More extreme precipitation in the world s dry and wet regions Markus G. Donat, Andrew L. Lowry, Lisa V. Alexander, Paul A. O Gorman, Nicola Maher Supplementary Table S1: CMIP5 simulations used in this study. Model Name Historical simulations (1951-2005): Number of Runs (and run identifier) RCP4.5 simulations (2006-2099): Number of Runs (and run identifier) RCP8.5 simulations (2006-2099): Number of Runs (and run identifier) ACCESS1-0 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) bcc-csm1-1-m 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) bcc-csm1-1 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) BNU-ESM 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) CanESM2 5 (r1i1p1, r2i1p1, r3i1p1, r4i1p1, r5i1p1) 5 (r1i1p1, r2i1p1, r3i1p1, r4i1p1, r5i1p1) 5 (r1i1p1, r2i1p1, r3i1p1, r4i1p1, r5i1p1) CCSM4 r6i1p1) r6i1p1) r6i1p1) CMCC-CM 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) CNRM-CM5 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) CSIRO-Mk3-6-0 FGOALS-s2 GFDL-CM3 1 (r1i1p1) 0 1 (r1i1p1) GFDL-ESM2G 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) GFDL-ESM2M 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) HadGEM2-CC 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) HadGEM2-ES 4 (r1i1p1, r2i1p1, r3i1p1, r4i1p1) 3 (r2i1p1, r3i1p1, r4i1p1) 4 (r1i1p1, r2i1p1, r3i1p1, r4i1p1) inmcm4 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) IPSL-CM5A-LR 4 (r1i1p1, r2i1p1, r3i1p1, r4i1p1) 4 (r1i1p1, r2i1p1, r3i1p1, r4i1p1) 4 (r1i1p1, r2i1p1, r3i1p1, r4i1p1) IPSL-CM5A-MR 1 (r1i1p1) 0 1 (r1i1p1) IPSL-CM5B-LR 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) MIROC-ESM-CHEM 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) MIROC-ESM 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) MIROC5 MPI-ESM-LR MPI-ESM-MR 1 (r1i1p1) MRI-CGCM3 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) NorESM1-M 1 (r1i1p1) 1 (r1i1p1) 1 (r1i1p1) NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 1
SI1. Sensitivity to different thresholds to identify the dry and wet regions We tested the sensitivity of our results to the specific threshold applied for the definition of the dry and wet regions. Observed and modelled changes for the historical period are very similar regardless of whether 20, 30, or 40 percent of the wettest/driest grid cells are used (compare Supplementary Figures S1 and S2 to Figure 1, Supplementary Figures S3 and S4 to Figure 2). The ensemble spread is somewhat larger, in particular for the dry-regions average, when we use fewer grid cells (e.g. the highest and lowest 20%) to define the wet and dry regions. Also the future projected changes show very similar ensemble mean differences regardless of threshold. But again the inter-model variability in particular for the dry-regions average increased when fewer grid cells are used (Supplementary Figure S5). Similarly, the linear regression of precipitation changes with global temperature increases results in higher p-values when averages over smaller regions (20%) are considered (Supplementary Figures S6 and S7). The slope values are however very similar regardless of whether we analyse the wettest and driest 20, 30, or 40 % of grid cells. 2 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S1: As Figure 1 but for the wettest and driest 20 percent of grid cells. Time series of PRCPTOT (annual precipitation totals) and Rx1day (the annual-maximum daily precipitation) for dry (a) and wet (b) regions in the HadEX2 observational dataset. Area-weighted average time series are shown for HadEX2 and the ensemble mean and spread of CMIP5 simulations. Precipitation indices were first normalized by calculating annual values as a fraction of the 1951-1980 local mean before calculating the dry and wet regions averages. Black lines: annual values from observations and ensemble mean, red lines: linear trend, blue dashed lines: 30-yr averages 1951-1980 and 1981-2010, grey shading: +/- one ensemble standard deviation. dprcptot and drx1day indicate the difference between the averages during 1981-2010 and 1951-1980, Slope is the linear trend Sen-slope estimate (see Methods, unit: decade -1 ), and the p-value the trend significance using a Mann-Kendall test. The masks (c) indicate the locations of the grid cells contributing to the average of the dry and wet regions, and the number n of grid cells contributing to the area-averages of dry and wet regions is given. Land grid cells that are less complete than 90% of the years 1951 to 2010 (see Methods) are marked in grey and excluded from this analysis; the masks with limited observational coverage have also been applied to the CMIP5 simulations. Area Fraction indicates the fraction of land area for which data are at least 90% complete. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 3
