 compound event-associated large-scale circulation pattern (figs. S8, A and B and S9). However, the probability changes of the circulation pattern (Fig. 4A), and the strength and changes of their governance on surface extremes (expressed by the variation of regression coefficients in Fig. 4B) are heavily model dependent. Specifically, the CanESM2 and GFDL-CM3 show that anthropogenic forcings may have led to more frequent occurrences of the June–July 2020-like regional circulations (lower TW scores), whereas the CESM-CAM5 and EC-EARTH show the opposite (Fig. 4A). For the remaining models (i.e., CSIRO-Mk3-6-0, GFDL-ESM2M, and MPI-ESM), the probability distribution of TW scores keeps essentially unchanged, indicative of no systematic response of favorable dynamic setup to warming. These results qualitatively hold when using SLP data from alternative reanalysis (52) or extending the analysis to future warmer periods (e.g., 2071–2100).

Fig. 4. Uncertain dynamic aspects of anthropogenic climate change in the event attribution.

(A) Histograms (shaded and blank bars) and fitted normal distributions (curve) of June–July mean TW scores (%) of the simulated monthly sea level pressure anomaly patterns as compared to the 2020 pattern based on ERA5 during historical (1961–1990; blue) and current (2006–2035; red) periods. (B) Scatterplots between the June–July mean TW scores (%) of the simulated sea level pressure anomaly patterns with reference to the 2020 pattern (ERA5) and concurrent precipitation anomalies (in percentage, relative to 1961–1990 climatology) in the MLRYR during historical (blue) and current (red) periods. The slope of the linear regression (b), the Pearson correlation coefficient (r), and its P value (P) are indicated in each panel. The sample size of each model is indicated in the parenthesis.

In these models, the circulation patterns showing a closer resemblance to the 2020 situation tend to produce larger excess precipitation in the MLRYR (Fig. 4B) and higher temperatures in SC (fig. S10) simultaneously, with the dynamical control being stronger for precipitation. Therefore, based on the probabilistic attribution method, the models that simulate decreasing (increasing) resemblance of large-scale dynamics to the 2020 pattern would likely conclude reduced (increased) probability and/or intensity of the 2020-like Mei-yu heavy precipitation and set the tone for attribution of the compound event. Thus, the presented notable intramodel differences in the dynamic responses highlight the imperative of constraining this highly uncertain part in the attribution of dynamically driven univariate and compound extremes when our understanding regarding dynamic aspects of anthropogenic climate change remains sparse, unmatured, and elusive.

DISCUSSION

The attribution of precipitation extremes has proven difficult and is well-known for large uncertainties (11, 12, 15, 54, 55). A key strength of our conditional attribution method lies in its capability of filtering the forcing signal in the observed precipitation anomaly from strong internal variability and pinpointing the source of attributable changes and uncertainties. This merit sets the stage for the attribution of precipitation extreme–involved compound events. Using the method, our analysis presents quantitative evidence that anthropogenic warming and wetting to date have intensified the record-shattering spatially compounding flood-heat event over eastern China in 2020. Given the dynamic occurrence of the event in warmer worlds, the end-of-century climate change following a moderate emissions scenario would further worsen the compound event by aligning a ~14% wetter Mei-yu season with a ~2.1°C warmer June–July period in SC, on top of their exceptional statuses in 2020.

So, as highlighted by previous studies (28, 53) and reinforced here, the starting procedure of constraining large-scale dynamics in the storyline attribution does not neglect the crucial role of the dynamic aspects of anthropogenic climate change in extreme event attribution; rather, in the absence of confident understanding of changing dynamics, it is advisable to alternatively frame the question as that under nearly identical large-scale circulations, how and why the current compound event would have unfolded in different ways in previous cooler or future warmer climates (36, 37, 56–62).

On the basis of different event definitions, models [e.g., HadGEM3-GA6 and Coupled Model Intercomparison Project Phase 6 (CMIP6)], and conditioning degrees, several probabilistic attribution analyses were conducted with respect to the 2020-like Mei-yu rainfall and consistently reported a reduced likelihood (by ~50%) due to anthropogenic climate change (63–65). Revisiting anthropogenic dynamic changes in the models used in the 2020 event attribution, we find that the decreased probability of the underlying circulation patterns in response to anthropogenic forcings is to blame (fig. S11), albeit with the robust thermodynamic expectation of warmer air holding more moisture. However, as revealed by the analysis based on the multimodel large ensemble archive, the consistency between only two groups of model simulations does not necessarily give rise to a reliable attribution statement. Rather, alternative plausibility for responses of dynamics to climate change in other models, undersampling of internal variability in CMIP6 limited runs, and unrealistic representation of aerosol forcings (66) add uncertainties to these existing attribution analyses but remain insufficiently communicated.

