Daily extreme precipitation events, exceeding a threshold for a 1-in-5-yr occurrence, were identified from a network of 935 Cooperative Observer stations for the period of 1908–2009. Each event was assigned a meteorological cause, categorized as extratropical cyclone near a front (FRT), extratropical cyclone near center of low (ETC), tropical cyclone (TC), mesoscale convective system (MCS), air mass (isolated) convection (AMC), North American monsoon (NAM), and upslope flow (USF). The percentage of events ascribed to each cause were 54% for FRT, 24% for ETC, 13% for TC, 5% for MCS, 3% for NAM, 1% for AMC, and 0.1% for USF. On a national scale, there are upward trends in events associated with fronts and tropical cyclones, but no trends for other meteorological causes. On a regional scale, statistically significant upward trends in the frontal category are found in five of the nine regions. For ETCs, there are statistically significant upward trends in the Northeast and east north central. For the NAM category, the trend in the West is upward. The central region has seen an upward trend in events caused by TCs.
Numerous studies have documented increases in U.S. extreme precipitation during the latter part of the twentieth century (e.g., Groisman et al. 2004, 2005; Kunkel et al. 2003, 2007). A recent paper examined the potential contribution of tropical cyclones (TCs) to the observed trends in the occurrence of daily extreme precipitation events, exceeding a threshold for a 1-in-5-yr occurrence (Kunkel et al. 2010). They found that an anomalously high number of events caused by TCs accounted for over one-third of the overall national annual anomaly during the period of 1994–2008. Knight and Davis (2009) also found increases in TC-caused events using another definition of extreme precipitation. The meteorological causes of the remaining extreme precipitation events have not been identified. This paper describes the results of a comprehensive analysis of the meteorological causes of secular variations in extreme precipitation event frequencies.
A set of 935 long-term National Weather Service Cooperative Observer (COOP) stations used for a series of recent studies was employed in this project. Daily extreme precipitation events were identified for each station based on exceedance of the threshold amount for a 1-in-5-yr recurrence interval over the period of 1895–2009. The threshold varies widely across the United States from around 25 mm in parts of the interior west to around 200 mm along the Gulf Coast (Fig. 1; a larger set of 3646 stations, with records spanning the shorter period of 1950–2010, was used in this figure to better illustrate the spatial variations). Because of large spatial variations in station density (see Kunkel et al. 2003, their Fig. 1), the station data were used to create a 1° × 1° gridded dataset of extreme precipitation events to achieve more even representation of areas. Grid “events” are defined as the number of events divided by the number of stations in each grid box. If all stations in a grid box experienced a qualifying event on a particular day, the assigned value for that grid box event was 1.0. Otherwise, the value is the fraction of stations experiencing an event. Both the gridded and station datasets were used for the following classification and analysis efforts.
The COOP dataset is a quite reliable source to evaluate trends in precipitation extremes. The observation equipment (8-in. gauge) and observational procedures have remained constant since the late nineteenth century. There are sources of errors, including recording or digitization errors. However, such errors tend to be random and thus are not a source of bias in trends. Furthermore, Kunkel et al. (2005) subjected the data to a number of quality control processes to detect and correct suspect values. There has been a shift over time in the relative number of stations taking their observations in the morning versus late afternoon. This is known to have an effect on temperature trends, but it is not known whether the distribution of precipitation values is affected. Any such effects could alter daily values but are not likely to affect multiday distributions. In the Kunkel et al. (2003) study, they found that overall trends were similar whether looking at daily amounts or 5-day accumulations. Thus, we do not expect that there are any overall trend biases arising from this shift.
The “precipitation” reports from COOP stations include liquid precipitation or liquid equivalent if all or part of the precipitation is frozen. The identification of extreme events used the precipitation reports and thus some extremes may be snowfall events. An analysis of snowfall data coincident with the precipitation reports indicated that just over 2% of the extreme station events were due wholly or partially to snowfall.
Manual analysis as well as one automated process was used to determine the causes of the extreme precipitation events. It was decided at the beginning of the project that only data that are available through the entire study period would be used, so as to limit bias over time since more observation sites are available later in the time period. Because of these constraints, only pressure (from two reanalyses sources) and temperature data as well as the National Oceanic and Atmospheric Administration (NOAA) U.S. Daily Weather Map Series (http://docs.lib.noaa.gov/rescue/dwm/data_rescue_daily_weather_maps.html) were used to supplement the precipitation records. For example, satellite data were not used to classify mesoscale convective systems (MCSs) since such data are only available from the 1960s onward.
