NowCastMIX: Automatic Integrated Warnings for Severe Convection on Nowcasting Time Scales at the German Weather Service

Paul M. James Deutscher Wetterdienst, Offenbach, Germany

Search for other papers by Paul M. James in
Current site
Google Scholar
PubMed
Close
,
Bernhard K. Reichert Deutscher Wetterdienst, Offenbach, Germany

Search for other papers by Bernhard K. Reichert in
Current site
Google Scholar
PubMed
Close
, and
Dirk Heizenreder Deutscher Wetterdienst, Offenbach, Germany

Search for other papers by Dirk Heizenreder in
Current site
Google Scholar
PubMed
Close
Open access

Abstract

NowCastMIX is the core nowcasting guidance system at the German Weather Service. It automatically monitors several systems to capture rapidly developing high-impact mesoscale convective events, including 3D radar volume scanning, radar-based cell tracking and extrapolation, lightning detection, calibrated precipitation extrapolations, NWP, and live surface station reports. Within the context of the larger warning decision support process AutoWARN, NowCastMIX integrates the input data into a high-resolution analysis, based on a fuzzy logic approach for thunderstorm categorization and extrapolation, to provide an optimized warning solution with a 5-min update cycle for lead times of up to 1 h. Feature tracking is undertaken to optimize the direction of warning polygons, allowing individual, tangentially moving cells or cell clusters to be tracked explicitly. An adaptive ensemble clustering is deployed to reduce the spatial complexity of the resulting warning fields and smooth noisy temporal variations to a manageable level for duty forecasters. Further specialized outputs for civil aviation and for a public mobile phone warning app are generated. Now in its eighth year of operation, a comprehensive and complete set of thunderstorm analyses and nowcasts over Germany has been created, which is of unique value for ongoing research and development efforts for improving the system, as well as for addressing climatological aspects of severe convection. Verification has shown that NowCastMIX has helped to significantly improve the quality of the official warnings for severe convective weather events when used within the AutoWARN process.

Denotes content that is immediately available upon publication as open access.

Publisher's Note: This article was revised on 8 October 2018 to display the authors' full names on the title page.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Dr. Paul M. James, paul.james@dwd.de

Abstract

NowCastMIX is the core nowcasting guidance system at the German Weather Service. It automatically monitors several systems to capture rapidly developing high-impact mesoscale convective events, including 3D radar volume scanning, radar-based cell tracking and extrapolation, lightning detection, calibrated precipitation extrapolations, NWP, and live surface station reports. Within the context of the larger warning decision support process AutoWARN, NowCastMIX integrates the input data into a high-resolution analysis, based on a fuzzy logic approach for thunderstorm categorization and extrapolation, to provide an optimized warning solution with a 5-min update cycle for lead times of up to 1 h. Feature tracking is undertaken to optimize the direction of warning polygons, allowing individual, tangentially moving cells or cell clusters to be tracked explicitly. An adaptive ensemble clustering is deployed to reduce the spatial complexity of the resulting warning fields and smooth noisy temporal variations to a manageable level for duty forecasters. Further specialized outputs for civil aviation and for a public mobile phone warning app are generated. Now in its eighth year of operation, a comprehensive and complete set of thunderstorm analyses and nowcasts over Germany has been created, which is of unique value for ongoing research and development efforts for improving the system, as well as for addressing climatological aspects of severe convection. Verification has shown that NowCastMIX has helped to significantly improve the quality of the official warnings for severe convective weather events when used within the AutoWARN process.

Denotes content that is immediately available upon publication as open access.

Publisher's Note: This article was revised on 8 October 2018 to display the authors' full names on the title page.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Dr. Paul M. James, paul.james@dwd.de

1. Introduction

Despite ongoing developments and increasing resolution, numerical weather prediction (NWP) models alone are not yet sufficient for forecasting high-impact convective weather events for warning services on nowcasting time scales (0–2 h). While such models can give a useful estimate of broad probabilities for severe convection over large areas, based on assessments of synoptic-scale background conditions, the location, timing, and intensity of individual model-generated thunderstorms often deviate considerably from reality. Hence, location- and time-specific verification scores for thunderstorms are typically very poor for NWP at present and hardly better than chance, as discussed by Sun et al. (2014), who describe typical limitations on performing nowcasting tasks with NWP, such as boundary layer details, microphysics, spinup, and a lack of appropriate observations.

Several nowcasting applications have been developed in the past decades to forecast convective events on time scales up to 2 h (Dixon and Wiener 1993; Johnson et al. 1998; Wilson et al. 1998, 2010; Pierce et al. 2000; Lang 2001; Mueller et al. 2003; Hering et al. 2004; Joe et al. 2012). Such rapidly updating nowcasting systems, built upon observational and remote sensing data (e.g., radar, satellite, lightning flashes, weather station reports), have useful verification scores for specific locations, that is, relatively high probability of detection (POD) and low false alarm ratios (FAR). However, when used alone, these nowcasting systems also have limitations. For the most part, they cannot capture dynamical aspects of mesoscale convective developments and mainly report back on what is currently happening. They have little capability in predicting how storm intensity might change, beyond that which can be inferred from empirical linear extrapolations. Hence, a future weather forecasting and warning process will need to consist of several components including nowcasting and ensemble-based NWP to provide seamless weather prediction and appropriate uncertainty information for customers (Heizenreder et al. 2015).

The nature of thunderstorms and their developmental life cycle, including the important aspects of intensity, the mode of convection (linear vs discrete) and accompanying attributes (hail, gusts, and rain), depends crucially on the local background conditions, referred to as the “near-storm environment” (NSE). Aspects of assessing the NSE are discussed, for example, by Doswell and Schultz (2006), while climatologies of NSEs have been constructed over the United States by Smith et al. (2012) and Thompson et al. (2010, 2012) and for the European region by Romero et al. (2007). Meanwhile, potential future changes in European NSEs have been examined by Púčik et al. (2017) based on regional climate models.

While observations-based nowcasting systems know little about this environment, NWP has much of the information required to build a useful NSE assessment of the potential for high-impact storm development within the model domain. In this sense, finding an optimal synthesis of real-time nowcasting data on current storm cells and NWP assessments of the background NSE is what is needed for improving forecast quality for warning services.

To provide optimal automatic decision support for the warning services (Koppert 2014), the German Weather Service [Deutscher Wetterdienst (DWD)] has developed the AutoWARN process (Reichert 2010, 2017; Reichert et al. 2015), covering all important severe weather events, such as gusts, heavy rain, snow, ice, frost, as well as thunderstorms. As summarized in Fig. 1, AutoWARN continually monitors data from observational sources and postprocessing systems based on NWP and remote sensing inputs. Whenever these data sources suggest that severe weather is occurring or is likely to occur within a predefined time frame, spatiotemporal warning proposals (polygons) are calculated by the AutoWARN Status Generator (ASG; Schröder et al. 2014) component, based on statistically postprocessed NWP model output and, in simplified form, by the Observation Proposal Generator (OPG). These warning proposals are sent on to the AutoWARN Status Editor (ASE), which the duty forecasters monitor and use for warning generation on their NinJo meteorological workstations (Koppert et al. 2004; Joe et al. 2005; Heizenreder and Haucke 2009). The polygons can be manually edited if required before the final automatic distribution to customers by the component PVW (from the German Produktion und Verteilung von Warnungen, meaning production and dissemination of warnings).

Fig. 1.
Fig. 1.

The AutoWARN process with arrows indicating the data flow direction, showing (left) the primary input data source; (center left) the core processing and editing components ASG, ASE, and OPG; and (center right) the postprocessing tool PVW for (right) the final dissemination of warnings to customers.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

To support the forecaster and provide a generic optimal solution for nowcast-based warnings in AutoWARN, all nowcast input data are preprocessed together in a single grid-based system: the NowCastMIX (NCM). This system combines input data from all available nowcasting systems, as well as NWP data for NSE assessments, utilizing both event and nonevent information, based on advanced fuzzy logic and clustering approaches. A single optimal and intelligent set of gridded warning fields is thus generated every 5 min, which is sent straight on to ASE, after conversion to a polygon format. The current operational AutoWARN setup thus uses NCM to generate all nowcasting warning information efficiently.

In this paper the NCM system is described in detail. In section 2 the system’s goals and core definitions regarding thunderstorm warnings are introduced. In section 3 the input data are described. In section 4, an overview of the combination methods and algorithms, including fuzzy logic, tracking, and clustering, is provided. Operational and technical aspects are discussed in section 5. A verification study is described in section 6, focusing on the improvements in warning forecast quality since the operational introduction of AutoWARN with NCM. A summary and discussion of future developments is given in section 7.

2. Goals and definitions

The DWD is responsible for issuing weather warnings to the public in Germany, including those for thunderstorms of any severity. Table 1 shows the general warning categorizations at DWD based on German climatology. Four basic warning severity groups are defined, and within these each warning category is given a respective two-digit “ii” code for reference. The yellow group (level 1) defines moderate events (ii = 31) that are unlikely to lead to dangerous conditions. The orange group (level 2) defines strong events that may lead to hazardous conditions temporarily, impacting traffic flow or hampering outdoor activities, while not usually being life threatening. In terms of attributes, either wind gusts in the range of 8–10 on the Beaufort scale (Bft) (63–103 km h−1) are likely (ii = 33), or more than 15 mm h−1 of rain is likely (ii = 34, 61), or both together with (ii = 38) or without (ii = 36) small hail pellets (<1.5 cm in diameter).

Table 1.

Definitions of available warning categories for convective events in NCM, in approximate order of severity (from top to bottom). The general severity level of the respective events is indicated on the left-hand side. The respective color codes for these severity level groups, as used in DWD, are yellow (moderate), orange (strong), red (severe) and violet (extreme). Heavy rain events are convective events without lightning. Hail is regarded as large hail when its diameter exceeds 1.5 cm.

Table 1.

Severe events in the red warning group (level 3) are those that may cause major disruptions to transport systems, possible damage to infrastructure, and may be life threatening. These occur with slow-moving storms with over 25 mm h−1 of rain, leading to an enhanced risk of flash flooding (ii = 42, 62), or with fast-moving storms having violent gusts reaching Bft 11 or more (ii = 40). The most common red thunderstorm type, however, combines over 25 mm h−1 of rain with larger hailstones (>1.5 cm in diameter) and strong gusts up to Bft 10 (89–103 km h−1; ii = 46).

Finally, extreme events in the violet warning group (level 4) are very dangerous, life-threatening situations that may cause significant structural damage and widespread disruption. On the one hand, slow-moving supercells lead to the potential for exceptionally heavy rainfall of over 40 mm h−1, accompanied by severe flash flooding (ii = 95, 66). Alternatively, faster-moving supercells and/or derechos may occur with over 25 mm h−1, combined with violent gusts reaching Bft 11 or more (>104 km h−1; ii = 48). In both cases such thunderstorms also often bring damaging hail, which can be very large in diameter (several centimeters). In rare cases, even tornadoes may be spurned near such events.

The primary goal of NCM is to support the weather warning operations at DWD. NCM produces, on a 5-min rapid-update cycle, categorical warning polygons with respect to a set of standard thunderstorm types and severities. These polygons cover regions where a significant risk of thunderstorms and/or heavy convective rainfall is present during the next 60 min, based on an extrapolation of current remote sensing data under the influence of the NSE. The polygons are optimized in size and orientation to maximize the event POD, while keeping the FAR acceptably low and damping temporal variability so as not to overwhelm duty meteorologists and subsequent customers with overly complex or rapidly changing warnings.

Clearly, finding an optimal solution to achieve these goals is a difficult task. The methods and input data deployed to this aim by NCM will now be described in detail, with a schematic overview shown in Fig. 2.

Fig. 2.
Fig. 2.

Schematic diagram for NowCastMIX. (left) The various input datasets (dark blue boxes). (center) The core processes that operate on the input data (white boxes). (right) The primary output targets (customers) (dark blue boxes).

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

NCM currently runs with a 1-km resolution on the 900 km × 900 km polar-stereographic Radar-Online-Aneichung (RADOLAN, which means radar online adjustment in English) domain (Weigl and Winterrath 2010), which covers Germany as well as a buffer zone to the west, illustrated in Fig. 3. The latter is included in the RADOLAN area to improve radar extrapolation forecasts, given that most convective rainfall events move in from the west or southwest. Many of the radar-based products that NCM uses are available consistently in high resolution on this domain.

Fig. 3.
Fig. 3.

Inner blue square enclosing Germany: RADOLAN domain (900 km × 900 km) used for the AutoWARN output of NCM. The whole area shown above is the European Radar Composite (EuRadCOM) area (2444 km × 2176 km) used for the output of NCM-Aviation, covering most of western and central Europe, as indicated in section 5.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

For convective weather events NCM is required to estimate wind gust potential, rainfall accumulations, and hail. As shown in Table 1, 10 predefined standard thunderstorm events are available for selection, covering typical thunderstorm scenarios across severity levels 1–4. These are distinguished according to different typical attribute combinations with respect to specific thresholds. A further three warning events are reserved for heavy convective rainfall without lightning or severe gusts.

