Water Resources Research

Hydroclimatology of flash flooding in Atlanta

Authors


Abstract

[1] The objective of this study is to characterize the climatology of extreme rainfall and flash flooding in Atlanta, Georgia using high-resolution land surface, rainfall, and discharge datasets. We examine nine urban watersheds in the Atlanta area that range in size from 3.7 to 225 km2 and exhibit a range of urban development and land-use characteristics. We develop a high-resolution 15 min, 1 km2 radar rainfall data set for the 2002–2010 period using the Hydro-NEXRAD system with volume scan reflectivity observations from the Atlanta WSR-88D radar and rainfall observations from a dense network of 72 U.S. Geological Survey rain gauges. Bias-corrected radar rainfall fields accurately capture the spatial and temporal structure of heavy rainfall. There is enhancement of heavy rainfall within and east of the urban core, and a rainfall minimum north and northwest of the city. There has been an increase in variability of annual flood peaks in Atlanta since the 1960s associated with urban impacts on runoff production. Flood response is dependent on a combination of basin size, drainage network structure, spatial distribution of land use, and basin storage in urban soils and storm water detention ponds. Future studies of urban rainfall modification in Atlanta and elsewhere should consider the influence of regional topography and other geographic features on the storm environment.

1. Introduction

[2] Economic damages and fatalities associated with urban flash flooding in the southeastern United States, including record flooding in Atlanta, Georgia in September 2004 and again in September 2009, point to the need for better flood risk assessment capabilities. The Atlanta metropolitan area has been one of the most rapidly urbanizing population centers in the United States since the mid-20th century, with population increasing from less than one million in 1950 to over five million in 2010 [United States Census Bureau, 2010]. This growth has increased regional vulnerability to thunderstorm-related hazards such as flash flooding, tornados and hail [Paulikas and Ashley, 2011]. Changes in urbanization have important implications for the urban hydrologic cycle and particularly for heavy rainfall [e.g., Diem and Mote, 2005; Niyogi et al., 2006; Ntelekos et al., 2008] and flood-producing runoff processes [e.g., Smith et al., 2002; Beighley and Moglen, 2003; Smith et al., 2005a; Villarini et al., 2009]. A densely instrumented observational network in Atlanta provides an excellent setting to examine the impacts of urbanization on these processes.

[3] The objective of this study is to characterize the climatology of extreme rainfall and flash flooding in the Atlanta metropolitan area, considering the dynamics of heavy rainfall and runoff and the temporal and spatial scales at which they interact in urban environments. Improved understanding of these processes should assist in the development of flood forecasting methods, storm water management policies, and land-use practices. Specific questions addressed in this study are:

[4] 1. What is the climatology of urban flooding in the Atlanta metropolitan area and how is it linked to the climatology of heavy rainfall?

[5] 2. How does urbanization affect storm event hydrologic response in urbanized watersheds?

[6] The rainfall and flood climatology across the southeastern United States is the result of a mixture of tropical cyclones, winter and springtime extratropical systems, and warm-season thunderstorm systems [Villarini and Smith, 2010]. There is a regional minimum in tropical cyclone-related annual flood peaks [Villarini and Smith, 2010] and heavy rainfall [Hart and Evans, 2001; Kunkel et al., 2011; Barlow, 2011] across much of Georgia attributable to the properties of extratropical transition of tropical cyclones. The topographic gradient from the Appalachian and Blue Ridge Mountains eastward to the Piedmont, coastal plain, and Atlantic Ocean plays an important role in the initiation and development of thunderstorms [e.g., Weisman, 1990a, 1990b; Murphy and Konrad, 2005; Ntelekos et al., 2007, 2009]. The largest flood peaks in small catchments across the eastern United States are often associated with organized warm season thunderstorm systems [Miller, 1990].

[7] Urbanization can influence heavy rainfall climatology in several ways [Shepherd, 2005]. As land becomes urbanized, an urban heat island (UHI) may develop due to the additional heat capacity of the built environment. Changing thermal properties due to the UHI can influence the initiation or subsequent behavior of convective rainfall. Additional aspects of urbanization that can impact rainfall mechanisms include increased surface roughness and elevated atmospheric aerosol concentrations [Rosenfeld et al., 2008; Ntelekos et al., 2009]. Urban rainfall climatology can also depend on regional features such as topography and land-water boundaries [Ntelekos et al., 2008].

[8] Numerous studies suggest that the UHI associated with the Atlanta metropolitan area has altered the spatial pattern of convective precipitation [Bornstein and Lin, 2000; Shepherd et al., 2002; Dixon and Mote, 2003; Diem and Mote, 2005; Mote et al., 2007; Rose et al., 2008; Diem, 2008; Shem and Shepherd, 2009; Bentley et al., 2010; Zhou and Shepherd, 2010; Bentley et al., 2011]. The interplay of regional features with urban effects has not been directly addressed in past studies of urban rainfall in the Atlanta area. This study highlights the localized impacts that regional topography may have on heavy rainfall in the Atlanta metropolitan area.

