Nowcasting Thunderstorms: a Status Report

Nowcasting Thunderstorms: A Status Report

James W. Wilson, N. Andrew Crook, Cynthia K. Mueller, Juanzhen Sun, and Michael Dixon National Center for Atmospheric Research,* Boulder, Colorado

ABSTRACT

This paper reviews the status of forecasting convective precipitation for time periods less than a few hours (nowcasting). Techniques for nowcasting thunderstorm location were developed in the 1960s and 1970s by extrapolating radar echoes. The accuracy of these forecasts generally decreases very rapidly during the first 30 min because of the very short lifetime of individual convective cells. Fortunately more organized features like squall lines and supercells can be successfully extrapolated for longer time periods. Physical processes that dictate the initiation and dissipation of convective storms are not necessarily observable in the past history of a particular echo development; rather, they are often controlled by boundary layer convergence features, environmental vertical wind shear, and buoyancy. Thus, successful forecasts of storm initiation depend on accurate specification of the initial thermodynamic and kinematic fields with particular attention to convergence lines. For these reasons the ability to improve on simple extrapolation techniques had stagnated until the present national observational network modernization program. The ability to observe small-scale boundary layer convergence lines is now possible with operational Doppler radars and satellite imagery. In addition, it has been demonstrated that high-resolution wind retrievals can be obtained from single Doppler radar. Two methods are presently under development for using these modern datasets to forecast thunderstorm evolution: knowledge- based expert systems and numerical forecasting models that are initialized with radar data. Both these methods are very promising and progressing rapidly. Operational tests of expert systems are presently taking place in the United King- dom and in the United States.

1. Introduction The primary tools for detecting convective storms are weather radar, lightning detectors, and satellite imag- This paper will review the history and status of ery. Very short period forecasting of the future loca- forecasting thunderstorms for very short time periods tion of convective storms has historically been based (nowcasting). The term nowcasting is used to empha- primarily on the extrapolation of radar reflectivity ech- size that the forecasts are time and space specific for oes. As will be discussed later the majority of indi- periods less than a few hours. Forecasts of this type vidual thunderstorms have lifetimes less than ~20 min, are particularly important to commercial and general thus forecast techniques based on the extrapolation of aviation, outdoor sporting events, the construction existing conditions are limited. For forecast periods industry, power utilities, and ground transportation. beyond ~20 min, techniques for forecasting the initiation, growth, and dissipation of convective storms are essential. Numerical simulation studies have contrib- *The National Center for Atmospheric Research is partially spon- uted significantly to our understanding of storm orga- sored by the National Science Foundation. nization and lifetime. This understanding is just Corresponding author address: Dr. James W. Wilson, Atmo- beginning to be used in modern nowcasting systems. spheric Technology Division, NCAR, P.O. Box 3000, Boulder, Two methods are presently under development for CO 80307-3000. E-mail: [email protected] forecasting storm evolution: knowledge-based expert In final form 14 May 1998. systems and explicit numerical forecast models that ©1998 American Meteorological Society are initialized with radar data. Section 2 discusses the

Bulletin of the American Meteorological Society 2079 lifetime, organization, and motion of convective and 47% lived longer than 1 h (see the right panel storms from both a numerical and observational view- of Fig. 1). point. Section 3 discusses forecasting techniques that Battan (1959) discussed how the individual ech- include extrapolation of existing storms, forecasting oes in a squall line generally moved to the left of the storm initiation and dissipation, and numerical predic- movement of the squall line itself. While the individual tion. Operational forecasting systems are described in echoes had lifetimes of tens of minutes, the squall line section 4 and the accuracy of the various techniques itself might have a lifetime of many hours. Newton and is discussed in section 5. Possible future directions are Fankhauser (1964) showed that echo motion varied contained in section 6. with storm size. In agreement with the above observational studies Wilson (1966) showed that lifetime and echo motion was dependent on the scale of the 2. Lifetime, organization, and motion of convective phenomena. Figure 2 from this study shows convective storms that well-organized systems (supercells and cyclones) have long lifetimes compared to unorganized thunder- Prior to describing nowcasting techniques and sys- storms and weak showers. Figure 3 shows that lifetime tems it is useful to review what is known from numeri- is highly related to scale size and that different scales cal and observational studies about the organization, have different motions. In this study small-scale fea- lifetime, and motion of convective storms. tures tended to move with the mean wind and larger A number of observational studies have shown that scales tended to move slower and to the right of the individual convective cells have mean lifetimes of mean wind. Thus when extrapolating echoes it is es- about 20 min. Battan (1953) and Foote and Mohr sential to consider the scale of the convective feature (1979) found mean durations of 23 and 21 min, re- and the length of the forecast period. For forecast pe- spectively. However, Battan reported that cells that riods greater than 20 min, extrapolation of individual merged with one another had longer durations. This cells within a multicellular system or alone is gener- agrees with Henry’s (1993) results from the High ally not reliable. However, large thunderstorms, super- Plains regions near Denver, Colorado. The left panel cells, storm complexes (mesoscale convective of Fig. 1 shows that 83% of the individual storm systems), and large frontal rainbands (Hill et al. 1977) cells lived < 30 min. However, if storms merged or can often be extrapolated for much longer time periods. split during their lifetime (i.e., were multicellular) only There is a significant body of knowledge based on 12% of the storm complexes existed for < 30 min numerical simulations that have a direct bearing on our physical understanding of storm type, organization, and lifetime. The group of Simple Tracks Complex Tracks papers of Thorpe et al. (1982), Weisman and Klemp (1986), Weisman et al. 80 568 complete simple 80 68 complete complex storm tracks storm tracks (1988), and Rotunno et al. (1988) dis- 83% lived < 30 minutes cusses how wind shear profiles and buoy- 12% lived < 30 minutes 60 60 ancy considerations can be used to estimate storm type (single cell, multicell, and supercell), storm initiation, and 40 40 storm longevity. Weisman and Klemp Percent Percent (1986) show the bulk Richardson num-

20 20 ber, which combines convective available potential energy (CAPE) and the surface to 6 km wind shear, can be used to 0 0 0.0 1.0 2.0 3.0 4.0 0.0 1.0 2.0 3.0 4.0 differentiate storm type, organization, and Duration (hrs.) Duration (hrs.) lifetime. Values of the bulk Richardson number between 10 and 40 favor super- FIG. 1. Histogram showing the lifetime of simple and complex storms ob- cells, while increasingly higher numbers served during the summer of 1991 near Denver, CO, based on data from an automated cell tracking system called TITAN. A simple storm is one that does not favor unsteady multicellular storms. merge or split during its lifetime and a complex storm is one that does (from CAPE and low-level shear have gained Henry 1993). wide acceptance by the forecaster com-

2080 Vol. 79, No. 10, October 1998 1.0 ) 1.0

) a) max max 0.8 > 40 mi

0.8 Supercell > 20 mi 0.6 20 - 40 mi > 10 mi Ea 0.6 st C oa st Cy 0.4 cl one 10 - 20 mi

original

L 5 - 10 mi i n e 0.2 0.4 o fWe U no East Coast Cyclone ak S rgan

how ized er T 0 40 80 120 160 s hunderstorms

0.2 Max cross-correlation coefficient (r Time (min)

Maximum cross-correlation coefficient (r 20 b) 0 2060 100 140 Time (min)

IG F . 2. Magnitude of the maximum cross correlation (rmax) between radar echo patterns as a function of time for different types 10 of storms (from Wilson 1966). Here rmax is a measure of echo lifetime and was obtained by cross-correlating echo patterns by the indicated time period. 5 - 10 mi munity as a means to identify days with potential for 0 > 20 mi

severe weather. North & South (knots) Thorpe et al. (1982) and Rotunno et al. (1988) show that the low-level vertical wind shear profile (surface to 2.5 km) directed normal to the gust front ∆ -10 ( u) is related to the extent and longevity of thunder- 0 5 10 15 storms. They find that long-lived, intense convection East (knots) results when an optimal shear occurs. The Rotunno IG et al. (1988) simulations show this optimal condition F . 3.(a) Magnitude of the maximum cross correlation (rmax) occurs when the import of positive vorticity associated between radar echo patterns as a function time for different spa- with the low-level shear just balances the net buoyant tial scales within a storm system. The data are for the unorganized thunderstorm case in Fig. 2. (b) Average motion for the 5–10 and generation of negative vorticity by the cold pool. This > 20-mile scales shown in (a) (from Wilson 1966). allows deep, vertical updrafts at the gust front. The depth of the updrafts are shown to increase considerably as ∆u increases from 0 to 20 m s−1. Weisman and convergence resulting in the storm’s rapid decay. Klemp (1986) indicate that both simulations and ob- Numerical simulations by Moncrieff and Miller (1976) servational studies show that when ∆u is > 25 m s−1 find that a steady long-lived squall line requires the and directed obliquely to the convergence line, long- propagation speeds of the density current and the cu- lived cells with supercell characteristics result. In low mulonimbus to be equal. wind shear situations the gust front tends to move Direct utilization of the above stability and verti- away from the storms, resulting in the demise of the cal wind shear concepts for nowcasting requires mea- storm updraft as it moves away from the near-surface suring them with sufficient time and space resolution

