Five of the nowcasting systems that were available during the Sydney 2000 Forecast Demonstration Project (FDP) were selected for evaluation. These systems, from the United States, the United Kingdom, and Australia, had the capability to nowcast the location and, with one exception, the intensity of convective storms. Six of the most significant convective storm cases from the 3-month FDP were selected for evaluating the performance of these state-of-the-art nowcasting systems, which extrapolated storms using a variety of methods, including cell and area tracking, model winds, and sounding winds. Three of the systems had the ability to forecast the initiation and growth of storms. Nowcasts for 30 and 60 min were evaluated, and it was found that even for such short time periods the skill of the extrapolation-only systems was often very low. Extrapolation techniques that allowed for differential motion performed slightly better, since high-impact storms often have motions different than surrounding storms. The ability to forecast initiation, growth, and dissipation is in its infancy. However, it was demonstrated that significant improvement in forecast accuracy was obtained for several of these cases when the locations of boundary layer convergence lines (sea breeze and gust fronts) were used in the nowcasts.
Based on the experiences during the FDP, and in forecast offices in the United States, a discussion is provided of the overall status of nowcasting convective storms. In addition, proposed future directions are discussed concerning the specificity of nowcast products, experimental test beds, and additional observations and research required.
The stated goals (Keenan et al. 2003) of the Sydney 2000 Forecast Demonstration Project (FDP) were to demonstrate the capability of modern forecast systems, and to quantify the associated benefits in the delivery of a real-time nowcast service. The term nowcasting is used to emphasize the short forecast period (from 0 to 6 h) and the high time and space specificity of the products. This paper will focus on the capability of the assembled systems, which represented the state of the art in operational nowcasting technology, to forecast convective weather in the 0–1-h range. Note that while some of these systems included numerical weather prediction (NWP) capabilities, they would not typically be classified as NWP systems, since they relied heavily on high-resolution observations, particularly radar data.
The Pierce et al. (2004, in this issue) and Ebert et al. (2004, in this issue) papers in this special issue evaluate the same nowcasting systems, but use statistical measures summed over extensive time periods to evaluate the nowcasts. The approach here was to examine in detail the nowcasts from only 6 days of the FDP, and then only for particularly high-impact events on those days. The intention was to gain insight into the strengths and weaknesses of the individual systems for high-impact weather events and for differing meteorological situations, so that specific comments could be made regarding the present state and future direction of nowcasting convective storms.
The official period of the FDP was from 4 September 2000 to 21 November 2000; however, because of the occurrence of two major convective events on 30 November and 1 December, the period for this paper was extended to include these dates. Convective storms with radar reflectivities exceeding 40 dBZ occurred within 100 km of Sydney on 24 days during this period. On many of these days, the storms were isolated and remained outside the Sydney basin, occurring instead over the Blue Mountains to the west or the Tasman Sea to the east. Six of these 24 days were selected for study, with selection being based on radar data availability, variability in storm evolution, storm intensity, and proximity to the Sydney metropolitan area.
Section 2 of this paper provides a brief description of each nowcast system, and section 3 contains a discussion of the six cases used for the evaluation. In section 4, the performance of each system for these six cases is described, to provide an insight into their comparative strengths and weaknesses. Section 5 examines post-FDP nowcasts made by the National Center for Atmospheric Research (NCAR) nowcasting system, for the purpose of illustrating future potential in the nowcasting field. Section 6 presents conclusions, and section 7 summarizes future directions for nowcasting.
2. Nowcast systems
Five nowcasting systems were evaluated:1 1) the Thunderstorm Identification, Tracking, Analysis, and Nowcasting (TITAN) system, developed by Dixon and Wiener (1993) and operated by the Bureau of Meteorology Research Centre (BMRC); 2) the Spectral Prognosis (S-PROG) system (Seed 2003), developed and operated by BMRC; 3) the Nowcasting and Initialisation for Modelling Using Regional Observation Data System (Nimrod; Golding 1998), developed and operated by the Met Office; 4) the Generating Advanced Nowcasts for Deployment in Operational Land-surface Flood Forecasts (Gandolf, Pierce et al. 2000), developed and operated by the Met Office; and 5) the thunderstorm Auto-nowcaster (ANC) system (Mueller et al. 2003), developed and operated by NCAR. Table 1 summarizes the basic features of each system.
All the systems employ some methodology for extrapolating storms as indicated in Table 1. We briefly describe the various methodologies here. Extrapolation systems called area trackers cross correlate two reflectivity images separated in time; the translation that yields the maximum correlation is used as an estimate of the echo displacement. This method may be used to obtain a field of extrapolation vectors (Rinehart 1981). In contrast, a cell tracker isolates individual cells and uses one of a variety of techniques to match cells at different times, thereby obtaining a motion vector for each cell. A third available extrapolation technique is to use observed winds from soundings or from numerical models to advect precipitation systems. All three types were employed during the FDP.
