1 Introduction

Waterspouts in the central Mediterranean have been the subject of research as long ago as 1749 when Bošković (1749) published his dissertation about a storm that hit Rome in the night between 11 and 12 June of 1749 and damaged a large part of the town. In his dissertation, Bošković also described several waterspout outbreak events he experienced as a child and stated that waterspouts are a quite common phenomenon for the Adriatic Sea (Fig. 1). However, the first detailed preliminary climatology of this phenomenon in the eastern Adriatic region was performed considerably later, in 2016, by Renko et al. (2016); it contained 359 waterspout events for the 13-year period from 2001 to 2013. A waterspout is generally defined as any tornado over a body of water and, in its most common form, as a non-mesocyclonic tornado over water. As such, it consists of an intense columnar vortex (usually appearing as a funnel-shaped cloud) that occurs over a body of water and is connected to a cumuliform cloud (American Meteorological Society, 2017). Several studies have discussed the physical processes, thermodynamical environment and meteorological conditions favourable for waterspout development in the Adriatic Sea (e.g., Sioutas and Keul 2007; Sioutas et al. 2013; Renko et al. 2013, 2016). Waterspouts usually form in environments that include pre-existing low-level vertical vorticity along convergence boundaries (Caruso and Davies 2005), steep low-level lapse rates and sufficient moisture in lower layers (Markowski and Richardson 2009; Keul et al. 2009; Renko et al. 2016). Most of the observed waterspouts in the Adriatic are F0 and F1 category tornadoes, i.e., weak tornadoes that are non-mesocyclonic (Renko et al. 2016). Waterspouts can pose a threat to people and property, particularly if they move onshore. Therefore, forecasting tornadoes over water, especially non-mesocyclonic ones, is important but still quite difficult (Caruso and Davies 2005; Dotzek et al. 2010) and presents a great challenge in operational meteorology. Antonescu et al. (2016) noticed that the vast majority of European meteorological services do not issue warnings and forecasts for tornadoes, probably because of a lack of widespread recognition of the threat of tornadoes to Europe.

Fig. 1
figure 1

Adriatic region with locations where selected waterspout cases were observed ((c) Google)

The most common method for forecasting deep moist convection and severe thunderstorms, as well as all accompanying meteorological phenomena, such as large hail, extreme rainfall, severe wind gusts, lightning and tornadoes, is the use of thermodynamic and kinematic parameters derived from numerical weather prediction (NWP) models (Taszarek et al. 2017). Together with recognition of common surface setup favourable for non-mesocyclone tornadoes and all ingredients supportive for the development of non-mesocyclone tornadoes, waterspout forecasts can be significantly improved. Still, the variety of atmospheric conditions in which tornadoes over water can occur makes forecasting very difficult; one of the best methods is the use of a combination of real-time radar observational data with observations of the near-storm environment (Markowski and Richardson 2009). Because of the lack of a radar network on the Adriatic coast, forecasters can only use observations from meteorological stations and satellite data for diagnostics and NWP for forecasting.

In 2005, Szilagyi (2009) proposed an empirically developed method, known today as the Szilagyi Waterspout Nomogram (SWN). This method was developed as a result of continuous investigations of waterspout activity over the Great Lakes of North America and is based on a large sample of observed waterspouts. Two instability parameters were judged to be most strongly correlated with waterspout occurrence (Szilagyi 2009): the temperature difference between the water surface (SST) and the 850-hPa level (T850) (ΔT = SST−T850), which represents the lapse rate in the lower atmosphere, and the convective cloud depth (ΔZ), which represents the strength of the updraft. One wind constraint was also proposed that can be thought of as a proxy for vertical wind shear below 1500 m (Keul et al. 2009): wind at the 850-hPa level should be ≤ 40 kts.

The SWN method showed promising (although different) results for waterspout forecasting for the Adriatic Sea (Keul et al. 2009; Renko et al. 2013). In the research of Renko et al. (2013), the authors used an operational NWP model from the National Meteorological and Hydrological Service of Croatia (Ivatek-Šahdan and Tudor 2004): ALADIN/HR, which has a horizontal grid spacing of 8 km. However, despite a very high hit rate for the Adriatic in Keul et al. (2009) when 96% waterspout cases satisfied the Nomogram, Renko et al. (2013) obtained a lower hit rate of 79% by which 15 cases of 19 events were successfully forecasted by SWN. This difference in the obtained results from the SWN method motivates further research.

