Journal of Geophysical Research: Atmospheres

Contribution of tropical cyclones to extreme rainfall events in the southeastern United States

Authors


Abstract

[1] Extreme precipitation has been increasing in the United States over the past century. In light of the associated impacts and possible linkages to climate change, this topic has garnered a great deal of attention from the scientific community and general public. Because tropical cyclones are a common source of heavy rainfall in the southeastern United States, we examined the contribution of tropical cyclone precipitation relative to overall extreme precipitation from all weather systems combined. We used a surface observation network over the period 1972–2007, consisting of first-order and Cooperative Observer Program weather stations. Furthermore, to account for precipitation that may be unmeasured by rain gauges because of windy conditions during tropical cyclones, we employed a wind-corrected data set and the North American Regional Reanalysis. According to several metrics of extreme precipitation, we found that extreme precipitation from tropical cyclones has been increasing over the past few decades. Additionally, the contribution of tropical cyclone precipitation to overall extreme precipitation has been significantly increasing by approximately 5%–10% per decade in the southeastern Atlantic coastal states. We attribute this rise in tropical cyclone contribution to an increase in both the storm wetness (precipitation per storm) and storm frequency over the period of record. There is little evidence that changes in storm duration are responsible for the increase. As such, we believe that an important factor in accurately projecting changes in extreme precipitation rests on whether tropical cyclone activity is driven more by natural decadal oscillations or by large-scale warming of the environment.

1. Introduction

[2] Extreme weather and climate events have received a great deal of attention because of resulting large loss of life and high damage costs [Karl and Easterling, 1999]. As populations shift to areas susceptible to flooding, storm damage, and extreme temperatures, society has become more vulnerable to extreme weather [Kunkel et al., 1999a]. Climate change scenarios anticipate the frequency of these extreme events will continue to increase in a warming environment [Intergovernmental Panel on Climate Change, 2007]. With respect to precipitation, changes in daily precipitation intensities are more likely to be influenced by anthropogenic climate change than the total precipitation amount [Trenberth et al., 2003]. Trends in 1 day and multiday heavy precipitation have been positive for much of the 20th century [Karl and Knight, 1998; Zhai et al., 1999; Kunkel et al., 1999b], and the frequency of 1 to 7 day precipitation totals have been increasing since the 1930s [Zhai et al., 1999; Kunkel et al., 1999b]. As such, changes in extreme precipitation are responsible for much of the 5%–10% increase in total precipitation in the United States over the past century [Karl and Knight, 1998; Groisman et al., 1999, 2001; Kunkel et al., 1999b; Easterling et al., 2000; Intergovernmental Panel on Climate Change, 2001; Semenov and Bengtsson, 2002; Kunkel, 2003; Groisman et al., 2005].

[3] A number of weather systems can produce extreme rainfall, including squall lines, mesoscale convective complexes, and tropical cyclones (TCs). TCs may be increasing in intensity in recent decades because of a positive trend in sea surface temperatures [Emanuel, 2005; Webster et al., 2005] as a result of anthropogenic activity [e.g., Mann and Emanuel, 2006] or natural variability [e.g., Landsea et al., 1999]. As increasing greenhouse gases result in a warming environment, some climate models project rises in TC intensity and near-storm precipitation rates [Knutson and Tuleya, 2004]. Therefore, since a connection has also been established between extreme precipitation and climate change, an interesting question is the extent to which observed changes in TC precipitation contribute to extreme precipitation arising from all systems. Our research seeks to gain a better understanding of this issue.

[4] There have been a few studies that have focused on the influence of TCs to overall warm-season precipitation. Groisman et al. [2004] used a 50.8 mm threshold to define extreme rainfall and identified all extreme rainfall events occurring within the zone of influence of each tropical cyclone over the past century for the extreme coastal regions of the southeastern United States. A storm's zone of influence was defined as a fixed distance from the storm center, independent of the shape of the actual rain shield. They found no significant century-long change in hurricane-related precipitation along the coast, despite the fact that total precipitation and extreme precipitation events unrelated to tropical cyclones increased [Groisman et al., 2004]. Shepherd et al. [2007] used satellite data from the Tropical Rainfall Measuring Mission (TRMM) and a newly defined metric termed the “millimeter-day” to determine the contribution of tropical cyclones to extreme rainfall just off the coastal southeastern United States. The millimeter-day, similar to the heating or cooling degree day, is referenced to the average daily precipitation on Mount Waialeale, Hawaii, one of the wettest locations on the planet. Major hurricanes were strongly correlated with the largest magnitude wet millimeter-days during the tropical cyclone season over the period 1998–2006. Other researchers also identified a strong correlation between increased rain rates and increased intensities of tropical cyclones [Velden, 1989; Rao and MacArthur, 1994; Rodgers et al., 2001]. Furthermore, a few regional studies examined the contribution of TC rainfall to overall extreme rainfall over the North Atlantic and western North Pacific using satellite data [Lau et al., 2008] and in coastal China using a surface observation network [Wu et al., 2007].

[5] In related work, we constructed a surface-based regional climatology of the contribution of tropical cyclones to the total hurricane season rainfall for 1980–2004 in the southeastern United States [Knight and Davis, 2007]. Temporal changes in tropical cyclone-induced rainfall were among the study's most interesting findings. Tropical cyclone rainfall increased significantly at 33% of stations, while only 5% of the stations exhibited significant trends in nontropical cyclone rainfall. In short, we showed that there is a positive trend in tropical cyclone rainfall in the southeastern United States. However, we did not examine tropical cyclone rainfall in terms of daily intensity in this climatology. Therefore, this raised a new question asking how much of the trend in extreme rainfall can be accounted for by tropical cyclones. In this work, we use both observational and reanalysis data from the entire southeastern United States, including inland areas, to explore TC rainfall trends. We present a variety of metrics to examine changes in the contribution of rainfall from TCs to the overall extreme rainfall in the region.

