ARTICLE https://doi.org/10.1038/s41467-019-08745-6 OPEN State-of-the-art global models underestimate impacts from climate extremes Jacob Schewe et al.# Global impact models represent process-level understanding of how natural and human systems may be affected by climate change. Their projections are used in integrated 1234567890():,; assessments of climate change. Here we test, for the first time, systematically across many important systems, how well such impact models capture the impacts of extreme climate conditions. Using the 2003 European heat wave and drought as a historical analogue for comparable events in the future, we find that a majority of models underestimate the extremeness of impacts in important sectors such as agriculture, terrestrial ecosystems, and heat-related human mortality, while impacts on water resources and hydropower are over- estimated in some river basins; and the spread across models is often large. This has important implications for economic assessments of climate change impacts that rely on these models. It also means that societal risks from future extreme events may be greater than previously thought. Correspondence and requests for materials should be addressed to J.S. (email: schewe@pik-potsdam.de). # A full list of authors and their affiliations appears at the end of the paper. NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 E stimation of the total damages caused by climate change Results requires a quantification of climate impacts across a large The 2003 European heat wave and drought. The 2003 EHWD range of economic and societal sectors. These sectors stretched over the entire summer, with large and persistent hot include agriculture1, water resources2, energy supply and anomalies especially during June and August (Supplementary demand3, human health4 and ecosystem services5. There are Fig. 1), and it extended across much of Western and Central approaches that integrate damages across sectors, such as Europe (Supplementary Fig. 2). The June–August average tem- the highly idealised damage functions used in integrated assess- perature anomalies (relative to 1961–1990) were extreme, ment modelling6, but also more sophisticated, coupled reaching 2 °C (2 standard deviations (σ) above the mean) aver- economic modelling frameworks that combine individual sectoral aged over Europe and more than 5 °C regionally (more than 3σ, models7–9. However, these approaches are centred on gradual and in some locations 5σ)17,18. Due to both decreased pre- changes in physical and biophysical indicators—such as crop cipitation and increased evapotranspiration, the high tempera- yields or water resources—and largely ignore the impacts of tures were accompanied by an intense drought, and the dry soils extreme climate and weather events. in turn amplified the heat wave19,20. This is a serious research gap because such events cause The EHWD had numerous impacts on the environment, enormous damages10. For addressing it, sectoral impact models economy and human health18. In this paper, we focus on large- must be able to credibly represent the impacts of extreme events. scale impact indicators which can be compared between The goal of this paper is to test whether this is the case in the observations and global-scale models (Fig. 1). In terms of these complex, process-based impact models that are routinely being indicators, the summer of 2003 is characterised by anomalies of applied in global-scale climate impact assessment1–3. While these up to 5σ in observed data. In particular, southern Europe saw models may be too costly to integrate them directly in cross- extreme reductions in ecosystem gross primary productivity sectoral economic models, they are the benchmark for any sim- (GPP) and large excess human mortality rates. Substantial relative pler models. More generally, these complex impact models, in reductions in crop yields, river flow and hydropower production conjunction with global climate models, form the basis of much were experienced across the different parts of Europe affected by of our current knowledge about future global climate change the EHWD. The ability of impact models to capture these large impacts, as reflected in the Intergovernmental Panel on Climate anomalies is summarised in Fig. 1, and discussed in the following Change reports, for instance. Whether they capture extreme sections. events well is therefore a key concern even beyond the application The EHWD was exceptional compared to the historic record, in economic assessments. but given continued global warming, comparable events are And yet, it is not known how well the current suite of models anticipated to occur about every 10 years by the middle of the can reproduce the multi-sectoral impacts of a given climatic 21st century, depending on future greenhouse gas emissions21–23. extreme event. Global process-based impact models have been Even when changes in the mean climate state are discounted— evaluated in terms of average quantities and sometimes in tantamount to assuming full adaptation to gradual climate change terms of inter-annual or intra-annual variability11–14, but their —relative deviations like the 2003 EHWD could still occur performance under extreme conditions has rarely been tested at compared to the background climate state in the late 21st large spatial scale15, and never—to our knowledge—in a multi- century24. Furthermore, the impacts of the 2003 event could have sector setting. And since events that are very rare today may been much greater if spring conditions had been drier19, e.g. become much more frequent in the future16, testing for varia- similar to those that prevailed in 201125, which underlines the bility alone may not be enough. risk of even more-extreme events in the future. Being able to Here, we choose the 2003 European heat wave and estimate the damages from such events is therefore crucial for drought (EHWD) event as a test case. The EHWD was sub- assessments of future climate change impacts. stantially stronger than previously observed events; it severely impacted several important sectors across a large geographical area, and its impacts are relatively well docu- River flow and water resources. Due to the prolonged rainfall mented. We examine the impacts of the EHWD in a large deficit (Supplementary Figs. 1–3), the summer of 2003 was ensemble of state-of-the-art impact models covering agri- anomalously dry in central Europe. This was visible in surface culture, water resources, terrestrial and marine runoff and river discharge levels26–28, which are indicators of ecosystems, energy, and human health, for the first time in a renewable freshwater availability for ecosystems and human uses, common modelling framework. For each of these sectors, we such as irrigation or the cooling of thermal power plants29. In identify key observed impacts of the 2003 EHWD reported in August 2003, discharge in both the Rhine and Elbe rivers reached the literature and/or recorded in public databases, and record low levels30. Navigation was impeded on the German examine how closely the models—driven by observations-based section of the Rhine on 37 days in 2003 due to low flow31. climate data—reproduce those impacts. As a common Satellite-based gravity measurements indicate an exceptional impact metric, we choose the deviation of 2003 from the his- depletion of terrestrial water storage compared to 200232. torical average, adjusted for long-term trends, and We examine monthly average river discharge at major gauging normalised by the historical standard deviation (except for stations across Europe where continuous data since 1979 is human health; see Methods). We thereby available from the Global Runoff Data Centre, and compare this circumvent potential biases in the baseline or the average inter- to an ensemble of global hydrological models (Methods). The annual variability, and instead focus on the models’ ability to observed data indicate negative anomalies of ~1.5–2σ during pick out the anomalous 2003 event from the rest of the time August 2003 at five of the stations: in the Rhine (Lobith) and the series. Danube basin (Bratislava, Achleiten, and Inn at Passau-Ingling), The Results section first provides a climatological analysis of as well as in the largest Swiss catchment, Aare (Figs. 1 and 2). the 2003 EHWD, and then presents impact model results for These larger anomalies are reproduced closely by most model each sector. In the Discussion section, we summarise and simulations, with a relatively small spread across the ensemble. evaluate our findings across sectors, and discuss their impli- The multi-model median is very close to the reported value at the cations for integrated assessments of climate change impacts Rhine, Tisza, Danube, and Aare stations. Conversely, many and for future model development. models overestimate the more moderate anomalies observed e.g. 2 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE Central/North Eastern Europe Germany, Europe Hungary, Maize Maize Germany, Inn at Passau Danube at Achleiten Romania, Wheat 4 Wheat 4 3 3 2 2 1 Rhine at Lobith 1 Danube at Bratislava Austria, Sweden 10 Wheat 20 30 40 Czech Republic London Finland Hungary, Belgium Wheat Latvia UK Slovakia Southwest Southeast Bulgaria, Europe Europe Wheat Spain, Italy, France, France, Greece, Maize Wheat Wheat Maize 6 Wheat 4 5 3 4 3 Italy, 2 Central Portugal, Wheat Aare at Stilli 2 Maize 1 1 10 10 Zurich 20 West 20 Rome Italy 30 30 40 40 50 60 Paris Slovenia Milan Barcelona Valencia Crop yield loss River discharge reduction Turin Ecosystem GPP reduction Excess mortality per 100,000 Hydropower usable capacity reduction Fig. 1 Multi-sector impacts of the 2003 EHWD. Black arcs represent observations, and colours represent model results. Units are standard deviations (black axis labels), except for human mortality which is given in excess deaths per 100,000 (grey axis labels). For river discharge, crop yields and ecosystem GPP, the thin red line marks the multi-model median; the dark-coloured segment marks the interquartile range; and the light-coloured segment marks the full range of model results. For hydropower, only one model is available which is marked by the thick blue arcs. For mortality, the red line and grey segments mark the median and the full range, respectively, across three climate forcing data sets and three different heat-mortality relations. Note the larger axis range for Southeast Europe. This figure only includes river discharge, crop yield, GPP and hydropower results for those locations where a negative anomaly larger than 1 standard deviation was observed. Figures 2–6 include further details on the data shown here, as well as additional data for locations with smaller or positive anomalies. The West and Central regions used for ecosystem GPP are defined in Methods in the Elbe and Oder rivers, and the ensemble spread tends to be streamflow and drought conditions in central and western Europe larger there (Fig. 2). Out of 12 stations, there are 5 stations where reported in earlier studies33,34. Indeed, in the hydrological models 75% or more of the models