The work reported in this paper used the Szilagyi Waterspout Index (SWI), obtained directly from the above-mentioned SWN method (Fig. 2), in a more operational and sophisticated way. Steps forward in this work were the following: (i) the use of a mesoscale numerical weather model (Weather and Research Forecasting, WRF model) with small horizontal grid spacing (500 m); (ii) the spatio-temporal SWI distributions in the WRF model grid, which allowed more detailed analysis of the SWN and SWI; and (iii) a more refined treatment of convection processes. Whereas Renko et al. (2013) used direct model values for SWN calculation; all thermodynamic and kinematic data in this study were computed from pseudo-soundings using the Open-Source Sounding Analysis Toolkit SHARPpy (Marsh and Hart 2012). The analysis was conducted on 10 waterspout outbreaks over the period of 2013–2016, one of which was extremely unique and well-known because of the detection of 40 waterspouts.

Fig. 2
figure 2

Szilagyi Waterspout: a Nomogram, b Index (Szilagyi 2009)

2 Dataset and Methodology

2.1 Study Area and Data

The semi-enclosed Adriatic Sea is in the northern–central part of the Mediterranean bordered by large parallel mountain ranges across the Balkan and the Apennine Peninsulas (Fig. 1). Characteristics of the Adriatic are the less-complex western Adriatic coast and the very complex eastern Adriatic (mainly) steep mountainous coast. The Adriatic Sea is shallow in its upper (northern) half with a maximum depth of 50 m, which suggests potential for annual and seasonal large spatial SST variations. The SST distribution during the summer months is more or less uniform along the basin (except for some coastal sites such as Senj, Telišman Prtenjak and Grisogono 2007), whereas the spatial gradient of the SST during the winter months increases southward and towards the open sea (Supić and Orlić 1992). The usual climatological value of SST in summer (i.e., in August) is approximately 24 °C, whereas during the winter period (i.e., in February), the SST is in the range of 9–13 °C (see Fig. 4 in Supić and Orlić 1992).

The described topography of the eastern Adriatic coast also frequently generates local winds with a wide range of speeds. For example, land breezes and weak bora flows (e.g., Telišman Prtenjak and Grisogono 2007; Belušić et al. 2017) often provide the formation of near-surface local convergence zones over the sea (Telišman Prtenjak et al. 2010; Ivančan-Picek et al. 2016). The highly-indented eastern coast of the Adriatic, with numerous various-sized islands and canals, as well as SST variations along the basin, certainly contribute to the frequency of waterspouts (Sioutas and Keul 2007). The northern Adriatic part is also a very active convective region, not only in the Mediterranean but in Europe as well (Meteorological Office 1962; van Delden 2001; Mikuš et al. 2012; Feudale and Manzato 2014).

Table 1 presents an overview of ten selected cases of waterspouts events, and Fig. 1 gives their approximate locations according to available records. Although the cases were chosen mostly arbitrarily, it is interesting to note that the appearance of waterspouts was consistent with the most typical observation times highlighted by Sioutas and Keul (2007) and Renko et al. (2016): from 06 to 10 UTC (8 out of 10). Most of the reports of cases used for this study of waterspout occurrence were collected from different media, articles about damage and the questionnaire “You saw a waterspout/tornado? Report to us!” on the official Web site of the National Meteorological and Hydrological Service of Croatia (meteo.hr) (Renko et al. 2016). All these reports were verified, mostly with photos. Only the report from 6 January 2016 was obtained from the official national SYNOP network; the waterspout formed near the town of Split and moved onshore at 14:50 UTC passing by the meteorological station Split–Marjan (14445). It is quite rare that a waterspout passes by a meteorological station at the time of observations.

Table 1 List of waterspout (WS) cases used in the study represented by date and time of WS occurrence, location and geographic coordinates of WS observations, presence of atmospheric electricity discharges and synoptic types for selected WS cases