A brief description of each nowcast system is provided below; for detailed descriptions the reader is referred to the indicated references.
The TITAN system (Dixon and Wiener 1993) was initially developed specifically for convective storm nowcasting. TITAN is a cell tracker that has the capability to grow and dissipate existing echoes based on past trends; however, this feature was not activated for the FDP. As operated by BMRC, it simply produced an outline of a given reflectivity threshold. Two instances of TITAN were run, using thresholds of 35 and 45 dBZ, but only the 35-dBZ threshold was evaluated for this paper. Nowcasts were generated every 10 min for periods between 10 and 60 min ahead.
The S-PROG system (Seed 2003) is best classified as an area tracker; during the FDP only one vector was produced to extrapolate the entire radar echo field. S-PROG exploits the notion that the lifetime of a feature is dependent on the scale of the feature. The evolution of each scale is modeled using a simple autoregressive (lag two) model, which automatically causes the forecast field to become smooth as the various scales age to reflect the uncertainty in the nowcast. Every 5 min, S-PROG produced a forecast on a 1-km grid, with lead times in the 0–1-h range and a forecast interval of 10 min.
Nimrod (Golding 1998) was specifically designed for nowcasting stratiform precipitation. Its ability to nowcast convective rainfall, beyond extrapolation, depends on the ability of the associated numerical model. As with other operational NWP models this capability is very limited particularly for the nowcasting period. Nimrod extrapolations are a mix of past echo motion and numerical weather prediction wind fields; with increasing time, the advection vectors and precipitation field relax to those of the numerical model forecast. During the first hour the numerical model has little impact on the nowcast. During the FDP the numerical weather prediction information was provided by the Australian Bureau of Meteorology Limited Area Prediction System (LAPS; Puri et al. 1998). The original precipitation field was derived from an analysis of mosaic radar and satellite data. Nimrod produced rainfall-rate nowcasts on a 5-km grid every 30 min for a 0–6-h period.
Gandolf (Pierce et al. 2000) was designed for the specific purpose of nowcasting large-scale multicellular convective storms. Gandolf is an object-oriented expert system, that analyzes 3D reflectivity data to identify convective entities. Entities are classified into stages of development, and then, based on a conceptual model of idealized storm evolution, future states are nowcast. Storm motion is predicted using steering-level winds from LAPS. The conceptual model has the capacity to initiate new cells or daughter cells along the flank of existing storms and in clear air, provided the boundary layer convergence from the NWP model is sufficiently strong. Every 10 min on a 2-km grid Gandolf produced a 0–2-h nowcast of the precipitation-rate field.
The ANC (Mueller et al. 2003) was developed and operated by NCAR. The ANC was specifically developed to nowcast convective storms. It is an expert system that utilizes fuzzy logic to combine forecast features based on physical understanding of storm evolution. Radar echoes are extrapolated primarily using an area tracker called Tracking Radar Echoes by Correlation (TREC) (Rinehart 1981; Tuttle and Foote 1990). However, the TITAN cell tracker is also used to identify storm characteristics. When insufficient data are available to obtain echo motion from reflectivity patterns, then sounding winds are used to define a steering-level wind to extrapolate the reflectivity field. Nowcasting storm initiation, growth, and dissipation is based on the interaction of storms and cumulus clouds with boundary layer convergence lines detected by radar, and wind field characteristics retrieved from a numerical model and its adjoint (Sun and Crook 2001). This wind retrieval system is called the Variational Doppler Radar Analysis System (VDRAS). Every 10 min it produced boundary layer winds on a grid with a horizontal spacing of 3 km and vertical spacing of 400 m. The ANC produced nowcasts of reflectivity and precipitation rate on a 1-km grid every 5 min, for 30- and 60-min nowcast periods.
Figure 1 is an example of the 30-min rain-rate nowcasts (in units of mm h−1) made by each of the systems for 26 September, verifying at 1030 UTC.2 The difference in resolution, and several other of the major features of each system, can be seen in Fig. 1: immediately apparent are the relatively high resolution of the ANC, the smoothing and dissipation of the rainfall-rate field by S-PROG, the lower resolution of Nimrod, the dissipation of the rainfall field by Gandolf, and the object-based character of TITAN.