The first step was to identify spatially contiguous precipitation regions (CPR) for each day using the daily gridded dataset of precipitation values that was generated from all COOP data. The CPRs were defined by finding the grid box with the greatest precipitation for each day and then searching for adjacent grid boxes with daily precipitation values greater than 12.5 mm. This search was continued until the values were lower than that threshold. The resulting region consists of contiguous grid boxes with precipitation greater than 12.5 mm, entirely surrounded by grid boxes with precipitation less than 12.5 mm. All of the boxes greater than the 12.5 mm threshold are part of the CPR. This process was repeated until all the grid boxes with precipitation greater than 12.5 mm were assigned to a CPR. The intent of this process was that CPRs represented areas where the precipitation resulted from the same cause throughout the CPR and thus a single evaluation would identify the cause for more than one station event. The somewhat arbitrary threshold was determined empirically by testing the results of the algorithm to identify unique CPRs with a range of thresholds on a small number of days with widespread precipitation. Although the extreme event threshold varies widely across the United States (Fig. 1), the selected fixed threshold of 12.5 mm resulted in CPRs similar to those identified by experts. In climatologically drier regions, some CPRs may not be identified by use of a fixed threshold. However, the only consequence of this is to increase the number of evaluations; there is no impact on the final results.
All COOP sites exhibiting extreme precipitation for a given day were assigned to the CPR in which their grid box resided. This allowed multiple extreme precipitation events caused by a single meteorological system to be classified together. A few COOP sites with extreme events could not be assigned because the grid box containing the station had a grid-averaged precipitation value less than 12.5 mm. These events were compared to precipitation maps for that day. If it still did not appear to be connected to any other substantial precipitation, the cooperative station observer forms, the Climatological Data publication, and other nearby cooperative station observations were double checked to determine if the extreme event was in fact real. The coordinates of the CPRs and COOP sites were then used to plot precipitation maps (Fig. 2). From these, the shape, size, spatial orientation, and other surrounding CPRs aided in the classification of the events.
The second step was to produce daily average surface pressure and temperature maps for the days with the extreme events. Surface pressure maps were computed from two different sources depending on the time period. One is a recent reanalysis effort that utilizes only surface pressure for input and thus is able to extend back into the nineteenth century (Compo et al. 2006, 2011; Whitaker et al. 2004); this source was used for events occurring prior to 1948. When the classification of events began, the reanalysis data only extended back to 1908 and this year was chosen as the initial year of analysis. For events occurring in 1948 and thereafter, NCEP–NCAR reanalysis data (Kalnay et al. 1996) were used; this reanalysis, which incorporates a much more extensive input dataset, was used so that future more in-depth analyses (e.g., upper-air patterns and thermodynamic conditions) could be performed on post-1948 extreme events if desired. Since the surface pressure patterns used in the classifications are constrained by surface pressure observations and both reanalyses use the same set of pressure observations, their patterns should be very similar. Daily average surface pressure maps for the day before, the day of, and the day after the CPR were plotted (Fig. 3). The grid boxes of the CPRs were also overlaid on the maps. The temperature maps were made in the same format (Fig. 4), using the same COOP gridded dataset as was used for the precipitation regions.
In addition to the pressure and temperature maps, NOAA U.S. Daily Weather Maps were included in the classification process. Winds were nearly always available on these maps. In the 1940s, frontal systems were added which allowed an “agreement checking” mechanism. Beyond this, cloud cover and dewpoint temperature observations were available for some years.
Using all these data, the extreme event CPRs were classified by cause. For the 1908–2009 period analyzed, there were 18 322 individual events that were aggregated into 9746 CPRs. The CPRs were divided into yearly groups, and then the order of completion for the years was randomized to avoid trends due to any biases arising from changes over time in the expert judgment process. The CPRs needing classification were mainly divided between two individuals, although unclear cases were considered in conference calls with all the project team individuals (authors of this paper). A large amount of collaboration occurred between the two individuals to help remove most of the bias due to differing decision processes.