Weather types were subjectively ascribed for all cases using the classification method already used in previous studies on waterspouts in this region (e.g., Sioutas and Flocas 2003; Sioutas and Keul 2007; Keul et al. 2009; Renko et al. 2016). Five synoptic types were considered: southwesterly flow (SW), long-wave trough (LW), short-wave trough (SWT), closed low (CLOSED) and non-gradient pressure field (NG). The circulation flow at the 500-hPa level and the position and orientation of pressure troughs and ridges in conjunction with surface features were used when identifying weather types (see Fig. 3 in Keul et al. 2009). Each weather type was represented by at least two cases, mostly with one from the warm part and one from the cold part of the year. In the analysis, near-surface meteorological measurements (SYNOP and METAR) from standard meteorological stations were used as well as radio sounding measurements (downloaded from http://weather.uwyo.edu/upperair/sounding.html). Additionally, waterspouts were divided into two categories: thunderstorm-related and “non-thunderstorm”. With this categorization we wanted to emphasize the difference between waterspouts. The first type is associated with deep moist convection and as such is often related to severe thunderstorms (National Oceanic and Atmospheric Administration 2016), whereas the second type can occur beneath cumulus congestus or even cumulus mediocris clouds, which are usually associated with convergence lines (Renko et al. 2016). The presence of atmospheric electrical discharge was verified by the Lightning Location System which is a part of the International Lightning Detection Network LINET (e.g., Betz et al. 2009). The mentioned system covers the entire Croatian territory and consists of six sensors (five along the eastern Adriatic coast and one in the hinterland). The detection efficiency of LINET sensors is very high, which enables the detection of strokes with peak currents below 5 kA (Betz et al. 2009); however, the sensitivity of the sensors decreases as the distances of the lightning strokes from the LINET sensors increase (Holler et al. 2009). Nonetheless, the statistical average location accuracy of LINET sensors is approximately 150 m (Betz et al. 2009).

Fig. 3
figure 3

Four horizontal model domains (A–D) with a horizontal resolution of 9 km a, 4.5 km b, 1.5 km c, and 0.5 km d are displayed on each subplot. The smallest domains include the wider area of a Dubrovnik (WS1, WS6), b Novigrad (WS3)), c Pula (WS2) and Ližnjan (WS4), d Hvar and Komiža (WS5, WS9), e Split (D1, WS7) and Murter (D2, WS8), and (f) Mali Lošinj (WS10)

2.2 Mesoscale Meteorological Model/SWI and SHARPpy

2.2.1 Mesoscale Meteorological Model

The non-hydrostatic WRF-ARW (WRF—Advanced Research version V3.7.1) model was used for the analysis of thermodynamic and kinematic parameters of instability and SWI. The WRF model (Skamarock et al. 2008) solves the fully compressible, non-hydrostatic equations of motion in an Arakawa-C grid with the terrain following vertical coordinates. The WRF model has been applied and tested for model setup in many studies for the wider Adriatic region; for the analysis of the convection and mesoscale flow (e.g., Ivančan-Picek et al. 2014; Poljak et al. 2014; Telišman Prtenjak et al. 2015; Kuzmić et al. 2015; Kehler-Poljak et al. 2017).

For this study, the model setup included four two-way nested domains with horizontal grid spacings of 9, 4.5, 1.5, and 0.5 km and a Lambert conformal projection. The largest domain (Fig. 3) covered the Apennine Peninsula and the western part of the Balkan Peninsula, whereas the positions of the inner ones were changed depending on the selected case, with the focus on the waterspout location. The finest resolution of the model (500 m) did not explicitly allow waterspout simulation; however, the aim was to better understand the meteorological conditions and the characteristics of the atmosphere in the area where the waterspouts developed. Although the finest grid spacing could be potential source of double-counting of turbulence in the lowermost atmosphere (Horvath et al. 2012) the model setup is acceptably chosen. The time of waterspouts occurrence is mostly in the early morning and/or during wintertime in the cloud environment when relatively shallow atmospheric boundary layer will not favour double-counting. A vertical resolution of 5 m was taken close to the surface and slowly stretched upwards to form 97 vertical levels (with 25 in the lowermost 1 km). The model parameterizations for several categories of physical processes used during the simulations are given in Table 2 and include parametrizations for long-wave and short-wave radiation, soil model, cumulus clouds, microphysics, surface layer and atmospheric boundary layer. Cumulus parameterization was switched on only in the largest domain.

Table 2 Settings of numerical simulations in the WRF model

Topographic data came from the 90-m resolution of the SRTM (Shuttle Radar Topographic Mission) digital topographic database, and land-use data came from the CORINE (Coordination of Information on the Environment Land Cover) database at 100-m resolution. Initial and lateral boundary conditions were taken from the ECMWF analyses at a resolution of 0.125°, refreshed every 6 h. The simulations lasted 36 h and started at noon on the previous day because of the spin-up time (here considered to be the first 12 h).

