1. Introduction
Flash flood forecasting presents many challenges. Quantitative precipitation forecasting remains a challenging forecasting task itself (Fritsch and Carbone 2004), yet flash flood forecasting also combine hydrological, topographical, and anthropogenic features that contribute uncertainty and nonlinearity, complicating the decision-making process. The National Weather Service (NWS) defines a flash flood as a flood that occurs within 6 h of the onset of the causative event (NWS 2006). This short lead time, combined with the complexity inherent to flash floods themselves, make it difficult to anticipate exactly where and when a flash flood will occur. Further compounding the forecast problem, flash floods are among the most dangerous natural hazards, with the potential to take human lives and cause extensive property damage.
Flash floods often affect the northeastern United States. Perhaps the most notorious flash floods in the region are the those in Johnstown, Pennsylvania, in 1889 and 1977 (Bosart and Sanders 1981; Zhang and Fritsch 1986; Zhang and Fritsch 1987; LaPenta et al. 1995). Part of the motivation for this study was the flash flooding within the Binghamton, New York (BGM), County Warning Area (CWA) during the summer of 2003, when three flash floods resulted in five deaths and over $10 million (2003 dollars) in damage (NOAA 2003). The BGM CWA, which comprises portions of northeast Pennsylvania and central New York, is particularly susceptible to flash floods owing to the presence of large population centers near the Susquehanna River and its tributaries, and the presence of varying topography in portions of the Catskill and Pocono Mountains. LaPenta et al. (1995) point out that New York and Pennsylvania experienced the most flood events among the northeastern states during the period 1955–88.
The meteorological literature offers numerous studies of flash floods. In a seminal paper in the flash flood literature, Maddox et al. (1979) classify flash flood events and describe the synoptic and mesoscale characteristics of each type. Doswell et al. (1996) take a fundamental approach, identifying the root cause of flash flooding as the combination of ample atmospheric moisture and a means of precipitating a large amount of this moisture over a single basin. In general, the lifting mechanisms to produce vigorous convection and, consequently, substantial precipitation include low-level convergence (often associated with frontal or mesoscale boundaries and low-level jets) and upslope flow (Doswell et al. 1996). The means of concentrating this precipitation in a small area include slow cell motion and cell training (Corfidi et al. 1996).
Flash flooding also involves the surface response to precipitation. The hydrological literature discusses the rainfall–runoff relationship, surface processes, and flood-routing mechanisms (e.g., Ogden et al. 2000; Smith et al. 2000; Smith et al. 2002; Zhang and Smith 2003). Operationally, hydrological processes are handled through the issuance of flash flood guidance (FFG; Ntelekos et al. 2006). FFG uses a soil moisture accounting model to reduce the soil moisture and stream conditions to an estimated value of average basin-wide precipitation required to induce flash flooding. Droegemeier et al. (2000) provide an overview of the recent state of flash flood forecasting from both a hydrological and a meteorological standpoint.
Although many studies examine flash flooding from a hydrometeorological perspective, few studies to date synthesize the relative contributions of meteorological processes and hydrological processes to flash flooding for a large number of cases. In addition, none of the studies described here compares the conditions of flash floods events with cases in which heavy rain, but not flash flooding, is reported.
This paper examines the environmental conditions prior to flash floods in the BGM CWA, as well as the conditions of similar events in which flash flooding was not reported. The variables evaluated in this study are drawn from the literature, with particular emphasis on those parameters listed in a flash flood checklist used at BGM prior to this study. A summary of the values of quantitative checklist parameters is provided in Table 1. In addition, the checklist describes several larger-scale features that often accompany flash floods. Low-level convergence provides a low-level lifting mechanism (Opitz et al. 1995). Upper-level divergence generates vertical motion via mass continuity (Funk 1991). Moderate vertical velocities support efficient microscale collision–coalescence warm rain processes. A sounding with a long, skinny CAPE is often indicative of these conditions (Davis 2001). However, large vertical velocities are still attainable with moderate instability through baroclinic processes (Emanuel 1982). Weak winds aloft and weak vertical speed shear favor slower storm motion, which may lead to greater rainfall accumulations (Opitz et al. 1995). Finally, low-level equivalent potential temperature patterns and precipitable water reflect the amount and transport of atmospheric moisture in the flash flood–producing storm’s environment (Funk 1991).
This paper will examine the relative importance of these processes through discriminant analysis and examine the potential for false alarms via comparison with similar nonflood events. The ultimate goal is to identify the statistical differences between cases for which flash flooding was or was not reported in order to better understand the relative importance of hydrological and meteorological processes.
Section 2 of this paper discusses the data and parameters and how they are evaluated in this study. Section 3 presents the results of the statistical analysis. Parameter values are compared with their thresholds on the BGM checklist, and comparisons between flood and nonflood events are made. Section 4 presents composite upper-air soundings and discusses these results in the context of the physical processes contributing to the events. Section 5 proposes modifications to the checklist.
2. Data and methodology
a. Case selection
The flash floods were selected from the National Climatic Data Center (NCDC) publication Storm Data (NOAA 1986–2003) for the years 1986–2003. The year 1986 was selected as the starting point owing to the relative lack of flash flood reports prior to this year. Presumably, flash flood events were classified more generally, as floods, prior to 1986. Thirty-six flash flood days labeled as “small stream” or “urban” floods were not included in this study due to the limited impact of these events. The spatial distribution of all flash floods per county from 1986 to 2003 can be found in Fig. 1, and their annual distribution is shown in Fig. 2.
