1. Introduction
The new generation of geostationary (GEO) scientific satellite missions brings lightning locating systems (LLSs) to GEO orbit. Somewhat similar sensors have been used on low-Earth-orbit satellites for several decades, namely, the Optical Transient Detector (OTD) (Boccippio et al. 2000) and the Lightning Imaging Sensor (LIS) on the Tropical Rainfall Measuring Mission (TRMM) satellite and on the International Space Station (ISS) (e.g., Blakeslee et al. 2020; Christian et al. 1999). Currently, the Geostationary Operational Environmental Satellites (GOES) GOES-16, GOES-17, and GOES-18 carry Geostationary Lightning Mappers (GLMs; Goodman et al. 2013) providing almost full-disk lightning observations over the Americas and adjacent oceans, and the Chinese Fengyun-4 satellite carries a lightning mapping imager (Yang et al. 2017) providing observations over China and its surroundings. In December 2022, the Meteosat Third Generation (MTG) Lightning Imager (LI) was successfully put in orbit to cover Europe, Africa, and adjacent regions (Dobber and Grandell 2014). These GEO satellite sensors feature high-frequency optical cameras to monitor and record cloud-top illuminations associated with (total) lightning activity including cloud-to-ground (CG) as well as inter- and intracloud (IC) lightning.
Total lightning information can play a key role when predicting thunderstorms and related occurrences of potentially harmful and severe weather events. The U.S. National Weather Service (NWS) defines severe weather as the presence of one or more of the three following phenomena: (i) hail at the ground of 2.54 cm or more in diameter, (ii) winds of at least 25 m s−1, and (iii) a tornado. In addition, heavy rain usually accompanies severe thunderstorms and may cause dangerous flash floods. Former studies linked total lightning activity and severe weather events (e.g., Williams et al. 1999; Goodman et al. 1988). Schultz et al. (2011) correlated lightning discharges and severe weather events. They demonstrated that the use of total lightning records yields a higher probability of detection, lower false alarm rate, and longer lead times than using CG strokes alone. Moderate to strong correlations between flash rates and radar-derived storm intensity attributes are described in the climatological study of Herzog et al. (2014). They used three different ground-based Lightning Mapping Arrays (LMAs) in the United States. Schultz et al. (2017) report increases in mixed-phase updraft volume and peak speed, and graupel mass prior to north Alabama LMA total lightning jumps manifested as abrupt increases in the storm’s flash rate. The term “lightning jump” (LJ) and its operational utility in severe weather nowcasting were first coined and studied by Williams et al. (1999). At least one LJ per hour is likely to be found in severe storms and can be indicative of an enhanced probability for tornadic development (Steiger et al. 2007a,b; Rudlosky and Fuelberg 2013).
Previous studies investigated means to identify and verify LJs. Schultz et al. (2009, 2011, 2016) correlated LJs to severe weather reports and studied the performance of different automated LJ algorithms on GLM proxy data,1 respectively. Based on the introduction of an automated LJ algorithm in Gatlin (2007) and Gatlin and Goodman (2010), they define the 2σ LJ algorithm that was adapted by multiple other authors on different continents. For example, Rudlosky and Fuelberg (2013) applied the 2σ LJ algorithm in the United States and found statistically more LJs in severe than in nonsevere storms (1.44 vs 0.92 LJ per hour). Chronis et al. (2015) tested this LJ algorithm with multiple activation and σ thresholds and showed that storm clusters with LJs are more organized, more intense, last longer, and exhibit more consistent lightning activity than storms without LJs. Stough et al. (2017) related 2σ LJs to supercell development. They found that supercell mesocyclone development and intensification are often correlated to LJ occurrences. In Europe, a primary interest was the potential of LJs as hail precursors. Mikuš Jurković et al. (2015) reported a sharp peak in total lightning prior to the occurrence of hail at ground. Wapler (2017) modified the 2σ LJ algorithm and introduced the σ level as LJ intensity measure. In their study, LJs often occur before severe hail, with a low LJ intensity of 0.1, more than 80% of the hail events are preceded by an LJ, and almost 100% of the hail events occur in relation to an LJ. These ratios decrease with a higher LJ intensity [see Fig. 6 of Wapler (2017)]. Using a modified σ LJ algorithm, Nisi et al. (2020) found that LJs aid discriminating ordinary and hail-producing thunderstorms. LJs had limited potential as a precursor of hail events. Farnell and Rigo (2020) studied the 2σ LJ algorithm and Rigo and Farnell (2022) identified storms with multiple LJs to be more intense and last longer than storms exhibiting a single LJ. Wu et al. (2018) applied a modified 2σ LJ algorithm in the Beijing region and report good skill to use LJs to forecast short duration rainfall, with lead times ranging from 36 to 52 min. The latter studies show that LJs were reported as indicators of storm intensification and potential severe weather. However, the lightning data were collected by ground-based instruments.
To assess the GLM performance relative to a certain ground-based LLS, typically the relative flash detection efficiency (DE) is used. It quantifies the fraction of all flashes observed by the ground-based LLS that is also detected by GLM. Former studies report a geographical pattern of GLM-16 DE. Murphy and Said (2020) and Marchand et al. (2019) found higher DE relative to the National Lightning Detection Network (NLDN), Global Lightning Detection Network (GLD360), and Earth Network Total Lightning Network (ENTLN) over the southern and eastern CONUS than over the northern and western CONUS. Erdmann et al. (2022) confirmed high GLM flash DE relative to NLDN over the southern and southeastern CONUS. The performance difference can be of technical nature, such as instrument characteristics, spatial resolution, and parallax effects (e.g., Bruning et al. 2019), that increase toward the edges of the field of view (FoV). The land–ocean distribution may have an impact on GLM DE as Rudlosky et al. (2018) found that oceanic flashes appear larger and brighter than continental flashes, thus, allowing higher GLM DE. Flash size and duration were identified as critical parameters for GLM flash detection (Marchand et al. 2019; Zhang and Cummins 2020; Erdmann 2020, 98–107) also in accordance with reduced GLM flash DE for high flash rates reported by Murphy and Said (2020). Rutledge et al. (2020) show lower GLM DE for small flashes, and for higher content of cloud water above the flash altitude as the scattered light can be masked by thick clouds. Light can be scattered farther to the cloud edges and stratiform cloud regimes where bright optical radiation can be detected by GLM (Peterson et al. 2020). In general, GLM DE is higher at nighttime than at daytime since the lower background luminosities allow the optical instrument to use lower detection energy thresholds at night.
This study focuses on using GLM lightning data to define LJs. These GEO-detected LJs may not provide exactly the same information as the ground-based counterpart. For example, Curtis et al. (2018) and Murphy and Said (2020) suggest that LMA recorded LJs are of a different nature than GLM-based LJs when compared with each other and with radar parameters. GLM detects lightning in a different way than the LMAs and does not record the same discharge processes even when the same situation and storms are observed. One advantage of the GLMs or equivalent GEO LLS in general constitutes the continuous total lightning detection over a wide domain. GLM resolves single storms and features with high temporal resolution to enable tracking of storms during their entire lifetime and identifying changes in the storm’s lightning activity with time.
The cited studies demonstrate the great potential of using total lightning locating GEO sensors in research in order to identify LJs. The main objective of this study is to determine the most consistent way of automatically detecting LJs in GLM lightning records for nowcasting of severe weather. Section 2 introduces our study cases, the datasets, and methodology. The three types of LJ algorithms used are defined in section 3. In the results (section 4), first a brief statistic about the identified LJs is presented, followed by an assessment of the performances of different LJ algorithms. The paper finishes with a discussion in section 5.
2. Data and methods
This section introduces the adopted cell tracking tool and the methodology to match severe weather reports to LJs associated with cloud cells. Figure 1 summarizes the processing steps starting with the tools and data (dark gray), to the matching of cloud cells and National Centers for Environmental Information (NCEI) reports (light gray), initial data processing (white) to get the LJs and ground-truth events that are analyzed (blue), the matching of LJs and NCEI events (green), the verification of LJs through NCEI events (yellow), and ending in the results (red; section 4).
Workflow chart to illustrate the processing steps until the results.
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
Data were collected for nine study cases, with each case composed of multiple days. The case selection is based on (i) the availability of GLM lightning records, (ii) severe weather occurrence that includes all three report types (i.e., wind, hail, and tornado reports) on at least two consecutive days, (iii) inclusion of cases from different seasons, and (iv) limiting the region to the CONUS and the FoV of GOES-16 and GLM-16. Table 1 presents the selected cases in more detail. Overall, this study includes more than 2.5 million cloud cell trajectories, of which about 48 200 cloud cells showed lightning activity and were, thus, classified as thunderstorms. Only periods with available NCEI reports, Advanced Baseline Imager (ABI) and GLM data are analyzed. This concerns, in particular, the days of 3–4 June 2020 when GOES-16 was offline from 1700 to 0130 UTC. As such, this period is discarded from our analysis.