2.2.2 SWI and SHARPpy

To start an SWI calculation, it was necessary to check the 850 hPa wind speed, which should be less than 20 ms−1 (Fig. 4). If this condition was met, other parameters needed for SWI calculation were: sea surface temperature (SST) and temperature at 850 hPa (T850)—for ΔT calculation; equilibrium level (EL) and lifting condensation level (LCL)—for convective cloud depth (ΔZ) calculation. For pairs of EL and LCL values, SWI is determined directly from the SWN shown in Fig. 2a (Szilagyi 2009). The values of SWI range from − 10 to + 10 and a waterspout is likely to occur when the SWI ≥ 0. In addition to the above parameters, the following were also calculated: convective available potential energy (CAPE), convective available potential energy integrated up to 3 km height (CAPE03), convective inhibition (CIN), K-index (KI) and the Total Totals index (TT). Several wind shear parameters were also calculated, namely, bulk shear 0–1 km (BS01), bulk shear 0–3 km (BS03) and bulk shear 0–6 km (BS06), which are the magnitudes of the wind vector difference between the surface and 1 km, between the surface and 3 km and between the surface and 6 km, respectively.

Fig. 4
figure 4

Box-and-whiskers plots of the distribution of: a full CAPE values, CAPE integrated up to 3 km height and CIN (Jkg−1); b the K-index (KI) and Total Totals (TT); c the wind speed at 850 hPa (ms-1) and the 0–1 and 0–6 km bulk shear (ms−1), i.e., the magnitude of the vector difference between the surface wind and the wind at 1 and 6 km AGL, respectively; and d convective cloud depth ΔZ (km), ΔT (°C) and SWI. Upper and lower parts of the box show 75th and 25th percentiles, whereas the horizontal line inside the box shows the median value and circle depicts the mean value; the crosses represent outliers

Together with SWI, all thermodynamic and kinematic parameters were derived from pseudo-soundings (NWP soundings), including profiles of air temperature, dew point temperature, pressure, wind speed and direction at 97 vertical levels. Pseudo-soundings 1 h before the event were used as an input for the open-source sounding analysis programme SHARPpy (Marsh and Hart 2012). SHARPpy is a Python implementation of the Storm Prediction Center’s (SPC) Skew-T and Hodograph Analysis Research Program (SHARP). All the indices, which depend on the parcel choice, were calculated by lifting a surface-based parcel; which used only the surface layer to define an air parcel, and virtual temperature correction was used (Doswell and Rasmussen 1994). Because the programme operates on single-user-supplied profiles of temperature, dewpoint, pressure, height, wind speed and wind direction (Marsh and Hart 2012), analysis is not restricted to particular sources (e.g., only radiosondes); the programme can use, like in this work, data from a model which provides vertical thermodynamic and kinematic profiles. Only 20 pseudo-soundings from the fine grid (D) model domain in the vicinity of waterspout location were used in this work. Results for SWI and all thermodynamic and kinematic parameters for 20 grid points in the vicinity of waterspout location for each event (Table 1) are presented and were analyzed using “box-and-whisker” diagrams.

3 Results and Discussion

3.1 Thermodynamic Environment and SWI for Selected Cases

Figure 4 displays the thermodynamic indices of instability (CAPE, CAPE03, CIN, TT, KI, ΔZ and ΔT) and kinematic parameters (BS01, BS06). As was expected, waterspouts can form in an environment with very different values of CAPE and CAPE03 (Renko et al. 2016) and they are comparable with those obtained from observational studies (based on radiosonde data) carried out in other parts of the Mediterranean (Giaiotti et al. 2007; Rodriguez and Bech 2017). Nonetheless, values obtained from the WRF model were in general higher than the ones obtained for the Adriatic by Renko et al. (2016). Reasons for this could have been (i) the larger number of vertical levels, which increased the ability of the model to accurately simulate CAPE, and (ii) the frequent overestimation of dewpoint temperatures in the WRF model, which led to higher moisture values in the lower troposphere (Holley et al. 2014). The latter can enlarge the values of CAPE regardless of model setup (Holley et al. 2014; Khodayar et al. 2016). Nonetheless, it is important to stress that the convective activity was reproduced in all simulated cases at the time of the observed waterspouts, which created favourable conditions. However, proximity soundings were used in Renko et al. (2016), some of which probably did not ideally represent the local convective waterspout environment and thus lowered median CAPE values.

Thunderstorm-related cases and cases in the warm part of the year had greater CAPE and CAPE03 values (Fig. 4a); the most notable case was from September 2015, WS5 (CAPE = 2350 Jkg−1, CAPE03 = 340 Jkg−1). This case was also characterized by very strong midlevel shear (BS03 = 26 ms−1) because of the passage of a cold front. This case will be discussed in more detail later.