simulate a negative discharge anomaly that include human interventions, irrigation water demand would larger than 1 standard deviation even though the observed increase under drought, but this has little effect on river flow anomaly is smaller. There is no station where the magnitude of when the flow is already low and limits the amount available for the observed anomaly is underestimated by many models. withdrawal (Supplementary Fig. 5). Results are similar in simulations that ignore the effects of Model performance differs little between June, July and August human land use, dams and reservoirs, and water withdrawals (Supplementary Fig. 6). While there is some variation in which (Supplementary Fig. 4), suggesting that present human modifica- stations are matched best in each month, in those cases where tions of the hydrological system do not substantially change the there is a large anomaly (such as in the Danube in July) the model response of the system to this type of extreme event. This is ensemble tends to reproduce that anomaly closely. Results are consistent with the moderate effect of human interventions on also insensitive to the choice of climate forcing data set NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 2 Discharge anomaly (standard deviations) 1 0 –1 –2 –3 –4 TA A G EN E E AU I N AU AP TH LL AV ED TR O N N EN H IT TI ST LI BI W SE SL UA H C LE ,S G LO O EB G C AR TI IN O IN , H S AL IN ZA A N E, I - T -D C ,L -F R U ,K TH N S ,A IN E, SA ,B EU EN R ,I TI ES N H EN BE SE ER BE AS ,N EN AT R AM G U E ,P U ES SA BE VI W N G N TH N DA SI W DA EN EL IN R H TE O N ,H U ER E, R D AA O Fig. 2 August average river discharge anomalies in 2003. Black circles are observed (GRDC) data. Grey numbers are the global hydrological models (see Methods); red lines indicate the median, and blue boxes the interquartile range, of the model ensemble. Stations are ordered by catchment size; the smallest catchment (Thames river at Kingston) has an area of about 10,000 km2, which corresponds to the size of four model grid cells (Supplementary Fig. 7). Our results are consistent with recent 1 standard deviation, there are 3 (4) countries where all models studies attesting a relatively good performance of global (more than 75% of models) underestimate that anomaly. For hydrological models for discharge and runoff variability in wheat, the numbers are 2 (7) out of 9 countries. The best- European catchments, compared to other regions12,35. At the matching models differ from country to country, in line with a same time, the tendency towards false-alarms is in agreement previous evaluation of these models11 which also found mixed with a dry bias induced by the models’ potential evaporation skill in reproducing overall inter-annual yield variability. There is schemes36. The fact that one of our models (number 6) appears broad agreement across models on the sign of the anomaly in consistently at the dry end of the ensemble may be because it most of the strongly impacted countries; except for wheat in Italy, applies a temperature-based evaporation scheme (Hamon) that France, Portugal and Greece, where nearly half of the models show has been shown to induce a particularly large bias when applied a positive anomaly. In terms of magnitude, the large anomalies in outside its calibration range36. France and Italy are underestimated by the entire ensemble for both crops. Generally, the agreement is somewhat better in Middle and Eastern European countries, such as Hungary, Bulgaria or Agriculture. Agricultural summer crop yields were poor in 2003 Poland, compared to Western Europe. These general results are due to the combined effect of the drought and the excessive robust against changing the climate forcing data set, although the temperatures recorded since June18,37. Low harvests incurred an yield loss in Italy is better reproduced with an alternative data set estimated €13 billion of uninsured losses in the European Union (Supplementary Fig. 8). We also note that COPA-COGECA (EU)38. At the EU level, maize and wheat were the most report a decline in maize yields in Austria by about 10% between damaged crops, with production shortfalls of 21 and 11%, 2002 and 2003, which is not reported in the FAOSTAT data respectively18. We examine maize and wheat crop yields simu- (Supplementary Table 2); thus, the real observed value in Fig. 3 lated by an ensemble of 12 global gridded crop models, and may be closer to the model ensemble mean than the FAOSTAT compare them with data reported by FAOSTAT (obtained from value shown by the black circle. http://faostat.fao.org/site/567/default.aspx on 30 August 2016) The poorer performance in Southwestern versus Eastern for central and southern European countries with continuous European countries may be due to more widespread irrigation in data since 1979. Consistent with previous studies39 and early the former. The crop models assume full irrigation in irrigated areas assessments by the COPA-COGECA agricultural association40 and do not account for potential limitations in water availability due (Supplementary Table 2), we find the 2003 EHWD was asso- to drought, which induces an overestimation of irrigated yields, and ciated with large negative yield anomalies in the FAOSTAT data thus biases total yields in countries with much irrigation, such as set in France, Germany and Italy for both maize and wheat; as France or Italy (Supplementary Fig. 9). This is in line with a recent well as in Spain for maize, and in Austria and Portugal for wheat study showing that extreme heat leads to strong declines in maize (Fig. 3, black circles). Additionally, we find substantial yield yield in both observations and models only under rainfed reductions of one standard deviation or more for both crops in conditions15. Moreover, vernalisation of winter crops (the require- Hungary, and for wheat in Bulgaria, Greece and Romania. ment of cold temperatures for flowering) is a known problem in Hungary and Romania in particular are among the EU’s largest crop models41, and the positive sign of the simulated wheat yield maize producers. anomalies in southern Europe in many models may be due to The ability of the crop model ensemble to reproduce these winter wheat flowering, maturing and being harvested early and impacts is mixed (Fig. 3, boxes and numbers). Out of 5 countries thus escaping the heat wave42. Finally, as is the case with the other with an observed negative maize yield anomaly larger than sectors, the crop model simulations were not specifically designed to 4 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE a Maize 2 1 Yield anomaly (standard deviations) 0 –1 –2 –3 –4 –5 ce ia ly ry y n ia d a e l ga an ai an tri ec Ita an ar ga an Sp tu s m re l lg om Po Au un Fr r er Bu G Po H R G b Wheat 2 1 Yield anomaly (standard deviations) 0 –1 –2 –3 ce y d ly n ia ia ry e a l ga an an ai ec tri Ita an ar ga an Sp tu s m re l lg Po om Au un Fr r er Bu G Po H R G Fig. 3 Crop yield anomalies in 2003. a Maize, b wheat. Black circles are observed data from FAOSTAT. Grey numbers are the 12 global gridded crop models; red lines indicate the median, and orange boxes the interquartile range, of the model ensemble. Countries are ordered by their total production (http://ec.europa.eu/eurostat/web/agriculture/data/database) in 2010, decreasing from left to right reproduce 2003 yields, and used standard management and anomaly was located; and one (Central) including Germany, phenology settings; for instance, fixed planting dates and a constant central Italy, and the western Balkans, which also saw large number of heat units required for a crop to mature. negative anomalies, but also including the Alps which saw a positive anomaly46. As observational benchmark we use MODIS47,48 GPP estimates derived from remotely sensed Terrestrial ecosystems. The summer of 2003 saw a major absorbed fraction of photosynthetically active radiation (fAPAR; reduction in overall GPP in parts of Europe, based on vegetation Methods). We find negative anomalies of 4 to 5σ in MODIS- model results39,43 and remote sensing greenness index data44,45. derived GPP during 2003 (Fig. 4b, black circles), consistent with The largest anomaly was situated in France, but negative previous analyses45. anomalies were also observed in northern Spain, Italy, and parts The magnitude of the observed anomaly is underestimated of the Balkan and Germany (Supplementary Fig. 10). by 75% of models in the Central region, and by all models in We evaluate an ensemble of global vegetation models over two the West region. In both regions the model ensemble spreads overlapping rectangular regions (Fig. 4a): one (West) comprising over a large range that also includes positive anomalies and near- France and northern Spain, where the centre of the negative GPP NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 a positive anomaly in 2007 and 2008, consistent with the MODIS estimates. The large spread across models for the reduction of GPP in the 2003 EHWD is despite a relatively high spatial correlation with MODIS-derived GPP globally in all these models49. Wes t Previous studies50,51 have suggested that many global vegeta- Central tion models do not capture the vegetation response to drought, that is, extremely low soil moisture and air humidity, especially for unprecedented or long-lasting cases. While the models adopt mechanistic, or process-based, approaches, most models include empirical parameterisations for temperature and moisture response functions, which may be less reliable outside the historical range of variability for which they have been calibrated. Moreover, not all ecosystem models account for heat stress effects on plant photosynthesis and water stress effects on respiration52, which might explain the underestimation of the EHWD’s impact by many models. Other possible reasons include the lack of b 6 differential species response to drought53, the lack of deep rooting access to water for forests, irrigation and more generally a poor description of the phenology of cultivated vegetation51. 