Study cases. The “No. of analyzed thunderstorm trajectories” column states the number of thunderstorms that remain after dropping the first 3 h of the case [see section 2d(2)]. The last column (No. of analyzed NCEI events) gives the number of NCEI events that were matched to a tracked cloud cell.
a. Satellite data
The GOES-R series constitutes the new generation of U.S. environmental satellites. The main instrument of GOES is the ABI, a passive multichannel imaging radiometer. ABI enables monitoring of various environmental elements on Earth’s surface such as trees, or within the atmosphere like clouds, moisture or smoke. Further, ABI application includes tracking and monitoring of cloud formation, atmospheric motion, convection, land surface temperature, ocean dynamics, flow of water, fire, smoke, volcanic ash plumes, aerosols and air quality, and vegetative health (NASA 2022). Its 16 different spectral bands include two visible channels (at 0.5- and 1.0-km resolution), four near-infrared channels (at 1.0-km resolution), and 10 infrared channels (at 2-km resolution) with on-orbit calibration. This study uses data from the GOES-16 satellite with a position nadir over the equator at 75.2° west, at an altitude of 35.786 km. The focus is on central and eastern CONUS. This study uses the full-disk scan data with 10-min temporal resolution as the default scan mode for GOES-16. The data are downloaded from the AWS archive, limited to the CONUS region, and ingested into the Satellite Application Facility (SAF) for Nowcasting and Very Short Range Forecasting (NWCSAF) software package (see section 2d). The 5-min rapid scan available over the CONUS could in general improve the accuracy of the cloud tracking; however, NWCSAF runs more efficiently with the 10-min update cycle when analyzing time periods of several days as in this study.
b. GLM data
GOES-16, GOES-17, and GOES-18 carry a GLM instrument. GLM functions like an optical sensor as a high-frequency camera. It detects total lightning activity as cloud-top illuminations, however, cannot directly discriminate between IC and CG lightning signals. The narrow 1-nm band sensitivity centered at the 777.4-nm oxygen line in the near-infrared enables the detection of lightning signals during both day- and nighttime. Additional spatial and temporal filtering optimizes GLM lightning detection especially during bright daytime conditions. A wide-FoV lens combined with a narrow-band interference filter is focused on a high-speed charge coupled device (CCD) focal plane with a nearly full-disk FoV coverage (1372 × 1300 pixels). The variable pitch pixel CCD results in pixels of about 8 km at nadir and only 14 km at the edge of the FoV (Goodman et al. 2013). The x, y coordinates of the focal plane are transformed to latitude and longitude coordinates using an estimated cloud-top ellipsoid as the lightning source layer with a height of 14 km at the equator and 6 km at the poles. Bruning et al. (2019) describes the effects of using this ellipsoid on GLM parallax with respect to ground-based references. GLM captures lightning continuously and in distinct time frames of 2 ms. GLM performance information can be found in Bateman et al. (2021), Rutledge et al. (2020), Murphy and Said (2020), Bateman and Mach (2020), Marchand et al. (2019), and others.
This study uses Level 2 GLM data. The different levels of data are based on clustering of the smallest detected elements, called events, as single illuminated pixels that pass the detection thresholds (Mach 2020). Event locations are set as the center of the illuminated CCD pixels in Earth coordinates. Adjacent events of the same time frame are merged into a group. Next, groups are combined into flashes [as in Mach et al. (2007)]. The GLM clustering algorithm uses a weighted Euclidean distance (WED) with limits of 16.5 km in latitude and longitude direction and 330 ms in time. The WED criterion is tested for pairs of events with one event in each group. If the WED remains below 1.0, the two groups belong to the same flash.
c. NOAA NCEI Storm Events database
Severe weather reports from the NOAA NCEI Storm Events database are utilized to evaluate the performance of the LJ algorithms. NCEI has recorded tornado reports since 1950 and expanded it to include hail and thunderstorm wind events starting in 1954. After 1996, the event database distinguishes 48 event types detailed in Murphy (2018).
This study evaluates tornadoes, large hail, and thunderstorm winds that are defined as severe weather by NOAA. The reports are used as ground truth although the NCEI archive, as a user-based report databank, can miss events or false alarms are possible. For example, at nighttime there are usually fewer observers than at daytime so nighttime severe events are more likely to be not reported. A human-based report database is not infallible; however, it remains the best approach available to record severe weather events over the entire CONUS. The NCEI reports are then clustered with a database scan (DBSCAN) algorithm (Scikit-learn Developers 2022) using a WED with spatial (10 km) and temporal (6 min) criteria as applied in Schultz et al. (2016) to avoid ambiguous reporting of a single severe weather event. Clustering is performed individually for each report type. The location and time of the first report of the cluster is kept and represents the event in the resulting data.
d. Cell tracking
The cell tracking is based on the GEO part of the nowcasting software from NWCSAF. NWCSAF belongs to the SAF Network and is part of the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) Ground Segment. This comprehensive package includes several modules. In the following, the nowcasting software is introduced in general and then the specific cell tracking module is described.
1) NWCSAF nowcasting
“The general objective of the SAFNWC is to provide operational services to ensure the optimum use of meteorological satellite data in Nowcasting and Very Short Range Forecasting by targeted users” (EUMETSAT 2022). The nowcasting software operates with satellite imagery, numerical weather prediction (NWP) input and auxiliary data such as observations (e.g., GLM lightning records). The software packages can process GEO and polar-orbiting meteorological satellite data. In total, 24 products are created that include cloud products [e.g., cloud mask (CMA), cloud type (CT), cloud-top temperature and height (CTTH), cloud microphysics (CMIC)], precipitation products [e.g., precipitating clouds and convective rain rates (CRR)], the convective products rapid developing thunderstorms (RDT) and convective initiation, satellite humidity and instability products, and wind products.
NWCSAF software handles the differences in channel lists and properties from various radiometers onboard GEO platforms such as the European Meteosat Second Generation SEVIRI (MSG-SEVIRI) and future MTG–Flexible Combined Imager (FCI) or the operational U.S. GOES-ABI. This study makes use of version NWCSAF v2018.1 with implementation of technical changes in common modules and on convection products to manage corresponding attributes in cloud cells, and the addition of a GLM data reader. The NWCSAF help desk (http://nwc-saf.eumetsat.int) provides all information needed to run the software tool. The detailed scientific and software documentation of all products can be found on the main NWCSAF web portal (https://www.nwcsaf.org). García-Pereda et al. (2019) summarize the NWCSAF GEO v2018.1 software.
2) RDT cell tracking module
The RDT Convective Warning (RDT-CW) package is developed at the French weather service Météo-France (Autones et al. 2020). This complex algorithm uses combined information from several satellite channels in a mix of physical and statistical approaches. Its objective is the identification of cloud cells and discrimination of convective and nonconvective cloud cells. Furthermore, it allows for thunderstorm detection, characterization of development stage and cloud properties, storm tracking, and future thunderstorm location prediction. The object-oriented approach provides several characteristics such as motion vector, cell contour, overshooting tops, and lightning activity for each cloud cell for any identified cloud system. RDT includes a parallax correction.
The algorithm consists of four main steps to track and characterize cloud systems. These steps are summarized in the following as described in the Algorithm Theoretical Basis Document (ATBD; Autones et al. 2020).
-
Detection of cloud cells: The complex RDT-CW package makes use of a variety of different inputs. For our study domain, the main sources are nine required ABI bands and the output from NWCSAF products CMA, CT, CTTH, CMIC that require another 7 ABI bands. Hence, 14 of the 16 ABI bands2 are used. Those consist of visible 0.64 μm; NIR 0.86, 1.37, 1.6, and 2.2 μm; IR 3.9 μm; water vapor bands 6.2, 6.9, and 7.3 μm; and IR 8.4, 10.3, 11.2, 12.3, and 13.3 μm. RDT and other NWCSAF products can also process NWP instability indices and therefore need data such as surface pressure and temperature, temperature profiles, relative humidity, dewpoint temperature at 2 m, the altitude, and pressure levels with geopotential, temperature, and relative humidity measures. Although those could be estimated from the satellite data, providing NWP files is highly recommended. Here, ECMWF NWP data are ingested as the u- and υ-component winds (horizontal wind); tropopause temperature; tropopause pressure; and lifted, K, and Showalter indices. NWP data are bilinearly interpolated. Other optional RDT inputs to improve the product quality include rainfall rates from NWCSAF products CRR and the CRR alternative CRR-Ph derived from cloud physics, as well as GLM flash records (section 2b) as observations.