A subset of 50 flash flood days out of 88 total flash flood days from approximately May through October was selected from the flash flood climatology to provide an independent sample of warm-season extratropical convective events. Limiting the selection to warm-season events precluded snowmelt and ice jam events. Also excluded were two flash flood days associated with named tropical cyclones or their remnants, events such as beaver dam breaks in which the meteorological influence was indirect, and multiple events occurring within 1 week of another. This last criterion preserves the independence of the synoptic-scale environments of the events. For those weeks with more than one flash flood day, a representative event was selected on the basis of the number of counties affected (multicounty events took precedence) and, secondarily, on the basis of the location of the flash floods (chosen to preserve the spatial distribution of all flash floods for the 1986–2003 period). This procedure removed 35 flash flood days and affected the temporal distribution of the sampled events, removing more events from the late summer months than from the spring and the fall, as multievent weeks were most common in the summer. Hereafter, the flash flood events included in the analysis will be abbreviated FF.
Three other datasets with the same criteria as above were also selected. The first, a randomly generated dataset (hereafter RAND), establishes a climatological comparison to the flash flood events consisting of an equal sample size and drawn from the same annual distribution. This dataset was constructed by retaining the dates and locations of the flash floods but randomly assigning a nonflood year between 1986 and 2003 to each event.
Another dataset consisted of 34 significant precipitation events between 1986 and 2003 that did not occur within 1 week of a reported flash flood or each other (hereafter SP). These cases were selected using data from the National Oceanic and Atmospheric Administration (NOAA) Hourly Precipitation Dataset (HPD; NCDC 2005). An SP event was defined as an hourly report of at least 2.54 cm (1 in.) occurring within a consecutive 6-h period reporting a total of at least 3.8 cm (1.5 in.) of precipitation. These thresholds were selected such that the sample size of the SP events was comparable to that of the FF events. These are events that were able to generate heavy rainfall rates conducive to flash flooding, but unable to sustain these rates long enough to generate sufficient runoff for flooding. In essence, these events are the heaviest observed rainfalls that did not produce flash flooding. The hourly precipitation stations are more numerous in the southern portion of the BGM CWA than in the northern portion, so there is a slight geographical bias in this dataset.
The final dataset consists of 17 flash flood watches issued for the BGM CWA between 1995 and 2003 that did not result in a flash flood report (hereafter WATCH). All WATCH events were separated from flash flood reports and each other by at least 6 days. This eliminated 32 nonverifying WATCH days. Counties reporting the largest hourly precipitation totals in the HPD were used to resolve the spatial and temporal scales of the watches to a one- to three-county area over a 6-h time period. This reduced their temporal and spatial scales such that they were comparable to the FF events. The WATCH dataset is considered representative of days on which the forecast meteorological and hydrological conditions satisfied the forecasters’ paradigm of flash flooding, yet flash flooding was not reported.
It is assumed that flash flooding did not occur during the SP and WATCH events. However, it is possible that flooding meeting the flash flood criteria occurred, but went unreported. This potentially introduces some uncertainty into the analysis. The lack of refereed literature examining the reporting biases of flash flooding limits the extent to which this uncertainty can be quantified. Hereafter, the term “nonflood event” refers to the combination of SP and WATCH events.
b. Precipitation data
Precipitation data came from the Northeast Regional Climate Center’s (NRCC) database of Cooperative Observer (COOP) daily precipitation data. A 48-h total summing the daily rainfall reports on the day of and the day after each flash flood report was compiled for every station in each county reporting a flash flood. Use of a 48-h accumulation eliminated complications resulting from different observation schedules. For each FF or nonflood event, the station with the highest storm total in each county represented that county. In constructing the SP and WATCH datasets, the station reporting the highest 48-h storm total in each county from the NRCC database was used, rather than the hourly totals from the HPD, which determined the members of the dataset, to provide consistency with the FF events.
A separate dataset consisted of radar data, which estimated storm total precipitation for FF and nonflood events for the period 1998–2003. This dataset was composed of archive level III hourly and storm total radar data from NCDC (2004). Hourly totals from the location in each county reporting the highest precipitation rates were used to determine the peak intensity of the precipitation, and storm total estimates were summed over the 48-h period used in the rain gauge analysis to supplement the gauge-derived storm totals for each county.
c. Environmental conditions
The National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) North American Regional Reanalysis (NARR; Mesinger et al. 2006) served as the primary dataset. The NARR has a spatial resolution of 32 km × 32 km and a temporal resolution of 3 h, which allows each event to be resolved approximately to the county in which it occurred. The locations of county-based flash flood reports were represented on the NARR grid as the area enclosed by grid points corresponding to the northwestern, northeastern, southwestern, and southeastern boundaries of each county. Meteorological and hydrological parameters of interest were averaged over this initial area at the NARR time preceding either the flash flood report in Storm Data (FF events) or the heaviest hourly rainfall report (nonflood events). The FF event times as reported in Storm Data compared favorably with the time of peak precipitation for nearby hourly rain gauges and, where available, hourly radar observations, suggesting that the flood response time was typically within 3 h of the heaviest precipitation.
A list of the parameters can be found in Table 2. They included measures of local conditions, such as thermodynamic properties, moisture, and winds, as well as synoptic conditions, such as divergence at upper and lower levels and equivalent potential temperature (θe) patterns. NARR-estimated soil moisture parameters in the upper soil layers were included to evaluate the near-surface hydrological conditions prior to each case.
The NARR soil moisture parameters are derived from the Noah land surface model (Mesinger et al. 2006). Although the NARR soil moisture parameters have some problems, the NARR has been shown to notably improve the Palmer drought severity index (Mo and Chelliah 2006). Other studies have also shown that the NARR provides acceptable soil moisture values at much higher spatial and temporal resolutions than are available from in situ data (Mo et al. 2005; Fan et al. 2006).