2. Data and Methods

[6] We used a network of 85 surface observation stations across the southeastern United States in this study (Figure 1 and Table 1). This area was chosen to represent the states that are commonly influenced by tropical systems originating in the North Atlantic Ocean, Caribbean Sea, and Gulf of Mexico [e.g., Knight and Davis, 2007]. We used rainfall observations from first-order weather station rain gauges when possible but also included a few stations from the Cooperative Observer Program (COOP) network to ensure a fairly evenly spaced coverage of the region. Surface rain gauge observation stations were selected to maximize data completeness over the period of record 1972–2007. Though we selected the network with the objective of even coverage, we acknowledge that convective rainfall events exhibit large spatial and temporal variability. As such, spatially uniform rainfall totals from tropical cyclones, for example, would not necessarily be expected [e.g., Keim, 1997]. We obtained surface rainfall observations from the National Climatic Data Center (NCDC) and used local climatological data (LCDs) for first-order stations and COOP forms for lower-order stations. This analysis required daily precipitation totals. There are a few caveats related to the use of daily precipitation. With respect to this study in particular, heavy rainfall from a single storm may occur within a 24 hour window but could be measured over 2 days. For example, heavy precipitation may occur from 2200 LT–0200 LT, and even though it is from the same system, this storm's rainfall would be split between 2 days. It is important to note that this potential problem is not restricted to tropical cyclone situations and would hold true for all precipitation events. Also, various stations measure precipitation within different 24 hour time intervals. Since this is a large-scale study spatially and temporally, this issue should not have a major influence on our results. In this analysis, we do not link the storm total precipitation to each storm but instead focus on the daily precipitation over the course of each storm. For a large spatial study over a long period of time, effects should be random and produce no bias.

Figure 1.

Network of 85 surface weather observation stations. Filled circles with centered crosses indicate first-order stations, and circles indicate Cooperative Observer Program (COOP) stations. Small filled points indicate the North American Regional Reanalysis (NARR) network. Grid points closest to the surface weather observation stations were selected to comprise our NARR data set.

Table 1. Station Network With Abbreviations, Longitude, Latitude, and Elevationa
StateCityStationLongitude (°W)Latitude (°N)Elevation (m)
  • a

    Cooperative Observer Program stations. All other stations are first-order weather stations.

AlabamaBirminghamBHM86.7533.56219
AlabamaHuntsvilleHSV86.7634.64214
AlabamaMobileMOB88.2430.6975
AlabamaMontgomeryMGM86.3932.3075
ArkansasFt. SmithFSM94.3735.33160
ArkansasLittle RockLIT92.2234.7389
DelawareDoverDOVa75.4739.1310
DelawareWilmingtonILG75.6139.6827
FloridaDaytona BeachDAB81.0629.1812
FloridaFt. MyersFMY81.8626.596
FloridaGainesvilleGNV82.2729.6952
FloridaJacksonvilleJAX81.6930.4910
FloridaKey WestEYW81.7624.551
FloridaMiamiMIA80.2925.793
FloridaOrlandoMCO81.3128.4333
FloridaPensacolaPNS87.1830.4741
FloridaTallahasseeTLH84.3530.4028
FloridaTampaTPA82.5327.989
FloridaWest Palm BeachPBI80.1026.686
GeorgiaAthensAHN83.3233.95275
GeorgiaAtlantaATL84.4333.64350
GeorgiaAugustaAGS81.9633.3749
GeorgiaBrunswickBQKa81.4731.269
GeorgiaColumbusCSG84.9432.52135
GeorgiaDawsonDAWa84.4031.70107
GeorgiaMaconMCN83.6532.69121
GeorgiaSavannahSAV81.2032.1317
GeorgiaTiftonTMAa83.4931.43121
KentuckyLexingtonLEX84.6138.04334
KentuckyLouisvilleLOU85.6638.23186
LouisianaBaton RougeBTR91.1530.5324
LouisianaLake CharlesLCH93.2230.135
LouisianaNew OrleansMSY90.2629.991
LouisianaShreveportSHV93.8332.4588
MarylandBaltimoreBWI76.6739.1850
MississippiGulfportGPTa89.0730.4110
MississippiJacksonJAN90.0832.31118
MississippiMeridianMEI88.7532.33101
MississippiUniversity (Oxford)UOXa89.5434.38154
New JerseyAtlantic CityACY74.5839.4626
New JerseyNewarkEWR74.1740.696
North CarolinaAshevilleAVL82.5435.44738
North CarolinaCape HatterasHSE75.6235.236
North CarolinaCharlotteCLT80.9435.21255
North CarolinaGreensboroGSO79.9436.10315
North CarolinaRaleigh/DurhamRDU78.7835.88148
North CarolinaWilmingtonILM77.9034.2711
OklahomaOklahoma CityOKC97.6035.39441
OklahomaTulsaTUL95.8936.20231
PennsylvaniaAllentownABE75.4440.65134
PennsylvaniaAvocaAVP75.7241.34328
PennsylvaniaErieERI80.1842.08250
PennsylvaniaHarrisburgCXY/MDT76.7640.19106
PennsylvaniaPhiladelphiaPHL75.2439.8712
PennsylvaniaPittsburghPIT80.2340.49410
PennsylvaniaWilliamsportIPT76.9241.24180
South CarolinaCharlestonCHS80.0432.9016
South CarolinaColumbiaCUB81.0033.9766
South CarolinaGeorgetownGGEa79.3233.3113
TennesseeBristol/Johnson/KingsportTRI82.4136.48518
TennesseeChattanoogaCHA85.2035.04233
TennesseeMemphisMEM89.9835.04116
TennesseeNashvilleBNA86.6836.12204
TexasAbileneABI99.6832.41610
TexasAmarilloAMA101.7135.221229
TexasAustinAUS97.6730.19185
TexasBrownsvilleBRO97.4325.917
TexasCorpus ChristiCRP97.5027.7715
TexasDallas/Ft. WorthDFW97.0432.90207
TexasDel RioDRT100.9329.37341
TexasHoustonIAH95.3429.9833
TexasLubbockLBB101.8233.661119
TexasMidlandMAF102.2031.94978
TexasSan AntonioSAT98.4729.53276
TexasVictoriaVCT96.9228.8539
TexasWacoACT97.2331.61176
TexasWichita FallsSPS98.4933.99347
VirginiaLynchburgLYH79.2037.33320
VirginiaNorfolkORF76.2036.899
VirginiaRichmondRIC77.3237.5157
VirginiaRoanokeROA79.9837.33400
Washington, D. C.DullesIAD77.4638.94107
Washington, D. C.ReaganDCA77.0438.855
West VirginiaCharlestonCRW81.5938.37334
West VirginiaElkinsEKN79.8638.89677

[7] Since this study is focused on precipitation events during tropical cyclones, we attempted to account for the windy conditions that may influence precipitation measurement accuracy. Raindrops fall at an angle in windy conditions, so the effective orifice size of the rain gauge changes [Rinehart, 1983; Hosking et al., 1985]. In addition to rainfall size distribution and gauge design, rainfall catch error depends on the ambient wind speed [e.g., Mueller and Kidder, 1972; Neff, 1977; Folland, 1988; Hanna, 1995; Nespor and Sevruk, 1999; Chang and Harrison, 2005] and is the single largest source of error in precipitation measurements [Legates, 1987, 1992]. This problem of wind-based undercatchment is especially apparent with tropical cyclones [Miller, 1958a]. There is little consensus with regard to a rain-wind relationship in observational studies [e.g., Koschmieder, 1934; Wilson, 1954; Dunn and Miller, 1960; World Meteorological Organization, 1962; Allerup and Madsen, 1979]. Medlin et al. [2007] completed a meta-analysis from these conflicting studies to compare rainfall totals from storms of different intensities (Table 2). We used the correction proposed by Medlin et al. [2007] to estimate precipitation undercatchment caused by wind and applied it to the surface rain gauge data set to create a second precipitation data set, hereafter referred to as the “wind-corrected data set.” The correction is based on the daily average wind speed, as found in the LCDs, over the 1984–2007 period of record. The period of record is shorter in length than the surface rain gauge data set because the daily average wind speed was not consistently reported until 1984. Referring to Table 2, if the daily average wind speed was 7 m/s at a station, for example, the measured precipitation total on that day would be increased by 30% for the wind-corrected data set. This method was applied to the entire year, independent of the presence of a TC. COOP stations' daily precipitation totals were adjusted using the nearest first-order station's daily average wind speed.