Table 2 provides a list of the six cases that were selected for study with a short description of the associated weather events. Factors used in selecting cases were storm severity, impact on the Sydney basin, forecast challenge, and data availability. Three of the days had particularly severe weather within the greater Sydney vicinity; two of these were associated with hail and flash flooding (30 November and 1 December) and the third with small tornadoes and large hail (3 November). Figure 2, and several later figures (e.g., Figs. 4, 6, 8, 11, and 14) show time series of four radar reflectivity images for each of the six cases.
Note that a general description of Sydney weather and the effects of terrain are given in May et al. (2004, in this issue) and Potts et al. (2000). The terrain in the vicinity of Sydney is shown later (Fig. 15).
4. Evaluation of FDP systems
The intent in this section is to examine in detail the system nowcasts for a few of the most significant events in order to obtain insight into the strengths and weaknesses of the systems in relationship to the meteorological situation. Emphasis is on the 60-min nowcasts.
Initially, a visual examination was made of all the 30- and 60-min nowcasts from each system for each of the six cases, which involved inspection of many hundreds of digital copies of the nowcasts made in real time. This inspection reinforced the general comments made in section 2 with respect to Fig. 1.
For four of the six cases, the most significant weather feature was examined; for example, the feature of interest for 25 September is the short intense line of thunderstorms that rapidly moved from 240° over the radar (white outline in Fig. 2). The evaluation considered five elements: errors in displacement, intensity bias, and ability to nowcast initiation, growth, and dissipation. Displacement errors were determined for the feature of interest using charts similar to some of the figures shown later (Figs. 3, 5, and 7).3 Evaluation of the other nowcast elements (intensity bias, initiation, growth, and dissipation) was based on qualitative analysis of radar images.
Not all five elements were present in each case. Also, not all elements could be evaluated for each system. TITAN did not nowcast intensity; therefore, only displacement errors could be considered. S-PROG does not have the capability to grow or initiate precipitation. Also since S-PROG always decreases the intensities, it was not evaluated for dissipation. ANC and Nimrod were evaluated for all five elements. Gandolf is not shown since it essentially produced no 60-min nowcasts, due to the unrealistic dissipation characteristic discussed earlier.
a. 25 September 2000
The feature of interest for 25 September is the rapidly moving short line of thunderstorms indicated in Fig. 2. This line moved at 80 km h−1 from 240° with very little change in intensity until it reached the coastline where it rapidly dissipated. Figure 3 shows the actual path of the center of the line (gray) and the 60-min nowcast locations of the line center for each system. Table 3 evaluates the five elements listed above.
Nimrod was the first system to provide a 60-min nowcast since it used a network of radars rather than one radar, which is particularly important for very rapidly moving echoes as in this case. It is apparent from Fig. 2 and Table 3 that TITAN and then the ANC provided the best location nowcasts. All systems tended to nowcast the position too far to the right (southwest), because the line of interest had a motion farther to the left (north) than other storms. The S-PROG positions were the farthest to the right (southwest), because this method used only one echo motion vector for the entire field. The other area-type trackers (Nimrod and ANC) allowed for differential motion vectors. After 1430 UTC the line started to dissipate and was less than 35 dBZ4 shortly after 1500 UTC. As mentioned above, the ANC and Nimrod are the only systems that had the ability to nowcast intensity changes. The ANC maintained the intensity of the line until 1440 UTC, and then successfully nowcast its dissipation as it moved off the coast. Nimrod weakened the system too much initially and then failed to nowcast the dissipation as it moved over the ocean. Although TITAN did not forecast rainfall intensity, it stopped extrapolating the storm after 1443 UTC, because the volume of the echo dropped below a prescribed limit. This was also the reason that the ANC dissipated the echo, since this latter system utilized computations of echo size and trend from TITAN as part of the logic for nowcasting dissipation.
b. 26 September 2000
The feature of interest on 26 September is the line of thunderstorms indicated in Fig. 4 that moves from 300° at 55 km h−1. Figure 5, similar to Fig. 3, shows the actual and nowcast positions of the center of this line. Table 4, similar to Table 3, provides an evaluation of the aformentioned five elements. There is an increase in the line extent and intensity between 0925 and 1015 UTC that is apparent in Fig. 4. This occurred when the line encounters the large-scale boundary [“southerly buster”; Colquhoun et al. (1985)] shown in Fig. 4 that was moving in from the southeast. This intensification occurred in spite of the boundary layer being very stable. The Doppler velocity data showed that outflow from this line of storms was able to penetrate to the ground prior to intersection with the southerly buster. Figure 4 shows this outflow boundary. There was a brief decrease in reflectivity when the line reached the coast, but then it quickly regained its original intensity.