Bulletin of the American Meteorological Society 2081 to define variations that are significant to nowcasting The tracking of individual cells began in the early storm evolution. As discussed in section 3b(2) the typi- 1970s by scientists at the National Severe Storms cal radiosonde and surface network is not sufficient for Laboratory (NSSL). They developed techniques for this purpose. However, there have been some recent identifying convective cells from radar reflectivity advances utilizing radar, satellite, and special data and tracked and extrapolated the cell centroid mesonetworks that will help to alleviate this measure- (Wilk and Gray 1970; Barclay and Wilk 1970; Zittel ment problem and are now being integrated into the 1976). Scientists at the Stanford Research Institute most recent nowcasting systems. refined these techniques in an attempt to handle the splitting and merging of echoes (Duda and Blackmer 1972; Blackmer et al. 1973). Wolf et al. (1977) and 3. Nowcasting techniques Endlich and Wolf (1981) used the Stanford Research Institute techniques to extrapolate clouds observed by Discussion of thunderstorm nowcasting techniques satellite. Dixon and Wiener (1993) added geometric is divided into three sections. The first (extrapolation) algorithms to handle the splitting and merging of is a historical treatment of thunderstorm extrapolation storms and developed a robust real-time cell tracking techniques first assuming no change in motion, size, and analysis system with the ability to grow or dissi- and intensity (steady-state assumption) and second pate detected cells based on their past trends. An ex- allowing for changes in size and intensity based on ample of the output from this algorithm called TITAN past trends. The second section (convection initiation/ (Thunderstorm Identification, Tracking, Analysis and dissipation) discusses recent research results that have Nowcasting) is shown in Fig. 4. provided hope that the initiation and dissipation of A centroid-type tracker developed and tested by thunderstorms can be forecast. The third section (nu- Bjerkaas and Forsyth (1979) was the original Next merical prediction) discusses the explicit numerical Generation Radar (NEXRAD) algorithm used to show prediction of thunderstorms with and without the use echo motion. This algorithm has been replaced by of radar to initialize the model. another centroid-type tracker, called Storm Cell Iden- tification and Tracking (SCIT) that shows the motion a. Extrapolation of multiple reflectivity maxima in a cell complex (Witt 1) STEADY-STATE ASSUMPTION and Johnson 1993). The notion of extrapolating radar echoes for the Goodman (1990) presents a methodology for clus- purpose of forecasting precipitation was started by tering and extrapolating lightning data based on a cen- Ligda (1953) before the availability of computers. troid tracker. He found that lightning data, when fused Noel and Fleisher (1960) and Hilst and Russo (1960) with radar, provided improved thunderstorm identifi- were the first to use computers to obtain echo motion. cation and continuity. The relationship between light- They correlated two digitized radar images taken at ning and storm characteristics and life cycle has been different times and used the location of the maximum explored by numerous investigators (Shackford 1960; cross-correlation value as a best estimate of the echo Livingston and Krider 1978; Williams 1985; Cherna pattern’s average motion. The echo field was then and Stansbery 1986; Goodman et al. 1988; Williams extrapolated with this vector without change in size et al. 1989). or intensity (steady-state assumption). Following these pioneering efforts a series of papers from the Travel- 2) INTENSITY AND SIZE TRENDING ers Research Center (Kessler and Russo 1963; Kessler Tsonis and Austin (1981) investigated the use of 1966; Wilson 1966) examined the predictability of trends in echo size and intensity to improve forecasts precipitation fields based on statistics of the echo pat- of cells that had already lived at least 30 min but found tern. The very perishable nature of small-scale features negligible improvement of skill even in elaborate non- was documented and predictability and echo motion linear time-trending schemes. We have examined this were related to scale size. Rinehart (1981) was the first further by using the TITAN echo extrapolation algo- to use cross-correlation techniques to obtain differen- rithm to compare forecasts with and without the in- tial motions within reflectivity fields instead of using clusion of echo size and intensity trending for forecast a single average vector for the entire precipitation pat- periods ranging between 6 and 36 min. These forecasts tern. This techniques was later used by nowcasting were made for 13 thunderstorm days from Colorado systems to be discussed in section 3a(2). during the summer of 1995. The results are shown in

2082 Vol. 79, No. 10, October 1998 FIG. 4. Sample output from the TITAN algorithm. The blue ellipses represent the 3D detection of the storms and the yellow ellipses represent successive 6-min forecasts of position and size. Based on trend history the cell to the north is forecast to increase in size and the cell to the south to decrease. The displays on the right side represent time histories for operator-selected storms of volume size, maximum reflectivity height profile, and histogram of reflectivity.

Fig. 5. As expected the forecast accuracy decreases these physical processes are often events occurring in with forecast length. However, while the trending tech- the boundary layer, such as convergence (Garstang and nique provides slightly higher probability of detection Cooper 1981). (POD), the false alarm ratio (FAR) values are also higher, resulting in nearly identical critical success b. Convection initiation/dissipation index (CSI) values for each technique (Donaldson Meteorological understanding of thunderstorm ini- et al. 1975). Tests were made using a variety of para- tiation has previously been insufficient to make pos- bolic curves to extrapolate the echo size (e.g., increase sible short-period detailed forecasts of thunderstorm echo size for 15 min then decrease); however, all these initiation (Carbone et al. 1990). However, in recent gave essentially identical results regardless of the years there have been significant breakthroughs for curve shape. This leads us to conclude, as Tsonis and those instances where boundary layer forcing is a Austin (1981) did, that essential physical processes “missing link” between an environment that seems that dictate the change in rainfall with time are not nec- ripe for convection and the actual development of a essarily observable in the past history of a particular thunderstorm at a specific location. Some aspects of echo development. In the case of convective storms this new insight had their origin in the pioneering sat-

Bulletin of the American Meteorological Society 2083 1.0 a) Trending 1450 18 May - 31 Jun No Trending 1 Jul - 15 Aug 0.8

2075

0.6 1950

0.4

POD

0.2 FAR

1575 0 6 1218243036 Forecast period (min)

FIG. 5. Comparison of POD and FAR statistics as a function 1700 of forecast period of extrapolation forecasts with and without trending echo size and intensity. Based on 13 thunderstorm days from Colorado during 1995 using the TITAN algorithm. 1825 ellite analyses of Purdom (1976) and collaborators. 1950 The development of Doppler weather radars and their ability to observe the clear-air boundary layer extended 0 25 km this insight. b) Prior to these observational advances, Byers and 60 Braham (1949) found low-level mesoscale conver- 50 Moving gence at the surface up to 30 min before the appear- Boundary ance of a radar precipitation echo. Ulanski and 40 Garstang (1978), Garstang and Cooper (1981), Behind Achtemeier (1983), and Watson and Blanchard (1984) 30 Boundary showed that a correlation existed between surface con- 20 vergence, as measured by mesonetworks of anemom- 10 eters, and the development of thunderstorms. They Number of Storms found lead times of 15–90 min between the onset of convergence and the onset of convective rainfall at the -50 -40 -30 -20 -10 0 10 20 30 40 50 surface. Purdom (1973, 1976, 1982) showed that con- Distance From Boundary (km) vergence lines were often visible on satellite imagery as lines of cumulus that could be observed to interact 50 50 with one another and produce thunderstorms. Wilson Stationary Colliding Boundary Boundary and Carbone (1984) proposed the use of sensitive 40 40 Doppler radars to monitor convergence lines even in the absence of clouds. Wilson and Schreiber (1986) 30 30 20 20

FIG. 6. (a) Storm initiation locations during the summer of 1984 10 10

along the Colorado Front Range. The contours are ground eleva- Number of Storms No Boundary tion in meters. (b) Frequency of storm initiation location relative to moving boundaries, stationary boundaries, and colliding bound- 0 10203040 01020 aries. The shading represents those cases classified as boundary Distance From Distance From initiated (from Wilson and Schreiber 1986). Boundary (km) Collision (km)

2084 Vol. 79, No. 10, October 1998 showed that 80% of the thunderstorms in the Denver area formed along radar-detected convergence lines. TABLE 1. Comparison of human, persistence, and extrapola- > Figure 6 shows two figures from their study. The first tion 30-min forecasts of radar reflectivity 40 dBZ for the summer of 1990 near Denver, CO. The forecasts were made for an indicates that the geographical location of storm ini- 8000 km2 area with a 5 km × 5 km grid spacing. The human fore- tiation appears random; however, the second shows caster attempted to forecast storm initiation, extrapolation, and dis- that when these same data are plotted relative to con- sipation. Persistence forecasts assumed the reflectivity field at vergence lines the location of initiation is very deter- forecast issue time would not move or change. The extrapolation ministic. These findings have lead to optimism that technique was based on that of Chornoby (1992) (from Wilson and Mueller 1993). very short-range forecasts of thunderstorm initiation and evolution are possible. It has been demonstrated (Mueller and Wilson Human Persistence Extrapolation 1989; Wilson and Mueller 1993) that forecasters could often anticipate thunderstorm initiation by monitor- POD 0.55 0.11 0.15 FAR 0.85 0.84 0.75 ing Doppler radar-detected boundary layer convergence lines (boundaries) together with visual monitoring of cloud development in the vicinity of the convergence line. Forecast experiments described in schematic example of this phenomenon from Atkins the above studies showed that human forecasters could et al. (1995). For very short period forecasting pur- do better than persistence or extrapolation forecasts poses it is necessary to know how fast a thunderstorm because of the ability to nowcast storm initiation and can develop. Knight et al. (1983) have shown from dissipation (see Table 1). However, forecasters often visual observations and model simulations that a cloud had difficulty in precisely timing and placing the loca- can grow, by the ice process, from first cloud to tion of storm initiation and nowcasting the evolution 30-dBZ echo in roughly 15 min. Henry and Wilson of existing storms. In addition, not all convergence (1993) found that it took between 8 and 16 min for lines initiated storms, even when they collided in ap- echoes to grow from 10 to 40 dBZ in the presence of parently conditionally unstable environments (e.g., convergence lines. However in nearly every instance Stensrud and Maddox 1988). Three reasons were ad- there was cloud present 30 min prior to the rapid in- vanced for these difficulties: 1) there is a basic defi- tensification. This agrees with the impressions of fore- ciency in our knowledge of the details of storm casters involved in the experiments described by initiation and evolution, 2) there is a need for detailed Wilson and Mueller (1993) that a thunderstorm would observation of boundary layer thermodynamics not initiate in 30 min unless cumulus clouds were al- and more detailed observation of cumulus cloud lo- ready present near the convergence line (in front or cation and growth, and 3) many of the forecaster ac- behind). An exception is likely when deep, intense tivities were manually intensive and prone to error. convergence is present. Considerable progress has recently been made in all As discussed in section 2, numerical simulations three of these areas. Each is discussed separately in the following sections. 12 August 1991 Cloud Development along the Sea-Breeze Front During an 1) IMPROVING BASIC UNDERSTANDING Off-Shore Flow Regime Some new understanding has recently emerged from the Convection Initiation and Downburst Experi- Coast - Line ment, which took place in Colorado in 1987 (Wilson et al. 1988), and the Convection and Precipitation– H Electrification Experiment, which took place in 1991 C Vector + R in Florida (Wakimoto and Lew 1993). Several stud- Shear ies have shown that cloud and storm initiation is fa- S vored at locations where horizontal convective rolls intersect convergence lines or preexisting cumulus FIG. 7. Schematic diagram showing the intersection between clouds (Kessinger and Mueller 1991; Crook et al. the sea breeze front and horizontal convective rolls and how it re- 1991; Wilson et al. 1992; Wakimoto and Atkins 1994; lates to cloud development. The sea breeze is delineated by the Atkins et al. 1995; Kingsmill 1995). Figure 7 shows a heavy, barbed line (from Atkins et al. 1995).