The initial detection of cloud systems makes use of the NWCSAF CT product to eliminate clear-sky areas, and then the 10.8-μm channel is used to identify towers of at least 6 K of vertical extension. The temperature thresholds are adaptive and cell specific. They are selected by the software between a given warm and cold temperature threshold to just allow the separation of cloud towers from other nearby cloud systems. The cell contours can be further adapted if needed. Figure 2 illustrates the principle of cloud cell detection based on temperature difference between the cloud tower and the warmest pixel in the environment.
-
Tracking of the cloud cells: The tracking is based on an overlap of consecutive satellite images. The RDT-CW software creates a movement guess cloud field from a given time step. This movement guess field is then compared with the following satellite image to identify the overlap and track cloud systems. An adaptation is applied if this forward approach does not find an overlap: a subsequent backward approach where the cells from the following image are backward advected in time and overlap is identified at the initial time step. Very small cells can be enlarged artificially when looking for the overlap. Cloud systems without any overlap after the full procedure define beginning trajectories. The tracking algorithm can handle splits and merges of cloud cells.
-
Discrimination of convective and nonconvective cells as well as identification of deep convection: An initial discrimination of convective and nonconvective clouds in RDT relies on an NWP convective mask to exclude widespread stratiform clouds from further processing and to avoid false alarm convective cells (e.g., in regions with thermally stable air). Then, a discrimination of convective and nonconvective cells is performed through testing several tens of statistical models that consider multiple satellite channels and are based on temperature transitions. Statistical models were tuned with MSG over France and since version 2018.1 adapted models for GOES data were also implemented. Including real-time lightning data can improve the discrimination quality. Details on the model tuning and the discrimination criteria are found in Autones et al. (2020). This study analyzes all cloud cells from intermediate RDT output products prior to the convective discrimination.
-
Prediction of the thunderstorm movement: RDT cell tracking benefits from analyzed motion vectors that are defined between the weighted gravity centers3 of cells. It provides a nonlinear prediction of cloud cells up to 1 h in the future, including forming, dissipation, merging, and splitting of cells.
The description of the NWCSAF and RDT cloud cell identification based on the temperature difference of a cloud tower and its environment, showing the principle of the detection algorithm for three successive slots with morphological evolution [as in Autones et al. (2020), their Fig. 10].
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
In general, RDT needs some spinup time to run smoothly and track trajectories reliably. Therefore, and after testing of this behavior, the first 3 h of each case study (Table 1) are not included in the further analysis. It was found that dropping the first 3 h of each case provides a good trade-off between keeping most of the data and avoiding any negative effect at the beginning of each RDT run.
RDT cell detection and tracking pairs GLM lightning observations with the identified cloud cells. To this end, the cloud cells are shifted backward in time to the time of the GLM flash using RDT motion vectors. The computed distances between the shifted cell position and the flash centroid must be smaller than a given threshold distance (here five satellite pixels, about 10–15 km) to pair the flash to a cloud cell. In this study, the 10-min time series of lightning flash count per minute is evaluated for each cell. The 10-min period represents the time between the previous and the recent, analyzed ABI images.
Table 2 summarizes the statistics of all cells tracked during the nine analyzed cases (Table 1). Separate statistics for all cloud cells and the thunderstorms show that thunderstorms have on average larger footprints and last significantly longer than the average of all cloud cells.
Statistics of the cloud and thunderstorm cells that were identified and tracked in this study using the NWCSAF RDT software.
e. NCEI events for RDT cloud cells
The initial matching (gray in Fig. 1) compares the location of the NCEI reports with the RDT cloud cells that occurred within one satellite update cycle (here 10 min) of the time of the report. These cloud cells are shifted to the exact time of the NCEI report using the RDT motion vector of the cloud cell. Hence, time is excluded as a factor for the matching. The cell contour line at the time of the report is then compared with the location of the report. The report is matched to a cloud cell if it is located inside the cloud cell contour. If a report could not be located inside any cloud cell, the matching criteria are relaxed so that the report must be located not more than 50 km from the cell contour. This distance should be sufficient to account for location uncertainties of the NCEI reports and also the uncertainty of the cloud cell contour. One report is only matched to the closest RDT cloud cell in case of multiple matches.
This work uses only those NCEI reports that could be matched to RDT cloud cells and is based on the following reasons: (i) The main objective of this study is to verify the capability of GLM LJ detection as a severe weather indicator. However, the cloud identification with cell tracking of NWCSAF software can miss a cloud cell. For example, de Laat et al. (2017) reported that RDT detects the young, growing cells well, but RDT is not always able to identify convective cells once the dynamic development ceases. As the LJ detection algorithms need a lightning time series of a storm, no LJ can be identified without cloud cells, even if GLM might observe lightning in the corresponding location. To this end, only reports that could be matched to an RDT cloud cell are included in our analysis. Results of this work indicate the potential of GLM LJs in nowcasting if all cloud cells were detected by the cloud tracking algorithm. It is, as such, an upper bound skill that is provided. The results are best representative of situations where cloud cells are clearly visible from space. As RDT may not detect all cells under high, uniform cloud shields, methods of this study are difficult to apply in such weather situations. (ii) Only the region within GOES-16 FoV is taken into account. (iii) NCEI reports without matched cloud cell are possibly false alarms or misreported in time/space and, thus, cannot verify any LJ. Ortega et al. (2009) studied the location and timing accuracy of crowdsourced weather reports, and state that the official storm data reports can be quite inaccurate in space (“thousands of km2”) and time (“30–60 min”). When comparing different databases, inconsistencies in the reported times and locations of severe weather will occur, as found by Trapp et al. (2006), Witt et al. (1998). This can also have an influence on the matching.
f. Matching of LJs and NCEI events
Matching of NCEI severe weather events to LJs is a crucial part of this work as the severe weather events should verify the identified LJs (green and yellow in Fig. 1). It needs to relate the two points given by the location of the LJ and the location of the severe weather event. The LJ location is defined as the position of the cloud cell’s geometric gravity center4 shifted to the time of the LJ. To assess this crucial task, this study distinguishes two different matching techniques that are explained in the following. (i) The trajectory-based matching considers LJs and NCEI severe events from the same trajectory only. It is the most physically sound matching approach. (ii) The WED-based matching compares the location and time of all severe weather events and LJs and finds matches on the basis of spatial and temporal criteria. It can be seen as an analysis of the local weather situation rather than a storm-based approach. Matches are not restricted to occur in the same cloud cell. In consequence, WED-based matching should lead to somewhat higher skill for each LJ detection algorithm than the trajectory-based matching.
1) Trajectory-based matching
This matching approach analyzes each RDT cell trajectory with its matched NCEI severe weather events and detected LJs. It is straightforward to compare the time of occurrence of the NCEI events and LJs of one trajectory. Figure 3 illustrates this approach. LJ and NCEI event are matched to cells of the same RDT cell trajectory. A time window is set around an LJ to find any NCEI event that could verify this LJ. For the statistics, a hit is counted for each NCEI severe weather event that falls inside such a time window around an LJ. NCEI severe weather events that cannot be allocated to any LJ matching time window are counted as misses. An LJ without any NCEI severe weather event within its matching time window is regarded as a false alarm.
Trajectory-based matching of LJ and NCEI event with an illustration of the matching time window around an LJ (below the dashed line).
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
In this study, the NCEI event to verify an LJ can occur up to 20 min before the LJ giving a negative lead time of the LJ, and up to 90 min after the LJ resulting in a positive LJ lead time. In the results (see section 4b) the consequences of this choice will be further discussed. It is evident that the trajectory-based matching of LJs and NCEI severe weather events does not need a distance criterion as LJ and NCEI event belong to the same thunderstorm. This matching technique can be compared with Schultz et al. (2016) who also used a thunderstorm trajectory-based matching; however, their matching time window included only a 45 min period after each LJ. The Schultz et al. (2016) matching method was also tested as a reference to our results. It is not further discussed as there are significantly different tools in use, in particular Schultz’ cell tracking methods is based on radar data, whereas satellite RDT cell tracking is used in this study.
2) WED-based matching
WED-based matching of LJ and NCEI event. The latitudinal distance dlat, longitudinal distance dlon, and time difference dt are shown. As in Fig. 3, the red symbol marks the LJ and the green symbol represents any first NCEI report of the NCEI event.