The NARR soil moisture parameters were supplemented with antecedent precipitation for periods of 2, 4, 7, and 30 days at the rain gauge reporting the highest storm total. Archived FFG issued for the BGM CWA for the years 1997–2003 was also examined (E. Boehmler 2006, personal communication; K. Hlywiak 2006, personal communication). FFG is computed daily, and at more frequent intervals during precipitation events, for each county. Like the NARR parameters, FFG values were averaged over all counties included in multicounty events.
Those variables representing ambient synoptic conditions were not confined to the counties with reported events. Rather, the entire domain defined by the coordinates (49.33°N, 80.71°W), (37.76°N, 85.94°W), (33.41°N, 72.29°W), (43.88°N, 64.70°W) was used. The ridge axis of θe at 850 hPa was determined by finding the maximum θe for each latitudinal row of data points. A regression line was fit to these points, with longitude as the predictor. This regression line approximated the position of the θe ridge. The strength of the θe ridge was determined by the proportion of variance explained (R2) by the above regression. The distance from the regression-based θe ridge axis to the midpoint of the flood quadrilateral (representing the location of the flood) was recorded as a measure of the distance between the ridge axis and the flood.
Divergence was computed for this study by transforming the NARR’s grid-relative winds to earth-relative winds. Divergence was calculated via a centered difference of the grid-relative winds, as the grid is approximately Cartesian for adjacent grid points. Maximum, minimum, and average values of divergence across the study domain at 850 and 200 hPa were included as predictors.
Discriminant analysis and cross validation were used to determine which parameters were best able to differentiate between the flash floods and each nonflood dataset. This approach is similar to that used in DeGaetano et al. (2002). Skill was assessed using the Kuipers skill score (KSS; Wilks 1995). The KSS is given as where I and J are the number of possible forecasts and events, respectively. The probabilities of each forecast and subsequent observation are given by p(yi) and p(oi), while the probability of each possible forecast–observation pair is p(yi, oi). The KSS is a desirable measure of forecast skill since it treats random and constant (e.g., always forecasting zero precipitation) forecasts equally, assigning them a score of 0. A perfect forecast is given a score of 100. The KSS also assigns higher skill to a correct forecast when the alternative forecast is more likely. Separate discriminant analyses were performed for a pairing of each nonflood dataset with the FF events.
Cross validation was used to check the validity of the discriminant classification with regard to independent data. Cross validation is a widely accepted method for obtaining estimates of predictive skill based on independent data, given a limited sample of developmental data. Both the discriminant analysis and the cross validation were performed for each variable, as well as for all possible combinations of two and three variables.
d. Soundings
Composites of upper-air soundings were constructed to supplement the NARR with in situ data. Upper-air data were collected from online data archives. Data were available for all but two FF events and two SP cases. Because the BGM CWA occupies a region centered among several upper-air stations (Buffalo, New York; Albany, New York; Pittsburgh, Pennsylvania; and Upton, New York), the station most closely corresponding to the upwind direction with respect to the NARR’s low-level flow prior to the event was selected for each case. The most recent sounding prior to the time of the flash flood report (FF events) or the time of the largest hourly rainfall report (nonflood cases) was used. The observations were linearly interpolated at 25-hPa intervals. Mean FF, SP, and WATCH soundings were then constructed from the interpolated temperature, dewpoint, and wind profiles for each upper-air station. This allowed a comparison of the upper-air conditions for each type of event based on the direction of low-level flow.
3. Results
a. Precipitation
Flash flooding tends to be associated with higher storm total precipitation than nonflood events (Fig. 3). This tendency is apparent in daily rain gauge data, for which sparse coverage restricts the resolution of the data, and in radar data, for which a comparatively short period of record exists. Rain gauge reports of greater than 7.6 cm were associated with 40% of the FF cases and with only 6% of nonflood cases. The gauged storm totals for approximately 25% of the FF events were less than 5 cm, and 5 FFs reported a storm total of less than 2.5 cm. In contrast, 62% of the nonfloods reported less than 5 cm, and 23% of the nonfloods reported less than 2.5 cm. For moderate precipitation totals of between 4 and 7.5 cm, FFs and nonfloods were equally likely. In these moderate precipitation cases, FFs were more common in the spring and early summer, while nonfloods were more common from midsummer through the fall (Fig. 3). As Sivapalan et al. (2005) have shown, flood frequency is subject to seasonality, such that the seasonal changes of soil moisture and vegetation can cause flash flooding at a lower precipitation threshold for part of the year.
Radar precipitation estimates provide an objective means of assessing the relative precipitation totals of FF and nonflood events with greater spatial resolution than did the rain gauge network. The results shown in Fig. 4 suggest that rain gauges often underestimate precipitation, due to the poor spatial resolution of the rain gauge network. It is also possible that the radar has overestimated the precipitation totals in some cases. In general, the FF events showed an even greater precipitation disparity in the radar precipitation estimates than in the rain gauge estimates. The radar storm total estimates for the FF events averaged approximately 7.9 cm, while those of the nonflood events averaged approximately 6.2 cm. FFs were also associated with higher rainfall rates than the nonflood events. Hourly radar products (not shown) show larger maximum hourly precipitation rates for FF events, especially in comparison to WATCH events. All WATCH events were associated with rainfall rates of less than 5 cm h−1. However, approximately 24% of FF events and 19% of SP events exceeded this threshold.
b. Checklist evaluation
The primary objective of the flash flood checklist that has been in place for the BGM CWA is to heighten forecaster awareness by calling attention to parameters that tend to depart from normal and to patterns that tend to occur on flash flood days (M. Evans 2006, personal communication). Thus, a good starting point for validating the discriminant analysis in the context of the checklist is a comparison of the FF events with the RAND events (Table 3). Measures of atmospheric moisture, including precipitable water and relative humidity, particularly at upper levels, were often much higher for FF events and are the primary factor that differentiates the FF events from the RAND events. The two moisture parameters on the checklist, precipitable water and 1000–500-hPa mean RH, achieve the highest KSS. The FF checklist thresholds for each parameter captured a majority of the FF events, with 82% of FFs above the mean RH threshold and 67% of floods above the precipitable water threshold. In general, the nonflood (i.e., SP and WATCH) days were also well above normal in terms of atmospheric moisture (see Table 4).