Bulletin of the American Meteorological Society 2085 show that storm initiation, organization, and lifetime 2) NEED TO IMPROVE STABILITY OBSERVATIONS is related to a balance between the horizontal vortic- Crook (1996) showed from numerical simulations ity on either side of convergence lines. The parameter that variations of 1 g kg−1 in boundary layer moisture ∆u (low-level shear directed normal to the conver- or 1°C in temperature were critical to whether storms gence line) is used to estimate this balance; values formed or not. Weckwerth et al. (1996) showed that near 20 m s−1 are considered optimum. Moncrieff and at least this magnitude of variability occurred routinely Miller (1976), Weisman and Klemp (1986), and in the convective mixed boundary layer over a distance Wilson and Megenhardt (1997) also show that initia- of only a few kilometers. This variability was a result tion, organization, and lifetime is dependent on the of horizontal convective rolls transporting dry air relative motion between the cumulus clouds and con- downward in the downdraft portion of the roll and vergence line. The boundary relative cell speed and bringing up moist air in the updraft portion (see ∆u are highly correlated. Figure 8 contrasts favorable Fig. 10). These studies showed that unless a sounding and unfavorable conditions for initiation and suste- happened to be taken in the updraft portion of a con- nance of convection. In the favorable situation the vective roll the potential for convection could be sig- wind directed normal to the gust front will increase nificantly underestimated. Since observing moisture with height resulting in a large ∆u and deep vertical fields with this sort of horizontal resolution is not yet updrafts. Also the storms would move at a speed simi- practical, Mueller et al. (1993) reasoned that monitor- lar to the gust front since the storm steering level ing cumulus cloud location and growth could often winds (~3-km wind) are similar to the gust front mo- provide a rough estimate of stability. Roberts (1996) tion. In the unfavorable situation the storms will move is developing automated techniques that utilize both in a direction opposite to the gust front and the wind satellite imagery and Doppler radar data to detect the directed normal to the gust front will increase with location of cumulus and monitor their growth from the height giving a negative ∆u and shallow tilted up- cumulus humilis stage to the cumulus congestus stage. drafts. Both Wilson and Megenhardt (1997) and Mueller et al. (1997a) show examples of storm dissi- 3) NEED TO AUTOMATE PARTS OF THE NOWCASTING pation as the cells move away from the convergence PROCESS line (see Fig. 9). Considerable progress has been made in auto- mating parts of the nowcasting process by Lincoln Laboratory, NSSL, and the National Center for Atmo- spheric Research (NCAR) partially under the support FAVORABLE FOR STORMS of the Federal Aviation Administration (FAA). For GL ) terminal and enroute aviation nowcasts the entire pro- 3 A cedure is being automated. These automated forecast 2 systems are discussed in the following section. When a forecaster is part of the nowcasting effort, such as 1 at National Weather Service Forecast Offices or the Kennedy Space Center, it is expected that the human HEIGHT ( km adds to the quality of the nowcast. The rationale for a UNFAVORABLE FOR STORMS human–computer nowcasting system was advanced GL ) by Browning (1980) and Wilson and Carbone (1984). 3 A Activities that are routine and/or prone to human er- 2 ror are best done by computers. In addition, the computers are needed to drive interactive displays where 1 forecasters can easily overlay and animate data on a common grid. Examples of such display systems are HEIGHT ( km given by Corbet et al. (1994), Johnson et al. (1995), FIG. 8. Illustration of conditions that are favorable and not fa- and Roberts et al. (1996). This allows the forecaster vorable for storm development. The wind vectors on the right side more time to use his/her physical reasoning and pat- represent the environmental wind profile, which is the same for each case. The dark shading represents a density current. The ar- tern recognition capabilities to assess data quality, row in the density current represents its motion. The other arrow evaluate automated forecast material, and apply broad represents the updraft tilt. meteorological understanding to the nowcasts.

2086 Vol. 79, No. 10, October 1998 c. Numerical prediction a)km b) km In this section we will briefly 20.0 20.0 describe the current status of 15.0 15.0 explicit numerical prediction of 10.0 10.0 thunderstorms. Numerical mod- 5.0 5.0 els have the ability to simulate 0.0 0.0 from the boundary

all the phases of thunderstorm Distance of the storm -5.0 -5.0 evolution described in the previ- 20:00 21:00 22:00 23:00 00:00 20:00 21:00 22:00 23:00 00:00 ous two sections; however, a significant amount of progress is c)km3 d) km3 required before explicit nu- 2000 2000 merical predictions show consis- 1600 1600 tent skill. For the sake of brevity, 1200 1200 and unless otherwise stated, we do not include forecasts of 800 800 Storm Volume larger-scale precipitation events 400 400 in which convection is generally 0 0 parameterized. The current field 20:00 21:00 22:00 23:00 00:00 20:00 21:00 22:00 23:00 00:00 Time Time of numerical thunderstorm prediction can essentially be FIG. 9. Time series of storm distance from the convergence line and corresponding storm volume. Panels (a) and (c) show storms that lived > 24 min. Panels (b) and (d) show storms divided into two categories, de- that lived < 24 min. Panels (a) and (b) are time series plots of storm distance from the bound- pending on whether or not ex- ary. Panels (c) and (d) are the storm volume. The different line types represent different plicit information about the storms. The vertical dashed lines indicate the transition from growth to dissipation of the convection is used in the initial individual long-lived storms. Note that cell volumes decrease as they move away from the conditions. boundary.

1) MODEL INITIALIZATION WITHOUT RADAR DATA This category can be considered as the extension relative cloud bases and depths of current NWP models down to resolutions of a few predicted from measurements actual cloud kilometers. Since these models currently do not in- directly beneath them clude storm data in the initial conditions, if storms are to be predicted by the model, they must be generated during the forecast period. A number of simulations in this category have been performed over the last decade, primarily in a research mode. Although the following list should not min in anyway be considered exhaustive, some examples q q in this category include the forecast of a squall line during SESAME using the Geophysical Fluid Dynam- ics Laboratory nonhydrostatic cloud model (Hemler et al. 1991); predictions of thunderstorms and torna- max q does along the dryline using RAMS (Cotton and Grasso 1996), MM5 forecasts of a squall line during TAMEX (Kuo and Wang, 1996) and PreSTORM FIG. 10. Schematic diagram showing the vertical distribution (Zhang et al. 1989), and forecasts of the 7 May 1995 of moisture in the mixed boundary layer containing horizontal roll squall line using the Advanced Regional Prediction circulations. Gray lines indicate roll circulations. The thick black lines are contours of moisture. Observed cloud-base heights are System (ARPS) at the Center for Analysis and Pre- shown by the solid curve. Dashed clouds represent relative cloud diction of Storms (CAPS) (Wang et al. 1996). Figure bases and depths expected if stability parameters are estimated 11 shows a 9-h forecast of the 7 May event along with from the moisture values directly beneath those clouds (from the verifying reflectivity. In realtime, CAPS has per- Weckwerth et al. 1996). formed explicit thunderstorm predictions at horizon-

Bulletin of the American Meteorological Society 2087 boundary specification of land use, soil moisture, evapotranspi- ration, and terrain. Clearly, if the initiation phase is not forecast, because of the sensitivities discussed above, predictions will be limited to the propagation and decay phases of the convection. Forecasting these stages will require the utilization of radar data, which is discussed in the next section.

2) MODEL INITIALIZATION WITH RADAR DATA FIG. 11. Reflectivity from an ARPS 9-h forecast on a 6-km grid (left) and corresponding As a first step toward in- radar observations at 0000 UTC on 8 May 1995. Simulation performed at CAPS; for fur- cluding explicit storm data, a ther details see Wang et al. (1996). number of researchers have examined methods to assimilate tal resolutions down to 3 km over the last few years surface rainfall observations into numerical models. starting with the VORTEX field program in 1995 In most of these simulations the convection has been (Droegemeier 1997). parameterized and the hydrostatic assumption made. On the positive side, these simulations are fairly These studies include (although again are not limited easy to perform since they only require an increase to) those by Fiorino and Warner (1981), Molinari in the resolution of current NWP models. However, the (1982), Wang and Warner (1988), Krishnamurti et al. main problems are associated with the fact that storm (1988), Puri and Miller (1990), and Carr and Baldwin data are not included in the initial conditions. Low- (1991). In general terms, these studies either adjust the level convergence in these simulations takes some initial conditions of moisture, divergence, heating time to spin up from the large-scale circulation and rates, etc., to match the observed rainfall or nudge the hence the models are not generally reliable for the first latent heating and moisture fields in a preforecast pe- 6 h or so. When storms do develop, there are often sig- riod. Recently, four-dimensional variational tech- nificant errors in timing and location that cannot be niques have been used to assimilate rainfall data into corrected unless new storm data are incorporated. numerical models (Zupanski and Mesinger 1995; Zou The cases where simulations in this category have and Kuo 1996). Operational centers are now starting shown some success are primarily convective events, to include rainfall data in the assimilation cycles of which are strongly forced by the large-scale environ- mesoscale forecast models (Lin et al. 1996; Jones and ment. This suggests that there is some reason for op- Macpherson 1997). In all of these simulations, the two- timism that skillful thunderstorm predictions will be dimensional field of rainfall data is converted into possible in cases where the large-scale forcing is three-dimensional fields of moisture, divergence, la- strong. The outlook, however, is less clear in cases of tent heating, etc., by making a number of implicit as- weak forcing. Recent results have suggested that fore- sumptions about the structure of convection. casts of convective initiation in weak forcing can be Simulations in which the explicit structure of thun- very sensitive to variations in low-level temperature derstorms is initialized have only been attempted in and moisture (Brooks et al. 1993; Crook 1996). This the last few years. The first published experiment was in turn suggests that accurate forecasts of convection a 15-min prediction of the Del City storm by Lin et al. initiation will require initial conditions for the dynami- (1993). Other experiments include a 30-min prediction cal, thermodynamical, and microphysical fields, which of the Arcadia, Oklahoma, supercell (Weygandt et al. in some cases are beyond our present observing capa- 1998) and predictions of a Florida multicell storm by bilities. These sensitivities also place strong require- Lazarus (1996). We have recently performed a 60-min ments on the lateral boundary conditions and the lower forecast of the Buffalo Creek, Colorado, flash flood

