Climatological Analyses of Thunderstorms and Flash Floods in the Baltimore Metropolitan Region

88 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

ALEXANDROS A. NTELEKOS AND JAMES A. SMITH Department of Civil and Environmental Engineering, Princeton University, Princeton, New Jersey

WITOLD F. KRAJEWSKI IIHR-Hydroscience and Engineering, University of Iowa, Iowa City, Iowa

(Manuscript received 9 February 2006, in final form 22 June 2006)

ABSTRACT

The climatology of thunderstorms and flash floods in the Baltimore, Maryland, metropolitan region is examined through analyses of cloud-to-ground (CG) lightning observations from the National Lightning Detection Network (NLDN) and discharge observations from 11 U.S. Geological Survey (USGS) stream gauging stations. A point process framework is used for analyses of CG lightning strikes and the occurrences of flash floods. Analyses of lightning strikes as a space–time point process focus on the mean intensity function, from which the seasonal, diurnal, and spatial variation in mean lightning frequency are examined. Important elements of the spatial variation of mean lightning frequency are 1) initiation of thunderstorms along the Blue Ridge, 2) large variability of lightning frequency around the urban cores of Baltimore and Washington D.C., and 3) decreased lightning frequency over the Chesapeake Bay and Atlantic Ocean. Lightning frequency has a sharp seasonal maximum around mid-July, and the diurnal cycle of lightning frequency peaks between 2100 and 2200 UTC with a frequency that is more than an order of magnitude larger than the minimum frequency at 1200 UTC. The seasonal and diurnal variation of flash flood occurrence in urban streams of Baltimore mimics the seasonal and diurnal variation of lightning. The peak of the diurnal frequency of flash floods in Moores Run, a 9.1-km2 urban watershed in Baltimore City, occurs at 2200 UTC. Analyses of the lightning and flood peak data also show a close link between the occurrence of major thunderstorms systems and flash flooding on a regional scale.

1. Introduction ments, which is tied to expansion of the drainage network through the storm drain system and increasing Flash flooding in urban drainage basins is an increas- impervious area, is characterized by increased sensitiv- ingly important hazard in terms of loss of life and prop- ity to short-duration rainfall rates (Smith et al. 2002, erty damage (i.e., Carpenter et al. 1999; Smith et al. 2005a,b). The altered hydrologic response associated 2002; Ogden et al. 2000). In this paper we examine the with urbanization leads to the hypothesis that warm- linkage between organized thunderstorm systems and season thunderstorm systems should become an in- flash floods in urban drainage basins through analyses creasingly important component of the climatology of of cloud-to-ground (CG) lightning observations and flash flooding. discharge data from the Baltimore, Maryland, metro- Alteration of the physical environment due to urban- politan region. ization also affects the climatology of precipitation in The questions that motivate this study arise from the urban environments (Landsberg 1956; Manley 1958; alteration of the physical environment through urban- Changnon et al. 1971; Diem and Mote 2005; Shepherd ization. The altered flood hydrology of urban catch- and Burian 2003; Moelders and Olson 2004). Observa- tional and numerical modeling studies suggest that the urban heat island (UHI), frictional effects associated Corresponding author address: James A. Smith, Princeton Uni- versity, Department of Civil and Environmental Engineering, with building canopy, and urban aerosols can affect the Princeton, NJ 08544. initiation and evolution of thunderstorm systems E-mail: [email protected] (Shepherd and Burian 2003; Bornstein and Lin 2000;

DOI: 10.1175/JHM558.1

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC

JHM558 FEBRUARY 2007 NTELEKOS ET AL. 89

Ϫ2 TABLE 1. USGS/BES stations used in this study along with the available data and the flood level, zo (cms km ), for each station.

Station ID Station name Available data zo 01583580 Baisman Run at Broadmoor, MD Jan 2000–Dec 2004 0.09 01589330 Dead Run at Franklintown, MD Jan 2000–Oct 2004 1.02 01589100 East Branch Herbert Run at Arbutus, MD Jan 2000–Aug 2000 1.02 01589197 Gwynns Falls near Delight, MD Jan 2000–Dec 2004 0.41 01589238 Gwynns Falls Tributary at McDonogh, MD Jan 2000–Aug 2003 0.15 01583980 Minebank Run at Loch Raven, MD Jan 2000–Aug 2003 0.51 01585230 Moores Run at Radecke Avenue at Baltimore, MD Jan 1998–Aug 2003 1.76 01585225 Moores Run tributary near Todd Avenue at Baltimore, MD Jan 1998–Aug 2003 2.30 01583570 Pond Branch at Oregon Ridge Jan 2000–Nov 2004 0.11 01589340 Rognel Hgts Storm Sewer Outfall at Baltimore, MD Jan 2000–Sep 2004 1.20 01585090 Whitemarsh Run Near Fullerton, MD Jan 2000–Jun 2003 0.99

