GoKawiilData
We used climate data from the Coupled Model Intercomparison Project Phase 6 (CMIP6)51 and combined the historical scenario (1851–2014) with the Shared Socioeconomic Pathway SSP5-8.5 from 2015 onwards (Extended Data Table 1). Using SSP5-8.5 allows for inspecting warming levels from moderate (2 °C) to extreme (4 °C) with the same consistent dataset. We also used the SSP2-4.5 to test the sensitivity of the results at 2 °C of global warming and found that—except for fisheries—results derived from different SSPs are similar (Extended Data Fig. 11).
We derived time series of annual maximum 5-day precipitation (from daily data; CMIP6 identifier of the used variable: day, pr), annual mean soil moisture over the total column (from monthly data; Lmon, mrso), annual maximum of daily maximum temperature (from daily data; day, tasmax), annual mean of sea surface temperature (from monthly data; Omon, tos), and—for identifying global warming levels—global-mean temperature (from monthly data; Amon, tas). Successively, gridded time series were interpolated to an equal 2.5° spatial grid before the next analyses with bilinear remapping for tasmax and tos (for tos, conservative remapping is used for the model AWI-CM-1-1-MR owing to an original unstructured grid type), and conservative remapping for mrso and pr (no interpolation is applied to tas). For the analysis of the FWI56,57, we used the already available annual maxima of daily FWI at 2.5° spatial grid from ref. 53, computed based on daily precipitation, wind (day, sfcWind), relative humidity (day, hurs), and maximum temperature. For all analyses, fields were interpolated to a finer grid at the end of the analyses for graphical purposes. We used one ensemble member per model: the r1i1p1f1 member when available for both the historical and the considered SSP scenario; otherwise the first ripf member in alphabetical order that was available for both historical and SSP (the ripf is an index used in CMIP6 to uniquely identify ensemble members of a given model, where r, i, p and f denote the realization, initialization method, physics parameterization and forcing index, respectively). Only when quantifying the contribution of internal climate variability to the uncertainties in projections, we used SMILEs58,59 (Extended Data Table 2), that is, the numbers reported in the text are unaffected by the SMILEs. We used CanESM5 and MIROC6 (each with 50 ensemble members available for all analysed sectors), which are shown as vertical lines on the left and right, respectively, in Figs. 2b, 3b and 4b and Extended Data Figs. 7b and 8b; these and additional SMILEs with fewer ensemble members are shown in Extended Data Fig. 2. For each sectoral analysis individually, we selected CMIP6 model ensemble members with data starting at least in 1851 both for the specific sectoral variable and the global-mean temperature required to identify global warming levels. The resulting set of models is shown in Extended Data Tables 1 and 2. We analysed individual ensemble members of SMILEs independently, which allows for consistent comparison with results from models with a single ensemble member. As we focus on climate change signals, in line with IPCC practice, we do not apply bias correction as results are not expected to be substantially affected; but users interested in projecting impacts should use bias-corrected data from individual climate outcomes.
Global warming levels
For each model ensemble member and target warming level, we selected the earliest 30-year window whose area-weighted global-mean temperature exceeds the area-weighted preindustrial (1851–1900) global-mean temperature by at least that target warming level. The same procedure was applied to the main time series used in the analyses—based on SSP5-8.5—and to those based on SSP2-4.5 used in Extended Data Fig. 11 to test the sensitivity of the results at 2 °C of global warming to the SSP. It is noted that the models used for analyses at different warming levels can vary because some models do not reach some of the highest warming levels.
Identifying extreme climate outcomes
For each sector, once the global climatic impact-driver f for climate outcomes is computed (equation (1)), see details below), models with f values below and above the 8th and 92nd percentiles (for wildfires 12th and 88th percentiles given that fewer models are available for FWI) are selected as best and worst-case climate outcomes, respectively (Fig. 1); here and elsewhere, percentiles are computed with the R-function quantile with the recommended60 algorithm type 8.
For the precipitation-related analysis, we derived highly populated areas in 2020 based on the GPWv4 population dataset52. We first obtained population density at each grid point (by dividing the population by the grid-point surface) and sorted grid points in terms of population density. Then, we derived the smallest land surface that includes 90% of the global population by selecting the first N grid points whose aggregated total population covers 90% of the global population. Then, only grid points where no models (among the runs used for the 2 °C projections under SSP5-8.5) show an average Rx5day during the preindustrial period below 25 mm (5 mm d−1) are retained. This avoids spuriously large percentage changes of Rx5day (ΔRx5day (%)) in very dry regions, which could disproportionately influence the global climatic impact-driver f (it is noted that the worst-case model at 2 °C exceeds the multimodel mean at 3 °C also without this filtering). To compute f, for each model, we first re-gridded ΔRx5day (%) (the percentage difference in the time-mean of annual maxima consecutive 5-day precipitation between the future and the preindustrial periods) to the population dataset’s grid via the nearest-neighbour approach and then computed the average (weighted by grid cell area) over the retained highly populated areas. It is noted that, in line with previous studies, we analyse local precipitation extremes as a proxy for local pluvial floods, assuming co-location and neglecting possible non-local flooding drivers.
For the drought analysis, we derived the total global breadbasket as the union of maize, wheat, soybean and rice breadbaskets from ref. 25. We computed f for each model as the average (weighted by grid cell area) of ΔDrought frequency (%) (the difference in the drought frequency between the future and the preindustrial periods) across grid points in the global breadbasket (where all models provide data). We defined grid-point droughts as annual average soil moisture low values occurring every 5 years on average (below the 20th percentile) in the preindustrial period. In Fig. 3a, Greenland, Antarctica and Iceland are excluded because these regions contain many locations for which not all models provide data.
For the FWI analysis, we used forest grid cells derived from the European Space Agency Land Cover dataset54,55,61, which was first re-gridded to a 0.25° spatial grid using the LC-CCI user tool as in ref. 62. Then, forest grid cells were defined where the sum of tree densities (broadleaf and needleleaf, both evergreen and deciduous) exceeds the sum of the density of all other classes for all years during 2010–2019. It is noted that the worst-case models at 2 °C exceed the multimodel mean at 3 °C even with a less restrictive forest definition—namely, where the sum of tree densities exceeds 20% of the sum of the density of all classes (excluding water) for all years during 2010–2019. To compute f, for each model, we first re-gridded ΔFWIx (the difference in the time-mean of annual maxima daily FWI between the future and the preindustrial periods) to the re-gridded land-cover dataset using the nearest-neighbour approach and then computed the spatially weighted average over forest grid cells.
... continue reading