[9] Urbanization also affects flood-producing runoff processes. Impervious surfaces, channel straightening, and elaboration of the drainage network through storm drain systems tend to decrease the hydrologic response time of urban basins and reduce infiltration, leading to larger flood peaks [e.g., Leopold, 1968; Graf, 1977; Bedient et al., 2000; Beighley and Moglen, 2002; Smith et al., 2002, 2005a, 2005b]. Soils in urban areas tend to infiltrate water at lower rates and have reduced storage volumes relative to natural soils, mainly due to compaction [Gregory et al., 2006]. The Clean Water Act (CWA) brought about a series of storm water management regulations that focus on infrastructure-based solutions to flood mitigation in the United States. Among these solutions, detention ponds are designed to decrease the severity of flooding at a site scale. Their effectiveness at mitigating flood peaks at a watershed scale, however, is unclear [e.g., Emerson et al., 2005; Goff and Gentry, 2006; Fang et al., 2010; Meierdiercks et al., 2010].

[10] Urbanization has impacted the hydrologic cycle in the Atlanta area, including lowered groundwater levels [Rose and Peters, 2001], reduced low stream flows, and increased high flows [Ferguson and Suckling, 1990]. Viger et al. [2011] projected annual surface water flows and groundwater flows to be reduced in the 21st century as a combined result of urbanization and climate effects. These studies focused on annual or seasonal trends over long periods and not on storm event rainfall and flood response. Analyzing rainfall-runoff relationships for small urban basins with rapid response times such as the Atlanta area watersheds presented here requires finer resolution data than is typically used for hydrologic analysis [Schilling, 1991; Berne et al., 2004b; Creutin et al., 2009; Emmanuel et al., 2012]. This is the first study to systematically examine high-resolution land surface, rainfall, and discharge data to describe the flash flood climatology of the Atlanta area.

2. Study Region

[11] Atlanta is located in the Piedmont physiographic province with the Blue Ridge Mountains to the north, the Valley and Ridge Appalachians to the west, and the Atlantic Coastal Plain to the south and the east (Figure 1). Land surface elevations in the Atlanta metropolitan area range from 250 to 300 m above sea level (masl) with minimal local relief, while the Blue Ridge Mountains to the north reach nearly 1000 masl and the Appalachians directly west reach approximately 400 masl.

Figure 1.

Study region. (top left) Northern Georgia; the Atlanta urban boundary and study watershed boundaries are shown in relation to regional topographic features, USGS and NCDC rain gauges and NWS radar observation sites. The red circle indicates the 200 km radius from the KFFC WSR-88D radar. (top right) Study area showing percent impervious land cover (year 2005), USGS stream gauging stations, and boundaries of the study watersheds. (bottom left) Study area showing year 1974 land use classifications, USGS stream gauging stations, and boundaries of the study watersheds. (bottom right) Study area showing year 2005 land use classifications, USGS stream gauging stations, and boundaries of the study watersheds. Figures 1 (top right) and 1 (bottom) share a common map scale.

[12] This study focuses on nine urban watersheds located within the Atlanta metropolitan area. Table 1 shows watershed characteristics including basin area, average slope, percentage and spatial distribution of impervious cover, and land use percentages. The watersheds range in size from 3.7 to 225 km2, with average slopes ranging from 6% to 11%, and impervious areas covering between 17% and 52% of basin area. The principal watershed in this and previous studies [Ferguson and Suckling, 1990; Rose and Peters, 2001] is the Peachtree Creek (225 km2), which drains a portion of downtown Atlanta and the northeastern part of the metropolitan area. Additional watersheds included in this study are Nancy Creek, Sope Creek, Proctor Creek, and Utoy Creek. Proctor Creek, along with the lower reaches of Peachtree Creek, drains part of downtown Atlanta and is highly urbanized. Nancy Creek consists of two drainage basins (henceforth referred to as Upper Nancy Creek and the Outlet of Nancy Creek) which are adjacent to and relatively comparable in terms of size and land use to the North Fork and South Fork of Peachtree Creek, but exhibit very different spatial distributions of urbanization (Figure 1, bottom). The upper and lower portions of Utoy Creek are heavily urbanized and are separated by an area of less intense development. Sope Creek has the highest percentage of low intensity residential development, much of it built after the enactment of the CWA, and therefore a higher percentage of its drainage area should be influenced by detention ponds.

Table 1. Physical Characteristics for Nine Atlanta-Area Urban Watershedsa
 Upper NancyOutlet NancyUpper PeachtreeN. Fork PeachtreeS. Fork PeachtreeOutlet PeachtreeSopeProctorUtoy
  • a

    See Figure 1 for watershed locations.

Area [km2]69983.79074225763588
Slope [%]9.09.96.28.57.78.49.09.610.8
Imperv. Area (2005) [%]272352302531193317
Imperv. Area (1991–2005) [% change]565634333432723155
Fractional Flow Distance of 50% Imperv. Area [-]0.610.730.580.660.490.520.680.710.58
 
Land Use (2005) [% of total]
Developed, Open Space333316313531382329
Developed, Low Intensity242124242524253222
Developed, Med. Intensity1412181511157177
Developed, High Intensity9735128124114
Forest18247162016221335
Other331222334

3. Data and Methods

[13] This study makes use of a variety of high-resolution datasets for land use, rainfall, and discharge.

3.1. Land Use and Land Cover Data

[14] Land use is derived from the 2005 National Land Cover Data set (NLCD) and from the University of Georgia Natural Resources Spatial Analysis Lab (NARSAL) which has created a set of land cover maps for Georgia spanning 1974 to 2005 and impervious cover maps from 1991 to 2005, allowing the assessment of spatial and temporal changes in land use.