Rozoff et al. 2003; Bentley and Stallins 2005; Orville et network records the date, time, latitude, and longitude al. 2001; Baik et al. 2001; Lowry 1998; Peterson 2003; Li (decimal degrees, three decimal points accuracy), po- et al. 2004; Hafner and Kidder 1999; Van den Heever larity (kA), and multiplicity (number of strokes per and Cotton 2004). Summer convective precipitation flash) of CG lightning strikes over the conterminous anomalies over urban environments have been exam- United States. Upon detection, only CG lightning ined over several cities in the United States and world- strikes that have an effective peak current magnitude wide [see Shepherd (2005) for a recent discussion and greater than 5 kA are recorded. The network has been summary]. Although there have been numerous studies extensively used for climatological analyses of lightning of precipitation anomalies in urban environments, and thunderstorms over the United States (Carey and there have been few studies that examine thunderstorm Rutledge 2003; Orville and Huffines 2001; Orville and and flash flood climatologies for urban regions [see Un- Silver 1997). The NLDN has an overall detection effi- derwood and Schultz (2004) and Holle and Bennett ciency of about 80%–90% and a location accuracy of (1997) for studies dealing with aspects of the issue]. approximately 1 km (Cummins et al. 1998; Idone et al. We use a point process framework (Karr 1991; 1998a,b). Error characteristics of the NLDN sensors Diggle 1983; Guttorp 1995; Smith and Karr 1985) for vary with geographic location, peak current magnitude, analyses of CG lightning strikes and the occurrences of season of the year, and density of the detectors (for flash floods. Analyses of lightning strikes as a space– additional details, see discussion in the references listed time point process focus on the mean intensity function, above). For this study, a record of 14 yr (1991–2004) of from which we examine the seasonal, diurnal, and spa- lightning data from the NLDN were used. tial variation in mean lightning frequency. Cloud-to- The Baltimore Ecosystem Study (BES) is a Long- ground lightning data from 1991 to 2004 from the Na- Term Ecological Research (LTER) project that focuses tional Lightning Detection Network (NLDN) are used on the impacts of urbanization on riparian ecosystem to study the spatial and temporal frequency of thunder- form and function (see Groffman et al. 2003). A major storm systems. Discharge data during the period 2000– observational component of the BES has been the de- 04 from 11 U.S. Geological Survey (USGS) stream velopment of a dense network of stream gauging sta- gauging stations are used to analyze flash flood occur- tions in small urban watersheds by the USGS. We use rences over the Baltimore metropolitan region. discharge observations for 11 of these stations (see The structure of the paper is as follows. In section 2 Table 1). The time resolution of discharge observations the study area and data are introduced. The point pro- ranges from 1 to 15 min. The discharge data used in this cess modeling framework used for statistical analysis of study cover the 5 year period from 2000 to 2003 or 2004, lightning and flash flood occurrences is presented in with the exception of the Moores Run station at section 3. Results follow in section 4 and conclusions Radecke Avenue (hereafter abbreviated as “Moores are summarized in section 5. Run”) and its tributary, for both of which the observing period is 1998–2003. 2. Data and study region We concentrate our study on a region centered over the Baltimore metropolitan area (Fig. 1). The urban The NLDN (Cummins et al. 1998; Orville et al. 2002) environment of Baltimore and Washington D.C., the consists of 106 CG lightning sensors operated and mountainous terrain of the Blue Ridge and Valley and maintained by Vaisala, Inc., in Tucson, Arizona. The Ridge province to the west, and the ocean–land bound-

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 90 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

FIG. 1. Study area (large white box) and subareas of focus (smaller white boxes).

ary to the east make this region a complex setting for {Tk, Yk} is a space–time point process defined on examining warm-season thunderstorm systems. [0, 1] ϫ D, where the unit interval represents time dur- For analyses of the diurnal cycle of thunderstorms ing the day and the domain of study is the two- and investigation of urbanization effects on thunder- dimensional region D. We choose the beginning and storm rainfall, we focus on three subareas. These areas ending time of the day, that is, times 0 and 1, to be 1200 are depicted in Fig. 1 with the three intermediate size UTC. squares of the same latitude marked as “Piedmont,” The CG lightning sample consists of observations “Urban,” and “Ocean.” The urban area covers the Bal- from every day of the year and from multiple years. Let timore and Washington D.C., metropolitan areas. The Nij denote the number of CG strikes in the study region Piedmont and ocean regions are of comparable size and on day j of year i. Days vary from 1 to 365 with 1 are located, respectively, to the west of the Baltimore– January being the first day of the year and 31 Decem- Washington region and over the open ocean east of ber being the last one. Leap years are neglected by Baltimore. The 11 USGS stream gauge stations used omitting 29 February (with no impact on subsequent for flash flood analyses (Table 1) are contained within analyses due to the low lightning frequency in Febru- the box covering Baltimore. ary). The number of years in the sample is denoted by Ͼ ∈ n. For Nij 0, that is, on days with lightning, Tijk [0, 1] denotes the time of the kth CG strike on day j of year 3. Methodology ∈ i, and Yijk D denotes the location in the domain D of We represent CG lightning strikes over the study re- the kth CG strike on day j of year i. The first day gion as a space–time point process. Because thunder- extends from 1200 UTC 1 January to 1200 UTC 2 Janu- storms are fundamentally linked to the diurnal cycle of ary. heating of the land surface, the basic time element of The counting process of CG strikes on day j of year the space–time model of lightning strikes is one day. i is given by

Each CG lightning strike is associated with a point in Nij ͑ ͒ ϭ ͑ Յ ∈ ͒ ͑ ͒ space and a time, that is, the spatial location at the Mij t, A ͚ 1 Tijk t, Yijk A , 1 ground surface of the CG strike and the time of the CG kϭ1 Յ ∈ strike. For a given day, let N denote the number of where 1(Tijk t, Yijk A) is the indicator function Ͼ Յ lightning strikes. For N 0, denote the time of the taking the value 1 if Tijk t and Yijk is in the subset A Յ kth strike (k N)byTk and the location by Yk. Then of the study domain D, otherwise the indicator function