Table 2. Underestimation of Rainfall Gauge Measurements According to Various Studies for Different Wind Speedsa
Wind Speed (m/s)Koschmieder [1934] (%)World Meteorological Organization [1962] (%)Wilson [1954] (%)Allerup and Madsen [1979] (%)Dunn and Miller [1960] (%)Estimated Error Used for Storm Comparisons (%)
  • a

    The rightmost column is an average of prior research rounded to the nearest 5% for each wind speed category. The 80% and 100% errors for the highest wind speed categories were subjectively estimated from a combination of errors from a limited number of studies and to maintain continuity with lower speed categories [Medlin et al., 2007].

4–5  10–20  15
5–1012–4015–56   30
10–1540–6750–83 50 60
15–250050100 80
25–35    50100

[8] A key difference between our approach and the one applied by Medlin et al. [2007] is that we used daily average wind speed rather than hourly wind speed. Medlin et al. [2007] adjusted the precipitation that occurred every hour and aggregated these adjusted values to determine the storm total. However, our observation network's wind and precipitation reporting is not at all consistent at the hourly time step for the entire period of record. Therefore, we elected to adjust the rainfall at the daily time step, as this was the next best option available. We completed a sensitivity analysis for Pensacola, Florida during 2004 to evaluate the difference in the hourly versus daily correction approaches. On days during which precipitation could be attributed to TCs, the hourly correction method resulted in a 9% greater precipitation amount than the daily correction method. Timing of precipitation was especially important in determining this sign—heavy rainfall during the windiest conditions resulted in a greater adjusted amount for the hourly method, and the opposite result was found when heavy precipitation occurred in advance of or following the windiest conditions. For the entire year, the hourly correction method resulted in 5% more total precipitation than the daily correction method. Putting this in perspective, however, the corrections shown in Table 2 are rounded to the nearest 5%, and adjustments as found in previous studies vary widely. Therefore, though the hourly method would be preferable, we believe using the daily average wind speed to correct the daily precipitation total is an acceptable substitute in an effort to account for wind loss.

[9] We also used a third precipitation data set generated from the North American Regional Reanalysis (NARR). According to Mesinger et al. [2006], the NARR is a long-term, dynamically consistent, high-frequency, high-resolution atmospheric and land surface hydrology data set that is deemed to be an improvement over previous global reanalysis data sets. The assimilation of observed precipitation from surface stations, converted into latent heat [Lin et al., 1999], is used in the initialization of the National Center for Environmental Prediction Eta model that produces a precipitation product at 32 km resolution [Mesinger et al., 2006]. Therefore, the model precipitation is close to the observed precipitation and is more realistic than if the model was free to forecast precipitation [Mesinger et al., 2006]. We include this data set in our analysis in another attempt to account for high wind conditions during tropical cyclones. Though the surface observations used to initialize the Eta model would be affected by wind, the precipitation amount produced by the model would presumably not be influenced. To adequately compare the NARR data set to the surface rain gauge and wind-corrected data sets, we selected the grid node closest to each observation station in the network shown in Figure 1. The period of record for the NARR data set in this study is 1979–2006.

[10] We identified days during which precipitation was directly or indirectly related to TCs. To do this, we used a suite of information including the National Hurricane Center's Tropical Cyclone Reports (see http://www.nhc.noaa.gov), NOAA's Daily Weather Map series, and the Hydrological Prediction Center's (HPC) tropical rain data set [Roth, 1972–2007] to track storms and determine areas receiving precipitation associated with TCs. The HPC tropical rain data set is a graphical product that displays a contour map of the area receiving rainfall for each storm, based on information from the National Weather Service River Forecast Centers and local forecast offices' posttropical cyclone reports. All precipitation related to the storm measured via rain gauges is included in the storm total. Precipitation that may arise from the influence of nearby fronts or precipitation that is produced after the storm becomes extratropical is included in the HPC's approach (D. Roth, personal communication, 2007). We use the HPC information as a guide to determine the area covered by the rain shield associated with each tropical system. The daily station rainfall for a given tropical cyclone is obtained from the LCDs, COOP forms, and NARR data set. For our purpose, in concurrence with the HPC tropical rain data set, precipitation from extratropical systems merging with TCs were included in our totals, so this approach depicts the maximum potential tropical cyclone-induced precipitation. In contrast, other studies have flagged precipitation as TC-induced if it occurred within a fixed distance from the storm's center, regardless of the actual rain shield [e.g., Groisman et al., 2004]. We did not include precipitation from Pacific tropical cyclones in this study—our emphasis is entirely on rainfall generated from storms originating in the North Atlantic sector. Events terminate at each station once the tropical cyclone moves away from the locale and does not produce measurable precipitation. Figure 2 displays a yearly climatology of the average number of TCs producing precipitation across our study region. If a TC's rain shield produced measurable precipitation at a station, that storm was included in this climatology. TC occurrence is most frequent along coastal regions, and more storms influence stations in the eastern portion than the western portion of the study region.

Figure 2.

Yearly climatology of the number of tropical cyclones (TCs) producing precipitation across the southeastern United States.

[11] There are a number of methods to determine how tropical cyclones contribute to extreme precipitation. We incorporated several of these in our study for all three data sets (surface rain gauge, wind-corrected, and NARR).

2.1. Top 10 Wettest Days

[12] Similar to the approach used by Michaels et al. [2004], we identified the number of TC days that occurred in the top 10 wettest days of each year at each station. We then examined network-wide temporal changes in the percentage of the wettest days of the year that were caused by TCs. This approach seeks to determine how the influence of TCs on the wettest days of the year may have changed over time. This approach does not require the use of fixed bins based on precipitation percentiles, a method that has been identified as potentially problematic in examining climate change [Michaels et al., 2004].

