On the possibility of the 2022-like spatio-temporally compounding event across the Yangtze River Valley

During July–September 2022, heatwaves, droughts, forest fires and floods hit the Yangtze River Valley successively, constituting a spatio-temporally compounding event. Understanding its risks matters to disaster preparedness. Through searching for event analogues in single-model initial-condition large-ensemble climate simulations, we report that the 2022 unprecedentedly widespread and intense hot drought might have occurred as early as in the 1970s, and would become increasingly possible and spatially extensive with warming. This tendency is also supported by the conventional multi-model (CMIP6) projection, especially evident in larger ensembles. Lower reaches of the valley and parts of Southwest China have greater chances of repeated exposure to the 2022-like heat—drought—fire—flood quadruple compound events. In the presence of favorable internal variability in line with future warming, it is plausible to see more than half of the valley at simultaneous risk of the 2022-like quadruple compound event. Our possibility projection highlights the urgency of accelerating the existing univariate extremes—oriented adaptation measures to better address emerging threats from unfamiliar compound hazards.


Introduction
The summer of 2022 was marked by multiple weather and climate extremes to the Yangtze River Valley (YRV), including heatwaves, droughts, forest fires, heavy precipitation and floods.These events are exceptional not only in magnitudes and/or durations but also in the form of spatially and temporally compounding (figure 1), which made their impacts multiply (Hao et al 2023).Starting in mid-July, one of the most severe, sustained and extensive hotdrought took place over much of the YRV, leading to power outages and harvest failures in the electricityand water-rich region.The hot-dry weather dried out forests, and sparked numerous wildfires (WMO 2023a).The hazard chain extended as some of the burned areas received heavy precipitation, where the flood and debris flow risks stayed high in the following months.
The rapid sequencing of multiple extremes shortens relief windows, enhances exposure/vulnerability of natural and human systems to the next hazard(s), and therefore might overwhelm their ability to respond.Such inter-connected extremes are referred to as compound events (Zscheischler et al 2018).The 2022 YRV case, characteristic of spatial synchrony, temporal concurrence and succession, is a mixture of several well-recognized types of compound events (Zscheischler et al 2020), and seems more proper to be termed as a spatio-temporally compounding event (Reichstein et al 2021).Whether the combination is just by chance or signals novel hazards is concerning to disaster preparedness.
Information on projected changes in such compound events remains sparse due to data and methodological challenges.First, an extreme event by definition is rare statistically, and multivariate events like the 2022 YRV case are far less sampled in decadeslong observational records (Liao et al 2021, Bevacqua et al 2023).Second, the multivariate statistical modeling techniques are far from mature to provide a reliable estimate for quadruple compound events' statistics, e.g.probability.Third, the long-standing pursuit for the 'likely range' based on models' consensus (IPCC 2021) deals inadequately with events that are distributed at the far tail of the ensemble range.The tail-distribution nature could arise from unprecedented magnitudes and/or unfamiliar combinations of events (Mankin et al 2020, McCollum et al 2020).Last but not least, the commonly practiced CMIP (Coupled Model Intercomparison Project) multimodel projections tend to treat internal variability as noise-a major source of inter-model projection uncertainty (Deser et al 2020); however, they are intrinsic to the climate system, and play a joint role with external forcings in altering both univariate extremes and the way of their clustering (Diffenbaugh and Davenport 2021, Touma et al 2022, Tan et al 2023).
The single-model initial-condition large ensembles (SMILEs) offer an opportunity to address these challenges.In each participating model, tens of ensemble members are run under the identical model physics and external forcings, and only differ in their initial conditions.So, the ensemble constitutes a broad range of physically plausible climates that better serve for the purpose of exploring possibility of extremes at a given warming level (Fischer et al 2023).The SMILEs have been used to investigate extremely rare events that either break previous records by a wide margin (e.g.Fischer et al 2021, Huang et al 2022) or consists of paired hazards (e.g.Swain et al 2018, Liao et al 2021).Inspired by these pioneer works, we project spatio-temporally compounding events analogous to the YRV case in 2022 mainly in the SMILE framework.Instead of focusing on event probability (quantification of likelihood and its change with huge uncertainty), we attempt to address the event possibility issue, i.e. whether we have the chance to foresee similar compound events in large-ensemble climate simulations and if yes how worse they could have behaved.