When the line first moved into range, there was little data available from which extrapolation vectors could be obtained; the ANC utilized sounding winds until there were sufficient data. TITAN was the first system to obtain sufficient data for quality extrapolations, after which both S-PROG and the ANC provided excellent extrapolations, with TITAN close behind in quality. The relatively poor performance of Nimrod is likely to be related to the small scale of the line: Nimrod used a 5-km resolution and it appeared that the line of interest lost definition and was smoothed with other echoes.
The incorrect extrapolation of the line of storms by the ANC prior to 1052 UTC hindered the ability of the ANC to forecast the growth. A correct storm extrapolation would have anticipated the line colliding with the southerly buster, and would have subsequently resulted in a nowcast of storm growth. Nimrod tended to decrease the intensity of the line throughout the period of interest.
c. 19 October 2000
During a 4-h period five cells formed in close proximity to each other along a quasi-stationary convergence line (Fig. 6), such that there was a tendency for the cells to pass over nearly the same locations. Figure 7 shows the track of each storm, and the inset shows the duration of each storm: storms A, B, and C lived between 35 and 65 min, meaning that by the time each system detected the storm and had sufficient information to extrapolate it, the storm was dissipating or dissipated, whereas storms D and E had sufficient lifetimes for reasonable 30- and 60-min nowcasts to be made. This was particularly true for nowcasts issued after 0500 UTC.
The ANC and TITAN tended to have the better nowcasts of position; Nimrod and S-PROG tended to move the storms too quickly. For Nimrod this was probably because its lower resolution combined dissipating storms with new storms. Displacement errors for the 30-min nowcasts were primarily in the 0–10-km range, and 0–20-km range for 60-min nowcasts. The ANC occasionally overestimated storm intensity, whereas Nimrod and S-PROG more consistently underestimated intensity. Note that the ANC provided only 5–10-min lead times on the initiation of the cells even though a boundary was present for cells C, D, and E. This lack of lead time is to be expected, since the cells were 70–80 km from the radar, and detection of growing congestus by the radar was minimal because of beam overshooting and intensities below minimum detection capabilities at that range. In this situation, initiation in ANC depended on evidence of cumulus in the vicinity of the convergence line (boundary), but satellite data were not available at sufficient frequency to provide this evidence. Because of the short lifetime of three of the five storms, there was not an easy way to present results like those in Table 4. Short-lived storms, as in this case, are rather common, pointing out the need to nowcast storm initiation and dissipation.
d. 3 November 2000
The feature of interest for this case was a storm that produced three small tornadoes and large hail in the Sydney vicinity. This storm at times exhibited supercell characteristics; note the well-defined clockwise hook echo in Fig. 8d at 0518 UTC [this case is described in detail elsewhere in this issue by Sills et al. (2004, in this issue)]. Nowcasts of the storm were examined from 0317 UTC, when it was first identified as a persistent feature, until 0622 UTC, when it dissipated to below 35 dBZ near the coastline. It formed in the vicinity of other storms that were moving from the west-southwest; however, the storm of interest had an average direction from 210° at 25 m s−1; during its most intense period (0450–0530 UTC) it moved almost due north. Its formation was influenced by the nearby intersection of a gust front with the sea breeze, as was its subsequent change to more northerly motion. Figure 9 shows the intersection of these two convergence lines at 0428 UTC just northwest of the cell of interest. The red arrow shows the 60-min ANC location of the storm, the yellow arrow shows the actual motion, and the white arrow shows the motion of the center of the maximum convergence associated with the intersection. It is apparent that the storm's direction of motion was similar to that of the maximum low-level convergence (where the convergence values were obtained from VDRAS). Elsewhere in this special issue, Crook and Sun (2004, in this issue) show the time history of the convergence fields at this time; while the ANC ingested this convergence field, no nowcast rules were in place to utilize the information in nowcasting the storm's position.
Although TITAN's initial nowcast (verifying at 0330 UTC) was too far to the east, Fig. 10 and Table 5 show that TITAN extrapolations were clearly superior to those of the other systems. This is unsurprising, since this was precisely the kind of situation in which a cell tracker should excel over an area tracker, that is, when an individual storm moves differently than surrounding storms. S-PROG, which used one vector for the entire area, would be expected to be least effective in this situation, and that was in fact the case. Obviously, nowcasting the northward movement of this storm was critical, since this direction took the storm into the western portions of Sydney, whereas the initial easterly track would have kept it well south of the city.