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
g. Statistical evaluation
3. LJ algorithms
This section introduces the three types of LJ detection algorithms used in this study. All algorithm types have variable parameters, referred to as algorithm thresholds. The thresholds are altered in order to test each algorithm for its optimal configuration. The automated LJ detection algorithms yield an instant in time whenever the algorithm-specific thresholds are met, referred to as raw LJ. Clustering of raw LJs with short temporal difference in the same storm should give more realistic LJ numbers with regard to the physical processes leading to the LJ. Raw LJs of a storm that occur within 6 min are merged to one longer LJ. LJs are also given a duration (see section 3a), and if a new LJ is identified whenever the previous LJ is still going on, these two LJs are merged. The time of the first LJ is kept as the beginning of the merged LJ. This temporal clustering was also used by (Schultz et al. 2009, 2011, 2016).
a. The σ LJ algorithm
The framework algorithm was provided by Gatlin (2007), Gatlin and Goodman (2010), and refined by Schultz et al. (2009) who coined the term sigma (σ) LJ detection algorithm. As a first step, the flash rate (FR) time series as flashes per minute of a thunderstorm is averaged over each 2-min time step for smoothing purposes (following Gatlin and Goodman 2010). The σ LJ algorithm includes two parameters that can be configured. One is an activation threshold: only if the FR exceeds this specified FR threshold, the following threshold calculation is initiated. This FR threshold avoids identifying LJs at times of very low electric activity of the thunderstorm. This FR threshold means a key difference between the σ LJ algorithm and the Gatlin algorithm and may be the reason for an improved skill of the former relative to the latter (see Schultz et al. 2009). Then, the discrete time derivative of FR (DFRDT) is calculated as the flash rate trend or time rate of change for consecutive 2-min time steps. The further processing evaluates DFRDT at a given time in comparison with the DFRDT of the previous five 2-min time steps. The standard deviation of these most previous 5 DFRDT history values yields the σ for testing DFRDT at the given time. An additional multiplier, often referred to as σ level, can be used to modify the threshold for detecting an LJ. A σ level of 2 means that the recent DFRDT value must exceed 2 times the corresponding σ value for marking an LJ. Figure 5 illustrates the workflow of the σ LJ algorithm using a σ level of 2. The σ algorithm evaluates 14 min of lightning records in total (six time steps for five previous DFRDT values plus the given time step) and, thus, needs this minimum storm history as spin up to identify any LJ. Schultz et al. (2009) suggested an FR activation threshold of 10 flashes per minute, and a σ-level threshold of 2. This study will test different combinations of threshold values for the algorithm.
Scheme to explain the σ LJ algorithm using a σ level of 2.0 [from Schultz et al. (2016), their Fig. 5].
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
The DFRDT time series is applied to define the duration of an LJ. If a raw LJ is identified by the algorithm, it continues during the following 2-min time steps until DFRDT drops below zero. The original σ LJ algorithm in Schultz et al. (2009) employed temporal clustering for LJs occurring within a 6-min time frame and considered to merge both a newly identified and an ongoing LJ within the same cell.
b. The FRarea LJ algorithm
The second LJ detection algorithm of this study is a modification of the σ algorithm (section 3a). It also starts with a certain FR threshold as activation criterion. Then, for the DFRDT and subsequent σ calculation, this new LJ algorithm takes the FR time series as averaged in 2-min time steps and divides each value by the most recent RDT cloud cell area.5 Hence, a DFRDT per cloud cell area time series is used to calculate and test the σ-level threshold. Due to the use of the cell area, this algorithm is referred to as the FRarea LJ algorithm in this study. FR trends become FR per cell area trends, thus, accounting for cases where cells merge or split during an ongoing trajectory. The approach should be a more general version of the common σ LJ algorithm as cell splitting and merging will not distort the FR trends and σ-level thresholds. LJ durations and temporal clustering of raw LJs can be processed in the same way as for the σ LJ algorithm.
c. The RIL LJ algorithm
4. Results
This section presents results and discusses them. First, the frequency of LJs and NCEI severe weather events on the selected days is evaluated. In the following, automated LJ algorithms for GLM lightning records are detailed and also compared for different scenarios, that is, summer versus winter, daytime versus nighttime, and trajectory-based versus WED-based matching of LJs and NCEI reports.
a. Lightning jump and severe weather frequency distribution
Figure 6a presents daily counts of LJs and NCEI severe weather events that were matched to the analyzed cloud cells. The test days are ordered chronologically (see Table 1) and numbered from 1 to 29. The time series shows that usually days with a high number of LJs correspond to days with a high number of NCEI severe events. In comparing the number of LJs with the number of NCEI events per day, it is seen that there is a large variability. Test days can have more LJs than NCEI events (e.g., days 15–24) or the number of NCEI events exceeds the number of LJs (test days 1 and 2, 6–12, and 26). Most severe weather days are dominated by wind events, and on test days 1, 4, 6, 17, and 26–28 hail events are dominant. Hence, the spring case (test days 25–27) exhibits a high number of hail events. Day 2, 11 January 2020, features an exceptionally high number of wintertime NCEI severe events, especially severe wind events. This day constitutes an unusual situation atypical for January. Summer and spring test days (days 6–17 and 25–29) exhibit usually both more LJs and more NCEI events than winter test days (days 1–5, 20–24) as expected due to more favorable conditions for convection.
(a) Histogram of the LJs (using the FRarea algorithm with FR threshold of 15 flashes per minute and σ threshold of 1.0) and analyzed NCEI events (clustered NCEI reports) on all test days (Table 1) that are sorted chronologically from 1 to 29. (b) The average diurnal cycle of LJ and NCEI event counts, using hourly binning with mean solar time.
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
Note that this section counts analyzed NCEI events that were matched to cloud cells. In Fig. 6a, 3 and 4 June 2020 (day 7 and 8) exhibit relatively low LJ counts as the period has been disregarded due to GOES-16 downtime when no cell tracking could be performed from 1700 UTC 3 June to 0130 UTC 4 June. This period includes CONUS local afternoon and evening; hence, the time with highest lightning activity per day is not covered.
Figure 6b presents the diurnal cycle of LJs and NCEI severe weather events averaged in hourly intervals of mean solar time. One can notice similar peak times of LJ and NCEI event counts in the local afternoon and evening. The LJ peak contains also early nighttime hours. During daytime, the number of NCEI events exceeds the number of LJs by a factor of 1.5 on average (different y axes). Contrarily, at nighttime the number of LJs is 2–4 times the number of NCEI events for the analyzed storms. The NCEI event counts at night are potentially lowered due to less observers than during daytime. The diurnal trends in individual severe weather event types, that is, wind, hail, or tornadoes, follow the overall findings. The results of this work agree in general with the long-term lightning climatologies provided by Cecil et al. (2014) and Christian et al. (2003) that are based on LIS and OTD observations.
Overall, this section provides encouraging results as there is a correlation between LJ occurrences and the number of NCEI severe events. This correlation serves as a solid basis for using LJs as warning indicators in severe weather nowcasting.
b. Lightning jumps as severe weather predictors
This section evaluates the hypothesis that an LJ indicates severe weather at the ground. NCEI events provide the ground-truth used to verify the detected LJs. Contingency tables are created for the trajectory-based and WED-based matching of LJs and NCEI weather events (see section 2f). Roebber diagrams (Roebber 2009) provide a concise way of illustrating POD, success ratio, CSI, and FBI in one diagram. Success ratio and POD are found on the x and y axes, respectively. The CSI and the FBI are included as isolines. The perfect forecast would be placed in the top-right corner of the diagram.
1) Analysis of the full dataset
Figure 7 shows the Roebber diagram for LJ algorithms using the σ approach, with trajectory-based matching (Fig. 7a) and WED-based matching (Fig. 7c). The RIL LJ algorithm results are shown in Figs. 7b and 7d for the two matching approaches, respectively. All four plots show that the LJ algorithms with the highest CSI feature an FBI close to 1.0. As this desired result is always achieved for the best LJ algorithms, the FBI is not further discussed here.
Roebber diagrams (Roebber 2009) for (a),(c) the σ (markers FR) and FRarea (markers FRa) LJ algorithms as well as (b),(d) the RIL LJ algorithm, using (top) trajectory-based matching of LJs and NCEI events and (bottom) WED-based matching. Colors indicate the σ parameter [in (a) and (c)] and the RIL parameter [in (b) and (d)] of the LJ algorithm. FR uses units of flashes per minute. CSI and FBI isolines are drawn in gray and are labeled on the diagonal as well as on the top and right side, respectively.