Data
For observations, we used quality-controlled monthly-mean maximum temperature (T max ), daily and monthly precipitation totals (P) over 1961-2022 observed at 2474 meteorological stations across mainland China, provided by the National Meteorological Information Center.
To calculate vapor pressure deficit (VPD), we also used air temperature, dew-point temperature and surface pressure variables from the ERA-5 reanalysis, with a resolution of 0.25 • in space and hourly in time (Hersbach et al 2020).The fire detection is based on the Collection 6.1 Level 2 Active fire products with coordinates and radiative power provided at an 1 km resolution (Giglio et al 2016).We only considered type-0 detections representative of vegetation fires.Following the data developers' recommendation (Giglio et al 2021), a confidence threshold for fire detection was set to 30%.
For simulations, we took variables as required in observations from two models in the SMILEs collection, i.e.CESM1 (Kay et al 2015) and CanESM2 (Kirchmeier-Young et al 2017).Only these two models from the seven-model archive (Deser et al 2020) were chosen because the others lack warranted daily variable outputs.The CESM1 has a horizontal resolution of 1.3 • × 0.9 • , and provides a 40-member ensemble; while the CanESM2 has a horizontal resolution of 2.8 • × 2.8 • , and provides a 50-member ensemble.Given the advantage of higher resolution in simulating localized extremes to aid hotspots mapping, we show the CESM-LE results in the main text, with CanESM2-based results presented in supplementary materials.
By experimental design of the SMILEs, historical simulations end in 2005 driven by past anthropogenic and natural forcings, and the 21st-century projection extends from 2006 to 2100.We also repeated the analysis using CMIP6 models (table S1), that provide all required variables at corresponding temporal scales.For the multi-model simulations, we adopted two commonly employed sampling protocols by taking (1) the first simulation member from each of the 26 models (table S1) that output all of required variables (i.e. the r1i1p1f1 scheme), and (2) the first three members from each of the 15 qualified models to assemble a 45-member ensemble (table S1, names in bold) comparable to the size of SMILEs being used (40-50).To be consistent with the scenario primarily considered in the SMILEs (RCP8.5,Deser et al 2020), only SSP5-8.5 projections were used in the CMIP6 results.To reconcile periods of historical simulations from the two groups, we artificially extended the historical period to 2020 by the RCP8.5/SSP5-8.5projections, and referred to it as 'the current climate' hereafter.