2.2. Frequency of Extreme Precipitation Days

[13] A different approach focused on the frequency of extreme rainfall events [e.g., Kunkel, 2003]. For this method, we counted both the number of TC days and all days that exceeded 5.08 cm (2 in) of precipitation and compared spatial and temporal patterns of each. This value was chosen in accordance with Groisman et al. [2004], but, as shown in subsequent approaches, the contribution of tropical cyclones is greater for stricter definitions of “extreme” precipitation. Rather than focusing on the magnitude of precipitation like some other methods [e.g., Karl and Knight, 1998], this approach generates information regarding the number of extreme events.

2.3. Top Percentile of Precipitation Days

[14] We also focused on the upper 10th percentile of TC and all precipitation in a variety of ways [e.g., Karl and Knight, 1998]. We compared trends in the average of the upper 10th percentile of precipitation on TC days to the upper 10th percentile on all precipitation days for each station. In addition, we observed how the top 10th percentile value fluctuated if TC days were removed from the data sets. For example, removing TC days from Raleigh, North Carolina's precipitation for 1999 reduced the upper 10th percentile value of precipitation by 30%. We did this for each year at each station to determine how the contribution of TCs to extreme rainfall, in terms of the top 10th percentile value, may be changing over time.

2.4. Trends in Percentage Contribution of Tropical Cyclones

[15] We also calculated the percent contribution of TCs to all observed precipitation in the upper 5th, 10th, and 20th percentiles of precipitation amounts for each year (similar to Lau et al. [2008]). We identified temporal changes in this contribution for each percentile at each station. This approach creates a physically intuitive value that indicates how important TC precipitation is to overall extreme precipitation.

[16] To examine the temporal behavior of rainfall in each approach, we used linear regression. Given our relatively small sample size (a single year value), we bootstrapped the regression slopes to develop confidence bounds and evaluated statistical significance [Wilks, 2006]. Specifically, for each station, we randomly drew 36 x-y data pairs (with replacement) from the 36 year sample (36 for the surface rain gauge data set's period of record, and fewer for the wind corrected and NARR data sets). Using these randomly drawn data pairs, we performed a simple least squares regression to determine the slope. We repeated this procedure 10,000 times, thus developing a 10,000-point distribution for the simple least squares regression trend for each station. We determined the two-tailed statistical significance based upon the 2.5th and 97.5th percentile observations in the ordered list of 10,000 regression slopes. The trend was deemed statistically significant if the 2.5th and 97.5th percentile values were both positive or both negative based on a two-tailed hypothesis [Wilks, 2006].

[17] To graphically display all data, we used a linear contour method within the geostatistical analyst of ArcMap Geographical Information System (GIS) (see http://www.esri.com). Although the resulting maps seemed easiest for the reader to visualize the data, it is important to note that these maps generally should not be used to interpolate between stations since precipitation variability over space is often nonlinear and can be influenced by complex terrain.

3. Results

3.1. Approach: Top 10 Wettest Days

[18] We present a yearly climatology of the number of TC days that are within the top 10 wettest days of the year in Figure 3a for the surface rain gauge data set. Not surprisingly, values are highest near the coastline because TC precipitation is more likely to occur in these areas [Knight and Davis, 2007]. In addition, storms generally make landfall on the coast of the Carolinas, eastern Florida, and the Gulf of Mexico stretching from east Texas to the Florida panhandle [Jarrell et al., 2008] (see http://www.nhc.noaa.gov), so we expect to see maxima in these areas. There is an overall maximum in the Carolinas because, in addition to being a high landfall area, storms making landfall along the Gulf Coast recurve toward this region once they become influenced by westerly flow. An interesting feature on the map is the bull's-eye maximum near Tampa, Florida. Storms track just to the west of the Florida peninsula often without necessarily making landfall, and their eastern outer bands can produce precipitation in the Tampa region. Overall, values may seem lower than expected because TCs do not affect every station every year (see Figure 2). However, the spatial pattern observed is fairly intuitive and consistent with the overall TC frequency (Figure 2).

Figure 3.

Yearly average of the number of top 10 wettest days that were TC days for the (a) surface rain gauge data set, (b) wind-corrected data set, and (c) NARR data set.

[19] Figures 3b and 3c display the same analysis for the wind-corrected and NARR data sets, respectively. Each shows a similar pattern to the surface rain gauge network, though the maximum areas expand along the coastlines in particular. As expected, the area of darker gray is most expansive for the wind-corrected data set. Since we applied a correction that increases daily precipitation totals based on wind speed, we would expect larger corrections along coastal areas where storms are most intense. Furthermore, since TCs produce some of the windiest conditions of the year for coastal states, we would expect precipitation on TC days to increase more than on other days once the correction is applied. Thus, the number of high-precipitation days that are associated with TCs may be greater for the wind-corrected data set than the surface rain gauge data set. The NARR data set results are between the surface rain gauge and wind-corrected analyses. The model initialization is based on actual surface observations [Mesinger et al., 2006], but the actual precipitation product would not be influenced by wind. As such, it follows that the area of maximum values along the coast is in between the surface rain gauge and wind corrected data sets. It is also important to note that each data set has a different period of record, so we would not expect an exact match between the three maps.

[20] We also examined how the number of TC days in the top 10 wettest days changed temporally for the whole network (Figure 4). The maximum possible value for each year is 850 days, which would occur if all of a year's top 10 wettest days were TC days at all 85 stations. We found a statistically significant positive trend for the surface rain gauge data set with a slope of 1.4 days per year. This indicates that more of the top 10 wettest days each year occur as a result of TCs across the southeastern United States. We also found statistically significant trends for both the wind-corrected and NARR data sets with slopes of 3.6 and 2.7 days per year, respectively (not shown). Again, each of these data sets has a different period of record, so we would not expect the slopes to align because of the year-to-year variability of storm frequency. However, all three data sets show an increase in the number of TC days occurring on the wettest days of the year.

Figure 4.

Time series and linear regression of the number of top 10 wettest days that were TC days for the southeastern United States for the surface rain gauge data set.

3.2. Approach: Frequency of Extreme Precipitation Days

[21] Our second metric of extreme precipitation focused on the frequency of events that exceed a precipitation threshold of 5.08 cm/d. Figures 5a and 5b show the number of days exceeding this threshold for TC days and all days over the period of record for the surface rain gauge data set, respectively. Each map's class intervals were defined using the quantile method, so membership within each gray scale band on a map is equal—thus, the maps are comparable despite different magnitudes of frequency. Two distinct spatial patterns are evident. Not surprisingly, the TC spatial pattern (Figure 5a) is characterized by higher frequencies of extreme events near the coastline, similar to the TC rainfall climatology established by Knight and Davis [2007]. There is a clear gradient along the Appalachian Mountains which act as a barrier to storms and enhance precipitation locally via orographic uplift [e.g., Schwartz, 1970; Bailey et al., 1975; Clark et al., 1987]. Contrastingly, the overall extreme precipitation frequency pattern (Figure 5b) does not have the same gradient along the Appalachian Mountains, and there is a broader maximum area within the central Gulf states. The overall precipitation map appears to treat the south and southeast as a large, fairly coherent rainfall region, while the TC map coincides with well-known storm tracks. The southern Mississippi River valley has the highest daily average rainfall from thunderstorms in the United States [Changnon, 2001], so the broad maximum area shown in Figure 5b is expected.