Given the recurrent large-scale monsoonal circulation patterns such as the 2020 case and the expectation of continued warming of the planet, our storyline attribution and projection results underscore the necessity of preparing for more intense spatially compounding flood-heat hazards over eastern China, which could lead to increased economic damage and loss of life if not planned properly.

MATERIALS AND METHODS

Model configuration

We used the WRF [version 4.4.2; (32)] to perform the storyline attribution experiments. In contrast to global climate models that typically use the hydrostatic approximation to simplify the equation of atmospheric vertical motion, WRF is fully compressible and nonhydrostatic, so it is more skillful in simulating the development of atmospheric instability and convective processes (67). Considering the spatial scale of the target compound event, we configured WRF with one domain encompassing eastern China (100° to 130°E; 15° to 40°N) that consists of 200 and 180 grid cells in zonal and meridional directions with a horizontal spacing of 15 km. We partitioned the model atmosphere into 40 vertical levels with the top pressure at 50 hPa, which proved to be high enough to capture the compound event (figs. S2 and S3). The selection of the physics scheme suite followed a recommendation of the National Center for Atmospheric Research, including the Thompson cloud microphysics scheme, the Tiedtke cumulus parameterization scheme, the Mellor–Yamada–Janjic planetary boundary layer scheme, the Rapid Radiative Transfer Model for general circulation models, and the Noah land surface model. To test the sensitivity of convection simulations to physical schemes, we replaced these schemes one by one with commonly used alternatives (i.e., the Purdue Lin cloud microphysics scheme, the Kain-Fritsch cumulus parameterization scheme, the Yonsei University planetary boundary layer scheme, and the Noah-MP land surface model). Together, we had 16 different combinations for the selection of the physics scheme suite. The simulation was run from Coordinated Universal Time (UTC) 00 27 May to 00 1 August 2020. To better account for the effects of the model’s initial conditions, we carried out two additional sets of simulations by changing the initialization time to UTC 00 25 May and 00 29 May 2020, respectively. Therefore, a large ensemble of 48-member (i.e., 16 different physics scheme suites × 3 perturbed initial conditions) simulations were generated for each of the experiments. No nudging or data assimilation was applied to the simulations.

Model inputs

We used the fifth-generation European Centre for Medium-Range Weather Forecasts atmospheric reanalysis of the global climate [ERA5; (68)] as the initial and lateral boundary conditions to drive the model. ERA5 provides 6-hourly global analyses of atmospheric, oceanic, and land variables at a horizontal resolution of 0.25°. We updated the SST field in ERA5 with the National Oceanic and Atmospheric Administration 0.25° daily Optimum Interpolation Sea Surface Temperature (OISST) dataset (69) throughout the simulation period, which is constructed upon multisource observations.

To calculate anthropogenically forced changes in thermodynamic variables, we used the monthly three-dimensional atmospheric temperature and SST outputs from the simulations of 13 global climate models participating in the Detection and Attribution Model Intercomparison Project of the CMIP6 (70, 71). The basic information about these models is provided in table S1. We used two sets of attribution experiments driven by all-forcings (ALL; i.e., anthropogenic greenhouse gases, aerosols, ozone, land use/land cover changes, and natural forcings) and natural forcings only (NAT; i.e., solar activities and volcanic aerosols), respectively. Since the ALL-forcing historical runs end in 2014, we extended them to 2020 with projections following a moderate emission scenario (SSP2-4.5). To better sample internal variability and extract the anthropogenically forced warming and wetting signals by means of a multimember average, each forced historical experiment of individual models was required to have at least three ensemble members.

To illustrate attribution uncertainties in the dynamic aspect, we also analyzed the monthly mean precipitation, temperature, and sea level pressure (SLP) data over the period 1961–2100 from the initial-condition large ensemble simulations of seven global climate models driven by historical all-forcings and the Representative Concentration Pathway 8.5 emission scenario [table S2; (31)].

Experimental setup

We applied the widely used pseudo-global warming approach to estimate the anthropogenically caused thermodynamic changes in the climate system (33–37, 56, 57, 61). Namely, we first computed the multimodel ensemble mean differences of decadal (2011–2020) mean air temperatures and SSTs in June and July between ALL- and NAT-forcing simulations in the CMIP6 archive (table S1). We then subtracted these man-made temperature perturbations from the historical initial and boundary thermodynamic conditions of the model and recalculated the three-dimensional atmospheric specific humidity under the assumption of constant relative humidity. Similarly, we estimated the temperature and humidity differences in June and July “climatological mean” between the time slices 2091–2100 and 2011–2020 as future anthropogenically forced thermodynamic perturbations, which were then added to the historical initial and boundary conditions.

