Observation-constrained projections reveal longer-than-expected dry spells

Abstract

Climate models indicate that dry extremes will be exacerbated in many regions of the world^1,2. However, confidence in the magnitude and timing of these projected changes remains low^3,4, leaving societies largely unprepared^5,6. Here we show that constraining model projections with observations using a newly proposed emergent constraint (EC) reduces the uncertainty in predictions of a core drought indicator, the longest annual dry spell (LAD), by 10-26% globally. Our EC-corrected projections reveal that the increase in LAD will be 42-44% greater, on average, than 'mid-range' or 'high-end' future forcing scenarios currently indicate. These results imply that by the end of this century, the global mean land-only LAD could be 10 days longer than currently expected. Using two generations of climate models, we further uncover global regions for which historical LAD biases affect the magnitude of projected LAD increases, and we explore the role of land-atmosphere feedbacks therein. Our findings reveal regions with potentially higher- and earlier-than-expected drought risks for societies and ecosystems, and they point to possible mechanisms underlying the biases in the current generation of climate models.

Fig. 1. Observed LAD climatology and climate-model projections of future LAD change.

a, Historical LAD climatology, averaged over seven observational datasets for the period 1998–2018. Hatched areas indicate regions with high uncertainty for which the observational standard deviation exceeds 30% of the observed climatological mean. b,c, Twenty-first-century relative LAD change (as a percentage of historical mean) as predicted by CMIP6 MEM under the SSP2-4.5 (b) and SSP5-8.5 (c) scenarios. Pixels where at least 70% of models agree on the sign of the change are marked with a dot. The rectangular boxes highlight seven identified hotspot regions with the highest relative change (names are defined in the main text). d, Mean and s.d. of zonally averaged land-only historical LAD bias in the CMIP6 (red) and CMIP5 (grey) models relative to the observational median (black vertical line). The s.d. of the zonally averaged historical LAD among seven observational datasets is indicated (dotted area, d). e,f, Inter-model mean and s.d. of the magnitude of future LAD change under the SSP2-4.5/RCP4.5 (e) and SSP5-8.5/RCP8.5 (f) scenarios, where the vertical black line shows zero change.

Fig. 2. Emergent constraint on future LAD change.

a,b, Inter-model relationship between historical LAD climatological mean and projected twenty-first-century LAD change from CMIP6 (red) and CMIP5 (grey) models under SSP2-4.5/RCP4.5 (a) and SSP5-8.5/RCP8.5 (b) scenarios. Every cross (CMIP5) and dot (CMIP6) represents the global (50° S–50° N) land average of a model. The corresponding MEMs are shown as vertical dashed lines. The observational mean (blue line) and standard deviation (blue shading) are shown. The bars on the right of each graph show the mean, 66% and 90% value range of the future LAD change before (empty bars) and after (filled bars) EC correction for the CMIP5 and CMIP6 model ensembles. The specific models in each graph are defined in Extended Data Fig. 4c,d and Supplementary Tables 2 and 3.

Fig. 3. EC-corrected regional CMIP6 LAD projections.

a,b, End of the twenty-first-century MEM bias of future LAD change projections for SSP2-4.5 (a) and SSP5-8.5 (b). Regions with small (less than 10%) twenty-first-century relative change or regions with invalid LAD data are shown as white (Methods). Regions with a significant local EC correlation (R > 0.4 and P < 0.05) are shown hatched. Hotspot regions are as indicated in Fig. 1. c,d, Time series of raw (grey) and EC-corrected (coloured) annual projections of LAD change in two global hotspot regions for the two climate-change scenarios indicated. Time series are smoothed using 20-year running means. The bars on the right show end-of-century statistics. Uncertainty ranges of bars and time series represent the 66% range of model ensemble values for the end-of-century and annual statistics, respectively.

Fig. 4. Link to physical mechanisms in CMIP6 models.

a, Global statistics of inter-model correlation between historical LAD and annual number of wet periods (days with more than 1 mm of rain, r1mm), total annual rainfall (Ptot) and horizontal resolution of the model (modRes). Horizontal resolution at the Equator is estimated on the basis of longitude resolution of a corresponding atmospheric model. Violin plots show the probability density shape of correlations; box plots show the interquartile range (IQR, box) and 1.5 × IQR (whiskers). b, Distribution of predicted future LAD changes in ‘dry’ (brown) and ‘wet’ (green) models in global hotspot regions and significance of the difference between them (grey bars) according to the P-value of an Anderson–Darling (AD) test. Box plots show IQR and 1.5 × IQR of LAD change in ‘dry’ or ‘wet’ models. c, Significant difference (P-value < 0.01) in projected future change in hydro-climatic variable between ‘dry’ and ‘wet’ models according to an Anderson–Darling test. Circle size indicates the magnitude of the difference (D-statistics) between two distributions in the Anderson–Darling test. The colour shows the sign of future variable change in ‘dry’ and ‘wet’ MEM. Variables are defined in Supplementary Table 4. Future changes for all variables were assessed for local dry periods (see Methods for details).