4 GPP anomaly (standard deviations) Energy. Low water availability and high water temperatures 2 during the 2003 EHWD affected electricity supply due to declines in the output of hydropower plants, and impaired operation of 0 thermoelectric power plants that require freshwater for cooling. Across Europe, more than 30 nuclear power plant units had to reduce their production in summer 2003 because of constraints –2 for cooling water uses54,55. Together with hydroelectric power restrictions, this led to a doubling in electricity spot market prices –4 compared to the previous summer30. Even wind power potential was at record low levels in several countries (Methods). We analyse model simulations of hydropower plants’ usable –6 capacity, i.e. the maximum available output at a given point in time (Methods). In the absence of observations of usable capacity, –8 we compare these simulations to reported annual hydropower generation from the global database of the U.S. Energy l t tra es W en Information Administration (EIA); while noting that a plant’s C actual power generation can be lower than its usable capacity Fig. 4 Summer (June–August) gross primary productivity (GPP) anomaly in depending on management. We exclude countries with less than 2003. a Outlines of the West and Central regions, overlaid on a map of the 20 plants or less than 100 MW total installed capacity. The observed GPP anomaly; see Supplementary Fig. 10 for a more detailed observed data show substantial negative anomalies of around version of this map. b Regionally averaged anomaly. Black circles show 1.5–2.5σ in 2003 for a number of countries, among them Italy MODIS remote-sensing-derived estimates. Grey numbers are the global which is one of the largest hydropower producers in Europe vegetation models; numbers offset to the left represent models that were (Fig. 5). Sweden, Finland, and the UK registered large negative run without considering any human influence except climate change, while anomalies, too; consistent with the observation that northern numbers offset to the right represent models that accounted for historical Scandinavia and the British Isles experienced dry conditions in land-use patterns and, in some cases, anthropogenic water withdrawal. Red summer 2003, even though they were not at the centre of the heat lines indicate the median, and green boxes the interquartile range, of the wave. At the same time, Spain, Portugal, and Greece registered model ensemble. One model (no. 4) did not report total GPP but only GPP above-average power generation; consistent with near-average for individual plant functional types (PFTs); for this model, we show the summer precipitation and slightly above-average precipitation sum of the four dominant PFTs in the relevant region during the rest of the year (Supplementary Figs. 2–3). The simulated anomalies are very close to the observed ones zero values (Fig. 4b, numbers). There is no systematic for Italy and the UK, relatively close for Finland and the Czech difference between models that include and exclude human- Republic, and also closely reproduce the positive deviation in induced changes to the natural vegetation pattern (land use Greece. Large observed negative anomalies in Latvia and Belgium change). The estimates are largely insensitive to the choice of are not reproduced. For the remaining countries, the model climate forcing data set (Supplementary Fig. 11), despite mostly overestimates the negative anomalies; underestimates the substantial differences in annual shortwave radiation between positive anomalies; and simulates below-average usable capacity data sets49. for some countries that saw near-average power generation, such Our finding of a smaller sensitivity of GPP to drought in as Switzerland and Germany. The latter may be because the models than in observations is consistent with a previous multi- underlying hydrological model does not capture enhanced glacial model study which, however, assessed a shorter time span45. melting during the heat wave, which could offset runoff deficits Inspecting individual models’ GPP time series (Supplementary downstream. In total, out of 9 countries with an observed Figs. 12 and 13), we note that those models that most closely negative hydropower generation anomaly larger than 1 standard reproduce the 2003 relative anomaly also exhibit a pronounced deviation, there are 3 countries where the usable capacity model 6 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE 3 Hydroelectric usable capacity anomaly (std.) 2 1 0 –1 –2 –3 –4 y .8) ed (1 ) Sw pa (12 ) itz in .4) er d ) Au ny 5) Po tria .4) Fi gal 6) ze G nd .9) te ep ce ( ) in lic ) d Ro om .0) ze nia .9) La a (1 ) Be tvia .8) Sl ium .5) Lu ul ia ( ) m r ia ) Po rg 2) H land .1) Sl ary .9) C ia 8) Be tia 7) s ) ) Sw ly .2 en 5.9 G lan 1.2 7 6 vi 1.9 B k .3 xe ga 1.3 ru 7 .6 . . U h R ree (2. ub 2. u . en (0. . la (0. wa 1 Ita (16 m (9 s (8 r t u (7 nl (4 gd (2 er a (1 lg (1 a 1 bo (1 (1 g 0 ro (0 (0 or (2 er (1 go ( ov ( ( N ce n a a a an ov m un S Fr K d c H ni C an ia sn Bo Fig. 5 Hydropower anomalies in 2003. Black circles show anomalies in annual hydropower generation reported by EIA (Methods). Blue bars show anomalies in simulated annual hydroelectric usable capacity. Countries are ordered by installed capacity (in GW) as included in the model, indicated in parentheses underestimates the magnitude of that anomaly. On the other There is no comprehensive and consistent database of city- hand, out of 25 countries, there are 10 countries where the model scale excess mortality to compare our estimates to. Instead, we simulates a negative discharge anomaly larger than 1 standard compile results from earlier studies that specifically estimated the deviation even though the observed anomaly is smaller in effect of the 2003 EHWD on natural mortality in European cities magnitude. (Supplementary Table 3). These are based on actual mortality A previous evaluation of the hydropower model, using the data reported for the 2003 summer e.g. by hospitals and city same climate input, found relatively low root mean squared error authorities, and are the closest available analogue to direct in southern European countries, but larger values in some north observations. In contrast, the models we applied60 were trained European countries, with particularly poor performance in on mortality and climate data from 1990–2000, and our estimates Latvia56. Our analysis broadly confirms this regional pattern, for 2003 are then derived based on the estimated model, reported while focusing explicitly on a drought event. We note the baseline mortality rates, and the climate forcing. complexity in comparing actual hydropower generation with We find that our statistical estimates are consistent with the hydropower usable capacity, which is simulated by the model. In range of literature values for Barcelona, London, and Rome, while particular, annual power generation is strongly affected by power they are substantially below the observed value(s) for Paris, Milan demand, which was above-average in the winter of 2003 due to and Turin (Fig. 6). For Turin, our estimate is zero for all three low temperatures. The hydropower model focuses on the physical climate forcing data sets, while an analysis61,62 based on observed impacts of changes in water resources on usable power plant mortality found more than 60 excess deaths per 100,000 capacity; economic feedbacks of the assessed water constraints people occurred during the 2003 EHWD. For Valencia and and related management or adaption options (e.g. on energy Zurich, our models also underestimate observed values, although prices and the supply-demand portfolio) are not modelled. More they come somewhat close when using the upper 95% confidence importantly, no monthly observational data are available, so that level for the slope of the temperature-mortality relation. we can only compare annual averages between model and Altogether, our models substantially underestimate the magni- observations. Simulated summer-only (June–August) anomalies tude of the impact in 5 out of 8 cities for which excess mortality are generally larger than for the annual average, and show a was observed. To our knowledge, this is the first independent similar relative distribution across countries, which suggests that evaluation of the performance of these models outside their the anomalies seen in the annual average are largely due to the calibration period. EHWD (Supplementary Fig. 14). Potential reasons for the discrepancies between our model estimates and observed mortality include compounding effects Human health. The 2003 EHWD caused a high human death toll that the statistical models do not account for, including: the mainly due to circulatory- and respiratory-related causes, espe- added effect of heat wave duration, the urban heat island effect, cially among the elderly population57. As many as 70,000 excess and air pollution (both with particulate matter and ozone); as well deaths were estimated for Europe, and 14,800 for France as the difference in spatial resolution between the climate data alone58,59. We apply a set of city-specific statistical models that used to train the models (local airport weather stations) and the describe a linear relationship between maximum daily apparent 0.5°-resolution forcing data sets (see Methods for an extended temperature (ATmax) and daily mortality (see Methods). We discussion). calculate ATmax from daily mean and maximum temperatures Interestingly, while results are similar across the different and relative humidity, from the same observations-based climate climate forcing data sets for most of the cities, we find a strong data set used in the other sectors above, as well as two additional dependence on the climate forcing for Milan. Unlike impacts on data sets (GSWP3 and PGFv2). water resources or plant growth, which integrate the weather NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 70 60 50 Excess deaths per 100,000 people 40 30 20 10 0 –10 A N AN IS E IN IA H N M O IC R C R O IL D RO PA EN TU R EL N M ZU LO L C VA R BA Fig. 6 City-specific estimates of the excess mortality attributable to the 2003 EHWD. Black symbols are observed estimates from the literature (Supplementary Table 3; note two very similar estimates for Paris). Circles denote studies that have reported both the number of excess deaths and the baseline population; diamonds denote studies which have only reported the number of excess deaths and where we have used the corresponding city population reported for 2003 in official statistics as a baseline. Grey bars and red lines are results from this study, for three different climate forcing datasets (left: GSWP3, middle: PGFv2, right: WFDEI). The bars indicate the results obtained by using the lower and upper 95% confidence intervals for the linear exposure response function slopes from ref. 60, and the red line indicates the result obtained using the central estimate for the slope effect over timespans on the order of weeks to months and over the models, nor in the reconstructed catch data (Supplementary larger areas, excess mortality can result from just a few days of Fig. 16). In a few cases, models show a drop in 2004, which may extreme heat in a particular location (city). Thus, small-scale be a lagged effect of the EHWD, but this is hard to verify in the deviations between the different climate forcing datasets or, for absence of a longer time series, and not supported by the that matter, between a gridded reanalysis dataset and the local reconstructed catch data. observations, can lead to large differences in the mortality Nonetheless, the ocean reanalysis dataset used to force the outcome. marine ecosystem models does capture the 2003 anomalous warming in the Mediterranean and North Sea with a similar magnitude as in the observed data (Supplementary Fig. 17). It Marine ecosystems. The 2003 EHWD also produced anom- also shows a local maximum in plankton net primary alously high sea surface temperatures (SST) in the seas around productivity (NPP) in 2003 in the Mediterranean and the North Europe, most pronounced in the north-western Mediterranean Sea, and captures an