Observational- and model-based trends and projections of extreme precipitation over the contiguous United States

Authors


Abstract

Historical and projected trends in extreme precipitation events are examined in Coupled Model Intercomparison Project 5 (CMIP5) models and observations, over the contiguous United States (CONUS), using several approaches. This study updates earlier studies that have used the extreme precipitation index (EPI) to assess observations and goes further by using the EPI to evaluate available climate model simulations. An increasing trend over the CONUS was found in the EPI, with large differences among seven subregions of the United States. Median of CMIP5 simulations also finds an increasing trend in the EPI, but with a smaller magnitude than the observations. Model spread is large and in most cases bigger than the model signal itself. Statistically significant (95th confidence level) increasing trends in the observational-based EPI occur over the Midwest and Eastern regions, while most decreasing trends occur over Western regions. Some models give negative correlation coefficients relative to observations. However, some ensemble members, for most models, show correlation coefficients greater than 0.5. Projections of extreme precipitation event frequency, for representative concentration pathway (RCP) scenarios 4.5 and 8.5, show increasing trends over the CONUS. Both scenarios give a steady increase throughout the period but the RCP 4.5 signal is smaller in magnitude. Overall, the RCP scenarios show an increase across all regions with the exception of some variability between decades in some regions for RCP 4.5. For the CONUS model spread is smaller than the projected signal. Regional analyses show overall agreement among models of a future increase in extreme precipitation event frequency over most regions.

1. Introduction

An increase in extreme precipitation events has been observed, using several different methods of detection and analysis, for much of the contiguous United States (CONUS) [Karl et al., 1995; Karl and Knight, 1998; Kunkel et al., 1999, 2003, 2007, 2012; Kunkel, 2003; Groisman et al., 2012; Alexander et al., 2006]. Extreme precipitation events can result in flooding, which can damage crops, cause property damage, and even loss of life. Thus, understanding current trends and future projections of such events is of great importance. A variety of methods have been implemented in past studies to define and analyze extreme precipitation events, including fixed thresholds, percentiles, indicators, and station-specific thresholds. Examples of prior use of each method are provided and for this study station-specific thresholds are used.

The climate extremes index (CEI) and the greenhouse climate response index (GCRI) were used by Karl et al. [1995] to assess changes in climate. Data from the National Climatic Data Center's (NCDC) climate division database with monthly resolution for the twentieth century were used. Daily changes in precipitation were found from a subset of U.S. Historical Climatology Network (HNC) stations. Because of sparse coverage in the Western United States, non-HNC stations were used to supplement where necessary. The CEI is a composite of five indicators of climate extremes. Three of these indicators relate to precipitation and two to temperature. The values in the CEI are given as percent of area covered by an extreme value (less than the tenth or greater than the ninetieth percentile) of the indicator, with a lower bound of 0% and an upper bound of 100%. The lower bound means that no part of the United States was affected by any of the extremes considered by the five indicators during a given year, whereas the upper bound means exactly the opposite. Karl et al. [1995] found that over the past century there have been large decadal variations in the CEI. Since about 1976 it has averaged about 1.5% higher than the previous 65 year average. The recent increase in the CEI is of longer duration than increases in prior periods and is mainly owing to an increase in the three precipitation indicators. The rates of change of these indices, at that time, were small enough that natural variability could not be rejected as the basis for the increase. The GCRI tracks expected changes in climate owing to increasing concentrations of greenhouse gases. The GCRI showed an increase in all of its indicators and these changes are consistent with trends expected from a climate affected by higher greenhouse gas concentrations. For the time period studied, their statistical analysis concluded that the change in this index is not adequately large or consistent enough to unequivocally rule out random variability in climate, although it was equated to about a 5% chance.

Karl and Knight [1998] used percentiles and daily thresholds to evaluate observations of precipitation trends over the CONUS. A combination of HNC stations and 48 additional stations for added coverage of the Western United States were used for their study. They found that the intensity and proportion of total annual precipitation resulting from heavier, extreme events have increased since 1910. The latter is relative to more moderate events. Their results showed that 53% of the total precipitation increase was a result of positive trends in the upper 10th percentile of the distribution. It was also found that for all categories of daily precipitation amounts, there has been an increase in the probability of occurrence.