Figure S2: As Figure 1 and Figure S1 but for the wettest and driest 40 % of grid cells. 4 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S3: As Figure 2 but for the wettest and driest 20 percent of grid cells. Time series of normalized PRCPTOT and Rx1day for dry (a) and wet (b) regions in the ensemble of CMIP5 simulations. The dry and wet grid cells were identified both for land regions and for all regions (land and ocean, excluding south of 60 S). Line colours and terminology as in Figures 1 and S1. The masks (c) indicate counts for each grid cell of how many simulations (out of 46) have a driest/wettest grid cell locally. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 5
Figure S4: As Figure 2 and Figure S3 but for the wettest and driest 40 % of grid cells. 6 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S5: As Figure 2 but for the wettest and driest 20 (left panel) and 40 (right panel) percent of grid cells. Simulated precipitation changes in dry and wet regions. Top row: PRCPTOT, bottom row: Rx1day, left column in each panel: precipitation averaged over the dry regions, right column in each panel: precipitation averaged over the wet regions. Historical changes show the average during 1981-2010, RCP4.5 and RCP8.5 the average during 2070-2099, all relative to the 1951-1980 average. Horizontal black lines represent the ensemble mean changes, the coloured boxes show +/- one ensemble standard deviation. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 7
Figure S6: As Figure 4 but for the wettest and driest 20 percent of grid cells. Scatter plots of precipitation index changes (top row: PRCPTOT, bottom row: Rx1day) over land against global mean temperature change in the different CMIP5 simulations. Left column: precipitation averaged over the dry regions, right column: precipitation averaged over the wet regions. Changes in RCP4.5 (2070-2099 minus 1951-1980) are shown in blue, and RCP8.5 (2070-2099 minus 1951-1980) in red. Dots represent the differences in the individual model runs, solid lines the linear regression best fit through all points of a scenario, dashed lines the 5-95% uncertainty range of the linear regression slope. Regression lines are only plotted if the associated p-value is below or equal to 0.1. 8 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S7: As Figure 4 and Figure S6 but for the wettest and driest 40 % of grid cells. SI2. Sensitivity to the ensemble construction: Weighting all models equally by using only one run from each model Different numbers of simulations are available from the different models, and therefore each model has a different weight in the ensemble average results presented in the main text. There are different ways to achieve equal weighting of all models. For example, sub-ensemble means averaging all runs of one model could be first calculated before calculating the ensemble mean over the sub-ensemble means from all models. However, this approach would reduce the noise related to climate variability that is included in each simulation, and therefore lead to artificially strong signals and would limit comparability of the model results to observations (of which only one climate realization exists). Therefore, we tested the robustness of our results to having equal weights for the different models by repeating all ensemble calculations using only one run (run identifier r1i1p1) from each model. We find that the results are virtually identical for the historical period (Figures S8, S9), regardless of whether exactly one run or several runs are used from the different models. Also the future projections show very similar changes in comparison to the ensemble results that include several NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 9
runs per model (Figure S10). Analyzing possible relationships between precipitation and global temperature changes, statistically significant linear regressions are found for the same cases, i.e. PRCPTOT and Rx1day changes in the dry regions, as those found in the larger ensemble including several runs per model. Slope values are about 1.5 to 2 % higher, but this within the expected range of uncertainty related to using fewer data points. Figure S8: As Figure 1 but only using one run (run identifier r1i1p1) from each model. Colours and terminology as in Figures 1 and S1. 10 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S9: As Figure 2 but only using one run (run identifier r1i1p1) from each model. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 11