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
In general, a higher POD also means an increase in FAR (lower success ratio). Pearson correlation coefficients R reach more than 0.96 for holding any σ level constant, and more than 0.85 for holding any FR level constant. These values of R are valid for both trajectory-based and WED-based matching and also for both the σ LJ and the FRarea LJ algorithm. The RIL LJ algorithm leads to R greater than 0.96 when any RIL is held constant and greater than 0.93 for any FR < 306 constant holds. The strong correlation between POD and FAR limits the upper CSI that can be achieved. Finding more LJs means an increase in POD, but, at the same time, also more LJs that are not matched to any NCEI event. Figure 7 shows that the POD can reach values close to 1 when using relaxed LJ algorithm thresholds (low FR and/or low σ level or RIL) for all three algorithm types, however, at the cost that also the FAR is close to 1.
Table 3 states the best performing algorithm of each type, that is, the algorithm with the highest CSI and FBI close to 1 as evidenced from Fig. 7. For all three algorithm types, similar best CSI of approximately 0.4–0.45 is found using trajectory- and WED-based matching, respectively. The ideal FR thresholds are in the range of 15–20 flashes per minute. A σ level of 1.5 is preferred for the Schultz et al. (2016) based σ algorithm, and a σ level of 1.0–1.5 is preferred for the FRarea LJ algorithm. The optimal sigma thresholds are thus lower than the 2.0 suggested by Schultz et al. (2016). They should be combined with FR thresholds slightly higher than suggested (FR of 15 or 20 flashes per minute instead of 10). Relatively low RIL of 1.1 gives the highest CSI for the RIL LJ algorithm, with FR of 20–25 flashes per minute. In all cases, the WED approach achieves higher CSI than the trajectory-based matching, with most occurrences having both higher POD and lower FAR. This behavior is expected as the WED matching is not restricted to a specific storm cell but uses matching distance and time criteria instead. The WED matching usually finds more matches than the trajectory-based approach since the number of LJs within those matching distance and time from the NCEI event is usually greater than the number of LJs within the same storm cell. The similar performance of the three algorithm types, in particular the good performance of the RIL LJ algorithm, is remarkable since the latter algorithm is easier to derive and interpret than the σ-based LJ algorithms.
The best LJ algorithm of each type, selected by the highest CSI. Matching approaches are separated, and the highest CSI value is provided as well as parameters [FR threshold (flashes per minute), σ, and RIL] of the best-performing algorithm. Clustered LJs are used. The results analyze all data.
Results of the individual NCEI event types (not shown) reveal lower CSI with significantly higher FAR for the rare hail and tornado events. FBI much greater than 5.0 is found for most LJ algorithms verifying the tornado events. The use of more strict thresholds in the LJ algorithm, that is a higher FR and/or a higher σ level or RIL is needed to detect an LJ, does not give higher skill when correlating LJs and these rare severe weather events. Hence, the strength of the LJ, that is defined by the maximum FR reached and the σ level or RIL, cannot be used to predict the type of NCEI event. High POD approaching 1 is possible for the hail events when the threshold in the LJ algorithms is relaxed. The overall CSI remains still low as this means also high FAR. However, during the April case (Table 1), that is dominated by hail events, the maximum CSI reached a high skill of 0.67 (daytime), with a POD of 0.90 and FAR of 0.28. These results suggest better performance when hail is the dominant severe weather hazard, which follows the general idea that frozen hydrometeors play a key role in cloud electrification. Analyzing wind events, that are more frequent and dominate the overall results, gives similar POD with about 0.15 higher FAR than the overall results. The increased FAR relative to the overall results causes a lower CSI value for wind events than overall. The FBI remains close to 1.0 for the best LJs algorithms for wind events (not shown) that, in fact, use slightly more strict thresholds than the best overall LJ algorithms (Table 3).
2) Comparison of daytime and nighttime
Tables 4 and 5 summarize the best skill for each LJ algorithm type at day- and nighttime, respectively. The LJ algorithm performs best during the daytime with the highest daytime CSI (0.42 and 0.50 for trajectory and WED matching, respectively) exceeding the best nighttime CSI (0.37 and 0.48). Higher POD and FAR are found at nighttime than overall (not shown). The recommended LJ algorithms at daytime agree with those recommended overall [section 4b(1)]. Nighttime LJ detection needs the strictest LJ thresholds for a given performance level. The FBI for the best nighttime LJ algorithms ranges from 1.3 to 1.6 and, thus, exceeds 1.0 indicating an overprediction of severe weather events. This finding is caused by a higher ratio of LJs to NCEI events during nighttime than daytime. Higher nighttime FAR might be expected because of less reliability of the human generated weather reports; however, other factors may play a role. Increased GLM sensitivity at night allows detection of smaller and dimmer flashes than during the day. This increased sensitivity explains the strict LJ thresholds needed at night and may also contribute to the relatively high nighttime FAR. The types of storms also differ between day and night, with nighttime storms less likely to produce severe weather. The higher nighttime FAR results from a combination of increased GLM sensitivity and fewer, less reliable severe storm reports. Consequently, an ensemble of LJ algorithms may be more useful than a single LJ algorithm to take into account the enhanced GLM DE at nighttime.
3) Comparison of summer and winter
The summer cases including April 2021 contain significantly more storms than do the winter cases, including November 2020 (Table 1). CSI reaches up to 0.41 and 0.50 for the trajectory and WED-based matching, respectively, during the summer cases. For the winter cases, CSI of 0.29 and 0.38, respectively, is found. All results for summer and winter are given in Tables 6 and 7, respectively. The best summer CSI exceeds the best winter CSI for each combination of LJ algorithm type and matching approach. This is caused by both lower POD and higher FAR in winter than in summer for both trajectory- and WED-based matching. The summer cases (Table 6) provide similar results to the overall analyses (Table 3) including both similar highest CSI values and the same LJ algorithm parameters giving this CSI. For the winter cases, however, the best CSI is achieved with increased FR thresholds relative to the summer cases. Higher FR thresholds in winter than summer are likely favored since winter storms analyzed in this work mainly occurred over the southern and southeastern CONUS. This region is characterized by high GLM flash DE, and relatively warm and moist climate allowing for convection. Summer storms were also analyzed over the northern and central CONUS where GLM flash DE is reduced. The optimal σ level giving the best LJ algorithms in winter is equal to 0.5–1.0, less than for the summer and overall results. The optimal RIL remains the same in winter and summer. The RIL LJ algorithm provides slightly more skill than both σ and FRarea LJ algorithm for the winter cases (Table 7). It should be mentioned that the winter CSI is heavily affected by 11 January 2020 with more NCEI events than all other winter days combined. If this unusual day is excluded from the statistics, the winter CSI increases to about 0.38 and 0.50 for the trajectory and WED matching, respectively. The skill is then the same as for the summer cases. The optimal winter LJ algorithms remain the ones explained above; in particular, higher FR thresholds in winter than summer give the highest CSI.
c. Distances and lead times between LJs and matched NCEI events
This section analyzes the distances and times between NCEI events and their matched LJs. Thus, the hits and the corresponding LJs make the statistic, and misses are not included. This differs from the U.S. NWS warning verification where all NCEI events that were not preceded by an LJ are given 0 lead time (e.g., Brooks and Correia 2018). Only the trajectory-based matching is considered. The distance is taken between the NCEI event and the LJ-producing cloud cell contour at the time of the NCEI event. The time difference is called lead time, and it is positive if the LJ precedes the first report of the NCEI event. The LJ to severe weather matching criteria would in general allow for negative lead times of up to 20 min, however, negative lead times have no physical meaning. Hence, negative lead times are set to zero for the statistics here as they can be triggered by the satellite image update cycle, inaccurate time stamps on the reports, and long-lasting severe weather events.7 As an NCEI event can be matched to more than one LJ, a strategy for creating 1-by-1 pairs for the statistics is needed. Two different strategies are applied here to estimate the range of distances and lead times: (i) The best match means the minimum WED [calculated as in Eq. (1)] for any NCEI event and its LJ matches. This WED assessment allows for ranking all matched pairs based on a combination of spatial and temporal discrepancy. (ii) The maximum lead time match considers that pair of an NCEI event and all matched LJ that gives the maximum lead time. The latter (ii) appears to be interesting for nowcasting purposes as here, in particular, longer lead times are desired. All presented results use the FRarea LJ algorithm with FR threshold of 15 flashes per minute and σ threshold of 1.0. The findings are similar for other LJ algorithms that perform equally well.