Groisman et al. [2012] defined two fixed daily thresholds for very heavy precipitation. A separate (larger) threshold of total precipitation was defined for “extreme” events for both multiday and single day events. More than one threshold for very heavy precipitation was necessary owing to differences in the upper 0.3% of mean daily precipitation for the various regions of the United States. When analyzing observed temporal changes over the central United States they found that moderately heavy precipitation events (12.7–25.4 mm) had decreased in frequency while heavier events (>25.4 mm) had increased for the period 1979–2009 compared with 1948–1978. Groisman et al. [2012] also found that in a comparison of the past 31 years (1979–2009) relative to 1948–1978, there have been significant increases above their thresholds of very heavy and extreme precipitation event frequencies over the central United States. For multiday extreme events, there has been up to a 40% increase in frequency, corresponding to a decrease in return periods for extreme events.

Another metric to quantify changes in extreme precipitation events is based on recurrence intervals, rather than fixed thresholds, and duration. Kunkel et al. [1999] examined extreme precipitation events of 1 and 5 year recurrence intervals using station-specific thresholds and durations of 1, 3, and 7 days. Statistical significance was found for positive trends in 1 year return and 7 day duration events as well as for 5 year 7 day events. Kunkel et al. [2003] named this metric as the extreme precipitation index (EPI) with a few changes to its methodology; at the same time, they analyzed a longer time series (1895–2000). Kunkel et al. [2003] arithmetically averaged the number of extreme events from stations over U.S. climate divisions and then those averages were further averaged using area weighting to find state values. Then the state values were averaged using state area weighting to get values for the whole CONUS, and this weighted average was named as the EPI. In this metric, areas of sparse or high station density are not given unduly low or high weighting. Kunkel et al. [2003] and Kunkel [2003] found that EPI values in the late 1800s and the early 1900s were similar to those in the early 1980s and 1990s. While acknowledging that natural variability could play a role, Kunkel [2003] theorized that there are substantial reasons to expect that the cause of the more recent increase in these extreme events may be related to human-induced changes in climate.

Kunkel et al. [2007] used a similar national heavy precipitation frequency index, except that the averaging was done on a grid rather than climate division. The station annual heavy precipitation counts were averaged for 1° latitude × 1° longitude grid cells. Then regional averages were calculated by averaging all grid cells with at least one station. Grid cells with no stations were not included in the calculation. They used a Monte Carlo sampling experiment to demonstrate that the difference between the recent very high frequency of extreme events and the moderately high frequency of events in the earlier portion of the record was statistically significant, with the highest confidence being for the shortest return period studied (1 year). Also, all return periods showed a positive nonzero trend at the 95% confidence level when the entire period (1895–2004) was used.

The “delivery method” of extreme precipitation is also an important aspect of understanding these changes in extreme events. Studies by Knight and Davis [2009] and Kunkel et al. [2010] found an increase in the number of extreme precipitation events connected to tropical cyclones. However, Groisman et al. [2012] found that tropical cyclones did not play a key role in the recent extreme precipitation increases. Groisman et al. [2012] also observed an inverse relationship between La Nina conditions and the very heavy and extreme precipitation days over the central United States. Another important weather system that affects the United States is the extratropical cyclone (ETC), which commonly produces precipitation. Kunkel et al. [2012] found that 54% of all extreme events were associated with the cold and warm fronts of ETCs and that these have been increasing. Another 24% of extreme events were associated with ETCs but not near the fronts; these have not been increasing. Other weather systems responsible for U.S. extreme events include tropical cyclones (13%) and mesoscale convective systems (5%), the North American monsoon (3%), air mass convection (1%), and upslope flow (0.3%); other than tropical cyclones, these have not been increasing.

Min et al. [2011] analyzed simulations of extreme precipitation from the Third Phase of the Coupled Model Intercomparison Project (CMIP3) and observations from the Hadley Centre global land-based gridded climate extremes data set (HadEX) for the time period of 1951–1999. They found that projected changes in extreme precipitation may be underestimated because models have a tendency to underestimate observed increases in heavy precipitation. They speculated that because of the possibility of a more rapid strengthening in extreme precipitation events than is currently being projected there could be more severe impacts.

