Forecasting the onset of cloud-to-ground lightning using radar and upper-air data in Romania

Authors


Abstract

A method for forecasting the onset of cloud-to-ground (CG) lightning flashes is examined. Reflectivity data and upper-air measurements were analysed for 49 convective storms observed in 36 different days over southern Romania during the convective season (May–September) from 2003 to 2005. The radar reflectivity data were associated with the CG lightning flash locations obtained from the Romanian National Lightning Detection Network. The CG lightning initiation signature is based on reflectivity thresholds (35 and 40 dBZ) at a given environmental temperature aloft (−10 and − 15 °C). We have found that the best predictor is the 35 dBZ radar reflectivity at the height of the − 10 °C isotherm, with a probability of detection of 95%, and mean lead time of 17 min before the first CG lightning. In previous studies, it was found that the onset of CG lightning flashes had greater threshold value for reflectivity (40 dBZ) and shorter lead time (13 min). The results presented in this study are currently used in the nowcasting operational activities at the Romanian National Meteorological Administration. Copyright © 2012 Royal Meteorological Society

1. Introduction

During the warm season (May–September) cloud-to-ground (CG) lightning flashes are produced in Romania mainly by airmass thunderstorms, or by thunderstorms initiated along or near surface airstreams boundaries, especially in southern Romanian (Stan-Sion and Antonescu, 2006). The purpose of this study is to examine a method for predicting the time of the first CG lightning flash associated with convective storms developing in the southern part of Romania, by using radar reflectivity and upper-air temperatures. This forecasting method can be useful for issuing severe weather warning, both for protecting human lives and for lightning sensitive operations, so that precaution may be taken. Because aircraft traffic and airport operations are affected by thunderstorms, this information would also be useful for local aviation administrations. Although there are previous studies on using the radar reflectivity to predict the onset of CG lightning (Table I), this is the first time, in author's knowledge, when the combined data from S-band and C-band radars are used to forecast the onset of CG lightning flashes over Eastern Europe.

Table I. Previous studies that have used the radar reflectivity threshold at a given environmental temperature aloft in order to predict the onset of cloud-to-ground lightning activity
StudyRegionPeriod of studyNumber of cases (number of days)CriteriaLead time (min)PODFARCSI
Dye et al. (1989)New Mexico (USA)July–August 198620 (18)40 dBZImmediately
    − 10 °C    
Buechler and Goodman (1990)Florida (USA)198615 (–)40 dBZ4–331.000.200.80
    − 10 °C    
Michimoto (1990)Hokuriku (Japan)December 198730 dBZ5
  August 1988 − 20 °C    
  January–March 1988–1990      
Hondl and Eilts (1994)Florida (USA)August 199028 (8)10 dBZ20
    0 °C    
Gremillion and Orville (1999)Florida (USA)May–September 1992–199739 (38)40 dBZ7.50.840.700.79
    − 10 °C    
Vincent et al. (2003)North Carolina (USA)2001–200250 (13)40 dBZ14.71.000.370.63
    − 10 °C    
Present studyRomaniaMay–September 2003–200549 (36)35 dBZ170.950.100.85
    − 10 °C    

The Romanian National Meteorological Administration acquired in 2002 (Ioana et al., 2004) eight Doppler weather radars, five of which are Weather Surveillance Radar-98 Doppler S-band systems, and three are C-band units. In this study we have used this radar network to collect data from thunderstorms which developed in the southern Romanian Plain, during the convective season (May–September) from 2003 to 2005. This study was developed for the Bucharest city, an area with two international airports, and a population of 2 million people, abundant with various outdoors activities. The study area is topographically homogeneous and the three Doppler radars covering this region have no beam blockage at any of the elevation angles (Figure 1).

Figure 1.