Parameters associated with vertical motion differed significantly from their climatological values on FF and nonflood days. Stability indices indicative of potential parcel movement, that is, surface-based CAPE, K index, and lifted index, were among the best predictors, as there was typically above-normal instability during both FF and nonflood events. Similarly, upper-level divergence was typically larger for all three types of precipitation events (Table 4). Lower-tropospheric vertical velocities, calculated in the NARR via mass conservation (G. Manikin 2006, personal communication), also demonstrated that the precipitation events were associated with greater vertical motion (Tables 3 and 4).
In addition to the meteorological parameters, soil moisture in the uppermost 10 cm exhibited one of the largest differences between the FF and RAND datasets, as FF events were associated with greater soil moisture than the climatological sample. Soil moisture deeper in the soil (10–40 cm) and weekly antecedent precipitation exhibited less skill, but remained among the better predictors (Table 3). The nonflood events also had greater soil moisture and antecedent precipitation in comparison with the RAND dataset.
Although the discriminant analysis did not pick up the distinction, wind directions often differed, with a northerly component more common for RAND events, and southeasterly to southwesterly low-level directions favoring moisture transport from the Atlantic Ocean and the Gulf of Mexico more common for FF, SP, and WATCH events (Fig. 5). An easterly wind component was also more common at lower levels for FF events than for RAND events.
Although wind directions differed, wind speeds were often comparable among all of the datasets, particularly in the lower troposphere (see ordinate in Fig. 5). The suggested threshold for the low-level jet (LLJ) for the BGM CWA prior to this study was 10.3 m s−1 (20 kt), a value exceeded in only 15 of the 51 FF events. This threshold was exceeded in a greater proportion of both sets of nonflood events, as 12 of the 36 SP events and 7 of the 17 WATCH events exceeded this LLJ threshold. This may suggest that the contribution of slower low-level winds to the production of slower storm motion is, in many cases, more important than the contribution of a strong, moist LLJ. Storm motion tended to be comparable in speed but more easterly and southerly in direction for FF events. Directional and speed shears at lower and midlevels were similar among the datasets. Although the sample of FF events does not contradict the wind speed, storm motion, and shear criteria in the checklist, alone these parameters are not strong indicators of flash flood occurrence. Rather, the underlying physical processes associated with them are important to understanding the mechanisms of flash flooding (Doswell et al. 1996). These processes will be examined in more detail in section 4.
In summary, the discriminant analysis succeeds in highlighting the properties of heavy precipitation: ample atmospheric moisture and a means of lifting that moisture above the condensation level. It also confirms the operational flash flood checklist, with the exception of the low-level wind speed. What remains to be shown is how the environmental conditions differ between days on which flash flooding occurred from those on which flash flooding was not reported in spite of favorable conditions and significant precipitation.
c. Comparison of FF and nonflood events
Soil moisture is one of the most skillful discriminators between FF and nonflood events, as FF cases are typically associated with wetter soil conditions (Table 5). This appears to result from the seasonality of soil moisture, as the sampled FF events outnumber the nonfloods in May and October, when soil moisture is often higher than in the summer (Fig. 6). Despite the FF events’ tendency to occur with greater soil moisture, the ability of antecedent precipitation to segregate among the datasets was not as strong. The spatial variability of precipitation and the simple one station per county method of specifying antecedent precipitation contributed to this apparent disparity between moisture in the top soil layer and antecedent precipitation.
The soil moisture parameter that BGM forecasters use operationally, FFG, indicated similar trends. In general, FFG values were comparable for FF and WATCH events, and significantly higher for SP events, indicating drier soil conditions for SP events (Fig. 7). Due to the limited period of record, FFG values were not included in the discriminant analysis; rather, they provided a means of evaluating the validity of the NARR soil moisture parameters. Three-hour FFG values had a correlation of −0.59 with the NARR 0–10-cm soil moisture values and a correlation of −0.73 with the NARR 10 40-cm soil moisture values (Fig. 7). Correlations between the NARR soil moisture parameters and FFG for 1- and 6-h durations were similar. In the absence of FFG, NARR surface soil moisture provides an acceptable surrogate.
Thermodynamically, FF events are often less vigorous convectively than many SP events. Both CAPE and K index were a skillful means of separating FF events from SP events (Table 4). Fewer than one-third of the SP events had a CAPE lower than 250 J kg−1, yet half of the FF events and roughly two-thirds of the WATCH events were in this range. This supports a checklist parameter that specifies a “tall and skinny” CAPE, which tends to yield small to moderate values. The K index also tended to be lower for FF events. Hence, flash flooding appears to favor warm-rain (collision–coalescence) processes, which are most efficient at moderate vertical velocities, as the coalescence process requires time for the rain droplets to grow. High CAPE and the consequentially large vertical velocities eject a high concentration of water vapor to the upper region of the storm, where ice processes and hail production dominate (Davis 2001). The high frequency of low-CAPE events in the WATCH dataset implies that this is also a potential scenario in which the checklist generates flash flood forecasts that do not verify.