As one aspect, this study provides an update to prior analyses of historical observations of extreme precipitation events over the CONUS, primarily using a version of the EPI. While previous studies have focused on observational-based changes, this study goes beyond that in order to assess new CMIP5 model capabilities at simulating extreme precipitation event frequency. With the availability of daily precipitation data for the United States, extending as far back as 1895, and new CMIP5 simulations, there is a unique opportunity to further assess the ability of current climate models to simulate historical extreme precipitation trends. In this process, the observational-based EPI is compared with the EPI determined from historical CMIP5 simulations.

This study also goes beyond prior studies by examining projected changes in extreme precipitation event frequency based on CMIP5 projections using the new RCP forcing scenarios. An index similar to the EPI is analyzed for CMIP5 model simulations using the RCP 4.5 and RCP 8.5 scenarios. This is done in order to evaluate the projected trends in extreme precipitation events over the 21st century.

2. Data

Observed daily average precipitation data were obtained from the U.S. Cooperative Observer Network, as included in the Global Historical Climate Network-Daily data set from NCDC, for 726 stations over the continental United States. The distribution of stations can be seen in Figure 1. In order to be consistent with the National Climate Assessment, the observational analyses were done from 1901 to 2012 and 1901 to 2010 for decadal EPI anomalies. However, for the purpose of comparing the historical model results to observed precipitation data, the EPI used in this study was also calculated from 1901 to 2005 because CMIP5 historical forcing model simulations end in 2005. A 1° latitude × 1° longitude grid was placed over the United States and each station was assigned to a grid cell. Observed data were screened prior to calculation of the EPI. For each period of analysis, a station was included if it had at least 90% available daily data. For each year in the station time series, the number of extreme events was calculated if there was at least 300 days for that year; otherwise, the station's value for that year was considered missing and not included in the grid cell average.

Figure 1.

Distribution of observing stations used in EPI analyses. All stations are given grid coordinates on a 1° latitude × 1° longitude grid.

Modeled data include results from the fifth phase of the CMIP (CMIP5) [Taylor et al., 2012]. All the models and their respective number of ensemble members used in this study are listed in Table 1. CMIP5 climate change modeling experiments included a long-term (century) integration [Hibbard et al., 2007; Meehl and Hibbard, 2007; Taylor et al., 2012] that was used in this study. The core set for the long-term integrations includes “historical” runs. The historical, or twentieth century, simulations cover most of the industrial period from mid-nineteenth century to near present (1850–2005) and include all forcings [Taylor et al., 2009, 2012]. The historical simulations are useful for the evaluation of model performance against current observed climate and climate change [Taylor et al., 2009]. Here the long-term historical simulations from CMIP5 are analyzed for comparison with the observed trends. Historical ensembles were run from the mid-1800s to at least 2005. Ensembles allow for statistical significance tests of variations between observations and simulations and between the simulations themselves [Taylor et al., 2009]. Thus, whenever possible we used all the available ensemble runs for each model.

Table 1. CMIP5 Models and Corresponding Ensemble Runs
ModelHistorical EnsemblesRCP 4.5 EnsemblesRCP 8.5 Ensembles
ACCESS1-0 11
CCSM4266
CESM1-BGC 11
CESM1-FASTCHEM3  
CMCC-CM   
CNRM-CM591 
CSIRO-Mk3-6-0101010
CanESM2555
FGOALS-g2   
FGOALS-s2333
GFDL-CM34  
GFDL-ESM2G 11
GFDL-ESM2M 11
GISS-E2-R42 
HadCM310  
HadGEM2-ES244
IPSL-CM5A-LR644
IPSL-CM5A-MR311
IPSL-CM5B-LR 11
MIROC-ESM311
MIROC-ESM-CHEM 11
MIROC5533
MPI-ESM-LR333
MPI-ESM-MR331
MRI-CGCM3511
NorESM1-M311
bcc-csm1-1311
inmcm4 11

The newly developed “representative concentration pathways” (RCPs) [Moss et al., 2010; Van Vuuren et al., 2011] were used to force future projection simulations, which were included within the core set of runs [Taylor et al., 2012]. RCPs are a set of scenarios with emission, concentration, and land-use trajectories. A set of four pathways were developed as a basis for long-term and near-term modeling. The RCPs are unique in that they allow for the exploration of impacts of different climate policies [Van Vuuren et al., 2011]. The RCPs used for this study are RCP 8.5, similar to a high emissions scenario, and RCP 4.5, which would be considered a midrange emission scenario including mitigation efforts. Although two other RCP scenarios are available, the two chosen for this study were the most readily available scenarios for CMIP5 simulations.