[15] The influence of land use change on hydrologic response must be considered within the context of storm water management policies and infrastructure as well as the spatial distribution of urban development within individual watersheds [Mejia and Moglen, 2010a, 2010b]. One useful measure of the spatial distribution of urban development is the impervious area curve [Meierdiercks et al., 2010]. The impervious area curve shows the percentage of the total impervious area within the watershed at a given flow distance from the outlet (see Figure 2). A curve with a steep slope at a given flow distance indicates a high degree of impervious development at that flow distance from the outlet. The impervious area curve provides a context for interpreting catchment hydrologic response since runoff from impervious areas generally reaches the channel network more quickly than runoff from pervious areas. Hydrologic response at the outlet of a basin will tend to be faster if impervious areas are located near to the outlet than if they are located farther upstream.

Figure 2.

Distribution of impervious area as a fraction of distance from (left) the basin outlet for the North Fork and South Fork of Peachtree Creek and (right) the outlet of Nancy Creek and the outlet of Peachtree Creek.

3.2. Rainfall Data

[16] The U.S. Geological Survey (USGS) maintains a network of 72 rain gauges around the metropolitan area (Figure 1). These gauges operate on a 15 min time resolution, which is aggregated to daily accumulations for archiving. The first gauges became operational in 1996 and most were operational by 2002. There are 11 National Weather Service (NWS) Cooperative Station Network rain gauges in the Atlanta metropolitan area that provide daily precipitation accumulations for the 1957–2010 period.

[17] In this study we develop a 9 year (2002–2010) 15 min, 1 km2 radar-estimated rainfall data set using the Hydro-NEXRAD system [Krajewski et al., 2010; Smith et al., 2012] for the February–October time period. The Hydro-NEXRAD processing system converts three-dimensional polar-coordinate volume scan reflectivity fields into two-dimensional Cartesian surface rainfall fields. The standard convective rainfall-reflectivity (Z-R) relationship ( inline image, where inline image, inline image, R is rain rate in inline image, and Z is the radar reflectivity factor in inline image), and several standard quality control algorithms were used [Seo et al., 2011]. The November through January period was not included due to complications arising from radar estimation of cold season precipitation [Borga, 2002; Rinehart, 2006; Hazenberg et al., 2011]. The location of the KFFC Weather Surveillance Radar, 1988, Doppler (WSR-88D) radar and of the 72 USGS rain gauges used for bias correction are shown in Figure 1 (top left).

[18] We compute a daily mean-field multiplicative bias correction based on daily rain gauge accumulations computed from the network of 72 USGS rain gauges in and around the Atlanta metropolitan area. The daily bias correction takes the form:

display math

where Gij is the daily rainfall accumulation for gauge j on day i, Rij is the daily rainfall accumulation for the radar pixel containing gauge j on day i and Si is the index of the rain gauge stations from which both rain gauge and radar have positive rainfall accumulation for day i. Each 15 min radar rainfall field from day i is then multiplied by the bias correction factor Bi. We apply a bias value different from unity only if there are at least five radar-gauge pairs with positive precipitation accumulations for day i. The bias correction procedure is the same as that used in the work of Smith et al. [2012].

[19] It should be noted that range-dependent bias was not explicitly addressed for most of the following analyses. Analyses of rainfall fields over the February–October period over the area encompassing the study watersheds show that range-dependent bias is small (not shown). Applications of radar rainfall estimates over larger areas or areas closer to or farther from the radar should consider the effects of range-dependent bias. For example, for the analyses of presented in section 4.3, which consider a large portion of the radar domain, the effects of long-term range-dependent bias were important and were removed by dividing the radar domain into concentric range-dependent 1 km rings, computing the mean quantity for each ring and then multiplying each cell in the radar domain by the ratio of its range ring mean quantity to the radar domain mean quantity. One additional challenge of using long-term radar rainfall fields for climatological analyses is the impact of missing records, which were particularly prevalent in the early part of the analysis period. The effect of missing records on seasonal averages was addressed by computing a seasonal mean cell-by-cell from available records and then multiplying it by the ratio of the total number of 15 min periods in the season to the number of available 15 min records. Better results could likely be obtained by integrating NEXRAD Stage IV operational products to fill time periods for which volume scan reflectivity data are unavailable. Alternate techniques for correcting range-dependent bias based on Vertical Profile of Reflectivity (VPR) corrections include Andrieu and Creutin [1995], Vignal and Krajewski [2001], Seo et al. [2000], Berne et al. [2004a], Bellon et al. [2007], Delrieu et al. [2009], Krajewski et al. [2011], and Germann et al. [2006].

3.3. Streamflow Data

[20] Urban watersheds in Atlanta were the subject of an extensive streamflow measurement campaign as part of the National Urban Runoff Program (NURP), directed by the Environmental Protection Agency and affiliated state and local groups in 1960s and 1970s. The purpose of the campaign was to establish flood frequency relationships for urban watersheds in the Atlanta area [Inman, 1983, 1988]. The program developed 2–10 year records of annual peak discharge for a large number of nested urbanized watersheds ranging from 2 to 340 km2. In addition, the USGS has maintained stream gauges at the outlets of Peachtree Creek since 1957 and Sope Creek since 1962. More recently, the stream gauging network has been expanded to all the watersheds described in section 2, and available data include instantaneous discharge measurements at 5–15 min resolution in addition to annual flood peaks. The instantaneous discharge records are available for several stations beginning in 1986, while the majority of the stations have instantaneous discharge records beginning after 2000. In this study, all instantaneous discharge values are linearly interpolated to 1 min and missing data periods of less than 2 h are filled via linear interpolation.