Methods
Instead of subjectively prescribing intensity thresholds and implicitly assuming their high impacts (i.e. the top-down perspective), we thoroughly examined the process of hazard cascading to sort out impact-relevant meteorological anomalies (i.e. a bottom-up perspective, Bevacqua et al 2021).
For the heat-drought event, both the magnitude of T max /P anomalies and the spatial extent of their concurrence are critical to the impacts on energy, health, water and agriculture.For instance, the spatially co-located droughts halved the energy production capacity in the hydropower-dominated region during the summer (International Energy Agency 2023); while the record heatwaves in multiple places substantially elevated the regional energy demand.The two stressors, together, resulted in a two-weeklong blackout.For heat-drought events of various intensity levels (figure S1), we report that their spatial extents are two to three times the previous records.In particular, the event with T max ⩾ 99th percentile & P ⩽ 10th percentile covered around 35% of the valley, more than four times as widespread as previous counterparts.Empirically, such a large margin beyond normal is rationale to be assumed impossible and hence falls out of the scope of routine disaster preparation.The intensity and areal extent thresholds are accordingly adopted to define the 2022-like heatwave-drought spatially compounding event.
The sustained hot-dry weather triggered forest fires during August-September, with the radiative power of severe cases exceeding 50 W m −2 (figure 1, symbol-×).Due to the lack of fire data in model outputs, we measured fire potential by VPD, which proves itself to be a good proxy for fuel moisture content in a range of forest and woodland biomes (Balch et al 2022, Clarke et al 2022).The daily VPD was calculated following the scheme of Yuan et al (2019).As guided by the severest fires during August-September 2022 observed within model grids, the grid-level VPD of the day mostly exceeded the historical 99th percentiles.Given that the coarse model resolution precludes a finer VPD-fire matching, the threshold-exceeding VPD should be interpreted as one of the necessary conditions for 2022-like severe fires somewhere in the grid.
Just in a month following the fires, some of the fire-stricken areas received heavy precipitation that reached up to 150-200 mm in a day (figure S2) or a total of 200-300 mm over the course of 5 d (shading, figure 1).With vegetation removed and soil properties changed (e.g.hydrophobic) by fires, flash floods and debris flows easily result even subject to moderate wet extremes (Thomas et al 2021, Jong-Levinger et al 2022).In this regard, the fire and heavy precipitation combination is typical of a temporally compounding event (Zscheischler et al 2020, Touma et al 2022).In addition to precipitation intensity and duration, post-fire hydro-geological hazards also depend on land cover and topography.Given the poor understanding on complex interactions amongst these factors, we did not strictly follow the absolute precipitation intensity as observed; instead, we defined post-fire heavy precipitation as those exceeding either the 99th percentiles for daily accumulation to represent short-lived, intense events or the 95th percentiles for five day totals to represent persistent events.Similarly, the current model resolution allows for neither identification of more localized precipitation extremes nor their matching to other extremes in situ.
To summarize, the 2022-like spatio-temporally compounding event is defined by the order of a widespread and intense hot-drought during July-August (T max ⩾ 99th & P ⩽ 10th covering at least 35% grids of the YRV); in any of the hot-drought grids, at least one extreme VPD day (VPD ⩾ 99th) during August-September; then at least one heavy precipitation event (daily P ⩾ 99th or five day total ⩾ 95th) within 30 d following any of extreme VPD events in the same grid.
A compound event is considered possible if it is simulated by any member from the CESM-LE or by any model in the CMIP6 multi-model context.Given more members/models (converted to fraction of nonzero event members, for fairer comparison) projecting the target event, its occurrence possibility is claimed to increase.In addition, we also explore how widespread future compound events might behave, i.e. the possibility for event areal extents.
Due to the lack of fit-for-purpose model evaluation metrics and bias correction methods tailored to quadruple compound events as well as unclear mechanistic understanding (see Discussion), no formal evaluation on models' performance in simulating the event's statistics and physics was conducted.Rather, the percentile-based thresholds serve as an implicit bias correction to improve the comparability of the simulations to relative changes in observations (Swain et al 2020, Poschlod et al 2020).