The potential causes of the extreme precipitation events were classified into one of the following seven categories: extratropical cyclones (ETCs), fronts (FRTs), North American monsoon (NAM), isolated thunderstorms occurring in convectively unstable air masses that will be denoted as air mass convection (AMC), MCSs, upslope flow precipitation (USF), and TCs. There were certain characteristics that were needed for each category. Fronts are usually associated with ETCs, so these categories are connected. The CPRs caused by fronts were one of the easiest to define. These required a temperature gradient that aligned approximately perpendicular to the long axis of the CPR. Frontal cases were also determined by wind shifts, local minima in the pressure fields, and changes in the dewpoint temperatures, if available, on the daily weather maps. ETC cases were defined as such when an event occurred in close proximity to the low pressure system center and was not aligned with a temperature gradient. Events such as winter west coast storms and nor’easters usually fell into this category. Categorization as an NAM event was subject to several constraints. First, the event had to occur in the southwestern part of the United States and be associated with widespread precipitation in that region. Second, time of occurrence was generally limited to the months of June–September. Additional indicators of an NAM event were low pressure near the Baja California peninsula or high pressure near Colorado or Utah. Air mass convection (Brooks et al. 2003) events were defined as being very small—one or two grid cells— and occurring in warm areas and times of year. Station proximity to mountains and airflow toward these mountains were needed for a CPR to be classified as upslope. The MCS category needed to be separated from frontal systems. While many MCSs are initialized along frontal boundaries (either surface or aloft), they frequently move away as intensification occurs. In practice, the events that were classified as MCSs were characterized by moderate-to-strong southerly winds but not always by anomalously warm temperatures. Because mesoscale convective complexes (MCCs) are a category of MCS, steps were taken to understand and identify these events based on data available for the present study. To aid in learning how to identify MCCs, the MCCs observed and documented in 1981 (Maddox et al. 1982), 1982 (Rodgers et al. 1983), 1983 (Rodgers et al. 1985), 1985 (Augustine and Howard 1988), 1986/87 (Augustine and Howard 1991), 1992/93 (Anderson and Arritt 1998), and 1997/98 (Anderson and Arritt 2001) were examined in temperature anomaly and pressure maps as well as the daily weather maps. The majority of the events identified in these studies did not coincide with heavy precipitation events, so they could not be used to make classifications of events used in the present study. Even with this learning process, it should be noted that an event was often classified as an MCS if no other category was appropriate.
An automated process for determining many of the TC-caused extreme events was used. The National Hurricane Center’s hurricane database archive (HURDAT; Jarvinen et al. 1984; Neumann et al. 1999) was used in the automation. If the extreme event was 5° or less away from the track of a documented tropical cyclone center in HURDAT for a given day, the event was classified as TC. Over 1200 extreme precipitation events caused by Atlantic or eastern Pacific TCs were categorized through this automated process. Some tropical cyclone events were not captured by the automation but were found through the manual analysis. When tropical cyclones interacted with extratropical systems, the classification decision was based on location of the extreme event with regard to the tropical cyclone and the frontal system. A tropical cyclone was deemed extratropical when it developed frontal characteristics.
Occasionally, multiple categories could be identified as potential causes of a CPR. This could be due to either insufficient data to fully determine the cause or could be due to multiple processes giving rise to an event. In these cases, a single category was chosen, reflecting the most likely cause or the apparent primary cause. A hierarchy was used in determining the primary causes. These decisions were typically based on the forcing mechanism scale, with the largest scales identified as the primary causes. For example, frontal systems were normally given priority when an event was also associated with another mechanism. Some CPRs that were classified as frontal events appeared to have been affected by ETC, AMC, USF, or MCS occurrence as well. The second priority cause was ETC. It was usually easily determined if the heavy event was near the center of the low, and thus classified as an ETC event. A combination of NAM, USF, MCS, and AMC could have occurred to cause heavy-event CPRs at certain times in the Southwest (SW) United States. If a heavy precipitation event in the SW United States was within widespread rainfall accompanied by an appropriate flow and pressure fields, it was classified as NAM even though the other classifications could have played a role. When the criteria for NAM events were not present, the forcing factors present were evaluated based on the other classification types. When small-scale isolated CPRs occurred, AMC was chosen. On the other hand, if the flow was perpendicular to the mountains in that area, the CPR was determined to be USF. Convection, especially on a larger scale, is a significant cause of heavy precipitation, but many times would not form without the influence of the other classification types. If no other classification was present and convection looked reasonable, MCS was typically chosen as the final cause of the heavy event.