2088 Vol. 79, No. 10, October 1998 case of 13 July 1996. Figure 12 Buffalo Creek, CO Flash Flood Event (July 13, 1996) is a plot of the reflectivity field for 24- and 48-min forecast periods and the verifying reflectivity for this case. The initial conditions for wind, thermodynamics, and microphysics was derived from the Denver, Colo- rado, WSR-88D using the retrieval technique of Sun and Crook (1997) that utilizes a cloud model and its adjoint. Verification analysis indicates that the numerical forecasts significantly improve over persistence and extrapolation in the 60-min time frame. Although these first experiments have shown some promise, there are a number of problems, both scientific and technical, that must be addressed before such forecasts show consistent skill. One of the main scientific problems is how to initialize the full dynamical, thermodynamical, and microphysical structure of the storm based on limited ob- FIG. 12. Numerical model forecast (right-hand panels) and verifying radar reflectivity servations (generally just (left-hand panels) of the Buffalo Creek Flash Flood that occurred in the Colorado moun- radar reflectivity and single tains. Forecast periods are t = 0, 24, and 48 min. The initial conditions for the model fore- Doppler velocity). This has casts are obtained from the adjoint method. Surface elevation is shown by the red contours prompted researchers to exam- and the watershed of Buffalo Creek is outlined in white. ine methods to retrieve the unob- served fields (cross-beam velocity, temperature, and it is difficult to maintain a storm once the initial pre- microphysical variables) from the available data. cipitation has fallen out. If the storm is forced from Retrieval techniques were first developed to estimate the boundary layer, then to maintain the storm it is the buoyancy from velocity fields obtained by dual- necessary to accurately observe and assimilate the Doppler analysis (Gal-Chen 1978; Hane and Scott boundary layer wind structure. Furthermore, because 1978). Since operationally, observations from two of the sensitivities described above it is also necessary Doppler radars are usually not available, a number of to accurately represent low-level temperature and researchers have developed techniques to retrieve the moisture structure. The persistence of a long-lived cross-beam velocity components from single Doppler storm clearly depends on the continual initiation of observations (see, e.g., Shapiro et al. 1995; Qiu and convection, thus raising the same questions about sen- Xu 1992; Sun and Crook 1994). Finally, methods have sitivities discussed in the previous section. Some of been developed to convert reflectivity observations the main technical issues include the computer re- into the microphysical fields carried by the model (see, sources needed for data assimilation and prediction at e.g., Rutledge and Hobbs 1983; Ziegler 1985; Sun and resolutions of around 1 km and the gathering and stor- Crook 1997). age of wideband data from multiple radars across the Another difficulty in explicit storm predictions is country. maintaining the storm’s strength. Our early results Finally, it should be noted that the future of storm- from these explicit storm forecasts have suggested that scale NWP lies in the successful combination of ap-

Bulletin of the American Meteorological Society 2089 proaches 1) and 2) described above. In other words, estimating radar echo motion (Collier 1991). Currently skillful numerical storm prediction will require that the Nowcasting and Initialization for Modeling using NWP models are run over large enough domains to Regional Observation Data system is being developed capture the large-scale forcing and high enough reso- and implemented to replace FRONTIERS. This sys- lution that the storm structure can be initialized from tem automates many of the labor-intensive operations radar data. Because of the timescale of most convec- in FRONTIERS such as data editing, extrapolation, tive cells (less than 20 min as shown in section 2) this and forecast preparation. This allows fully automated will require that the models be run with very rapid up- 30-min forecast cycles as well as forecaster input on date cycles. Experiments to test different data inser- hourly and three-hourly forecasts (Collier 1991). tion strategies for the 7 May 1995 squall line case have The U.S. National Weather Service is moving to- recently been performed by Xue et al. (1998). ward interactive tools that allow display and editing of a variety of data using the Weather Forecast Office- Advanced System (WFO Advanced) developed by the 4. Operational nowcasting systems Forecast Systems Laboratory (Roberts et al. 1996). In addition, the WDSS (Warning Decision Support In this section we discuss two types of operational System) developed by the NSSL is currently being nowcasting systems. The first are those based prima- tested at a few WFOs. The WDSS provides display and rily on the extrapolation of radar echoes while the sec- a suite of algorithms for calculating storm tracks, ond also includes forecasts of thunderstorm initiation detecting hail, detecting mesocyclones, tornadoes, and dissipation. The use of expert systems is included damaging winds, estimating precipitation accumula- in this second section. All the systems discussed tion, and evaluating near-storm environment (Johnson handle convective storms, although the Canadian and et al. 1995). In France the Approach Synthétique de U.K. systems discussed in section 4a frequently handle la Prevision Immediate au SMIRIC Project provides nonconvective systems. display capabilities for lightning, radar, and satellite data along with automated linear extrapolation tech- a. Extrapolation niques for lightning and radar data (Juvanon du Vachat The first automated operational nowcasting system and Cheze 1993). The METEOTREND project at the was implemented in 1976 utilizing the McGill Slovak Hydrometeorological Institute in Czechoslo- Weather Radar; products were sent to the Atmospheric vakia allows interactive analysis of satellite features Environment Service Forecast Centre, Quebec Region. and algorithm parameters used to produce a 2-h ex- McGill University scientists (Austin and Bellon 1974; trapolation forecast (Podhorsky 1987). The Japan Me- Bellon and Austin 1978; Bellon et al. 1980) adopted teorological Agency produces hourly forecast a version of the cross-correlation technique to forecast precipitation accumulation maps for up to 3 h in ad- precipitation amounts called Short-term Automatic vance based on radar, rain gauge, and satellite data Radar Prediction. A later version of this system called (Hirasawa 1991). RAINSAT (Austin and Bellon 1982; Austin et al. There is at least one fully automated system where 1990) was developed by McGill and implemented in the product goes directly to the user without forecaster both Canada and Spain (Nevado 1990). It used satel- intervention. That is the Integrated Terminal Weather lite and radar data and a cross-correlation scheme to System (ITWS) developed and tested by Massachu- make 1–6-h forecasts of rainfall. setts Institute of Technology (MIT)–Lincoln Labora- The U.K. Meteorological Office implemented tory for the FAA. It provides several aviation-specific Forecasting Rain Optimized using New Techniques of products including a 10- and 20-min forecast of the Interactively Enhanced Radar and Satellite data leading edge of thunderstorm activity based on cross- FRONTIERS in the early 1980s. This highly interac- correlation analysis (Evans and Ducot 1994). A more tive system provides 1–6-h forecasts of precipitation. thorough explanation and review of most of the op- Interactive capabilities allow the forecaster to a) edit erational system discussed here as well as other sys- spurious radar echoes, b) add orographic rain not de- tems can be found in Conway (1992). tected by radar, c) modify precipitation rates based on rain gauges, d) reregister satellite data, e) invoke al- b. Convection/dissipation gorithms to calculate precipitation amounts, and In the U.K. Meteorological Office a fully auto- f) prepare forecasts using one of several techniques for mated precipitation forecasting system is being devel-

2090 Vol. 79, No. 10, October 1998 oped that includes the forecasting of storm growth initialized with Doppler radar data, to forecast the (Hand and Conway 1995; Hand 1996). The system, movement and characteristics of boundary layer con- called Generating Advanced Nowcasts for Deploy- vergence lines. The rules used in the Auto-nowcaster ment in Operational Land Surface Flood Forecast are based primarily on the research findings discussed (GANDOLF), uses object-oriented convective storm in section 3b. nowcasting procedures. Radar data and Meteosat IR NSSL has been utilizing the NEXRAD SCIT al- satellite data are used to analyze convective cells in gorithm, which is a centroid-type extrapolator. They all stages of growth; subsequent movement and devel- are examining the predictive value of storm cell infor- opment up to 3 h ahead are predicted using a concep- mation, particularly storm rotation, for forecasting tual life cycle model combined with mesoscale NWP thunderstorm lifetime. data. Case study analyses show encouraging results. In another recent collaborative effort NCAR, Utilizing the recent advances in forecasting con- NSSL, and the National Weather Service (NWS) are vective precipitation discussed in section 3b, several involved in an operational experimental program at operational forecasting systems are under develop- National Weather Service Forecast Offices called Sys- ment and testing. The largest effort in the United States tem for Convection Analysis and Nowcasting. The is a collaborative effort involving MIT–Lincoln Labo- purpose is to automatically detect and analyze current ratory, NSSL, and NCAR under the sponsorship of the weather, and generate short-term forecasts and warn- FAA. An automated convective storm nowcasting ing guidance for NWS forecasters within the Ad- system is being developed for use in the airport ter- vanced Weather Interactive and Processing System minal area that forecasts storm growth and decay, as (AWIPS) environment. This effort involves the NCAR well as extrapolation (Mueller et al. 1997b). The fol- Auto-nowcaster discussed above, the NSSL WDSS lowing three paragraphs briefly describe the contribu- discussed in section 4a, and the NWS AWIPS Thun- tion of each group. derstorm Product discussed in Kitzmiller (1996) and MIT–Lincoln Laboratory is utilizing components Smith and Churma (1996). of their ITWS Microburst Prediction algorithm with The NCAR Auto-nowcaster is also being tested new capabilities to test the feasibility of short-term under U.S. Army Test and Evaluation Command sup- predictions of convection. Their prototype involves port at the White Sands Missile Range in New Mexico. machine-intelligent image processing and data fusion Many of the operations on a range are sensitive to techniques to determine cell growth and decay lightning and winds; therefore forecasts of convection (Wolfson et al. 1994), storm tracking to determine both are important both for the safety of personnel and the cell and storm envelope motion, probabilistic esti- efficiency of operations. A complete meteorological mates of cell lifetime based on short-term trends, and dataset is available at the range. Observations include estimates of the cumulus cloud amounts based on Geo- high-density surface observations, WSR-88D radar, stationary Operational Environmental Satellite data. boundary layer profilers, lightning network, and The algorithm produces a continuous forecast out to soundings. Extensions to other ranges are expected in 30 min by simultaneously solving five 6-min forecast the near future. problems. Lincoln Laboratory is also exploring the use of elliptical filters to separate radar echo features by scale size, with the purpose of extrapolating only those 5. Accuracy of nowcasting techniques scales that can be expected to live for the length of the forecast period. a. Extrapolation NCAR has been developing techniques for precise, Evaluation and comparison of the accuracy of short-period forecasts of thunderstorms initiation, nowcasts is very difficult. Statistics such as POD and movement, and dissipation. This knowledge-based FAR do not adequately represent performance. For expert system is called the Auto-nowcaster. The example, no credit is given for correctly forecasting a system utilizes radar, satellite, surface, and upper- nonevent or slightly missing a forecast in either time air weather observations. Algorithms identify and or space. However, these statistics are useful for com- extrapolate thunderstorms (centroid and cross- paring techniques that are evaluated precisely in the correlation extrapolators), detect and extrapolate con- same manner. Significantly different numbers can be vergence lines, retrieve boundary layer winds from obtained by just using different grid spacings and giv- single Doppler data, and use a numerical model, ing credit to forecasts that are near misses. It is with