Figure S10: As Figure 3 and Figure S5 but only using one run (run identifier r1i1p1) from each model. 12 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S11: As Figure 4 but only using one run (run identifier r1i1p1) from each model. SI3. Changes in tropical and extra-tropical latitudes Precipitation characteristics may be different between high and low latitudes, due to different weather systems at play and very different specific humidity. Therefore, we tested the precipitation changes separately for the dry and wet grid cells that are located in tropical and sub-tropical (30 S- 30 N) latitudes, and in extra-tropical (north of 30 N or south of 30 S) latitudes. Precipitation changes for wet and dry grid cells in extra-tropical latitudes are very similar to the global changes presented in the main text (Supplementary Figures S12, S14). However, for grid cells located in tropical latitudes the results are less robust, and no significant changes are found based on observations (Figure S13). Note that observational coverage is very limited in the 30 S to 30 N latitude band. For example, for Rx1day only thirteen grid cells contribute to each wet and dry-region time series. Using complete coverage from the CMIP5 simulations, the results in both the tropical and extra-tropical regions are qualitatively similar to the global analyses (Supplementary Figures S14, S15), showing total precipitation increases in the dry regions and only small changes in the wet regions while precipitation extremes increase in both the dry and wet tropical regions. If only one run is used for each model, then there is a decreasing trend in PRCPTOT for the dry regions when only tropical oceans are considered, consistent with supplemental ref. 1, but we find that this decreasing trend is not robust to the inclusion of more model runs (not shown). NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 13
The future projections on average over the wet and dry grid cells in extra-tropical latitudes are very similar to the global analysis (Supplementary Figure S16 (left) in comparison to Figure 3), but the ensemble spread is generally smaller in the extra-tropics. Also the projections for the tropical dry and wet regions show similar ensemble mean changes as the global analyses, but inter-model spread is very large in particular for the PRCPTOT dry-areas average when ocean grid cells are also included (Supplementary Figure S16, right). Accordingly, the linear regression of precipitation with global temperature changes is highly significant for extra-tropical grid cells, while in the tropical regions only the increase of PRCPTOT in the dry regions shows significant regression relationship (Supplementary Figures S17, S18). Figure S12: As Figure 1 but for wet and dry regions that are located in extra-tropical latitudes north of 30 N and south of 30 S. 14 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S13: As Figures 1 and S12, but for the dry and wet grid cells that are located in the tropical latitude band 30 S to 30 N. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 15
Figure S14: As Figure 2 but for wet and dry regions that are located in extra-tropical latitudes north of 30 N and south of 30 S. 16 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S15: As Figures 2 and S14, but for the dry and wet grid cells that are located in the tropical latitude band 30 S to 30 N. Figure S16: As Figure 3 but for wet and dry regions that are located in extra-tropical latitudes north of 30 N and south of 30 S (left panel) and in the tropical latitude band 30 S to 30 N (right panel). NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 17
Figure S17: As Figure 4 but for wet and dry regions that are located in extra-tropical latitudes north of 30 N and south of 30 S. 18 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S18: As Figures 4 and S17, but for the driest and wettest 30% of grid cells that are located in the tropical latitude band 30 S to 30 N. SI4. Extreme precipitation changes in dry and wet regions identified from total precipitation The dry and wet regions masks for PRCPTOT and Rx1day are to a large extent similar but also show some differences (see Figure 1, 2). As a further test of robustness we also calculated the Rx1day changes using the dry and wet region masks derived from PRCPTOT. This approach leads to increased variability in the dry region observational time series and also increased inter-model spread when masked to grid boxes with observational coverage (compare Figure S19 to Figure 1 bottom panel). Using complete coverage from the CMIP5 models, the Rx1day increase in the dry regions is slightly smaller (about 80% using land and ocean grid cells and about 90% over land) when using the dry grid cells mask identified for PRCPTOT as compared to the Rx1day mask (compare Figure S20 to Figure 2 bottom panel). Changes in the wet regions average are almost identical regardless of which mask is used. The future projected changes are almost identical between both masks (Figure S21 in comparison to Figure 3 bottom row). Accordingly, when relating the Rx1day increases to the simulation-specific NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 19
temperature increases (Figure S22), also the slope values are almost identical between the results based on the different masks. Figure S19: As Figure 1 (bottom row), but Rx1day averaged over the wet and dry regions identified for PRCPTOT from the HadEX2 dataset. 20 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S20: As Figure 2 (bottom row), but Rx1day averaged over the wet and dry regions identified for PRCPTOT from each CMIP5 simulation. Figure S21: As Figure 3 (bottom row), but Rx1day averaged over the wet and dry regions identified for PRCPTOT from HadEX2 (Observational Grid Boxes) and each CMIP5 simulation (Land and Land+Ocean). NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 21