observed phytoplankton abundance peak in Sea (Supplementary Fig. 15). This surface warming affected the English Channel71 (Supplementary Fig. 18). At the same time, subsurface temperatures, vertical mixing, and currents, with simulated NPP exhibits large year-to-year variability, and the potential effects on marine ecosystems and, by extension, on 2003 value is generally less exceptional in a long-term context fisheries63. Indeed, several impacts, positive as well as negative, on than for temperature (Supplementary Fig. 19). local and regional marine ecosystems have been reported in the Thus, literature, data and models appear to agree that the 2003 literature, including a die-off of Posidonia oceanica seagrass EHWD affected physical ocean parameters as well as plants and habitats64 and rocky benthic communities65; and effects on the planktonic NPP, but that the signal did not propagate to higher reproduction of some fish species like Bluefin tuna66, anchovy trophic levels on a large scale immediately. This may be because and round sardinella67. However, no direct impacts on com- the event was too brief or too weak to cause a significant change mercial fisheries are reported in the literature due to the EHWD, in upper trophic levels. Another reason may be that temporary despite evidence for longer-term effects of warming on fisheries warming can act on fish biomass in two opposite ways: on the one landings68. hand, higher temperatures tend to increase the amount of food We simulated total consumer biomass at a 0.5° horizontal that fish need to maintain growth rates, as well as their mortality resolution with an ensemble of four marine ecosystem models69. due to predation, disease or senescence. On the other hand, short- As there is no widely accepted long-term database of fish biomass term warming increases phytoplankton growth, which has a appropriate to the scales relevant to the EHWD, to gauge the positive effect on fish biomass. The net effect may thus be small. impact of the event on marine ecosystems and fisheries, we Further evaluation of marine ecosystem model simulations compare relative change in reconstructed catch70 to the relative resolving individual species, and/or forced with higher- change in simulated consumer biomass, taking the assumption resolution climate data e.g. from regional climate models, might that harvest is a reasonable indicator of biomass. We find no illuminate whether the models agree with the data for the right discernible impact of the 2003 EHWD on fish biomass in any of reasons; and longer simulation runs would be needed to account 8 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE for any lagged effects beyond 2004. Nonetheless, our results 2σ in the Weser, Elbe and Oder rivers, where no substantial suggest that this modelling framework is applicable to extreme anomalies were observed. The hydropower model shows large events like the 2003 EHWD in a similar fashion as in the reductions in usable capacity in Switzerland, Poland and Ger- terrestrial impact sectors. many, while reported hydropower generation was near average in all these countries. On a regional level, the underestimation of impacts across sectors is particularly evident in southern Europe; Discussion while in northern and eastern Europe, impacts on crop yields and Our study is unique in assessing the ability of several state-of-the- river flow are matched relatively well, and some hydropower art impact models, for multiple sectors, to replicate the impacts impacts are even greatly overestimated (Fig. 1). from a recent, high-impact climatic extreme event, the 2003 Whilst our study is the most comprehensive systematic com- EHWD. When impacts occur simultaneously in different sectors, parison to date between simulated and observed impacts for an it is important that models replicate not only the average varia- extreme climatic event, it is not exhaustive in the treatment of bility of sectoral impact indicators, but the actual timing and different sectors, e.g. we did not include livestock production, magnitude of particular events. The rationale of our study was vector-borne diseases or forest fires18. Assembling an even larger therefore to approach the assessment of model ability differently multi-disciplinary team could have partly addressed this. For —not through a classical evaluation of impact statistics over some sectors suitable observational data for model comparison several decades11,14,35—but instead to answer a what-if question: were unavailable (nuclear power generation). In all sectors, we what if these models are used to predict the impacts of a single have focused on large-scale impact indicators—e.g. average extreme event like the EHWD?. monthly discharge or country-level crop yield—which are close to The answer, according to our results, is that the total impacts of what the models are designed to represent, and which we assume such an event would likely be seriously underestimated. This is to reflect the total societal costs and key environmental impacts of despite some encouraging results in individual sectors. For instance, the 2003 EHWD to a large extent. Smaller-scale characteristics of the observed low water levels in the Rhine, Danube, and Aare rivers some of the impacts (e.g. minimum daily river flow or maximum are matched very well by the median across our hydrological daily excess mortality) could have important implications not models, and the spread across models is moderate. Furthermore, in captured by the more aggregate indicators, for example the most countries where a substantial reduction in hydropower gen- overburdening of critical infrastructure for a short time. Repre- eration is observed, our hydropower model also simulates a sub- senting such processes accurately will be a substantial challenge stantial reduction in usable capacity. With respect to human health, for the relatively large-scale models applied here. our statistical models yield heat-related excess mortality rates of a We have investigated a single climate event, but our results are similar magnitude seen in some observed estimates for Barcelona, relevant beyond the case of the 2003 EHWD. In 2015 and 2018, London and Rome. We also find that generally the choice of climate central Europe was again hit by extremely hot and dry summer forcing data set is not a major source of uncertainty. conditions, and projected trends in temperature and precipitation Importantly though, severe impacts on agriculture and eco- suggest that such events will likely occur more frequently in the system productivity are underestimated by a large margin by future72. While each event is unique, similar events produce most of the sector models. None or few of the crop models similar biophysical impacts: future drought-heat wave events will (depending on the climate forcing) reproduce the extremeness of again lead to low river flows, reduce ecosystem productivity, and the negative maize yield anomalies, of three and four standard damage crop yields. As long as the model shortcomings exposed deviations, respectively, in France and Italy, which are both by our analysis persist, they will likely lead to biased estimates of among the largest maize producers in Europe. The crop model extreme event impacts. ensemble also systematically under-predicts the severity of the These shortcomings are related to the representation of both EHWD’s impacts for wheat in Western Europe, implying that natural processes and human management in the models. For increasing temperatures and more frequent heatwaves due to instance, human adaptation (or the lack thereof) can affect the climate change could have substantially more severe impacts in severity of impacts for a given extreme event73. Indeed, excess the region than typically assumed. Similar results are found for human mortality during the 2006 heatwave in Europe was much natural ecosystem primary productivity: only a few of the global lower than 2003, possibly because of better preparedness57. This vegetation models come close to the enormous negative anomaly might go some way towards explaining the underestimation of seen in remote sensing-derived GPP in Western Europe, and observed impacts by the heat-related mortality models: the 2003 some models even display a positive anomaly. For both the crop EHWD was outside the climatological range that the models were and GPP models, neither the multi-model mean nor any indivi- trained on, and the range that the health system was prepared for. dual model reliably reproduce observations. The mortality models In process-based models, the capacity to respond to extreme estimate only half of the observed rate of heat-related excess climate conditions can be overestimated if resource constraints mortality in Paris, which saw by far the most fatalities of all cities. are not accounted for. An example is the lack of constraints on Taking all sectors together, and across all countries, cities, irrigation water in our crop model simulations. For instance, regions and rivers studied, severe negative impacts (deviation of streamflow in the Vienne basin, which contains large shares of more than 1σ, or more than 10 excess deaths per 100,000 people) irrigated maize cropland74, was anomalously low in summer 2003 were observed in 38 cases (Fig. 1). In 20 of them, the severity of (Fig. 2); and indeed, the maize yield anomaly in France is impacts is underestimated by the respective models, counting in simulated more realistically when irrigation is switched off, than those countries where 75% or more, but not all, of the crop in the default model setup with irrigation (Supplementary Fig. 9). models underestimate yield impacts. Leaving out those countries, Accounting for such cross-sectoral interactions75 and reflecting we are left with 14 cases where the entirety of models under- changing levels of societal adaptation in models will be critical for estimates the observed impacts. This summary statistic is neces- more accurate impact and damage estimates in the future. sarily crude, but highlights the magnitude of the problem. Important natural processes such as heat stress effects on While many of the largest impacts of the EHWD are thus ecosystem photosynthesis are also missing in some models, while underestimated by the sectoral model ensembles, we also find a other processes such as potential evapotranspiration are para- number of “false alarms”, mainly for water resources and meterised through empirical relationships that may not hold hydropower. Hydrological models show flow reductions of up to outside past ranges of climatic variability36. Such shortcomings NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 will affect impact assessments of other climatic events, too. In for the construction of dams and reservoirs over time). This is similar to addition, there are known issues that have not emerged in our simulations under future climate change scenarios, where e.g. agricultural management choices in any particular year are not known. Therefore, if models are study, but may prove problematic when assessing the impacts of found to accurately