The Romanian National Doppler Weather Radar Network—composed from five Weather Surveillance Radar-98 Doppler S-band systems located at Timisoara, Oradea, Barnova and Medgidia, and three C-band units operational at Bucharest, Craiova and Oradea (not shown on this map). The figure also shows the Romanian National Detection Network (RNLDN) and the study area (black box)

For understanding how radar reflectivity can be used to forecast the first CG lightning flash, in Section 2 we examine the main mechanism for storm electrification. Section 3 presents the data set and methodology and Section 4 contains the results. The conclusions are presented in Section 5.

2. Previous studies

In their review chapter on charging mechanisms, MacGorman and Rust (1998) stated that in order to produce electrical charges inside the cloud, and thus electrical discharges, a strong updraft associated with vigorous convection growth should occur in the lower part of the mixed-phased region of the cloud. The mixed-phase region of the cloud, situated approximately between − 10 and − 40 °C isotherm, is the region where supercooled liquid water and ice coexist. This is a favourable environment for a non-inductive charging mechanism that involves rebounding collisions between cloud ice particles and riming graupel in the presence of supercooled liquid water. The non-inductive graupel-ice charging mechanism is the only one that laboratory and modeling studies have indicated it capable of causing clouds to be electrified enough to create CG lightning flashes (Saunders, 1993; MacGorman and Rust, 1998). The observed dependence of the non-inductive charging mechanism on environmental parameters (Takahashi, 1978; Saunders and Peck, 1998) appears to explain why some storms are thunderstorms and others are not. Through laboratory experiments Takahashi (1978) found that the graupel-ice mechanism depends on the temperature and liquid water content. For typical condition, low water content on the order of 1 g m−3 and temperature colder than − 10 °C, the graupel tend to be negatively charged and the small ice crystals became positive charged. The positively charged crystals are lifted in the upper regions of the storm, and the negatively charged graupel is suspended in the updraft in the mid-levels of the storm. MacGorman et al. (1989) suggested that this mechanism operates at the 7–9 km level, where the temperature of the environment is − 20 to − 40 °C and there are ice crystals in sufficient quantities (Michimoto, 1990). Polarimetric techniques showed a correlation between radar reflectivity values and various precipitations types (Doviak and Zrnic, 2006). Thus, the appearance of 30–40 dBZ radar echoes between − 10 and − 15 °C indicates the possible presence of graupel or hail suspended in an updraft containing ice crystals.

Bright et al. (2005) indicated that three ingredients are necessary for CG lightning flash production: (1) for ensuring the presence of supercooled liquid water the lifted condensation level must be warmer than − 10 °C, considering that in general ice begins to nucleate at temperature colder than − 5 to − 10 °C; (2) in order to have ice nucleation the equilibrium level must be colder than − 20 °C; (3) convective available potential energy (CAPE) must be greater than 100–200 J kg−1 in the layer between 0 and − 20 °C in order to have sufficient vertical motion in the mixed-phase region. Similar ingredients for CG lightning flashes were used by van der Broeke et al. (2005). They agree with the first two ingredients for lightning production proposed by Bright et al. (2005), but they stated that CAPE should be present in the lower mixed-phase region of the cloud between − 10 and − 20 °C.

Previous studies, as far back to 1970s, have used the capability of weather radar to forecast the onset of CG lightning flashes. One of the first studies on using radar data to forecast and detect the lightning flashes was the study by Larsen and Stansbury (1974). They generated radar maps of precipitation with a time step of 10 min, at 7 km height, outlining the regions exceeding 43 dBZ. They showed that this region (Larsen area) was the source of lightning flashes throughout 90% of the life time of the thunderstorms, and that they could account for 75% of the flashes observed. In a subsequent study, Marshall and Radhakant (1978) used radar maps at the height of 6 km (approximately − 17 °C) and at 5-min intervals, with 4 dBZ outlines intervals between 30 and 58 dBZ. They showed that ‘Larsen area’ within the 38 dBZ outline is a representation of the convectively active regions of the thunderstorms.