Figure 5.

Number of days exceeding 5.08 cm of precipitation per day for (a) TC days and (b) all days for the surface rain gauge data set.

[22] We also determined how the yearly frequencies of both TC and all extreme (greater than 5.08 cm) precipitation events changed temporally over the study region. The frequency of daily TC precipitation over 5.08 cm increased significantly over the 1972–2007 period of record at a rate of 0.73 d/yr. There was no statistically significant change in the frequency of all extreme precipitation days over the same period. The frequency approach applied to the wind-corrected data set resulted in similar findings (not shown). In summary, we found a positive trend in the number of TC extreme precipitation days and no trend in overall extreme precipitation days for both the surface rain gauge and wind-corrected data sets.

[23] The NARR data set had far fewer days eclipse the 5.08 cm/d threshold. Representation of Atlantic hurricanes is a potential weakness of the NARR as its precipitation analysis over the North Atlantic is not as accurate as it is over land [Mesinger et al., 2006]. Because the NARR is a fairly new product, its relative strengths and weaknesses are still being documented [e.g., Mo et al., 2005; West et al., 2007]. Grumm and Holmes [2007] found that though the NARR is a good starting point for precipitation analysis, it likely underestimates regions of heavy precipitation. Sun and Barros [2009] similarly found that the NARR reproduces spatial patterns of heavy precipitation events well but is less accurate in reproducing the magnitude and frequency of heavy events. As such, the NARR climatology would suggest lower extremes than rain gauge climatologies. Our findings are similar, as the overall frequency of daily precipitation events greater than 5.08 cm was lower than in the other data sets. In a sensitivity analysis, reducing the threshold for the NARR to 3.56 cm/d produced maps with more similar spatial patterns to those in Figures 5a and 5b than the patterns using the 5.08 cm threshold. Our results with the NARR tend to confirm previous work [Grumm and Holmes, 2007; Sun and Barros, 2009]—the NARR effectively reproduces extreme precipitation patterns but is less successful in reproducing their magnitudes. A potential cause of this is that the NARR is an area-averaged product, and localized extremes may be smoothed out of the precipitation fields. However, since the NARR is still fairly new, literature focused on its validation is limited [e.g., Mo et al., 2005; West et al., 2007].

[24] Finally, we examined the contribution of TC events to overall extreme (greater than 5.08 cm) precipitation events by taking the ratio of Figures 5a and 5b for the surface rain gauge network. As shown in Figure 6, there is a definite spatial pattern where the ratio is greater in eastern portions of the study region and lesser in western portions. This suggests that TCs are more important in the east with respect to extreme precipitation events than they are in the west. Along the Atlantic coast, TCs have caused approximately 20%–25% of the extreme precipitation events occurring over the past few decades. Relative to the Atlantic coast states, the contribution of TC events to overall extreme rainfall is lower for the Gulf Coast region because the lower Mississippi River valley receives more extreme rainfall from weather systems other than tropical cyclones, such daily ordinary air mass thunderstorms [Changnon, 2001].

Figure 6.

Ratio of the number of TC extreme rainfall events to overall extreme rainfall events (greater than 5.08 cm) for the surface rain gauge network.

3.3. Approach: Top Percentile of Precipitation Days

[25] For a third metric, we focused on the top 10th percentile of TC precipitation and all precipitation to compare how each has changed temporally. Figures 7a and 7b display the trend in the average precipitation of the top 10th percentile for TCs and all precipitation for the surface rain gauge data set, respectively. Both maps were made according to the same scale, so the gradations are directly comparable. The greatest contrasts between the two maps are evident in the eastern portion of the study region. TC extreme rainfall has been increasing more rapidly than the overall extreme rainfall, if the overall extreme rainfall has been increasing at all. There are more statistically significant changes in the TC rainfall case than in total (extreme) rainfall. Only 4 stations showed significant trends in the overall extreme precipitation case—approximately the number expected by random chance. Given the amount of research and IPCC projections showing positive trends across the Southeast in overall total and extreme rainfall, this is a surprising result. However, precipitation trends over much of the Southeast are negative or relatively steady during this study's period of record (Figure 8). The Intergovernmental Panel on Climate Change [2007] indicates that the frequency of heavy precipitation events likely increased because precipitation totals increased significantly from 1900 to 2005. Since total precipitation showed little sign of change during this period of record in the southeastern United States, our finding of no trend in extreme precipitation is consistent with the existing literature.

Figure 7.

Trend (in centimeters per year) in the average precipitation of the top 10th percentile for (a, c) TCs and (b, d) all precipitation for the surface rain gauge data set and NARR data set, respectively. Large filled triangles indicate stations that exhibit a statistically significant trend at the 95% confidence level based on bootstrapped linear regression.

Figure 8.

Trend in total precipitation, 1972–2007. Large filled triangles indicate statistically significant changes at the 95% confidence level based on bootstrapped linear regression.

[26] The same percentile analysis is shown for the NARR data set in Figures 7c and 7d. There are many more statistically significant positive trends in the TC extreme precipitation for the NARR data set than with the surface rain gauge data set. In addition, the overall extreme precipitation increases significantly over a larger portion of the study region than with the surface rain gauge data set. This finding is consistent with prior research indicating that total precipitation over landmasses has increased over the past century [Bradley et al., 1987; Diaz et al., 1989; Dai et al., 1997; Karl and Knight, 1998; New et al., 2001], much of which can be attributed to an increase in the number of extreme precipitation events [Karl et al., 1995a; Karl and Knight, 1998]. Nonetheless, comparing Figures 7c and 7d, we note that the TC extreme rainfall has been increasing more rapidly than the overall extreme rainfall, similar to findings with the surface rain gauge data set. There are distinct differences between Figures 7a and 7b and between Figures 7c and 7d. Disparities between the two data sets can be explained by the varying periods of record—the surface rain gauge data set includes 1972–2007 and the NARR analysis is from 1979 to 2006. Precipitation in the United States shows large decadal variability. During the early and mid 1970s, the area average of the annual precipitation over the contiguous United States was above normal. In the late 1970s, the departure from normal was approximately zero [Karl et al., 1995b], so it is not surprising that overall positive trends were more evident in the data set beginning in 1979 than in the one beginning in 1972. To account for this discrepancy, we also calculated trends in the surface rain gauge data set from 1979 to 2006. The number of statistically significant changes in this analysis is more like that in Figure 7c rather than 7a, so period of record is an important factor in determining the magnitude or, in some cases, the sign of overall extreme precipitation trends.