4. Results and Discussion

4.1. Land Use and Land Cover

[21] Land use derived from the 2005 NLCD ranges from moderately developed in Utoy Creek (88 km2), where 60% of the watershed is composed of urban land types, to highly developed in Upper Peachtree Creek (3.7 km2) with urban land types covering approximately 90% of the watershed (Table 1).

[22] Analyses of NARSAL land use datasets reveal a marked change in land use across the study watersheds between 1974 and 2005, with forest and agricultural land being replaced by urban land types (Figure 1, bottom). The increase in urbanization has been more dramatic in watersheds that were less developed prior to 1974. For example, Sope Creek saw a 106% increase in urban land cover from 1974 to 2005, from 36% to 74% of basin area, whereas Upper Peachtree Creek saw a 30% increase, from 71% to 92% of basin area over the same period.

[23] The USGS maintains four stream gauging stations along Peachtree Creek, defining four gauged subwatersheds ranging from 3.7 to 225 km2. As seen in Figure 1, all four subcatchments have undergone varying degrees of urbanization since 1974, and exhibit very different spatial distributions of urban development. Upper Peachtree Creek (3.7 km2) is the smallest of the nine study watersheds, was the most urbanized basin prior to the enactment of the CWA, has the highest percentage of impervious area (52%), and has the highest percentage of high-intensity development (35%) of the study watersheds. The North Fork and South Fork of Peachtree Creek are similar in terms of area (90 km2 and 74 km2, respectively), and have similar land use characteristics (30% and 25% impervious, respectively), but differ in the spatial distribution of impervious area as shown by their respective impervious area curves (Figure 2, left). The North Fork has a greater percentage of its impervious area situated in the upper reaches of the watershed than the South Fork. Downtown Atlanta is located downstream of the North Fork and South Fork stream gauges, and so the hydrologic response from urban core is reflected only in the farthest downstream gauge, henceforth referred to as the Outlet of Peachtree Creek, which encompasses a total drainage area of 225 km2 and is 31% impervious.

[24] Impervious area curves for the entire Peachtree Creek and Nancy Creek watersheds show that urban development is concentrated closer to the basin outlet in Peachtree Creek than in Nancy Creek (Figure 2, right). A convenient way to tabulate information contained in the impervious area curve is to specify the fraction of total flow distance from the outlet at which 50% of the total impervious area is situated. For example, a value of 0.60 would mean that 50% of the total impervious area is located closer than 0.60 times the longest flowpath in the watershed. This metric ranges from 0.49 for the South Fork of Peachtree Creek to 0.73 for the outlet of Nancy Creek (Table 1), suggesting that in the South Fork of Peachtree Creek, more of the urban development has occurred closer to the basin outlet than in Nancy Creek.

4.2. Rainfall Estimation

[25] The daily bias correction removes systematic bias due to variability in Z-R relationships and radar calibration errors [Villarini and Krajewski, 2010]. In this study, bias correction improves the daily coefficient of determination (R2) from 0.74 to 0.88, decreases the daily root-mean-square error (RMSE) from 9.34 to 5.74 mm, and decreases the daily mean absolute error (MAE) from 2.64 to 0.42 mm, and eliminates the systematic underestimation of rainfall by the radar (Figure 3, top). This data set will be made available to other researchers by request.

Figure 3.

(top) Comparison of colocated rain gauge and radar-estimated daily rainfall accumulations prior to (Figure 3, top left) and after mean-field bias correction (Figure 3, top right) over 2002–2010 February–October period. (bottom) Comparison of colocated rain gauge and bias-corrected radar-estimated daily rainfall accumulations for 3 May 2010 storm (Figure 3, bottom left) and 19–22 September 2009 storm (Figure 3, bottom right).

[26] Of particular interest for this study is the ability of radar to capture the spatial and temporal dynamics of heavy rainfall (see Baeck and Smith [1998], Ntelekos et al. [2008], and Villarini et al. [2010]). Agreement is good between the daily bias-corrected radar and rain gauge accumulations for the 3 May 2010 and 19–22 September 2009 storms (Figure 3, bottom left and bottom right, respectively). Although USGS rain gauge observations are typically retained only as daily accumulations, time series of 15 min accumulations were available for several events. The 15 min basin averaged radar-estimated rainfall rates for two Peachtree Creek subwatersheds compared with 15 min time series of rain gauge observations for the 3 May 2010 and 19–22 September 2009 storms show good agreement between bias-corrected radar and rain gauges (Figure 4). The radar captures the temporal variability represented in the rain gauge observations, while providing a more complete spatial description of rainfall than is possible using only rain gauges.

Figure 4.

Comparison of rain gauge and bias-corrected radar time series for (top) Upper Peachtree Creek for 19–23 September 2009 and (bottom) the Northfork of Peachtree Creek for 3 May 2010.

4.3. Hydroclimatology

[27] Bias-corrected Hydro-NEXRAD radar rainfall fields offer improvements over previous datasets for conducting analyses of rainfall climatologies [Smith et al., 2012]. Previous studies in the Atlanta area have used operational radar rainfall fields [e.g., Mote et al., 2007], radar reflectivity fields [e.g., Mote et al., 2007], or rain gauge networks [e.g., Diem and Mote, 2005; Diem, 2008]. Operational radar rainfall fields are developed for different applications than those pursued in this study.