Once the classification was finished for all the CPRs, the cause associated with a specific CPR was assigned to all of the individual gridbox extreme precipitation events in that CPR.
The following results are all based on the “grid event” data, which should minimize bias that would otherwise arise because of the uneven spatial distribution of stations. The largest single cause of extreme precipitation events in the United States was found to be frontal, accounting for about 54% of all grid events. ETCs are associated with 24% of the events, followed by tropical cyclones at 13% and MCSs at 5%. About 3% of the events are associated with NAM and 1% with air mass convection. Only about 0.3% of the events were found to be caused primarily by upslope flow.
The spatial variability in causes is illustrated in Fig. 5a, which shows the annual percentage breakdown for the primary causes in each of the nine climate regions defined by Karl and Knight (1998). Generally only the causes accounting for the highest percentages are listed; thus, the percentages do not add to 100%. In addition, in the case of those causes that are minor in a national context (USF, AMC, NAM, and TC), percentages are given for those regions where they most frequently occur. In the Northwest (NW) and West (W) regions, ETCs account for 80% or more of the events, with FRTs accounting for most of the rest. The FRT category is the dominant cause in the remaining regions with the exception of the Southeast (SE), where TCs are the most frequent cause. In the continental interior regions of the West North Central (WNC) and East North Central (ENC), the combination of FRTs and ETCs account for around 90% or more of the events, with MCSs the third more frequent cause. TCs are a prominent cause in the Northeast (NE; 36%) and South (S; 17%) and also contribute in the Central (C, 9%) and SW (3%). The NAM is responsible for 21% of the events in the SW. The minor categories of AMC and USF occur primarily in the SE (2%) and SW (2%), respectively.
It should be noted that the percentages for FRT, ETC, and TC are somewhat inflated, since the largest-scale cause was chosen for events with multiple possible causes. However, these events with multiple apparent causes were infrequent. The effect of this hierarchical process, therefore, is thought to be minimal in most locations. However, a few regions may be significantly influenced. For example, clear frontal signatures were often not present in the W and NW regions—due in part to the complex topography and limited over-ocean observations. Events associated with ETCs and associated frontal systems were generally classified as ETCs in these regions.
The seasonal progression of the regional results is shown in Figs. 5b–e. The total percentage of events occurring in the winter (Fig. 5b) is high in the W and NW; quite low in the SW, S, C, NE, and SE; and insignificant in the interior regions of WNC and ENC. ETCs are the dominant cause in the western regions (W, NW, and SW) and FRTs are the primary cause elsewhere. In the spring (Fig. 5c), the total percentages of events increase, relative to winter, in all regions except the NW and W. The most frequently occurring causes remain the same except in the SW, where the FRT category replaces ETCs. MCSs make contributions in the S (14%), C (4%), and SE (8%). AMCs make a minor contribution in the SE (3%). The summer percentages (Fig. 5d) are the highest of the four seasons in the WNC (61%), ENC (66%), C (44%), SW (43%), and the NE (46%); the lowest in the NW (16%) and the W (4%); and the causes are the most varied in all regions except the NW. FRTs remain the dominant category in the SW (44%), WNC (70%), ENC (79%), C (73%), S (51%), and NE (49%). TCs are the dominant cause in the SE (58%) and the second most frequent cause in the S (26%) and NE (35%). NAM events are nearly as frequent (41%) as FRT events in the SW. The MCS category is the second most frequent in the ENC (10%) and C (13%) and third most frequent in the WNC (3%), S (15%), and SE (9%). USF events occur most frequently in the summer in the SW (3%). The fall season (Fig. 5e) total percentages are highest of the four seasons in the SE (46%)—where TCs are by far the largest contributor (71%)—and the S (35%). Total percentages are second highest in the NE (44%), S (35%), SW (35%), C (29%), ENC (23%), NW (32%), and W (19%). TCs are also the dominant fall contributor in the NE (44%) and second highest in the S (22%) and C (18%).