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 91

n 365 Nij takes the value 0; Mij (t, A) is the number of strikes in ␭ˆ ͑ ͒ ϭ Ϫ1 ͑|| Ϫ ||͒ area A, up until time t on day j of year i. For a fixed day j t, x n ͚͚͚ K1 j l ϭ ϭ ϭ j, it is assumed we have n independent and identically i 1 l 1k 1 и ͑|| Ϫ ||͒ и ͑|| Ϫ ||͒ ͑ ͒ distributed (i.i.d.) copies of the space–time point pro- K2 t Tijk K3 x Yijk , 6 cess of CG strikes for that day (see section 4 for analy- where K , K , and K are kernel functions. The simplest ses pertaining to this assumption). For a fixed year i, the 1 2 3 form of kernel functions would be to take K (||jϪl||)to identically distributed assumption breaks down. The 1 be1ifl equals j and 0 otherwise; K (||tϪT ||) equal to space–time point process of CG strikes for a day in 2 ijk ␦tϪ1 if ||tϪT || is less than ␦t; and K (||xϪY ||) equal to mid-July, for example, may not have the same distribu- ijk 3 ijk ␦AϪ1 if ||xϪY || is less than ␦A. The fundamental prop- tion as that of CG strikes in early February (they are, in ijk erty of kernel functions is that they are probability den- fact, quite different). The independence assumption sity functions, so K integrates to 1 over the entire do- may also break down. Cloud-to-ground strikes on day j 3 main D, K integrates to 1 over the time interval [0, 1], of year i may not be independent of CG strikes on day 2 and K sums to 1 over the integers from 0 to 364. For j ϩ 1 of year i. 1 estimating the mean intensity function, we will use The distributional properties of the space–time point Gaussian kernels, which take the form process of CG lightning strikes exhibit pronounced clustering in space and time and are difficult to com- 1 1 s 2 K͑s͒ ϭ expͭϪ ͩ ͪ ͮ. ͑7͒ pletely specify. We will not attempt to completely ͑2␲␴2͒1ր2 2 ␴ specify the distribution of CG strikes, but rather attempt to characterize the mean intensity of CG strikes The parameter ␴ determines the width of the kernel ␭ through the intensity function j(t, y), which represents estimator, that is, how close or far away points must be the mean rate of occurrence of lightning (in CG strikes to receive significant weight in the estimator. The larger kmϪ2 dayϪ1) at time t during the day and location y on the parameter ␴, the greater the smoothing applied to day j of the year. It is defined by the following relation- the data [for more information on the smoothing tech- ship: niques see Scott (1992)]. To estimate the diurnal cycle of lightning on day j of t ͓ ͑ ͔͒ ϭ ͵ ͵ ␭ ͑ ͒ ͑ ͒ the year, estimators simplify to E Mij t, A j s, y ds dy. 2 A 0 n 365 Nij ͑ ͒ ϭ Ϫ1 ͑|| Ϫ ||͒ ͑|| Ϫ ||͒ We will examine the climatology of CG strikes through mˆ j t n ͚ ͚ ͚ K1 j l · K2 t Tilk iϭ1 lϭ1 kϭ1 statistical analyses of integral functions of the mean intensity. 365 n Nij ϭ ͑|| Ϫ ||͒ Ϫ1 ͑|| Ϫ ||͒ The diurnal cycle of mean CG strikes on day j over ͚ K1 j l n ͚ ͚ K2 t Tilk lϭ1 iϭ1 kϭ1 the entire domain D is given by 365

Ϫ ϭ K ͑||j Ϫ l||͒ m˜ ͑t͒. ͑8͒ ͑ ͒ ϭ | | 1͵ ␭ ͑ ͒ ∈ ͓ ͔ ͑ ͒ ͚ 1 l mj t D j t, y dy; t 0, 1 , 3 lϭ1 D Thus the kernel estimator of the diurnal cycle on day j where |D| is the area (km2) of the domain D. The mean is a smoothed form of the kernel estimators m˜ (t) for seasonal cycle of lightning is specified by l each day l, with the smoothing reflecting how close day 1 l is to day j. ϭ ͵ ͑ ͒ ϭ ͑ ͒ mj mj s ds; j 1,...,365. 4 The kernel estimator for spatial variation of the 0 mean annual lightning is given by At any location y, the temporally averaged rate of oc- n 365 Nij currence of lightning activity is given by the function ˆ͑ ͒ ϭ Ϫ1 ͑|| Ϫ ||͒ h x n ͚ ͚ ͚ K3 x Yijk iϭ1 lϭ1 kϭ1 365 1 ͑ ͒ ϭ ͵ ␭ ͑ ͒ ͑ ͒ n N h y ͚ j t, y dt. 5 ij jϭ1 0 ϭ Ϫ1 ͑|| Ϫ ||͒ ͑ ͒ 365 n ͚ ͚ K3 x Yijk . 9 ␭ iϭ1 kϭ1 To estimate the intensity function j (t, x), we will simply count the number of CG strikes that are close to The estimator for summer lightning (following section) the location x and occur around time t on a day that is or monthly lightning takes essentially the same form. near day j of the year. To do so, we will use “kernel The mean intensity function does not completely estimators” of the form specify the distribution of the space–time point process.

