Climate extremes

Changes in U.S. temperature extremes under increased CO2 in millennial-scale climate simulations 

Changes in extreme weather may produce some of the largest societal impacts from anthropogenic climate change. (At present, weather damages are dominated by rare events that happen only every several decades or more.) However, predicting future changes in those rare events is not possible using only the short observational record.  Insight on changes in extremes must come from climate models, where we can generate long simulations.

In this project, we use millennial runs from the Community Climate System Model version 3 (CCSM3) in equilibrated pre-industrial and future (700 and 1400 ppm CO2) conditions to examine both how extremes change and how well these changes can be estimated as a function of run length.

Figure 1: Illustration of how changes in extreme TEMPERAtures can differ from changes in OVERALl temperature distributions. Model output FrOM 1000-year RUNS OF CCSM3 with CO2 at pre-industrial (PI) and 2.5 x PI (700 ppm) levels, from a mid-latitudes location (in Idaho).

(a, bottom): daily temperatures in Winter (DJF), with PI in blue and future in  RED. (B, TOP). annual cold extremes (Annual Winter TMIN), on same temperature scale. Note that Winter extreme cold TEmperatures WARM MORE strongly than The mean temperature shift. (From Huang et al, 2015)

Extreme value theory provides a means of estimating the far tails of distributions. We estimate changes to distributions of future temperature extremes by fitting annual maximum and minimum temperatures to generalized extreme value (GEV) distributions.  

Using 1000-year preindustrial and future timeseries of temperatures in the contiguous United States, we show that changes in extremes are different in summer and winter. In winter, cold extremes generally warm much more than the mean shift in wintertime temperatures (Figure 1), while in summer, warm extremes generally warm only as much as the shift in means. 

The changes in winter extremes involve more than a simple shift in magnitudes. The scale and shape of their distributions also change. This effect is best demonstrated by plotting the "return level" of extreme events, i.e. the magnitude of an extreme that recurs on a specified timescale. Figure 2 below shows changes in the 2-year to 100-year return levels for wintertime cold extremes. Over ocean regions (red lines at lower left are the Pacific off California; those on right are the Gulf of Mexico and the Atlantic), changes are relatively flat across time, suggesting a simple shift in the distribution of extremes. Over land regions, however, changes show complex behavior. In the inland Southwest U.S.,  changes in the 100-year "extreme extremes" are larger than changes in the 2-year "moderate extremes". Further north, this pattern is reversed.

FIGURE 2: ESTIMATED CHANGES IN RETURN LEVELS OF WINTER TMIN  EXTREMES IN 700 PPM CO2 VS. PRE-INDUSTRIAL MODEL RUNS FOR THE NORTH AMERICAN REGION ( 6 X 16 GRID CELLS). EACH OF THE 16 PANELS REPRESENTS A LONGITUDE; IN EACH PANEL, THE 6 LATITUDES ARE DENOTED BY COLOR. THE X AXIS OF EACH PANEL IS THE RETURN PERIOD AND Y AXIS THE CHANGE IN  RETURN LEVEL.THICK DASHED LINES SHOW ESTIMATED RETURN LEVEL CHANGES; THE ENVELOPES SHOW ASSOCIATED UNCERTAINTIES (ESTIMATED FROM BLOCK BOOTSTRAPPED SAMPLES WITH RESAMPLED YEARS).  FOR EACH GRID CELL, THE CORRESPONDING MEAN TEMPERATURE CHANGE IS MARKED WITH A SYMBOL (CROSSES FOR LAND AND BOXES FOR OCEAN LOCATIONS). (FROM HUANG ET AL., 2015.)

The 1000-year runs used here allow us to accurately determine changes in even 100-year extremes, but in practice, most modeling studies must rely on shorter runs. GEV modeling should allow estimation of infrequent events using relatively short time series, but it is important to understand its limitations for climate data. We therefore repeat the estimation of selected return levels (20-, 50-, and 100-year extremes) using segment of the timeseries of varying length. The resulting estimates can become very poor when the timeseries is of comparable length or shorter than the return period of interest: that is, a 100-year model run cannot be used to reliably estimate changes in 100-year extremes. These results suggest caution when attempting to use short observational records or model runs to infer changes in extreme events. 

People:

Whitney Huang | Elisabeth Moyer | Michael Stein | Shanshan Sun | David McInerney | Hao Zhang