Bulletin of the American Meteorological Society 2091 reluctance that we show statistics since they are often Doswell et al. 1990) and the Gilbert Skill Score (GSS; incorrectly interpreted by the casual user. Statistical see Schaefer 1990). The extrapolation techniques were measures should never be taken at face value unless clearly better than persistence; however, all the skill one understands precisely how the verification was done. scores showed there was little difference between the We use them here to compare techniques that are veri- two extrapolation techniques. It should be noted that fied in precisely the same manner; no attempt should the TITAN technique did extrapolate echo size and be made to compare statistics between experiments. intensity. However, as noted previously in section Elvander (1976) was one of the first to do a com- 3a(2) trending of echo size and intensity does not im- prehensive evaluation of different extrapolation tech- prove results. niques. He compared 1) the cross correlation, 2) echo A review of forecast quality from the above, centroid tracking, and 3) a combined individual echo coupled with comprehensive reviews by Bellon and and area motion tracking technique developed by Duda Austin (1978), Browning et al. (1982), and Collier and Blackmer (1972). He examined accuracies for (1989) indicate that the major cause of poor extrapo- forecast periods between 10 and 60 min for convec- lation forecasts is not due to errors in forecast displace- tive precipitation systems and found a sharp fall off ment, but to decay and growth occurring during the in accuracy with increasing time period. The simple forecast period. For individual convective storms, ex- cross-correlation technique gave slightly better results. trapolation forecasts deteriorate very rapidly, becom- Dixon and Wiener (1993), who evaluated the TITAN ing of little value for periods beyond 20–30 min. technique for forecast periods between 6 and 30 min, However, for supercell type storms and the general lo- also showed a sharp decrease in accuracy with increas- cation of squall lines or storm complexes, extrapola- ing forecast period. tion forecasts may be useful for periods up to several Henry et al. (1996) compared extrapolation and hours. For large-scale precipitation systems that are persistence forecasts for convective weather events primarily stratiform in nature, Browning et al. (1982) from Denver and Atlanta for time periods between 6 showed skill with echo extrapolation techniques out and 30 min. They grouped days into two categories to 6 h. depending on area of echo coverage. The forecasts for days with the larger echo coverage had considerably b. Initiation/dissipation better results. For Denver, extrapolation forecasts were Favorable results are just beginning to emerge with better than persistence forecasts for all time periods. techniques that attempt to forecast storm initiation and However, for Atlanta there was little difference be- dissipation. As previously shown in Table 1, Wilson tween the two techniques for all forecast periods. Brown and Brandes (1997) have made compari- sons between the Forecast Systems Laboratory cross- TABLE 2. Comparison of 30-min forecast made with a cross- correlation extrapolator (Jackson 1993; Jackson and correlation extrapolator Forecast Systems Laboratory (FSL) and Jesuroga 1995) and the TITAN storm motion type a storm cell extrapolator (TITAN). Persistence forecasts (PER) are extrapolator (Dixon and Wiener 1993). The evaluation also given. The evaluation is on a national scale from 11 convec- was on a national scale including observations from tive storm days. Verification skill measures are provided for POD, 11 days in 1994 and 1996. Cases were selected to rep- FAR, CSI, HSS, and GSS (from Brown and Brandes 1997). resent all seasons except winter with a heavy emphasis on convective-type weather situations. The Algorithm Kavouras mosaic of WSR-88D data was used as input data. The mosaic provided 2 km × 2 km grids of Statistic TITAN FSL PER five reflectivity levels at 5-min intervals. Thirty- minute forecasts were made for reflectivity levels POD 0.37 0.35 0.23 > 40 dBZ. The verification was carried out on a FAR 0.66 0.63 0.77 2 km × 2 km grid, a very demanding test, thus the scores are low. Forecasts of persistence were also in- CSI 0.22 0.22 0.13 cluded for comparison. Several verification measures were used to compare skills and are shown in Table 2. HSS 0.35 0.36 0.23 The familiar measures POD, FAR, and CSI are shown GSS 0.22 0.22 0.13 together with the Heidke Skill Score (HSS; see

2092 Vol. 79, No. 10, October 1998 and Mueller (1993) found for 30-min periods that human forecasts that included predictions of storm ini- TABLE 3. Comparison of POD and FAR statistics for 30-min tiation were better than persistence and extrapolation forecasts for four methods. “Extrapolation” uses the steady-state assumption. “Trending” also extrapolates size trends. “Rule 1” is forecasts. The initiation forecasts increased the POD the same as extrapolation except small storms are dissipated if they considerably over simple extrapolation while FAR are not near a boundary. remained nearly the same. Statistics are now being obtained from the NCAR Method POD FAR Auto-nowcaster and the Lincoln Laboratory growth and decay systems described in section 3b(2). Persistence 0.13 0.85 Experiments with the Auto-nowcaster have been made for 12 storm days in Colorado from 1995. A variety Extrapolation 0.27 0.59 of forecast rules were tested for forecast periods between 6 and 30 min. Comparison of POD and FAR Trending 0.33 0.70 values are shown in Table 3 for 30-min forecasts. Rule Rule 1 0.25 0.48 1 has the lowest FAR and trending has the highest POD. It was judged that the rule 1 technique provided the best results. This was because in comparison to extrapolation there was a 9% decrease in the area of de- reflectivity field at a height of 1 km at 2030 UTC and tection but a decrease of 42% in the area of false alarm. the overlaid 30-min forecast for echo greater than Rule 1 differed from the extrapolation forecast only 35 dBZ. The forecast and verification reflectivity field in that it dissipated TITAN extrapolated cells that is shown in Fig. 14b. For ease of discussion the fore- were < 400 km3 and were located outside of a zone ex- cast is divided into three regions labeled A, B, and C tending from 5 km in front of the convergence line in Fig. 14. Although this is only one case it nicely il- to 20 km behind. There are no statistics presented for lustrates some of the present strengths and weaknesses convection initiation forecasts because of difficulties of the system. Region A at issue time indicates a south- detecting cumulus with the radar. Recalling from sec- west–northeast–oriented line of strong storms. By tion 3c, storm initiation is not forecast unless cumu- valid time, the southwestern half of the line has dissi- lus clouds are detected near the convergence line. In pated and the northeastern half has maintained itself Colorado midlevel stratiform cloud is often generated by anvils from mountain storms; the automated tech- 24 July 1996 niques for detecting cumulus cloud near convergence 1.0 persistence lines were frequently misinterpreting this cloud as cu- trending mulus cloud and led to high false alarm rates. Thus Auto-nowcaster rules effort is under way to develop radar–satellite tech- 0.6 niques to identify cumulus (Roberts 1996). POD Recent experiments suggest that the difficulties in 0.2 Colorado of misinterpreting stratiform echo for cumulus echo may not be as serious in Memphis. For ex- 22:00 23:00 00:00 ample, statistics are presented in Fig. 13 for a Memphis 1.0 case that occurred on 24 July 1996. The results were obtained completely in an automated manner with the 0.6 Auto-nowcaster. They are for a 1.5-h time period as two convergence lines collide near the Memphis air- FAR persistence port. Figure 13a shows that the Auto-nowcaster rule 0.2 trending Auto-nowcaster rules 2 technique, which includes storm initiation, has a considerably higher POD than either persistence or ex- 22:00 23:00 00:00 trapolation forecasts. At the same time, the FAR Time of day (GMT) (Fig. 13b) for this technique (rule 2) is only slightly FIG. 13. Forecast statistics for a 1.5-h period on 24 July 1996 higher than for extrapolation forecast. near Memphis, TN, from the thunderstorm Auto-nowcaster. The Figure 14 shows an example forecast for one time top panel is POD and the bottom FAR. Forecast are for persis- period for this case. Figure 14a shows the initial tence, extrapolation, and storm initiation/dissipation.

Bulletin of the American Meteorological Society 2093 B is an area of strong growth and little movement because of the collision of the two convergence lines, which is handled fairly well by the convection initiation rules. Region C is an area of stratiform echo well behind the convergence lines. However, because of its large area it is classified as a large storm with a long lifetime. Techniques are presently under development to dissipate relatively large stratiform echoes that are well behind convergence lines. The accuracy of convective storm forecasting is summarized qualitatively in Fig. 15. It is assumed that the forecasts are being made on a spatial scale accurate to a few kilometers. Figure 15 is adapted from similar figures by Browning (1980), Doswell (1986), and Austin et al. (1987). Extrapolation refers to linear extrapolation techniques, the skill of which rapidly decreases to very low values after 30 min. Explicit storm models refer to small-scale (~1 km) nonhydrostatic numerical models initialized with radar data as discussed in section 3c(2). Expert systems refer to automated or manual techniques that depend heavily on current observations, particularly from remote sensors, and rules generated by humans based on meteorological research and experience. The Auto- nowcaster and GANDOLF represent such systems. Expert systems may also make use of numerical model output and utilize fuzzy logic techniques and neural networks. These techniques are shown to outperform all other techniques for the short period. Large-scale models are those typically run at meteorological centers like the National Centers for Environmental Prediction. Since these models do not initially contain information on storm locations, they have very little accuracy in the first few hours. These models are likely only to

FIG. 14. Example forecast from the thunderstorm Auto- be able to forecast the large-scale thunderstorms sys- nowcaster from Memphis, TN, on 24 July 1996. (a) The radar tems that are strongly forced by large-scale events. reflectivity field at forecast issue time, the automated detected con- The very rapid fall off in accuracy in Fig. 15 dur- vergence lines (yellow barbed lines), and the 30-min forecast ing the first 3 h represents the decreasing capability (heavy white lines) for > 35 dBZ echo (see dBZ scale). (b) The radar reflectivity field 30 min after (a) or the verification field. A of extrapolation techniques and our inadequate scien- perfect forecast would have all the > 35 dBZ echo within the white tific knowledge and lack of suitable observations to outlines. The letters A, B, and C and the associated thin black lines precisely forecast the initiation and dissipation of con- designate regions discussed in the text. vection. The relatively low accuracy at time periods beyond a few tens of minutes is a general acknowl- edgment that precisely forecasting the time and loca- and grown. The dissipation is the result of the conver- tion of convective storms is extremely difficult. Expert gence line moving away from the line of storms, which systems are shown to outperform other techniques is handled well by the Auto-nowcaster. The northeast- since they would incorporate forecasts from extrapo- ern half of the line is not handled as well because the lation techniques, large-scale models, mesoscale mod- automated convergence line detection systems did not els, explicit numerical models, statistical techniques, detect the convergence line that is critical for main- and the latest scientific knowledge on physical pro- taining small storms and initiating new storms. Region cesses that may not be adequately capture by the mod-

2094 Vol. 79, No. 10, October 1998 time are just beginning but can be expected to accel- erate rapidly during the next few years. Both numerical modeling and expert system techniques would benefit substantially from improved high-resolution wind and buoyancy fields. As discussed in section 3, techniques that assimilate radar

Extrapolation and mesonet data into numerical models (using, e.g., the adjoint method) are very promising and are ex- s Expert system pected to progress rapidly in the next few years.