Despite low Kuiper skill scores for 850-hPa wind direction between FF and nonflood events, differences in the lower-tropospheric wind field can be observed qualitatively (Fig. 5). The subset of events with an easterly component favored FFs almost exclusively. In contrast, events with westerly to northwesterly 850-hPa winds tended to favor the SP and WATCH categories. In general, wind directions, directional shear, and speed shear demonstrated little skill at other levels. Midlevel wind speeds—500 and 700 hPa—tended to be faster for SP events, contributing to greater midtropospheric wind shear. Wind shear tends to decrease the precipitation efficiency via the entrainment of dry air (Doswell et al. 1996).
Differences in the atmospheric moisture variables were only weakly evident. This result is not surprising, as atmospheric moisture has previously been found to be quite similar regardless of precipitation amount in the northeastern United States (Harnack et al. 2001). Mean RH exhibited minimal skill in differentiating nonflood events from FF events, and precipitable water displayed no skill. In general, WATCH events had the highest mean RH values, followed by FF events, while SP events typically had the smallest mean RH (Table 4). This is due to a difference in the vertical distribution of moisture, as WATCH events tended to be more moist in the lower troposphere, and SP events tended to be drier in the middle troposphere, which is often more indicative of a severe weather environment (Davis 2001).
Notable differences also exist in the 850-hPa θe field. Both the strength of the θe ridge and the location of the θe axis are found to differ, particularly between FF and SP events (Table 4). The FF events typically exhibit a weaker correlation (R2) of the θe axis and a computed θe axis location closer to the BGM CWA in comparison to the nonfloods. Statistically, this was a skillful means of discriminating both SP and WATCH events from the FF events (Table 5).
Composite maps of 850-hPa θe for low, middle, and high values of the R2 parameter (0–0.2, 0.4–0.6, and 0.8–1.0, respectively) provide evidence for both of these trends (Fig. 8 for FF events; Fig. 9 for SP events). The 850-hPa θe maxima in the FF composites tend to be farther west at the latitude of the BGM CWA than those in the SP composites. The θe ridge axis for the FF events with high R2 (Fig. 8c) passes through Virginia and central New York, while that for the SP events (Fig. 9c) lies farther to the east, along the East Coast. Similarly, the 850-hPa θe maximum lies farther to the west (and closer to the BGM CWA) in Figs. 8a and 8b than in Figs. 9a and 9b, respectively.
The tendency for 850-hPa θe to exhibit a lower R2 suggests that the 850-hPa θe ridge is often less defined for FF events, as evidenced by the increased “patchiness” and “waviness” of the low-R2 composites (Figs. 8a and 9a). That this pattern is most common in the FF events suggests that the 850-hPa θe is often more diffuse for FF events. The high R2 FF composite (Fig. 8a) exhibits hints of a narrow, well-defined 850-hPa θe ridge axis that has been somewhat smoothed out as a result of compositing. This is less apparent in the SP composite, which consists of more members and was thus subject to more smoothing of the features of individual events. The FF events also display a weaker 850-hPa θe gradient than do the SP events.
In all three panels of the FF 850-hPa θe–height composite (Fig. 8), a trough is located upwind of the θe ridge, generating positive or neutral θe advection over the BGM CWA. In all three panels of the SP 850-hPa θe–height composite (Fig. 9), a trough is located just to the west of the θe ridge, generating neutral or negative θe advection over the BGM CWA. These observations suggest that in FF situations, θe advection is stronger than in SP cases. Taken as a whole, the properties of the θe field suggest that strong 850-hPa θe advection is less important than a relatively large area of relatively high 850-hPa θe with weak to moderate 850-hPa θe advection in which dry air entrainment will have less influence. It also appears to be important that the location of heaviest precipitation is located near the location of highest 850-hPa θe.
Other qualitative parameters on the checklist are less able to distinguish between FF and nonflood events. Divergence at low and high levels showed low skill in discriminating the FF from the nonflood datasets. Lower-level divergence had weak skill for FF and SP events. In general, temperatures, thickness, and isobaric heights tended to be lower for the FF events owing to differences in their temporal and seasonal distributions compared to the nonflood events.
4. Discussion
Composite soundings are presented to gain further insight into the physical processes involved. Nine FF events with an easterly component were represented by the Upton, New York (OKX), upper-air station (Fig. 10). These events fit the description in Opitz et al. (1995) of convective systems with boundary layer transport of Atlantic Ocean moisture in which low-level warm air advection serves as a forcing mechanism. The distinguishing feature of the OKX sounding is the substantial veering throughout the wind profile, which is indicative of warm-air advection. Directional shear, and veering with height in particular, have also been shown to favor propagation of convective systems. In the presence of an east–west low-level boundary, a veering wind profile will tend to favor low-level convection normal to the boundary and upper-level flow parallel to the boundary, which generates an extended region of heavy precipitation along the boundary (Davis 2001). However, too large of a directional shear may result in significant dry entrainment, reducing the precipitation efficiency (Doswell et al. 1996). This large directional shear is the distinguishing feature of the mean OKX sounding for the two WATCH events with easterly low-level flow (not shown). Some of these cases may be elevated thunderstorms, which typically form north of an east–west boundary and may feature training cells rooted in the boundary layer above the frontal inversion (Colman 1990a, b). These storms feature a strongly baroclinic environment with strong lower- to midtropospheric wind shear and warm-air advection, no surface-based CAPE, and often occur in a hydrostatically stable environment, a description that fits the sounding in Fig. 10 (Colman 1990a).