[27] Extreme TC precipitation is increasing at a faster rate than overall extreme precipitation. For the NARR data set, we detrended overall extreme precipitation (Figure 7d) for trends in TC extreme precipitation (Figure 7c). We only did this for stations displaying statistically significant changes in Figure 7d. On average, we found that TCs account for one-third of the trend in overall extreme precipitation at these stations. Because extreme precipitation has not changed over the 1972–2007 period of record, we did not complete a similar analysis for the surface rain gauge data set.

[28] We next examined how the upper 10th percentile value of precipitation would change if TC days were removed from the record. Figure 9a shows this analysis for the entire period of record of the surface rain gauge data set. The largest reductions are along the Atlantic and Gulf coastlines and interior regions of the Carolinas and Virginia. The upper 10th percentile value would decrease by approximately 5%–10% in these areas if TC precipitation days were removed. This map is quite similar to the maps shown in Figure 2 because this analysis essentially removes TC precipitation days that were in the top 10 wettest days of the year. However, this approach creates a time series with a value other than zero every year—the previous approach assigned a zero if there were no TC precipitation days in a given year. We ran a regression analysis on this time series at each station, where the percent reduction of the upper 10th percentile value after the removal of TC precipitation days is the variable of interest (Figure 9b). The percent reduction of the upper 10th percentile value has increased over time over much of the southeastern United States. In other words, precipitation generated by TCs is contributing an increasing percentage of the overall extreme precipitation over time. This result corroborates previous studies that focused on different regions and used different methods [e.g., Shepherd et al., 2007; Lau et al., 2008]. Both the wind-corrected and NARR data sets show similar spatial patterns to the surface rain gauge data set results (not shown), though magnitudes and number of significant stations vary as before.

Figure 9.

(a) Percentage reduction of the upper 10th percentile value of precipitation if TC precipitation days are removed from the overall period of record for the surface rain gauge. (b) Temporal trend (percentage per year) in the percentage reduction (1972–2007). Large filled triangles indicate stations that exhibit a statistically significant trend at the 95% confidence level based on bootstrapped linear regression.

3.4. Approach: Trends in Percentage Contribution of TCs

[29] The final approach uses the metric of percent contribution of TC precipitation to overall extreme precipitation at the upper 20th, 10th, and 5th percentiles. Figures 10a10c show how the percent contribution has changed temporally at each of these percentiles for the surface rain gauge data set, respectively. The gradations are consistent between the three percentiles, so direct comparisons can be made across maps. The wind-corrected and NARR data sets had similar findings after accounting for their differing periods of record, so they are not displayed. With the exception of western and central Texas, the contribution of TC precipitation to overall extreme precipitation at each percentile has been increasing over the 1972–2007 period of record. There is a distinct band of stations with an increasing TC contribution stretching from Florida along the Atlantic coast to Virginia in all three percentiles.

Figure 10.

Temporal trend of the percentage contribution of TC precipitation to overall precipitation within the (a) upper 20th percentile, (b) 10th percentile, and (c) 5th percentile for the surface observation data set (1972–2007). Large filled triangles indicate stations that exhibit a statistically significant trend at the 95% confidence level based on bootstrapped linear regression.

[30] Contrasting the three maps, the magnitude of the trend generally increases as precipitation becomes more extreme. Using Norfolk, Virginia as an example, there is a 6.0% per decade trend at the 20th percentile, a 7.8% per decade trend at the 10th percentile, and a 9.9% per decade trend at the 5th percentile. In general, this pattern of an increasing trend value is apparent across the study region moving from Figures 10a and 10b (20th percentile to 10th percentile) and also from 10b to 10c (10th percentile to 5th percentile). Therefore, the contribution of TC rainfall to extreme rainfall is increasing more rapidly as the definition of extreme becomes stricter. This finding is consistent with Shepherd et al. [2007] who developed the wet millimeter-day to examine extreme rainfall in four minibasins near the coastal southeastern United States. They found that TC rain accounted for more wet millimeter-days than non-TC rain and that strong TCs accounted for the highest wet millimeter-days in the study.

4. Discussion

[31] We used a variety of metrics to understand how tropical cyclones contribute to extreme rainfall in the southeastern United States. A brief summary of the unique findings from each approach is shown in Table 3. We consistently found that both the frequency and magnitude of extreme TC events are increasing faster than the frequency and magnitude of all extreme precipitation events. Furthermore, precipitation arising from tropical cyclones has accounted for a greater percentage of the overall extreme precipitation in the southeastern United States throughout the past few decades.

Table 3. Summary of Results From Each Metric Used to Examine Extreme Precipitationa
Extreme Precipitation MetricMajor Findings
  • a

    TC, tropical cyclone.

Top 10 wettest daysTCs are responsible for more of the top 10 wettest days in coastal areas than inland areas.
The number of TC days comprising the top 10 wettest days of the year has increased.
Frequency of extreme precipitation daysMaps showing the frequency of extreme events from all events show broad regional patterns, but maps of TC extreme events show storm-track patterns.
The frequency of extreme TC events has increased, but there has been no change in the frequency of all extreme events over the period of record.
TCs have caused 20%–25% of the extreme precipitation events along the East Coast.
Top percentile of precipitation daysThe upper 10th percentile of TC rainfall has been increasing more rapidly than overall extreme rainfall.
The upper 10th percentile value of precipitation would be reduced by 5%–10% if TCs were removed.
Precipitation from TCs is becoming more important to the overall extreme precipitation in terms of magnitude.
Trends in percentage contribution of TCsThe contribution of TC precipitation to overall extreme precipitation has been increasing.
The contribution of TC precipitation to overall extreme precipitation increases more rapidly as the definition of “extreme” becomes stricter.