The distances and lead times between trajectory-based matches of LJs and NCEI events are shown in Fig. 8, with the best match in Figs. 8a and 8b and the maximum lead time match in Figs. 8c and 8d, respectively. The distance distributions Figs. 8a and 8c sharply peak from 0 to 1 km meaning that the NCEI event was located within the LJ-producing cloud cell contour. Distances for the best match approach (Fig. 8a) rarely exceed 20 km since the distance is part of the minimum WED calculation. Larger distances between matched NCEI events and LJs are possible in the maximum lead time match (Fig. 8c) as this pairing is based on the lead times only. The lead time distributions in Figs. 8b and 8d represent these different approaches. The lead time distribution for the best match in Fig. 8b peaks for lead times close to 0 min. The frequencies are monotonically decreasing for longer lead times, and lead times longer than 40 min are rare. With distances close to 0 km for the majority of LJ–NCEI pairs, the lead times become the key point when looking for the WED minimum, thus, the best match approach approximates a minimum lead time approach. The distribution mean and median are found at 7.9 and 0 min, respectively. The maximum lead time match lead time distribution (Fig. 8d) also features a sharp peak for lead times of less than 5 min. However, the distribution shows many LJs preceding severe weather events and lead times greater than 0, with a significant amount of LJ–NCEI event pairs having lead times longer than 1 h. The short-lead-time peaks, where LJs occur almost at the time of the severe weather report, are partly happening during short lived thunderstorms, that is, short RDT trajectories. Longer lead times indicate that severe weather producing storms exhibit first LJ activity well before the severe weather impacts the ground. The longest lead times are limited by the time matching criterion (90 min). The curves of individual NCEI event types follow the overall distribution shape. However, it is noticed in Figs. 8b and 8d that LJ lead times for tornadoes tend to be shorter than for wind and hail events. Tornado events were still rare in comparison with wind and hail events, and this finding would need further proof. The mean and median lead times for the maximum lead time match equal 37.5 and 34.0 min, respectively.
Histograms of the (a),(c) distances and (b),(d) lead times between matched LJs (using the FRarea algorithm with FR threshold of 15 flashes per minute and σ threshold of 1.0) and NCEI events with the trajectory-based matching. Results for (top) best match and (bottom) maximum lead time match are shown for all tested days (Table 1). Separate NCEI event types and the overall counts (all) are given as shown in the legend.
Citation: Journal of Applied Meteorology and Climatology 62, 11; 10.1175/JAMC-D-22-0144.1
Overall, LJs may occur close in space and time to a given severe weather event but lead times of several tenths of minutes can also be expected. Assessing the true physics-based lead times needs a one-to-one physical correspondence between LJs and severe weather events that was beyond the scope of this paper. Hence, this work did not investigate whether optical GLM LJs indicate a physical change inside the thundercloud that causes severe weather. The long lead times seen in the maximum lead-time match are likely nonphysical when it comes to a direct relation to a severe weather event, however, they may have a high value for nowcasting severe weather as they indicate relatively early that a storm can turn severe. The best match pairs of LJs and severe weather events may help in analyzing the physical relation between GLM LJs and severe weather events in the future.
5. Summary and final thoughts
Almost 50 000 thunderstorm trajectories were identified by the NWCSAF tool and its rapid developing thunderstorm package. The storms’ lightning records as detected by the Geostationary Lightning Mapper on GOES-16 are analyzed over the central and eastern CONUS. Three different lightning jump algorithm types are tested in different configurations, namely, the σ LJ algorithm, the FRarea LJ algorithm as a modification of the former, and the simple RIL LJ algorithm. The ground truth is constituted by NCEI severe weather archive reports clustered in space and time to NCEI severe weather events. They are used to verify detected LJs, that are matched to severe weather events, and count statistical hits, misses, and false alarms. Critical success index values reach ∼0.5, with a probability of detection as high as ∼0.7 and a false alarm ratio of ∼0.4. The best CSI values are obtained using the FRarea LJ algorithm, with a flash rate (FR) activation threshold of 15 flashes per minute and a σ threshold of 1.0–1.5. The novel RIL LJ algorithm can achieve similar skill with an FR threshold of 20 flashes per minute and the RIL threshold of 1.1. At nighttime the ratio of the number of LJs to the number of NCEI events is twice as high as at daytime, thus, more strict LJ detection thresholds can help to effectively reduce the relatively high FAR with the overall recommended thresholds. Certain situations allow for high CSI of 0.67 (POD: 0.90; FAR: 0.28), as the daytime analysis of three days in April 2021 demonstrated. While matching LJs and NCEI events, the lead times exhibit a large variability, from no time difference between LJ and NCEI event up to lead times of more than 1 h. Averaging the upper bound of derived lead times, an LJ precedes a matched NCEI weather event by about 35 min. However, this calculation includes some long, nonphysical lead times and, thus, average lead times for physically related LJs and NCEI events would be shorter (not part of this work).
Overall, the CSI skill of LJ-based nowcasting of severe weather is higher when there is more lightning (e.g., during daytime and during spring and summer). This could point to a weakness of the GLM and LJ algorithms, or it might be related to cloud and lightning characteristics. Convection reaches higher altitudes in a warmer atmosphere (day vs night; summer vs winter) and lightning can discharge at high altitudes. In turn, light emitted by these high-altitude flashes is efficiently detected by GLM. Additionally, the noninductive cloud electrification in winter might work differently as under colder conditions more frozen hydrometeors are present even at lower altitudes. Discharge processes can happen more often in the lower cloud parts or shallow convection. This is in agreement with the results of Yoshida et al. (2019) for storms in Japan. Thus, the FAR in winter is higher than during summer as LJs are detected even for shallow convective clouds that do not produce severe weather. Further research of cases in different seasons under consideration of cloud tops and lightning altitudes might provide more insight into this aspect. Additionally, GLM has detection limitations within the edges of its CCD, which will miss low energy/smaller flashes that are commonly observed in severe storms (Cummins 2021). An ensemble of LJ algorithms may replace a single LJ algorithm with fixed threshold in order to receive the highest skill depending on the time of day and the season. Further research could confirm this idea.
Satellite-based cloud cell detection and tracking has some downsides. The results of this work best represent situations where cloud cells are clearly visible from space. Weather systems with high, uniform cloud shields are not ideal to be analyzed with the methods of this study. The assumed ground truth, the NCEI reports, are likely missing evidence for some severe weather events, and false reports are also possible. Their location and time are based on observer reports and may not be perfectly accurate, and it can be expected that some severe weather events at nighttime are not reported. The scores shown here—that is, POD, success ratio, and CSI—are lowered by the previous issues. It was not further tested to what extent the scores are altered.
LJs are correlated to NCEI reports in general, however, no relation is found to predict the type of severe weather that is most likely to happen. For example, one cannot expect that a strong LJ indicates that a tornado would happen soon. Certainly, LJs happen in storms with large hail and tornadoes as a high POD value indicates. However, simultaneous high FAR values suggest using complementary information to the LJs for issuing hail or tornado warnings. LJs can guide the forecasters that severe weather is more likely than in storms without LJs, but we cannot state what kind of severe weather to expect. The frequency of LJs, expressed as an LJ rate per time period, may be useful as a predictor of the severe weather type. This was not tested here and could be the subject of further investigation.
GLM lightning detection and the flash detection efficiency depend on the region where storms occur within the GLM field-of-view. This work has the objective of creating a large statistical dataset and does not discriminate regions of GLM DE differences. A complementary analysis was conducted (not shown) in which LJs and NCEI weather events were analyzed in a region of reduced GLM DE. Regions of GLM low and high DE were separated by selecting the states accordingly. The criterion uses the extended abstract of Cummins (2021) and sets a threshold of 60% daytime flash DE, equal to an energy threshold of at least 3 fJ. There are sufficient LJs and NCEI events in the region of low GLM DE for a statistical analysis (about ⅓ of the cases in section 4a). The vast majority of them occurred during the summer test days, and therefore results for the low-GLM DE region are compared with CONUS summer and spring results. The CSI of the best algorithms is slightly higher than for the full CONUS summer/spring analysis (0.52 vs 0.50). Somewhat lower LJ algorithm thresholds, in particular a lower FR activation threshold relative to the full CONUS analysis, give the highest CSI. The GLM flash DE may affect the strength of the LJs expressed as the σ level or RIL, but there is low impact on the number of detected LJs. The relative change from a low flash rate to a high flash rate should still be captured, even with reduced flash DE. In the future, an equivalent study will be conducted also in the region of high GLM DE to address the question if one would suggest different LJ algorithm thresholds there, and to assess the maximum CSI in that region of high GLM DE.
The lightning data from LMA records are modified to provide synthetic GLM observations.
Excluding channels visible (0.47 μm) and ozone band (9.6 μm).
Temperature-based weights with higher impact of low-temperature pixels to represent 3D cloud morphology Autones et al. (2020).