Another approach to forecast the CG lightning flashes is based on the outputs of the numerical weather prediction (NWP) models (Mazany et al., 2002; Burrows et al., 2005). Burrows et al. (2005) developed statistical models based on tree-structured regression for predicting the probability of lightning, using lightning data from the North American Lightning Detection Network and predictors (e.g. CAPE, convective inhibition, Showalter, precipitable water, 700 hPa vertical motion, 500–1000 hPa layer thickness) derived from Global Environmental Multiscale model output at the Canadian Meteorological Center. Real-time forecast in 3 h intervals for lightning probability was made in 2003 and 2004, and results demonstrate that tree-structured regression is a viable method for building statistical probability forecast models. The main issues with their approach is that this can be used for finding regions favourable for thunderstorm formation, and can be used just as guideline for nowcasting and short range forecast. Shafer and Fuelberg (2008) also use statistical schemes to forecast warm-season lightning over Florida (USA). Based on analysis data from Rapid Update Cycle and lightning data form National Lightning Detection Network, and using a perfect prognosis technique, Shafer and Fuelberg (2008) developed a gridded forecast guidance product for CG lightning activity. They also classified the amounts of CG lightning in two categories, with one or more CG flashes, using binary logistic regression. The scheme is applied to output from three mesoscale models during an independent test period, and the results showed that although the temporal and spatial forecast lightning was not perfect, there was a ‘generally good agreement between the forecasts and their verification, with most of the observed lightnings occurring within the higher forecast probability contours’ (Shafer and Fuelberg, 2008). McCaul et al. (2009) have proposed and developed two methods for quantitative short-range forecast of lightning threat based in the ice-phase hydrometeor fields from regional cloud-resolving numerical simulations. The first method is based on the upward fluxes of precipitating hydrometeors at − 15 °C level. The second method is based on the vertically integrated amount of hydrometeors for each model grid column. Because the first method is able to represent better the temporal variability of lightning threat, and the second method is representing more accurate the areal coverage of the threat, McCaul et al. (2009) have proposed a ‘blended’ forecast method. This ‘blended’ forecast method was designed to retain most of the temporal sensitivity of the first method, while adding the improved spatial coverage of the second. As was noted by McCaul et al. (2009), the main issues associated with their forecast methods are represented by: (1) imperfect location and timing of convective storms from the individual simulations and (2) NWP output needs to be recalibrated against observed lightning data for every change of the model grid mesh or parameterization schemes.

More recent studies (Table I) show that an evidence exists concerning the correlation between radar reflectivity factor at certain altitude where the temperature assumes typical values and time lapse before the occurrence of the first CG lightning flash.

These methods, besides the fact that are straightforward and easy to implement in operational environments, represent an alternative to the use of complex schemes of NWP models. Also, it can be shown that these nowcasting methods for CG lightning forecasting based on radar data, eliminate some problems associated with the forecast based on NWP data, e.g. location and timing of the lightning activity. But, similar with NWP forecast methods based on sets of convective parameters that have different values for different regions, the methods based on radar data can be regional dependent. With one exception, Michimoto (1990), all the previous studies from Table I have been carried out in the United States. The majority of the previous studies have found similar results for criteria for the onset of CG lightning activity, but with different lead times ranging from immediately up to 33 min.

3. Data and methods

3.1. Radar data

The Romanian National Radar Network (RNRN) includes five S-band radars and three C-band radars. The S-band radars from RNRN are based on the technology and meteorological algorithms developed over more than 30 years in the USA NEXRAD network. The radar data used in this study were obtained from Bucharest, Medgidia and Craiova radars (Figure 1) using the VCP21 (Volume Coverage Pattern) scan strategy.