reproduce the impacts of the 2003 EHWD using the present extreme events of different types, or outside Europe. For instance, set-up, this increases confidence in their ability to quantify the impacts of future the global ecosystem models do not consider the potentially extreme events represented in climate projections. amplifying interactions between climate change and forest dis- turbances, such as wind storms76; and some global hydrological Impact metric. In all sectors except human health, we measure impacts as a given models struggle to capture the magnitude and timing of processes variable’s deviation in 2003 from the historical average, adjusted for long-term such as transmission loss and snowmelt accurately35. Finally, our trends, and normalised by the historical standard deviation (see below for more models may perform worse in parts of the developing world details). That is, we measure the extremeness of the event in a given variable, not the absolute magnitude of change in that variable. If a model correctly simulated where constraints from observational data are more limited. the absolute magnitude of change in a variable in 2003, but simulated the same Our results emphasise the importance of considering the magnitude of change every other year, then the agreement in 2003 may not be for uncertainties of current biophysical impact estimates. Modelling the right reasons, and there would be little confidence in the model’s response to studies of future climate change impacts on mean conditions in extreme events in future climate scenarios. In the paper we use the term “impact” to refer to the relative deviation, or extremeness, of an indicator, unless otherwise different sectors1,2,77,78 have shown that the uncertainties of stated. An exception is heat-related excess mortality, which is already defined as a biophysical impact estimates are of a similar magnitude as climate deviation from the expected baseline mortality. modelling uncertainty. A common practice has since been to use multi-model ensembles rather than single impact models, as is Water resources. Monthly discharge data was obtained from the Global Runoff standard in climate modelling. We have shown that for an Data Centre (GRDC, http://www.bafg.de/GRDC/EN/02_srvcs/21_tmsrs/ extreme event, even an entire ensemble of impact models can riverdischarge_node.html) for gauging stations along major rivers in the geo- sometimes completely miss observations. This urgently calls for a graphic region that was affected by the 2003 EHWD. Only stations with catch- shift in the agenda of model development and evaluation away ments of ~10,000 km2 (equivalent to ~4 model grid cells) or larger were selected. Stations also had to have continuous monthly data available from GRDC for the from mean conditions towards extremes. In the meantime, a whole period 1979–2008. Simulated daily discharge data was first averaged to precautionary interpretation of impact estimates is recom- monthly values. Then, for both simulated and reported discharge, a single month of mended, where the most pessimistic model in an ensemble must interest was extracted (e.g., August), resulting in a yearly series of values for that be considered just as plausible as those in the centre of the month over 1979–2008. From this series, a linear trend was removed, and the standard deviation calculated. The 2003 relative discharge anomaly was then cal- distribution. culated as the difference between 2003 and the 1979–2008 average, divided by the For quantifying the ultimate impacts of climate change on standard deviation (with 2003 excluded from the calculation of the standard different aspects of the economy, or parts of society, different deviation). We note that none of the models consider changes in glacial meltwater metrics need to be applied beyond the (mainly) biophysical runoff, which was an important factor during the 2003 EHWD in Alpine indicators studied here; e.g., macro-economic damages, public catchments26. health burdens, or effects on societal equality and political sta- bility. However, any such analysis needs to rest on a firm Agriculture. We analyse crop model simulations covering the period 1980–2009 assessment of the biophysical impact, and will be unreliable if the under the “default” setting of the ISIMIP2a simulation protocol, where modelling teams individually determined how sowing dates, harvesting dates, fertiliser biophysical impact is not captured accurately. Our study thus application rates and crop varieties are parameterised in their model, with the provides important context to any subsequent analysis of eco- objective of best overall performance for the simulation period. Total crop yield in nomic and societal impacts. each grid cell was calculated as the weighted sum of irrigated and rainfed yields, with the weights being the respective cell fractions under irrigated and rainfed cultivation around the year 2000, as reported in the MIRCA2000 data set82. The Methods 2003 crop yield anomaly was calculated from the simulated and observed series of Climate forcing data. The primary climate forcing dataset used in this study is the annual harvests, analogously to the discharge anomaly, as described above; except latest WATCH meteorological forcing data set (WFDEI)79, which is based on ERA- that rather than a linear trend, a 7-year moving average was removed from the Interim reanalysis data. Two other data sets are used for comparison, for some FAOSTAT reported crop yields, in order to account for non-linear changes sectors: the Global Soil Wetness Project phase 3 forcing data set (GSWP3; http:// induced e.g. by technological and management changes. Such factors are absent in hydro.iis.u-tokyo.ac.jp/GSWP3/)80, based on the 20th century reanalysis (20CR); the crop model simulations, but we applied the same de-trending method to the and the Princeton Global Forcing data set version 2 (PGFv2)81, based on NCEP/ simulation data for consistency. The size of the moving average window was NCAR Reanalysis. The magnitude and spatio-temporal pattern of the 2003 sum- decreased at the edges to preserve the length of the time series. Alternative results mer temperature and precipitation anomaly is similar in all three data sets (Sup- with PGFv2 climate forcing and with linear de-trending are shown in Supple- plementary Figs. 1 and 2), and in agreement with other reanalysis products17,18. mentary Figs. 8 and 20, respectively. There is some ambiguity in assigning wheat harvests to calendar years, but our results are robust against shifting the relevant time series by one year (Supplementary Methods). Model ensemble. Our multi-sectoral ensemble of impact models (Supplementary Table 1) includes 7 global hydrological models, 12 global gridded crop models, and 8 global vegetation models. In addition, we analysed the impacts on hydropower Terrestrial ecosystems. The regions used for analysis are West (7°W–7°E, 42° production as well as heat-related mortality, using single models; and potential −50°N) and Central (0°−20°E, 42°−53°N). Unlike in some other sectors, the impacts on fisheries using four marine ecosystem models. All terrestrial model observational time series for GPP is relatively short—the MODIS data set extends simulations were driven by WFDEI climate forcing for the period 1979–2010, and from 2000–2011. For a consistent model-data comparison, the models’ GPP followed a common protocol (ISIMIP2a, www.isimip.org), allowing us to compare anomaly was therefore calculated in terms of the standard deviation over models within a sector (such as agriculture) and to apply a common method of 2000–2010, after removing a linear trend over the period 1979–2010. We note that analysis across sectors. For some sectors, additional simulations were conducted, the standard deviation over this short period, and thus the GPP anomaly metric, driven with PGFv2 or GSWP3 climate forcing. In those cases, ensembles are may be rather sensitive to inclusion of additional years. Calculating the standard slightly different for the different climate forcings because not all models were run deviation of the models’ GPP anomaly over the entire simulated period 1971–2010, with all climate data sets. In Figs. 2–4, as well as the corresponding supplementary rather than just 2000–2010, has little effect on the models’ relative distribution, but figures, models are uniquely identified by numerals (see Supplementary Table 1). contracts the spread across models, as can be expected (Supplementary Fig. 21). In contrast to the terrestrial models, the marine ecosystem models need not Note that the conversion of the MODIS space-borne measurements of the fraction only atmospheric but also oceanic forcing variables, and are therefore driven by a of absorbed photosynthetically active radiation (fAPAR) into GPP involves addi- reanalysis-based climate model run. While thus not completely consistent with the tional uncertainties. To obtain GPP from fAPAR, a light use efficiency (LUE) factor other sectors, these simulations demonstrate how our analysis framework can be is applied. The MODIS LUE model is based on simple functions of climatic data expanded to the marine realm. (i.e., temperature and vapour pressure deficit) calibrated only at a few sites83. The Importantly, the model simulations were not specifically designed to capture a fact that the MODIS LUE model ignores soil moisture variations and CO2 phy- particular event. Models were set up for best overall performance, and parameters siological effects makes it prone to GPP biases, especially for drought. On the other were kept constant throughout the simulation period, with climate as the only hand, the negative 2003 summer anomaly found in MODIS GPP is consistent with source of annual-scale variations (except that the hydrological simulations account eddy covariance measurements of carbon fluxes at different sites across Europe45. 