Extended Data Fig. 1. Uncertainty of observational LAD estimates.

a,b, Global mean and standard deviation of inter-product mean absolute difference (MAD) (a) and Pearson correlation (R) (b) among 12 precipitation data sets: 11 pre-selected FROGS data sets and an independent merged TRMM–CMORPH product (see Methods and Table S1). Prior to constructing bar plots, MAD and R obtained per dataset are averaged over the global land (50° S–50° N). The 7 products which are finally selected based on MAD, R and dataset characteristics (see Methods) for further study are highlighted. c, Local standard deviation of LAD of the 7 selected precipitation products.

Extended Data Fig. 2. Characteristics of global LAD distribution.

Global land area fraction corresponding to a particular LAD climatological value range (blue dots). Cumulative population count over the same LAD range bins, calculated from either observational (black) or CMIP6 model mean (red), is shown as lines. Colour code of LAD bins corresponds to the colour bar in Fig. 1a. Population data are extracted from GPW-v4 for the year 2020.

Extended Data Fig. 3. Future LAD change.

a,b,d,e 21st century change in LAD predicted by CMIP6 (b,e) and CMIP5 (a,d) MEM under SSP5-8.5 (RCP8.5) and SSP2-4.5 (RCP4.5) scenarios. Pixels where at least 70% of models agree on the sign of the change are marked with a dot. c,f Difference in the future LAD change between CMIP6 and CMIP5 ensemble means under the SSP2-4.5 (RCP4.5) (c) and SSP5-8.5 (RCP8.5) (f) scenarios.

Extended Data Fig. 4. Linear past-future LAD relationship.

a,b, Correlation between historical and future LAD climatology in CMIP5 and CMIP6 ensembles for the ‘mid-range’ (a) and vhigh-end’ (b) emission scenarios; bar plots show mean, 66% and 90% range of future LAD statistics. c,d, EC relationship between historical LAD climatology and its 21st century change under the ‘high-end’ emission pathway scenario in CMIP5 (c) and CMIP6 (d) models. Every dot (CMIP6) and cross (CMIP5) represent the global land average of a model. Corresponding MEMs are shown as vertical dashed lines. Observational mean and uncertainty, 1 standard deviation (blue shading), is given. Model names corresponding to the numbers are given in Tables S2–S3.

Extended Data Fig. 5. EC relationship sensitivity to the time period.

a,b, Inter-model Spearman rank correlation coefficient (R) and slope of the regression line between historical LAD and its future change (+80 years from the historical) for CMIP5 (a) and CMIP6 (b) models. The regions are split to future LAD increase (red line) and LAD decrease (blue line) regions. The time series represent 20-year moving averages. Different starting dates for CMIP5 and CMIP6 are defined by data availability. c, CMIP5-based EC relationship assessed for two different historical periods of 1951–1998 (grey) and 1951–2016 (red). Here, REGEN dataset is selected as the observational dataset with the longest time-period. Every dot/cross represents the global land average of a model. Corresponding MEMs are shown as vertical dashed lines. Observational mean and uncertainty, 1 standard deviation (blue shading), is given. Bars show a mean, 66 and 90% value range of the future LAD prior (empty bar) and post (filled bar) EC correction.

Extended Data Fig. 6. Global EC strength and consistency.

a,b, Inter-model correlation between historical LAD in CMIP6 model ensemble and its corresponding magnitude of the 21st century change for (a) SSP5-8.5 and (b) SSP2-4.5 emission scenarios. Correlations higher than 0.4 (p-value < 0.05) have brighter colour. Regions with a non-small relative future LAD change (>10% of historical value) are hatched. c, Global statistics of correlations between historical LAD and future LAD change under SSP5-8.5 separated per p-value range are shown as box plots. Box plots show interquartile range (IQR, box) and 1.5 times IQR (whiskers) of data.