[28] There is a local maximum in both mean rainfall and mean number of days with rainfall accumulations greater than 25 mm for the 2002–2010 February–October time period downwind (east) of the city (Figure 5), consistent with the conclusions based on rainfall estimates from rain gauges [Diem and Mote, 2005; Diem, 2008], operational radar rainfall products [Mote et al., 2007], and from spaceborne sensors (Shepherd et al. [2002]; see Diem et al. [2004] for a discussion of the limitations of satellite precipitation estimates in urban rainfall studies).

Figure 5.

(left) Average February–October rainfall in mm and (right) number of days with greater than 25 mm of rain from 2002–2010. Black outline shows the Peachtree Creek watershed; gray outline shows the border of the Atlanta metropolitan area. The radar data are bias corrected and range adjusted.

[29] Mean wind direction during the summer months [Shepherd et al., 2002] and during periods of thunderstorm activity [Bentley et al., 2010] is westerly to southwesterly, so elevated rainfall east of the urban core is consistent with the description of rainfall enhancement downwind of urban areas presented in the work of Changnon et al. [1971]. There is a rainfall minimum to the northwest and southwest of Atlanta (Figure 5), probably associated with the complex nature of regional eastward flow and thunderstorm initiation over the ridge-valley system of the Appalachians and Blue Ridge Mountains as mentioned in the work of Murphy and Konrad [2005] and Weisman [1990a, 1990b]. It is possible that topographically influenced dynamics of thunderstorm initiation and motion also contribute to the observed patterns of elevated rainfall in and around Atlanta. Previous studies of urban enhancement of precipitation in Atlanta have not considered the possible impacts of regional topography on precipitation within and near to the urban core (see, for example, Ntelekos et al. [2007]).

[30] Annual peak discharge for Sope Creek and the outlet of Peachtree Creek show an increase in variability since 1960 (Figure 6, bottom). Linear trends were computed using the nonparametric Thiel-Sen estimator [Sen, 1968]. Rain gauge data were examined for the same time period to determine whether the increase in annual flood peak variability could be attributed to increased variability of extreme rainfall at daily and subdaily scales. Since the NWS raingages are not necessarily colocated with stream gauging stations, the date of occurrence of the largest rainfall accumulations oftentimes did not correspond with the date of the annual peak floods. Instead, to examine changes in extreme rainfall climatology over the 1957–2009 period, daily rainfall accumulations from the 11 NWS rain gauges were pooled and the single maximum daily accumulation retained for each year. No change in variability is evident (Figure 6, top). The unpooled annual maximum daily accumulations for each of the 11 NWS gauges were also examined, leading to the same conclusions. Pooled and unpooled annual maximum 2 day rainfall accumulations (not shown) are very similar to the pooled annual maximum 1 day accumulations. Only one NWS rain gauge, at Hartsfield International Airport, reported hourly values for the entire 1957–2009 period. The annual maximum rain rates for durations of 1 h, 3 h (Figure 6, middle), 6 h, 12 h, and 24 h (not shown) were also calculated to examine the possibility of changing variability of extreme rainfall at the subdaily time scales which are important in urban hydrology.

Figure 6.

(top) Maximum annual 1 day rainfall accumulation selected from pooling records for 11 Atlanta-area NWS Cooperative rain gauges with long-term records. (middle) Maximum annual 1 h and 3 h rainfall rate for Hartsfield International Airport NWS Cooperative rain gauge. (bottom) Maximum annual flood peaks normalized by drainage area at the Sope Creek and Peachtree Creek stream gauge stations. Linear trends are calculated using Theil-Sen Estimator.

[31] Based on the Mann–Kendall test [Mann, 1945; Kendall, 1975], both pooled and unpooled annual maximum daily rainfall data do not exhibit significant trends at the 5% level. The same is true for annual peak rain rate for the Hartsfield International Airport gauge at several subdaily time scales. On the other hand, statistically significant trends in annual peak discharge were detected at the 5% level for both the Sope Creek and Peachtree Creek stream gauging stations.

[32] The increased variability in annual peak discharge is greater in Sope Creek than in Peachtree Creek. Land use changes show that low-intensity and high-intensity urban development increased by 106% in Sope Creek between 1974 and 2005, compared to 47% in Peachtree Creek. Likewise, between 1991 and 2005, the impervious area in Sope Creek increased by 72%, compared to 32% in Peachtree Creek. Therefore, it is likely that the increased variability in annual flood peaks can be attributed to the effects of increased urbanization on flood response and not to an increase in extreme rainfall volumes. This is consistent with the results in the work of Villarini et al. [2009], who showed a large impact of urbanization on the magnitude of flooding in the Charlotte, North Carolina metropolitan area with no significant change in the magnitude of extreme rainfall (see also Villarini et al. [2012] for similar results for Chicago). Because significant suburban development has taken place in Sope Creek since the enactment of the CWA, the watershed contains a large number of detention ponds. While no data are available on the number, size, or spatial distribution of detention ponds in Sope Creek, the increase in variability of annual flood peaks over that period suggests that detention ponds have not been effective at mitigating the effects of land use change on annual flood peaks. The impact of detention ponds on flood response as a function of event magnitude is discussed further in section 4.4.