10 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE Potential biases in absolute GPP values are also circumvented by our use of computed by using data from a spatially corresponding 0.5° climate dataset-grid standard deviations as unit of comparison. cell as input. To the extent that extreme values may be attenuated at coarser spatial scales, this may bias our results towards lower mortality. We note that this Energy. Hydropower: The anomalies in observed hydropower generation and limitation exists in other sectors, too (e.g. when crop models are calibrated on site- level weather variables but forced with grid-level climate data), and for a given simulated usable capacity were calculated analogously to the discharge anomaly, as location it may be alleviated through bias-correction techniques. However, spatially described above; except that instead of a linear trend, a constant mean value was distributed projections of future climate are typically available on a grid level only, subtracted, since there were no substantial trends. Annual country-level hydro- and thus our methodology reflects how the impacts models are typically applied to power generation data was obtained from the EIA database (https://www.eia.gov/ project future climate change impacts98,99. To reiterate, we do not assess how beta/international/data/browser/). For some countries, EIA data is limited to the models perform under ideal conditions for a specific event or location, but how time since their independence; the shortest series (starting 1993) is available for well the model and data framework presently available for climate impacts Slovakia and the Czech Republic. Hydropower usable capacity was simulated for the period 1981–2010 based on plant level regulated streamflow simulations from simulations performs for extreme events like the 2003 EHWD. The ERFs and input climate data do not account specifically for the urban heat the VIC hydrological model forced with WFDEI climate, hydraulic head, total island effect, which has been shown to increase local temperature, and in turn, efficiency of the power generating unit, density of freshwater and the gravitational mortality rates100. Including this effect would therefore increase our modelled acceleration (see supplementary information of ref. 56). Simulations were per- estimates for mortality. However, we do not attempt to include it because the ERFs formed for hydropower plants in Europe using data from the World Electric Power were trained using local airport weather data and not city-centre weather data60. Plant Database84 and simulated usable capacity was then aggregated to country The use of the former for training the ERFs means that the mortality estimates level. The hydropower usable capacity model56 simulates the physical impacts of from the ERFs are actually representative of temperatures outside the urban heat changes in water resources on usable power plant capacity. Economic feedbacks of island, so it would be inconsistent to add an urban heat island effect to the the assessed water constraints and related adaptation options (e.g. on energy prices, temperatures that are input to the ERFs. This limitation might explain why our supply demands portfolio) are not modelled. estimates are much lower than the literature values for some cities, as might the Wind and solar power: In addition to modelling hydropower impacts, we assessed the anomalies in wind power potential capacity factors assuming a potential attenuation of extreme temperatures in the coarse-resolution climate forcing data, as mentioned above. Another potential reason for the discrepancies is homogeneous capacity installed. This assumes the use of a single wind turbine and that the heat effects on summer mortality in 2003 were probably compounded by extrapolation of wind at hub height using the methodology of ref. 85. We found a high ozone concentrations and resulting poor air quality18, whereas air quality record low capacity factor in several Western European countries in summer effects have been removed in the derivation of the ERFs60. (June-August) 2003, relative to the 1979–2014 average, using the WFDEI climate data set. In France the anomaly was –39%. While installed wind power capacity was very low in 2003, these results suggest that comparable future climatic events, Marine ecosystems. Within ISIMIP2a, an ensemble of global and regional marine in a world with much more installed wind power capacity, would induce large ecosystem models was run with forcing from a global coupled ocean circulation negative anomalies in wind power generation. In contrast, we found a positive and planktonic ecosystem model101, which in turn was forced with an atmospheric short-wave radiation anomaly of 5–20% in Western European countries, which reanalysis dataset102,103. This set-up is used as an analogue to the observational would have been relevant for solar power generation. The consequence of all these forcing of the terrestrial impacts models, given that three-dimensional oceanic anomalies would therefore have been a very unusual pattern of energy systems observations or reanalysis data are not available. Marine ecosystem models were impacts from weather, a situation that transmission operators have to manage in run without fishing, in order to capture only climate effects on fish biomass. No order to balance supply and demand. ocean acidification was assumed. Human health. Models: We calculate excess mortality attributable to heat for 8 Code availability. Data processing scripts are available from the corresponding cities for which comparable estimates of observed mortality were found in the author upon request. Availability of impact model code varies by model and is literature. Our calculations are based on previously published models called epi- indicated in Supplementary Table 1. demiological exposure-response functions (ERF). These ERFs were originally derived from a set of non-linear curves developed from a flexible parametric approach60, and summarised as the percentage change in daily mortality (b) per Data availability degree increase in maximum daily apparent temperature (ATmax) above a Simulation data from gridded impact models is available through https://esg.pik- threshold. Based on a relative risk framework86, daily excess mortality attributable potsdam.de/projects/isimip2a/ and citable using the following DOIs: https://doi.org/ to heat (Dattr) can then be derived from the ERFs as follows: 10.5880/PIK.2017.002 (terrestrial ecosystems); https://doi.org/10.5880/PIK.2017.006 (agriculture); https://doi.org/10.5880/PIK.2017.010 (hydrology); https://doi.org/10.5880/ Dattr ¼ Dobs ð1  ebΔATmax Þ PIK.2018.004 (marine ecosystems, regional); https://doi.org/10.5880/PIK.2018.005 (marine ecosystems, global). The city-level mortality model data and country-level hydropower model data are available from the authors upon request. where Dobs is the observed average daily mortality in the warm season (as reported in ref. 60). We sum Dattr over all days in June to August 2003 on which ATmax was above the threshold ðΔATmax > 0Þ. A similar approach has previously been used in Received: 9 January 2018 Accepted: 28 January 2019 climate change impact assessments by refs. 73 and 87. Here we use the central estimate and the upper and lower 95% confidence intervals of the slope (b) reported for each city in ref. 60. The ERFs used here have been applied for modelling adaptation to heat stress under future climate change73, but here we assume no adaptation, i.e. the ERFs remain unchanged between the reference period 1990–2000 and 2003. ATmax is calculated from daily mean and maximum temperatures and relative humidity, References following Tetens’ equation88. As in ref. 60, we use 3-day averages of ATmax in all of 1. Rosenzweig, C. et al. Assessing agricultural risks of climate change in the 21st our analyses to account for possible lag-effects. century in a global gridded crop model intercomparison. Proc. Natl Acad. Sci. In contrast to the sectors above, excess mortality results are presented not in USA 14, 1–6 (2014). terms of standard deviations from the long-term average, but in terms of a city’s 2. Schewe, J. et al. Multimodel assessment of water scarcity under climate total population. This is because excess mortality is already defined as an excursion change. Proc. Natl Acad. Sci. USA 111, 3245–3250 (2014). from a baseline mortality rate, and is not a continuous variable that could be 3. van Vliet, M. T. H. et al. Multi-model assessment of global hydropower and assumed as approximately normally distributed. cooling water discharge potential under climate change. Glob. Environ. Chang Caveats: We note that the studies we use for evaluating the health models 40, 156–170 (2016). (Supplementary Table 3) differ somewhat in methodology among each other as 4. Whitmee, S. et al. Safeguarding human health in the Anthropocene epoch: well as compared to our approach. For instance, while most studies61,62,87,89–92— report of The Rockefeller Foundation–Lancet Commission on planetary including ours—analyse the whole summer (June–August), two studies93,94 health. Lancet 386, 1973–2028 (2015). considered only a shorter period during which the actual heat wave occurred in the 5. Pecl, G. T. et al. Biodiversity redistribution under climate change: Impacts on respective city. However, since excess mortality occurred mainly during the heat ecosystems and human well-being. Science (80-) 355, eaai9214 (2017). wave days, this difference in approach is expected to have only a minimal effect on 6. Pindyck, R. S. Climate change policy: what do the models tell us? J. Econ. Lit. the magnitude of the difference between modelled and reported mortality 51, 860–872 (2013). estimates. In addition, our approach only considers the effects of hot temperatures 7. Hsiang, S. et al. Estimating economic damage from climate change in the on mortality on each day independent from other days. The models do not account United States. Science (80-) 356, 1362–1369 (2017). for the potential compounding effects on mortality of heat-wave duration95,96, nor 8. Monier, E. et al. Toward a consistent modeling framework to assess multi- do they model mortality displacement97. sectoral climate impacts. Nat. Commun. 9, 660 (2018). Another important caveat is that the ERFs were trained on local weather data obtained from airport weather stations60, but the outputs from the ERFs are NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 11 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 9. Harrison, P. A. et al. Differences between low-end and high-end climate 38. Parry, M. L., Canziani, O. F., Palutikof, J. P., van der Linden, P. J. & Hanson, change impacts in Europe across multiple sectors. Reg. Environ. Chang. C. E. (Eds) in Climate Change 2007: Impacts, Adaptation and Vulnerability. https://doi.org/10.1007/s10113-018-1352-4 (2018). Contribution of Working Group II to the Fourth Assessment Report of the 10. Franzke, C. L. E. Impacts of a changing climate on economic damages and Intergovernmental Panel on Climate Change. 843–868 (Cambridge University insurance. Econ. Disasters Clim. Chang. 1, 95–110 (2017). Press, Cambridge, 2007). 11. Müller, C. et al. Global gridded crop model evaluation: benchmarking, skills, 39. Ciais, P. et al. Europe-wide reduction in primary productivity caused by the deficiencies and implications. Geosci. Model Dev. 10, 1403–1422 heat and drought in 2003. Nature 437, 529–533 (2005). (2017). 40. Copa-Cogeca. Assessment of the impact of the heat wave and drought of the 12. Beck, H. E. et al. Global evaluation of runoff from 10 state-of-the-art summer 2003 on agriculture and forestry http://docs.gip-ecofor.org/libre/ hydrological models. Hydrol. Earth Syst. Sci. 21, 2881–2903 (2017). COPA_COGECA_2004.pdf (2004). 13. Hattermann, F. F. et al. Cross‐scale intercomparison of climate change 41. van Bussel, L. G. J., Stehfest, E., Siebert, S., Müller, C. & Ewert, F. Simulation of impacts simulated by regional and global hydrological models in eleven large the phenological development of wheat and maize at the global scale. Glob. river basins. Clim. Change 141, 561–576 (2017). Ecol. Biogeogr. 24, 1018–1029 (2015). 