3. Methodology

This study uses the EPI as a metric for the frequency of heavy precipitation events for a given duration and recurrence interval. The United States is examined as a whole (CONUS) and is also broken up into seven subregions (outlined in Figure 2) that are consistent with those used in the U.S. National Climate Assessment. Duration refers to the number of days that precipitation is accumulated over and recurrence interval (return) is an average of the number of years between events. The method of finding the top extreme events is station specific, essentially using a station-specific threshold, and is the same as that described in Kunkel et al. [2003]. For any station time series, the largest event is found and identified as the rank-one event and those days are then omitted from the time series. The second largest event is then found, identified as the rank-two event, and the event day(s) are also omitted. This process continues iteratively until N events have been identified (N being the number of years of data in the time series divided by the return period in years). Thus, the number of extreme events in a given time series depends only on the extreme event return time and the length of the time series.

Figure 2.

Map shows the seven subregions of the United States, over which the EPI is calculated. This choice of regions is consistent with the USGCRP 2009 (and the new ongoing) National Climate Assessment.

For this paper, the EPI is calculated in a manner similar to Kunkel et al. [2007]. The EPI is found from the annual number of extreme events at each station. The annual counts were averaged for all the stations in each grid cell to get an EPI time series for each possible grid cell. If any grid cell contained no stations with usable data, they were omitted from further calculations. Then the EPI for each grid cell was averaged over the regions in Figure 2 or over the continental United States. Interpreting what the EPI tells us is fairly straightforward; if the EPI is increasing, for a given region of observing stations, the number of events is also increasing. Or put more simply, the frequency of such events is increasing.

The EPI is presented as a 10 year running average (to reduce excess noise over individual years) and as anomalies by decade. EPI anomalies are fractional deviations from the long-term mean (LTM), which is the average EPI for the period 1901–1960. Observational-based EPI and EPI anomalies were calculated from 1901 to 2012 and 1901 to 2010, respectively, for an up-to-date picture of extreme precipitation event frequency. Statistical significance in EPI trends was also analyzed using a Poisson's distribution significance test. First, the EPI was calculated for individual grid cells for the time period 1901–2012 for a 2 day 5 year duration and return. Statistical significance (95% confidence level) was found for each grid cell time series using a linear regression fit to a Poisson's distribution.

Historical simulations of precipitation in CMIP5 end in 2005 so the observational-based EPI had to be recalculated for 1901–2005 in order to properly compare it to model simulations, given that the EPI is dependent on the time period for which it is calculated. In order to keep the analysis on a decadal time scale, EPI anomalies were found from 1906 to 2005. The same LTM of 1901–1960 was used for this. In order to get the model median of EPI anomalies, the EPI was calculated from 1901 to 2005 for all ensembles of each model, and then fractional EPI anomalies were found for each ensemble time series. Decadal averages from 1906 to 2005 were found for each ensemble, and all ensembles for each model were averaged, giving one decadal time series for each model. The model median and standard deviation were then calculated across all models. Ensemble runs for each model were also graphed individually using EPI anomalies by decade. A comparison of how well each model performed in respect to observations was done using a correlation coefficient analysis. Observational-based and simulated EPI values were averaged by decade from 1906 to 2005 and then correlation coefficients for the models and observations were calculated for each model.

The analysis of RCP-forced projections was done using the future simulations for RCP 4.5 and RCP 8.5. We analyzed models for which both historical simulations and projections were available (see Table 1) and which contained data for the time period 2006–2100. We developed a similar index methodology based upon the EPI. From the historical simulations we determined thresholds for a given duration and return. Those thresholds were then applied to the projections in order to find the top events exceeding the thresholds. First, the top N events, for a given duration and return, for our historical time period (1901–2005) were found for each grid cell. This was done for every ensemble of every model. The final Nth event (the smallest in magnitude of the extreme events) for each grid cell for each ensemble member was recorded and then averaged over ensemble members per model. These averages will hereby be referred to as “reference thresholds” for the future scenarios. Next, for every ensemble member per model, all events that were larger than the reference threshold were flagged and set to missing from the time series in the same iterative manner used previously. This resulted in a time series of annual counts of extreme events by grid cell for every ensemble member per model. The grid cell time series were then averaged by region using area weighting for each ensemble member per model. The averages of all ensemble members per model were calculated, and the model median and standard deviation were found by decade for all seven subregions and the CONUS.