The daily gridbox events were first summed for each grid box for each year (the result representing the number of events per station for each year in each grid box), then arithmetically averaged for each cause for each year, for the United States as a whole and for each region. Figures 6 and 7 show time series of the annual averages for each cause. It should be noted that the vertical scales are different between these two figures, illustrating the differing relative frequencies of the causes. There is a sizeable upward trend in the number of events caused by fronts (Fig. 6). There is also an upward trend in the events caused by tropical cyclones, as was discussed in Kunkel et al. (2010) and section 2. For the five other causes, there is not an overall trend.
Table 1 gives the magnitude and statistical significance of the trends for each meteorological cause and nine regions defined by Karl and Knight (1998). Figure 8 shows the time series for the frontal category for the nine regions. Statistically significant upward trends in the frontal category are found in five of the nine regions (Table 1): NE, ENC, C, WNC, and S.
For the six causes other than frontal, regional trends are not statistically significant, with the following exceptions (Table 1). For ETCs, there are statistically significant upward trends in the NE and ENC. For the NAM (monsoon) category, the trend in the West is upward. The Central region has seen an upward trend in events caused by tropical cyclones.
Given the overall upward trend in total events and in events caused by fronts and tropical cyclones, a question arises whether there are more systems causing extreme events or whether there are more extreme events per system. Figure 9 shows a time series of the total annual number of CPRs with at least one extreme event and of the average annual number of events in each CPR. There is a statistically significant (at the p = 0.01 level) upward trend in each of these. The slope is 1.8% per decade for the number of events per CPR and 2.4% per decade for the total number of CPRs with extreme events. A closer examination indicates that the time series for the total number of CPRs with extremes is characterized by a step increase around 1940 and, in fact, the trend since 1940 is not statistically significant. However, the number of events per CPR, while exhibiting substantial interannual variability, is quasi-linear and the trend is statistically significant both for the entire period and the period after 1940. Although the above analysis examined the overall statistics for all extremes, the results for the frontally caused events is similar (not shown). The number of events per CPR for tropical cyclone events (Fig. 10) is approximately double that for all CPRs identified in this study, and also exhibits a statistically (at the p = 0.01 level) significant increase.
The assignment of a meteorological cause to the thousands of extreme events was a very large undertaking, but has now been completed for the period of 1908–2009. These results are based on consistently applied definitions of causes described earlier and the ability to identify the causes from the available data. The following key points were identified:
The largest single cause of extreme precipitation events in the United States was found to be frontal, accounting for about 54% of all grid events.
ETCs are associated with 24% of the events, followed by tropical cyclones at 13% and MCSs at 5%. About 3% of the events are associated with NAM and 1% with air mass convection. Only about 0.3% of the events were found to be caused primarily by upslope flow.
In the Northwest and West regions, ETCs account for 80% or more of the events. The FRT category is the dominant cause in the remaining regions with the exception of the Southeast, where TCs are the most frequent cause. MCSs are the third most frequent cause in the West North Central and East North Central. TCs are a prominent cause in the Northeast and South. The NAM is responsible for 21% of the events in the Southwest. The minor categories of AMC and USF occur primarily in the Southeast (2%) and Southwest (2%), respectively.
The upward trends appear to be primarily driven by increases in events caused by fronts and tropical cyclones. Statistically significant upward trends in the frontal category are found in five of the nine regions, mainly in central and northern regions.
For ETCs, there are statistically significant upward trends in the Northeast and East North Central. For the NAM category, the trend in the West is upward.
The Central region has seen an upward trend in events caused by tropical cyclones. Landfalling tropical cyclones have not increased (Kunkel et al. 2008, their Fig. 2.17), while the number of events per contiguous precipitation region associated with TCs has increased. It is not known whether there is a trend in extratropical cyclone frequency (and associated fronts), but is the subject of a separate investigation.
The potential role of water vapor trends is also being investigated.
This work was partially supported by National Oceanic and Atmospheric Administration Climate Program Office award NA07OAR4310063. We thank Anthony Arguez for helpful discussions during project planning. Any opinions, findings, and conclusions are those of the authors and do not necessarily reflect the views of NOAA or the institutions for which they work.
Current affiliation: Department of Atmospheric Sciences, Colorado State University, Fort Collins, Colorado.