3. Input data

NCM combines different input data values associated with a thunderstorm cell or cell cluster, yielding a best-guess consensus view of its properties and likely risks. This section describes the main input data used.

a. 3D weather radar data

DWD operates a network of 17 weather radar systems delivering 2D and 3D radar data in 5-min intervals (Helmert et al. 2014). Sixteen radar sites are equipped with dual-polarimetric systems, while at Emden (northwest Germany) a single-polarimetric radar system is in operation. The volume scan is composed of 10 elevation scans with elevation angles ranging from 0.5° to 25°. To exploit the polarimetric data adequately, a software framework called Polarimetric Radar Algorithms (POLARA) was developed (Rathmann and Mott 2012), containing various algorithms, such as quality control, hydrometeor classification, and quantitative precipitation estimation (QPE). In particular, the vertically integrated liquid water (VIL) product of the POLARA system, based on Greene and Clark (1972) and Graham and Struthwolf (1999), is deployed as a key component for estimating thunderstorm cell severity within NCM, as shown in more detail later.

b. Thunderstorm detection systems

The detection of significant thunderstorms at DWD is done with radar-based cell detection and tracking systems every 5 min [Konvektive Entwicklung in Radarprodukten (KONRAD) and CellMOS] and a lightning detection system in near–real time (LINET). NCM monitors these three systems continuously. In principle, any of these systems could be replaced by comparable ones, if available.

KONRAD (Lang 2001) examines radar data for clusters of high-reflectivity pixels, indicating the presence of convective cells with heavy precipitation, producing empirical estimates of cell severity as a function of reflectivity maxima and spatial extent. An empirical tracking and motion forecast is also provided. CellMOS (Hoffmann 2008; Wapler et al. 2012; Trepte 2014) also detects cells, combining lightning and radar data, but their forecast severity properties and track vectors are estimated statistically by model output statistics, referencing historical data with NWP estimates of the midtropospheric steering flow. While KONRAD requires several pixels to exceed a reflectivity of 46 dBZ to define a cell, CellMOS cells can be triggered sooner at 37 dBZ, but require nearby lightning flashes to have been detected also.

For a rapid response when new thunderstorms are beginning to develop, NCM also monitors lightning flashes directly. Lightning flash data are provided in real time by the LINET system (Betz et al. 2009; Wapler 2013), which provides accurate flash locations with a negligible time delay. The goal here is to generate and disseminate warnings early: as soon as the first lightning has occurred, even before KONRAD or CellMOS have been triggered.

c. Precipitation analyses and extrapolation

DWD produces high-resolution QPEs via its RADOLAN system (Weigl and Winterrath 2010), combining surface precipitation observations and radar-based precipitation estimates. Gauge adjustment is performed hourly, utilizing real-time data from the 17 German radar stations and approximately 1300 conventional precipitation measurement devices. DWD also runs a comprehensive quantitative precipitation forecast (QPF) system, RadVOR (Winterrath et al. 2012). At its core is an algorithm for advecting precipitation reflectivities, AutoRadSatW (Winterrath and Rosenow 2007), which generates vector fields from consecutive radar images, enhanced with satellite and NWP motion data.

d. Point observations

In addition to the thunderstorm detection systems mentioned above, NCM monitors real-time reports from DWD’s synoptic observation network. In nowcasting it is essential to receive information with a high temporal frequency. Information about recent weather conditions sent from stations at regular primary reporting times (e.g., hourly, half hourly) is not sufficient for monitoring rapidly developing convective cells. In view of this, DWD set up an extra method for stations to report severe weather in real time, as soon as it occurs. These DWD-specific reports are referred to as “MREPs” and indicate the occurrence of hail, heavy rain, and/or wind gusts exceeding certain predefined thresholds, respectively. These thresholds have been set appropriately to help distinguish between the standard thunderstorm severity levels, as defined in official guidelines for the warning services. As soon as a severe event exceeds a threshold, an MREP is sent. Note, however, that nonofficial observations (e.g., from social media and/or storm spotters) are not currently ingested into the system due to potential problems with inconsistency and quality control.

e. Mesocyclone detection

A further system that produces space–time point events is the Mesocyclone Detection Algorithm (MDA; Hengstebeck et al. 2018; Wapler et al. 2016; Zrnić et al. 1985). This DWD algorithm derives and processes azimuthal shear information as a proxy for low- and/or midlevel rotational flow from dual-polarization Doppler radar data. Strong rotational signals can infer the presence of deep, rotating updrafts associated with supercells. The severity of detected mesocyclones is assessed by a fuzzy logic scheme, resulting in five possible severity levels, while a sophisticated quality-assurance algorithm successfully removes dealiasing artifacts without eliminating mesocyclonic vortices.

f. Numerical weather prediction

DWD operates the nonhydrostatic convection-permitting limited area model COSMO-D2 (Baldauf et al. 2018). This runs every 3 h for a forecast period of 27 h on a domain centered on Germany and covering an area of approximately 1430 km × 1590 km. There are 65 vertical levels, while the horizontal resolution is approximately 2.2 km, allowing a direct simulation of severe weather events triggered by deep moist convection. COSMO-D2 is also run operationally in an ensemble mode with 20 members, yielding probabilistic estimates of convective activity.

4. NowCastMIX methodology

a. Analysis of current thunderstorm activity

Any convective cells detected by the KONRAD or CellMOS systems are read in by NCM to be processed and used in subsequent warnings. A circle of 7.5-km radius is first placed around each cell to initiate the process of generating warning polygons. This optimal radius has been chosen as representing a typical cell size in terms of warning relevance.

NCM defines further cells wherever lightning flashes are seen outside of preexisting cell circles and draws new circles around these. To minimize the risk of spurious lightning flash data leading to false cell detections (e.g., erroneous flash detections caused by human activities not related to thunderstorms, such as local discharges on electric railways and from overhead electricity pylons), NCM requires that either at least 1 mm of VIL is present at the flash location, indicating that precipitation is occurring there, or that at least two flashes have been detected within 15 min and within 7.5 km of each other.

NCM generally operates with a 20-min time window when setting up the analysis of current cells, increasing temporal stability, which is an important aspect for warning services, and preventing cells from disappearing too quickly after they first appear. Nevertheless, cells are removed from the analysis after 10 min in cases where no more lightning activity is occurring in the vicinity, indicating that the cell has dissipated.

b. Fuzzy logic combination of input data

Thunderstorm severity can be categorized by the intensity of three core attributes: wind gusts, heavy rain, and hail. Their potential for causing dangerous conditions depends on many meteorological factors, such as cell speed, the size and intensity of the convective overturning, low-level moisture content, vertical shear, etc. Hence, a core task of NCM is estimating the likelihood of attribute occurrence, based on an assessment of current severity levels as a function of different threshold levels. This is done by combining the different input data values associated with a cell meaningfully, yielding a best-guess consensus view of its properties and likely risks.

For diverse input data there are no preexisting equations for estimating risk probabilities. Hence, a useful approach is to employ fuzzy logic, as derived from the fuzzy set theory of Zadeh (1965). This is an ideal tool for converting attributes into warning categories because it allows nonsharp boundaries to be defined for each input attribute and can evaluate the contribution of nonevents just as effectively as events.

Fuzzy logic has been used widely in operational meteorology (e.g., Murtha 1995). Applications include forecasting cloud ceiling heights and visibilities at airports (Hansen 2007), predicting fog formation (Mitra et al. 2008), and the classification of cloud types from satellite data (Baum et al. 1997). Fuzzy logic has also improved the quality of radar precipitation estimates by helping to filter out nonprecipitating noise (Berenguer et al. 2006). The National Severe Storms Laboratory’s Warning Decision Support System (WDSS; Lakshmanan et al. 2007) and the work of Roberts et al. (2006), who developed a fuzzy-logic-based prototype for nowcasting severe thunderstorms, are further examples. Meanwhile, the operational NCAR Auto-Nowcast System (Mueller et al. 2003) also uses fuzzy logic rules and is of particular relevance to NCM. Also relevant is a fuzzy logic method for classifying convective storm severity over Finland (Rossi et al. 2014).

At each storm cell center NCM deploys a hierarchy of fuzzy sets, as shown in Fig. 4, to yield likelihoods that the thunderstorm attributes are reaching certain severity thresholds. The resulting values, referred to here as fuzzy potentials, are then combined to produce a final category for the storm cell in question, according to the guideline thresholds in Table 1.

Fig. 4.
Fig. 4.

Fuzzy logic hierarchy in NCM. Dark blue cells represent the various input datasets that are processed in the first level of fuzzy sets (light blue). These yield respective fuzzy potentials (fp) that feed into higher levels of fuzzy sets. The highest level of fuzzy sets (light green) provides the final fuzzy potentials for the thunderstorm attributes (heavy rain, hail, and severe wind gusts) at two or three specific severity thresholds, respectively. These are then combined to generate a most appropriate final thunderstorm warning category, as indicated in Table 1.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

Typically two or three different parameters are assessed in each fuzzy set, with input values derived at the respective cell centers. For each input parameter, empirical functions have been created, defining the degree of the parameter’s membership to a range of severity levels. These functions have been successively optimized over the first years of NCM operation. The respective thresholds, which place their main focus on extreme values, have been made consistent by examining the observed statistical distributions of parameter values for all thunderstorms over several convective seasons.

To simplify the task of creating modifiable and consistent fuzzy sets, a fuzzy set creation suite has been developed for NCM. Here, required severity thresholds are selected for each parameter, while the relative weighting and linearity of the combination rules can be adjusted empirically, before final automatic rule generation. The deployment of this software suite is vital for NCM, since several input datasets are involved, any of which could be upgraded or replaced, making a manual creation of fuzzy sets very laborious.

The first fuzzy set combines lightning density with VIL to yield a basic fuzzy potential for intense convection. This potential is then fed into subsequent fuzzy sets for heavy rain, hail, and wind gusts. The initial fuzzy set for strong gusts combines the estimated speed of the cell’s motion vector (see next section for more details) with a background estimate of likely wind gusts in the lower troposphere, based on the COSMO-D2 limited area model. Here, it is assumed that there is a significant positive correlation between cell motion speed and likely maximum gusts due to the transfer of horizontal momentum from the midtroposphere down to the surface (e.g., Geerts 2001).

To estimate maximum local rainfall amounts, several factors play a role and two fuzzy sets are required. First, rainfall intensity is assessed via the already calculated fuzzy potential for intense convection, combined with available precipitable water according to COSMO-D2 and cell size according to KONRAD, noting that a large cell will take longer to traverse a location, resulting in more rain there. Second, cell speed must be assessed, since a slow-moving storm can result in far more rain locally than a fast-moving one. Hence, cell speed is combined with 1-h QPFs from RadVOR. A calibrated QPE (i.e., of rain already fallen over the projected storm track) is also added to this second fuzzy set, based on RADOLAN. This is important because if successive storms move over the same location or when a convective cell is embedded in larger-scale rain, rainfall could exceed the threshold levels more readily.

The probability of hail occurrence, as well as potential hail size, is less easy to estimate than wind gusts or heavy rainfall. KONRAD provides an empirical estimate of already falling hail via pixel counts of very high reflectivity (Wapler 2017). This is combined in a fuzzy set with the intensity of convection fuzzy potential from VIL and lightning density.

Ultimately, the various basic fuzzy potentials above are combined to produce final fuzzy potentials for heavy rain, wind gusts, and hail for the various standard specific thresholds, respectively. Note that a tornado potential is not directly provided by NowCastMIX. Although around 10–25 tornadoes occur in Germany per year on average (Dotzek 2001), many of these are relatively weak events. Severe tornadoes, leading to major destruction and loss of life, are much less frequent in Europe, as a whole, than in the U.S. Midwest. Groenemeijer and Kühne (2014) have indeed estimated that the annual mean death toll from tornadoes is around 5 times smaller in Europe than in the United States. Due to their rarity in Germany, tornadoes are extremely hard to forecast with a fuzzy logic approach calibrated with verification or observational data. Nevertheless, a high tornado risk can be inferred indirectly in cases where NowCastMIX’s fuzzy sets yield the highest thunderstorm category (ii = 48), which implies a very high combined risk of severe gusts and hail.

We note that the fuzzy potentials above are not absolute, calibrated probabilities of occurrence, but rather are qualitative numerical values only used internally within the fuzzy hierarchy. These have been tuned to yield an acceptable frequency distribution of storm categories. Based on typical operational practice for extreme weather in Germany and Austria, red category storms should have a typical frequency of up to three events per year on the spatial scale of the administrative districts in these countries. Similarly, violet (extreme) category storms should not occur more often than once in four years on this scale on average. For reference, the average size of an administrative district is around 870 km2 in Germany and 890 km2 in Austria, although there are large variations. Meanwhile NCM analyses show that there are approximately 60 thunderstorm events per year per administrative district on average across Germany.