For most cases, convective inhibition (Fig. 4a) was quite small, and only a small amount of lift was thus needed for convection initiation. However, strong lift was necessary for two winter cases: 14.1.2016. (WS9) and 11.2.2016. (WS10). These two cases were both connected to shallow convection, significant lapse rates in the lower atmosphere and consequently, to the lowest values of SWI. The WS9 case was also special because it was the only case with negative values of KI (KI = − 14 °C). Large negative values could have had several causes. On the previous day (WS8), cold and drier air was advected in the middle troposphere and unstable air was located in a very shallow layer near the sea surface; all CAPE was thus concentrated in the lowest 3 km (Fig. 4a). Additionally, the simulated convection was observed to be more eastern than the reported waterspout location. KI for other cases (Fig. 4b) ranged between 20 and 35, which accords with the previous studies of Sioutas and Keul (2007), Keul et al. (2009), Sioutas et al. (2013) and Renko et al. (2013). The same applies for TT (Fig. 4b), whose values ranged from 40 to 60 except for the WS9 case (TT = 32).

Figure 4c demonstrates that the low-level shear values were similar to ones obtained by Groenemeijer and van Delden (2007), Pucik et al. (2015), Renko et al. (2016), Taszarek et al. (2017) and Rodriguez and Bech (2017), which suggests that most of the analyzed waterspouts were non-mesocyclonic; the intensity of a tornado strongly depends on the increase of low-level shear (Groenemeijer and van Delden 2007; Doswell and Evans 2003; Pucik et al. 2015; Taszarek et al. 2017). The WS7 case stands out, with BS01 values centered at approximately 10 ms−1. Deep layer shear (BS06) values for most cases were in the expected range (Groenemeijer and van Delden 2007; Pucik et al. 2015; Renko et al. 2016; Taszarek et al. 2017 and Rodriguez and Bech 2017) but two winter cases (WS9 and WS10) had very high values, with deep layer shear greater than 30 ms−1. As previously mentioned, waterspouts for these two cases were connected to shallow convection; the difference between EL and LCL, i.e., convective cloud depth, was less than 5 km. Environmental conditions for these cases could be classified as high-shear low-CAPE (HSLC environments), where large low-level lapse rates and strong synoptic forcing also play a significant role (Sherburn and Parker 2014; Sherburn et al. 2016). For other waterspout events the range was approximately from 6 to 11 km, which agrees with ΔZ values obtained by Keul et al. (2009) and Sioutas et al. (2013) for the Eastern Mediterranean (Ionian Sea, Aegean Sea, Adriatic Sea) and Renko et al. (2016) for the Adriatic Sea. Mean ΔZ value for waterspout outbreaks (more than two waterspouts occurring on a single calendar day and generally produced by the same weather system) over the Great Lakes of North America was 5751 m (Sioutas et al. 2013), which is very similar to the mean ΔZ values for non-thunderstorm waterspouts in Renko et al. (2016). As can be seen from Figs. 2 and 5 for waterspouts that developed under conditions supportive for deep moist convection, the values of ΔT could be rather small. The WS9 case that had the smallest ΔZ value had the biggest difference between SST and T850T = 15.8 °C). However, most cases had smaller ΔT values and, according to the Nomogram, for SWI > 0 and ΔT < 15 °C, ΔZ should have been greater than 4500 m. Since the cloud depth for events in the cold part of year can be rather shallow (WS9), we can conclude that a minor correction of the SWI to the slight negative values would be appropriate.

Fig. 5
figure 5

SWI values for 20 pseudo-soundings for each waterspout event (note: ΔZ values are shown in km, whereas in the original units of ft were used)

Results obtained for SWI in this work are in accordance with SWI values obtained by Keul et al. (2009), Sioutas et al. (2013) and Renko et al. (2013). Although the sample used in this paper was considerably smaller, the hit rate, 9 of 10 cases with SWI > 0, is quite satisfactory, which confirms that SWI is a tool that can improve and help waterspout forecasts when used together with all knowledge on atmospheric conditions favourable for waterspout formation.

3.2 Case studies

3.2.1 WS1: 28 August 2013, South Adriatic, Dubrovnik

The largest documented waterspout outbreak in the Adriatic Sea was observed on the 28th of August 2013, with 41 waterspouts and funnels spotted along the Eastern Adriatic, from which at least 30 were observed near Dubrovnik (thanks to the strong coordination with a storm chasing group “Storm Chasers Dubrovnik”).

