The most at-risk regions in the world for high-impact heatwaves

The study addresses where in the world regional daily maximum temperature records (TXx) are statistically likely to be exceeded, identifying communities most at risk from unprecedented heatwaves. The context is that record-shattering heat events, exemplified by the June 2021 western North America heatwave, can cause severe societal and environmental impacts, yet planning and adaptation often lag until after such events occur. Estimating the likelihood of rare extremes is challenging due to limited observational records and potential nonstationarity from anthropogenic climate change. The authors motivate a global assessment using extreme value theory (EVT) to evaluate return periods of current regional heat records and to determine which regions may have been comparatively ‘lucky’ so far and thus are potentially underprepared. They justify the use of TXx (annual maximum of daily maximum temperature), recommended by the WMO for assessing heatwaves, while noting alternative metrics (e.g., thresholds, percentiles, humidity-inclusive indices, nighttime temperatures) may be more impact-relevant in specific regions. The purpose is to quantify statistical headroom beyond current records and highlight regions where unprecedented extremes are both likely and potentially high-impact due to limited preparedness and growing exposure.

The paper situates its work within literature on extreme heat events and statistical attribution. Prior studies have shown the devastating impacts and rarity of events like the 2021 Pacific Northwest heatwave and the 2003/2006 European heatwaves, and the benefits of heat action plans in reducing mortality. EVT is widely used to assess probabilities and return periods of meteorological extremes in observations and models, including nonstationary approaches that relate extremes to global mean surface temperature (GMST). However, the literature notes pitfalls: potential nonlinearity and regional variability in relationships with GMST; local factors (soil moisture feedbacks, aerosols, irrigation) that complicate global application; dependence in data due to decadal variability; and limited sampling of rare mechanisms, which complicates extrapolation to very long return periods. Reanalysis datasets (e.g., ERA5, JRA-55) are commonly used but have known discrepancies, particularly in data-sparse regions (e.g., parts of Africa). Model-based large ensembles offer an alternative for exploring very rare events and understanding statistical ‘implausible’ extremes. The authors build on region-definitions designed for climate extremes to align with policy-relevant scales.

- Data sources: Reanalysis datasets ERA5 (1959–present) and JRA-55 (1958–present) provide gridded daily maximum temperature. Consistency between reanalyses is required: for each region (assessed over 1990–2022), the ERA5 record year must appear within the top five JRA-55 record years to be included. GMST is taken from NASA GISS. - Regions: Predefined 0.5 Mm² political/economic regions from Stone (2019) are used to align with impacts and policy scales. From 237 initial regions, Antarctic regions are removed (217), and after reanalysis consistency screening, 136 regions remain. - Heat metric: TXx, the annual maximum of daily maximum temperature, is used (WMO-recommended). A block-maxima approach avoids threshold subjectivity and clustering biases. - EVT framework: Generalized Extreme Value (GEV) distributions are fitted to TXx time series per region. Scale and shape parameters are assumed constant. To account for nonstationarity due to warming, the location parameter is adjusted via a linear relationship with GMST; regional data are adjusted to GMST = 1°C (approximate current climate). - Record exclusion approach: For each region, the statistical fit excludes the record year to assess the return period of the observed record and to estimate a ‘statistical maximum’. Data from years after the record are retained to better constrain the distribution. Uncertainty is quantified via bootstrap resampling (100 iterations) to derive 5th–95th percentile ranges. - Statistical maximum: Defined operationally as the 1-in-10,000-year return level from the GEV fit (excluding the record year), approximating the asymptotic maximum; larger return periods change results minimally. - Case study: The Alberta (Canada) region illustrates the approach around the June 2021 western North America heatwave. The 2021 TXx is shown to lie beyond the GEV fit; the prior (2018) record corresponds to ~166-year return period after GMST adjustment. - Climate model large ensembles: Historical simulations (1950–2014) from CanESM5 and MIROC6 (50 realizations each) are analyzed. Each model’s TXx is adjusted by its ensemble-mean GMST to align trends and reduce internal variability. Two techniques are used: (1) per-member GEV fits producing 50 fits per region; (2) merged-member fit combining all years from all members per region. Both techniques assess how often current records fall outside GEV fits and compute return periods where possible. - Population and exposure: The study links statistical likelihood to exposure using 2020 population data and projected growth to 2050 (SSP5) to highlight regions where limited preparedness and growing exposure may amplify impacts. - Quality control and caveats: Regions with inconsistent ERA5/JRA-55 extremes are excluded from the main global assessment; differences between reanalyses are documented. Spatial scale considerations and model resolution limitations are acknowledged.

