Potential for surprising heat and drought events in wheat-producing regions of USA and China

The study addresses the risk of historically unprecedented heat and drought events in major winter wheat-producing regions of the USA and China, questioning the adequacy of risk assessments based solely on historical climate–yield relationships. Given global food system interdependence, simultaneous shocks in breadbasket regions can strongly affect food availability and prices. Traditional analyses extrapolate from past climate–yield associations, but climate change introduces novel and extreme events that can cross critical physiological thresholds in wheat, leading to nonlinear impacts not captured by historical correlations. The purpose is to visualize and quantify current-day risks of unprecedented temperature and precipitation extremes during sensitive wheat growth phases (March–May) using the UNSEEN large-ensemble approach, thereby informing adaptation planning and highlighting potential underestimation of risk when relying on history alone.

Prior work on multiple breadbasket failures largely extrapolates from historical climate–yield relationships, estimating risks of simultaneous climate shocks based on past data. Studies have shown substantial portions of yield variability attributable to weather, and both process-based and statistical models generally agree that warming reduces wheat yields. Physiological studies identify critical wheat growth phases sensitive to heat (e.g., stem elongation, anthesis, grain filling), with thresholds around 27.8 °C for stress and above ~32.8 °C for enzyme breakdown; temperatures above ~34 °C are damaging in some contexts. Novel approaches such as UNSEEN and climate storylines have been used to reassess extremes, often anchored to past events; they can also generate synthetic, unprecedented events if models are credible. Given that adaptation decisions are often driven by experienced extremes, visualizing plausible but unobserved extremes can support preparedness.

Study design followed UNSEEN protocol to ensure credibility. Regions: USA Midwest winter wheat area (approx. 105–95.5 W, 35–40 N) including western Kansas, eastern Colorado, NW Oklahoma; China winter wheat region in the northeast/east (110–122.5 E, 30–45 N) including Hebei, Shandong, Henan, Jiangsu, Anhui. Season of interest: March–May (MAM), coinciding with sensitive development stages (anthesis, grain filling). Variables: (1) TXx, seasonal maximum daily temperature; (2) number of days with Tmax > 27.8 °C (stress); (3) number of days with Tmax > 32.8 °C (enzyme breakdown); (4) total precipitation in MAM. Calculations at native dataset resolution with area-averaging over land points in each region. Large-ensemble data: ECMWF SEAS5 seasonal forecast archives (initialized monthly, 7-month integrations). Ensemble size: 25 members per init from 1981–2016; 51 members from 2017 onward. Leads including full MAM: November–March inits (leads 5 to 1). This yields 125 realizations per year up to 2016 and 255 per year from 2017 onward across five lead times. Observational/reanalysis references: USA—DayMet v4 (1 km daily); China—ERA5-Land (9 km). All datasets scaled to 1° for analysis. UNSEEN evaluation: - Stability across lead times: checked for model drift; none found (Supp. Figs. 3–4). - Independence across ensemble members: excluded lead 1 for all variables due to dependence. Removed any lead where pairwise rank correlations median > 0.25; occurred once (TXx USA lead 2 removed). - Fidelity tests: created 10,000 proxy series matching observational length from UNSEEN; compared mean, SD, skewness, kurtosis. Applied additive bias correction for China MAM total precipitation (UNSEEN minus 24 mm to match observed mean). TXx USA observed kurtosis slightly below 95th percentile of UNSEEN; interpret with caution. For derived threshold-count variables, stats generally within 95% range, except China “stress days” below 5th percentile—interpret cautiously. Statistical analyses: - Visual inspection of time series with boxplot statistics for UNSEEN vs observations. - Threshold exceedance counting to estimate probabilities and joint likelihoods (compound heat–dryness). - Extreme value analysis: fitted Gumbel, GEV, and non-stationary GEV to TXx for observations and UNSEEN; location and scale linearly dependent on time; model selection via likelihood-ratio tests; parameters via MLE; return periods and magnitudes estimated for 1981 and 2020. - Dynamical diagnostics: composites of 500 hPa geopotential height (GPH) and wind anomalies for 10 driest, wettest, and hottest (by enzyme-breakdown days) seasons in each region, using available monthly SEAS5 data (leads 1–4). Compound-event analysis: assessed correlations between regions and identified ensemble members producing concurrent extremes; composited associated circulation patterns.