VCP21 is a standard precipitation scan and it is used when the storms are situated at a certain distance from radar to avoid sampling errors when the storms are near the radar. The radar performs a volume scan using nine sweeps at different tilts (0.5°, 1.45°, 2.4°, 3.35°, 4.3°, 6.0°, 9.9°, 14.6°, 19.5°) every 6 min, and has low vertical coverage for areas close to the radar. Although this scan strategy could potentially miss radar observation in the mixed-phase region close to the radar site, the operational forecasters use VCP21 due to the short time between observation and the availability of data. One application that integrates the data from the two types of radars (S-band and C-band) is called Principal User Processor and was developed for real-time interpretation of radar data in nowcasting and forecasting environments. Each storm has been observed either with the C-band radars from Bucharest and Craiova or with the S-band radar operational at Medgidia according to the distance criteria (i.e. the closest radar).

3.2. CG lightning data

One of the new networks installed in 2002 in the frame of National Integrated Meteorological System (SIMIN, Romania) (Ioana et al., 2004) is the Romanian National Lightning Detection Network (RNLDN). The RNLDN consist of eight SAFIR3000 Total Lightning Automatic Detection Stations manufactured by Vaisala Oyj (Helsinki, Finland), and provides national coverage, with a horizontal accuracy lower than 1 km for the CG lightning flashes. The detection stations perform two types of detection, one being based on the very high frequency interferometry, to obtain accurate angular localization of intra-cloud and CG lightning flashes. The second type of detection is provided by a wide-band low frequency electric-field sensor, and the data from this sensor are used to obtain the characteristics of CG lightning flashes.

In this paper, we use only the CG lightning data, which are detected with a detection efficiency estimated by the manufacturer at 90%. During the study period (2003–2005) the working parameters of the RNLDN were constant, therefore no corrections were applied for the detection efficiency of the CG lightning data. The CG lightning data used in this study were recorded by the RNLDN within the box 43.58°—45.00°N and 22.83°—28.08°E (Figure 1), situated in a plain region in the southern part of Romania. The main reason for choosing this area was to develop a forecast technique to monitor the thunderstorms developing near Bucharest city.

3.3. Forecasting the onset of CG lightning flashes

On the basis of the non-inductive charging mechanism proposed by Takahashi (1978) several criteria have been proposed in previous studies to examine the onset of CG lightning activity. In this study, four criteria for the onset CG lightning activity were tested. These criteria are based on the existence of radar reflectivity factor with values of 35–40 dBZ, showing the possible presence of graupel suspended in the updraft, in a region situated between − 10 and − 15 °C. The analysed sample consists of 49 isolated storm events from 36 lightning days, storms detected at more than 30 km from the radar.

In order to identify the individual storm cells, the Storm Cell Identification and Tracking algorithm (Johnson et al., 1998) was used. After the storm identification, the lightning data from RNLDN were overlaid with the storm track. This has been done to verify if a storm produced lightning and to identify the first CG lightning flash. The storms were post-analyzed using radar software environments and a similar procedure with that described by Vincent et al. (2003). This procedure was used because it can be easily applied during real-time operational context. A four-panel plan position indicator (PPI) display of the reflectivity factor at the lowest four elevation angles (0.5°, 1.45°, 2.4°, 3.35°) was used to visualize the radar data. For a specific sample, the values of radar reflectivity factor and echo height are read directly from the radar display. The values of isotherm heights were obtained from the Bucharest soundings (12 UTC). Only the storms being within 185 km from the sounding release point, and during 6 h period centred at 12 UTC sounding were analysed.

For the cases in which the height of − 10 or − 15 °C isotherm was not directly observed on the PPI elevation angle, cubic spline interpolation was used to obtain the elevation. The spline interpolation is a form of interpolation that allows each segment to have a unique equation while still constraining the curve to fit the data properties (De Boor, 1978). This interpolation is preferred over polynomial interpolation because of the small interpolation errors. In this study, the cubic spline interpolation was performed between the closest elevation angles at which the beam centres are found above and below the level of the − 10 or − 15 °C isotherm.