[32] Our results are comparable to those of Lau et al. [2008] who used satellite data to determine the contribution of TCs to extreme rainfall in the North Atlantic and northwestern Pacific basins. Using the surface rain gauge data set, we truncated our period of record to 1979–2005 to match that of Lau et al. [2008] and then calculated a network-wide average of the temporal trend over this new period of record (Table 4). The percentages from our work are comparable in magnitude to the northwestern Pacific and less than those in the North Atlantic. There are a few potential reasons for differences between this work and Lau et al. [2008]. First, the data set used by Lau et al. [2008] only includes June–November precipitation, whereas we include the entire year. As such, there is a greater likelihood of having extreme precipitation on a non-TC day in our study because our analysis is not restricted to the hurricane season. Second, for the North Atlantic region in particular, the percent contribution of TC precipitation to all precipitation in Lau et al.'s [2008] study has been shown to be greater than other studies using satellite data [Rodgers et al., 2001; Shepherd et al., 2007], so methodological differences may be responsible for differing trends. Third, and most importantly, our study region includes land areas. Tropical cyclones weaken after making landfall because they move away from their oceanic heat source [Bergeron, 1954; Hubert, 1955; Miller, 1964], and rainfall intensity tends to coincide with storm intensity [Shepherd et al., 2007]. This is evident in all of our analyses that show a greater contribution of TCs to extreme rainfall in coastal versus inland areas. As expected, the individual station trends shown in Figures 10a10c approach the average North Atlantic trend as the stations' distance from the coast decreases. It is important to note that the trends in Figures 10a10c increase slightly with the adjusted period of record. For example, the trends in Norfolk, Virginia would increase to 7% per decade at the 20th percentile, 9% per decade at the 10th percentile, and 12% per decade at the 5th percentile for the period coinciding with Lau et al. [2008].

Table 4. Temporal Trend of the Percentage Contribution of TC Precipitation to the Upper 5th Percentile of All Precipitation, Expressed as the Percentage Change per Decade, Over the Period of Record 1979–2005a
PercentilePercentage
  • a

    The southeastern United States is based on an entire year and the surface observation data set, and the North Atlantic and northwestern Pacific are based on July–November (data from Lau et al. [2008]).

  • b

    From this study.

  • c

    From Lau et al. [2008].

Southeastern United Statesb
20th3
10th4
5th5
 
North Atlanticc
20th12
10th15
5th18
 
Northwestern Pacificc
20th3
10th4
5th4

[33] There are several potential explanations for the observed increases in extreme precipitation from TCs. Increases in storm frequency, duration, or wetness (precipitation amount per storm) would all result in a higher likelihood of an extreme precipitation event. Storm frequency has increased over the eastern portion of the study region over the 1972–2007 period of record. There is a distinct band of stations stretching from Florida to Pennsylvania showing statistically significant increases. Changes in the storm duration would also affect the TC contribution to extreme precipitation. Potential explanations for this include varying storm translational speed or storm size—larger, slower-moving storms would result in longer storm precipitation durations at stations, thus providing the opportunity for more TC rainfall. Much of the study region is characterized by trends toward longer-lasting storms, though the Texas panhandle, Arkansas, and the western mid-Atlantic states have had storms of shorter duration. We also analyzed the data for trends in storm wetness, or the average annual amount of precipitation produced per storm, in all three data sets. We found that storms have generally increased in wetness over much of the study region, but, as in previous analyses, the period of record determines the number of stations displaying a statistically significant trend.

[34] To determine the extent to which the storm frequency, duration, and wetness influence the percent contribution of TCs to extreme precipitation, we ran a multiple regression analysis. The TC percent contribution at the upper 20th, 10th, and 5th percentiles for each data set were the dependent variables in each regression analysis. If an independent variable was not significant at the 95% confidence level or it was collinear with another variable (VIF greater than 5), it was removed and the regression was run again to determine the standardized betas (partial correlations). Figure 11 shows the relative importance of each independent variable using the standardized betas in determining the percent contribution of TCs to extreme rainfall at various percentiles for the surface rain gauge data set. Results are similar for the wind-corrected and NARR data sets (not shown). On average, these factors explained approximately 80% of the variability in the percent contribution of TCs to extreme precipitation. Results clearly show that storm duration is not an important factor in determining the percent contribution. It should be noted that when the storm duration was a significant factor, the standardized beta was negative. Storms with longer durations at a station may spread precipitation over more days, so an individual day may not be categorized as an extreme precipitation day.

Figure 11.

Standardized betas (slopes) from multiple regression analyses for the surface rain gauge data set. The TC percentage contribution to extreme precipitation at the upper (a) 20th, (b) 10th, and (c) 5th percentiles are the dependent variables. Independent variables are storm frequency, duration, and wetness, and the betas relative to each other are displayed by the pie charts. For example, in Norfolk, Virginia (Figure 11a), storm wetness had a larger effect on the TC contribution to extreme precipitation than storm frequency since there is more black than gray, and storm duration had no effect since there is no white. If an independent variable was not significant at the 95% confidence level, it was removed and the regression was run again to determine the betas. The duration variable always resulted in a negative beta; this is shown in the pie charts for comparative purposes.

[35] Storm wetness is the most important determinant of TC percent contribution in the west and northeast portions of the study area. In the east, wetness and frequency are fairly equal in determining storm contributions. Previous work has attributed changes in TC-related precipitation to storm frequency [e.g., Landsea et al., 1999; Knight and Davis, 2007; Lau et al., 2008], but there has not been as much of a focus on storm wetness. Our result is consistent with the finding of Shepherd et al. [2007] that strong TCs account for the highest wet millimeter-days. Over our period of record, there has been an increase in the number of intense storms, possibly linked to an increase in sea surface temperatures [Emanuel, 2005; Webster et al., 2005]. As such, this result suggests that storm intensity may be just as important in determining the contributions of TCs to extreme precipitation as storm frequency.

[36] Identification of the causes behind multidecadal trends in tropical cyclone activity has been a contentious topic among atmospheric scientists. It is well understood that sea surface temperatures are directly related to tropical cyclone occurrence and potential intensities of storms [e.g., Palmen, 1948; Miller, 1958b; Wendland, 1977; Merrill, 1988; Zhang et al., 1990; Evans, 1993; DeMaria and Kaplan, 1994; Shapiro and Goldenberg, 1998]. The number of tropical cyclones since the 1970s has increased, perhaps because of a period of high sea surface temperatures as part of the positive phase of the Atlantic Multidecadal Oscillation (AMO) [Goldenberg et al., 2001]. The multidecadal shifts in sea surface temperatures as indicated by the AMO closely resemble changes in Atlantic tropical cyclone activity [Landsea et al., 1999] and are a potential explanation for the observation of increasing TC numbers. Since there has only been one AMO cycle during the observed hurricane record, the reliability of the AMO-hurricane link is uncertain. However, there have been efforts to use climate models to reproduce shifts in Atlantic hurricane activity [e.g., Knight et al., 2006; Zhang and Delworth, 2006].