4. Results

4.1. Historical Extreme Precipitation Observations

Figure 3 shows the 10 year running average of the EPI over the CONUS for the 2 day duration and the 5 year, 10 year, and 20 year return periods for 1901–2012. This updates the EPI analysis to the most current complete data available. An increasing trend can be seen in all three returns. This increasing trend is similar to that found previously by Kunkel et al. [2003] from 1895 to 2000. Figure 4 shows the EPI for the seven subregions of the United States in the same format as Figure 3. The Midwest, South Great Plains, Northeast, and Southeast regions all show an increasing trend in the EPI through 2012 for all three returns. The Northeastern region tends to stand out as having the most drastic increase in the EPI in the most recent decade. The other regions that show an increase more closely match the CONUS in that their increase is more gradual and steady over time. The Northwest, Southwest, and North Great Plains regions have no discernible trend either up or down. It is worth noting that the lack of station density in the Western regions could have an influence on the data in those regions.

Figure 3.

Extreme precipitation index for the period 1901–2012 for the CONUS, shown as 10 year running averages for 2 day 5 year, 2 day 10 year, and 2 day 20 year durations and returns.

Figure 4.

Observational-based EPI for the seven U.S. regions for the time period 1901–2012. EPI is shown as a 10 year running average for 2 day 5 year, 2 day 10 year, and 2 day 20 year durations and returns.

Fractional deviations from the LTM (1901–1960) or “EPI anomalies” show an increasing trend in recent decades for the CONUS from 1901 to 2010. Looking at these anomalies by decade helps to smooth out year-to-year variation that may mask the presence of trends. Results of these EPI anomalies for the CONUS are shown in Figure 5. Anomalies of 2 day 5 year, 10 year, and 20 year duration and returns are plotted and all three returns show agreement of an increasing trend over the last six decades, largely in agreement with Kunkel et al. [2003, 2007]. The lowest values occurred in the 1930s, followed by relatively high values for the 1941–1950 decade. This is followed by comparatively lower values in the next decade and then a steady increase thereafter. The large positive anomalies in the two most recent decades greatly exceed the early period maximum of 1941–1950. Our findings also agree with those of Groisman et al. [2012] who found an overall increase over the past three decades in the frequency of intense precipitation events. They found that the frequency increase was the greatest for extreme rain events with their highest threshold totals. Our analysis and this agreement among studies tell us that not only is the frequency of extreme events increasing but also that increase is greater than that we have observed in the past.

Figure 5.

EPI anomalies, or deviations from the long-term-mean (1901–1960) for the period 1901–2010 shown by decade for 2 day 5 year, 2 day 10 year, and 2 day 20 year durations and returns.

To evaluate the significance of the EPI trends, a statistical significance trend test was done using a linear regression fit to a Poisson's distribution with P < 0.05 considered statistically significant. Figure 6 shows 2 day duration and 5 year return trends in the EPI for each grid cell, for the period 1901–2012, over the United States. Black dots denote no significant trend, red upward pointing triangles denote positive significant trends, and blue downward pointing triangles denote negative significant trends. Triangle size is proportional to the magnitude of the trend. It can be clearly seen in Figure 6 that there is an abundance of positive significant trends in the Middle and Eastern United States. The strongest and most frequent significant negative trends are in the Western United States. This largely supports what we see when plotting the EPI by region. Kunkel et al. [1999] also looked at statistically significant trends (using a Kendal τ statistical test) in the EPI by grid cell for percent anomalies, although for a shorter period than we examined in this study. They found that for the period 1931–1996 (relative to the period mean) most increasing statistically significant trends occurred in the central Great Plains through the Middle Mississippi River extending into the Southern Great Lakes region. They also observed some increases in the Southwest and Southern regions. Our analysis also shows increases in those general areas but a bit more evenly spread out among the whole eastern half of the country and decreasing trends in parts of the Western United States. The relatively minor differences are likely due to the differing period of analysis.