The primary intensification occurred between 0400 and 0430 UTC. The 60-min nowcasts from the ANC intensified the storm during this period, but then decreased it somewhat between 0430 and 0500 UTC; in reality this dissipation did not occur. This incorrect nowcast was based on the prediction that the storm would move away from the collision point between the two boundaries (Fig. 9). Nimrod tended to slightly decrease the intensity at each forecast period. The actual primary dissipation occurred between 0600 and 0630 UTC as the storm reached the coast. The ANC correctly dissipated the system but was only able to give a 40-min lead time. Nimrod only slightly decreased the echo intensity, which it did for the entire lifetime of the storm.
This is the one case where GANDOLF did not completely dissipate all the echoes by 60 min. However, there was still so much dissipation that it was not possible to confirm if the storm of interest was present.
e. 30 November 2000
The primary feature of interest for this case was the initiation of a large area of convection between 0338 and 0438 UTC (Fig. 11); this is also shown by the white contour near cell B in Fig. 12. The reflectivity image is for 0338 UTC and the white contours outline the area of >35 dBZ at 0438. Also shown are the convergence lines at 0338 UTC (yellow). The majority of the initiation took place in the region of the boundary collisions; there are four clusters of cells at 0338 UTC labeled A, B, C, and D in Fig. 12. Cells A, C, and D all dissipated within the hour. Figure 13 shows the extrapolation vector for each system for all four cells; Nimrod was not available for this case since it occurred after the official end of the FDP. Cell A dissipated at 0410 UTC, then initiation took place in the general vicinity as the boundary from the southwest intersected existing small echoes. All the systems incorrectly nowcast cell A's motion and lifetime. Interestingly, although TITAN produced an incorrect extrapolation forecast, a new echo initiated where cell A was extrapolated, giving a hit for the wrong reason. Cell B was the only cell that lived for 60 min, and it merged into the large area of initiation. All systems had the wrong extrapolation for this cell. The motion of cell B was influenced by the strong convergence center that developed as the three boundaries approached each other, in common with the 3 November case. Cell C dissipated by 0345 UTC, which was correctly nowcast by the ANC. TITAN never identified it for extrapolation. Cell D dissipated at 0415 UTC. The ANC significantly decreased the size of the cell but did not dissipate it completely. All systems extrapolated cell D in about the same direction and with similar accuracy.
The 0438 UTC nowcast produced by ANC initiated and then grew a large area of storms that overlapped with the actual new area (although it was centered farther to the east). This initiation nowcast is discussed below in section 5a. The misplacement of the initiation is similar to the 3 November problem, where extrapolation moved the storms too rapidly away, from the center of the convergence. Information concerning the magnitude of the convergence was available from VDRAS, and the location of boundary collisions was also available to the ANC; thus, this problem could theoretically be remedied with some modified ANC logic.
In summary, only storm B survived for 60 min, and it underwent major growth. The other three cells dissipated. All the systems incorrectly nowcast the motion of storm B. The ANC did nowcast extensive growth and initiation, but positioning errors were 10–30 km. In addition, the ANC correctly dissipated two of the three storms.
f. 1 December 2000
Figure 14b shows three cells that developed rapidly 10–30 min prior to the analysis, near the eastern edge of the west slope of the Blue Mountains. There were at least 19 storms that initiated in this area between 0300 and 0700 UTC, as shown in Fig. 15. The sea breeze advanced through this area about 0300 UTC, and the sounding taken at 0300 UTC at the airport in the sea-breeze air showed that the low-level flow from the south and southeast was very stable, while the air above 1500 m was unstable and moving from the west. The heights of the Blue Mountains are about 1000–1500 m. It is theorized that the upper-level westerly winds were being lifted over the relatively cool sea-breeze air and initiating the storms. The only systems available for this case were ANC, S-PROG, and TITAN, among which only the ANC had the capability to forecast the initiation of storms. To initiate a storm, the ANC required at least the radar detection of cumulus clouds near the boundary. However, since the sea breeze was moving west, it was soon out of the area where the initiation took place, and the ANC had no other logic available for initiating storms. The ANC did provide 5–10-min lead times for some of the storms, but that was only because reflectivity greater than 45 dBZ was detected aloft. All three systems often had difficulty extrapolating the storms for the first few time periods, because the movement was slow and erratic during the formation process. Once the storms were fully formed they moved east and dissipated within 1–2 h over the ocean. Thus, since initiation was not predicted and the extrapolations were initially erratic, there was very little useful information provided by the any of the systems prior to the storms moving over the ocean. Again, the importance of nowcasting initiation is apparent.