The synoptic situation that day (Fig. 6) was characterized by a large cyclonic vortex over Central Europe that slowly moved eastwards. On the periphery of the vortex several short-wave troughs could be spotted. In the early morning a strong southwesterly flow affected the southern Adriatic and a short-wave trough moved from Italy to the southern Balkans during the day. The weather type for this case was SW turning into SWT during the day. All ingredients for convection were present: very moist air in the lower troposphere, moderately steep midlevel lapse rates due to advection of midlevel air from Africa, together with weak or moderate lift that could be expected because of orography and with the displacement of the short-wave trough. In an environment such as this, a significant amount of CAPE could be released (Fig. 6c). Still, this case was classified as a “non-thunderstorm”.

Fig. 6
figure 6

a Surface pressure chart (source: Deutscher Wetterdienst DWD), b ERA-Interim reanalysis for 500 hPa geopotential height (black lines) and 1000–500 hPa thickness field (m, colour shaded) for 28 August 2013 at 06 UTC, c Skew-T log-p diagram of the radiosounding at Brindisi, Italy on 28 August 2013 at 00 UTC (source: http://weather.uwyo.edu/upperair/sounding.html)

WRF-ARW CAPE values (Fig. 7) near Dubrovnik ranged from 1200 to 1900 Jkg−1. Wind shear at low levels is clearly visible in Fig. 7, and strong SE winds near the surface shifted to weak SW winds. Additionally, BS03 (not shown) and BS06 had values that supported more organized deep convection. Together with SWI values that ranged from 6 to even 9 near the coast and low LCL (not shown) at approximately 500 m, conditions for waterspout development were met. Because of the very slow displacement of the system, favourable conditions lasted most of the day, with the convergence zone placed very locally near the town of Dubrovnik. Clearly, topography also played an important part in the role of this major waterspout outbreak.

Fig. 7
figure 7

a CAPE (Jkg−1), b wind (ms−1) at 10 m, c 0–6 km wind shear and d SWI for the innermost domain D, forecasted by WRF for 28 August 2013 at 07 UTC

3.2.2 WS5: 24 September 2015, Middle Adriatic, Hvar

Information about four waterspouts near the town of Hvar on the island of Hvar was obtained from a media report (http://www.crometeo.hr/hvar-pijavice-jedrilici-raskidale-jedra-jahta-udarila-o-rivu-foto-video/) (Fig. 8a). Two waterspouts produced damage; one affected a sailboat in the port and the other directly moved onshore to the hotel ‘Bodul’ and produced damage estimated as F0 or F1 (Fig. 8b).

Fig. 8
figure 8

a Estimated waterspout trajectories of four waterspouts (source: Crometeo) and b waterspout and damage at hotel ‘Bodul’ (photo: Jakša Kuzmičić) on 24 September 2015 near the town Hvar on Hvar island

A quite intensive cut-off low in upper tropospheric levels travelled that morning from central Italy to the southern Adriatic and produced a cyclone with frontal systems that moved along the Adriatic Sea (Fig. 9). A strong jet stream was present at mid-levels. A strong convergence zone was also present because of the passage of a cold front. Warm and moist air was pulled ahead of the cold front, while advection of cold air began at mid-levels. Warm seas and rich maritime moisture that overlapped with steep lapse rates, together with strong forcing because of the passage of the front, resulted in a potential for deep moist convection.

Fig. 9
figure 9

a Surface pressure chart (source: Deutscher Wetterdienst DWD) and b ERA-Interim reanalysis for 500 hPa geopotential height (black lines) and 1000–500 hPa thickness field (m, coloured shaded) for 24 September 2015 at 06 UTC, c Skew-T log-p diagram of the radiosounding at Brindisi, Italy on 24 September 2015 at 00 UTC (source: http://weather.uwyo.edu/upperair/sounding.html), d lightning strikes as detected by LINET network sensors between 06 and 10 UTC on 24 September 2015

As previously noted, this case was characterized by the highest amount of model CAPE and the highest values of SWI (Fig. 10). Overlap of significant CAPE values with strong wind shear, together with low LCL, resulted in high EL (10 < ΔZ < 11 km). Very strong convection was confirmed with lightning activity detected by the LINET sensors (Fig. 9). Since the convective cloud depth was also the largest for this event, even quite small ΔT values would be sufficient to satisfy the Nomogram; however, with a ΔT of approximately 10 °C, the WS5 case had SWI values in a range from 8.5 to 9.5 for all 20 pseudo-soundings from the fine grid (D) model domain near the waterspout. As the cold front passed, conditions for waterspout development were noticeably reduced (SWI < 0 in the lower left corner in Fig. 10).