The fuzzy logic sets have been tuned so that around 6%–7% of the resulting thunderstorms are in the red category in the analysis of administrative districts, while around 0.4% are in the violet category. The tuning for the weaker yellow and orange category storms has been guided by general warning strategy considerations, with just over 50% of storms in the yellow category and around 40% in the orange category. There is a strong interannual variability in thunderstorm category frequencies, depending on the synoptic character of each convective season. However, over the seven years that NCM has so far run, a sufficient range of situations has occurred, so that a satisfactory tuning of the fuzzy sets has been achieved.

c. Calculating cell motion vectors

A core task in NCM is the calculation of a cell motion vector field (CVF), since it is the basis for selecting downstream regions requiring warnings and for advecting various input events with differing observation times onto a common analysis time for consistency.

Three systems are available to NCM for computing a CVF. First, the general vector field from AutoRadSatW within the RadVOR QPF system provides an initial first-guess CVF, although it is not specifically weighted toward strong reflectivities. The second and third input vectors are then provided by the cell tracks in KONRAD and CellMOS, respectively. These systems specifically follow cells of strong reflectivity, providing the focus that NCM needs. Each cell vector is added successively to the CVF using a systematic weighting scheme with a Gaussian spatial decay, optimized with respect to a comprehensive verification of NCM warning polygons against its own analyses.

Finally, the CVF is successively refined by testing each input KONRAD and CellMOS vector against the local background CVF vector. This is necessary, since the tracking system in KONRAD, in particular, is empirical and occasionally produces erroneous vectors as a result of a failure to capture appropriate centers of thunderstorm cells, especially in complex clusters. Outlying vectors, whose speeds and/or directions fall outside a predefined sector, also optimized by verification studies, relative to the CVF are rejected and the CVF is readjusted accordingly.

d. Explicit cell tracking

While the CVF provides a useful and accurate estimate of cell motion for the vast majority of cases, it cannot capture specific vectors of outlier cells moving at a significantly different speed or direction than other nearby cells. The Gaussian mapping inevitably smooths out these outlying vectors to a greater extent, even when KONRAD has been able partially to track them. This weakness is especially important for the case of dangerous right-moving supercells, where the warned regions may deviate significantly from those regions that are eventually affected (e.g., Bunkers et al. 2000).

To address this weakness an explicit cell tracking algorithm has been built into NCM, which runs after the general CVF calculation. Here, cell objects are formed by combining KONRAD, CellMOS, and lightning flash events, using the standard DBSCAN clustering algorithm (Ester et al. 1996) in an ensemble mode (see also section 4f) to identify closely connected cell clusters. The DBSCAN method is applied in a hierarchical sense by successively removing the weakest cells to ensure that the strongest cell cores can be clearly isolated from each other and not be joined together erroneously by a bridge of weaker activity. Failure to achieve this could impact strongly on the quality of subsequent tracking vectors, since the correct identification of a cell core is essential for calculating meaningful tracking vectors.

The clustered cell objects thus identified are then tracked by comparing their positions with cell objects during the previous time steps. Multiple possible tracks are allowed using a weighting scheme that yields probabilities for each possible track, closely similar to a multiple hypothesis tracking (MHT) devised by Reid (1979) and implemented efficiently by Cox and Hingorani (1996).

The forecast motion vectors for each cell are then calculated based on a weighted combination of vectors derived from the cell’s probable tracks over the last six time steps, resulting in temporally smoothed vectors. For newly developing cells, the standard CVF is used to provide an initial vector for the missing time steps. This ensures that all cells start with a reasonably accurate vector, even when they first appear. This is a major improvement over the tracking algorithm in KONRAD, which starts all new cells with a zero motion vector for the first time step.

e. Formation of warning polygons

After detecting cells, estimating attributes, calculating tracking vectors, and drawing circles around their centers, these circles are extended in the direction of motion to create warning regions for the next hour. These will form the warning proposals that the forecasters receive within the AutoWARN process for possible manual modification and then dissemination to customers. Each cell has an explicit motion vector, which is used to form a warning cone. The 7.5-km-radius circle around the cell is successively superimposed in the direction of motion, gradually increasing in size with an expansion angle of 5° to account for intrinsic uncertainties in the motion vectors, until a forecast time of 60 min has been reached. The expansion angle of 5° has been chosen empirically, based on experience. It is a constant parameter since there is currently no clear objective basis for varying the angle as some function of uncertainty.

Since all analyzed cells generate cones, clusters of cells typically result in many overlapping cones. Since only one warning status per grid point is allowed, a weighting scheme is used to determine which warning category should dominate in each case as a function of relative forecast time and severity. In most cases, the highest severity dominates, except in cases where a higher severity near the end of the forecast time range (i.e., end of the warning cone) overlaps a slightly lower severity early in the forecast period: the latter being considered more probable and more immediately relevant. After superimposition some awkward, thin filaments and small cutout polygons may result, having undesirable properties as warning areas. A filtering algorithm is thus applied that tests the size and ratio of the polygon area to the edge length. Each polygon is tested against a matrix of thresholds, as a function of event category. Those that are too thin and/or small are rejected and absorbed by neighboring polygons, whose properties then extend across the rejected polygon. Finally, a Gaussian smoothing is applied to remove any remaining sharp edges, if present.

f. Temporal and spatial optimization via an adaptive clustering ensemble

While the method of producing warning polygons described above is accurate and objective, if left unfiltered, their severity levels would fluctuate rapidly as each individual storm cell grows and decays. Indeed, during major thunderstorm outbreaks hundreds of individual cells may be present, forming large, complex clusters that fluctuate both in spatial form and intensity. As a result, the duty meteorologists would be confronted by many newly significant warning polygons every few minutes. In complex situations this can become unmanageable and frustrating, while strongly fluctuating warnings would be confusing and unacceptable to many customers.

A temporal filter is therefore required to minimize severity level changes, while still allowing rapid responses to new storm developments. Within the context of the explicit cell tracking discussed in section 4d, a two-level hierarchy is deployed using DBSCAN clustering to group cells within an ensemble of alternative clusterings to find an optimal cluster set for reducing the temporal instability of track centers and the spatial configuration of clusters. DBSCAN is an appropriate method here because cluster numbers depend only on the relative cell positions and need not be defined in advance.

DBSCAN is applied to all cells registered during a 20-min time window, with its parameters optimized to yield an acceptable distribution of cluster sizes. After cluster formation, the highest cell severity level observed within each cluster is mapped onto all cells in that cluster. However, to avoid excessive overwarning at the highest severities, red/violet category storms are only mapped onto other red/violet cells in the cluster. Hence, severe warnings remain more tightly focused, in accordance with DWD warning strategies, and have a smaller spatial scale than lower category warnings on average.

This approach results in higher severities lasting at least 20 min within a cluster, before they can fall back, reducing the temporal variability of the final output. In practice, however, the spatial extent of the clusters themselves can vary significantly with time, resulting in short-lived severity jumps covering quite large areas. Such effects have no physical basis but result from the hard decision trees in the clustering method, where small spatial shifts can result in different clusterings affecting a much larger scale.

To overcome this problem, an ensemble of clusterings is performed at each run time. To generate the different members, the cell positions are randomly shifted in some direction within a 6-km radius of the original positions and DBSCAN is run on each member, producing differing cluster configurations. Each alternative clustering is then compared to the current clustering of the preceding run, accumulating penalty points as a function of dissimilarity and spatial variability of the cluster centroids used to form cell object tracks. To increase efficiency, the cell position randomizations are biased successively toward certain directions, based on running penalty totals, so that subsequent clusterings can concentrate increasingly on useful regions of phase space. This approach is called an adaptive clustering ensemble (ACE; Topchy et al. 2004). Finally, the clustering with the fewest penalties is used, reducing the temporal variability of the cluster configurations and tracks to an acceptable level.

The ACE method has been tested and refined based on many case studies with complex thunderstorm clusters, improving the usability of warning polygons significantly. The full process of generating warning polygons, from cell analysis, via expansion into the forecast period, to spatially clustered regions, is illustrated schematically in Fig. 5.

Fig. 5.
Fig. 5.

Schematic illustration showing the production of warning polygons in NCM. (a) Current positions and severity categories of thunderstorm cells are analyzed (colored regions), and the cell motion vector field is used to construct overlapping warning cones showing regions at risk of thunderstorms in the next 60 min (regions initially in gray here). (b) Clustering of analyzed cells (strongly colored areas) allows orange categories to expand over yellow categories within respective clusters, whereas red and violet categories do not extend into lower category areas. The overlapping forecast cones are now colored opaquely according to clustered category. (c) The final warning polygons for the next 60 min are now shown fully colored.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

g. Quality assurance based on real-time observations and mesocyclone detections

In the vast majority of situations, an integrated assessment of automatic remote sensing data produces a sufficiently balanced and accurate summary of convective development for nowcasting purposes. Nevertheless, a small risk remains that an essential system could be offline or produce erroneous data at a crucial time, leading to severity underestimation and failure to release a timely warning.

To reduce this risk, NCM reads in all recent MREPs and cross references them with convective cells detected from the radar and lightning systems. The MREP is assumed to have been triggered by a nearby cell. The time and position of the MREP is read and the resulting point is advected forward to the current analysis time, using the CVF. As shown in Fig. 6, all cells within a radius of 25 km for hail and precipitation, or 35 km for wind gusts, from the MREP’s current virtual position, are examined. A weighting scheme takes the cell–MREP distance and also the fuzzy logic potentials for the respective attribute (hail, rain, or gusts) into account. The probability of a cell being selected is proportional to the fuzzy potential and to the inverse of the cell–MREP distance. The highest weighted cell is selected, and the MREP property is tagged onto it. The automatic estimate of this cell’s severity is compared to the MREP. If the report suggests that the automatic category is too low, it is raised accordingly. Since the cell then takes part in the subsequent clustering procedure, as described earlier, the MREP attribute can spread to other cells in the vicinity.

Fig. 6.
Fig. 6.

Schematic diagram showing the method for assigning a typical space–time point observation to a thunderstorm cell. A report of severe weather at m minutes before the analysis time t has been received (red circle, Rt-m). Its virtual location at time t (pale red circle, R′t) is estimated via the CVF, using a vector length of m minutes. In this example, three nearby cells are present (yellow circles). Cell 1 is the most probable candidate for having triggered the severe event, given that it is much closer to the virtual report position than the other two, noting that fuzzy logic probabilities of the relevant storm attribute are also taken into account at each cell (not shown here) in the assignment algorithm.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

Furthermore, NCM monitors the input data for mesocyclones, using all events of severity level 3 or higher. Although mesocyclone detection via MDA can be used to assess potential threats related to tornadoes and hail (Hengstebeck et al. 2018), its use in NCM is limited to estimating the potential for supercell-generated wind gusts. MDA events are converted to an equivalent gust level and then integrated into the MREP assignment algorithm, as above. The MDA-based gust detection is especially useful, because severe gusts can be detected almost anywhere within the radar coverage area, whereas MREP gust reports can only be sent when a gust has directly affected one of the relatively sparsely distributed synoptic stations.

h. Heavy rain warnings beyond nowcasting time scales

As mentioned earlier, NCM can warn for heavy rainfall for the next hour in cases where no lightning is observed. Such events, however, are rare: only about 0.25% of cells are classified as nonthunderstorm heavy rain. Much more frequent and significant in terms of dangerous weather are events on somewhat longer time scales, extending beyond nowcasting and reaching up to several hours. Typically, convective cell cores are embedded within a larger-scale, slow-moving precipitation cluster. Although hourly rainfall totals may not be extreme in themselves, in comparison with short-lived thunderstorm events, the total accumulation after a number of hours can become extreme.

An example of such an event was the catastrophe in Simbach, Bavaria, on 1 June 2016, where over 100 mm of rain fell widely in the surrounding catchment areas within a few hours. This caused a severe flash flood that killed seven people and resulted in massive damage to buildings and infrastructure. Although some lightning occurred during the heavy rain period, the main cause of the disaster was the persistence and near stationarity of a cluster of heavy precipitation.

At the time of the event, NCM only produced occasional warnings for, at most, orange-category thunderstorms with heavy rain. But the ongoing dangerous development was not easy to spot, even for duty meteorologists. The postevent analysis thus led to a new set of heavy rainfall warning events being added in NCM on a nominal time scale of 6 h. These heavy rain warnings are sent to a different postprocessing channel as thunderstorm warnings and can exist simultaneously and overlap the latter. Table 2 shows the standard warning thresholds for 1- and 6-hourly heavy rain events, for each of the three severity levels.

Table 2.

Thresholds deployed at DWD for heavy rainfall events on 1- and 6-hourly time scales and their equivalent severity levels.

Table 2.