TITAN and the ANC tended to have the lowest displacement errors, and S-PROG and Nimrod the highest. This was because the storms of interest tended to have motion vectors different than the surrounding storms; thus, the importance of allowing for differential motion vectors is demonstrated. This feature was not in place for S-PROG during the FDP, although it has since been implemented. Nimrod had the coarsest resolution—5 km—and it is likely that this caused the motion of the smaller-scale, high-impact features to be smoothed with the motion of the larger-scale features. Nimrod was designed for stratiform rain, hence, the relative coarseness of the grid. Therefore, for convective storm nowcasting, it is necessary for extrapolation systems to allow for differential motion and to use a grid resolution of 1–2 km.
The ANC tended to have little bias in the rainfall-rate or reflectivity nowcasts, whereas Nimrod and S-PROG5 routinely underestimated it. This suggests that both S-PROG and Nimrod were smoothing the data more than necessary. The ANC was the only system that showed any success in nowcasting initiation, growth, and decay, with the decay nowcasts showing the greatest skill. The success that the ANC had in nowcasting storm evolution is likely to be linked to its ability to utilize knowledge of convergence line locations in the nowcasts. Therefore, in the following section, data from the same six cases are used to show that if boundary layer convergence lines are identified and characterized, then improvement over extrapolation can be obtained.
5. Experimental results with the ANC
In this section, ANC results are presented based on reruns made after the FDP. For the reruns a complete set of human-determined convergence lines were entered, and additional experiments were conducted with different forecast parameters. Identification of boundaries is critical to activating the portions of the ANC that will cause initiation and growth. Note that the ANC contains an automated boundary detection algorithm; however, this was judged unreliable prior to the Sydney FDP, and therefore a human boundary insertion method was installed. Boundaries were entered in real time for only 3 of the 6 days, as the 30 November and 1 December cases occurred after the FDP, and half of the 26 September case occurred after forecasters left for the day, hence, the reason for reentering the boundaries for this exercise.
Table 6 lists the predictors that were used to activate storm initiation, growth, and dissipation within the ANC (the table includes references where the variable has been previously discussed). All of the predictors except growth rate are only activated in the vicinity of convergence lines, and nowcasts are based on the collocation of the various predictors. Thus, if appropriate predictor fields are aligned in space (after extrapolation), a nowcast for storm initiation or growth is issued. Storm growth rate is used to dissipate storms in the absence of boundaries. Details describing the ANC are given by Mueller et al. (2003) and therefore are not presented here.
Figure 16 is an example of a 60-min reflectivity nowcast that includes initiation, growth, and dissipation of the original reflectivity field. This is for the approximate time period shown in Figs. 11c and 11d and Fig. 12 when three boundaries were colliding with each other. The large brown and yellow areas in Fig. 16b represent 60-min nowcasts of storm initiation for reflectivities of 35 and 44 dBZ, respectively. Initiation and growth were based on the parameters boundary relative steering flow (Ubsf), boundary collision (BC), VDRAS vertical velocities (MaxW), and boundary–storm collision (B/S), listed in Table 6. All these parameters require the presence of at least one boundary. The storm to the south was correctly nowcast to dissipate as a result of negative growth rate and the absence of a boundary. Comparison of nowcasts in Fig. 16a (extrapolation only) and Fig. 16b (ANC) with Fig. 16c (verification) makes it apparent that the ANC dissipation and initiation nowcasts were superior to the extrapolation-only nowcasts. The nowcast labeled growth in Fig. 16b was wrong because the storm was extrapolated too far to the south.
Figure 17 shows a time series plot of the critical success index (CSI) (Donaldson et al. 1975) for four of the six cases for both the ANC forecasts (dark line) and extrapolation-only forecasts (light line). Results are not provided for 25 September and 1 December since the boundaries on these days did not affect the ANC nowcasts, which were therefore the same as extrapolation forecasts. The nowcast parameters are the same as those used in real time during the FDP. The verification statistics used here are based on the use of a 1-km grid. No credit is given even if the nowcast is off by just one grid point.
First it is apparent from Fig. 17 that there is considerable falloff in the statistical scores from 30 to 60 min (right versus left column). It is clear that the ability to make precise 60-min nowcasts with 1-km accuracy is very low. However, there was considerable skill in nowcasting the general area in which the initiation and growth occurred. Second the ANC is superior to extrapolation.
Improvement over extrapolation was most evident during the 30 November case. This case is used below to demonstrate in greater detail the reasons for improvement and insight into future avenues for improvement.
a. 30 November
Figure 18 shows that the ANC, utilizing the FDP nowcasting parameters, is nearly always superior to the extrapolation-only nowcasts. This is also true for the probability of detection (POD) and false alarm ratio (FAR) statistics (not shown). As mentioned above there are several nowcast parameters that make it possible to initiate and grow storms. Figure 18 shows a subjective determination of when each of these variables was contributing to the nowcast. It is not possible to separate out what variables are the most important or least important since they contribute with different magnitudes at different times. A third technique called experimental (dashed gray) has been added to Fig. 18. Following the FDP, modified nowcast parameters have been tested particularly for the 60-min nowcasts. The parameters used to initiate storms in Sydney were very dependent on the detection of nonprecipitating cumulus clouds by radar. For a 60-min nowcast such clouds are often not yet present. Thus it would be expected that as the nowcast period increases the importance of boundary characteristics will increase.