Fig. 10
figure 10

a CAPE (Jkg−1), b 0–3 km wind shear (ms−1) and c SWI for the innermost domain D, forecasted by WRF for 24 September 2015 at 07 UTC

3.2.3 WS7: 6 January 2016, Middle Adriatic, Split

This winter waterspout was quite unique because it passed by the main meteorological station Split–Marjan at the time of SYNOP observations, exactly at 14:58 UTC on the 6th of January 2016. The observer of the Meteorological and Hydrological Service of Croatia, Mate Pavić, followed its development and successfully photographed it (Fig. 11). The waterspout approached the meteorological station from the south–southwest and the anemometer measured a strongest wind gust of 24 ms−1 (Fig. 12). Tangential wind speeds in waterspouts usually exceed 20 ms−1 (Golden 1974). From Fig. 12, the pressure drop was nicely visible at the time of waterspout passage and showed one of the main waterspout features, a low pressure core. The observer estimated the cloud base height at 300 m, which could indicate that low LCL also played an important role in the waterspout development. Wind measurements at meteorological stations (Fig. 13) in the surroundings revealed the presence of a pronounced convergence zone, namely, southwesterly winds turns to northwesterly north of the city of Split. This type of wind setup is in accordance with the conceptual model of non-mesocyclonic tornadoes of Caruso and Davies (2005). Environments supportive of non-mesocyclonic tornadoes are composed of a weak, slow-moving or stationary surface front with little temperature contrast but a sharp wind shift from the south or southwest to the northwest. Additionally, the frontal wind shift boundary was oriented northeast to southeast, which is nicely visible in the satellite image and SYNOP data (Fig. 13) which also show developing thunderstorms along this convergence line.

Fig. 11
figure 11

Waterspout on 6 January 2016 near the main weather station Split–Marjan; photo by DHMZ observer Mate Pavić

Fig. 12
figure 12

Wind, atmospheric pressure and temperature measurements at the main weather station Split–Marjan on 6 January 2016 between 14:00 and 16.00 UTC (the black line depicts the time the waterspout was observed near the station)

Fig. 13
figure 13

a Meteosat SEVIRI Sandwich product (High-resolution visible satellite image overlaid with semi transparent colour enhanced infrared IR 10.8 μm) on 6 January 2016 at 14:50 UTC (the white dot represents the Split location), b SYNOP data at the main weather station Split–Marjan (the black circle) and surrounding stations on 6 January 2016 at 15:00 UTC

The synoptic setting (Fig. 14) involved an upper long-wave trough over the Balearic Sea progressing eastwards. The Middle and South Adriatic were under the influence of a strong southwesterly flow but the cold air within the trough was approaching the Adriatic at mid-levels, which produced steep lapse rates.

Fig. 14
figure 14

a Surface pressure chart (source: Deutscher Wetterdienst DWD), b ERA-Interim reanalysis for 500 hPa geopotential height (black lines) and 1000–500 hPa thickness field (m, coloured shaded) for 6 January 2016 at 12 UTC

Instability was also seen in the model CAPE values (Fig. 15); although not as high as in the previous two cases, it was still quite high for winter and more than sufficient for the development of convection. The highest values of CAPE (934 Jkg−1) were located over the open sea where they overlapped with strong low-level shear (Fig. 4) and significant deep layer shear (Fig. 15) that could produce tornadoes of F1 intensity (Groenemeijer and van Delden 2007; Doswell and Evans 2003). According to eyewitnesses, the waterspout weakened entering the mainland and the damage was light to moderate according to the Fujita Tornado Damage Scale (http://www.spc.noaa.gov/faq/tornado/f-scale.html); this event could thus be assigned an F0 or even F1 category.

Fig. 15
figure 15

a CAPE (Jkg−1) and b 0–6 km wind shear (ms−1) for the innermost domain D, forecasted by WRF for 6 January 2016 at 13:45 UTC

The rather small ΔT, approximately 10.5 °C, and ΔZ between 7 and 8 km had low SWI values according to the Nomogram (3 < SWI < 4), similar to other winter cases (Fig. 16). The crucial part in waterspout formation was played by a strong convergence zone and a strong low-level shear together with low LCL. The median value for BS01 was 10 ms−1, which was the highest median value of BS01 for all 10 cases.