To detect such events, NCM combines recent hourly QPEs from RADOLAN with 1-h QPFs from RadVOR and NWP ensemble forecasts of rainfall for the coming hours from COSMO-D2-EPS, based on the 75th percentiles of the hourly precipitation totals. At each grid point the maximum rainfall value in any 6-hourly window including the analysis time is calculated. If this exceeds at least one of the 1-hourly rainfall thresholds, the Shannon entropy [Shannon (1948), an empirical measure of the distribution of points within a finite number of bins] of the individual hourly values, is examined to determine whether the rainfall is of a longer-lasting character with steady rain over several hours (high entropy) or whether most rain falls in a short space of time (low entropy).

This addresses a key problem wherein intense short-lived thunderstorms may produce enough rain locally to exceed 6-hourly rainfall thresholds. In such cases it is unnecessary to create heavy rain warnings as well as thunderstorm warnings for the same event. Hence, low entropy, below a certain threshold, signals the suppression of heavy rain warnings whenever these overlap or touch other thunderstorm warnings.

For 6-h events with higher entropy, warnings are suppressed locally wherever RadVOR suggests that little or no rainfall will fall over the next hour. Thus, warnings are avoided in locations where it is currently (almost) dry. In this case, either the heavy rain event has already ended, or the warnings would be based on future rainfall anticipated by NWP alone and hence outside of the scope of nowcasting.

5. Operational aspects and output

Warning polygons are converted into a standardized XML format known as the Common Alerting Protocol (CAP; OASIS Emergency Management Technical Committee 2010). The polygon edge coordinates are extracted from the original gridpoint data and embedded within a CAP file, which includes detailed header information, showing the warning type and validity time, for example. With this standard format, NCM output can be shared easily with external users, such as the European Severe Storms Laboratory (ESSL), which regularly assesses NCM warnings each summer within the context of its ESSL Testbed (Groenemeijer et al. 2017).

The CAP warning proposals are sent on to the AutoWARN Status Editor which is continually monitored operationally on NinJo workstations. Polygons indicating a change of warning status are made visible to the forecasters, who can edit their shape and/or category, if necessary, before dissemination to customers. The forecasters can even delete a proposal if they have concrete reasons for rejecting it, based on parallel monitoring of other data sources on the workstations. Forecasters can also monitor NCM output via the “Nowcasting Layer” as part of the severe weather tools on the NinJo workstations (Joe et al. 2005), as shown in Fig. 7. For diagnostic purposes this information can be overlaid with other indicators, such as KONRAD, radar composites, satellite data, and NWP output.

Fig. 7.
Fig. 7.

Screenshot of NowCastMIX warning cones in NinJo, overlaid with VIL (black–gray–orange–red color range) and radar-based QPE from the RADOLAN system (light blue–green–yellow–orange color range), showing fast-moving thunderstorms over the Ruhr district of northwestern Germany at 2000 UTC 21 Aug 2011. The dark red lines and patches are political state boundaries and urban conglomerations, respectively. Major rivers are indicated in blue. The east–west extent of the geographical area shown is approximately 400 km.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

Real-time NCM output is also shown on the DWD-Intranet, which acts as an instantaneous and easily accessible monitor of the NCM status for both developers and forecasters. An example screenshot is shown in Fig. 8, showing current and nowcast regions of thunderstorm and heavy rain activity, with past tracks and forecast motion vectors. A text-based list of all analyzed cells with values of input data and fuzzy logic potentials is also linked. Furthermore, another linked page shows a detailed breakdown of the production times of each NCM subprocess, allowing insights into the possible sources of any production problems that may occur. However, thanks to its robust architecture, NCM has had negligible downtime, on the order of <0.01% in its seven years of operation so far.

Fig. 8.
Fig. 8.

Intranet monitor image, showing NCM output during a severe thunderstorm outbreak at the analysis time 1800 UTC 18 Aug 2017. Current thunderstorm activity is shown as solidly colored polygons whose centers are indicated by black-outlined circles of a specific severity, colored inside according to the categories (ii) on the left-hand side of the graph. The past tracks of each cluster object are shown with small joined circles. The respective tracking vectors are given as black arrows and can be compared to the gray arrows, which show the standard, smooth CVF vector field. The resulting integrated warning areas for the next 60 min are shown as opaque-colored polygons. Heavy rain areas are shown as green polygons that are hatched where they overlap with thunderstorm polygons. The availability of input data for the current NCM run is indicated at the bottom right by green (data available) or red (data missing) buttons. The total area of each warning category (km2) is indicated at the top right.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

Beyond AutoWARN, simplified NCM warnings are sent automatically for real-time display in DWD’s smartphone weather app: WarnWetter. Since neither polygon complexity nor temporal stability is an issue here, no clustering or spatial smoothing is applied. The 10 thunderstorm categories for AutoWARN have been reduced to four basic severity classes (yellow, orange, red, and violet) for the app. Nevertheless, a detailed written description, relating to the original categorical information, can be viewed by tapping on a polygon.

NCM also provides specialized output to civil aviation authorities: NCM-Aviation. Here, a detailed analysis of the convection status is provided (four thunderstorm categories, plus an extra category for radar reflectivities above 37 dBZ without lightning). The CVF is used to extrapolate this analysis linearly to T + 60 min with a 5-min forecast step. Thunderstorm development aspects are not considered, however, in the current version. Cell intensity remains constant through the forecast. The resulting analysis and 12 forecast time steps are converted into polygons that are then displayed in real time at various German airport weather advisory centers.

NCM-Aviation has recently been expanded to cover a much larger area (around 6 times the size of RADOLAN) over much of western and central Europe, shown in Fig. 3, albeit with a reduced quality outside of Germany, where fewer data are available. KONRAD, VIL, and QPF estimates, in particular, are largely unavailable to DWD outside of Germany. Hence, a domain expansion for the AutoWARN output is currently not yet considered viable.

NCM runs operationally on a high-performance Linux cluster every 5 min, typically requiring about 2 min in total to complete all tasks. It consists of shell scripts controlling data inputs/outputs and a core Fortran-90 executable. Since radar composite data and KONRAD become available within 4 min after the analysis time, NCM is started at T + 4 min. Some other input data, such as VIL, CellMOS, and RadVOR, have slightly longer production times and are read from their preceding runs (i.e., 5 min old). Here, the CVF advects older features forward to their expected positions at the common analysis time. By contrast, lightning flash data are available from a data bank in near–real time, such that recent flashes can be utilized up to over 4 min after the analysis time, right up until the core Fortran executable starts. For consistency, such flashes are advected backward in time with the CVF to their nominal positions at the common analysis time.

6. Verification

NCM has yielded a complete analysis of categorized thunderstorm activity over Germany for seven consecutive years at the time of writing. In addition to archived warning polygons, a comprehensive text-based analysis of thunderstorm cell positions and motion vectors has been created, giving values of various input fields and the resulting fuzzy potentials at the respective cell centers. Furthermore, the newly implemented explicit cell tracking algorithm has been applied retrospectively to the entire analysis period, resulting in a complete reanalysis of storm tracks over the convective season, nominally April–September, for 2011–17. These datasets are certainly the most comprehensive analyses of thunderstorm activity over Germany in existence, yielding a holistic view of categorized thunderstorms that simpler datasets, such as lightning flash data, cannot provide. As such, these data have considerable value for nowcasting research and development and may indeed be globally unique.

While climatological aspects of thunderstorms are of great interest, these are beyond the scope of the current work and will be discussed in a follow-up paper. Here, we raise an important verification aspect, namely the comparison between the fully automatic warning proposals generated by NCM and the final official warnings issued by the duty forecasters. In particular, we focus on whether the operational introduction of AutoWARN has led to a verifiable quality improvement in official thunderstorm warnings.

NCM polygons are mapped onto the 411 administrative districts of Germany, which provide a common baseline for the verification. A district is assumed to have a warning if at least 3% of its area is covered by polygons, based on an agreed-upon threshold with the forecasters. Where more than one warning category is present in a district, the highest severity category having an area of at least 3% is used (or just the highest category if no area exceeds 3%).

Before the introduction of AutoWARN in 2014, the Editing, Production and Monitoring (EPM) NinJo workstation module was used by forecasters to disseminate warnings. The EPM system did not allow for free polygon creation like AutoWARN. Instead, warnings were created manually via selecting whole administrative districts, with no intradistrict discrimination possible. Warnings subsequently produced within AutoWARN, while being potentially free from spatial restrictions, have also been mapped onto the common administrative districts, so that NCM proposals and EPM- and AutoWARN-based official warnings can be all compared consistently.

In Fig. 9 we first show standard verification scores for NCM’s automatic 1-h warnings as a function of lead time, using the NCM analyses on administrative regions as the baseline. Hence, the mean POD and FAR are computed by shifting the analysis and forecast time series relative to each other, respectively. Since NCM generates warnings about 5 min after the respective analysis time, the analyses have been shifted by 5 min backward relative to the forecasts, such that the highest mean POD occurs at a lead time of −5 min.

Fig. 9.
Fig. 9.

Verification scores for NCM thunderstorm warnings against NCM’s own thunderstorm analyses as a function of lead time on the German administrative districts, showing POD (solid) and FAR (dashed) for all thunderstorms (yellow), orange severity storms or higher (orange), and red severity storms or higher (red).

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

When all thunderstorms are examined together, irrespective of severity, the POD stays above the FAR up to a lead time of 40 min, for which the CVF quality plays a vital role. Not surprisingly, higher severity events have a higher FAR and lower POD, with the POD of red-level events dropping below the FAR in under 15 min. However, since POD and FAR are scale dependent, this result has only limited relevance in absolute terms. More importantly, the mean spatial variations across Germany of POD, FAR, and the extremal dependency index (EDI; Ferro and Stephenson 2011) have been calculated for a lead time of 30 min and are shown in Fig. 10. The EDI is a useful base-rate independent forecast quality index for rare binary events. As such, it provides a much more trustworthy measure of skill for thunderstorm verifications than some other commonly used indices, such as the Heidke skill score or the critical success index (as discussed in the above citation).

Fig. 10.
Fig. 10.

Mean verification scores on the German administrative regions for NCM thunderstorm warnings (any severity) for a lead time of T + 30 min against NCM’s own analyses, 2011–17 (April–September), showing (a) POD, (b) FAR, and (c) EDI. Values are normalized by removing the (slight) dependency on the area of the administrative districts, as determined via linear regressions, and are then shown in percent (different scales, as shown). Note that the color scale for FAR is inverted, so that warm colors represent better scores in all cases.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

As seen in Fig. 10, the mean skill of NCM warnings increases toward northern Germany, probably due to orographic influences. While much of northern Germany is relatively flat, central and southern Germany are more mountainous. It is conceivable that complex terrain disrupts the linear progression of convective systems (e.g., Markowski and Dotzek 2011), making extrapolation-based predictions of future trajectories harder. Lower forecast skill is also seen in the far southwest and far northwest. This may result from an increasing lack of available radar data products westward of Germany. Since many thunderstorms move in from the west to the southwest (around 55% of all storms have trajectories between 210° and 280° according to NCM’s tracking analyses), a lack of data will inevitably have a negative impact on forecast quality downstream.

AutoWARN was made operational for the start of the 2014 convective season. However, after some unexpected technical problems and handling issues arose, it was taken back offline toward the end of that season. The operational warning services returned to using EPM and continued doing so until shortly after the 2015 convective season. Specific improvements to AutoWARN were undertaken, and it is now being used operationally on a permanent basis. As a result, it has become possible to compare official nowcasting warning performance for three seasons using EPM (2012, 2013, and 2015) against three seasons using AutoWARN (2014, 2016, and 2017). Although NCM was available as guidance during the EPM periods, the crucial difference is that there were no warning proposal polygons available to the forecasters. NCM warnings could be seen in the NinJo workstations, but the forecasters had to enter all thunderstorm warnings into manually chosen administrative districts by hand under EPM, whereas in AutoWARN NCM warnings are already available for direct use to start with.

Although it is not possible to reconstruct the actions of the forecasters in modifying polygons exactly, since the required information was not all stored, just under half of all NCM warnings on administrative regions remained unchanged after processing by the forecasters during the AutoWARN period, while a close balance between raising and reducing the severity of the remaining warnings is seen, as shown in Table 3. During the EPM period, a slightly higher frequency of changes to NCM polygons can be inferred.

Table 3.

Mean percentage changes to automatic NCM thunderstorm warnings across the German administrative regions on a 5-min time scale, following processing by the duty forecasters to create official warnings, showing the EPM (2012, 2013, and 2015) and AutoWARN (2014, 2016, and 2017) periods, respectively. For this analysis, the warnings have been grouped into three basic groups: yellow, orange, and red/violet. Subtle changes to the ii code within a group do not count as a change here.

Table 3.

Since mean synoptic conditions differ from season to season, impacting on quality indicators, the absolute warning quality cannot be compared directly between the EPM and AutoWARN periods. Instead, NCM forecast quality forms a baseline, so that we compare the relative quality of official warnings during the AutoWARN and EPM periods to NCM, respectively. Here, lightning flash data are used as a neutral verification basis for all warnings and only a straightforward yes/no verification for thunderstorms is performed, irrespective of severity.