Figure 6.

Statistically significant trends in the EPI by grid cell for the time period 1901–2012 for a 2 day duration and 5 year return. Statistical significance is determined via a linear regression using Poisson's distribution. Black circles represent no significant trend. Red upward pointing triangles represent significant positive trends. Blue downward pointing triangles represent significant negative trends. The size of the triangle depends on the steepness of the slope of the trend.

4.2. Observational Versus Model Output for Extreme Precipitation

EPI anomalies were also calculated for historical model simulations. The EPI itself was calculated in the same manner as observations except for 1901–2005. Each CMIP5 model had a number of ensemble runs. To incorporate all ensembles for each model an average EPI anomaly time series was calculated over all ensembles for each individual model. CMIP5 historical forcing model data stop in 2005, so in order to get a proper comparison the observational-based EPI was recalculated for the same time period as the models. EPI anomalies were calculated from 1906 to 2005. This change in calculation shifts the decades by 5 years compared to the previous figure showing just observations (Figure 5). This causes the previous signal of increased EPI anomalies over the past six decades (for the CONUS) to appear as a four decade increase instead. This comparison was done for the CONUS and all seven regions.

Figure 7 shows EPI anomalies for the model median and observations for the CONUS for 2 day 5 year, 2 day 10 year, and 2 day 20 year durations and returns. Models do capture the increase in EPI anomalies; however, the recent increases are less than that of observations, i.e., the models tend to underestimate the observed trend in extreme precipitation. Figure 7 also highlights the similarities between the various return periods. This underestimation by CMIP5 models supports prior findings by Min et al. [2011] who found that CMIP3 models also tended to underestimate observed increases in extreme precipitation events. This could have further ramifications with projections of extreme precipitation event frequency if models continue this underestimation with projections. Given that this has been found in both the prior generation of CMIP3 models and the current CMIP5 models it is clear that this is a significant and continuing issue with climate models. The spread among models is also large, as shown by the error bars in Figure 7. In fact, model spread is larger than the signal itself for all decades. EPI anomalies of observations and model median for the seven separate regions are shown in Figure 8 for a 2 day 5 year duration and return. Once again there is a large spread among models. The observations show all regions, except for the North Great Plains, with some increase in EPI anomalies for at least the most recent decade. Some regions show a steady increase, like the Southeast, and some a more drastic increase, like the Northeast. The North Great Plains region remains fairly consistent with small anomalies except for a large spike during the 1986–1995 decade. Models seem to capture the overall trend for all regions; however, the signal is again decreased compared to observations. Discrepancies among models and observations in the Western regions could be in part due to the lack of station density in the Western regions of the United States. Also, the North American monsoon affects the Southwest region of the United States and finer resolution regional models tend to struggle with correctly simulating the monsoon [Adams and Comrie, 1997]. Although there have been recent advancements made, difficulties still exist that could influence the results further for that region [Gutzler et al., 2005].

Figure 7.

EPI anomalies for the CONUS for a 2 day 5 year, 10 year, and 20 year duration and returns, shown by decades from 1906 to 2005. Blue bars are observational-based EPI anomalies and red bars are model median EPI anomalies. Error bars are ±1 standard deviation.

Figure 8.

EPI anomalies for the seven regions of the United States for a 2 day 5 year duration and return. Anomalies are shown by decade from 1906 to 2005 for each region. Blue bars are observational-based EPI anomalies and red bars are model median EPI anomalies. Error bars are ±1 standard deviation.

Given the large spread among the CMIP5 models, we also examined them individually in comparison to observations. Through a correlation coefficient test we found that some models agreed much better with observations than others. Figure 9 shows model correlation coefficients for 2 day 5 year, 10 year, and 20 year duration and returns. Each ensemble correlation coefficient is shown separately for each model. This figure highlights the large spread among ensemble members and models. However, most of the models have at least one ensemble member showing a correlation coefficient greater than 0.50. Some models do very well, with correlation coefficients nearing 0.8 or 1.0. Among those are FGOALS-s2 and MPI-ESM-LR, although they have only a small number of ensemble members for each model. At least 8 of the 19 models (depending on return time) have some ensemble runs that give negative correlation coefficients. HadGEM2-ES is the only model that is exclusively negatively correlated to observations. The drastic differences between models and individual ensemble members emphasize the large spread among models seen in previous figures, but also give some insights into which models are causing these variations. The discovery of these variations could call for further analysis in this area to investigate why some models have more variation among ensemble members than others and why there is such a spread among models in general. This spread among models and among ensembles could play a factor in the overall underestimation of increases in extreme precipitation events by CMIP5 models.