Forecast Accuracy Explicit storm models Because of the above efforts it can now be expected Large scale models that significant advances in the ability to nowcast con- 36121824vective storms and possibly precipitation amounts will Forecast Period (hrs.) materialize during the next five years.

FIG. 15. Qualitative assessment of forecast accuracy as a function of forecast period for convective precipitation on a spatial Acknowledgments. This paper was originally prepared as an scale of a few kilometers. Adapted from Browning et al. (1980), invited contribution to the Convective Weather Forecasting Work- Doswell (1986), and Austin et al. (1987). shop at the Seventh Conference on Aviation, Range and Aero- space Meteorology. The workshop was organized by Marilyn Wolfson of MIT Lincoln Laboratory. We thank Marilyn for her encouragement to prepare this paper. Many of NCAR’s activities els but are observable in the datasets. As discussed in reported here were supported by the FAA Aviation Weather Re- this paper the modernization of the national observa- search Program to whom we are most appreciative. We are grate- tional network coupled with rapid progress in mod- ful to Dan Megenhardt, who prepared most of the figures, and eling and observational research of thunderstorm Tammy Weckwerth for reviewing the paper. We are also grateful evolution makes it possible now to implement very to CAPS for providing Fig. 11. This research is partially supported by the National Science Foundation through an Interagency short period nowcasting systems that should provide Agreement in response to requirements and funding by the Fed- a real economic benefit. eral Aviation Administration’s Aviation Weather Development Program. The views expressed are those of the authors and do not necessarily represent the official policy or position of the U.S. 6. Future government.

It is apparent that both expert system and numeri- References cal modeling techniques will continue to be developed for nowcasting convective storm evolution. Because Achtemeier, G. L., 1983: The relationship between the surface of the great need to improve flash flood warnings and wind field and convective precipitation over the St. Louis area. to provide stream flow estimates for hydrologic plan- J. Appl. Meteor., 22, 982–999. ning purposes it is natural that both techniques will Atkins, N. T., R. M. Wakimoto, and T. M. Weckwerth, 1995: evolve to include the nowcasting of quantitative pre- Observations of the sea breeze front during CaPE. Part II: Dual- Doppler and aircraft analysis. Mon. Wea. Rev., 123, 944–969. cipitation amounts. It can be expected that these two Austin, G. L., and A. Bellon, 1974: The use of digital weather techniques will be combined so that the output from records for short-term precipitation forecasting. Quart. J. Roy. the explicit numerical model forecasts of convective Meteor. Soc., 100, 658–664. storms will become one of the data fields used by the ——, and ——, 1982: Very-short-range forecasting of precipi- expert system. tation by objective extrapolation of radar and satellite data. The primary difficulties with present expert sys- Nowcasting, K. Browning, Ed., Academic Press, 177–190. ——, ——, P. Dionne, and M. Roch, 1987: On the interaction be- tems are the automated detection and monitoring of tween radar and satellite image nowcasting systems and me- convergence lines and cumulus clouds from radar and soscale numerical models. Proc. Mesoscale Analysis & Fore- satellite data. To capitalize on the strengths of each casting, Vancover, BC, Canada, European Space Agency, system it is expected that efforts will increase to com- 225–228. bine the datasets and to develop very robust techniques ——, A. Kilambi, and H. P. Biron, 1990: The operational use of a high-speed interactive radar data processing system (RDPS). to monitor convergence lines and cumulus clouds. Part 1: A technical description. Weather Radar Networking: Efforts to ingest radar, satellite, and high-density Seminar on COST Project 73, C. G. Collier and M. Chapuis, mesonet data into numerical models that can run in real Eds., Kluwer Academic Publishers, 237–241.

Bulletin of the American Meteorological Society 2095 Barclay, P. A., and K. E. Wilk, 1970: Severe thunderstorm radar Conway, B. J., 1992: Current developments in very short range echo motion and related weather events hazardous to aviation weather forecasting. Short Range Forecasting Research Divi- operations. ESSA Tech. Memo. ERLTM-NSSL 46, 63 pp. sion, PWR 1, 33 pp. [Available from Meteorological Office, [Available from National Information Service, Operations London Road, Berkshire, RG12 2SZ, United Kingdom.] Division, Springfield, VA 22161.] Corbet, J., C. Mueller, C. Burghart, K. Gould, and G. Granger, Battan, L. J., 1953: Duration of convective radar cloud units. Bull. 1994: Zeb: Software for integration, display, and management Amer. Meteor. Soc., 34, 227–228. of diverse environmental datasets. Bull. Amer. Meteor. Soc., ——, 1959: Radar Meteorology. The University of Chicago Press, 75, 783–792. 161 pp. Cotton, W. R., and L. D. Grasso, 1996: The predictability of se- Bellon, A., and G. L. Austin, 1978: The evaluation of two years vere/tornadic supercell thunderstorms. Preprints, 18th Conf. on of real time operation of a short-term precipitation forecasting Severe Local Storms, San Francisco, CA, Amer. Meteor. Soc., procedure (SHARP). J. Appl. Meteor., 17, 1778–1787. 381–384. ——, S. Lovejoy, and G. L. Austin, 1980: Combining satellite and Crook, N. A., 1996: Sensitivity of moist convection forced by radar data for the short-range forecasting of precipitation. Mon. boundary layer processes to low-level thermodynamic fields. Wea. Rev., 108, 1554–1566. Mon. Wea. Rev., 124, 1767–1785. Bjerkaas, C. L., and D. E. Forsyth, 1979: Real-time automative ——, T. L. Clark, and M. W. Moncrieff, 1991: The Denver cy- tracking of severe thunderstorms using Doppler weather ra- clone. Part II: Interaction with the convective boundary layer. dar. Preprints, 11th Conf. on Severe Local Storms, Kansas City, J. Atmos. Sci., 48, 2109–2126. MO, Amer. Meteor. Soc., 573–576. Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm identi- Blackmer, R. H., R. O. Duda, and R. Reboh, 1973: Application fication, tracking, analysis and nowcasting—A radar-based of pattern recognition to digitized Weather Radar Data. Final methodology. J. Atmos. Oceanic Technol., 10, 785–797. Rep., Contract 1-36072, SRI Project 1287, Stanford Research Donaldson, R. J., Jr., R. M. Dyer, and M. J. Kraus, 1975: An ob- Institute, Menlo Park, CA, 89 pp. [Available from National jective evaluator of techniques for predicting severe weather Information Service, Operations Division, Springfield, VA events. Preprints, Ninth Conf. on Severe Local Storms, 22161.] Norman, OK, Amer. Meteor. Soc., 321–326. Brooks, H. E., C. A. Doswell III, and L. J. Wicker, 1993: Doswell, C. A., III, 1986: Short-range forecasting. Mesoscale Me- STORMTIPE: A forecasting experiment using a three-dimen- teorology and Forecasting, P. Ray, Ed., Amer. Meteor. Soc., sional cloud model. Wea. Forecasting, 8, 352–362. 689–719. Brown, B. G., and E. Brandes, 1997: An intercomparison of 2D ——, R. Davies-Jones, and D. L. Keller, 1990: On summary mea- storm motion extrapolation algorithms. Preprints, 28th Conf. sures of skill in rare event forecasting based on contingency on Radar Meteorology, Austin, TX, Amer. Meteor. Soc., 495– tables. Wea. Forecasting, 5, 576–585. 496. Droegemeier, K. K., 1997: The numerical prediction of thunder- Browning, K. A., 1980: Local weather forecasting. Proc. Roy. Soc. storms: Challenges, potential benefits, and results from London, Ser. A, 371, 179–211. realtime operational tests. WMO Bulletin, 46, 324–336. ——, C. G. Collier, P. R. Larke, P. Menmuir, G. A. Monk, and Duda, R. O., and R. H. Blackmer, 1972: Applications of pattern R. G. Owens, 1982: On the forecasting of frontal rain using a recognition techniques to digitized weather radar data. Final Rep. weather radar network. Mon. Wea. Rev., 110, 534–552. Contract I-36092, SRI Project 1287, Stanford Research Institute, Byers, H. R., and R. R. Braham Jr., 1949: The Thunderstorm. U.S. Menlo Park, CA, 135 pp. [Available from National Informa- Govt. Printing Office, 187 pp. tion Service, Operations Division, Springfield, VA 22161.] Carbone, R. G., and Coauthors, 1990: Convective dynamics: Elvander, R. C., 1976: An evaluation of the relative performance of three Panel report. Radar in Meteorology, D. Atlas, Ed., Amer. Me- weather radar echo forecasting techniques. Preprints, 17th Conf. on teor. Soc., 391–400. Radar Meteorology, Seattle, WA, Amer. Meteor. Soc., 526–532. Carr, F. H., and M. Baldwin, 1991: Incorporation of observed pre- Endlich, R. M., and D. E. Wolf, 1981: Automatic cloud tracking cipitation estimates during the initialization of synoptic and applied to GOES and METEOSAT observations. J. Appl. Me- mesoscale storms. Preprints, First Int. Symp. on Winter Storms, teor., 20, 309–319. New Orleans, LA, Amer. Meteor. Soc., 71–75. Evans, J. E., and E. R. Ducot, 1994: The Integrated Terminal Cherna, E. V., and E. J. Stansbery, 1986: Sferics rate in relation Weather System. Lincoln Lab. J., 7, 449–474. to thunderstorm dimensions. J. Geophys. Res., 91, 8701– Fiorino, M., and T. T. Warner, 1981: Incorporating surface winds 8707. and rainfall rates into the initialization of a mesoscale hurri- Chornoboy, E. S., 1992: Storm tracking for TDWR: A correlation cane model. Mon. Wea. Rev., 109, 1914–1929. algorithm design and evaluation. Project Rep. DOT/FAA/NR- Foote, G. B., and C. G. Mohr, 1979: Results of a randomized hail 91/8, MIT Lincoln Laboratory, 83 pp. [Available from National suppression experiment in northeast Colorado. Part VI: Post Information Service, Operations Division, Springfield, VA hoc stratification by storm type and intensity. J. Appl. Meteor., 22161.] 18, 1589–1600. Collier, C. G., 1989: Applications of Weather Radar Systems. Gal-Chen, T., 1978: A method for the initialization of the anelas- Wiley and Sons, 294 pp. tic equations: Implications for matching models with obser- ——, 1991: NIMROD—A system for nowcasting and initializa- vations. Mon. Wea. Rev., 106, 587–606. tion for modeling using regional observational data. Preprints, Garstang, M., and H. J. Cooper, 1981: The role of near surface out- 25th Int. Conf. on Radar Meteorology, Paris, France, Amer. flow in maintaining convective activity. Proc. Nowcasting-I Symp., Meteor. Soc., 21–24. Copenhagen, Denmark, European Space Agency, 161–168.