Results
Comprehensively considering the intensity and areal extent, the 2022-like heat-drought event is extremely rare even in ∼2400 model years in the CESM-LE (40 ensemble members × 60 years) during the historical period .However, alternative modes of internal variability in combination with historical anthropogenic forcings could have caused similar events much earlier (in the 1970s) or more spatially extensive (∼55% of the YRV, the simulated case in 2017) (figure 2).The earlier emergence and wider coverage of similar hot droughts are still possible when considering different thresholds (figure S3) or using the CanESM2-LE simulations instead (figure S4(a)).Amongst the 26-member CMIP ensemble, only two models (figure 2(a)) suggest the possible occurrence of the target hot-drought before 2021.By the IPCC confidence language (Intergovernmental Panel on Climate Change (IPCC) 2021), the simulated events lie outside the likely range of the multi-model ensemble from a probabilistic perspective.In the multi-model context, one is rationale to attach more confidence to the impossibility of historical occurrence consistently suggested by the remaining 24 models.There should be accordingly no need to adopt proactive preparedness against the 'unlikelysibility of historical occurrence consi hazard' .
Taking the multi-member mean (CESM-LE, red curve) or multi-model ensemble mean (CMIP, blue curve) to approximate anthropogenically-forced changes (figure 2(a)), we report that anthropogenic warming in general favors spatial expansion of concurrent heatwaves and droughts of the 2022 event's intensity.With diverse plausible modes of internal variability superimposed, we project a general increase in the occurrence possibility of the 2022-like hot droughts, as indicated by growing numbers of ensemble members in the CESM-LE having the chance to see the event (figure 2(b), red bars).Specifically, the number of non-zero event members in the CESM-LE rises from 5 before 2021 to 32 in the next 80 years in total.As for areal extent, there would be greater possibilities to see cases affecting more than half of the domain, with the most widespread case covering ∼80% of the valley.Given the same forcings and model physics across the CESM-LE, the member-variant compound events during a given period could be reasonably ascribed to the underlying unforced natural variability.Also note that the increases for both the event number and the spatial extent are not monotonous with time.This further points to significant modulation of internal variability on the risk of low-likelihood compound events.
The tendency of increasing possibility for 2022like intense and widespread hot droughts is also supported by the CanESM2-LE (figure S4(a)) and CMIP6 projections (figure 2, blue bars).In particular, by the multi-model framework, expanding the ensemble member to 45 (see data) brings to light slightly higher possibility of earlier and wider historical hot droughts of the 2022 intensity (figure S5(a), comparable to counterparts in the CESM-LE).The possibilities for future occurrence of 2022-like hot droughts and exceptionally widespread analogues also grow with the size of ensemble members in the multimodel framework (figure S5(b) vs. Figure 2(b), blue bars).This comparison, again, underlines the necessity of sufficiently sampling internal variability to better visualize the risks of unfamiliar compound hazards.However, a further attribution of distinct events simulated by the multi-model ensembles to different expressions of internal variability is not straightforward because model-dependent climate sensitivities and resolution-relevant physical parameterizations are also confounded (table S1).
In broad swaths of the valley just subject to hot droughts, forest fires could be ignited at any location with elevated fuel aridity, and heavy precipitation tends to behave in a more random way depending on chaotic storm activities and topography.We next examine the possibility of their sequencing down to a local scale.There is a spatially heterogeneous pattern for the ensemble-member aggregated occurrences of heat-drought-extreme VPD-heavy rainfall quadruple compound events, with one hotspot in the populous lower reaches and the other in the mountainous and denselyforested Southwest China (figure 3).The identified hotspots are largely consistent with the affected communities in the summer of 2022 (figure 1), but are much wider than observed.This warns emerging threats of similar cascading hazards to the unexposed areas in the past.Despite diverse plausible configurations between internal variability and the considered emission scenario, there will be only one realization to be observed.Within the CESM-LE, we pick up the most frequently-occurring member for each grid (figure 3(b)).Given the most favorable internal variability modes under the RCP8.5 scenario, much of the lower reaches and patches of Southwest China might see more than five 2022-like quadruple compound events in the next few decades.(1961-1980, 1981-2000, 2001-2020……2081-2100).
By contrast, only sporadic cases, i.e. 1-2 at most within 80 years, are expected in the upper reaches of the YRV and South China.At the local scale, anthropogenic climate change, mainly warming, set the stage for the four extremes compounding, as shown by more than half of ensemble members projecting at least one quadruple compound event across much of the valley in contrast to less than five nonzero event members in cooler climates (figure 3(c) vs. figure S6).
Though forest fires and heavy precipitation are local hazards, their outbursts in multiple sites within the same season could put more strains on rescue and relief resources (McGinnis et al 2023), thus leading to 'loss amplification' (Otto et al 2020).This motivates us to investigate the seasonal areal extent of fire and flood-triggering conditions also from a spatially compounding perspective.For the heat-drought-extreme VPD combination, most historical events (before 2021) could have covered 35%-45% of the valley, and near-future (2021-2040) cases would somehow behave less extensively (figure 4, red).Higher warming levels (the 2040s onwards) would not necessarily make the triple compound events more extensive than those in previous decades, but do boost the chance of experiencing exceptionally widespread cases (⩾50% of the valley), based on both the SMILEs (figure 4(a The quadruple-whammy events, with the fourth component-heavy precipitation involved, typically have smaller sizes in space, griping 20%-40% of the valley for most seasons.The spatial extent of tripleand quadruple compound events shows no systematic increase, though the threshold-exceeding T max and VPD are projected to occur in almost every grid every summer after 2080 in the high-end emission scenario.