14. Traore, A. K. et al. Evaluation of the ORCHIDEE ecosystem model over Africa 42. Smith, P. C. et al. European-wide simulations of croplands using an improved against 25 years of satellite-based water and carbon measurements. J. Geophys. terrestrial biosphere model: 2. Interannual yields and anomalous CO 2 fluxes Res. Biogeosci. 119, 1554–1575 (2014). in 2003. J. Geophys. Res. 115(G4), G04028 (2010). 15. Schauberger, B. et al. Consistent negative response of US crops to high 43. Vetter, M. et al. Analyzing the causes and spatial pattern of the European 2003 temperatures in observations and crop models. Nat. Commun. 8, 13931 carbon flux anomaly using seven models. Biogeosciences 5, 561–583 (2008). (2017). 44. Zaitchik, B. F., Macalady, A. K., Bonneau, L. R. & Smith, R. B. Europe’s 2003 16. Dosio, A., Mentaschi, L., Fischer, E. M. & Wyser, K. Extreme heat waves under heat wave: a satellite view of impacts and land–atmosphere feedbacks. Int J. 1.5 °C and 2 °C global warming. Environ. Res. Lett. 13, 054006 (2018). Climatol. 26, 743–769 (2006). 17. Schär, C. et al. The role of increasing temperature variability in European 45. Reichstein, M. et al. Reduction of ecosystem productivity and respiration summer heatwaves. Nature 427, 332–336 (2004). during the European summer 2003 climate anomaly: a joint flux tower, remote 18. García-Herrera, R., Díaz, J., Trigo, R. M., Luterbacher, J. & Fischer, E. M. A sensing and modelling analysis. Glob. Chang. Biol. 13, 634–651 (2007). review of the European summer heat wave of 2003. Crit. Rev. Environ. Sci. 46. Jolly, W. M., Dobbertin, M., Zimmermann, N. E. & Reichstein, M. Divergent Technol. 40, 267–306 (2010). vegetation growth responses to the 2003 heat wave in the Swiss Alps. Geophys. 19. Seneviratne, S. I. et al. Swiss prealpine Rietholzbach research catchment and Res. Lett. 32, L18409 (2005). lysimeter: 32 year time series and 2003 drought event. Water Resour. Res. 48, 47. Running, S. W. et al. A continuous satellite-derived measure of global W06526 (2012). terrestrial primary production. Bioscience 54, 547–560 (2004). 20. Fischer, E. M., Seneviratne, S. I., Vidale, P. L., Lüthi, D. & Schär, C. Soil 48. Zhao, M., Heinsch, F. A., Nemani, R. R. & Running, S. W. Improvements of moisture–atmosphere interactions during the 2003 European summer heat the MODIS terrestrial gross and net primary production global data set. wave. J. Clim. 20, 5081–5099 (2007). Remote Sens. Environ. 95, 164–176 (2005). 21. Beniston, M. The 2003 heat wave in Europe: a shape of things to come? An 49. Ito, A. et al. Photosynthetic productivity and its efficiencies in ISIMIP2a biome analysis based on Swiss climatological data and model simulations. Geophys. models: benchmarking for impact assessment studies. Environ. Res. Lett. 12, Res. Lett. 31, L02202 (2004). 085001 (2017). 22. Stott, P. A., Stone, D. A. & Allen, M. R. Human contribution to the European 50. Huang, Y., Gerber, S., Huang, T. & Lichstein, J. W. Evaluating the drought heatwave of 2003. Nature 432, 610–614 (2004). response of CMIP5 models using global gross primary productivity, leaf area, 23. Zhao, C. et al. Temperature increase reduces global yields of major crops in precipitation, and soil moisture data. Glob. Biogeochem. Cycles 30, 1827–1846 four independent estimates. Proc. Natl Acad. Sci. USA 114, 9326–9331 (2017). (2016). 24. Bador, M. et al. Future summer mega-heatwave and record-breaking 51. Van Oijen, M. et al. Impact of droughts on the C-cycle in European temperatures in a warmer France climate. Environ. Res. Lett. 12, 074025 vegetation: a probabilistic risk analysis using six vegetation models. Biogeosci. (2017). Discuss. 11, 8325–8371 (2014). 25. Whan, K. et al. Impact of soil moisture on extreme maximum temperatures in 52. Chen, M. et al. Regional contribution to variability and trends of global gross Europe. Weather Clim. Extrem. 9, 57–67 (2015). primary productivity. Environ. Res. Lett. 12, 105005 (2017). 26. Zappa, M. & Kan, C. Extreme heat and runoff extremes in the Swiss Alps. Nat. 53. De Kauwe, M. G. et al. Do land surface models need to include differential Hazards Earth Syst. Sci. 7, 375–389 (2007). plant species responses to drought? Examining model predictions across a 27. Hannaford, J., Lloyd-Hughes, B., Keef, C., Parry, S. & Prudhomme, C. mesic-xeric gradient in Europe. Biogeosciences 12, 7503–7518 (2015). Examining the large-scale spatial coherence of European drought using 54. IAEA. Operating experience with nuclear power stations in Member States in regional indicators of precipitation and streamflow deficit. Hydrol. Process. 25, 2003. www-pub.iaea.org/MTCD/Publications/PDF/Pub1219_web.pdf (IAEA, 1146–1162 (2011). Vienna, 2004). 28. Gudmundsson, L. & Seneviratne, S. I. Observation-based gridded runoff 55. Rübbelke, D. & Vögele, S. Impacts of climate change on European critical estimates for Europe (E-RUN version 1.1). Earth Syst. Sci. Data 8, 279–295 infrastructures: the case of the power sector. Environ. Sci. Policy. 14, 53–63 (2016). (2011). 29. Oki, T. & Kanae, S. Global hydrological cycles and world water resources. 56. van Vliet, M. T. H., Wiberg, D., Leduc, S. & Riahi, K. Power-generation system Science 313, 1068–1072 (2006). vulnerability and adaptation to changes in climate and water resources. Nat. 30. Fink, A. H. et al. The 2003 European summer heatwaves and drought Clim. Chang. 6, 375–380 (2016). -synoptic diagnosis and impacts. Weather 59, 209–216 (2004). 57. Murray, V. et al. in Managing the Risks of Extreme Events and Disasters to 31. Umweltbundesamt. Monitoringbericht 2015 zur Deutschen Advance Climate ChangeAdaptation (eds Field, C. B. et al.) 487–542 Anpassungsstrategie an den Klimawandel (Dessau-Roßlau, Germany) http:// (Cambridge University Press, Cambridge and New York, 2012). www.umweltbundesamt.de/publikationen/monitoringbericht-2015 (2015). 58. Robine, J.-M. et al. Death toll exceeded 70,000 in Europe during the summer 32. Andersen, O. B., Seneviratne, S. I., Hinderer, J. & Viterbo, P. GRACE-derived of 2003. C. R. Biol. 331, 171–178 (2008). terrestrial water storage depletion associated with the 2003 European heat 59. Pirard, P. et al. Summary of the mortality impact assessment of the 2003 heat wave. Geophys. Res. Lett. 32, L18405 (2005). wave in France. Euro. Surveill. 10, 153–156 (2005). 33. Haddeland, I. et al. Global water resources affected by human interventions 60. Baccini, M. et al. Heat effects on mortality in 15 european cities. Epidemiology and climate change. Proc. Natl Acad. Sci. USA. 111, 3251–3256 (2013). 19, 711–719 (2008). 34. Wada, Y., van Beek, L. P. H., Wanders, N. & Bierkens, M. F. P. Human water 61. Michelozzi, P. et al. The impact of the summer 2003 heat waves on mortality consumption intensifies hydrological drought worldwide. Environ. Res. Lett. 8, in four Italian cities. Eur. Surveill. 10, 161–165 (2005). 034036 (2013). 62. Michelozzi, P. et al. Temperature and summer mortality: geographical and 35. Zaherpour, J. et al. Worldwide evaluation of mean and extreme runoff from temporal variations in four Italian cities. J. Epidemiol. Community Health 60, six global-scale hydrological models that account for human impacts. Environ. 417–423 (2006). Res. Lett. 13, 065015 (2018). 63. Olita, A. et al. Effects of the 2003 European heatwave on the Central 36. Milly, P. C. D. & Dunne, K. A. A hydrologic drying bias in water-resource Mediterranean Sea: surface fluxes and the dynamical response. Ocean Sci. 3, impact analyses of anthropogenic climate change. JAWRA. 53, 273–289 (2007). 822–838 (2017). 64. Marbà, N. & Duarte, C. M. Mediterranean warming triggers seagrass 37. van der Velde, M., Tubiello, F. N., Vrieling, A. & Bouraoui, F. Impacts of (Posidonia oceanica) shoot mortality. Glob. Chang. Biol. 16, 2366–2375 extreme weather on wheat and maize in France: evaluating regional crop (2009). simulations against observed data. Clim. Change 113, 751–765 (2012). 12 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 ARTICLE 65. Garrabou, J. et al. Mass mortality in Northwestern Mediterranean rocky 92. Tobías, A. et al. Short-term effects of extreme hot summer temperatures on benthic communities: effects of the 2003 heat wave. Glob. Chang. Biol. 15, total daily mortality in Barcelona, Spain. Int. J. Biometeorol. 54, 115–117 1090–1103 (2009). (2010). 66. García, A. et al. Climate-induced environmental conditions influencing 93. Le Tertre, A. et al. Impact of the 2003 heatwave on all-cause mortality in 9 interannual variability of Mediterranean bluefin (Thunnus thynnus) larval french cities. Epidemiology 17, 75–79 (2006). growth. Fish. Oceanogr. 22, 273–287 (2013). 94. Johnson, H. et al. The impact of the 2003 heat wave on mortality and hospital 67. Maynou, F., Sabatés, A. & Salat, J. Clues from the recent past to assess admissions in England. Health Stat. Q 25, 6–11 (2005). recruitment of Mediterranean small pelagic fishes under sea warming 95. Hajat, S. et al. Impact of high temperatures on mortality. Epidemiology 17, scenarios. Clim. Change 126, 175–188 (2014). 632–638 (2006). 68. Tzanatos, E., Raitsos, D. E., Triantafyllou, G., Somarakis, S. & Tsonis, A. A. 96. Gasparrini, A. & Armstrong, B. The impact of heat waves on mortality. Indications of a climate effect on Mediterranean fisheries. Clim. Change 122, Epidemiology 22, 68–73 (2011). 41–54 (2014). 97. Gosling, S. N., Lowe, J. A., McGregor, G. R., Pelling, M. & Malamud, B. D. 69. Tittensor, D. P. et al. A protocol for the intercomparison of marine fishery and Associations between elevated atmospheric temperature and human mortality: ecosystem models: Fish-MIP v1.0. Geosci. Model Dev. 11, 1421–1442 (2018). a critical review of the literature. Clim. Change 92, 299–341 (2009). 70. Pauly, D. & Zeller, D. Catch reconstructions reveal that global marine fisheries 98. Gosling, S. N., McGregor, G. R. & Lowe, J. A. The benefits of quantifying catches are higher than reported and declining. Nat. Commun. 7, 10244 climate model uncertainty in climate change impacts assessment: an example (2016). with heat-related mortality change estimates. Clim. Change 112, 217–231 71. Gómez, F. & Souissi, S. The impact of the 2003 summer heat wave and the (2012). 2005 late cold wave on the phytoplankton in the north-eastern English 99. Kingsley, S. L., Eliot, M. N., Gold, J., Vanderslice, R. R. & Wellenius, G. A. Channel. C. R. Biol. 331, 678–685 (2008). Current and projected heat-related morbidity and mortality in rhode island. 72. Orth, R. et al. Record dry summer in 2015 challenges precipitation projections Environ. Health Perspect. 124, 460–467 (2015). in Central Europe. Sci. Rep. 6, 28334 (2016). 100. Smargiassi, A. et al. Variation of daily warm season mortality as a function of 73. Gosling, S. N. et al. Adaptation to climate change: a comparative analysis of micro-urban heat islands. J. Epidemiol. Community Health 63, 659–664 modeling methods for heat-related mortality. Environ. Health Perspect. 125, (2009). 1–45 (2017). 101. Stock, C. A., Dunne, J. P. & John, J. G. Global-scale carbon and energy flows 74. van der Velde, M., Wriedt, G. & Bouraoui, F. Estimating irrigation use and through the marine planktonic food web: an analysis with a coupled effects on maize yield during the 2003 heatwave in France. Agric. Ecosyst. physical–biological model. Prog. Oceanogr. 120, 1–28 (2014). Environ. 