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 92 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

We also examine aspects of the lightning distribution 4. Results that are not directly linked to the mean intensity func- Statistical analyses of lightning and flash flood occur- tion. To examine the initiation of thunderstorms, we rences in the Baltimore metropolitan region are pre- analyze the distribution of the location of the first CG sented in this section. We begin with the space–time strike Yij1 on a given day. We are particularly interested in the initiation of thunderstorms on major outbreaks point process representation of CG lightning strikes. of thunderstorms, so analyses will be conditioned on These results are followed by analyses of the occur- total CG strikes for the day. We examine the evolution rence of flash floods and an examination of the joint of CG strikes in the following section through two ex- occurrence of major thunderstorm systems and regional amples involving major flash flood episodes. To char- flash flooding. acterize the evolution of CG strikes over the course of a. Lightning climatology a day, it is necessary to completely specify the distribution through the “conditional intensity function” (Karr The annually averaged rate of occurrence of light- 1991). Heuristically, the conditional intensity function ning hˆ(y) (Fig. 2) over the study area varies from 1.5 to specifies the rate of occurrence of lightning strikes at a 4.6 strikes kmϪ2 yrϪ1. The values are lowest in the Val- point in time t and a point in space x, conditioned on the ley and Ridge physiographic province, which lies west history of CG lightning strikes up until time t. Analyses and north of the Blue Ridge. An elongated region of in the following section point to useful directions to maximum lightning strikes extends from the southern pursue in specifying explicit models for the conditional portion of the Blue Ridge, in our study region, to the intensity function. Baltimore–Washington metropolitan region. The sharp We will also examine the occurrence of flash floods gradients that parallel the strike of the lower portion of in a drainage basin in a point process framework. Let the Blue Ridge, with lower flash densities to the west in

Zij denote the maximum daily discharge, as a unit dis- the Valley and Ridge and higher flash densities to the charge (m3 sϪ1 kmϪ2), at a stream gauging station on east over the Piedmont province point to the role of the day j of year i. Denote the time of the peak discharge on “first ridge” for the climatology of thunderstorms in the day j of year i by Uij, with time again ranging from 0 to region. As will be shown below, the Blue Ridge is a 1, with the beginning and ending times equal to 1200 region with a high frequency of thunderstorm initiation. UTC. The peak discharge on each day will not neces- Lightning frequency varies in complex fashion sarily be a “flood.” We introduce a discharge threshold around the Baltimore–Washington metropolitan re- (m3 sϪ1 kmϪ2) to define a flood. gions (Fig. 2), with areas of larger flash densities south The counting process for a flood on a given day is and east of the cities. The impacts of urbanization on specified by the binary random variable: the thunderstorm climatology of the region are linked to the impacts of sea-breeze circulations associated with ͑ ͒ ϭ ͑ Յ Ͼ ͒ ͑ ͒ Xij t 1 Uij t; Zij z0 . 10 the Atlantic Ocean, Chesapeake Bay, and the Potomac estuary. Analogous to the previous model, we can character- If the time window is restricted to the summer season ize the occurrence process through a rate function: (the months of June, July, and August), the estimated rate of occurrence of lightning hˆ(y) (Fig. 3) shows simi- t ͓ ͑ ͔͒ ϭ ͵ ␥ ͑ ͒ ͑ ͒ lar patterns to the annual analyses, but with markedly E Xij t j s ds. 11 Ϫ Ϫ2 Ϫ1 0 larger rates (4.3 15.1 strikes km yr ). These values of lightning frequency are rates. A value of 15.1 strikes Ϫ Ϫ In this case the rate function is given by km 2 yr 1, for example, is the value that would result at the spatial location of maximum frequency if the mean d summer rate of occurrence persisted for the entire year. ␥ ͑t͒ ϭ P{U Յ t|Z Ͼ z } P{Z Ͼ z }. ͑12͒ j dt ij ij o ij o Since the summer season consists of 3 months, the summer climatology (strikes kmϪ2 summerϪ1 in this case) The rate of occurrence of floods at time t on day j is the could be obtained by dividing the values of Fig. 3 by a probability of a flood on that day times the conditional factor of 4. If the values of Fig. 3 are divided by 4 they density function of the time (during the day) of a flood are close to the annual values of Fig. 2, implying that peak, given that the peak exceeds z0. The probability of lightning is concentrated in the summer months (see a flood on day j is given by also Fig. 4). Amplification of CG flash rates southeast of Balti- ␥ ϭ Ն ͑ ͒ j P{Zij z0}. 13 more and Washington D.C. is a more pronounced fea-