Figure 9.

Correlation coefficients for models versus observations over the CONUS for each model; shown for individual ensembles for 2 day duration and 5, 10, and 20 year returns.

To explore individual models further, an analysis of each model's ensemble runs was performed. This was done by simply calculating the EPI anomalies for each ensemble run in the same manner that model and observational-based EPI anomalies were calculated for a 2 day 5 year duration and return. Ensemble EPI anomalies were plotted on individual graphs per model. For example, the HadCM3 model has 10 ensemble members and there is considerable variability in past decades. The most recent decade, however, shows a strong positive anomaly for every ensemble member, which is a sharp difference from most of the other decades. This increase of anomalies in the last decade was not a unique feature to this particular model; in fact, most of the other models showed a similar increase in their ensemble runs for the most recent decade.

4.3. Projections of Future Extreme Precipitation Events

Projections of extreme precipitation events were analyzed in a manner similar to historical simulations using an index based on the EPI. RCP 4.5 and RCP 8.5 were used as forcing scenarios for the CMIP5 model projections. The entire historical period (1901–2005) was used as the reference period for projections.

The event frequency index was calculated for projections from 2005 to 2100 but shown by decade from 2011 to 2100. Figure 10 shows the model median and model spread of event frequency for CONUS by decade from 2011 to 2100 for 2 day durations and all three returns. This index can be interpreted in the same manner as the EPI, showing if storms of a given duration and return are happening more often than they were historically. Figure 10 shows that extreme precipitation events are increasing in frequency out to at least 2100 for all three returns for both RCP 4.5 and RCP 8.5. As expected, RCP 8.5 shows a steeper increase over the period than RCP 4.5. Model spread is not as large, in comparison to the signal, as it was with historical simulations. This shows there is agreement among models that there will be an overall increase in extreme precipitation event frequency in the future.

Figure 10.

Extreme precipitation event frequency for RCP 4.5 (green) and RCP 8.5 (blue) for 2 day duration and 5, 10, and 20 year returns for the CONUS. The event frequency index was calculated for 2006–2100 but decadal anomalies are shown from 2011 to 2100. Error bars are ±1 standard deviation.

When comparing projections of event frequency for the subregions for a 2 day duration and 5 year return (Figure 11), all the regions show an increasing trend for RCP 8.5. The Northern regions and the Midwest region show steeper increasing trends than the other regions for RCP 8.5. The event frequency index for RCP 4.5 is smaller in magnitude in comparison but still shows an increasing trend for most of the regions. The same regions that show a steeper increase (Northeast, Northwest, and Midwest) for RCP 8.5 also show more variability between decades for RCP 4.5. Model variability is a bit larger by region, for some regions, than it is for the CONUS as a whole. However, model spread is still smaller than the signal itself for the majority of the regions and decades. Overall, all regions show an across-the-board increase in extreme precipitation event frequency for RCP 8.5 and most regions show an increase for RCP 4.5 for the future.

Figure 11.

Regional extreme precipitation event frequency for RCP 4.5 (green) and RCP 8.5 (blue) for a 2 day duration and 5 year return. Calculated for 2006–2100 but decadal anomalies begin in 2011. Error bars are ±1 standard deviation.