In Figs. 11a and 11b POD and FAR are shown against lead time for the EPM and AutoWARN periods, respectively, for both the official forecaster warnings (FCR) and for NCM. During the EPM period, the forecaster POD was much lower than that of NCM, while the FAR was much higher. In part, this apparent poor performance has a strategic source, since the warning strategy requires spatially broader warnings, which should not change too often, than NCM tends to generate. In some cases, however, duty forecasters missed some newly developing storms at first, disseminating necessary warnings too late. This inevitably suppresses the POD. Another problem with EPM is that the predefined shapes of the administrative districts lead to suboptimal subjective choices for which region gets a warning, whereas NCM’s polygons are optimally drawn relative to the motion vectors and then mathematically objectively mapped onto these regions.

Fig. 11.
Fig. 11.

Verification scores (POD and FAR) for thunderstorm warnings (any severity) against lightning flash data across the German administrative districts as a function of lead time, showing automatic NCM warnings (NCM, solid) and manually generated operational warnings from duty forecasters (FCR, dashed). Mean values over three convective seasons (April–September) are shown, respectively, for seasons during which forecasters (a) used older background guidance systems without AutoWARN (2012, 2013, and 2015) or (b) had the benefit of real-time direct warning proposals from AutoWARN (2014, 2016, and 2017). (c) Relative EDI for forecaster minus NCM for seasons with AutoWARN guidance (solid green) and seasons without AutoWARN guidance (dashed green). A lead time of, e.g., 30 min shown above is equivalent to T + 30 min as referred to in the text.

Citation: Weather and Forecasting 33, 5; 10.1175/WAF-D-18-0038.1

During the AutoWARN period, the POD of the forecaster warnings has improved and now lies very close to that of NCM. In particular, the forecaster POD decays less quickly with lead time than during the EPM period. Meanwhile, the forecaster FAR also deteriorates (increases) less quickly with lead time during the AutoWARN period. Nevertheless, the forecaster FAR starts at a higher level at small lead times, relative to NCM, than during the EPM period and remains relatively high up to a lead time of 1 h. In spite of this, the strong improvement in the forecaster POD outweighs the effect of a higher FAR, such that the forecaster EDI, relative to that of NCM, has improved notably between the EPM and AutoWARN periods (Fig. 11c) for lead times of T + 10 min and beyond, with larger improvements at longer lead times. Nevertheless, the forecaster EDI still remains generally lower than the NCM EDI in absolute terms.

A t test was carried out by dividing each season into six periods, each with equal numbers of thunderstorms. The forecaster EDI differences between the EPM and AutoWARN periods are highly significant over a range of lead times. In particular, the null hypothesis (that the verification scores of the EPM and AutoWARN periods do not differ fundamentally) can be rejected at the 99th percentile for the forecaster EDI for lead times of T + 30 min and longer. For shorter lead times, the null hypothesis can still be rejected at the 95th percentile down to approximately T + 15 min, although such comparisons become irrelevant as the lead time approaches zero insofar as forecast quality is concerned.

The relative improvement in POD for the official warnings occurs because the duty forecasters are practically forced to notice and respond quickly to developing thunderstorms when NCM pushes direct warning proposals in real time into the AutoWARN process. Furthermore, they can use NCM’s optimal polygons directly and/or modify these freely, without having to consider subjectively the implications for underlying administrative districts. Indeed, as a result of this technical improvement, warnings can now be given on the scale of smaller municipal regions (there are about 11 000 of these in Germany), improving the spatial accuracy of the final warnings still further. A slight increase in FAR in AutoWARN still occurs (Fig. 11c), presumably because missing fewer new storm developments results in a higher mean area with warnings, while the warning strategy of requiring broad and temporally stable warnings remains in place.

7. Summary and future developments

NowCastMIX is the core nowcasting system at DWD, assessing input data from several remote sensing, NWP, and postprocessing systems to generate integrated, optimized warnings for moderate and high-impact convection events. These are produced automatically with a high update frequency for DWD warning services, civil aviation, and a public mobile warning app, benefiting from advanced thunderstorm categorization, clustering, and tracking techniques. It has run for seven years at the time of writing, consistently producing valuable warnings for several user groups. Furthermore, a comprehensive and complete set of thunderstorm analyses over Germany has been generated, which has unique value for ongoing research and development efforts for improving NCM and addressing climatological aspects of severe convection.

Within the context of the AutoWARN process, NCM has become an essential tool for providing guidance to duty forecasters on nowcasting time scales, complementing the available severe weather tools in the NinJo workstation, short-range guidance from NWP, and ensemble and MOS-based postprocessing systems. Verification has shown that NCM has helped to significantly improve the quality of the official warnings for severe convective weather events, when used in conjunction with AutoWARN. With further improvements and extensions in development or in planning, as indicated below, NCM is becoming one of the leading nowcasting systems in Europe.

NCM is currently being extended to cover key aspects of wintertime nowcasting, namely snowfall and freezing rain. For snowfall, radar extrapolations are combined with synoptic station reports and NWP to estimate snow limits and potential snow rates. Similarly, the freezing rain risk is estimated by looking for regions where input data suggest warm air aloft over a frozen surface and where radar extrapolations indicate falling rain. These winter nowcasting developments are beyond the scope of this paper and are less advanced than those for summertime convective events. Indeed, they will benefit from ongoing work in radar meteorology at DWD to develop advanced algorithms for extrapolating dual-polarization hydrometeor classifications (Steinert 2014) from the height of the radar beams down to the surface. These algorithms will also benefit nowcasting of convective events, notably by improving hail detection and refining QPE.

For convective events, ensemble prediction system (EPS) data from regional NWP models, in particular COSMO-D2-EPS and ICON-EPS (Zängl et al. 2015), may prove valuable for improving NCM warnings in the near future. DWD has started the seamless integrated forecasting system SINFONY project (Blahak et al. 2017) to improve ensemble-based seamless prediction from minutes to hours for high-impact convective events. Improved seamless EPS data for nowcasting and NWP could provide probability fields for high-impact convection, which could be blended with NCM analyses to yield optimized prewarning areas (known as “watches” in some countries) covering a time frame of up to 2 h and including appropriate uncertainty information. These would complement the primary warning cones, which NCM already produces, by expanding the areas that receive a warning or advanced warning and making the expected system locally less sharp (i.e., less on/off in nature). EPS data could also help improve the optimization of warning levels, notably in preventing NCM from starting new warnings often with the lowest (yellow) category, when newly developing convective cells first appear, in unstable situations where severe category storms (orange or higher) are very likely. This would also improve the usability of the output warnings.

Detection and nowcasting algorithms based on radar and satellite data will undoubtedly be improved in the near future to allow NCM to improve lead times for thunderstorm prewarnings in the initial stages of storm development. Currently, NCM has to wait until the first lightning flashes have occurred and/or precipitation has become sufficiently strong to trigger cell detection systems, before giving out primary warnings, while noting that the proposed EPS-based prewarnings have not yet been implemented. Improvements in cell detection and nowcasting with KONRAD3D (Werner 2017) could help to increase warning lead times for NowCastMIX further. Rapid-scan satellite data certainly have potential for detecting new cells during the convective initiation (CI) phase (e.g., Weckwerth and Parsons 2006), several minutes earlier than is currently achieved in NCM, especially with the expected higher temporal and spatial resolution of Meteosat Third Generation data. However, in spite of ongoing improvements, typical CI methods tend to have a high FAR (e.g., Walker et al. 2012). This hinders direct integration into NCM, for which robust indicators of events for triggering primary warnings are essential. Hence, a method is needed for reducing the FAR, for example, by including NWP data to identify environments more likely to lead to CI (Mecikalski et al. 2015). In this sense, satellite-based CI data may still prove useful in the future with careful tuning, despite the high FAR, as an extra input for calculating probabilities for generating prewarnings, next to EPS probabilities for severe convection and extrapolations of current storm positions.

Acknowledgments

The authors thank the many duty forecasters and supervisors in the DWD forecast center who have spent considerable time evaluating NCM output, giving ongoing feedback to the developers to help improve the system. The advice of numerous research and development specialists is also greatly appreciated with respect to implementing data from NWP, postprocessing, and remote sensing systems into NCM. Furthermore, the support of the technical staff in charge of the IT infrastructure and the NinJo workstation environment has been vital and is gratefully acknowledged.

REFERENCES

  • Baldauf, M., C. Gebhardt, S. Theis, B. Ritter, and C. Schraff, 2018: Beschreibung des operationellen Kürzestfristvorhersagemodells COSMO-D2 und COSMO-D2-EPS und seiner Ausgabe in die Datenbanken des DWD. Deutscher Wetterdienst (DWD), Offenbach, Germany, 115 pp., https://www.dwd.de/SharedDocs/downloads/DE/modelldokumentationen/nwv/cosmo_d2/cosmo_d2_dbbeschr_version_1_0_201805.html.

  • Baum, B. A., V. Tovinkere, J. Titlow, and R. M. Welch, 1997: Automated cloud classification of global AVHRR data using a fuzzy logic approach. J. Appl. Meteor., 36, 15191540, https://doi.org/10.1175/1520-0450(1997)036<1519:ACCOGA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Berenguer, M., D. Sempere-Torres, C. Corral, and R. Sánchez-Diezma, 2006: A fuzzy logic technique for identifying nonprecipitating echoes in radar scans. J. Atmos. Oceanic Technol., 23, 11571180, https://doi.org/10.1175/JTECH1914.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Betz, H. D., K. Schmidt, P. Laroche, P. Blanchet, W. P. Oettinger, E. Defer, Z. Dziewit, and J. Konarski, 2009: LINET—An international lightning detection network in Europe. Atmos. Res., 91, 564573, https://doi.org/10.1016/j.atmosres.2008.06.012.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Blahak, U., and Coauthors, 2017: Development of a new seamless prediction system for very short range convective-scale forecasting at DWD. 38th Conf. on Radar Meteorology, Chicago, IL, Amer. Meteor. Soc., KS3.1, https://ams.confex.com/ams/38RADAR/webprogram/Paper321209.html.

  • Bunkers, M. J., B. A. Klimowski, J. W. Zeitler, R. L. Thompson, and M. L. Weisman, 2000: Predicting supercell motion using a new hodograph technique. Wea. Forecasting, 15, 6179, https://doi.org/10.1175/1520-0434(2000)015<0061:PSMUAN>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cox, I. J., and S. L. Hingorani, 1996: An efficient implementation of Reid’s multiple hypothesis tracking algorithm and its evaluation for the purpose of visual tracking. IEEE Trans. Pattern Anal. Mach. Intell., 18, 138150, https://doi.org/10.1109/34.481539.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm Identification, Tracking, Analysis and Nowcasting—A radar-based methodology. J. Atmos. Oceanic Technol., 10, 785797, https://doi.org/10.1175/1520-0426(1993)010<0785:TTITAA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Doswell, C. A., III, and D. M. Schultz, 2006: On the use of indices and parameters in forecasting severe storms. Electron. J. Severe Storms Meteor., 1 (3), http://www.ejssm.org/ojs/index.php/ejssm/article/viewArticle/11/12.

    • Search Google Scholar
    • Export Citation
  • Dotzek, N., 2001: Tornadoes in Germany. Atmos. Res., 56, 233251, https://doi.org/10.1016/S0169-8095(00)00075-2.

  • Ester, M., H. P. Kriegel, J. Sander, and X. Xu, 1996: A density-based algorithm for discovering clusters in large spatial databases with noise. Proc. Second Int. Conf. on Knowledge Discovery and Data Mining, Portland, OR, AAAI, 226–231, https://www.aaai.org/Papers/KDD/1996/KDD96-037.pdf.

  • Ferro, C. A. T., and D. B. Stephenson, 2011: Extremal dependence indices: Improved verification measures for deterministic forecasts of rare binary events. Wea. Forecasting, 26, 699713, https://doi.org/10.1175/WAF-D-10-05030.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Geerts, B., 2001: Estimating downburst-related maximum surface wind speeds by means of proximity soundings in New South Wales, Australia. Wea. Forecasting, 16, 261269, https://doi.org/10.1175/1520-0434(2001)016<0261:EDRMSW>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Graham, R. A. and M. Struthwolf, 1999: VIL density as a potential hail indicator across northern Utah. NWS Western Region Tech. Attachment 99-02, 8 pp., https://www.weather.gov/media/wrh/online_publications/TAs/ta9902.pdf.

  • Greene, D. R., and R. A. Clark, 1972: Vertically integrated liquid water—A new analysis tool. Mon. Wea. Rev., 100, 548552, https://doi.org/10.1175/1520-0493(1972)100<0548:VILWNA>2.3.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Groenemeijer, P., and T. Kühne, 2014: A climatology of tornadoes in Europe: Results from the European Severe Weather Database. Mon. Wea. Rev., 142, 47754790, https://doi.org/10.1175/MWR-D-14-00107.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Groenemeijer, P., and Coauthors, 2017: Severe convective storms in Europe: Ten years of research at the European Severe Storms Laboratory. Bull. Amer. Meteor. Soc., 98, 26412651, https://doi.org/10.1175/BAMS-D-16-0067.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hansen, B. K., 2007: A fuzzy logic–based analog forecasting system for ceiling and visibility. Wea. Forecasting, 22, 13191330, https://doi.org/10.1175/2007WAF2006017.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Heizenreder, D., and S. Haucke, 2009: Das meteorologische Visualisierungs- und Produktionssystem “NinJo.” Promet, 35, 5769.