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 93

Ϫ Ϫ FIG. 2. Regional contour map of mean annual lightning flash density in strikes km 2 yr 1 over the study area. Data are from 1991 to 2004. ture in the summer analyses. Lightning strikes increase The diurnal cycle of the Piedmont domain is gener- to about 13.5 strikes kmϪ2 yrϪ1 to the southeast of the ally similar to that of the Baltimore–Washington cities and the amplification extends for distances of 50– domain, except that the peak diurnal flash rate is ap- 60 km. Future studies will examine whether this ampli- proximately 1 h earlier and slightly smaller. The timing fication of CG strikes is tied to urbanization impacts on difference reflects the preferential initiation of thunder- thunderstorm evolution. storms along the Blue Ridge and their eastward pro- The seasonal occurrence of thunderstorms is tightly gression over the course of the day, as discussed in concentrated during the warm season (Fig. 4), as re- more detail below. The diurnal cycle of thunderstorms flected in estimates of the seasonally varying intensity is strikingly different for the eastern region over the function mj for the entire study region. Lightning flash Atlantic Ocean, with its 0300 UTC peak at a flash rate density peaks during the end of June and beginning of that is almost 5 times smaller than the Baltimore– July with flash rates over the region averaging 10 strikes Washington maximum [see Toracinta et al. (2002) and kmϪ2 yrϪ1. Lightning flash rates are close to 0 from Nesbitt et al. (2000) for analyses and discussion of October to March. mechanisms associated with land–ocean contrasts in The diurnal cycle of warm-season lightning occur- lightning frequency]. rences over the Baltimore–Washington domain (Fig. 1), Of particular interest in examining flash floods are as represented by the spatially averaged rate of occur- major thunderstorm systems. We examine below the Ϫ2 Ϫ1 rence mj (t), has a sharp peak of 38 strikes km yr linkage between these storms, as reflected in total re- between 2100 and 2200 UTC. For the analyses in Fig. 5, gional lightning counts, and the occurrence of flash the diurnal flash rate mj (t) is averaged over the warm floods. The 10 largest thunderstorms for the broader season (June–August). Cloud-to-ground flash rates are study area and for the area covering just the city of close to 0 from 0900 to 1500 UTC, which is a principal Baltimore are listed in Tables 2 and 3, respectively. The rationale for beginning the day of our lightning model largest storm for the larger study region occurred on 16 at 1200 UTC. August 2003 and produced a flash density of 0.6 strikes

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 94 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

Ϫ FIG. 3. Regional contour map of mean summer (June–August) lightning flash density in strikes km 2 yrϪ1 over the study area. Data are from 1991 to 2004. kmϪ2 over the 56 066 km2 study region. For the Balti- major thunderstorm events has been higher during the more area, the largest storm occurred on 7 July 2004 2000–04 period than for the preceding time period. and produced a mean flash density of 1.31 strikes kmϪ2 For the major storm events (Table 3) in the Balti- over the 1973 km2 region. This result is enlightening in more region (Fig. 1), the distribution of initiation loca- understanding the magnitude of this particular thunder- tions {Yij1} was examined (Fig. 6). We computed the storm. In a matter of few hours, approximately 40% of probability that a grid of resolution 0.02° by 0.02° was the mean annual lightning activity of the area was ob- the initiation location by taking the first 100 CG strikes served. The storm produced record flooding over the for each of the 10 days and computing the frequency by Baltimore metropolitan region (Smith et al. 2005a, grid. All of the 10 largest storms for the Baltimore re- 2007). Only two of the events from the larger area are gion initiated far from the city and propagated over it. also included in the list of events impacting the city of The initiation locations cluster along and adjacent to Baltimore (16 June 1994 and 25 May 2004). the Blue Ridge. This conclusion is not sensitive to the Under the i.i.d. assumption made for the space–time number of storms. There is a strong link between the point process model in section 2, it is presumed that Blue Ridge and initiation of thunderstorm systems that there are no time trends in lightning from year to year. pass over the Baltimore region. In this connection, it is notable that 7 of the 10 largest events for the broader study region occurred during the b. Life cycle of warm-season thunderstorm systems last 5 yr. Updates to the NLDN (Cummins et al. 1998) The life cycle of thunderstorm systems that produce could possibly lead to reduced detection prior to 1995. flash floods in the Baltimore metropolitan region ex- Even so, in the list of the 10 largest events here there is hibit systematic elements in their initiation and evolu- only one between 1995 and 1999. Even with the differ- tion. These include initiation of thunderstorms along ent set of events from the smaller Baltimore study re- the Blue Ridge (as described above), eastward propa- gion the conclusions are the same. The frequency of gation of thunderstorm systems over the Piedmont, and

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 95

FIG. 4. Seasonal cycle of lightning flash density over the study Ϫ Ϫ FIG. 5. Diurnal cycles of summer (June–August) lightning flash region in strikes km 2 yr 1. Original data have been smoothed density in strikes kmϪ2 yrϪ1 for three subregions of Fig. 1. The with a Gaussian filter (␴ ϭ 10 days). Data are from 1991 to 2004. black, dark gray, and light gray lines are respectively for the boxes covering the Piedmont, the Baltimore–Washington metropolitan, interaction of thunderstorm systems with the urban en- and the ocean sites of the same latitude (see Fig. 1). Data are from vironment and sea-breeze circulation systems. These 1991 to 2004. properties are illustrated by storms on 7 July 2004 (Fig. 7) and 13 June 2003 (Fig. 8). These storms produced the conditional intensity function (Karr 1991). Future stud- floods of record in, respectively, Dead Run (Table 1; ies will explore the climatology of thunderstorms see also Smith et al. 2005a, 2006) and Moores Run (see through analyses based on explicit models of the con- analyses below and Table 1; see also Smith et al. 2005b). ditional intensity function. The 7 July 2004 storm ranks first in CG flash density over the Baltimore metropolitan region, and the 13 c. Joint occurrence of lightning and flash floods June 2003 storm is not in the list of the 20 largest lightning producing storms over Baltimore. The relationship between thunderstorm climatology The life cycle of thunderstorm systems is reflected in and flash flood climatology is examined through dis- quantities derived from the mean intensity function, charge observations from 11 USGS stream gauging sta- like the spatial rate of occurrence function h(y) (Figs. 2 tions in the Baltimore metropolitan region and the and 3) and the diurnal frequency mj(t) (Fig. 5). The point process framework introduced in section 2. To mean intensity function does not, however, capture the examine the occurrence of flash floods, we select the complete distribution of CG lightning strikes. In par- discharge threshold z0 such that there are 10 events per ticular, those elements that are associated with propa- year on average at each stream gauging station. The gation and interaction of thunderstorm systems are not BES network comprises gauging stations that span a represented. To completely specify the space–time broad spectrum of urban development. The threshold, 3 Ϫ1 Ϫ2 point process of CG strikes requires specification of the z0 (Table 1) ranges from 1.76 m s km for Moores

TABLE 2. Ten largest lightning events over the study region. TABLE 3. Ten largest lightning events over Baltimore City. Data Data are from 1991 to 2004. are from 1991 to 2004.