- Western North America 2021 event: The June 2021 heatwave (e.g., Lytton, 49.6°C) was far beyond the GEV-estimated plausible range for Alberta when excluding 2021, consistent with prior rapid attribution findings. The prior (2018) Alberta record had an estimated return period of ~166 years (after GMST adjustment), implying the 2021 event was exceptionally beyond expectations. - Global reanalysis assessment (1959–2021): Among 136 regions with ERA5/JRA-55 consistency, 41 regions (31% of land area considered) have current records that fall outside the GEV fit (i.e., statistically implausible under the fitted distribution excluding the record). These regions show no clear spatial clustering and occur throughout the record, with more events in later decades possibly due to data availability and/or nonlinear climate change. - Statistical headroom: The 2021 western North America heatwave stands out as nearly 2°C further beyond the statistical maximum compared with any other region analyzed. Some regions have not experienced events beyond a 1-in-100-year magnitude within the 62-year period, implying low sampling of the upper tail and vulnerability to future record-shattering events. - Regions with most likely near-term records: Eight regions have record return periods below ~100 years (adjusted to current climate), indicating high likelihood of breaking records soon: Far Eastern Russia (70.6y), Central America (78.1y), Afghanistan (83.9y), Papua New Guinea (89.6y), Central Europe (91.4y), Northwestern Argentina (91.7y), Queensland, Australia (94.2y), and Beijing region, China (99.8y). These regions often have minimal headroom between current records and 1-in-100-year levels (0.1–0.5°C), and some face rapid population growth (e.g., Afghanistan, Central America). - Climate model ensembles: Using 50-member ensembles (1959–2015), ‘implausible’ records occur in at least one realization for every region globally, with up to 50% of members in some regions. Across >10,000 fits, exceptional extremes occur in 26% of regions (CanESM5) and 24% (MIROC6) when fitting per member; when merging members, 18% (CanESM5) and 22% (MIROC6) of regions still show exceptional extremes. Spatial patterns differ from reanalysis but indicate susceptibility worldwide. This supports the conclusion that any region can experience record-shattering events beyond the fitted statistical distribution. - Preparedness and exposure: Regions with limited health and energy infrastructure and rapid population growth (e.g., Afghanistan and parts of Central America) are particularly vulnerable despite having lower observed records, underscoring the need for proactive heat action planning.

The analysis demonstrates that many regions have current heat records that are either statistically ‘implausible’ under standard EVT fits or have short return periods, implying that unprecedented events could occur virtually anywhere and at any time. This directly addresses the research aim by identifying regions where current records provide poor sampling of the tail and thus where communities may be underprepared for truly extreme events. The June 2021 western North America event exemplifies the potential for rare dynamics and nonlinear processes to produce extremes far exceeding statistical expectations. The lack of strong spatial/temporal clustering among implausible records in reanalysis, coupled with confirmation from model ensembles, emphasizes the broad exposure to risk. The policy significance is clear: planning based on historical records alone is insufficient; heat action plans must anticipate events beyond observed maxima. Socioeconomic context modulates risk: developing regions with limited resources and rapid population growth are likely to see disproportionate impacts even if the statistical likelihood is similar. The findings support a precautionary approach to adaptation and resilience planning that accounts for record-shattering possibilities and considers alternative metrics (e.g., humidity, multi-day heat) that better reflect impacts.

The study identifies regions where record-breaking heat extremes are statistically more likely due to poorly sampled upper tails of the temperature distribution and shows that statistically implausible extremes have already occurred in 31% of analyzed regions. Climate model large ensembles corroborate that exceptional extremes can occur virtually anywhere (18–26% of regions depending on method), reinforcing the need for universal preparedness. The work highlights several regions (e.g., Central America, Afghanistan, Beijing region, parts of Europe and Australia) where new records are relatively likely in the near term and where rapid population growth may exacerbate vulnerability. Future research should: (1) improve understanding of the atmospheric and land-surface processes behind the most extreme heatwaves and potential nonlinearities; (2) refine statistical frameworks for nonstationary extremes beyond linear GMST relationships; (3) leverage large ensembles, proxy-region approaches, and ensemble boosting/reinitialization to better quantify plausible physical maxima; and (4) integrate multiple heat metrics (humidity, multi-day duration, nighttime temperatures) to inform impact-relevant adaptation planning.

- EVT assumptions: The approach assumes constant scale and shape parameters and a linear relationship between the location parameter and GMST; these may be regionally invalid or nonlinear due to processes like soil moisture–temperature feedbacks, aerosols, or irrigation. - Independence and stationarity: EVT requires independent block maxima; multi-year variability and decadal modes could introduce dependence, affecting fit quality. - Extrapolation uncertainty: Historical observations may not sample mechanisms producing the most extreme events, making extrapolation to 1-in-10,000-year levels uncertain, and some observed events fall beyond fitted distributions. - Record exclusion bias: Excluding the record year aids sensitivity assessment but introduces selection bias; including it may bias return periods differently. - Reanalysis discrepancies: ERA5 and JRA-55 differ, especially over data-sparse regions (notably parts of Africa), leading to exclusion of many regions and potential biases. - Metric and scale limitations: Only TXx is analyzed; other impact-relevant metrics (e.g., humidity-informed indices, multi-day heat, nighttime heat) may tell different risk stories. Spatial aggregation (0.5 Mm²) may dampen local extremes; model coarse resolution (CanESM5, MIROC6) may limit fidelity. - Temporal coverage: Analyses focus on 1959–2021 (reanalyses) and 1950–2014 (models); changing dynamics post-2014/2021 are not directly captured.