Precipitation variability increases in a warmer climate — TEI-tagged view

Angeline G. PendergrasspersName, Reto KnuttipersName, Flavio LehnerpersName, Clara DeserpersName, Benjamin M. SandersonpersName (2017date). “Precipitation variability increases in a warmer climate.” Scientific Reportstitle·j.

Object of study — precipitation variabilityterm·dom from daily to decadal timescales in climate modelterm·key ensembles and station data

Research question — how precipitation variabilityterm·dom, less studied than mean and extreme precipitationterm·dom, responds to warmingterm·dom

Contribution — increased precipitation variabilityterm·dom is a robustterm·mth emergent aspect of the changing water cycle

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Precipitation variabilityterm·dom increases in a warmer climate Abstract Understanding changes in precipitation variabilityterm·dom is essential for a complete explanation of the hydrologic cycleterm·key’s response to warmingterm·dom and its impacts. While changes in mean and extreme precipitationterm·dom have been studied intensively, precipitation variabilityterm·dom has received less attention, despite its theoretical and practical importance. Here, we show that precipitation variabilityterm·dom in most climate modelsterm·key increases over a majority of global land area in response to warmingterm·dom (66% of land has a robustterm·mth increase in variability of seasonal-mean precipitation). Comparing recent decades to RCP8.5 projections for the end of the 21 st century, we find that in the global, multi-model mean, precipitation variabilityterm·dom increases 3–4% K −1 globally, 4–5% K −1 over land and 2–4% K −1 over ocean, and is remarkably robustterm·mth on a range of timescales from daily to decadal. Precipitation variabilityterm·dom increases by at least as much as mean precipitationterm·dom and less than moistureterm·key and extreme precipitationterm·dom for most models, regions, and timescales. We interpret this as being related to an increase in moistureterm·key which is partially mitigated by weakening circulationterm·key. We show that changes in observed daily variability in station data are consistent with increased variability.

Introduction Precipitation variabilityterm·dom is a crucial climatic factor for society, agriculture, and the environment; increased precipitation variabilityterm·dom can reduce agricultural yield 1 , and in developing countries can affect growth of children 2 . Precipitation variabilityterm·dom also connects extreme wet and dry events, floods and droughts, which pose threats to the environment and society 3 . Yet it is sometimes assumed that precipitation variabilityterm·dom does not change in a warmingterm·dom climate [e.g., refs 4 , 5 ], or that mean precipitationterm·dom and its variability change at the same rate 6 . While temperature variabilityterm·dom does not change systematically in response to projections of global warmingterm·dom in most regions 7 , 8 , it is not clear that this should hold for precipitation.

The reigning conventional wisdom of how precipitation changes with warmingterm·dom is that mean precipitationterm·dom change is energetically constrained to about 2% K −1 9 , while extreme precipitationterm·dom change is driven primarily by the change in near-surface moistureterm·var with little change in coincident circulationterm·key, with a magnitude of about 6% K −1 for the 99.9 th all-day percentile of daily precipitation 10 . To formulate a complete theory of how precipitation changes with warmingterm·dom, we must connect explanations for the mean and extremes by addressing how variability responds to warmingterm·dom. Remarkably few studies have undertaken this charge.

Two decades ago, early climate modelterm·key simulationsterm·mth indicated that daily to interannual precipitation variabilityterm·dom increases in response to a doubling of carbon dioxide 11 – 13 . This notion was supported by station observations showing increased variability on decadal timescales 14 and confirmed for interannual timescales by studies with small ensembles of model simulationsterm·mth 15 , 16 . But in the mid-2000s, the changes in interannual to decadal variability were deemed slight in the face of large present-day variability 17 , and subsequent work downplayed its importance 18 , 19 . Very recent work has focused on changes in precipitation variabilityterm·dom at regional scales, including the monsoons 20 , and one study has examined decadal prediction and identified increases in interannual precipitation variabilityterm·dom 21 . However, most work in the last decade focused on the robustterm·mth response of ENSO-related precipitation change across models, which occurs despite disagreement on changes in ENSO-related sea-surface temperature variabilityterm·dom 22 – 27 ; these studies are all limited to interannual timescales and focused on the tropics.