[37] Other studies have taken a different approach in exploring the tropical cyclone–sea surface temperature link. Mann and Emanuel [2006] attribute tropical cyclone fluctuations to forced large-scale warming (mainly anthropogenically induced) and an offsetting cooling from anthropogenic troposphere aerosol forcing instead of the AMO. Michaels et al. [2006] found that the number of major hurricanes may increase in a warming environment, but the intensity of major hurricanes would not become greater beyond a certain sea surface temperature threshold. Webster et al. [2005] and Emanuel [2005] both find an increase in storm intensity and frequency over the past thirty years and suggest linkages to global warming. Hoyos et al. [2006] also directly link the trend in category 4 and 5 hurricanes to increasing sea surface temperatures. Vecchi and Soden [2007] use the relationship between sea surface temperatures and the potential intensity of storms to reconstruct changes in potential intensity using 20th century sea surface temperatures. Despite having all-time high sea surface temperatures currently, potential intensities peaked in the 1930s and 1950s, and current values resemble the historical average. Nyberg et al. [2007] showed that the recent enhanced hurricane activity since 1995 is not unusual compared to other periods constructed via proxy records of wind shear and sea surface temperatures. On the basis of this constructed data set, variability in the wind shear primarily controlled Atlantic hurricane activity over the past 270 years, and the enhanced recent activity represents a recovery to normal hurricane levels rather than a response to rising sea surface temperatures [Nyberg et al., 2007]. Current research focuses on the oceanic heat content as measured from the sea surface to the depth of the 26°C isotherm (known as the tropical cyclone heat potential). This may be a better predictor than sea surface temperatures for determining the intensification of tropical cyclones [Scharroo et al., 2005], and new satellite products are now being used to approximate this variable [Shepherd and Knutson, 2007]. Finally, a new technique that downscales tropical cyclone climatologies from global models suggests that global warming may reduce hurricane frequency but cause increases in intensity for some locations [Emanuel et al., 2008].

[38] In short, there are many conflicting perspectives on the effect that changing sea surface temperatures may have on tropical cyclones. Some scientists argue that tropical cyclone patterns change with a natural cycle of sea surface temperatures, while others point to anthropogenically induced warming as the reason for increasing sea surface temperatures. Regardless of the cause, if storm frequency and intensity, and thus storm wetness, continue to increase, we would anticipate the contribution of TCs to extreme precipitation to continue to rise as well.

5. Conclusion

[39] We implemented several metrics of extreme precipitation to determine how TCs contribute to the overall extreme precipitation in the southeastern United States. Each of these metrics, including an analysis of the top wettest days, the frequency of extreme days, changes in the top tenth percentile, and trends in the percent contribution of TC rainfall to overall extreme rainfall, shed new light on the overall research goal. However, all metrics consistently showed similar spatial patterns of extreme precipitation. In addition, these approaches all showed that extreme precipitation from TCs is increasing across the southeastern United States, often more rapidly than extreme precipitation from all sources, both in magnitude and frequency. Prior research indicates that total precipitation and thus extreme precipitation has been increasing over the past century. However, total precipitation in the southeastern United States has gone largely unchanged or even decreased since 1972, this study's period of record. Hence, we found little to no change in extreme precipitation over this period and could therefore not draw a conclusion about the extent to which TCs drive extreme precipitation trends since there is no trend. In a shortened period of record using the NARR data set (1979–2006), extreme precipitation increased significantly at approximately one-fifth of the stations. For these stations, TCs account for approximately one-third of the trend in extreme precipitation.

[40] Various metrics provide evidence that the contribution of precipitation from TCs to overall extreme precipitation has become more important over time for all three data sets. We showed positive trends in the percent contribution by as much as 5%–10% per decade in coastal regions. This conclusion is noteworthy because a similar result was recently found using a different methodology in different parts of the world [Lau et al., 2008]. We attribute changes in the contribution to storm wetness in the northeast and west and to both storm wetness and storm frequency in the east over our period of record. It is apparent from other research that storm frequency is an important driver of hurricane precipitation. Examining changes in storm wetness has not been as thoroughly vetted in the literature but is physically intuitive in a warming environment. TC extreme rainfall magnitudes, frequency, and contributions to overall extreme rainfall have increased since 1972, while overall extreme rainfall from all weather systems has gone unchanged. Thus, our results suggest that extreme precipitation contributions from other systems have decreased. If tropical cyclone activity was not increasing over the past few decades, extreme rainfall may have actually declined.

[41] We also made an effort to account for the effects of taking surface measurements in high-wind environments. In addition to a data set with measurements from first-order surface rain gauge stations, we applied a correction based on a station's daily average wind speed. Spatial patterns between the surface rain gauge data set and the wind-corrected data set are generally consistent among the various analyses, though wind-corrected precipitation amounts in coastal regions were higher during TCs because a storm's intensity decreases as it moves inland. Since tropical cyclone-force winds are rare in inland areas, corrections would be lower away from the coast, and there would be minimal differences between the wind-corrected and surface rain gauge data sets. We also used a data set constructed from the NARR and found it to be useful in generating spatial patterns of extreme precipitation. However, consistent with previous research [Grumm and Holmes, 2007; Sun and Barros, 2009], the NARR data set did not represent magnitudes of extreme events as well.

[42] Results in this study are subject to a few uncertainties. Since our methodology includes precipitation after a TC becomes extratropical and merges with a front, our estimates are probably the maximum contribution that TCs could have on extreme precipitation. An improvement to this approach would be the development of an objective way to determine the source of a station's precipitation. Perhaps this approach could involve a back-trajectory analysis at each storms' periphery to determine the origin of air and establish a clear spatial boundary of air originating in the tropics. We also showed that the period of record is important in detecting trends. For example, differences in patterns between the NARR and surface observation data sets were often driven by differing periods of record. Since all of these periods of record begin during a time characterized by a low TC frequency and end in a time of high TC frequency, it would be interesting to extend this study backward several decades to test for century-long trends. However, there is evidence that the hurricane data set is missing storms prior to the implementation of geostationary satellites, so caution should be taken when extending the period of record prior to 1966 [Landsea, 2007]. Furthermore, it would be interesting to identify change points in the time series of the contribution of TCs to overall extreme rainfall. This would determine if there is a stepwise function in accordance to the AMO shift in 1995 or if there is a smooth trend over the period of record. Nonetheless, our results consistently show increasing contributions of TC precipitation to overall extreme precipitation in the southeastern United States. Projections indicate that the frequency of extreme events will continue to rise in a warming environment [Intergovernmental Panel on Climate Change, 2007]. Our work suggests that the accuracy of these projections largely depends on whether TCs are driven more by natural decadal oscillations in the AMO or by large-scale warming of the environment. If it is the former, this work suggests that projections of extreme rainfall may need to better account for TC contributions.

Acknowledgments

[43] We thank Stephan DeWekker, Patricia Wiberg, David Hondula, and Luke Sitka for offering suggestions and ideas throughout this project. Comments from three anonymous reviewers also improved this manuscript. D.K. was funded in part by an AMS Industry Fellowship sponsored by the Lockheed Martin Corporation.