Discussions
The combination of the impact-centric definitional perspective and the SMILEs establishes a paradigm for evaluating risks of low-likelihood compound events.It could be used to explore the possibility of events compounding both through physical linkage Despite the merits, the analytic tool is not exempt from limitations.The 2022 event exposed the previously overlooked vulnerability of the society to interconnected extremes (Hao et al 2023) and is therefore more persuadable and acceptable to stakeholders for guidance of adaption planning.However, taking a single case as a benchmark may underestimate the diversity of combinations amongst events (e.g.intensity, duration, extent and interval) that are capable of leading to comparable or severer impacts.Machine learning models trained by similar impacts recorded in some disaster databases or derived from impact modeling provide a promising way forward (Sweet and Zscheischler 2022, Yu et al 2022).In addition, the results are model-dependent.Constrained by the availability of required variables, only CESM and CanESM2 large ensembles were used here.Comparisons to projections from other models, when available, are still needed to enhance the robustness of conclusions.Also, the projected worst scenario in only 40-50 ensemble members might only represent a conservative estimate for future risks of spatio-temporally compounding events.It is worthwhile of revisiting the upper-limits of impact-relevant characteristics (e.g.areal extent) when more ensemble members (e.g.∼100) are ready.
Another limitation of the study pertains to a fragmentary understanding on the physical inter-linkage among components.The huge impacts of the 2022 spatio-temporally compounding event, along with our possibility projections, underline that using the ever-experienced extremes to guide adaptation planning is far from sufficient and effective.For instance, despite proactive adaptation measures against common heatwaves and droughts in the YRV (Hao et al 2023), the unprecedentedly intense and widespread 2022 event still caused disastrous consequences.Also, much of attention in infrastructure designing has been paid to strong individual extremes (e.g.∼100 year events), but back-to-back strikes from multiple extremes have been considerably less concerned.As a matter of fact, the quadruple compound event is not unique to the YRV, but was also reported in other regions, e.g.Southern California (AghaKouchak et al 2020).Identifying hotspots and visualizing the worst case through the lens of possibility projection would help prepare for and enhance local communities' resilience against the emerging compound hazard.

Conclusions
We used initial-condition large-ensemble simulations to explore possibilities of the 2022-like heat-drought-fire-flood spatio-temporally compounding event, in terms of chance of occurrence and spatial extent.Overall, a warmer climate is conducive to the spatial concurrence and temporal cascade of these extremes in the YRV, and favors the emergence of more widespread compound events there.Further accounting for the significant modulation of internal variability, we report that the unprecedentedly widespread and intense heat-drought event in July-August 2022 might have had a chance to occur much earlier than observed.The chance of the four extremes happening in the 2022 summer manner would be higher in the lower reaches of the YRV and parts of Southwest China.In the context of favorable internal variability modes under the high-end emission scenario, the most widespread quadruple compound event might affect more than half of the valley within the same season.

Figure 1 .
Figure 1.Observed spatially and temporally compounding pattern.The colored dots show joint percentiles, as detailed in the right panel legend, of July-August mean Tmax and precipitation total, with grey ones indicating no hot droughts of the considered intensities.The black numbers in each bin of the joint percentiles show the fraction of stations (%) that observed excess heat and precipitation deficit of the intensity (x left ⩽ Tmax < x right & yupper ⩽ P < y lower ).Symbols-X in the map locate severe forest fires with radiative power above 50 W m −2 during August-September.The color-filled grids in the map display the maximum post-fire (within a month) five day precipitation totals amongst stations within the model grid, referring to the colorbar beneath.

Figure 2 .
Figure 2. The occurrence and spatial extent of the 2022-like hot drought.(a) The observed (black), multi-member (red), and multi-model (blue) ensemble mean spatial extent (curves) of hot droughts defined by Tmax ⩾ 99th percentile and P ⩽ 10th percentile.Simulated cases from different modeling groups that are equivalently or more widespread compared to the observed extent (grey dashed line) are shown by the symbol-X of corresponding colors.(b) The fraction of non-zero events CESM-LE members (red) and CMIP6 models (blue) in each 20 year period(1961-1980, 1981-2000, 2001-2020……2081-2100).

Figure 3 .
Figure 3. Hotspots for spatio-temporally compounding events as identified by CESM-LE 40 member-aggregated occurrences (a), event numbers in the ensemble member that projects the most frequent occurrence for each grid (b), and numbers of non-zero event ensemble members (c), through 2021-2100.