Order Date Strikes kmϪ2 Order Date Strikes kmϪ2 1 16 Aug 2003 0.60 1 7 Jul 2004 1.31 2 14 Jul 2000 0.53 2 25 May 2004 1.18 3 17 Jun 2004 0.51 3 13 May 2000 1.10 4 16 Jun 1994 0.43 4 18 Apr 2002 1.09 5 25 May 2004 0.42 5 22 Aug 2003 1.00 6 27 Aug 2003 0.41 6 6 Jul 2003 0.92 7 24 Jul 1996 0.40 7 16 Jun 1994 0.88 8 11 Aug 1992 0.35 8 2 Aug 2002 0.86 9 26 Aug 2003 0.34 9 9 Sep 1999 0.75 10 22 Jul 2001 0.32 10 17 Sep 1991 0.73

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 96 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

FIG. 6. Raster map of the initiation locations, expressed in probability (%), of the first 100 CG lightning strikes of the 10 largest lightning-producing thunderstorm events over Baltimore City (see Table 3). Each thunderstorm day starts at 1600 UTC. Raster resolution is 0.02° by 0.02°. Data are from 1991 to 2004.

Run to 0.09 m3 sϪ1 kmϪ2 for Baisman Run (the value of and 2200 UTC. The maximum frequency is more than 2.30 m3 sϪ1 kmϪ2 is for a tributary of the 9.1 km2 a factor of 3 larger than the minimum frequency. The Moores Run watershed, which drains less than 5% of maximum frequency of flash floods corresponds with the basin). Moores Run is located in northeastern Bal- the maximum frequency of lightning, reinforcing the timore City and is characterized by residential devel- link between flash flood occurrence and thunderstorm opment constructed prior to implementation of storm occurrence. water management regulations. It has one of the high- To examine the joint occurrence of major thunder- est frequencies of occurrence of 1 m3 sϪ1 kmϪ2 flood storm systems and regional flash flooding, counts of CG peaks in the conterminous United States (Smith et al. lightning strikes were compared with counts of flash 2005b). Baisman Run is the forest reference watershed flood events, based on observations from 11 USGS for the BES. Flood response for this basin reflects hy- stream gauging stations (Table 1). The joint occurrence drologic response of the region prior to urban develop- is examined through comparisons of the number of ment. stream gauging stations with floods on a given day and The seasonal cycle of flood response in Moores Run the number of CG strikes for the Baltimore region (see (Fig. 9) mimics the seasonal cycle of lightning occur- Fig. 11 in which results for the year 2000 are illus- rence with a sharp maximum during June and July. The trated). The observations show a strong correspon- frequency of events ranges from a maximum of 25 dence between lightning counts and flash floods during events yrϪ1 during June and July to frequencies that are the summer months (June–August). more than an order of magnitude smaller in Novem- The correspondence between intensity of thunder- ber–February. storms and flash flooding is illustrated through a scat- The diurnal cycle of flash floods in Moores Run (Fig. terplot (Fig. 12) of the total number of lightning strikes 10) exhibits a sharp maximum frequency between 2100 in a day over Baltimore (expressed in strikes kmϪ2

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 97

FIG. 7. Hourly CG lightning strikes (white x marks) for the 7 Jul 2004 event. Gray-shaded areas with black outline denote the urban regions of Baltimore and Washington D.C. See also Fig. 1.

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 98 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

FIG. 8. Same as in Fig. 7 but for 13–14 Jun 2003. yrϪ1) versus the number of stations with peak discharge number of lightning strikes increases there is a corre- exceeding the flood threshold for the same day during sponding increase in regional flash flooding, especially the summer months of 2000–03. This analysis includes as CG strike frequency increases beyond 0.05 strikes all the summer days during the 4-yr period. As the kmϪ2 or about 100 strikes over the area covering Bal-

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 99

FIG. 11. Daily CG lightning strikes (solid black line) over the FIG. 9. Seasonal cycle of flash floods for USGS gauging station area covering Baltimore City and of daily number of USGS gaug- “Moores Run at Radecke Avenue in Baltimore City, MD.” Origi- ing stations exceeding the flood threshold, z0 (vertical gray bars nal data have been smoothed with a Gaussian filter (␴ ϭ l month). ending with solid circles) for the year 2000. Data from 11 stations were used (see Table 1). timore for a certain day. There are a total of 56 days during the 4 yr (total of 368 summer days) with one or tions. A point process modeling framework serves as more stream gauging stations reporting floods. Of those the foundation for theses analyses. A space–time point 56 days, 50 were days for which lightning was observed process representation of CG lightning strikes is intro- over the area. This translates to a 90% conditional duced for the study of the climatology of thunder- probability that lightning occurs given that a flash flood storms. Principal conclusions of the study are the fol- has occurred. lowing: 1) The spatial frequency of CG lightning strikes for 5. Summary and conclusions thunderstorms that affect the Baltimore metropolitan region is characterized by pronounced spatial The climatology of thunderstorms and flash floods in heterogeneity. Elements of the spatial heterogeneity the Baltimore metropolitan region has been examined of thunderstorms include a maximum frequency through analyses of CG lightning observations and discharge observations from USGS stream gauging sta-