5. Conclusions

The EPI shows an overall increasing frequency of observed extreme precipitation events over the CONUS from 1901 to 2012. This increasing trend is consistent with prior studies by Kunkel et al. [1999, 2003] and Groisman et al. [2012]. The Midwest, South Great Plains, Northeast, and Southeast regions also show an increasing EPI over time for the same time period. Most regions show a gradual increase except for the Northeast, which sees a more drastic increase in the most recent decade. The North Great Plains, Southwest, and Northwest show no clear trend in the EPI with time. It should be noted that the lack of station coverage in the Western United States could influence some results. When changes in time period and regional divisions are considered Kunkel et al. [1999] showed similar increases by region out to 1997. Observational-based anomalies of the EPI for the CONUS also show a steady increasing trend over the past six decades, which is also consistent with prior studies looking at trends in extreme precipitation event frequency [Kunkel et al., 2003, 2007; Groisman et al. 2012]. When the EPI is examined by grid cell, the majority of statistically significant (to the 95th confidence level) increasing trends can be found in the Eastern United States and the majority of statistically significant decreasing trends in the Western United States, near the coast, for a 2 day and 5 year duration and return. Kunkel et al. [1999] also showed most of the statistically significant increasing trends in the Midwest to Northeastern regions. We found that statistically significant increasing trends are the most frequent and more evenly spread out over most of the Eastern United States. However, these minor differences are likely owing to the differing periods of analysis between the two studies.

CMIP5 model simulations, in comparison to observations, also show an overall increasing trend in the EPI for the CONUS, although with a reduced signal compared to that of the observations. There is much variability when analysis is done by region. However, when shown as EPI anomalies, observations seem to show an increase for at least the most recent decade in all but the North Great Plains. Model signal is reduced compared to observations but overall trends seem to be captured. In general, the CMIP5 models tend to underestimate the observed trends in extreme precipitation, which could potentially lead to further underestimation in projections. This reduced signal by models agrees with a prior study by Min et al. [2011] that showed a similar result using CMIP3 models and thus shows that there is an ongoing trend with models to underestimate the observed increases in extreme precipitation events. Some models perform significantly better than others at simulating extreme precipitation events. Many models have ensemble members that show greater than a 0.5 correlation coefficient over the CONUS for 2 day duration and 5, 10, and 20 year returns. However, many of the models also have ensemble members that show negative correlation coefficients as well. Some models do particularly well with correlation coefficients nearing 0.8 and 1.0 (FGOALS-s2 and MPI-ESM-LR). There is a high variability in EPI anomalies among ensembles with one striking trend; many ensembles show an agreement for an increase in EPI anomalies for the most recent decade. This trend shows that although there is large variability within models and ensemble members, overall there is model agreement that the most recent decade has seen a significant increase in extreme precipitation events.

There are a number of possible causes for the observed increase in extreme precipitation events over the past several decades. One hypothesis is that the observed increase in water vapor in the atmosphere due to overall warmer surface temperatures is leading to an increase in extreme precipitation events [Karl and Trenberth, 2003; Trenberth et al., 2003; Emori and Brown, 2005; Willett et al., 2007; Kunkel et al., 2013]. Willett et al. [2007] found a significant increase globally in specific humidity at the surface. They found that this increase is largely owing to human activities as opposed to natural forcings. They identified that specific humidity has increased owing to increasing temperatures while relative humidity has stayed relatively unchanged. One theory Willett et al. [2007] proposed was that owing to this increase in atmospheric moisture at the surface and lower atmosphere, there could likely be important changes to extreme precipitation. Santer et al. [2007], using CMIP3 simulations, found evidence for a human-induced signal in the Earth's atmospheric moisture content along with the moisture cycling among ocean, land, and atmosphere. In a recent study by Groisman et al. [2012] an increase in the frequency of extreme and intense precipitation days and events was found to be collinear to those connected to changes in mean annual temperature over the Northern Hemisphere.

Extreme precipitation events are also projected to continue to increase in frequency over the CONUS for the future, more so for higher emission scenarios. Regional projections of extreme precipitation events also show overall increases for RCP 8.5, with RCP 4.5 showing a reduced signal for all regions and some variability between decades for certain regions. Some regions do show more of a spread among models than the CONUS as a whole, and there is also variability in the amount of increase among regions. However, all future simulations show an increasing trend in event frequency for RCP 8.5, regardless of region.

Acknowledgments

We acknowledge support in part from NASA project NASA NNX12AF32G on using CMIP5 results in Climate Analyses for the United States. We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modeling groups for producing and making available their model output. For CMIP the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI), and the WCRP's Working Group on Coupled Modelling (WGCM) for their roles in making available the WCRP CMIP3 multimodel data set. Support of this data set is provided by the Office of Science, U.S. Department of Energy. Data can be found at: http://www-pcmdi.llnl.gov/.