4.4. Flood Hydrology

[33] The median annual flood peak was computed for all the NURP stream gauges with four or more years of annual peak discharge data. Although 4 years of peak annual discharge is inadequate to study upper tail flood behavior, it is sufficient to establish average flood behavior across spatial scales. Median annual flood peaks scale linearly with basin area on a log-log scale (Figure 7). Mean annual flood peaks show similar results (not shown). The populations of subwatersheds of Nancy Creek, Peachtree Creek, and Proctor Creek all scale consistently, while the subwatersheds in Utoy Creek are separated into two distinct populations, one with larger median annual peaks than the other for a given basin size. The subwatersheds with higher median annual peaks are located on the south branch of Utoy Creek, which is characterized by shorter flow paths and greater urbanization than the northern branch. Peachtree Creek also consists of a north and south branch, with similar flowpath lengths and similar magnitudes but differing spatial distributions of urbanization, but does not divide into two distinct populations. This suggests that urban flood response is highly dependent on watershed area and flowpath length as well as on the extent of urban development.

Figure 7.

Mean annual flood peaks as a function of basin area for four Atlanta area basins. Linear trends for Nancy Creek, Peachtree Creek, and Proctor Creek are calculated using ordinary least squares.

[34] Rainfall-runoff response is examined by developing a sample of flood events at each station based on a peak discharge threshold. The threshold was selected to extract an average of five flood events per year from USGS instantaneous discharge records for seven study watersheds (Utoy Creek and Proctor Creek were omitted due to short instantaneous discharge records). Depending on the watershed, these records span either from 2002, 2003, or 2004 to 2010, leading to 35–45 flood peaks per watershed. Once a peak was selected, no other peak occurring in a 48 h period centered on the time of the selected peak could be included, ensuring that events with multiple peaks were not double counted. Basin-averaged rainfall time series were generated using the bias-corrected Hydro-NEXRAD rainfall fields to compare rainfall-runoff response between watersheds. No significant seasonality was observed in flood peak counts (not shown). Therefore, stratification of flooding by season was deemed unnecessary.

[35] The median peak discharge, response time, and volume-to-peak ratio for the seven basins, as well as the coefficient of variation (CV) for each quantity were computed (Table 2). The response time is defined as the interval between the time of the centroid of the maximum rainfall volume for a 12 h period containing the flood peak and the time of occurrence of the flood peak. The volume-to-peak ratio is a measure of hydrograph “peakiness” and was computed as the maximum runoff volume in a 12 h period containing the time of occurrence of the peak divided by the peak discharge value. The resulting ratio is in units of time and converted to hours. Using a 12 h runoff volume to compute the volume-to-peak ratio, a ratio approaching 12 h indicates relatively low peaks and slow hydrograph recession, while smaller ratios indicate higher peaks and faster recession. For more details on the use of flood-to-peak ratio as a measure of hydrologic response, see Bradley and Potter [1992]. Mean values of these rainfall-runoff metrics were also computed and were similar to the median values. The response time values computed from 6 h rainfall volumes were similar to the response time values computed from 12 h rainfall volumes.

Table 2. Rainfall-Runoff Relationships for Peak-Over-Threshold Flood Eventsa
WatershedArea [km2]Median Peak Discharge [m3 s−1 km−2]Median Response Time [h]Median Volume-to-Peak Ratio [h]
  • a

    Quantities in parenthesis are the unitless coefficient of variation.

Sope760.66 (0.72)3.92 (0.55)5.25 (0.70)
Upper Peachtree3.74.52 (0.68)0.74 (0.63)1.77 (2.10)
North Fork Peachtree900.62 (0.48)3.63 (0.71)6.72 (0.39)
South Fork Peachtree740.98 (0.30)2.36 (1.20)5.91 (0.43)
Outlet Peachtree2250.53 (0.38)3.84 (0.74)7.13 (0.31)
Upper Nancy690.67 (0.39)3.72 (0.70)5.59 (0.45)
Outlet Nancy980.32 (0.70)6.13 (0.45)8.07 (0.38)

[36] Upper Peachtree Creek has the highest median peak discharge per unit area, shortest median response time, and “peakiest” flood response (lowest median volume-to-peak ratio). The high CV for the volume-to-peak ratio for Upper Peachtree Creek is the result of its small size and rapid response, since the 12 h runoff volume is often the result of several distinct flood peaks rather than a single peak as in larger basins. The South Fork of Peachtree Creek has the highest variability in response time, likely reflecting a difference in flood response of the largely impervious lower portion of the watershed and the more lightly developed upper portion.

[37] The median peak discharge and response time for Sope Creek are consistent with other similarly sized basins (the North Fork and South Fork of Peachtree Creek and Upper Nancy Creek) but the peak discharge and volume-to-peak ratio display greater variability. As previously noted, post-CWA development in Sope Creek has been more extensive that in other watersheds and thus there is a higher density of detention ponds. It is likely that detention ponds are controlling flood peaks and runoff volumes for small events but have less influence on large events, thus leading to greater variability in flood response. The outlet of Nancy Creek has the smallest median flood peaks per unit area, longest median response times, and largest median volume-to-peak ratio of any watershed. This is likely due to the fact that much of the development in the watershed is situated far from the outlet. The high variability in peak discharge may be due to the interaction between basin structure and the spatiotemporal structure of individual rainfall events. More precise statements on the impact of spatiotemporal rainfall structure on flood response is difficult based solely on empirical studies.