FIG. 12. Scatterplot of regional flooding (expressed in number Ն of stations for which z z0) vs total number of lightning strikes Ϫ FIG. 10. Diurnal cycle of flash floods for USGS gauging station over Baltimore City (in strikes km 2) for days with lightning ac- “Moores Run at Radecke Avenue in Baltimore City, MD.” Origi- tivity. Only the summer months are included (June–August). nal data have been smoothed with a Gaussian filter (␴ ϭ 2 h). Data are from 2000 to 2003.

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC 100 JOURNAL OF HYDROMETEOROLOGY VOLUME 8

over the Blue Ridge, a minimum frequency to the Carey, L. D., and S. A. Rutledge, 2003: Characteristics of cloud- west of the Blue Ridge in the Valley and Ridge to-ground lightning in severe and nonsevere storms over the province, large variation in lightning frequency central United States from 1989–1998. J. Geophys. Res., 108, 4483, doi:10.1029/2002JD002951. around the Baltimore and Washington metropolitan Carpenter, T. M., J. A. Sperfslage, K. P. Georgakakos, T. areas, and decreased thunderstorm frequencies to Sweeney, and D. L. Fread, 1999: National threshold runoff the east over the open ocean. These analyses indi- estimation utilizing GIS in support of operational flash flood cate that the climatology of thunderstorms, and con- warning systems. J. Hydrol., 224, 21–44. sequently the climatology of flash floods, is strongly Changnon, S. A., Jr., F. A. Huff, and R. G. Semonin, 1971: METROMEX: An investigation of inadvertent weather dependent on orographic precipitation mechanisms modifications. Bull. Amer. Meteor. Soc., 52, 958–968. along the Blue Ridge. The climatology of thunder- Cummins, K. L., M. J. Murphy, E. A. Bardo, W. L. Hiscox, R. B. storms may also include significant effects associ- Pyle, and A. E. Pifer, 1998: A combined TOA/MDF technol- ated with urbanization (UHI, frictional effects asso- ogy upgrade of the U.S. National Lightning Detection Net- ciated with urban canopy, and urban aerosols) and work. J. Geophys. Res., 103, 9035–9044. Diem, J. E., and T. L. Mote, 2005: Interepochal changes in sum- sea-breeze circulation systems. mer precipitation in the southern United States: Evidence of 2) The climatology of thunderstorms in the Baltimore possible urban effects near Atlanta, Georgia. J. Appl. Me- metropolitan region exhibits pronounced seasonal teor., 44, 717–730. and diurnal cycles. These cycles are reflected in the Diggle, P. J., 1983: Statistical Analysis of Spatial Point Patterns. climatology of flash floods in Baltimore. Academic Press, 148 pp. Groffman, P. M., and Coauthors, 2003: Down by the riverside: 3) There are systematic elements to the life cycle of Urban riparian ecology. Front. Ecol. Environ., 1, 315–321. thunderstorm systems that produce flash floods in Guttorp, P., 1995: Stochastic Modeling of Scientific Data. Chap- the Baltimore metropolitan region. These include man and Hall, 384 pp. initiation of thunderstorms along the Blue Ridge, Hafner, J., and S. Q. Kidder, 1999: Urban heat island modeling in eastward propagation of thunderstorm systems over conjunction with satellite-derived surface/soil parameters. J. the Piedmont, and interaction of thunderstorm sys- Appl. Meteor., 38, 448–465. Holle, R. L., and S. P. Bennett, 1997: Lightning ground flashes tems with the urban environment and sea-breeze associated with summer 1990 flash floods and streamflow in circulation systems. Tucson, Arizona: An exploratory study. Mon. Wea. Rev., 125, 4) Most of the large thunderstorm events impacting the 1526–1536. study region occurred between 2000 and 2004 in a Idone, V. P., D. A. Davis, P. K. Moore, Y. Wang, R. W. Hender- dataset that extends back to 1994. The interannual son, M. Ries, and P. F. Jamason, 1998a: Performance evaluation of the U.S. National Lightning Detection Network in variation of lightning climatology warrants addi- eastern New York. 1: Detection efficiency. J. Geophys. Res., tional study. 103 (D8), 9045–9055. 5) There is a strong link between the occurrence of ——, ——, ——, ——, ——, ——, and ——, 1998b: Performance regional flash flooding and major thunderstorm out- evaluation of the U.S. National Lightning Detection Network breaks (see also item 2 above). Urbanization in Bal- in eastern New York. 2: Location accuracy. J. Geophys. Res., 103 (D8), 9057–9069. timore, and along the major urban corridors of the Karr, A. F., 1991: Point Processes and Their Statistical Inference. eastern United States, has made warm-season thun- 2d ed. Marcel Dekker, 490 pp. derstorm systems the principal agent of flash flood- Landsberg, H. E., 1956: The climate of towns. Man’s Role in ing in urban drainage basins. Changing the Face of the Earth, Vol. 2, W. L. Thomas Jr., Ed., The University of Chicago Press, 584–606. Acknowledgments. The research was supported by Li, Q., H. Zhang, X. Liu, and J. Huang, 2004: Urban heat island effect on annual mean temperature during the last 50 years in the National Science Foundation (NSF Grants EAR- China. Theor. Appl. Climatol., 79, 165–174. 0208269, EAR-0409501, and ITR-0427325). This sup- Lowry, W. P., 1998: Urban effects on precipitation amount. Prog. port is gratefully acknowledged. Phys. Geogr., 22, 477–520. Manley, G., 1958: On the frequency of snowfall in metropolitan Quart. J. Roy. Meteor. Soc., REFERENCES England. 84, 70–72. Moelders, N., and M. A. Olson, 2004: Impact of urban effects on Baik, J. J., Y. H. Kim, and H. Y. Chun, 2001: Dry and moist con- precipitation in high latitudes. J. Hydrometeor., 5, 409–429. vection forced by an urban heat island. J. Appl. Meteor., 40, Nesbitt, S. W., E. J. Zipser, and D. J. Cecil, 2000: A census of 1462–1475. precipitation features in the tropics using TRMM: Radar, ice Bentley, M. L., and J. A. Stallins, 2005: Climatology of cloud-to- scattering, and lightning observations. J. Climate, 13, 4087– ground lightning in Georgia, USA, 1992–2003. Int. J. Clima- 4106. tol., 25, 1979–1996. Ogden, F. L., H. O. Sharif, S. U. S. Senarath, J. A. Smith, M. L. Bornstein, R., and Q. Lin, 2000: Urban heat islands and summer- Baeck, and J. R. Richardson, 2000: Hydrologic analysis of the time convective thunderstorms in Atlanta: Three case stud- Fort Collins, Colorado, flash flood of 1997. J. Hydrol., 228, ies. Atmos. Environ., 34, 507–516. 82–100.