For the original FDP nowcasts, the presence of radar-observed cumulus (Cu) near a convergence line with a favorable Ubsf was given the most weight. Boundary collision received the second most weight and MaxW the least. The nowcast labeled experimental in Fig. 18 did not make use of any cumulus information for storm initiation, rather weight was given about equally to Ubsf, MaxW, and BC. As Fig. 18 shows, the nowcasts for the “experimental” run are superior indicating the importance of boundary characteristics at 60 min. These same nowcast parameters were tested on the 60-min nowcasts for the other three cases; the result was significant improvement for 3 November, no difference for 19 October, and slightly worse for 26 September. This suggests that for nowcast periods ≥60 min, boundary characteristics take on more importance for storm initiation nowcasts than does early detection of cumulus clouds.
b. All cases
Table 7 summarizes the contribution of the ANC variables given in Table 6 to the improvement in the 60-min nowcasts, compared to extrapolation only, for the four cases presented in Fig. 17. The contribution of the various variables was manually determined by overlaying the magnitude of each forecast variable field on the reflectivity field at validation. Table 7 shows that boundary characteristics played a primary role in contributing to the improved nowcasts while the presence of cumulus clouds and storm growth rate contributed little to the nowcast. Growth rate and radar- or satellite-detected cumulus increase in importance for the 30-min nowcast. Analysis not shown found that for the 30 November case the VDRAS vertical velocities contributed the most to the improvement in the nowcasts.
The analysis here indicates that nowcasting of high-impact convective storm events requires an extrapolation system that provides for differential storm motion and a grid resolution of about 1–2 km. This is because high-impact events often have motions different than surrounding storms. Cell trackers are likely to have a slight advantage over area trackers for these events.
The accuracy of nowcasts even for periods ≤60 min is quite low. This is because of inaccuracies in the extrapolations and even more so because of storm initiation, growth, and dissipation. The primary finding from this paper is that skill above extrapolation occurs when boundary layer convergence lines can be identified and utilized by a nowcasting system to nowcast storm evolution. The nowcasting of precise details of storm location and intensity (i.e., within a few kilometers and few minutes) are very low and we believe are not likely to improve significantly in the near future. Supercells and large squall lines are possible exceptions during period when they have reached a steady state. However, there is skill in nowcasting the general area and intensity of storms. For example Fig. 19 shows a 60-min nowcast verifying at 0509 UTC on 30 November. Note from Fig. 18 that the critical success index, verified on a 1-km grid, for this nowcast at this time is quite low at 0.20. However, irrespective of the skill score the nowcast at this time provides considerable skill in outlining the area of strongest convection and would provide useful information to forecasters and aviation users.
In summary there is demonstrated capability that 0–1-h nowcasts of convective storms can be made that improve over extrapolation. However, this capability is in its infancy and there are numerous questions related to basic meteorological understanding, improved observations, algorithm improvement, product development, forecaster involvement, user participation, and verification procedures yet to be answered. Some of these questions are investigated in the next section.
We choose to speculate on what research and approaches might be necessary in the future to improve convective storm nowcasting particularly with the user in mind. Users would include aviation, agriculture, construction, emergency managers, the outdoor entertainment industry, and the transportation industry. In the United States with the exception of severe storm warnings only very limited nowcasting products are disseminated to users by the National Weather Service. The Federal Aviation Administration is sponsoring a progressive program under its Aviation Weather Research Program to produce convective storm nowcasts for aviation terminal and en route operations. The Met Office provides rainfall nowcasts to flood authorities. With these few exceptions and perhaps others there seems little emphasis among government meteorological agencies world wide in issuing nowcasts to targeted users. No doubt there are 0–2-h nowcast technologies available today that, if implemented, could greatly profit certain users, providing the appropriate products were developed.
We imagine that in the foreseeable future that convective storm nowcasts for the 0–6-h period will be based on an expert system approach that combines numerical prediction with observational and statistical techniques. The specificity of the nowcasts will decrease with increasing lead times and be dependent on the mechanisms forcing the convection. For example, the specificity or probabilities associated with a squall line, forced by a synoptic cold front, will be much higher than those for isolated convective storms in a nonsynoptically forced environment where individual gust fronts are the only forcing mechanism. Nowcasting is complicated by the fact that multiple-scale forcing mechanisms are often operating at the same time. This is evident from time lapse images of radar and satellite data.