Mean of latitude and longitude at the temperature threshold defining the cloud cell (Autones et al. 2020).
The area included in the cloud cell contour is determined by the RDT software based on a colder temperature than the environment (Fig. 2).
An FR threshold of 30 flashes per minute in combination with high RIL thresholds finds hardly any LJs, and the POD is close to 0 (as seen in Figs. 7b,d).
Only the beginning of the event goes in the statistic. An intensification, which would point to dynamical changes inside the cloud and might have a relation to a detected LJ, is not considered as a separate event.
Acknowledgments.
The work of author Erdmann was supported by the fellowship “Towards an automated severe weather warning tool based on MTG-LI and FCI data” from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). The hosting institution of this fellowship is the Royal Meteorological Institute of Belgium (RMIB). The authors thank F. Autones, M. Claudon, and J.-M. Moisselin for providing, before an official release, the NWCSAF software package with included GLM data reader, for their expertise on the RDT-CW package, and for the review of the article. We thank N. Clerbaux for setting up the new software version of NWCSAF at the RMIB and for downloading necessary satellite data. We thank ECMWF for providing the necessary NWP data. We thank C. J. Schultz and two anonymous reviewers for excellent work and suggestions.
Data availability statement.
The NWCSAF software is available on the NWCSAF website (https://www.nwcsaf.org). ABI data are available online via NASA EARTHDATA (https://search.earthdata.nasa.gov/portal/idn/search?fi=ABI). GLM data are available online via NASA CLASS (https://www.avl.class.noaa.gov/saa/products/search?sub_id=0&datatype_family=GRGLMPROD&submit.x=22&submit.y=2). Access to ECMWF data requires a user account and access token. The NCEI weather reports are online (https://www.ncdc.noaa.gov/stormevents/).
REFERENCES
Autones, F., M. Claudon, and J.-M. Moisselin, 2020: Algorithm Theoretical Basis Document for the convection product processors of the NWC/GEO. NWCSAF Tech. Rep., 61 pp., https://www.nwcsaf.org/Downloads/GEO/2018.1/Documents/Scientific_Docs/NWC-CDOP2-GEO-MFT-SCI-ATBD-Convection_v2.2.pdf.
Bateman, M., and D. Mach, 2020: Preliminary detection efficiency and false alarm rate assessment of the Geostationary Lightning Mapper on the GOES-16 satellite. J. Appl. Remote Sens., 14, 032406, https://doi.org/10.1117/1.JRS.14.032406.
Bateman, M., D. Mach, and M. Stock, 2021: Further investigation into detection efficiency and false alarm rate for the Geostationary Lightning Mappers aboard GOES-16 and GOES-17. Earth Space Sci., 8, e2020EA001237, https://doi.org/10.1029/2020EA001237.
Blakeslee, R. J., and Coauthors, 2020: Three years of the lightning imaging sensor onboard the International Space Station: Expanded global coverage and enhanced applications. J. Geophys. Res. Atmos., 125, e2020JD032918, https://doi.org/10.1029/2020JD032918.
Boccippio, D. J., and Coauthors, 2000: The Optical Transient Detector (OTD): Instrument characteristics and cross-sensor validation. J. Atmos. Oceanic Technol., 17, 441–458, https://doi.org/10.1175/1520-0426(2000)017<0441:TOTDOI>2.0.CO;2.
Brooks, H. E., and J. Correia Jr., 2018: Long-term performance metrics for National Weather Service tornado warnings. Wea. Forecasting, 33, 1501–1511, https://doi.org/10.1175/WAF-D-18-0120.1.
Bruning, E. C., and Coauthors, 2019: Meteorological imagery for the Geostationary Lightning Mapper. J. Geophys. Res. Atmos., 124, 14 285–14 309, https://doi.org/10.1029/2019JD030874.
Cecil, D. J., D. E. Buechler, and R. J. Blakeslee, 2014: Gridded lightning climatology from TRMM-LIS and OTD: Dataset description. Atmos. Res., 135–136, 404–414, https://doi.org/10.1016/j.atmosres.2012.06.028.
Christian, H. J., and Coauthors, 1999: The lightning imaging sensor. 11th Int. Conf. on Atmospheric Electricity, Guntersville, AL, NASA, 746–749, https://ntrs.nasa.gov/citations/19990108789.
Christian, H. J., and Coauthors, 2003: Global frequency and distribution of lightning as observed from space by the optical transient detector. J. Geophys. Res., 108, 4005, https://doi.org/10.1029/2002JD002347.
Chronis, T., L. D. Carey, C. J. Schultz, E. V. Schultz, K. M. Calhoun, and S. J. Goodman, 2015: Exploring lightning jump characteristics. Wea. Forecasting, 30, 23–37, https://doi.org/10.1175/WAF-D-14-00064.1.
Cummins, K. L., 2021: On the spatial and temporal variation of GLM flash detection and how to manage it. 10th Conf. on the Meteorological Application of Lightning Data, Online, Amer. Meteor. Soc., 692, https://ams.confex.com/ams/101ANNUAL/meetingapp.cgi/Paper/382589.
Curtis, N., L. D. Carey, and C. Schultz, 2018: An analysis of the lightning jump algorithm using geostationary lightning mapper flashes. 25th Int. Lightning Detection Conf./Seventh Int. Lightning Meteorology Conf., Fort Lauderdale, FL, NASA, 6 pp., https://ntrs.nasa.gov/api/citations/20180003505/downloads/20180003505.pdf.
de Laat, A., and Coauthors, 2017: Analysis of geostationary satellite-derived cloud parameters associated with environments with high ice water content. Atmos. Meas. Tech., 10, 1359–1371, https://doi.org/10.5194/amt-10-1359-2017.
Dobber, M., and J. Grandell, 2014: Meteosat Third Generation (MTG) Lightning Imager (LI) instrument performance and calibration from user perspective. Proc. 23rd Conf. on Characterization and Radiometric Calibration for Remote Sensing (CALCON), Logan, UT, Utah State University, 13 pp., https://digitalcommons.usu.edu/cgi/viewcontent.cgi?filename=0&article=1088&context=calcon&type=additional.
Erdmann, F., 2020: Préparation à l’utilisation des observations de l’imageur d’éclairs de météosat troisième génération pour la prévision numérique à courte échéance (Preparation for the use of Meteosat Third Generation Lightning Imager observations in short-term numerical weather prediction). Ph.D. thesis, Université Toulouse 3–Paul Sabatier, 279 pp.
Erdmann, F., O. Caumont, and E. Defer, 2022: A geostationary lightning pseudo-observation generator utilizing low-frequency ground-based lightning observations. J. Atmos. Oceanic Technol., 39, 3–30, https://doi.org/10.1175/JTECH-D-20-0160.1.
EUMETSAT, 2022: NWCSAF general information. NWCSAF, accessed 8 March 2022, https://www.nwcsaf.org/web/guest/nwcsaf-general-information.
Farnell, C., and T. Rigo, 2020: The lightning jump algorithm for nowcasting convective rainfall in Catalonia. Atmosphere, 11, 397, https://doi.org/10.3390/atmos11040397.
García-Pereda, J., and Coauthors, 2019: Use of NWCSAF NWC/GEO software package with MSG, Himawari-8/9 and GOES-13/16 satellites. 2019 Joint EUMETSAT/AMS/NOAA Conf., Boston, MA, Amer. Meteor. Soc., 17B.2, https://ams.confex.com/ams/JOINTSATMET/mediafile/Handout/Paper358940/EUMCONF2019Boston-JavierGarciaPereda-NWCSAFSatellites.pdf.
Gatlin, P. N., 2007: Severe weather precursors in the lightning activity of Tennessee Valley thunderstorms. M.S. thesis, University of Alabama in Huntsville, 87 pp.
Gatlin, P. N., and S. J. Goodman, 2010: A total lightning trending algorithm to identify severe thunderstorms. J. Atmos. Oceanic Technol., 27, 3–22, https://doi.org/10.1175/2009JTECHA1286.1.
Goodman, S. J., H. J. Christian, and W. D. Rust, 1988: A comparison of the optical pulse characteristics of intracloud and cloud-to-ground lightning as observed above clouds. J. Appl. Meteor., 27, 1369–1381, https://doi.org/10.1175/1520-0450(1988)027<1369:ACOTOP>2.0.CO;2.
Goodman, S. J., and Coauthors, 2013: The GOES-R Geostationary Lightning Mapper (GLM). Atmos. Res., 125–126, 34–49, https://doi.org/10.1016/j.atmosres.2013.01.006.