  • Heizenreder, D., P. Joe, T. Hewson, L. Wilson, P. Davies, and E. De Coning, 2015: Development of applications towards a high-impact weather forecast system. Seamless Prediction of the Earth System from Minutes to Months, G. Brunet, S. Jones, and P. M. Ruti, Eds., WMO 1156, 419–443.

  • Helmert, K., and Coauthors, 2014: DWDs new radar network and post-processing algorithm chain. Proc. Eighth European Conf. on Radar in Meteorology and Hydrology, Garmisch-Partenkirchen, Germany, DWD–DLR, 4.4, http://www.pa.op.dlr.de/erad2014/programme/ShortAbstracts/237_short.pdf.

  • Hengstebeck, T., K. Wapler, D. Heizenreder, and P. Joe, 2018: Radar network–based detection of mesocyclones at the German Meteorological Service. J. Atmos. Oceanic Technol., 35, 299321, https://doi.org/10.1175/JTECH-D-16-0230.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hering, A., C. Morel, G. Galli, S. Sénesi, P. Ambrosetti, and M. Boscacci, 2004: Nowcasting thunderstorms in the Alpine region using a radar based adaptive thresholding scheme. Proc. Third European Conf. on Radar Meteorology and COST-717 Final Seminar, Visby, Sweden, KNMI, 206–211.

  • Hoffmann, J. M., 2008: Entwicklung und Anwendung von statistischen Vorhersage-Interpretations-Verfahren für Gewitternowcasting und Unwetterwarnungen unter Einbeziehung von Fernerkundungsdaten. Ph.D. dissertation, Freie Universität Berlin, Berlin, Germany, 205 pp., https://refubium.fu-berlin.de/handle/fub188/11705?show=full.

  • Joe, P., H.-J. Koppert, D. Heizenreder, B. Erbshaeusser, W. Raatz, B. Reichert, and M. Rohn, 2005: Severe weather forecasting tools in the NinJo workstation. World Weather Research Programme’s Symp. on Nowcasting and Very Short Range Forecasting, Toulouse, France, WMO, 7.13, http://www.ninjo-workstation.com/fileadmin/files/downloads/publications/Pub_Joe_Toulouse_Severe_Weather_Forecasting.pdf.

  • Joe, P., and Coauthors, 2012: Automated processing of Doppler radar data for severe weather warnings. Doppler Radar Observations—Weather Radar, Wind Profiler, Ionospheric Radar, and Other Advanced Applications, J. Bech and J. L. Chau, Eds., InTech Open, 33–74, https://doi.org/10.5772/39058.

    • Crossref
    • Export Citation
  • Johnson, J., P. MacKeen, A. Witt, M. DeWayne, E. Stumpf, M. Eilts, and K. Thomas, 1998: The storm cell identification and tracking algorithm: An enhanced WSR-88D algorithm. Wea. Forecasting, 13, 263276, https://doi.org/10.1175/1520-0434(1998)013<0263:TSCIAT>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koppert, H.-J., 2014: Designing semi-automatic systems for weather warning operations. World Weather Open Scientific Conf., Montreal, QC, Canada, WMO World Weather Research Programme, SCIPS177.01.

  • Koppert, H.-J., T. S. Pedersen, B. Zürcher, and P. Joe, 2004: How to make an international meteorological workstation project successful. 20th Int. Conf. on Interactive Information and Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Seattle, WA, Amer. Meteor. Soc., 11.1, https://ams.confex.com/ams/84Annual/techprogram/paper_71789.htm.

  • Lakshmanan, V., T. Smith, G. J. Stumpf, and K. Hondl, 2007: The Warning Decision Support System–Integrated Information. Wea. Forecasting, 22, 596612, https://doi.org/10.1175/WAF1009.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lang, P., 2001: Cell tracking and warning indicators derived from operational radar products. 30th Int. Conf. on Radar Meteorology, Munich, Germany, Amer. Meteor. Soc., 245–247.

  • Markowski, P. M., and N. Dotzek, 2011: A numerical study of the effects of orography on supercells. Atmos. Res., 100, 457478, https://doi.org/10.1016/j.atmosres.2010.12.027.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0–1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather prediction model data. J. Appl. Meteor. Climatol., 54, 10391059, https://doi.org/10.1175/JAMC-D-14-0129.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mitra, A. K., S. Nath, and A. K. Sharma, 2008: Fog forecasting using rule-based fuzzy inference system. J. Indian Soc. Remote Sens., 36, 243253, https://doi.org/10.1007/s12524-008-0025-z.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mueller, C., T. Saxen, R. Roberts, J. Wilson, T. Betancourt, S. Dettling, N. Oien, and J. Yee, 2003: NCAR Auto-Nowcast System. Wea. Forecasting, 18, 545561, https://doi.org/10.1175/1520-0434(2003)018<0545:NAS>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Murtha, J., 1995: Applications of fuzzy logic in operational meteorology. Scientific Services and Professional Development Newsletter, Canadian Forces Weather Service, Ottawa, ON, Canada, 42–54, http://www.chebucto.ns.ca/Science/AIMET/archive/murtha.pdf.

  • OASIS Emergency Management Technical Committee, 2010: Common Altering Protocol version 1.2. OASIS, Burlington, MA, http://docs.oasis-open.org/emergency/cap/v1.2/CAP-v1.2-os.html.

  • Pierce, C. E., P. J. Hardaker, C. G. Collier, and C. M. Haggett, 2000: GANDOLF: A system for generating automated nowcasts of convective precipitation. Meteor. Appl., 7, 341360, https://doi.org/10.1017/S135048270000164X.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Púčik, T., and Coauthors, 2017: Future changes in European severe convection environments in a regional climate model ensemble. J. Climate, 30, 67716794, https://doi.org/10.1175/JCLI-D-16-0777.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rathmann, N., and M. Mott, 2012: Effective radar algorithm software development at the DWD. Seventh European Conf. on Radar in Meteorology and Hydrology, Toulouse, France, Météo-France, http://www.meteo.fr/cic/meetings/2012/ERAD/extended_abs/NET_316_ext_abs.pdf.

  • Reichert, B. K., 2010: AutoWARN—Automatische Unterstützung der Herausgabe von Unwetterwarnungen. Promet, 35, 98103.

  • Reichert, B. K., 2017: Forecasting and nowcasting severe weather using the operational warning decision support system AutoWARN at DWD. Ninth European Conf. on Severe Storms, Pula, Croatia, European Severe Storms Laboratory, ECSS2017-2.

  • Reichert, B. K., and Coauthors, 2015: The Decision Support System AutoWARN for the Weather Warning Service at DWD. EMS Annual Meeting Abstracts, Vol. 12, EMS2015-221-2, http://meetingorganizer.copernicus.org/EMS2015/EMS2015-221-2.pdf.

  • Reid, D., 1979: An algorithm for tracking multiple targets. IEEE Trans. Automat. Contr., 24, 843854, https://doi.org/10.1109/TAC.1979.1102177.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Roberts, R. D., D. Burgess, and M. Meister, 2006: Developing tools for nowcasting storm severity. Wea. Forecasting, 21, 540558, https://doi.org/10.1175/WAF930.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Romero, R., M. Gayà, and C. A. Doswell III, 2007: European climatology of severe convective storm environmental parameters: A test for significant tornado events. Atmos. Res., 83, 389404, https://doi.org/10.1016/j.atmosres.2005.06.011.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rossi, P. J., V. Hasu, J. Koistinen, D. Moisseev, A. Mäkelä, and E. Saltikoff, 2014: Analysis of a statistically initialized fuzzy logic scheme for classifying the severity of convective storms in Finland. Meteor. Appl., 21, 656674, https://doi.org/10.1002/met.1389.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schröder, G., R. Hess, B. Reichert, and D. Heizenreder, 2014: Automated weather warning proposals based on post-processed numerical weather forecasts. World Weather Open Scientific Conf., Montreal, QC, Canada, WMO World Weather Research Programme, SCI-PS177.04.

  • Shannon, C. E., 1948: A mathematical theory of communication. Bell System Tech. J., 27, 379423, 623–656.

  • Smith, B. T., R. L. Thompson, J. S. Grams, and C. Broyles, 2012: Convective modes for significant severe thunderstorms in the contiguous United States. Part I: Storm classification and climatology. Wea. Forecasting, 27, 11141135, https://doi.org/10.1175/WAF-D-11-00115.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Steinert, J., 2014: Hydrometeor classification for the DWD Weather Radar Network: First verification results. Eighth European Conf. on Radar in Meteorology and Hydrology, Garmisch-Partenkirchen, Germany, DWD–DLR, http://www.pa.op.dlr.de/erad2014/programme/ExtendedAbstracts/225_Steinert.pdf.

  • Sun, J., and Coauthors, 2014: Use of NWP for nowcasting convective precipitation: Recent progress and challenges. Bull. Amer. Meteor. Soc., 95, 409426, https://doi.org/10.1175/BAMS-D-11-00263.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thompson, R. L., B. T. Smith, J. S. Grams, A. R. Dean, and C. Broyles, 2010: Climatology of near-storm environments with convective modes for significant severe thunderstorms in the contiguous United States. 25th Conf. on Severe Local Storms, Denver, CO, Amer. Meteor. Soc., 16B.6, https://ams.confex.com/ams/25SLS/webprogram/Paper175727.html.

  • Thompson, R. L., B. T. Smith, J. S. Grams, A. R. Dean, and C. Broyles, 2012: Convective modes for significant severe thunderstorms in the contiguous United States. Part II: Supercell and QLCS tornado environments. Wea. Forecasting, 27, 11361154, https://doi.org/10.1175/WAF-D-11-00116.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Topchy, A., B. Minaei-Bidgoli, A. K. Jain, and W. F. Punch, 2004: Adaptive clustering ensembles. Proc. 17th Int. Conf. on Pattern Recognition, Cambridge, United, Kingdom, IEEE, 272–275.

    • Crossref
    • Export Citation
  • Trepte, S., 2014: Gewitterzellenvorhersage mit CellMOS. Die Thüringische Sintflut von 1613 und ihre Folgen für heute, Schriften der Deutsche Wasserhistorische Gesellschaft e. V., Band 22, 141–156.

  • Walker, J. R., W. M. MacKenzie, J. R. Mecikalski, and C. P. Jewett, 2012: An enhanced geostationary satellite-based convective initiation algorithm for 0–2-h nowcasting with object tracking. J. Appl. Meteor. Climatol., 51, 19311949, https://doi.org/10.1175/JAMC-D-11-0246.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., 2013: High-resolution climatology of lightning characteristics within central Europe. Meteor. Atmos. Phys., 122, 175184, https://doi.org/10.1007/s00703-013-0285-1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., 2017: The life-cycle of hailstorms: Lightning, radar reflectivity and rotation characteristics. Atmos. Res., 193, 6072, https://doi.org/10.1016/j.atmosres.2017.04.009.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., M. Goeber, and S. Trepte, 2012: Comparative verification of different nowcasting systems to support optimization of thunderstorm warnings. Adv. Sci. Res., 8, 121127, https://doi.org/10.5194/asr-8-121-2012.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., T. Hengstebeck, and P. Groenemeijer, 2016: Mesocyclones in central Europe as seen by radar. Atmos. Res., 168, 112120, https://doi.org/10.1016/j.atmosres.2015.08.023.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weckwerth, T. M., and D. B. Parsons, 2006: A review of convection initiation and motivation for IHOP_2002. Mon. Wea. Rev., 134, 522, https://doi.org/10.1175/MWR3067.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weigl, E., and T. Winterrath, 2010: Radargestützte Niederschlagsanalyse und –vorhersage (RADOLAN, RADVOR-OP). Promet, 35, 7886.

  • Werner, M., 2017: KONRAD3D: A new tool for detection and nowcasting of convective cells at DWD. Second European Nowcasting Conf., Offenbach, Germany, EUMETNET, 15–16, http://eumetnet.eu/wp-content/uploads/2017/07/enc_book_of_abstract.pdf.