We carried out four experiments based on the above model configurations and settings. In the first experiment (referred to as A2020), the model’s initial and boundary conditions are from ERA5 and OISST datasets, representing the actual weather and climate conditions before and during the event. The model parameter settings of the other three experiments were the same as A2020, except for the model’s initial and boundary conditions subject to thermodynamic adjustments. Specifically, three-dimensional atmospheric temperature, specific humidity, and SST were adjusted to construct a nonanthropogenic-warming climate (N2020) and plausible warmer worlds at the end of the 21st century under SSP1-2.6 and SSP2-4.5 (“SSP126” and “SSP245”). In addition, we updated the concentrations of major well-mixed and long-lived greenhouse gases (e.g., carbon dioxide, nitrous oxide, and methane) in the model to the level in 2020 for the first experiment and adjusted them to the pre-industrial level in 1850 for the second and the 2100 levels under SSP1-2.6 and SSP2-4.5 scenarios (table S3).

Observations and model validation

We used a high-resolution (0.25° × 0.25° or 28 km × 28 km) daily gridded observation dataset of China [CN05.1; (72)] to assess the model’s performance and calculate temperature and precipitation anomalies. As a cross-validation, we used ~0.1° resolution satellite-gauge merged daily rainfall from the Integrated Multi-satellitE Retrievals for the Global Precipitation Measurement Mission [GPM_3IMERG; (73)] and the monthly mean gridded temperature data of a 0.5° × 0.5° spacing from the Global Historical Climatology Network and the Climate Anomaly Monitoring System [GHCN_CAMS; (74)]. The GPM_3IMERG and GHCM_CAMS were regridded onto the 0.25° × 0.25° CN05.1 grid cells using areal conservative remapping and bilinear interpolation algorithms, respectively.

Atmospheric moisture and land-surface energy budgets

To understand the modeled change in regional precipitation in the MLRYR due to anthropogenic warming and wetting, the vertically integrated atmospheric moisture budget was examined (42), which reads as

$$P^{\mathrm{\prime }}=-⟨\mathrm{\nabla }\cdot {\mathit{V}q⟩}^{\mathrm{\prime }}+E^{\mathrm{\prime }}-{\mathrm{\partial }}_{\mathrm{t}}{⟨q⟩}^{\mathrm{\prime }}+\mathrm{\delta }$$ (1)

where P is precipitation, V is three-dimensional winds, q is atmospheric specific humidity, and E is land surface evapotranspiration. 〈X〉 = ${\displaystyle {\int }_{p_{\mathrm{t}}}^{p_{\mathrm{s}}}}X\frac{\mathrm{d}p}{g}$ is the calculator for column integration throughout the troposphere (i.e., from p_s = 1000 hPa to p_t = 300 hPa), the “∇·” represents the divergence calculator and the ()′ denotes the change caused by anthropogenic warming and wetting (difference between other experiments and A2020). The –∂_t〈q〉′ is the local change in atmospheric moisture storage, and the δ is a residual term that associated with changes in transient eddies and other nonlinear effects.

The convergence of moisture flux –〈∇ · Vq〉 can be divided into two parts: the vertical moisture advection –〈ω∂_pq〉 and the horizontal moisture advection –〈V_h · ∇_hq〉. Thus, the Eq. 1 can be reorganized as

$$P^{\mathrm{\prime }}=-{⟨\mathrm{\omega }{\mathrm{\partial }}_{\mathrm{p}}q⟩}^{\mathrm{\prime }}-{⟨{\mathit{V}}_{\mathrm{h}}\cdot {\mathrm{\nabla }}_{\mathrm{h}}q⟩}^{\mathrm{\prime }}+E^{\mathrm{\prime }}-{\mathrm{\partial }}_{\mathrm{t}}{⟨q⟩}^{\mathrm{\prime }}+\mathrm{\delta }$$ (2)

where ω is pressure velocity and V_h is horizontal wind vector.