[38] The maximum rain rate for each flood event was computed for time periods ranging from 0.25 to 12 h, ensuring that each period contained the time of peak discharge. Then the Spearman correlation coefficient was computed relating peak discharge to maximum rain rate for each period and for each basin (Figure 8). Except for Upper Nancy Creek, correlation between maximum rain rate and peak discharge exhibits either a local or global maxima for rainfall periods close to the median response time shown in Table 2. Correlation between maximum rainfall rate and peak discharge in Upper Peachtree Creek is the highest for rainfall periods less than about 1.5 h, and decreases significantly for longer rainfall periods, since the response time is very short and the flood wave passes by the stream gauge very quickly. Both the outlet and the South Fork of Peachtree Creek have high correlations between about 1.25 to 3 h reflecting the rapid runoff response of downtown Atlanta, and then again for time periods longer than 6 h when runoff from the upper reaches of the basins is contributing to discharge measurements.

Figure 8.

Spearman correlation coefficient of maximum rain rate calculated for periods ranging from 0.25 to 12 h to peak discharge for seven watersheds.

[39] The temporal correlation structure for Upper Nancy Creek is surprising, with correlation increasing to at least 12 h. The rate of increase, however, diminishes after about 4 h, corresponding with the median response time of 3.72 h shown in Table 2. In the North Fork of Peachtree Creek and the outlet of Nancy Creek, correlations are relatively constant after about 3 h and 4.5 h, respectively, when runoff from the upper reaches is contributing to discharge measurements. Correlations in these basins decrease slightly after 7 to 8 h. The temporal correlation is relatively constant for Sope Creek for periods longer than about 2 h, with somewhat higher correlations for periods longer than 7 h. This may indicate flood peak attenuation due to the relatively high degree of detention pond storage, or may reflect the variable influence of detention ponds on the hydrologic response of events of different magnitudes.

[40] In general, correlation between peak discharge and rain rate does not show a uniform dependence on the spatial distribution of impervious area. For example, urbanization in both the Outlet of Nancy Creek and the North Fork of Peachtree Creek is situated far from the basin outlet, while urbanization in Upper Nancy Creek and the South Fork of Peachtree Creek is closer to the basin outlet. However, correlations are higher for Upper Nancy Creek than for the Outlet of Nancy Creek, whereas correlations are higher for the North Fork of Peachtree Creek than for the South Fork. It is possible that the spatiotemporal rainfall structure affects these correlations differently in different basins.

[41] Rainfall-runoff relationships can also be examined in terms of runoff ratio, which is defined as the percentage of rainfall over the basin that leaves the basin as discharge at the basin outlet. The maximum rainfall and runoff volumes and corresponding runoff ratios as well as the coefficient of variation for 1, 3, 6, and 12 h periods containing the time of peak discharge were computed (Table 3). Upper Peachtree Creek shows the highest runoff ratios for all time periods but similar variability to other watersheds. The outlet of Nancy Creek shows the lowest runoff ratios for all time periods, but runoff ratios are more variable at 1, 3, and 6 h than at 12 h. This is due to the fact that the median response time of the outlet of Nancy Creek is greater than 6 h, and so for shorter time periods, the streamflow at the outlet does not include runoff from the upper reaches of the watershed. Peak discharge in upper Nancy Creek and the outlet of Peachtree Creek show a significant increase in sensitivity to 12 h rain rate than 6 h rain rate (Figure 9). For the other basins, differences between peak discharge sensitivity to 6 h and 12 h rain rates are less striking. A simple interpretation of this result is elusive, but may be related to the interactions of spatiotemporal rainfall structure with heterogeneous land surfaces. The wide spectrum of observed rainfall-runoff relations can be useful for formulating hypotheses and validating results from hydrologic or hydraulic models.

Figure 9.

(left) Maximum 6 h rainfall rate versus peak discharge and (right) maximum 12 h rainfall rate versus peak discharge for peaks over threshold for six basins. Linear trends are calculated using Theil-Sen Estimator. Upper Peachtree Creek is omitted to improve clarity.

Table 3. Median Flood Water Balance Components for Peak-Over-Threshold Data for Seven Watershedsa
WatershedMax 1 h Rainfall [mm]Max 1 h Discharge [mm]Runoff Ratio for 1 h [-]Max 3 h Rainfall [mm]Max 3 h Discharge [mm]Runoff Ratio for 3 h [-]Max 6 h Rainfall [mm]Max 6 h Discharge [mm]Runoff Ratio for 6 h [-]Max 12 h Rainfall [mm]Max 12 h Discharge [mm]Runoff Ratio for 12 h [-]
  • a

    Quantities in parenthesis are the unitless coefficient of variation.

Sope18.6 (0.51)2.2 (0.75)13 (0.59)28.3 (0.45)4.7 (0.80)18 (0.54)34.7 (0.48)6.5 (0.80)19 (0.45)42.3 (0.56)8.8 (1.00)21 (0.49)
Upper Peachtree17.5 (0.54)7.4 (0.85)43 (0.40)27.4 (0.61)11.5 (1.00)46 (0.44)32.7 (0.68)12.5 (0.98)45 (0.40)35.8 (0.72)14.4 (1.11)43 (0.35)
North Fork Peachtree17.2 (0.53)2.0 (0.50)14 (0.43)27.8 (0.54)5.7 (0.56)25 (0.29)33.7 (0.58)8.9 (0.65)31 (0.29)36.0 (0.57)12.2 (0.79)33 (0.33)
South Fork Peachtree19.5 (0.45)3.3 (0.32)16 (0.43)28.6 (0.42)7.3 (0.38)26 (0.33)37.8 (0.43)11.5 (0.50)34 (0.28)43.1 (0.44)13.8 (0.59)34 (0.32)
Outlet Peachtree15.9 (0.46)1.9 (0.39)12 (0.45)26.4 (0.45)4.9 (0.43)20 (0.33)31.7 (0.46)7.3 (0.53)26 (0.31)39.3 (0.41)9.3 (0.58)25 (0.31)
Upper Nancy18.1 (0.50)2.4 (0.41)15 (0.41)28.4 (0.51)5.7 (0.39)20 (0.39)34.7 (0.54)7.6 (0.51)24 (0.39)40.9 (0.51)10.5 (0.63)24 (0.38)
Outlet Nancy14.1 (0.50)1.1 (0.70)9 (0.70)22.9 (0.42)3.0 (0.71)14 (0.68)27.2 (0.46)4.8 (0.71)18 (0.72)36.3 (0.54)7.3 (0.66)20 (0.33)