The mean soundings for the Pittsburgh, Pennsylvania (PIT), upper-air station represent the majority of events—those with a NARR-based southerly to west-southwesterly 850-hPa flow (Figs. 11 –13). These events generally occur in the warm sector of an extratropical cyclone and may represent the lowering of isobaric heights in the middle troposphere as a result of adiabatic ascent (Sanders 1971). The FF events have the lowest mean temperatures and dewpoints throughout the sounding. This is likely a result of the spatial and seasonal differences between this dataset and the nonflood datasets. There was a southern bias to the SP and WATCH events owing to the selection of events based on precipitation amount (see section 2a) and a seasonal bias owing to a higher frequency of FF events in spring and fall.
Although all three PIT soundings have nearly unidirectional wind profiles, the FF profile (Fig. 11) has a somewhat more southerly wind direction than the nonflood profiles, particularly at midlevels. The tendency for more southerly midlevel winds for FF events suggests that midlevel moisture transport from the Gulf of Mexico is somewhat larger during FF events. Differences in speed shear among the three profiles are more pronounced. Midtropospheric vertical speed shear is much larger for SP events (Fig. 12) than FF events. The differences in speed shear have implications for precipitation efficiency. Larger shear may reduce the precipitation rates achieved in these SP events through dry-air entrainment. The slower lower- and middle-tropospheric wind speeds of FF events entail less dry-air entrainment, resulting in greater precipitation efficiency, and create slower storm motion, resulting in greater point rainfall accumulations for equivalent rainfall rates (Cotton 1990). However, this does not guarantee higher precipitation totals, as storm motion does not account for multicellular structures such as cell training. In the upper troposphere, winds average 5 m s−1 stronger for FF cases than WATCH cases (Fig. 13). This suggests a stronger upper-level jet and hence stronger upper-level divergence, resulting in greater forcing for the FF events. It may also indicate a tendency in WATCH cases for upper-level forcing anticipated by forecasters to be weaker than expected.
The sounding observations are generally true for the NARR data as well. A discriminant analysis consisting of only the PIT events (not shown) yielded the greatest differences between the FF and SP events in midlevel directional shear, 500-hPa wind speed, storm motion speed, K index, and CAPE. As discussed above, the greatest differences in the soundings included differences in wind direction and speed (especially at midlevels). Some of the greatest differences in the convective parameters listed in Figs. 11 and 12 lie in the K index, CAPE, and precipitable water. In short, the statistical analysis reveals most of the same trends as a more conventional sounding analysis. However, the soil moisture parameters were similarly skillful in classifying the PIT FF and SP events. While physical processes lend insight into whether a given precipitation event will cause flash flooding, it appears that the surface hydrological conditions play an equally important role.
Current operational procedures reflect this knowledge, as FFG is an integral part of flash flood forecasting (Davis 2001). This work has revealed some trends that may improve our understanding of the hydrological component of flash flooding in the northeastern United States.
Antecedent soil moisture appears to have a seasonal influence in that flash flooding is more common for the BGM CWA in June and less common in July and August (Fig. 2). To a lesser extent, there is also a disparity in flash flood occurrence between May and early autumn (September–October). This is likely associated with the annual soil moisture cycle (Ropelewski and Yarosh 1998), as soil moisture is typically lower in late summer and autumn than in the spring and early summer (see Fig. 6). Furthermore, Sivapalan et al. (2005) show that flood frequencies are greatest when the seasonality of high antecedent soil moisture and high precipitation coincide. In the case of the BGM CWA, this situation favors June, when relatively high climatological soil moisture combines with summertime convection to produce a distinct annual peak in flash flood frequency (Fig. 2).
As De Michele and Salvadori (2002) have shown, flood frequency is directly proportional to antecedent soil moisture. A short-term link between soil moisture and flash flooding appears to be possible in late July–August. The 2- and 4-day antecedent precipitation totals (not shown) are generally higher for FF events than nonflood events during this period. In addition, NARR-based soil moisture differs most significantly for FF events from SP and RAND events during this time (Fig. 6). This was also the time of year during which the coincidence of multiple flash flood events within a period of 7 days, resulting in the exclusion of one or more cases from the NARR dataset (see section 2a), was most frequent. A single large precipitation event amidst an otherwise dry period may result in flash flooding, as evidenced by several FF events with anomalously low soil moisture. However, the results discussed above suggest that middle to late summer flash flooding is more frequent when short-term precipitation compensates for the climatological decline in soil moisture.
This discussion of the probable effects of soil moisture on flash flood occurrence is presented with the caveat that although the NARR and FFG data used in this study represent the best available means of diagnosing the soil moisture, they contain errors that cannot be accurately determined due to a lack of in situ measurements for validation.
The greatest limitation in this work is the uncertainty generated by potential reporting biases. Communication systems, public awareness, and observational tools have improved tremendously since the beginning of the period of study. It is certainly possible that some events classified as nonflood events may have produced flooding that was either not observed or not reported, particularly in earlier years. A future task may involve investigating historical streamflow records to consider whether any events have been misclassified.
Another difficulty of this type of climatological study is that the floods themselves vary in intensity from minor street flooding to significant property damage and loss of life. Embedded within the scale differences of the floods themselves lie differences in the hydrological response from basin to basin due to nonmeteorological factors. Flood thresholds vary due to basin size, topography, land use, vegetation, soil type, and other properties. The significance of soil moisture suggests that the precipitation necessary to generate flash flooding varies temporally as well, dependent upon seasonal and short-term preconditioning of surface conditions. From the perspective of an operational forecaster, these complications are an unavoidable part of the flash flood warning decision process. That the meteorology yields trends that may help a forecaster to anticipate whether a given heavy precipitation event will cause flash flooding is of significance.