Here we quantify the change in precipitation variabilityterm·dom on timescales from daily to decadal with warmingterm·dom in three ensembles of global climate modelterm·key simulationsterm·mth: the Coupled Model Intercomparison Project version 5term (CMIP5term) ensemble, and two large initial-condition ensembles using the CESM1 and GFDL models (see Methods for details). We use standard deviationterm·mth of precipitation as a metric of its variability (following refs 4 , 5 ; alternative metrics are discussed in the Supplementary Information). We primarily focus on land because this is where most impacts will be felt. The changes in precipitation variabilityterm·dom are interpreted in the framework for the changing distributions of moistureterm·key, vertical velocity, and precipitation presented in ref. 28 .

Results Spatially-aggregated precipitation variabilityterm·dom change Are there important changes in precipitation variabilityterm·dom in response to climate change? An earlier study 17 asserted that changes in variability were not important by visual inspection of precipitation time series at individual grid points because modern precipitation variabilityterm·dom is already large. To assess whether models predict an increase in interannual precipitation variabilityterm·dom with warmingterm·dom, we examine the spatially-aggregated change 29 in the interannual standard deviationterm·mth of seasonal-mean precipitation. Spatial aggregationterm·mth is useful for expressing the risk of climate impacts probabilistically over land areas, because variability at each particular grid cell is high and models do not accurately simulate the location of all precipitating features. Grid cells are binned according to their change, and the resulting histogram is plotted with the frequency represented as a fraction of land area. Figure 1 shows these histograms of the change in interannual standard deviationterm·mth of precipitation by land area from each member of the CMIP5term and CESM1 ensembles from the end of the 20 th century to the end of the 21 st century. As a reference case to quantify the expected internal variabilityterm·key in precipitation when no change in climate occurs, the difference between members of the CESM1 ensemble drawn randomly without replacement during the historical period is also included. Figure 1 Spatially aggregated precipitation variabilityterm·dom change. Land area fraction experiencing a given change in the interannual standard deviationterm·mth of seasonal-mean precipitation over extra-tropical land in ( a ) summer and ( b ) winter, and tropical land in ( c ) JJA and ( d ) DJF at individual grid points from 1976–2005 to 2071–2100 forced by the RCP8.5 scenario, and changes expected from natural variability sampled as differences between randomly-drawn members of the CESM1 single-model ensemble for 1976–2005 (grey shading). Shading indicates the 5–95% confidence interval.

Spatial aggregationterm·mth of precipitation variabilityterm·dom reveals a robustterm·mth change in response to warmingterm·dom. The change in interannual standard deviationterm·mth of precipitation is distinct from present-day internal variabilityterm·key. The interannual standard deviationterm·mth of precipitation increases over a larger land fraction than it decreases in the CMIP5term and CESM1 ensembles in most regions, seasons, and models (see Fig. S1 for results from the GFDL ensemble; which are similar in the extratropics but have some differences in the tropics, addressed later). In the extratropics in summer, internal variabilityterm·key is relatively small, and variability across models is only slightly higher (see Methods for region and season definitions). Agreement across CMIP models as well as with CESM1 (and GFDL: Fig. S1 ) is higher over extratropical land than over tropical land. In the tropics, CESM1 shows a larger increase in variability than the CMIP5term models.

Spatial patterns of precipitation mean, variability, and extremes The spatial patterns of the fractional changes in mean, interannual and daily variability, and extreme precipitationterm·dom share many features (Fig. 2 ). All show maxima in the ITCZ and monsoon regions, as well as at high latitudes of both hemispheres. This is consistent with previous research 30 showing increasing interannual monsoon variability in response to a doubling of carbon dioxide. Mean, variability, and extreme precipitationterm·dom all have minima in the subtropics, particularly the eastern ocean basins, extending to adjacent land areas. The regions of decrease are most extensive for mean precipitationterm·dom (26% of land area), less for variability (7.3% of land for seasonal and 2.9% for daily) and least for extreme precipitationterm·dom (1.5%). Figure 2 Spatial patterns of change in mean, variability and extreme precipitationterm·dom. Change in precipitation ( a ) mean, ( b ) seasonal standard deviationterm·mth, ( c ) daily standard deviationterm·mth, and ( d ) daily maximum for the CMIP5term multi-model mean in 2071–2100 relative to 1976–2005 scaled by the change in global mean surface air temperature. Stippling indicates gridpoints where at least 67% of models agree on the sign of the change. Maps were generated using NCL 53 .