Fig. 16
figure 16

a ΔZ (km) and b ΔT (°C) for the innermost domain D, c and d SWI for two inner domains C and D, forecasted by WRF for 6 January 2016 at 13:45 UTC

4 Summary

Waterspouts are highly common in the Adriatic Sea and pose a threat mostly to smaller crafts but also endanger larger boats, people and property (Figs. 8 and 11). Since it is necessary to continuously improve existing warning procedures, methods and tools, as well as develop new ones; in this study, we tested a waterspout forecasting technique originally developed for the Great Lakes of North America. However, although this method was tested in Keul et al. (2009), Renko et al. (2013), and Sioutas et al. (2013), the approach in this study was improved because convection processes were treated in a more sophisticated way, with a numerical weather model of much finer resolution. All thermodynamic and kinematic data were computed from pseudo-soundings from a mesoscale numerical weather model with 500-m horizontal grid spacing (the WRF model) using the Open-Source Sounding Analysis Toolkit SHARPpy (Marsh and Hart 2012). Additionally, all indices, which depended on the parcel choice, were calculated by lifting a surface-based parcel, and a virtual temperature correction was used. A step forward was made in index presentation; SWI was in the form of spatio-temporal fields and is currently more operational.

Ten waterspout cases from 2013 to 2016 were analyzed and three were discussed in detail. This study affirmed that waterspouts can form in environments with very different values of CAPE and CAPE03, which usually suggests environments that can sustain convection and non-severe weather (e.g., Groenemeijer and van Delden 2007; Pucik et al. 2015; Renko et al. 2016; Taszarek et al. 2017). Only for two cases, and especially for the WS9 case, stronger lift was needed because of stronger convective inhibition. This case also stood out because of negative values of KI, whereas others were in the expected range (20 < KI < 35); the same applied to TT. Comparing BS01 (median value of BS01 for most cases < 7 ms−1) with that of previous studies (e.g., Groenemeijer and van Delden 2007; Doswell and Evans 2003; Pucik et al. 2015; Taszarek et al. 2017), we can conclude that most of the analyzed waterspouts were non-mesocyclonic. Only four cases (WS1, WS4, WS9 and WS10) had BS06 values greater than 20 ms−1, which usually indicates that mesocyclonic tornadoes can occur (Rasmussen and Blanchard 1998; Thompson et al. 2003; Groenemeijer and van Delden 2007; Pucik et al. 2015; Taszarek et al. 2017). Nonetheless, our two winter cases (WS9 and WS10) were connected to shallow convection (ΔZ < 5 km) and were most likely non-mesocyclonic. Complete verification about the type of waterspout, either mesocyclonic or non-mesocyclonic, (Davies-Jones et al. 2001) would require radar measurements that are still not available along the whole Eastern Adriatic Coast.

Pairs of ΔZ and ΔT values, and consequently SWI, generally accorded with values obtained by Keul et al. (2009), Sioutas et al. (2013) and Renko et al. (2013), which again confirms that SWI can be useful in waterspout forecasting. However, this study showed that SWI use was better for the summer cases; some deviations were present for the winter cases. SWI is a forecasting technique developed for the Great Lakes of Canada. Therefore, some adjustments for the Adriatic could be made. For example, the typical SST winter temperatures are greater than on the Great Lakes, and cold outbreaks over the Mediterranean do not create as large vertical temperature differences (ΔT). Therefore, one proposal that arises from this study is to allow slightly negative SWI values as relevant predictors, especially for winter waterspouts with shallow convection and smaller low-level lapse rates. Additionally, one can expect a high rate of false alarms, which can be reduced in the future by adding additional parameters to SWI fields, such as wind convergence and moisture convergence. Non-mesocyclonic tornadoes are typically more difficult to anticipate and forecast than supercell tornadoes (Caruso and Davies 2005), which makes waterspout forecasting quite challenging. Of course, SWI must be used together with the knowledge of all ingredients and environmental setup supportive for waterspout development. However, further analysis of the cases and the verification of the method through its operational application (Horvat et al. 2017) are necessary in the future. Because verification of rare events is very challenging (Ferro and Stephenson 2011), one should have in mind the difficulties that will arise: (i) the small number of events that may be observed and could lead to large uncertainty about forecast quality, (ii) most verification measures necessarily degenerate to trivial values as event rarity increases and (iii) events may be observed inaccurately because of short space–time scales, and non-events can even pass unrecorded (Ferro 2007). Recognizing all of these difficulties, non-events should also be considered in further method development.