Figure 4 .
Figure 4.The spatial extent of simulated heat-drought-extreme VPD triple compound events (a) and heat-droughtextreme VPD-heavy precipitation quadruple compound events (b).For each period, if the simulated sample size is larger than 10, the distribution of their spatial extents is shown in the form of box-and-whisker plots; otherwise, all cases are directly shown by sample dots with the median overlaid.The dashed line delimits the half-of-the-valley level, beyond which the events are termed as exceptionally widespread cases.
Large-scale atmosphericoceanic teleconnections (Chen and Li 2023, Jiang et al 2023, Wang et al 2023) and strong regional landair interactions (Hao et al 2023, Jiang et al 2023) have been identified as major drivers for the July-August heatwaves and droughts.Both the heat-drought and fire might arise partially from the springtime vigorous plant growth (WMO 2023b), which acted to accelerate summertime soil moisture depletion and prepared abundant fuels (Zhang et al 2021, Gloege et al 2022).The inter-dependence between fire and heavy precipitation remains largely elusive.Neither process-based understanding nor multivariate statistical modeling enables us to confirm the nature of their sequencing, i.e. a causal connection due to firerelated thermodynamic disturbances (Cunningham and Reeder 2009, Zhang et al 2019) or just coincidence (AghaKouchak et al 2020, Touma et al 2022).
WorkingGroups I, II and III tothe Sixth Assessment Report of the Intergovernmental Panel on Climate Change (Cambridge University Press) (available at: www.ipcc.ch/ar6-syr/)Intergovernmental Panel on Climate Change (IPCC) 2021 Climate change 2021: the physical science basis Contribution of Working Group I to the Sixth assessMent Report of the Intergovernmental Panel on Climate Change ed V Masson-Delmotte et al (Cambridge University Press) Riede F and Frank D 2021 More floods, fires and cyclones-plan for domino effects on sustainability goals Nature 592 347-9 Swain D L, Langenbrunner B, Neelin J David and Hall A 2018 Increasing precipitation volatility in twenty-first-century California Nat.Clim.Change 8 427-33 Swain D L, Wing O E, Bates P D, Done J M, Johnson K A and Cameron D R 2020 Increased flood exposure due to climate change and population growth in the United States Earth's Future 8 Sweet L and Zscheischler J 2022 Using interpretable machine learning to identify compound meteorological drivers of crop yield failure EGU General Assembly Conf.Abstracts pp EGU22-5464 Tan X, Wu X, Huang Z, Fu J, Tan X, Deng S, Liu Y, Gan Yew T and Liu B 2023 Increasing global precipitation whiplash due to anthropogenic greenhouse gas emissions Nat.Commun.14 2796 Thomas M A, Rengers F K, Kean J W, McGuire L A, Staley D M, Barnhart K R and Ebel B A 2021 Post wildfire soil-hydraulic recovery and the persistence of debris flow hazards J. Geophys.Res.126 e2021JF006091 Touma D, Stevenson S, Swain D L, Singh D, Kalashnikov D A and Huang X 2022 Climate change increases risk of extreme rainfall following wildfire in the western United States Sci.Adv. 8 eabm0320 Wang Z, Luo H and Yang S 2023 Different mechanisms for the extremely hot central-eastern China in July-August 2022 from a Eurasian large-scale circulation perspective Environ.Res.Lett.18 024023 World Meteorological Organization (WMO) 2023a State of the Global Climate in 2022 (available at: https://public.wmo.int/en/our-mandate/climate/wmo-statement-state-of-globalclimate) World Meteorological Organization (WMO) 2023b State of Global Water Resources 2022 (available at: https://public.wmo.int/en/our-mandate/water/state-of-global-water-resources-2022) Yu Y, Mao J, Wullschleger S D, Chen A, Shi X, Wang Y, Hoffman F M, Zhang Y and Pierce E 2022 Machine learning-based observation-constrained projections reveal elevated global socioeconomic risks from wildfire Nat.Commun.13 1250 Yuan W et al 2019 Increased atmospheric vapor pressure deficit reduces global vegetation growth Sci.Adv. 5 eaax1396 Zhang Y, Fan J, Logan T, Li Z and Homeyer C R 2019 Wildfire impact on environmental thermodynamics and severe convective storms Geophys.Res.Lett.46 10082-93 Zhang Y, Keenan T F and Zhou S 2021 Exacerbated drought impacts on global ecosystems due to structural overshoot Nat.Ecol.Evol. 5 1490-8 Zscheischler J et al 2018 Future climate risk from compound events Nat.Clim.Change 8 469-77 Zscheischler J et al 2020 A typology of compound weather and climate events Nat.Rev. Earth Environ. 1 333-47