There are also differences between the spatial patterns of changes in interannual variability and mean precipitationterm·dom. The increase in variability is smaller at high latitudes than mean precipitationterm·dom change, in winter in the Arctic and both winter and summer over the Southern Ocean (Fig. S2 ). The increase in mean precipitationterm·dom in winter in the Arctic coincides with increased warmingterm·dom which has been attributed to sea ice loss 31 . The magnitude of changes in daily precipitation variabilityterm·dom and mean precipitationterm·dom at high latitudes are similar.

Magnitude of precipitation variabilityterm·dom change By how much do climate modelsterm·key predict precipitation variabilityterm·dom to change in response to changes in climate? We consider the magnitude of the change in precipitation variabilityterm·dom with warmingterm·dom in the context of two null hypotheses. The first hypothesis arises from a statistical argument for a positive-definite variable like precipitation. The simplest form of change for precipitation is to scale an initial timeseries by a constant, which leads to a change in standard deviationterm·mth equal to the change in its mean 11 . Our first hypothesis, then, is that the change in standard deviationterm·mth follows mean precipitationterm·dom. The predicted rate of increase of global-mean precipitation is energetically constrained to about 2% K −1 9 .

Instead of following the mean, extreme precipitationterm·dom increases much more with warmingterm·dom, at a rate similar to the increase in near-surface moistureterm·var 10 . The second hypothesis is that variability is also driven by the change in near-surface moistureterm·var and would be similar in magnitude to its change 15 , 32 . A simple model assuming that precipitation rate is proportional to near-surface moistureterm·var and vertical velocity, which can each be represented by an appropriate distribution 28 , predicts that if moistureterm·key were to increase with no corresponding change in circulationterm·key or relative humidity, the change in standard deviationterm·mth of precipitation would equal the rate of moistening, 6–7% K −1 (this is also noted by ref. 33 ).

For the change in precipitation variabilityterm·dom to follow the change in moistureterm·key, the circulationterm·key must not change, according to the model of ref. 28 . If precipitation rate is proportional to moistureterm·key and vertical velocity (when vertical velocity is upward), then a strengthening of the circulationterm·key (and thus the average magnitude of vertical velocity) would increase precipitation variabilityterm·dom, while a weakening of the circulationterm·key and decrease in magnitude of vertical velocity would decrease precipitation variabilityterm·dom. Climate modelsterm·key project a decrease in variability of the tropical circulationterm·key in response to warmingterm·dom 34 related to an increase in stability 35 – 37 , and so we should expect the influence of circulationterm·key to decrease precipitation variabilityterm·dom below the magnitude that would be predicted from the influence of moistureterm·key alone.

Figure 3 shows the change in spatially-averaged precipitation variabilityterm·dom over successively warmer decades of the CMIP5term simulationsterm·mth (single-model ensemble results are shown in Fig. S3 ). The rate of change of precipitation interannual standard deviationterm·mth is generally at least as large as mean precipitationterm·dom and smaller than moistureterm·key for the CMIP5term multi-model mean over land (also for CESM1 and GFDL). In the tropics and extratropical summer, the increase in standard deviationterm·mth is greater than the increase in mean precipitationterm·dom and less than near-surface moistening. In extratropical winter it increases at a rate similar to mean precipitationterm·dom change. Changes over ocean have slightly smaller magnitudes but are otherwise similar (not shown). The relationship between standard deviationterm·mth and global-mean temperature is essentially linear when taken in average over large areas and ensembles of model simulationsterm·mth. In contrast to precipitation variabilityterm·dom, temperature variabilityterm·dom (Fig. S4 ) does not change systematically with warmingterm·dom, with the exception of high latitude winter where it decreases in association with sea ice decline 38 , 39 . Figure 3 Rate of change of variability with warmingterm·dom. Change in seasonal mean (dashed lines), standard deviationterm·mth (solid lines), and moistureterm·key (dotted lines) averaged over extra-tropical land in ( a ) summer and ( b ) winter, and tropical land in ( c ) JJA and ( d ) DJF as a function of global-mean surface temperature for the CMIP5term multi-model mean. Each marker indicates a 30-year period centered on consecutive decades between 2006 and 2086 relative to the 1976–2005 base period.