2096 Vol. 79, No. 10, October 1998 Goodman, S. J., 1990: Predicting thunderstorm evolution using Kessinger, C. J., and C. K. Mueller, 1991: Background studies and ground-based lightning detection networks. NASA Tech. nowcasting Florida thunderstorm activity in preparation for the Memo NASA TM-103521, 193 pp. [Available from National CaPOW forecast experiment. Preprints, 25th Conf. on Radar Information Service, Operations Division, Springfield, VA Meteorology, Paris, France, Amer. Meteor. Soc., 67–70. 22161.] Kessler, E., 1966: Computer program for calculating average ——, D. E. Buechler, and P. J. Meyer, 1988: Convective tendency lengths of weather radar echoes and pattern bandedness. J. images derived from a combination of lightning and satellite Atmos. Sci., 23, 569–574. data. Wea. Forecasting, 3, 173–188. ——, and J. A. Russo, 1963: Statistical properties of weather ra- Hand, W. H., 1996: An object-oriented technique for nowcast- dar echoes. Preprints, 10th Weather Radar Conf., Washing- ing heavy showers and thunderstorms. Meteorol. Appl., 3, 31– ton, D.C., Amer. Meteor. Soc., 25–33. 41. Kingsmill, D. E., 1995: Convection initiation associated with a ——, and B. J. Conway, 1995: An object-oriented approach to sea-breeze front, a gust front and their collision. Mon. Wea. nowcasting showers. Wea. Forecasting, 10, 327–341. Rev., 123, 2913–2933. Hane, C. E., and B. C. Scott, 1978: Temperature and pressure per- Kitzmiller, D. H., 1996: One-hour forecasts of radar-estimated turbations within convective clouds derived from detailed air rainfall by an extrapolative-statistical method. TDL Office motion information: Preliminary testing. Mon. Wea. Rev., 106, Note 96-1, 26 pp. [Available from NWS/TDL, W/OSD2, 1325 654–661. East–West Highway, Silver Spring, MD 20910.] Hemler, R. S., F. B. Lipps, and B. B. Ross, 1991: A simulation of Knight, C. A., W. D. Hall, and P. M. Roskowski, 1983: Visual a squall line using a nonhydrostatic cloud model with a 5 km cloud histories related to first radar echo formation in north- horizontal grid. Mon. Wea. Rev., 119, 3012–3033. east Colorado cumulus. J. Climate Appl. Meteor., 22, 1022– Henry, S. G., 1993: Analysis of thunderstorm lifetime as a func- 1040. tion of size and intensity. Preprints, 26th Conf. on Radar Me- Krishnamurti, T. N., H. S. Bedi, W. Heckley, and K. Ingles, 1988: teorology, Norman, OK, Amer. Meteor. Soc., 138–140. Reduction of the spin-up time for evaporation and precipita- ——, and J. W. Wilson, 1993: Developing thunderstorm forecast tion in spectral model. Mon. Wea. Rev., 116, 907–920. rules utilizing first detectable cloud radar echoes. Preprints, Kuo, Y.-H., and W. Wang, 1996: Simulations of a prefrontal Fifth Conf. on Aviation Weather Systems, Vienna, VA, Amer. rainband observed in TAMEX IOP 13. Proc. Seventh Conf. Meteor. Soc., 304–307. on Mesoscale Processes, Reading, United Kingdom, Amer. ——, C. K. Mueller, and J. W. Wilson, 1996: Base-line statistical Meteor. Soc., 335–338. evaluation of the automated thunderstorm nowcaster. Preprints, Lazarus, S., 1996: Assimilation and prediction of a Florida mul- 15th Conf. Weather Analysis and Forecasting, Norfolk, VA, ticell storm using observed single-Doppler data. Ph.D. dis- Amer. Meteor. Soc., 344–346. sertation, School of Meteorology, University of Oklahoma, Hill, F. F., K. W. Whyte, and K. A. Browning, 1977: The contri- 340 pp. [Available from the School of Meteorology, Univer- bution of a weather radar network for forecasting frontal pre- sity of Oklahoma, 100 East Boyd, Room 1310, Norman, OK cipitation; A case study. Meteorol. Mag., 106, 68–89. 73019.] Hilst, G. R., and J. A. Russo, 1960: An objective extrapolation Ligda, M. G., 1953: The horizontal motion of small precipitation technique for semi-conservative fields with an application to areas as observed by radar. Tech. Rep. 21, Department of Me- radar patterns. Tech. Memo 3, The Travelers Research Cen- teorology, M.I.T., 60 pp. [Available from Library, Massachu- ter, Hartford, CT, 34 pp. setts Institute of Technology, 77 Massachusetts Ave., Cam- Hirasawa, M., 1991: Nowcasting and very-short-range forecast- bridge, MA 02139.] ing of precipitation in Japan Meteorology Agency. JMA Tech. Lin, Y., P. S. Ray, and K. W. Johnson, 1993: Initialization of a Memo. modeled convective storm using Doppler radar-derived fields. Jackson, M. E., 1993: An echo motion algorithm for air traffic Mon. Wea. Rev., 121, 2757–2775. management using a national radar mosaic. Preprints, Fifth Int. ——, K. E. Mitchell, E. Rogers, and M. E. Baldwin, 1996: Im- Conf. on Aviation Weather Systems, Vienna, VA, Amer. Me- pact of hourly precipitation input on the NCEP Eta model. Pre- teor. Soc., 299–303. prints, 11th Conf. on Numerical Weather Prediction, Norfolk, ——, and R. T. Jesuroga, 1995: The ATMS convective area guid- VA, Amer. Meteor. Soc., 268–269. ance product. Preprints, Sixth Int. Conf. on Aviation Weather Livingston, J. M., and E. P. Krider, 1978: Electric fields produced Systems, Dallas, TX, Amer Meteor. Soc., 78–82. by Florida thunderstorms. J. Geophys. Res., 83, 385–401. Johnson, J. T., and Coauthors, 1995: Operational testing of Molinari, J., 1982: Numerical hurricane prediction using assimi- enhanced WSR-88D algorithms and display concepts in lation of remotely sensed rainfall rates. Mon. Wea. Rev., 110, National Weather Service offices. Preprints, 27th Conf. on 553–571. Radar Meteorology, Vail, CO, Amer. Meteor. Soc., 170– Moncrieff, M. W., and M. J. Miller, 1976: The dynamics and simu- 172. lation of tropical cumulonimbus and squall lines. Quart. J. Roy. Jones, C. D., and B. Macpherson, 1997: A latent heat nudging Meteor. Soc., 102, 373–394. scheme for the assimilation of precipitation data into an op- Mueller, C. K., and J. W. Wilson, 1989: Evaluation of the TDWR erational mesoscale model. Meteorol. Appl., 4, 269–277. nowcasting experiment. Preprints, 24th Conf. on Radar Me- Juvanon du Vachat, R., and J. L. Cheze, 1993: Nowcasting storms teorology, Tallahassee, FL, Amer. Meteor. Soc., 224–227. and precipitations over Ile-de-France. The ASPIC project. First ——, J. W. Wilson, and N. A. Crook, 1993: Utility of soundings European Conf. on Applications of Meteorology, Oxford, and mesonets to forecast thunderstorm initiation. Wea. Fore- United Kingdom, Amer. Meteor. Soc., 95–97. casting, 8, 132–146.