4. Results

To study the potential use of these criteria for forecasting thunderstorms in the densely populated region of Bucharest, each criterion was tested using a 2 × 2 contingency table, to obtain the probability of detection (POD), the false alarm ratio (FAR) and the critical success index (CSI; Wilks, 2006). The CSI, unlike POD and FAR, takes into account both false alarms and missed events, thus being a more balanced score. Its range is 0–1, where 1 represents a perfect forecast. The best criterion must be valuable from an operational forecast perspective (i.e. long lead time) and in the same time accurate (i.e. high value for POD and CSI, and low values for FAR).

The results show that the best prediction criterion was when the 35 dBZ echo was detected at the level of − 10 °C isotherm (Figure 2), with a POD of 95% and a CSI of 85%. Recent studies (Dye et al., 1989; Buechler and Goodman, 1990; Gremillion and Orville, 1999; Vincent et al., 2003) showed that the best criterion for the onset of CG lightning flashes for United States was 40 dBZ at − 10 °C level. For Japan, Michimoto (1990) showed that the 30 dBZ at − 20 °C level as the best criteria for the CG lightning initiation.

Figure 2.

Results of the statistical analysis for the onset of cloud-to-ground lightning activity in this study; the probability of detection (POD), the false alarm rate (FAR) and the critical success index (CSI) are represented

The results for southern Romania showed that increasing the reflectivity criteria from 35 to 40 dBZ at − 10 °C level results in an increase of the FAR (12.5%) and a decrease of the POD (83.3%) and CSI (79%) (Figure 2). Increasing the isotherm level from − 10 to − 15 °C has less impact on the statistical scores. Thus, the 35 dBZ at − 15 °C was the second best criterion (Figure 2) based on POD (90%) and CSI (80%).

Vincent et al. (2003) used a third variable, namely the persistence of criteria for two consecutive volume scans, resulting eight sets of criteria. The use of a third criterion could result in a decrease of the lead time between detection of the thunderstorm initiation signature and the first CG lightning flash, but it may also help reduce FAR. Two consecutive volume scans criteria were also tested in the present study. Applying this criterion for the 35 dBZ at − 10 °C level results in lower values of POD (88.6%) and higher values of FAR (11.3%) in comparison with the single scan criterion, and a lead time of 10 min.

In this study the mean lead time obtained was 17 min for 35 dBZ at − 10 °C criteria and 13 min for 40 dBZ at − 10 °C (Figure 3). For 35 dBZ at − 15 °C, the second best criterion based on FAR and CSI, the lead time was 12 min. These findings are consistent with the results obtained by Vincent et al. (2003) for North Carolina (USA).

Figure 3.

Results of the statistical analysis for the mean time interval between the lightning initiation signature and the time of the first CG lightning flash

5. Conclusion

The analysis of radar data for 49 thunderstorms shows that the radar reflectivity and the sounding data can be used to forecast the onset of CG lightning activity. In this study, the best predictor for the onset of CG lightning activity was the 35 dBZ radar reflectivity at the level of − 10 °C isotherm, with a POD equal to 95%, a FAR of 10% and CSI of 85%. The mean lead time was 17 min. Gremillion and Orville (1999) found that the best criterion for CG was the 40 dBZ echo detected at the − 10 °C level, with an average lead time of 7.5 min (Table I). They also noted that the 35 dBZ radar echo at the − 10 °C isothermal level was the second best predictor, with POD of 88% and FAR of 20%. In this study was found that the best criterion has a lower radar reflectivity threshold than those found in previous studies, and this would act by increasing the lead time. Contrary to the expectation, the increase of lead times has not increased FAR. Vincent et al. (2003) found that the best predictor of CG lightning, based on CSI, was the presence of a 40 dBZ echo at the − 10 °C, with 37% FAR, 100% POD, 63% CSI, and an average lead time of 14.7 min (Table I). Despite several potential sources of error (e.g. detection efficiency of lightning detection network, radar detection errors, and interpolation between different radar elevation angles to find the proper height of isotherms) the results from this study compares well with previous studies.

Ancillary