Because the magnitude of precipitation variabilityterm·dom increase is in most cases greater than the change in mean precipitationterm·dom, the first hypothesis can be rejected as an explanation. The second hypothesis, that precipitation variabilityterm·dom is driven by moistureterm·key, can provide a starting point to explain its changes. Because in most models and regions the change in variability is smaller than the change in moistureterm·key, the simple model discussed above implies that circulationterm·key must also weaken. The mean, variability, and extreme of precipitation each measure a different aspect of the distribution of precipitation, but they are not independent of one another. In order to reconcile the differing changes in variability with extreme precipitationterm·dom, there must also be a change in the shape of the distribution of circulationterm·key, because merely changing its strength does not break the symmetry between the changes in mean precipitationterm·dom, variability, and the magnitude of extreme precipitationterm·dom change. Reference 28 showed that an increase in the skewness of the vertical velocity distribution reconciles these differences.

Timescale dependence Do the magnitude of and mechanisms driving changes in precipitation variabilityterm·dom depend on the timescale considered? Fig. 4 shows the change in variability of precipitation at temporal resolutions from 1 day to 3 years averaged over all land grid points for CMIP5term (Fig. S5 shows CESM1 and GFDL, extending to multi-decadal time periods; Fig. S6 shows the change in power spectral density). Changes in precipitation variabilityterm·dom averaged over global ocean grid points are similar, with somewhat smaller magnitude; we primarily focus on land because this is where most impacts on humans will be felt, but we do not expect the dominant mechanisms for change to differ substantially over ocean. Figure 4 Precipitation variabilityterm·dom change by timescale. The change in standard deviationterm·mth of precipitation averaged over all land grid points divided by the change in global-mean surface air temperature in 2071–2100 relative to 1976–2005 in the CMIP5term multi-model mean for time scales ranging from 1 day to 3 years (note that the value at 3 years represents the change in 2050–2100 relative to 1955–2005; see text for details). Gray envelope denotes the 95% confidence interval according to a two-tailed student’s t- test.

The magnitude of precipitation variabilityterm·dom change varies remarkably little across this wide range of timescales, maintaining an increase between 4 and 5% K −1 in the CMIP5term multi-model global land mean, with a modest decrease in magnitude toward longer timescales. The global average (including ocean) has a similar pattern with a slightly smaller increase, 3–4% K −1 . The smaller increase over ocean might be related to the smaller warmingterm·dom there 40 , to soil-moisture feedbacks over land 41 , or to compensating regional changes over ocean.

A notable feature of the precipitation variabilityterm·dom change is that no individual timescale stands out, including ENSOterm·key timescales (>3 y; see also Fig. S6 ). Recent work shows that the ENSO-related precipitation increases in magnitude with warmingterm·dom, maintaining a similar spatial pattern 22 – 24 , 27 . The consistent response of precipitation variabilityterm·dom across timescales indicates that this reflects processes whose influence is felt on all timescales, like the increase in moistureterm·key, rather than processes related predominantly to ENSOterm·key. Indeed, if we think of precipitation as a red-noise processterm·mth, the effects of moistureterm·key on the change in the characteristics of this process project onto all timescales. That said, ENSOterm·key does modulate the difference in interannual precipitation variabilityterm·dom among models, particularly in the tropics in DJF. The CESM1 and GFDL models happen to bracket the responses of precipitation variabilityterm·dom across all CMIP5term models (Figs 3 and S3 ); but the GFDL model has an ENSOterm·key variability response that is an outlier among CMIP5term models 42 . Removing the influence of ENSOterm·key by isolating neutral ENSOterm·key years brings the change in precipitation variabilityterm·dom in these two models into good agreement (Fig. S7 ).

The uncertainty in precipitation variabilityterm·dom due to internal variabilityterm·key alone (Fig. S5 ) is small for short timescales and increases dramatically moving to longer timescales. This stands in contrast to the uncertainty across the CMIP5term multi-model ensemble, which is relatively large at all timescales. This indicates large structural uncertainty at short timescales, and that changes in short timescale variability would be detectable in the observational record before changes in variability at longer timescales.