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC FEBRUARY 2007 NTELEKOS ET AL. 101

Orville, R. E., and A. C. Silver, 1997: Lightning ground flash den- ——, M. L. Baeck, J. E. Morrison, P. Sturdevant-Rees, D. F. sity in the contiguous United States: 1992–95. Mon. Wea. Turner-Gillespie, and P. D. Bates, 2002: The regional hydrol- Rev., 125, 631–638. ogy of extreme floods in an urbanizing drainage. J. Hydro- ——, and G. R. Huffines, 2001: Cloud-to-ground lightning in the meteor., 3, 267–282. United States: NLDN results in the first decade, 1989–98. ——, ——, K. L. Meierdiercks, P. A. Nelson, A. J. Miller, and Mon. Wea. Rev., 129, 1179–1193. E. J. Holland, 2005a: Field studies of the storm event hydro- ——, G. Huffines, and J. Nielsen-Gammon, 2001: Enhancement logic response in an urbanizing watershed. Water Resour. of cloud-to-ground lightning over Houston, Texas. Geophys. Res., 41, W10413, doi:10.1029/2004WR003712. Res. Lett., 28, 2597–2600. ——, A. J. Miller, M. L. Baeck, P. A. Nelson, G. T. Fisher, and ——, G. R. Huffines, W. R. Burrows, R. L. Holle, and K. L. Cum- K. L. Meierdiercks, 2005b: Extraordinary flood response of a mins, 2002: The North American Lightning Detection Net- small urban watershed to short-duration convective rainfall. work (NALDN)—First results: 1998–2000. Mon. Wea. Rev., J. Hydrometeor., 6, 599–617. 130, 2098–2109. ——, M. L. Baeck, K. L. Meierdiercks, A. J. Miller, and W. F. Peterson, T. P., 2003: Assessment of urban versus rural in situ Krajewski, 2007: Radar rainfall estimation for flash flood temperatures in the contiguous United States: No difference forecasting in small urban watersheds. Adv. Water Resour., in found. J. Climate, 16, 2941–2959. press. Rozoff, C. M., W. R. Cotton, and J. O. Adegoke, 2003: Simulation of St. Louis, Missouri, land use impacts on thunderstorms. J. Toracinta, E. R., D. J. Cecil, E. J. Zipser, and S. W. Nesbitt, 2002: Appl. Meteor., 42, 716–738. Radar, passive microwave, and lightning characteristics of Scott, D. W., 1992: Multivariate Density Estimation: Theory, Prac- precipitating systems in the Tropics. Mon. Wea. Rev., 130, tice, and Visualization. John Wiley and Sons, 317 pp. 802–824. Shepherd, J. M., 2005: A review of the current investigations Underwood, S. J., and M. D. Schultz, 2004: Patterns of cloud-to- of urban-induced rainfall and recommendations for the ground lightning and convective rainfall associated with future. Earth Interactions, 9. [Available online at http:// postwildfire flash floods and debris flows in complex terrain EarthInteractions.org.] of the western United States. J. Hydrometeor., 5, 989–1003. ——, and S. J. Burian, 2003: Detection of urban-induced rainfall Van den Heever, S. C., and W. R. Cotton, 2005: Modeling the anomalies in a major coastal city. Earth Interactions, 7. impacts of urban aerosol on convection and precipitation. [Available online at http://EarthInteractions.org.] Preprints, 16th Conf. on Planned and Inadvertent Weather Smith, J. A., and A. F. Karr, 1985: Statistical inference for point Modification, San Diego, CA, Amer. Meteor. Soc., CD- process models of rainfall. Water Resour. Res., 21, 73–79. ROM, 2.3.

Unauthenticated | Downloaded 09/29/21 09:15 AM UTC