Bulletin of the American Meteorological Society 2097 ——, F. McDonough, and J. W. Wilson, 1997a: Forecasting the feeder” process in warm frontal rainbands. J. Atmos. Sci., 40, extent and lifetime of thunderstorm complexes for aviation ap- 1185–1206. plications. Preprints, Seventh Conf. on Aviation, Range and Schaefer, J. T., 1990: The critical success index as an indicator Aerospace Meteorology, Long Beach, CA, Amer. Meteor. of warning skill. Wea. Forecasting, 5, 570–575. Soc., 244–248. Shackford, C. E., 1960: Radar indications of a precipitation-light- ——, and Coauthors, 1997b: The Memphis ITWS Convective ning relationship in New England thunderstorms. J. Meteor., Forecasting Collaborative Demonstration. Preprints, 7th Conf. 17, 15–17. on Aviation, Range and Aerospace Meteorology, Long Beach, Shapiro, A., S. Ellis, and J. Shaw, 1995: Single Doppler velocity CA, Amer. Meteor Soc., 249–257. retrievals with Phoenix II data: Clear air and microburst wind Nevado, T., 1990: Composite radar and satellite data process- retrievals in the planetary boundary layer. J. Atmos. Sci., 52, ing in the meteorological system in Spain. Weather Radar 1265–1287. Networking: Seminar on COST Project 73, C. G. Collier and Smith, S. B., and M. E. Churma, 1996: An overview of the M. Chapuis, Eds., Kluwer Academic Publishers, 136–146. AWIPS thunderstorm product. Preprints, 15th Conf. on Newton, C. W., and J. C. Fankhauser, 1964: On the movements Weather Analysis and Forecasting, Norfolk, VA, Amer. Me- of convective storms, with emphasis on size discrimination teor. Soc., 297–300. in relation to water-budget requirements. J. Appl. Meteor., 3, Stensrud, D. J., and R. A. Maddox, 1988: Opposing mesoscale 651–668. circulations: A case study. Wea. Forecasting, 3, 189–204. Noel, T. M., and A. Fleisher, 1960: The linear predictability of Sun, J., and N. A. Crook, 1994: Wind and thermodynamic retrieval weather radar signals. Research Rep. 34, Department of Me- from single-Doppler measurements of a gust front observed teorology, M.I.T., 46 pp. [Available from Library, Massachu- during Phoenix II. Mon. Wea. Rev., 122, 1075–1091. setts Institute of Technology, 77 Massachusetts Ave., Cam- ——, and ——, 1997: Dynamical and microphysical retrieval bridge, MA 02139.] from Doppler radar observations using a cloud model and its Podhorsky, D, 1987: Use of satellite imagery for forecasting the adjoint. Part I: Model development and simulated data experi- evolution of mesoscale phenomena. Preprints, Workshop on ments. J. Atmos. Sci., 54, 1642–1661. Satellite and Radar Imagery Interpretation, Reading, United Thorpe, A. J., M. J. Miller, and M. W. Moncrieff, 1982: Two- Kingdom, EUMETSAT, 415–436. dimensional convection in nonconstant shear: A model of mid- Purdom, J. F. W., 1973: Satellite imagery and the mesoscale con- latitude squall lines. Quart. J. Roy. Meteor. Soc., 108, 739–762. vective forecast problem. Preprints, Eighth Conf. on Severe Tsonis, A. A., and G. L. Austin, 1981: An evaluation of extrapo- Local Storms, Denver, CO, Amer. Meteor. Soc., 244–251. lation techniques for the short-term prediction of rain amounts. ——, 1976: Some uses of high resolution GOES imagery in the Atmos.–Ocean, 19, 54–65. mesoscale forecasting of convection and its behavior. Mon. Ulanski, S. L., and M. Garstang, 1978: The role of surface di- Wea. Rev., 104, 1474–1483. vergence and vorticity in the life cycle of convective rainfall. ——, 1982: Subjective interpretations of geostationary satellite Part I: Observations and analysis. J. Atmos. Sci., 35, 1047– data for nowcasting. Nowcasting, K. Browning, Ed., Academic 1062. Press, 149–166. Wakimoto, R. M., and J. K. Lew, 1993: Observations of a Florida Puri, K., and M. J. Miller, 1990: The use of satellite data in the waterspout during CaPE. Wea. Forecasting, 8, 412–423. specification of convective heating for diabatic initialization ——, and N. T. Atkins, 1994: Observations of the sea-breeze front and moisture adjustment in numerical weather prediction mod- during CaPE. Part I: Single-Doppler, satellite and cloud pho- els. Mon. Wea. Rev., 118, 67–93. togrammetry analysis. Mon. Wea. Rev., 122, 1092–1114. Qiu, C.-J., and Q. Xu, 1992: A simple adjoint method of wind Wang, D. H., and T. T. Warner, 1988: Use of four-dimensional analysis for single-Doppler data. J. Atmos. Oceanic Technol., data assimilation by Newtonian relaxation and latent-heat forc- 9, 588–598. ing to improve mesoscale-model precipitation forecast. A case Rineheart, R. E., 1981: A pattern-recognition technique for use study. Mon. Wea. Rev., 116, 2593–2613. with conventional weather radar to determine internal storm ——, M. Xue, V. Wong, and K. Droegemeier, 1996: Predictions motions. Recent Progress in Radar Meteorology, R. Carbone, and simulations of convective storms during Vortex95. Pre- Ed., National Center for Atmospheric Research, 105–118. prints, 11th Conf. on Numerical Weather Prediction, Norfolk, Roberts, R. D., 1996: Integration of WSR-88D and GOES-8 data VA, Amer. Meteor. Soc., 301–303. for detecting and forecasting storm initiation. Preprints, 15th Watson, I. W., and D. O. Blanchard, 1984: The relationship be- Conf. on Wea. Analysis and Forecasting, Norfolk, VA, Amer. tween total area divergence and convective precipitation in Meteor. Soc., 287–290. south Florida. Mon. Wea. Rev., 112, 673–685. Roberts, W. F., P. C. Kucera, C. M. Lusk, D. C. Walker, and L. E. Weckwerth, T. M., J. W. Wilson, and R. M. Wakimoto, 1996: Johnson, 1996: 1995 real-time forecast exercise for WFO-Ad- Thermodynamic variability within the convective boundary vanced. Preprints, 12th Int. Conf. on Interactive Information layer due to horizontal convective rolls. Mon. Wea. Rev., 124, and Processing Systems for Meteorology, Oceanography, and 769–784. Hydrology, Atlanta, GA, Amer. Meteor. Soc., 198–201. Weisman, M. L., and J. B. Klemp, 1986: Characteristics of iso- Rotunno, R., J. B. Klemp, and M. L. Weisman, 1988: A theory lated convective storms. Mesoscale Meteorology and Fore- for strong, long-lived squall lines. J. Atmos. Sci., 45, 463–485. casting, P. S. Ray, Ed., Amer. Meteor. Soc., 331–358. Rutledge, S. A., and P. V. Hobbs, 1983: The mesoscale and ——, ——, and R. Rotunno, 1988: Structure and evolution of microscale structure and organization of clouds and precipi- numerically simulated squall lines. J. Atmos. Sci., 45, 1990– tation in a mid-latitude cyclone. VIII: A model for the “seeder- 2013.

2098 Vol. 79, No. 10, October 1998 Weygandt, S. S., A. Shapiro, and K. K. Droegemeier, 1998: The of boundary layer convergence zones and horizontal rolls in use of wind and thermodynamic retrievals to create initial fore- the initiation of thunderstorms: A case study. Mon. Wea. Rev., cast fields from single-Doppler observations of a supercell 120, 1758–1815. thunderstorm. Preprints, 16th Conf. on Weather Analysis and Witt, A., and J. T. Johnson, 1993: An enhanced storm cell identifi- Forecasting, Phoenix, AZ, Amer. Meteor. Soc., 286–291. cation and tracking algorithm. Preprints, 26th Int. Conf. on Ra- Wilk, K. E., and K. C. Gray, 1970: Processing and analysis tech- dar Meteorology, Norman, OK, Amer. Meteor. Soc., 514–521. niques used with the NSSL weather radar system. Preprints, Wolf, D. E., D. J. Hall, and R. M. Endlich, 1977: Experiment in 14th Conf. on Radar Meteorology, Tucson, AZ, Amer. Me- automatic cloud tracking using SMS-GOES data. J. Appl. Me- teor. Soc., 369–374. teor., 16, 1219–1230. Williams, E. R., 1985: Large-scale charge separation in thunder- Wolfson, M. M., F. L. Delanoy, B. E. Forman, F. G. Hallowell, clouds. J. Geophys. Res., 90, 6013–6025. M. L. Pawlak, and P. D. Smith, 1994: Automated microburst ——, M. E. Weber, and R. E. Orville, 1989: The relationship be- wind-shear prediction. Lincoln Lab. J., 7, 399–426. tween lightning type and convective state of thunderclouds. Xue, M., D. Wang, D. Hou, K. Brewster, and K. K. Droegemeier, J. Geophys. Res., 94, 13 213–13 220. 1998: Prediction of the 7 May 1995 squall lines over the cen- Wilson, J. W., 1966: Movement and predictability of radar ech- tral US with intermittent data assimilation. Preprints, 16th Conf. oes. Tech. Memo ERTM-NSSL-28, National Severe Storms on Weather Analysis and Forecasting, Phoenix, AZ, Amer. Laboratory, 30 pp. [Available from National Information Ser- Meteor. Soc., 161–163. vice, Operations Division, Springfield, VA 22161.] Zhang, D. L., K. Gao, and D. B. Parsons, 1989: Numerical simu- ——, and R. Carbone, 1984: Nowcasting with Doppler radar: The lation of an intense squall line during 10–11 June 1985 Pre- forecaster-computer relationship. Nowcasting II, K. Browning, STORM. Part I: Model verification. Mon. Wea. Rev., 117, Ed., European Space Agency, 177–186. 960–994. ——, and W. E. Schreiber, 1986: Initiation of convective storms Ziegler, C. L., 1985: Retrieval of thermal and microphysical vari- by radar-observed boundary layer convergent lines. Mon. Wea. ables in observed convective storms. Part I: Model devel- Rev., 114, 2516–2536. opment and preliminary testing. J. Atmos. Sci., 42, 1487– ——, and C. K. Mueller, 1993: Nowcasts of thunderstorm initia- 1509. tion and evolution. Wea. Forecasting, 8, 113–131. Zittel, W. D., 1976: Computer applications and techniques for ——, and D. L. Megenhardt, 1997: Thunderstorm initiation, or- storm tracking and warning. Preprints, 17th Conf. on Radar ganization and lifetime associated with Florida boundary layer Meteorology, Seattle, WA, Amer. Meteor. Soc., 514–521. convergence lines. Mon. Wea. Rev., 125, 1507–1525. Zou, X., and Y.-H. Kuo, 1996: Rainfall assimilation through an ——, J. Moore, B. Foote, B. Martner, A. Rodi, T. Uttal, and optimal control of initial and boundary conditions in a limited- J. Wilczak, 1988: Convection initiation and downburst ex- area mesoscale model. Mon. Wea. Rev., 124, 2859–2882. periment (CINDE). Bull. Amer. Meter. Soc., 69, 1328–1348. Zupanski, D., and F. Mesinger, 1995: Four-dimensional varia- ——, G. B. Foote, J. C. Fankhauser, N. A. Crook, C. G. Wade, tional assimilation of precipitation data. Mon. Wea. Rev., 123, J. D. Tuttle, C. K. Mueller, and S. K. Krueger, 1992: The role 1112–1127.

Bulletin of the American Meteorological Society 2099