135, 90–97 (2010). 102. Griffies, S. M. et al. Coordinated ocean-ice reference experiments (COREs). 75. Frieler, K. et al. A framework for the cross-sectoral integration of multi-model Ocean Model. 26, 1–46 (2009). impact projections: land use decisions under climate impacts uncertainties. 103. Large, W. G. & Yeager, S. G. Diurnal to Decadal Global Forcing For Ocean and Earth Syst. Dyn. 6, 447–460 (2015). Sea-Ice Models: The Data Sets and Flux Climatologies. NCAR Technical Note 76. Reyer, C. P. O. et al. Are forest disturbances amplifying or canceling out NCAR/TN-460+STR (National Center for Atmospheric Research, Boulder, climate change-induced productivity changes in European forests? Environ. 2004). Res. Lett. 12, 034027 (2017). 77. Caminade, C. et al. Impact of climate change on global malaria distribution. Proc. Natl Acad. Sci. USA 111, 1–6 (2014). Acknowledgements 78. Friend, A. D. et al. Carbon residence time dominates uncertainty in terrestrial The work was supported within the framework of the Leibniz Competition (SAW-2013- vegetation responses to future climate and atmospheric CO2. Proc. Natl Acad. PIK-5), the EU FP7 project HELIX (grant no. FP7–603864–2), the FP7 project Sci. USA 111, 3280–3285 (2014). IMPACT2C (grant agreement#282746) and by the German Federal Ministry of Educa- 79. Weedon, G. P. et al. The WFDEI meteorological forcing data set: WATCH tion and Research (BMBF, grant no. 01LS1201A1). H.K. acknowledges support by Grant- Forcing Data methodology applied to ERA-Interim reanalysis data. Water in-Aid for specially promoted research 16H06291 from JSPS. Guided access is provided Resour. Res. 50, 7505–7514 (2014). via www.isimip.org. We thank the Global Runoff Data Centre (GRDC) for making 80. van den Hurk, B. et al. LS3MIP (v1.0) contribution to CMIP6: the Land available their river discharge data. Surface, Snow and Soil moisture Model Intercomparison Project – aims, setup and expected outcome. Geosci. Model Dev. 9, 2809–2832 (2016). Author contributions 81. Sheffield, J., Goteti, G. & Wood, E. F. Development of a 50-year high- J.Sc. designed the research, analysed and interpreted data, and wrote the paper. S.N.G., resolution global dataset of meteorological forcings for land surface modeling. C.R., F.Z., P.C., J.E., L.F., V.H., H.K.L., S.I.S., M.T.H.v.V., R.V. and Y.W. helped design J. Clim. 19, 3088–3111 (2006). the research and interpret results. J.Sc., J.E., K.F., L.W., F.Z., C.R., T.D.E., D.P.T., P.C., 82. Portmann, F. T., Siebert, S. & Döll, P. MIRCA2000-Global monthly irrigated H.M.S. and S.N.G. coordinated impact model simulations. J.E., L.F., S.N.G., M.T.H.v.V., and rainfed crop areas around the year 2000: a new high-resolution data set R.V., Y.W., F.Z., D.A.C., J.C., M.C., D.D., A.d.W., T.D.E., C.F., D.G., N.H., A.I., N.K., for agricultural and hydrological modeling. Glob. Biogeochem. Cycles 24, P.L., C.Mo., C.M., H.M.S., S.O., Y.P., T.A.M.P., G.S., Y.S., E.S., T.S., J.Stee., J.Stei., Q.T., H. GB1011 (2010). T. and X.W. performed impact model simulations. L.F. and H.K. provided observational 83. Running, S. W. & Zhao, M. User's Guide: Daily GPP and annual NPP data. M.B. and J.V. processed data. L.B., D.A.C., P.C., M.C., T.D.E., J.E., C.F., L.F., A.D.F., (MOD17A2/A3) products NASA Earth Observing System MODIS land D.G., S.N.G., L.G., N.H., V.H., A.I., N.K., H.K.L., C.Mo., C.Mü., H.M.S., R.O., T.A.M.P., algorithm. Version 3.0. https://www.ntsg.umt.edu/files/modis/ C.R., E.S., S.I.S., T.S., J.Stee., J.Stei., Q.T., H.T., D.P.T., M.T.H.v.V., R.V. and Y.W. pro- MOD17UsersGuide2015_v3.pdf (2015). vided comments and guidance on the manuscript. 84. UDI. World electric power plants database http://www.platts.com/ (2013). 85. Tobin, I. et al. Assessing climate change impacts on European wind energy from ENSEMBLES high-resolution climate projections. Clim. Change 128, Additional information 99–112 (2015). Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- 86. Gasparrini, A. & Leone, M. Attributable risk from distributed lag models. Bmc 019-08745-6. Med. Res. Methodol. 14, 55 (2014). 87. Mitchell, D. et al. Attributing human mortality during extreme heat waves to Competing interests: The authors declare no competing interests. anthropogenic climate change. Environ. Res Lett. 11, 074006 (2016). 88. Monteith, J. L. & Unsworth, M. Principles of Environmental Physics. (Elsevier Reprints and permission information is available online at http://npg.nature.com/ Science, San Diego, 2007). reprintsandpermissions/ 89. Borrell, C. et al. Socioeconomic position and excess mortality during the heat wave of 2003 in Barcelona. Eur. J. Epidemiol. 21, 633–640 (2006). Journal peer review information: Nature Communications thanks the anonymous 90. Grize, L., Huss, A., Thommen, O., Schindler, C. & Braun-Fahrländer, C. Heat reviewers for their contribution to the peer review of this work. Peer reviewer reports are wave 2003 and mortality in Switzerland. Swiss Med. Wkly 135, 200–205 available. (2005). 91. Martínez Navarro, F., Simón-Soria, F. & López-Abente, G. Valoración del Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in impacto de la ola de calor del verano de 2003 sobre la mortalidad. Gac. Sanit. published maps and institutional affiliations. 18(Supl.1), 250–258 (2004). NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications 13 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-019-08745-6 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/. © The Author(s) 2019 Jacob Schewe 1, Simon N. Gosling 2, Christopher Reyer1, Fang Zhao3, Philippe Ciais 4, Joshua Elliott5, Louis Francois6, Veronika Huber7, Heike K. Lotze 8, Sonia I. Seneviratne 9, Michelle T.H. van Vliet 10, Robert Vautard 4, Yoshihide Wada 11, Lutz Breuer 12,13, Matthias Büchner1, David A. Carozza 14,43, Jinfeng Chang 4, Marta Coll 15, Delphine Deryng16,17, Allard de Wit 18, Tyler D. Eddy 8,19,20, Christian Folberth 11, Katja Frieler1, Andrew D. Friend21, Dieter Gerten 1,22, Lukas Gudmundsson 9, Naota Hanasaki 23, Akihiko Ito 23, Nikolay Khabarov 11, Hyungjun Kim 24, Peter Lawrence 25, Catherine Morfopoulos26, Christoph Müller 1, Hannes Müller Schmied 27,28, René Orth29,30, Sebastian Ostberg 1, Yadu Pokhrel 31, Thomas A.M. Pugh32,33, Gen Sakurai34, Yusuke Satoh10,23, Erwin Schmid35, Tobias Stacke36, Jeroen Steenbeek 37, Jörg Steinkamp 28,44, Qiuhong Tang 38, Hanqin Tian 39, Derek P. Tittensor8,40, Jan Volkholz1, Xuhui Wang4,41,42 & Lila Warszawski1 1 Potsdam Institute for Climate Impact Research, Member of the Leibniz Association, 14473 Potsdam, Germany. 2School of Geography, University of Nottingham, Nottingham NG7 2RD, UK. 3School of Geographic Sciences, East China Normal University, Shanghai 200241, China. 4Laboratoire des Sciences du Climat et de l’Environnement, CEA-CNRS-UVSQ, 91191 Gif-sur-Yvette, France. 5University of Chicago and ANL Computation Institute, 5735S. Ellis Ave, Chicago, IL 60637, USA. 6Institut d’Astrophysique et de Géophysique/U.R. SPHERES, Université de Liège, B-4000 LIEGE, Belgium. 7 Department of Physical, Chemical and Natural Systems, Universidad Pablo de Olavide, Ctra. de Utrera 1, 41013 Sevilla, Spain. 8Department of Biology, Dalhousie University, Halifax, NS B3H 4R2, Canada. 9ETH Zurich, Land-Climate Dynamics, Institute for Atmospheric and Climate Science, 8092 Zurich, Switzerland. 10Water Systems and Global Change group, Wageningen University, PO Box 47, 6700 AA Wageningen, The Netherlands. 11International Institute for Applied Systems Analysis, Schlossplatz 1, A-2361 Laxenburg, Austria. 12Institute for Landscape Ecology and Resources Management (ILR), Research Centre for BioSystems, Land Use and Nutrition (iFZ), Justus Liebig University Giessen, Heinrich-Buff-Ring 26, 35390 Giessen, Germany. 13Centre for International Development and Environmental Research (ZEU), Justus Liebig University Giessen, Senckenbergstraße 3, 35392 Giessen, Germany. 14Department of Earth and Planetary Sciences, McGill University, Montreal H3A 0E8, Canada. 15 Institute of Marine Sciences (ICM - CSIC), Barcelona E-08003, Spain. 16Leibniz Centre for Agricultural Landscape Research (ZALF), Müncheberg 15374, Germany. 17IRI THEsys, Humboldt University of Berlin, 10117 Berlin, Germany. 18Wageningen Environmental Research, 6700 AA Wageningen, The Netherlands. 19Nereus Program, Institute for the Oceans and Fisheries, University of British Columbia, Vancouver V6T 1Z4 BC, Canada. 20Nereus Program, Institute for Marine & Coastal Sciences, School of the Earth, Ocean, and Environment, University of South Carolina, Columbia 29208 SC, USA. 21Department of Geography, University of Cambridge, Cambridge CB2 3EN, UK. 22Geography Department, Humboldt- Universität zu Berlin, 10099 Berlin, Germany. 23National Institute for Environmental Studies, 16-2 Onogawa, Tsukuba, Ibaraki 305-8506, Japan. 24 Institute of Industrial Science, the University of Tokyo, Tokyo 153-8505, Japan. 25Terrestrial Science Section, National Center for Atmospheric Research, 1850 Table Mesa Drive, Boulder, CO 80305, USA. 26Imperial College of London, Department of Life Science, Silwood Park Campus Buckhurst Rd, Berks SL5 7PY, UK. 27Institute of Physical Geography, Goethe-University Frankfurt, Altenhöferallee 1, 60438 Frankfurt am Main, Germany. 28Senckenberg Biodiversity and Climate Research Centre (SBiK-F), Senckenberganlage 25, 60325 Frankfurt, Germany. 29Department of Physical Geography, Bolin Centre for Climate Research, Stockholm University, SE-10691 Stockholm, Sweden. 30Department of Biogeochemical Integration, Max Planck Institute for Biogeochemistry, D-07745 Jena, Germany. 31Department of Civil and Environmental Engineering, Michigan State University, MI 48824, USA. 32School of Geography, Earth & Environmental Sciences, University of Birmingham, Birmingham B15 2TT, UK. 33 Birmingham Institute of Forest Research, University of Birmingham, Birmingham B15 2TT, UK. 34Institute for Agro-Environmental Sciences, National Agriculture and Food Research Organization, 3-1-3 Kannondai, Tsukuba, Ibaraki 305-8604, Japan. 35University of Natural Resources and Life Sciences, Vienna, Feistmantelstrasse 4, 1180 Vienna, Austria. 36Max Planck Institute for Meteorology, 20146 Hamburg, Germany. 37Ecopath International Initiative, Bellaterra 08193, Spain. 38Key Laboratory of Water Cycle and Related Land Surface Processes, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, 100101 Beijing, China. 39School of Forestry and Wildlife Sciences, Auburn University, 602 Duncan Drive, Auburn, AL 36849, USA. 40UN Environment Programme World Conservation Monitoring Centre, 219 Huntingdon Road, Cambridge CB3 0DP, UK. 41Sino-French Institute of Earth System Sciences, College of Urban and Environmental Sciences, Peking University, Beijing 100871, China. 42Laboratoire de Météorologie Dynamique, Université Pierre et Marie Curie, Paris 75005, France. 43Present address: Department of Mathematics, Université du Québec à Montréal, Montreal H2X 3Y7, Canada. 44Present address: Johannes Gutenberg-University, Anselm-Franz-von-Bentzel-Weg 12, 55128 Mainz, Germany 14 NATURE COMMUNICATIONS | (2019)10:1005 | https://doi.org/10.1038/s41467-019-08745-6 | www.nature.com/naturecommunications