Extended Data Fig. 7. Sensitivity of EC-corrected future LAD change bias to observational spread.

a,b, Future LAD change bias calculated using mean value of either three observational data sets with lowest global mean LAD, i.e. GPCP, TRMM-CMORPH, PERSIANN (see Methods) (a) or three observational data sets with highest global mean LAD, i.e. REGEN, GPCC, GSMAP (see Methods) (b) for EC correction. Masked in white are the regions with small (<10%) 21st century relative change or regions with invalid LAD data (see Methods). Regions with a significant local EC correlation (R > 0.4 and p-value < 0.05) are hatched. Boxes show two regions where change in sign between two groups is observed. c, Difference in historical LAD climatological mean between two observational groups. d, Difference in EC-corrected future LAD change bias between both groups.

Extended Data Fig. 8. Calibrated future LAD projections for CMIP6 models.

a,b End-of-century raw and EC-corrected MEM LAD change averaged per hotspot region under (a) SSP2-4.5 and (b) SSP5-8.5 scenarios. Bar plots show the 66% of model ensemble values for raw (grey) and EC-corrected (red) LAD change projections. c, Spatial shift (before and after EC correction) in global areas of 2 to 4 month climatological LAD under the SSP5-8.5 scenario for the end of the century. e, EC-corrected future (SSP5-8.5) LAD climatology for 2080–2100. Hatched areas show regions with a significant local EC correlation (R > 0.4 and p-value < 0.05). d,f,g, Time-series of raw (grey) and EC-corrected (colour) annual projections of LAD change in three global hot-spot regions for two emission scenarios. Time-series are smoothed using 20-year running mean. Bars and time-series uncertainty ranges represent the 66% of model ensemble values for the end-of-century and annual statistics, respectively.

Extended Data Fig. 9. Local correlation of LAD to precipitation and global consistency in EC correlation.

a,b, Pearson correlation between historical LAD and total annual number of (a) wet days (rain <1 mm/day), and (b) total annual precipitation. c-f, Inter-model correlation of historical CMIP6 LAD climatology in a pixel from (c) Australia, (d) East Africa, (e) South Atlantic Convergence Zone (SACZ), and (f) CE-Asia to the historical CMIP6 LAD scatter in rest of the world. Note, SACZ and CE-ASIA appear as locally correlated features decoupled from the rest of regions with high local EC correlation.

Extended Data Fig. 10. Inter-model correlation of historical LAD to the corresponding future projected change in hydro-climatic variables.

The value of Pearson correlation assessed for every domain separately is shown in colour. Correlations significant at p-value < 0.01 level are marked with a cross. Names of variables are decoded in Table S4. Future changes for the variables are assessed during local dry periods (see Methods).

Extended Data Fig. 11. Differences in future projections of hydro-climatic variables between ‘dry’ and ‘wet’ models.

a,b, KDE-based distribution frequencies of future projected changes in selected hydro-climatic variables in CMIP6 models (see Methods) following the SSP5-8.5 scenario for locally defined ‘dry’ (red) and ‘wet’ (grey) models in regions of future LAD (a) increase and (b) decrease. Each distribution is built on model simulated data, which can include up to ten ensemble members, depending on their availability (see Methods). Shaded areas show the +/− 1 standard deviation data range; dotted line shows the median value.

Extended Data Fig. 12. Sensitivity of the EC relationship to the minimum daily rainfall threshold, model re-gridding procedure and future uncertainty estimation approach.

a-c, KDE-based frequency distribution across global land (50°S–50°N) daily rain rates in CMIP6 models (grey) and GPCP observations (red) (see Methods) over 1998–2014. Three panels represent sensitivity of results to data being upscaled to different grid resolutions. d, Global EC-relationship estimated for CMIP6 models at their original grid resolution. e, Globally averaged future LAD change projections before (empty bar), and after (filled bars) EC-correction using either PI-based approach as in Fig. 2 (gray filled bar) or K-L divergence approach (orange filled bar). 66 years of REGENAllSat dataset (see Methods) and CMIP5 data of the same period are used to estimate the mean and variance of every model and observational data, and then to estimate weights to correct the future model projections using K-L divergence approach. In d,e, every dot/cross represents the global land average of a model. Corresponding MEM is shown as vertical dashed line. Blue shading shows the standard deviation of observed LAD climatology across observational data sets (d) and 66 years of REGENAllSat (e). Bars show the mean, 66 and 90% range of the future LAD prior (empty bar) and post (filled bar) EC correction.