  • Wilson, J. W., N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998: Nowcasting thunderstorms: A status report. Bull. Amer. Meteor. Soc., 79, 20792099, https://doi.org/10.1175/1520-0477(1998)079<2079:NTASR>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilson, J. W., Y. Feng, M. Chen, and R. D. Roberts, 2010: Nowcasting challenges during the Beijing Olympics: Successes, failures, and implications for future nowcasting systems. Wea. Forecasting, 25, 16911714, https://doi.org/10.1175/2010WAF2222417.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Winterrath, T., and W. Rosenow, 2007: A new module for the tracking of radar-derived precipitation with model-derived winds. Adv. Geosci., 10, 7783, https://doi.org/10.5194/adgeo-10-77-2007.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Winterrath, T., W. Rosenow, and E. Weigl, 2012: On the DWD quantitative precipitation analysis and nowcasting system for real-time application in German flood risk management. Proc. Symp. on Weather Radar and Hydrology, Exeter, United Kingdom, IAHS Publ. 351, 323–329, https://www.dwd.de/DE/leistungen/radolan/radolan_info/Winterrath_German_flood_risk_management_pdf.pdf?__blob=publicationFile&v=4.

  • Zadeh, L. A., 1965: Fuzzy sets. Inf. Control, 8, 338353, https://doi.org/10.1016/S0019-9958(65)90241-X.

  • Zängl, G., D. Reinert, P. Ripodas, and M. Baldauf, 2015: The ICON (ICOsahedral Non-hydrostatic) modelling framework of DWD and MPI-M: Description of the non-hydrostatic dynamical core. Quart. J. Roy. Meteor. Soc., 141, 563579, https://doi.org/10.1002/qj.2378.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zrnić, D. S., D. W. Burgess, and L. D. Hennington, 1985: Automatic detection of mesocyclonic shear with Doppler radar. J. Atmos. Oceanic Technol., 2, 425438, https://doi.org/10.1175/1520-0426(1985)002<0425:ADOMSW>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
Save
  • Baldauf, M., C. Gebhardt, S. Theis, B. Ritter, and C. Schraff, 2018: Beschreibung des operationellen Kürzestfristvorhersagemodells COSMO-D2 und COSMO-D2-EPS und seiner Ausgabe in die Datenbanken des DWD. Deutscher Wetterdienst (DWD), Offenbach, Germany, 115 pp., https://www.dwd.de/SharedDocs/downloads/DE/modelldokumentationen/nwv/cosmo_d2/cosmo_d2_dbbeschr_version_1_0_201805.html.

  • Baum, B. A., V. Tovinkere, J. Titlow, and R. M. Welch, 1997: Automated cloud classification of global AVHRR data using a fuzzy logic approach. J. Appl. Meteor., 36, 15191540, https://doi.org/10.1175/1520-0450(1997)036<1519:ACCOGA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Berenguer, M., D. Sempere-Torres, C. Corral, and R. Sánchez-Diezma, 2006: A fuzzy logic technique for identifying nonprecipitating echoes in radar scans. J. Atmos. Oceanic Technol., 23, 11571180, https://doi.org/10.1175/JTECH1914.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Betz, H. D., K. Schmidt, P. Laroche, P. Blanchet, W. P. Oettinger, E. Defer, Z. Dziewit, and J. Konarski, 2009: LINET—An international lightning detection network in Europe. Atmos. Res., 91, 564573, https://doi.org/10.1016/j.atmosres.2008.06.012.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Blahak, U., and Coauthors, 2017: Development of a new seamless prediction system for very short range convective-scale forecasting at DWD. 38th Conf. on Radar Meteorology, Chicago, IL, Amer. Meteor. Soc., KS3.1, https://ams.confex.com/ams/38RADAR/webprogram/Paper321209.html.

  • Bunkers, M. J., B. A. Klimowski, J. W. Zeitler, R. L. Thompson, and M. L. Weisman, 2000: Predicting supercell motion using a new hodograph technique. Wea. Forecasting, 15, 6179, https://doi.org/10.1175/1520-0434(2000)015<0061:PSMUAN>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cox, I. J., and S. L. Hingorani, 1996: An efficient implementation of Reid’s multiple hypothesis tracking algorithm and its evaluation for the purpose of visual tracking. IEEE Trans. Pattern Anal. Mach. Intell., 18, 138150, https://doi.org/10.1109/34.481539.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Dixon, M., and G. Wiener, 1993: TITAN: Thunderstorm Identification, Tracking, Analysis and Nowcasting—A radar-based methodology. J. Atmos. Oceanic Technol., 10, 785797, https://doi.org/10.1175/1520-0426(1993)010<0785:TTITAA>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Doswell, C. A., III, and D. M. Schultz, 2006: On the use of indices and parameters in forecasting severe storms. Electron. J. Severe Storms Meteor., 1 (3), http://www.ejssm.org/ojs/index.php/ejssm/article/viewArticle/11/12.

    • Search Google Scholar
    • Export Citation
  • Dotzek, N., 2001: Tornadoes in Germany. Atmos. Res., 56, 233251, https://doi.org/10.1016/S0169-8095(00)00075-2.

  • Ester, M., H. P. Kriegel, J. Sander, and X. Xu, 1996: A density-based algorithm for discovering clusters in large spatial databases with noise. Proc. Second Int. Conf. on Knowledge Discovery and Data Mining, Portland, OR, AAAI, 226–231, https://www.aaai.org/Papers/KDD/1996/KDD96-037.pdf.

  • Ferro, C. A. T., and D. B. Stephenson, 2011: Extremal dependence indices: Improved verification measures for deterministic forecasts of rare binary events. Wea. Forecasting, 26, 699713, https://doi.org/10.1175/WAF-D-10-05030.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Geerts, B., 2001: Estimating downburst-related maximum surface wind speeds by means of proximity soundings in New South Wales, Australia. Wea. Forecasting, 16, 261269, https://doi.org/10.1175/1520-0434(2001)016<0261:EDRMSW>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Graham, R. A. and M. Struthwolf, 1999: VIL density as a potential hail indicator across northern Utah. NWS Western Region Tech. Attachment 99-02, 8 pp., https://www.weather.gov/media/wrh/online_publications/TAs/ta9902.pdf.

  • Greene, D. R., and R. A. Clark, 1972: Vertically integrated liquid water—A new analysis tool. Mon. Wea. Rev., 100, 548552, https://doi.org/10.1175/1520-0493(1972)100<0548:VILWNA>2.3.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Groenemeijer, P., and T. Kühne, 2014: A climatology of tornadoes in Europe: Results from the European Severe Weather Database. Mon. Wea. Rev., 142, 47754790, https://doi.org/10.1175/MWR-D-14-00107.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Groenemeijer, P., and Coauthors, 2017: Severe convective storms in Europe: Ten years of research at the European Severe Storms Laboratory. Bull. Amer. Meteor. Soc., 98, 26412651, https://doi.org/10.1175/BAMS-D-16-0067.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hansen, B. K., 2007: A fuzzy logic–based analog forecasting system for ceiling and visibility. Wea. Forecasting, 22, 13191330, https://doi.org/10.1175/2007WAF2006017.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Heizenreder, D., and S. Haucke, 2009: Das meteorologische Visualisierungs- und Produktionssystem “NinJo.” Promet, 35, 5769.

  • Heizenreder, D., P. Joe, T. Hewson, L. Wilson, P. Davies, and E. De Coning, 2015: Development of applications towards a high-impact weather forecast system. Seamless Prediction of the Earth System from Minutes to Months, G. Brunet, S. Jones, and P. M. Ruti, Eds., WMO 1156, 419–443.

  • Helmert, K., and Coauthors, 2014: DWDs new radar network and post-processing algorithm chain. Proc. Eighth European Conf. on Radar in Meteorology and Hydrology, Garmisch-Partenkirchen, Germany, DWD–DLR, 4.4, http://www.pa.op.dlr.de/erad2014/programme/ShortAbstracts/237_short.pdf.

  • Hengstebeck, T., K. Wapler, D. Heizenreder, and P. Joe, 2018: Radar network–based detection of mesocyclones at the German Meteorological Service. J. Atmos. Oceanic Technol., 35, 299321, https://doi.org/10.1175/JTECH-D-16-0230.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hering, A., C. Morel, G. Galli, S. Sénesi, P. Ambrosetti, and M. Boscacci, 2004: Nowcasting thunderstorms in the Alpine region using a radar based adaptive thresholding scheme. Proc. Third European Conf. on Radar Meteorology and COST-717 Final Seminar, Visby, Sweden, KNMI, 206–211.

  • Hoffmann, J. M., 2008: Entwicklung und Anwendung von statistischen Vorhersage-Interpretations-Verfahren für Gewitternowcasting und Unwetterwarnungen unter Einbeziehung von Fernerkundungsdaten. Ph.D. dissertation, Freie Universität Berlin, Berlin, Germany, 205 pp., https://refubium.fu-berlin.de/handle/fub188/11705?show=full.

  • Joe, P., H.-J. Koppert, D. Heizenreder, B. Erbshaeusser, W. Raatz, B. Reichert, and M. Rohn, 2005: Severe weather forecasting tools in the NinJo workstation. World Weather Research Programme’s Symp. on Nowcasting and Very Short Range Forecasting, Toulouse, France, WMO, 7.13, http://www.ninjo-workstation.com/fileadmin/files/downloads/publications/Pub_Joe_Toulouse_Severe_Weather_Forecasting.pdf.

  • Joe, P., and Coauthors, 2012: Automated processing of Doppler radar data for severe weather warnings. Doppler Radar Observations—Weather Radar, Wind Profiler, Ionospheric Radar, and Other Advanced Applications, J. Bech and J. L. Chau, Eds., InTech Open, 33–74, https://doi.org/10.5772/39058.

    • Crossref
    • Export Citation
  • Johnson, J., P. MacKeen, A. Witt, M. DeWayne, E. Stumpf, M. Eilts, and K. Thomas, 1998: The storm cell identification and tracking algorithm: An enhanced WSR-88D algorithm. Wea. Forecasting, 13, 263276, https://doi.org/10.1175/1520-0434(1998)013<0263:TSCIAT>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koppert, H.-J., 2014: Designing semi-automatic systems for weather warning operations. World Weather Open Scientific Conf., Montreal, QC, Canada, WMO World Weather Research Programme, SCIPS177.01.

  • Koppert, H.-J., T. S. Pedersen, B. Zürcher, and P. Joe, 2004: How to make an international meteorological workstation project successful. 20th Int. Conf. on Interactive Information and Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Seattle, WA, Amer. Meteor. Soc., 11.1, https://ams.confex.com/ams/84Annual/techprogram/paper_71789.htm.

  • Lakshmanan, V., T. Smith, G. J. Stumpf, and K. Hondl, 2007: The Warning Decision Support System–Integrated Information. Wea. Forecasting, 22, 596612, https://doi.org/10.1175/WAF1009.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lang, P., 2001: Cell tracking and warning indicators derived from operational radar products. 30th Int. Conf. on Radar Meteorology, Munich, Germany, Amer. Meteor. Soc., 245–247.

  • Markowski, P. M., and N. Dotzek, 2011: A numerical study of the effects of orography on supercells. Atmos. Res., 100, 457478, https://doi.org/10.1016/j.atmosres.2010.12.027.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0–1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather prediction model data. J. Appl. Meteor. Climatol., 54, 10391059, https://doi.org/10.1175/JAMC-D-14-0129.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mitra, A. K., S. Nath, and A. K. Sharma, 2008: Fog forecasting using rule-based fuzzy inference system. J. Indian Soc. Remote Sens., 36, 243253, https://doi.org/10.1007/s12524-008-0025-z.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mueller, C., T. Saxen, R. Roberts, J. Wilson, T. Betancourt, S. Dettling, N. Oien, and J. Yee, 2003: NCAR Auto-Nowcast System. Wea. Forecasting, 18, 545561, https://doi.org/10.1175/1520-0434(2003)018<0545:NAS>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Murtha, J., 1995: Applications of fuzzy logic in operational meteorology. Scientific Services and Professional Development Newsletter, Canadian Forces Weather Service, Ottawa, ON, Canada, 42–54, http://www.chebucto.ns.ca/Science/AIMET/archive/murtha.pdf.

  • OASIS Emergency Management Technical Committee, 2010: Common Altering Protocol version 1.2. OASIS, Burlington, MA, http://docs.oasis-open.org/emergency/cap/v1.2/CAP-v1.2-os.html.

  • Pierce, C. E., P. J. Hardaker, C. G. Collier, and C. M. Haggett, 2000: GANDOLF: A system for generating automated nowcasts of convective precipitation. Meteor. Appl., 7, 341360, https://doi.org/10.1017/S135048270000164X.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Púčik, T., and Coauthors, 2017: Future changes in European severe convection environments in a regional climate model ensemble. J. Climate, 30, 67716794, https://doi.org/10.1175/JCLI-D-16-0777.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rathmann, N., and M. Mott, 2012: Effective radar algorithm software development at the DWD. Seventh European Conf. on Radar in Meteorology and Hydrology, Toulouse, France, Météo-France, http://www.meteo.fr/cic/meetings/2012/ERAD/extended_abs/NET_316_ext_abs.pdf.

  • Reichert, B. K., 2010: AutoWARN—Automatische Unterstützung der Herausgabe von Unwetterwarnungen. Promet, 35, 98103.

  • Reichert, B. K., 2017: Forecasting and nowcasting severe weather using the operational warning decision support system AutoWARN at DWD. Ninth European Conf. on Severe Storms, Pula, Croatia, European Severe Storms Laboratory, ECSS2017-2.