The predictability of convection is partially dependent on the predictability of the forcing or triggering mechanism. A major limitation to convective storm predictability is the inability to predict the emergence of individual gust fronts and their characteristics. The problem is particularly acute considering this must be done for convective storms that form after the nowcast has been issued, that is, nowcasting gust fronts from secondary convection. No known technique will be able to do this in the foreseeable future. Thus nowcasts beyond ∼1–2 h will not be able to be specific to a particular storm. Fortunately long-lived (>2 h) supercell storms are sometimes exceptions.
In recognition of the limitations to precise prediction of convective storms for periods ≥60 min, the detail of the nowcast needs to be degraded and the use of probabilities needs to be considered. Figure 20 is a proposed example of what a nowcast of the future might look like. The area of nowcast storms would be outlined and coded by the expected density of storms and severity. Roberts et al. (2001) have already demonstrated some success in anticipating storm severity by using the NSSL Warning Decision Support System (WDSS) and the ANC.
We speculate that future nowcasting systems will utilize the following:
High-resolution boundary layer wind analyses that include individual thunderstorm outflows up to synoptic-scale fronts. These analyses will be based on a combination of wind retrievals from radar and surface stations similar to those produced by VDRAS that are shown in Fig. 11 but covering a network of radars. Short period nowcasts (1–3 h) of the wind field utilizing the VDRAS numerical model are under development (Crook and Sun 2004) that may be very beneficial in nowcasting the evolution of convergence line intensity and depth.
Higher-resolution stability fields based on improved water vapor measurements. A variety of new water vapor measurement techniques tested during the International H2O Project (Weckwerth et al. 2004) will be examined to determine if they have a near-term operational potential. These include (a) water vapor retrievals from radar refractivity measurements (Fabry et al. 1997), (b) satellite water vapor retrievals, (c) low-cost vertical pointing differential absorption lidar (DIAL), and (d) radiometers.
Numerical prediction of synoptic and mesoscale parameters such as convergence, convective available potential energy (CAPE), and convective inhibition (CIN). Presently Megenhardt et al. (2000) make use of equivalent potential temperature gradients, vorticity, and convergence diagnosis from the Rapid Update Cycle model to identify frontal regions that are used in initiating and growing storms for a national 2-h convective storm nowcasting product. As the model resolutions and data assimilations improve, the utility of diagnostics and predictions from operational models should improve.
Presently research is very limited into the fundamental questions related to improving convective storm nowcasting and defining the limits of predictability; this research needs to be expanded. Topics include triggering of elevated convection not associated with forcing by boundary layer convergence, observation and evolution of small-scale variations in boundary layer water vapor, interaction of dynamic triggering features on multiple scales within and above the boundary layer, and the detection, evolution, and prediction of boundary layer convergence lines.
Progress in accelerating the development of improved convective storm nowcasting systems and accelerating the transfer to operations would benefit from the development of regional test beds with access to unique research and observational data, archived data, and involving users as partners. These test beds would be follow-ons to the Sydney 2000 FDP effort where experimental products were sent to targeted users (Anderson-Berry et al. 2004, in this issue). There is a particular need to determine how best to represent probability nowcasts to users. The test beds would serve as vehicles to evaluate new techniques and products of benefit to end users, for training of forecasters with forecasters as partners, and to serve as a pathway to operations.
We would particularly like to thank Terry Bentancourt of NCAR for her exceptional efforts in developing new algorithms for the ANC in time for the FDP and installing the ANC system in Sydney. We thank Dan Megenhardt for final preparation of the figures. The NCAR effort was funded by the USWRP and NCAR base NSF funds.
Corresponding author address: Dr. James Wilson, National Center for Atmospheric Research, P.O. Box 3000, Boulder, CO 80307-3000. Email: firstname.lastname@example.org
A sixth system, the Storm Cell Identification and Tracking (SCIT) algorithm (Johnson et al. 1998), developed and operated by the National Severe Storms Laboratory (NSSL), also produced extrapolation nowcasts of individual cells, but these were not saved in a format that could be easily evaluated.
All times in this paper are in UTC. Local time was 12 h ahead.
The Ebert and McBride (2000) common radar area technique gave equivalent results for the same storm features.
Note that 35 dBZ was used as the threshold for a storm. Once a cell dropped below this threshold it was classified as having dissipated.
Since the FDP, a new scheme has been implemented to maintain the average intensity of the rainfall.