5. Summary and Conclusions

[42] In this study we have examined the hydroclimatology of flash flooding for the Atlanta metropolitan area. The major findings of this paper are as follows.

[43] 1. The Atlanta metropolitan area has undergone rapid urbanization since the 1950s. This urbanization has impacted both rainfall and runoff processes. Nine Atlanta-area urban watersheds are examined, ranging from 3.7 to 225 km2 and exhibiting a range of land use characteristics. The percentage of impervious areas range from 19% in Sope Creek to 52% in Upper Peachtree Creek. The rate of urbanization has varied significantly as well. Sope Creek (76 km2) saw a 106% increase in urban land cover from 1974 to 2005, from 36% to 74%, whereas Upper Peachtree Creek (3.7 km2) saw a 30% increase, from 71% to 92% over the same period.

[44] 2. A high-resolution (1 km2, 15 min) bias-corrected radar rainfall data set was developed for 2002–2010 February–October period using the Hydro-NEXRAD system. Daily bias correction improves R2 from 0.74 to 0.88, decreases RMSE from 9.34 to 5.74 mm, and decreases MAE from 2.64 to 0.42 mm. Long-term bias-corrected radar rainfall datasets capture the spatial and temporal features of heavy precipitation and are useful for a variety of meteorological and hydrological applications, several of which are presented in this study.

[45] 3. A local maximum in both average annual rainfall and number of days of heavy rain is located within and east of the city, consistent with previous studies in Atlanta and other urban areas, which generally show rainfall enhancement downwind of the urban center. There is a rainfall minimum to the northwest and southwest of Atlanta, probably associated with the complex nature of flow over the ridge-valley system of the Appalachians and Blue Ridge Mountains. It is possible that this complex flow contributes to the elevated rainfall observed in and around the city. Future studies of urban impacts on rainfall should examine the influence of regional topography on storm motion and evolution. Bias-corrected Hydro-NEXRAD rainfall fields can provide a useful tool for evaluating such influences.

[46] 4. There has been an increase in the variability of annual flood peaks in Peachtree Creek and Sope Creek since 1960, with no apparent corresponding increase in the variability of annual maximum rainfall at hourly to daily time scales. It is likely that the increased variability in annual flood peaks can be attributed to the effects of urbanization on runoff processes. The increase in variability has been more pronounced in Sope Creek than in Peachtree Creek. In contrast to Peachtree Creek, urbanization in Sope Creek largely took place after the implementation of modern storm water management regulations. The greater increase in annual flood peak variability in Sope Creek suggests that these regulations have not been effective at mitigating the effects of urbanization on flood response.

[47] 5. Flood peak scaling analyses are presented for a diverse sample of urban watersheds in Atlanta. The populations of subwatersheds of Nancy Creek, Peachtree Creek, and Proctor Creek all scale consistently, while the subwatersheds in Utoy Creek are separated into two distinct populations. The subwatersheds with higher median annual floods are located on the southern branch of Utoy Creek, which is characterized by shorter flow paths and greater urbanization than the northern branch, suggesting that urban flood response is dependent on watershed area, flowpath length, and the extent of urban development.

[48] 6. A storm event data set of rainfall and runoff response for 35 to 45 storms per watershed was developed for seven of the study watersheds. Each watershed has a characteristic response time that is both a function of flow distance and land use characteristics. The median response time ranges from 0.74 h for Upper Peachtree Creek (3.7 km2) to 6.13 h for the outlet of Nancy Creek (98 km2), and is dependent on the degree and spatial distribution of urbanization as well as basin area. For example, the outlet of Peachtree Creek (225 km2) has a shorter median response time (3.84 h) than smaller Sope Creek (3.92 h; 76 km2) due to the rapid runoff response from the highly urban downtown.

[49] 7. Upper Peachtree Creek has the highest area-normalized peak discharge and most peaky hydrographs, whereas the outlet of Nancy Creek has the lowest area-normalized peak discharge and least peaky hydrographs. In most basins, peak discharge is highly correlated with peak rainfall durations roughly equal to the median response time. For example, peak discharge in Upper Peachtree Creek (median response time of 0.74 h) is highly correlated with maximum 0.25 h to 1.5 h maximum rainfall rates and less correlated with longer peak rainfall durations. Correlations between peak discharge and rain rate do not show a uniform dependence on the spatial distribution of impervious area.

Acknowledgments

[50] The research was supported by the National Science Foundation (grant CBET-1058027), the Willis Research Network, the NOAA Cooperative Institute for Climate Science, NASA GPM, and the U.S. Geological Survey (USGS), Department of the Interior, under USGS award G11AP20215. The authors thank the three anonymous reviewers for their helpful comments.

Ancillary