5. Conclusions
Statistical analysis has yielded trends that can refine the checklist that was in use for the BGM CWA prior to this study. Several parameters—especially atmospheric moisture and convective indices—proved to be effective at separating the FF and nonflood events from the climatology. In a revised checklist, these parameters would indicate the potential for heavy precipitation. A secondary set of parameters would further suggest the conditions under which flash flooding is more favorable. Most significant among these is soil moisture, which was shown to be the most skillful predictor of flash flood occurrence. The preference for either flooding or nonflooding with certain low-level wind directions reflects differences in the underlying physical processes of mesoscale storm development as described in section 4. Flash flooding was also found to be more likely with small to moderate CAPE, demonstrating a preference for collision–coalescence warm rain processes, and with large areas of near-constant upwind 850-hPa θe, which ensures an ample low-level supply of moisture and reduces dry-air entrainment caused by vertical wind shear. The FF events are associated with stronger upper-level winds than are the WATCH events, and less vertical speed shear than the SP events. These tendencies should help forecasters, particularly those in the northeastern United States and climatically similar regions, to better anticipate whether a given potential heavy precipitation scenario will produce flash flooding.
Acknowledgments
This work was funded through the COMET program of the University Corporation for Atmospheric Research as COMET Outreach Project S05-52254. The authors thank Mike Evans at NWS BGM, Bob Davis as NWS PIT, and two anonymous reviewers for their helpful suggestions on the manuscript. They would also like to thank Ashley Coles, Andrew Shook, and Shaun Walter for their assistance in compiling data for this project.
REFERENCES
Bosart, L. F., and Sanders F. , 1981: The Johnstown flood of July 1977: A long-lived convective system. J. Atmos. Sci., 38 , 1616–1642.
Colman, B. R., 1990a: Thunderstorms above frontal surfaces in environments without positive CAPE. Part I: A climatology. Mon. Wea. Rev., 118 , 1103–1122.
Colman, B. R., 1990b: Thunderstorms above frontal surfaces in environments without positive CAPE. Part II: Organization and instability mechanisms. Mon. Wea. Rev., 118 , 1123–1144.
Corfidi, S. F., Meritt J. H. , and Fritsch J. M. , 1996: Predicting the movement of mesoscale convective complexes. Wea. Forecasting, 11 , 41–46.
Cotton, W. R., 1990: Storms. Vol. 1. Geophysical Science Series, Aster Press, 158 pp.
Davis, R. S., 2001: Flash flood forecast and detection methods. Severe Convective Storms, Meteor. Monogr., No. 50, Amer. Meteor. Soc., 481–525.
DeGaetano, A. T., Hirsch M. E. , and Colucci S. J. , 2002: Statistical prediction of seasonal East Coast winter storm frequency. J. Climate, 15 , 1101–1117.
De Michele, C., and Salvadori G. , 2002: On the derived flood frequency distribution: Analytical formulation and the influence of antecedent soil moisture condition. J. Hydrol., 262 , 245–248.
Doswell C. A. III, , Brooks H. E. , and Maddox R. A. , 1996: Flash flood forecasting: An ingredients-based methodology. Wea. Forecasting, 11 , 560–581.
Droegemeier, K. K., and Coauthors, 2000: Hydrological aspects of weather prediction and flood warnings: Report of the Ninth Prospectus Development Team of the U.S. Weather Research Program. Bull. Amer. Meteor. Soc., 81 , 2665–2680.
Emanuel, K. A., 1982: Inertial instability and mesoscale convective systems. Part II: Symmetric CISK in a baroclinic flow. J. Atmos. Sci., 39 , 1080–1097.
Fan, Y., Van den Dool H. M. , Lohmann D. , and Mitchell K. , 2006: 1948–98 U.S. hydrological reanalysis by the Noah land data assimilation system. J. Climate, 19 , 1214–1237.
Fritsch, J. M., and Carbone R. E. , 2004: Improving quantitative precipitation forecasts in the warm season: A USWRP research and development strategy. Bull. Amer. Meteor. Soc., 85 , 955–965.
Funk, T. W., 1991: Forecasting techniques utilized by the Forecast Branch of the National Meteorological Center during a major convective rainfall event. Wea. Forecasting, 6 , 548–564.
Harnack, R., Apffel K. , Georgescu M. , and Baines S. , 2001: The determination of observed atmospheric differences between significant and light precipitation events in New Jersey, USA. Int. J. Climatol., 21 , 1529–1560.
LaPenta, K. D., and Coauthors, 1995: The challenge of forecasting significant rain and flooding throughout the Eastern Region of the National Weather Service. Part I: Characteristics and events. Wea. Forecasting, 10 , 78–90.
Maddox, R. A., Chappell C. F. , and Hoxit L. R. , 1979: Synoptic and mesoscale aspects of flash flood events. Bull. Amer. Meteor. Soc., 60 , 115–123.
Mesinger, F., and Coauthors, 2006: North American Regional Reanalysis. Bull. Amer. Meteor. Soc., 87 , 343–360.
Mo, K. C., and Chelliah M. , 2006: The modified Palmer drought severity index based on the NCEP North American Regional Reanalysis. J. Appl. Meteor. Climatol., 45 , 1362–1375.
Mo, K. C., Chelliah M. , Carrera M. L. , Higgins R. W. , and Ebisuzaki W. , 2005: Atmospheric moisture transport over the United States and Mexico as evaluated in the NCEP regional reanalysis. J. Hydrometeor., 6 , 710–728.
NCDC, cited. 2004: NCDC NEXRAD data inventory search. [Available online at http://www.ncdc.noaa.gov/nexradinv.].
NCDC, cited. 2005: Precipitation data, hourly—U.S. and some non-U.S. (DS3240). [Available online at http://cdo.ncdc.noaa.gov/CDO/cdo.].