- Rising temperature extremes: UNSEEN ensembles show increasing MAM TXx in both regions; in the US Midwest, ensemble upper quartiles approach ~35 °C with maxima near 40 °C vs observed maxima ~37 °C; in China, UNSEEN includes high-30s °C vs observed ~35 °C. - Increased heat-threshold exceedances: Modeled and observed increases in days above stress (27.8 °C) and enzyme-breakdown (32.8 °C) thresholds. UNSEEN shows discrete events far beyond observed, including >20 enzyme-breakdown days in US Midwest and >10 in China. - Return period shifts: A 1-in-100-year MAM TXx event in 1981 now occurs about every 6 years in the US Midwest (≈17% annual chance) and about every 16 years in China (≈6% annual chance) by 2020. Table 1 100-year TXx magnitudes (UNSEEN): USA 35.7 °C (1981) to 38.7 °C (2020); China 34.7 °C (1981) to 36.1 °C (2020). - Heat–dryness coupling: Extreme heat strongly associated with low precipitation; very wet seasons do not co-occur with extreme heat. In the US Midwest: 161 UNSEEN record-breaking hot seasons (by enzyme-breakdown days); 14% had rainfall lower than the worst observed drought year (2014). Among 31 UNSEEN droughts worse than observed record, 71% also had record-breaking heat. In China: 63% of UNSEEN record-breaking droughts were also record-breaking heat seasons. - Circulation drivers: Hot/dry extremes in both regions linked to 500 hPa circulation with wind anomalies advecting dry continental air into regions and large-scale high-pressure anomalies. Wet extremes linked to onshore flow (southerly/easterly) bringing moisture from Gulf/Atlantic (USA) and from the south/east (China) associated with strong highs to the northeast. - Compound hot extremes across regions: Weak but positive correlation in TXx between USA and China in UNSEEN (r ≈ 0.06; 95% CI 0.03–0.09), likely reflecting global warming influence. Identified 10 ensemble members generating simultaneous extreme heat in both regions; composite pattern suggests a zonal wavenumber-3 disturbance with highs over both regions. The most extreme compound event (2018 ensemble member) produced regional averages of 12.9 enzyme-breakdown days (USA; observed record 8.5) and 5.2 (China; observed record 2.9). - Precipitation trends: No clear MAM rainfall trend in US Midwest; UNSEEN contains events drier than 2014 observed drought. In China, UNSEEN includes drier-than-observed seasons. - Implication: Recent observed extremes, especially in US Midwest, lie at the lower end of today’s plausible extreme distribution, implying a high potential for surprise if relying on historical records.

The findings demonstrate that relying on historical climate–yield relationships risks underestimating present-day hazards to wheat in major U.S. and Chinese breadbaskets. The UNSEEN approach reveals that unprecedented, physiologically critical heat levels are now plausible with substantially increased frequency, and often coincide with drought, posing significant threats to yields despite weak historical temperature–yield correlations. The study links these extremes to robust large-scale circulation patterns, providing physically consistent storylines that can be monitored as precursors. By quantifying shifts in return periods and illustrating compound hot–dry and trans-breadbasket events, the results address the research question of how current risk has diverged from historical experience, emphasizing the need for adaptation and revised risk perception among stakeholders.

This study applies the UNSEEN large-ensemble approach to current-climate risk in two key winter wheat regions, revealing substantial increases in the likelihood and magnitude of record-breaking heat—often concurrent with drought—beyond historical experience. It provides dynamically consistent storylines for unprecedented single-region and compound cross-region events, improving risk awareness and informing adaptation planning. Future research should: enhance model fidelity and perturbation methods to better sample low-likelihood high-impact events; integrate crop process models with threshold responses to unprecedented heat; assess multi-hazard interactions and socioeconomic drivers (e.g., conflict, trade); and develop monitoring/forecasting tools for the identified circulation precursors. Continued work on heat- and drought-tolerant cultivars, agronomic adjustments (planting/maturity dates), irrigation and soil moisture management, and potential shifts in agricultural zones will be critical to reduce vulnerability.

- Model fidelity and sampling: Despite stability and independence checks, the UNSEEN ensemble may not capture the full spectrum of plausible events. Certain variables required caution: USA TXx kurtosis marginally outside the 95% range; China stress-days below the 5th percentile; China precipitation required mean bias correction (−24 mm). - Lead-time dependence: Lead 1 excluded for dependence; USA TXx lead 2 removed due to interdependence—potentially reducing sample diversity. - Observational proxies: China used ERA5-Land rather than dense in situ observations; scaling to 1° may smooth local extremes. - Scope: Focus on MAM season and selected regions/indices; yield impacts are inferred via physiological thresholds, not directly simulated with crop models. - External drivers: Non-climatic stressors (pests, policy, conflict, irrigation practices) not explicitly modeled, limiting direct translation to yield or economic impacts.