Observed daily precipitation variabilityterm·dom change Are changes in precipitation variabilityterm·dom already occurring? Daily precipitation variabilityterm·dom can be estimated with greater certainty than interannual precipitation variabilityterm·dom because the number of samples during the period of observations is larger (Fig. S5 shows uncertainty due to internal variabilityterm·key alone, which is irreducible, across timescales). An earlier study 43 shows “mildly significant” increases in the observed US interannual precipitation variabilityterm·dom. We calculate observed daily variability change over the second half of the 20 th century from station observations (Fig. 5 ). Precipitation variabilityterm·dom has increased at most stations (though this increase in variability is similar in magnitude to the increase in mean precipitationterm·dom), consistent with model projections of increased variability with warmingterm·dom. The median observed change is 13% K −1 , compared to 8.0% K −1 during the same period in CMIP5term models, subsampled at grid boxes with observing stations. Figure 5 Observed daily precipitation variabilityterm·dom change. Change in the standard deviationterm·mth of daily precipitation from 1960–1969 to 1990–1999 per degree global mean surface temperature change, ( a ) at station locations and ( b ) aggregated as a histogram. Dark gray line in ( b ) is the median change across stations. See Methods for details. The map was generated using NCL 53 .

Discussion In summary, precipitation variabilityterm·dom simulated by climate modelsterm·key from daily and monthly to interannual and decadal timescales increases robustly over almost all land areas in response to anthropogenic warmingterm·dom in most models. For longer timescales, particularly interannual and longer, the forced increase may well be compensated for or amplified by natural variability (Fig. S5 ). We also show that observed daily precipitation variabilityterm·dom increased over the second half of the 20 th century. In most models, regions, and seasons, the change in precipitation variabilityterm·dom (measured by the mean standard deviationterm·mth) is higher in magnitude than the change in mean precipitationterm·dom and lower in magnitude than the change in moistureterm·key and extreme precipitationterm·dom (heaviest day of rainfall each year). In most regions and seasons, mean precipitationterm·dom increase sets a lower bound on the change in precipitation variabilityterm·dom; the increase in precipitation variabilityterm·dom is at least as large and in many cases greater than the increase in mean precipitationterm·dom. We attribute this to the role of increasing moistureterm·key counteracted by weakening circulationterm·key.

In our analysis, we quantify precipitation variabilityterm·dom by its standard deviationterm·mth, and we examine its change in a set of model simulationsterm·mth subject to just one forcing scenario. While the standard deviationterm·mth is a crude metric of precipitation variabilityterm·dom, especially on short timescales with many dry periods, precipitation variabilityterm·dom also increases when quantified by other metrics (Fig. S8 ). The standard deviationterm·mth has the advantage of being well-defined, with a straightforward behavior in response to simple scaling of the distribution. Our analysis focuses on model simulationsterm·mth driven by a large increase in greenhouse gas forcing along with changes in aerosol forcing. Changes in variability will differ with the type of forcing, e.g., in an extreme example of geoengineering by stratospheric sulfate aerosol injection, where global-mean temperature does not change but mean precipitationterm·dom decreases 44 . Finally, in our interpretation we have not considered mechanisms which might be important regionally, like soil moistureterm·key feedbacks.

Aggregated globally and over long timescales, the increased precipitation variabilityterm·dom from daily all the way to decadal timescales is a robustterm·mth emergent aspect of the water cycle that is changing as a result of anthropogenic warmingterm·dom, yet has received little attention. Both theoretical work and impact studies are required to complete and connect the understanding of changes in mean, variability, and extreme precipitationterm·dom, as well as the risks they pose to society.

Methods Model simulationsterm·mth Model simulationsterm·mth include the CMIP5term multi-model ensemble 45 (the r1i1p1 member only, 37 models with monthly data and 34 with daily data), CESM1 large ensemble 46 (40 members), and GFDL ESM2M large ensemble 47 (30 members). For all ensemble simulationsterm·mth, fully coupled simulationsterm·mth are forced by historical and RCP8.5 scenarios. We used output starting from 1951 because the GFDL large ensemble starts in 1950 and this allows for 1 year of spinup, and ending in 2100 when the GFDL and CESM1 simulationsterm·mth end.