The change in vertical moisture advection can be further decomposed into three terms that, from left to right of the Eq. 3 below, are solely due to a change in atmospheric moisture profile, a change in vertical pressure velocity, and a nonlinear interaction term

$$-{⟨\mathrm{\omega }{\mathrm{\partial }}_{\mathrm{p}}q⟩}^{\mathrm{\prime }}=-⟨{\mathrm{\omega }}_{\mathrm{ref}}{\mathrm{\partial }}_{\mathrm{p}}{q}^{\mathrm{\prime }}⟩-⟨{\mathrm{\omega }}^{\mathrm{\prime }}{\mathrm{\partial }}_{\mathrm{p}}{q}_{\mathrm{ref}}⟩-⟨{\mathrm{\omega }}^{\mathrm{\prime }}{\mathrm{\partial }}_{\mathrm{p}}{q}^{\mathrm{\prime }}⟩$$ (3)

where ω_ref and q_ref are the vertical pressure velocity and specific humidity in the simulations of reference experiment when calculating changes (i.e., N2020 for A2020 versus N2020; A2020 for SSP126 or SSP245 versus A2020).

To understand modeled temperature changes in SC, the land-surface energy budget was analyzed, which is written as

$$R_{\mathrm{n}}={\mathrm{SW}}_{\mathrm{d}}+{\mathrm{SW}}_{\mathrm{u}}+{\mathrm{LW}}_{\mathrm{d}}+{\mathrm{LW}}_{\mathrm{u}}$$ (4)

where R_n is net surface radiation, SW and LW represent shortwave and longwave radiation on the land surface, and the subscripts “d” and “u” denote downward (incoming) and upward (outcoming) directions.

The net surface radiation is balanced by latent heat flux λE, sensible heat flux H, and ground heat flux G

$$R_{\mathrm{n}}=\mathrm{\lambda }E+H+G$$

where λ is the latent heat of vaporization.

Constructed flow analog approach

We first reconstructed the dynamically-driven anomalies of daily precipitation in the MLRYR (105° to 121°E; 29° to 35°N) and concurrent daily temperature anomalies in SC (105° to 121°E; 22° to 28°N) during June–July 2020. To this end, we identified the patterns for daily SLP anomalies (SLP_ano; relative to 1961–1990) in the key circulation region (102.5° to 124.5°E; 17.5° to 37.5°N) based on ERA5. Then, we calculated the daily anomalies (relative to 1961–1990) of regional area-weighted mean precipitation in the MLRYR (P_ano) and surface air temperature in SC (T_ano), and removed their secular linear trends on monthly basis using the Theil-Sen slope estimator (75). For each day in June–July 2020, we selected the top 50 analog days that were most similar to the target SLP_ano patterns from samples within a 31-day window centered on the given day but in other years. The degree of pattern similarity was measured by the TW score (51). Twenty five of the 50 (i.e., 50%) SLP_ano analogs were randomly subsampled and linearly combined to form an “ensemble mean” analog by applying the Moore-Penrose pseudoinverse to the linear regression model (48, 50). The derived linear regression coefficients were applied to the P_ano and T_ano of the corresponding analog days to obtain the optimal linear combinations of surface weather. We repeated this random sampling procedure 1000 times and took the median of the constructed P_ano and T_ano as the best estimates for the dynamically induced components of the observed P_ano and T_ano.

The TW score compares the two SLP_ano patterns (constructed versus observed) by considering their zonal and meridional pressure gradients rather than the absolute anomalies in each grid cell (51). Let g(k, t) denotes the difference in pressure gradient at grid cell k between the observed (at day u) and the candidate analog (at day t) SLP_ano fields that include I × J grid cells in zonal and meridional directions

$$\begin{array}{r}g(k,t)=\mid [{\mathrm{SLP}}_{\mathrm{ano}}(k-1,u)-{\mathrm{SLP}}_{\mathrm{ano}}(k+1,u)]\\ -[{\mathrm{SLP}}_{\mathrm{ano}}(k-1,t)-{\mathrm{SLP}}_{\mathrm{ano}}(k+1,t)]\mid \end{array}$$ (6)

And let G(k, t) represents the larger gradient of the two SLP_ano fields

$$G(k,t)=\max [\mid {\mathrm{SLP}}_{\mathrm{ano}}(k-1,u)-{\mathrm{SLP}}_{\mathrm{ano}}(k+1,u)\mid ,\mid {\mathrm{SLP}}_{\mathrm{ano}}(k-1,t)-{\mathrm{SLP}}_{\mathrm{ano}}(k+1,t)\mid ]$$ (7)

The TW score, i.e., TWs, for day t is

$$\mathrm{TWs}(t)=\frac{{\displaystyle \sum _{k=2}^{I-1}}g(k,t)+{\displaystyle \sum _{k=2}^{J-1}}g(k,t)}{{\displaystyle \sum _{k=2}^{I-1}}G(k,t)+{\displaystyle \sum _{k=2}^{J-1}}G(k,t)}$$ (8)

We converted it into percentage form. By design, a smaller TW score indicates a higher circulation pattern similarity.

Supplementary Materials

This PDF file includes:

Tables S1 to S3

Figs. S1 to S11