  • Reichert, B. K., and Coauthors, 2015: The Decision Support System AutoWARN for the Weather Warning Service at DWD. EMS Annual Meeting Abstracts, Vol. 12, EMS2015-221-2, http://meetingorganizer.copernicus.org/EMS2015/EMS2015-221-2.pdf.

  • Reid, D., 1979: An algorithm for tracking multiple targets. IEEE Trans. Automat. Contr., 24, 843854, https://doi.org/10.1109/TAC.1979.1102177.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Roberts, R. D., D. Burgess, and M. Meister, 2006: Developing tools for nowcasting storm severity. Wea. Forecasting, 21, 540558, https://doi.org/10.1175/WAF930.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Romero, R., M. Gayà, and C. A. Doswell III, 2007: European climatology of severe convective storm environmental parameters: A test for significant tornado events. Atmos. Res., 83, 389404, https://doi.org/10.1016/j.atmosres.2005.06.011.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rossi, P. J., V. Hasu, J. Koistinen, D. Moisseev, A. Mäkelä, and E. Saltikoff, 2014: Analysis of a statistically initialized fuzzy logic scheme for classifying the severity of convective storms in Finland. Meteor. Appl., 21, 656674, https://doi.org/10.1002/met.1389.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schröder, G., R. Hess, B. Reichert, and D. Heizenreder, 2014: Automated weather warning proposals based on post-processed numerical weather forecasts. World Weather Open Scientific Conf., Montreal, QC, Canada, WMO World Weather Research Programme, SCI-PS177.04.

  • Shannon, C. E., 1948: A mathematical theory of communication. Bell System Tech. J., 27, 379423, 623–656.

  • Smith, B. T., R. L. Thompson, J. S. Grams, and C. Broyles, 2012: Convective modes for significant severe thunderstorms in the contiguous United States. Part I: Storm classification and climatology. Wea. Forecasting, 27, 11141135, https://doi.org/10.1175/WAF-D-11-00115.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Steinert, J., 2014: Hydrometeor classification for the DWD Weather Radar Network: First verification results. Eighth European Conf. on Radar in Meteorology and Hydrology, Garmisch-Partenkirchen, Germany, DWD–DLR, http://www.pa.op.dlr.de/erad2014/programme/ExtendedAbstracts/225_Steinert.pdf.

  • Sun, J., and Coauthors, 2014: Use of NWP for nowcasting convective precipitation: Recent progress and challenges. Bull. Amer. Meteor. Soc., 95, 409426, https://doi.org/10.1175/BAMS-D-11-00263.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thompson, R. L., B. T. Smith, J. S. Grams, A. R. Dean, and C. Broyles, 2010: Climatology of near-storm environments with convective modes for significant severe thunderstorms in the contiguous United States. 25th Conf. on Severe Local Storms, Denver, CO, Amer. Meteor. Soc., 16B.6, https://ams.confex.com/ams/25SLS/webprogram/Paper175727.html.

  • Thompson, R. L., B. T. Smith, J. S. Grams, A. R. Dean, and C. Broyles, 2012: Convective modes for significant severe thunderstorms in the contiguous United States. Part II: Supercell and QLCS tornado environments. Wea. Forecasting, 27, 11361154, https://doi.org/10.1175/WAF-D-11-00116.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Topchy, A., B. Minaei-Bidgoli, A. K. Jain, and W. F. Punch, 2004: Adaptive clustering ensembles. Proc. 17th Int. Conf. on Pattern Recognition, Cambridge, United, Kingdom, IEEE, 272–275.

    • Crossref
    • Export Citation
  • Trepte, S., 2014: Gewitterzellenvorhersage mit CellMOS. Die Thüringische Sintflut von 1613 und ihre Folgen für heute, Schriften der Deutsche Wasserhistorische Gesellschaft e. V., Band 22, 141–156.

  • Walker, J. R., W. M. MacKenzie, J. R. Mecikalski, and C. P. Jewett, 2012: An enhanced geostationary satellite-based convective initiation algorithm for 0–2-h nowcasting with object tracking. J. Appl. Meteor. Climatol., 51, 19311949, https://doi.org/10.1175/JAMC-D-11-0246.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., 2013: High-resolution climatology of lightning characteristics within central Europe. Meteor. Atmos. Phys., 122, 175184, https://doi.org/10.1007/s00703-013-0285-1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., 2017: The life-cycle of hailstorms: Lightning, radar reflectivity and rotation characteristics. Atmos. Res., 193, 6072, https://doi.org/10.1016/j.atmosres.2017.04.009.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., M. Goeber, and S. Trepte, 2012: Comparative verification of different nowcasting systems to support optimization of thunderstorm warnings. Adv. Sci. Res., 8, 121127, https://doi.org/10.5194/asr-8-121-2012.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wapler, K., T. Hengstebeck, and P. Groenemeijer, 2016: Mesocyclones in central Europe as seen by radar. Atmos. Res., 168, 112120, https://doi.org/10.1016/j.atmosres.2015.08.023.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weckwerth, T. M., and D. B. Parsons, 2006: A review of convection initiation and motivation for IHOP_2002. Mon. Wea. Rev., 134, 522, https://doi.org/10.1175/MWR3067.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Weigl, E., and T. Winterrath, 2010: Radargestützte Niederschlagsanalyse und –vorhersage (RADOLAN, RADVOR-OP). Promet, 35, 7886.

  • Werner, M., 2017: KONRAD3D: A new tool for detection and nowcasting of convective cells at DWD. Second European Nowcasting Conf., Offenbach, Germany, EUMETNET, 15–16, http://eumetnet.eu/wp-content/uploads/2017/07/enc_book_of_abstract.pdf.

  • Wilson, J. W., N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, 1998: Nowcasting thunderstorms: A status report. Bull. Amer. Meteor. Soc., 79, 20792099, https://doi.org/10.1175/1520-0477(1998)079<2079:NTASR>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilson, J. W., Y. Feng, M. Chen, and R. D. Roberts, 2010: Nowcasting challenges during the Beijing Olympics: Successes, failures, and implications for future nowcasting systems. Wea. Forecasting, 25, 16911714, https://doi.org/10.1175/2010WAF2222417.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Winterrath, T., and W. Rosenow, 2007: A new module for the tracking of radar-derived precipitation with model-derived winds. Adv. Geosci., 10, 7783, https://doi.org/10.5194/adgeo-10-77-2007.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Winterrath, T., W. Rosenow, and E. Weigl, 2012: On the DWD quantitative precipitation analysis and nowcasting system for real-time application in German flood risk management. Proc. Symp. on Weather Radar and Hydrology, Exeter, United Kingdom, IAHS Publ. 351, 323–329, https://www.dwd.de/DE/leistungen/radolan/radolan_info/Winterrath_German_flood_risk_management_pdf.pdf?__blob=publicationFile&v=4.

  • Zadeh, L. A., 1965: Fuzzy sets. Inf. Control, 8, 338353, https://doi.org/10.1016/S0019-9958(65)90241-X.

  • Zängl, G., D. Reinert, P. Ripodas, and M. Baldauf, 2015: The ICON (ICOsahedral Non-hydrostatic) modelling framework of DWD and MPI-M: Description of the non-hydrostatic dynamical core. Quart. J. Roy. Meteor. Soc., 141, 563579, https://doi.org/10.1002/qj.2378.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zrnić, D. S., D. W. Burgess, and L. D. Hennington, 1985: Automatic detection of mesocyclonic shear with Doppler radar. J. Atmos. Oceanic Technol., 2, 425438, https://doi.org/10.1175/1520-0426(1985)002<0425:ADOMSW>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Fig. 1.

    The AutoWARN process with arrows indicating the data flow direction, showing (left) the primary input data source; (center left) the core processing and editing components ASG, ASE, and OPG; and (center right) the postprocessing tool PVW for (right) the final dissemination of warnings to customers.

  • Fig. 2.

    Schematic diagram for NowCastMIX. (left) The various input datasets (dark blue boxes). (center) The core processes that operate on the input data (white boxes). (right) The primary output targets (customers) (dark blue boxes).

  • Fig. 3.

    Inner blue square enclosing Germany: RADOLAN domain (900 km × 900 km) used for the AutoWARN output of NCM. The whole area shown above is the European Radar Composite (EuRadCOM) area (2444 km × 2176 km) used for the output of NCM-Aviation, covering most of western and central Europe, as indicated in section 5.

  • Fig. 4.

    Fuzzy logic hierarchy in NCM. Dark blue cells represent the various input datasets that are processed in the first level of fuzzy sets (light blue). These yield respective fuzzy potentials (fp) that feed into higher levels of fuzzy sets. The highest level of fuzzy sets (light green) provides the final fuzzy potentials for the thunderstorm attributes (heavy rain, hail, and severe wind gusts) at two or three specific severity thresholds, respectively. These are then combined to generate a most appropriate final thunderstorm warning category, as indicated in Table 1.

  • Fig. 5.

    Schematic illustration showing the production of warning polygons in NCM. (a) Current positions and severity categories of thunderstorm cells are analyzed (colored regions), and the cell motion vector field is used to construct overlapping warning cones showing regions at risk of thunderstorms in the next 60 min (regions initially in gray here). (b) Clustering of analyzed cells (strongly colored areas) allows orange categories to expand over yellow categories within respective clusters, whereas red and violet categories do not extend into lower category areas. The overlapping forecast cones are now colored opaquely according to clustered category. (c) The final warning polygons for the next 60 min are now shown fully colored.

  • Fig. 6.

    Schematic diagram showing the method for assigning a typical space–time point observation to a thunderstorm cell. A report of severe weather at m minutes before the analysis time t has been received (red circle, Rt-m). Its virtual location at time t (pale red circle, R′t) is estimated via the CVF, using a vector length of m minutes. In this example, three nearby cells are present (yellow circles). Cell 1 is the most probable candidate for having triggered the severe event, given that it is much closer to the virtual report position than the other two, noting that fuzzy logic probabilities of the relevant storm attribute are also taken into account at each cell (not shown here) in the assignment algorithm.

  • Fig. 7.

    Screenshot of NowCastMIX warning cones in NinJo, overlaid with VIL (black–gray–orange–red color range) and radar-based QPE from the RADOLAN system (light blue–green–yellow–orange color range), showing fast-moving thunderstorms over the Ruhr district of northwestern Germany at 2000 UTC 21 Aug 2011. The dark red lines and patches are political state boundaries and urban conglomerations, respectively. Major rivers are indicated in blue. The east–west extent of the geographical area shown is approximately 400 km.

  • Fig. 8.

    Intranet monitor image, showing NCM output during a severe thunderstorm outbreak at the analysis time 1800 UTC 18 Aug 2017. Current thunderstorm activity is shown as solidly colored polygons whose centers are indicated by black-outlined circles of a specific severity, colored inside according to the categories (ii) on the left-hand side of the graph. The past tracks of each cluster object are shown with small joined circles. The respective tracking vectors are given as black arrows and can be compared to the gray arrows, which show the standard, smooth CVF vector field. The resulting integrated warning areas for the next 60 min are shown as opaque-colored polygons. Heavy rain areas are shown as green polygons that are hatched where they overlap with thunderstorm polygons. The availability of input data for the current NCM run is indicated at the bottom right by green (data available) or red (data missing) buttons. The total area of each warning category (km2) is indicated at the top right.

  • Fig. 9.

    Verification scores for NCM thunderstorm warnings against NCM’s own thunderstorm analyses as a function of lead time on the German administrative districts, showing POD (solid) and FAR (dashed) for all thunderstorms (yellow), orange severity storms or higher (orange), and red severity storms or higher (red).

  • Fig. 10.

    Mean verification scores on the German administrative regions for NCM thunderstorm warnings (any severity) for a lead time of T + 30 min against NCM’s own analyses, 2011–17 (April–September), showing (a) POD, (b) FAR, and (c) EDI. Values are normalized by removing the (slight) dependency on the area of the administrative districts, as determined via linear regressions, and are then shown in percent (different scales, as shown). Note that the color scale for FAR is inverted, so that warm colors represent better scores in all cases.

  • Fig. 11.

    Verification scores (POD and FAR) for thunderstorm warnings (any severity) against lightning flash data across the German administrative districts as a function of lead time, showing automatic NCM warnings (NCM, solid) and manually generated operational warnings from duty forecasters (FCR, dashed). Mean values over three convective seasons (April–September) are shown, respectively, for seasons during which forecasters (a) used older background guidance systems without AutoWARN (2012, 2013, and 2015) or (b) had the benefit of real-time direct warning proposals from AutoWARN (2014, 2016, and 2017). (c) Relative EDI for forecaster minus NCM for seasons with AutoWARN guidance (solid green) and seasons without AutoWARN guidance (dashed green). A lead time of, e.g., 30 min shown above is equivalent to T + 30 min as referred to in the text.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 5565 1980 63
PDF Downloads 2465 468 29