Standard deviationterm·mth We quantify precipitation variabilityterm·dom by its standard deviationterm·mth. We remove the linear trend from each time series before calculating its standard deviationterm·mth. This largely removes the effect of transient warmingterm·dom at the end of the 21 st century in the RCP8.5 scenario.

Extreme precipitationterm·dom We define extreme precipitationterm·dom (shown in Figs 2 and 3 ) as the maximum daily precipitationterm·dom each year (rx1dayterm·dom 48 ), averaged over the same time periods as mean precipitationterm·dom and precipitation variabilityterm·dom.

Statistical significance Error bars or envelopes in Figs 1 and 4 show the 95% confidence interval of the multi-model mean calculated from a two-tailed Student’s t -test, using the standard deviationterm·mth among ensemble members where the degrees of freedom is the number of ensemble members.

Robustness across climate modelsterm·key A change is considered “robustterm·mth” when at least 67% of climate modelsterm·key agree on the sign of the change.

Spatial aggregationterm·mth Spatial aggregationterm·mth follows ref. 29 . We use each model’s own land mask and grid cell area fields to calculate the land fraction contributed by each grid cell. Where summer and winter seasons are identified, summer seasons are defined as June to August (JJA) north of the equator and December to February (DJF) south of it; vice versa for winter. Tropics and extratropics are defined as latitudes less and greater than 30°, respectively. Internal variabilityterm·key is quantified by comparing two members of the CESM1 ensemble drawn randomly without replacement during the base period.

Regridding For maps of the CMIP5term multi-model mean, calculations (mean, standard deviationterm·mth, or extreme precipitationterm·dom) are made on each model’s grid, data are regridded to a common 2.5 × 2.5° grid, the multi-model mean initial field is calculated, each model’s change is scaled by its change in global-mean surface air temperature, and then the multi-model mean absolute change is divided by the initial multi-model mean.

Spatial averaging For Figs 3 and 4 , the spatial-average change in each of mean precipitationterm·dom, standard deviationterm·mth of precipitation, or mean moistureterm·key is calculated by first calculating the variable at each grid cell. In the case of standard deviationterm·mth, it is squared. Then, the spatial average of the variable for the initial and final time periods is calculated, and in the case of standard deviationterm·mth, the square root is taken. Next, the difference of the spatial average between the time periods is calculated, followed by scaling by each model’s global-mean surface air temperature change over the appropriate period of time (in Fig. 4 ), and finally calculating the multi-model mean.

Timescale dependence For each timescale, the standard deviationterm·mth is calculated for an initial and final time period. Spatial averaging is carried out as described above. For daily through annual timescales, the initial period is 1976–2005 and the final period is 2071–2100. Daily data for 34 CMIP5term models are used. For the 1-month timescale, the annual cycle is removed before the standard deviationterm·mth is calculated. For the 3-month timescale, monthly data with annual cycle removed are averaged in 3-month seasons and the standard deviationterm·mth is calculated. For the 1-year timescale, monthly data is averaged over years. For the 3-year timescale, monthly data are averaged in 3-year periods over the 51-year periods 1955–2005 and 2050–2100 to achieve sufficient sampling.

Observational analysis Daily precipitation station data come from the Global Historical Climatology Network Daily 49 dataset, version 3.22 50 and use the subset of stations in the Global Climate Observing System (GCOS) Surface Network (GSN). This subset of stations has better spatial sampling than the total network. Only stations with data for at least 50% of each period are used. Global-mean surface temperature observations are calculated from NOAA GlobalTemp 51 , 52 . Change in the standard deviationterm·mth of daily precipitation is calculated from 1960–1969 to 1990–1999. The time periods are chosen to maximize the number of stations; we repeated the analysis for start dates in the 1950s and 1970s and end dates from the 1990s through 2016, with similar results.

Data availability CMIP5term simulationsterm·mth are available from PCMDI through the Earth System Grid (ESG) repository at https://esgf-node.llnl.gov/projects/cmip5term/ . CESM1 simulationsterm·mth are available from the ESG at https://www.earthsystemgrid.org/dataset/ucar.cgd.ccsm4.CESM_CAM5_BGC_LE.html . GFDL large ensemble data can be requested from the corresponding author upon reasonable request and with permission of K.R. Rodgers. GHCN-Daily data are available from NOAA at 10.7289/V5D21VHZ.

Electronic supplementary material Supplementary Information