NOAA, 1986–2003: Storm Data. Vols. 28–45.
Ntelekos, A. A., Geogakakos K. P. , and Krajewski W. F. , 2006: On the uncertainties of flash flood guidance: Toward probabilistic forecasting of flash floods. J. Hydrometeor., 7 , 896–915.
NWS, 2006: Definitions and general terminology. National Weather Service Manual 10-950, 4 pp. [Available online at http://www.nws.noaa.gov/directives/sym/pd01009050curr.pdf.].
Ogden, F. L., Sharif H. O. , Senarath S. U. S. , Smith J. A. , Baeck M. L. , and Richardson J. R. , 2000: Hydrologic analysis of the Fort Collins, Colorado, flash flood of 1997. J. Hydrol., 228 , 82–100.
Opitz, H. H., and Coauthors, 1995: The challenge of forecasting significant rain and flooding throughout the Eastern Region of the National Weather Service. Part II: Forecast techniques and applications. Wea. Forecasting, 10 , 91–104.
Ropelewski, C. F., and Yarosh E. S. , 1998: The observed mean annual cycle of moisture budgets over the central United States (1973–92). J. Climate, 11 , 2180–2190.
Sanders, F., 1971: Analytic solutions of the nonlinear omega and vorticity equations for a structurally simple model of disturbances in the baroclinic westerlies. Mon. Wea. Rev., 99 , 393–407.
Sivapalan, M., Bloschl G. , Merz R. , and Gutknecht D. , 2005: Linking flood frequency to long-term water balance: Incorporating effects of seasonality. Water Resour. Res., 41 .W06012, doi:10.1029/2004WR003439.
Smith, J. A., Baeck M. L. , Morrison J. E. , and Sturdevant-Rees P. , 2000: Catastrophic rainfall and flooding in Texas. J. Hydrometeor., 1 , 5–25.
Smith, J. A., Baeck M. L. , Morrison J. E. , Sturdevant-Rees P. , Turner-Gillespie D. F. , and Bates P. D. , 2002: The regional hydrology of extreme floods in an urbanizing drainage basin. J. Hydrometeor., 3 , 267–282.
Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences. Academic Press, 465 pp.
Zhang, D., and Fritsch J. M. , 1986: Numerical simulation of the meso-β scale structure and evolution of the 1977 Johnstown flood. Part I: Model description and verification. J. Atmos. Sci., 43 , 1913–1944.
Zhang, D., and Fritsch J. M. , 1987: numerical simulation of the meso-β scale structure and evolution of the 1977 Johnstown flood. Part II: Inertially stable warm-core vortex and the mesoscale convective complex. J. Atmos. Sci., 44 , 2593–2612.
Zhang, Y., and Smith J. A. , 2003: Space–time variability of rainfall and extreme flood response in the Menomonee River Basin, Wisconsin. J. Hydrometeor., 4 , 506–517.
Spatial distribution of flash flooding (by county) for the BGM CWA, 1986–2003. Smaller values in parenthesis indicate flash floods per county normalized by area.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Annual distribution of flash floods within the BGM CWA, 1986–2003.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Rain gauge–estimated storm total precipitation as a function of time of year (cm; F = flood; S = significant; W = watch). Averages are 6.53, 4.46, and 4.08 cm, respectively.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Radar-estimated storm total precipitation (cm) and rain gauge–estimated storm total precipitation (cm) for flood, significant, and watch events where both datasets are available.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
The 850-hPa wind speed (m s−1) and 850-hPa wind direction (°). Symbols are the same as in Fig. 3, with the addition of dots to represent the RAND events. Averages are F: 8.8 m s−1 and 198.4°, S: 9.4 m s−1 and 226.8°, W: 10.0 m s−1 and 225.1°, and RAND: 8.15 m s−1 and 249.19°.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Soil moisture in the uppermost 10 cm (fraction) as a function of time of year (symbols same as Fig. 5). Averages are F: 0.36, S: 0.32, W: 0.35, and RAND: 0.31.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Three-hour flash flood guidance (cm) and 0–10-cm depth soil moisture (fraction; symbols same as Fig. 3). Averages for FFG are F: 7.5 cm, S: 12.3 cm, and W: 7.5 cm.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Composite maps of 850-hPa θe (shaded) and 850-hPa geopotential height (contoured) for flash flood cases with R2 values in the range (a) 0–0.2 (17 members), (b) 0.4–0.6 (11 members), and (c) 0.8–1.0 (8 members).
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Composite maps of 850-hPa θe (shaded, K) and 850-hPa geopotential height (contoured, m) for nonflooding significant precipitation cases with R2 values in the range (a) 0–0.2 (3 members), (b) 0.4–0.6 (9 members), and (c) 0.8–1.0 (12 members).
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Mean OKX sounding, hodograph, and selected convective parameters for FF events.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Mean PIT sounding, hodograph, and selected convective parameters for FF events.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Mean PIT sounding, hodograph, and selected convective parameters for SP events.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Mean PIT sounding, hodograph, and selected convective parameters for WATCH events.
Citation: Weather and Forecasting 23, 1; 10.1175/2007WAF2006066.1
Quantitative parameters on the BGM checklist prior to this study.
Discriminant analysis results. Columns indicate which dataset is paired with the RAND events. Numerical values given, in order, are Kuipers skill score, cross-validated Kuipers skill score, and cross-validated percent correct.
Quartile and median values for selected parameters: FF (F), SP (S), WATCH (W), and RAND (R) events.
Discriminant analysis results. Columns indicate which dataset is paired with the FF events. Numerical values given, in order, are Kuipers skill score, cross-validated Kuipers skill score, and cross-validated percent correct.