Herzog, B. S., K. M. Calhoun, and D. R. MacGorman, 2014: Total lightning information in a 5-year thunderstorm climatology. XV Int. Conf. on Atmospheric Electricity, Norman, OK, NOAA/NSSL, https://www.nssl.noaa.gov/users/mansell/icae2014/preprints/Herzog_325.pdf.
Mach, D. M., 2020: Geostationary lightning mapper clustering algorithm stability. J. Geophys. Res. Atmos., 125, e2019JD031900, https://doi.org/10.1029/2019JD031900.
Mach, D. M., H. J. Christian, R. J. Blakeslee, D. J. Boccippio, S. J. Goodman, and W. L. Boeck, 2007: Performance assessment of the optical transient detector and lightning imaging sensor. J. Geophys. Res., 112, D09210, https://doi.org/10.1029/2006JD007787.
Marchand, M., K. Hilburn, and S. D. Miller, 2019: Geostationary lightning mapper and Earth networks lightning detection over the contiguous United States and dependence on flash characteristics. J. Geophys. Res. Atmos., 124, 11 552–11 567, https://doi.org/10.1029/2019JD031039.
Mikuš Jurković, P., N. S. Mahović, and D. Počakal, 2015: Lightning, overshooting top and hail characteristics for strong convective storms in Central Europe. Atmos. Res., 161–162, 153–168, https://doi.org/10.1016/j.atmosres.2015.03.020.
Murphy, J. D., 2018: Storm data preparation. NOAA Tech. Rep. NWSI 10-1605, 110 pp., https://www.nws.noaa.gov/directives/sym/pd01016005curr.pdf.
Murphy, M. J., and R. K. Said, 2020: Comparisons of lightning rates and properties from the U.S. National Lightning Detection Network (NLDN) and GLD360 with GOES-16 Geostationary Lightning Mapper and Advanced Baseline Imager data. J. Geophys. Res. Atmos., 125, e2019JD031172, https://doi.org/10.1029/2019JD031172.
NASA, 2022: Instruments: Advanced Baseline Imager (ABI). NASA, accessed 24 March 2022, https://www.goes-r.gov/spacesegment/abi.html.
Nisi, L., A. Hering, U. Germann, K. Schroeer, H. Barras, M. Kunz, and O. Martius, 2020: Hailstorms in the Alpine region: Diurnal cycle, 4D-characteristics, and the nowcasting potential of lightning properties. Quart. J. Roy. Meteor. Soc., 146, 4170–4194, https://doi.org/10.1002/qj.3897.
Ortega, K. L., T. M. Smith, K. L. Manross, K. A. Scharfenberg, A. Witt, A. G. Kolodziej, and J. J. Gourley, 2009: The severe hazards analysis and verification experiment. Bull. Amer. Meteor. Soc., 90, 1519–1530, https://doi.org/10.1175/2009BAMS2815.1.
Peterson, M., S. Rudlosky, and D. Zhang, 2020: Changes to the appearance of optical lightning flashes observed from space according to thunderstorm organization and structure. J. Geophys. Res. Atmos., 125, e2019JD031087, https://doi.org/10.1029/2019JD031087.
Rigo, T., and C. Farnell, 2022: Characterisation of thunderstorms with multiple lightning jumps. Atmosphere, 13, 171, https://doi.org/10.3390/atmos13020171.
Roebber, P. J., 2009: Visualizing multiple measures of forecast quality. Wea. Forecasting, 24, 601–608, https://doi.org/10.1175/2008WAF2222159.1.
Rudlosky, S. D., and H. E. Fuelberg, 2013: Documenting storm severity in the mid-Atlantic region using lightning and radar information. Mon. Wea. Rev., 141, 3186–3202, https://doi.org/10.1175/MWR-D-12-00287.1.
Rudlosky, S. D., S. J. Goodman, K. S. Virts, and E. C. Bruning, 2018: Initial geostationary lightning mapper observations. Geophys. Res. Lett., 46, 1097–1104, https://doi.org/10.1029/2018GL081052.
Rutledge, S. A., K. A. Hilburn, A. Clayton, B. Fuchs, and S. D. Miller, 2020: Evaluating geostationary lightning mapper flash rates within intense convective storms. J. Geophys. Res. Atmos., 125, e2020JD032827, https://doi.org/10.1029/2020JD032827.
Schultz, C. J., W. A. Petersen, and L. D. Carey, 2009: Preliminary development and evaluation of lightning jump algorithms for the real-time detection of severe weather. J. Appl. Meteor. Climatol., 48, 2543–2563, https://doi.org/10.1175/2009JAMC2237.1.
Schultz, C. J., W. A. Petersen, and L. D. Carey, 2011: Lightning and severe weather: A comparison between total and cloud-to-ground lightning trends. Wea. Forecasting, 26, 744–755, https://doi.org/10.1175/WAF-D-10-05026.1.
Schultz, C. J., L. D. Carey, E. V. Schultz, and R. J. Blakeslee, 2017: Kinematic and microphysical significance of lightning jumps versus nonjump increases in total flash rate. Wea. Forecasting, 32, 275–288, https://doi.org/10.1175/WAF-D-15-0175.1.
Schultz, E. V., C. J. Schultz, L. D. Carey, D. J. Cecil, and M. Bateman, 2016: Automated storm tracking and the lightning jump algorithm using GOES-R Geostationary Lightning Mapper (GLM) proxy data. J. Oper. Meteor., 4, 92–107, https://doi.org/10.15191/nwajom.2016.0407.
Scikit-learn Developers, 2022: sklearn.cluster.dbscan. Accessed 11 March 2022, https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html.
Steiger, S. M., R. E. Orville, and L. D. Carey, 2007a: Total lightning signatures of thunderstorm intensity over North Texas. Part I: Supercells. Mon. Wea. Rev., 135, 3281–3302, https://doi.org/10.1175/MWR3472.1.
Steiger, S. M., R. E. Orville, and L. D. Carey, 2007b: Total lightning signatures of thunderstorm intensity over North Texas. Part II: Mesoscale convective systems. Mon. Wea. Rev., 135, 3303–3324, https://doi.org/10.1175/MWR3483.1.
Stough, S. M., L. D. Carey, C. J. Schultz, and P. M. Bitzer, 2017: Investigating the Relationship between lightning and mesocyclonic rotation in supercell thunderstorms. Wea. Forecasting, 32, 2237–2259, https://doi.org/10.1175/WAF-D-17-0025.1.
Trapp, R. J., D. M. Wheatley, N. T. Atkins, R. W. Przybylinski, and R. Wolf, 2006: Buyer beware: Some words of caution on the use of severe wind reports in postevent assessment and research. Wea. Forecasting, 21, 408–415, https://doi.org/10.1175/WAF925.1.
Wapler, K., 2017: The life-cycle of hailstorms: Lightning, radar reflectivity and rotation characteristics. Atmos. Res., 193, 60–72, https://doi.org/10.1016/j.atmosres.2017.04.009.
Williams, E., and Coauthors, 1999: The behavior of total lightning activity in severe Florida thunderstorms. Atmos. Res., 51, 245–265, https://doi.org/10.1016/S0169-8095(99)00011-3.
Witt, A., M. D. Eilts, G. J. Stumpf, E. D. W. Mitchell, J. T. Johnson, and K. W. Thomas, 1998: Evaluating the performance of WSR-88D severe storm detection algorithms. Wea. Forecasting, 13, 513–518, https://doi.org/10.1175/1520-0434(1998)013<0513:ETPOWS>2.0.CO;2.
Wu, F., X. Cui, and D.-L. Zhang, 2018: A lightning-based nowcast-warning approach for short-duration rainfall events: Development and testing over Beijing during the warm seasons of 2006–2007. Atmos. Res., 205, 2–17, https://doi.org/10.1016/j.atmosres.2018.02.003.
Yang, J., Z. Zhang, C. Wei, F. Lu, and Q. Guo, 2017: Introducing the new generation of Chinese geostationary weather satellites, Fengyun-4. Bull. Amer. Meteor. Soc., 98, 1637–1658, https://doi.org/10.1175/BAMS-D-16-0065.1.
Yoshida, S., E. Yoshikawa, T. Adachi, K. Kusunoki, S. Hayashi, and H. Inoue, 2019: Three-dimensional radio images of winter lightning in Japan and characteristics of associated charge structure. IEEJ Trans. Electr. Electron. Eng., 14, 175–184, https://doi.org/10.1002/tee.22795.
Zhang, D., and K. L. Cummins, 2020: Time evolution of satellite-based optical properties in lightning flashes, and its impact on GLM flash detection. J. Geophys. Res. Atmos., 125, e2019JD032024, https://doi.org/10.1029/2019JD032024.