Figure S22: As Figure 4 (bottom row), but Rx1day averaged over the wet and dry regions identified for PRCPTOT from each CMIP5 simulation. SI5. Future projected precipitation changes accounting for shifts in the location of the wet and dry regions The geographical distribution of rainfall may also change in a warming climate, and the past wet and dry regions may not necessarily be the wet and dry regions in the future climate. Given the uncertainties about future changes in wet and dry regions, presenting changes for the past wet and dry regions seems more useful for adaptation purposes and is therefore discussed in the main text. Here, however, to account for possible spatial shifts in the wet and dry regions, we also present changes for grid cells that are among the 30% driest/wettest grid cells in both the past (1951-1980) and future (2080-2099) climate periods. This approach leads to generally very similar results to those presented in the main text, but smaller ensemble spreads in the future projections. A smaller spread is expected as the dry regions average cannot be inflated (towards higher precipitation values) by including grid cells that are no longer in the dry regions, and similarly for the wet regions average this approach avoids including grid cells that are no longer among the wettest 30 percent. This also explains that the ensemble-average amplitude of increase in the dry regions is mostly smaller compared to the results on the past mask only, with mean changes in Supplementary Figure S23 being only about 60% to 90% (depending on region mask and measure) of the changes in Figure 3. For the wet regions average the ensemble mean change for Rx1day is up to 15% larger if also the future mask of wet and dry grid cells is taken into account (Supplementary Figure S23 in comparison to Figure 3, and Supplementary Figure S24 in comparison to Figure 4). 22 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S23: As Figure 3, but using only grid cells that are included in the past (1951-1980) and future (2070-2099) masks of dry and wet regions. NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 23
Figure S24: As Figure 4, but using only grid cells that are included in the past (1951-1980) and future (2070-2099) masks of dry and wet regions. SI6. Precipitation measures over all global grid cells as a function of global temperature increases Here we present global average precipitation changes (over all land only grid cells and over all land and ocean grid cells, north of 60 S) as a function of global mean temperature changes. Projected total annual precipitation changes over all global land grid cells show a statistically significant linear regression with the global temperature increase (Supplementary Figure S25 top left), and the regression lines show slopes between 5 and 7 %K -1. When we perform the same analysis also including ocean grid cells (Supplementary Figure S25 top right), we obtain slope values of about 2 to 3 %K -1, comparable with previous studies 2,3, however the linear fits are not statistically significant (p>0.1). Similarly Allen and Ingram 2, based on CMIP2 models (their Figure 2), find rather scattered dp vs dt changes in the transient simulations, whereas the signal is stronger (and the linear fit significant) when they use equilibrium CO 2 -doubling experiments. Global average Rx1day changes, both over land only and when including ocean grid cells, also show significant linear relationships with global temperature changes, with regression slopes between roughly 3 and 5 %K -1 (Supplementary Figure S25, bottom). 24 NATURE CLIMATE CHANGE www.nature.com/natureclimatechange
SUPPLEMENTARY INFORMATION Figure S25: Scatter plots of precipitation indices changes over all global land (left) and global land and ocean (right) grid cells as a function of global mean temperature change in the different CMIP5 simulations. Top row: PRCPTOT, bottom row: Rx1day. Changes in RCP4.5 (2070-2099 minus 1951-1980) are shown in blue, and RCP8.5 (2070-2099 minus 1951-1980) in red. Dots represent the differences in the individual model runs, solid lines the linear regression best fit through all points of a scenario, dashed lines the 5-95% uncertainty range of the linear regression slope. Regression lines are only plotted if the associated p-value is below or equal to 0.1. References 1. Allan, R. P., Soden, B. J., John, V. O., Ingram, W. & Good, P. Current changes in tropical precipitation. Environ. Res. Lett. 5, 025205 (2010). 2. Allen, M. R. & Ingram, W. J. Constraints on future changes in climate and the hydrologic cycle. Nature 419, 224 32 (2002). 3. Lambert, F. H. & Webb, M. J. Dependency of global mean precipitation on surface temperature. Geophys. Res. Lett. 35, L16706 (2008). NATURE CLIMATE CHANGE www.nature.com/natureclimatechange 25