The instrumentation deployed over the Adriatic TA during SOP1 belonged mainly to the DHMZ observational network. DHMZ deployed a ground observation operational network that included automatic, climatological and rain gauge stations; two radio-soundings, Zagreb-Maksimir (station ID = 14240, H = 123 m a.s.l., φ= 45∘49′ N, λ= 16∘02′ E) and Zadar–Zemunik (station ID = 14430, H = 88 m a.s.l., φ= 44∘5′ N, λ= 15∘21′ E), and two radars, Bilogora and Osijek. The locations mentioned in the text are indicated in Fig. 1b.

The meteorological measurements and observations from 58 SYNOP stations (31 of which were automatic stations) were made every hour and reported in real time during SOP1. All the automatic stations measured data at 10 min intervals and reported the measured data in real time. However, not all 63 automatic stations measured all the meteorological parameters. Twenty-one automatic stations only reported wind parameters (average 10 min speed and direction and wind gust speed measured in the previous 10 min). Five additional stations measured wind parameters, temperature and relative humidity. All real-time surface measurements (SYNOP and automatic station data) and available radar figures were stored at the HyMeX data centre.

The dense network of climatological stations (120 stations with an average distance of 20 km) was the source of temperature, humidity and wind speed, cloudiness and visibility were estimated from observations only three times per day at 06:00, 13:00 and 20:00 UTC and accumulated rainfall and snow height were measured at 06:00 UTC (more than 500 stations reported accumulated 24 h rainfall).

In addition to operational radiosoundings in Zadar–Zemunik at 00:00 and 12:00 UTC, several extra radiosoundings were deployed through the Data Targeting System (DTS) upon request of the HOC. These targeted radiosoundings, among others in the western Mediterranean, were activated during IOP16, which caused heavy precipitation, strong winds and snow in the eastern Adriatic. Requests for additional radiosoundings at 06:00 and 18:00 UTC were carried out under the EUMETNET Observation Programme. Sounding data measured at Zadar–Zemunik, located on the eastern coast of the Adriatic Sea at the southern end of Velebit mountains, provided information on the vertical structure of the troposphere to monitor the upstream flow of the precipitation events in the Adriatic region. The selection of sensitive area predictions (SAPs), that is, predictions for regions where observations are expected to have the largest impact on the forecasts for the verification, used methods developed by ECMWF and Meteo-France (Prates et al., 2009). The verification area selected for SAP calculations was centred over the northern and/or central Adriatic.

To complement the ground-based observations, the data from two radars in Croatia (Bilogora (H = 270 m a.s.l., φ= 44∘53′ N, λ= 17∘12′ E) and Osijek (H = 89 m a.s.l., φ= 45∘30′ N, λ= 18∘34′ E)) and one in Slovenia (Lisca; H = 944 m a.s.l., φ= 46∘04′ N, λ= 15∘17′ E) were made available operationally in graphical form. Estimates of the instantaneous surface rain rates from the Lisca and Bilogora radars were provided to the HyMeX web server in real time. North-western Croatia, particularly Rijeka and Istria, is covered by operational radars over Croatia, Slovenia and Italy, but the area is on the edge of the ranges and behind a mountain.

Standard Meteosat Second Generation (MSG) Spinning Enhanced Visible and Infrared Imager (SEVIRI) data are available in intervals of 15 min, and Rapid Scan Service (RSS) data are available in 5 min intervals. The abundance of remote sensing data on the HyMeX server encourages detailed analyses of all the cases that produced HPEs over Croatia during SOP1.

Satellite-derived precipitation data from the Tropical Rainfall Measuring Mission were used (TRMM, Huffman et al., 2007). In particular, we used the 3 h accumulated precipitation data from the 3B42RT product to compute the 24 h accumulated rainfalls for the period from 06:00 to 06:00 UTC the next day, and 1 h precipitation data from the 3B41RT product were compared with the precipitation forecasts developed using operational NWP models.

A short description of the model characteristics and the operational set-up during SOP1 is given here.

During SOP1, DHMZ provided the products from the operational forecast (Tudor et al., 2013). At the time, the numerical weather prediction system (NWP) was based on the hydrostatic and non-hydrostatic ALADIN models.

The ALADIN hydrostatic model (Aladin International Team, 1997; Tudor et al., 2013) was run twice per day on a domain with 8 km resolution (Fig. 1a), starting from 00:00 and 12:00 UTC up to a 72 h lead time. The operational suite used lateral boundary conditions from the global model ARPEGE run by Meteo-France. The initial fields were obtained using a data assimilation procedure (Stanešić, 2011). The operational ALADIN model is a limited-area model that applies Fourier spectral representation of the model variables using fast Fourier transforms (FFTs) in both directions with a quadratic elliptic truncation (Machenhauer and Haugen, 1987), which ensures an isotropic horizontal resolution and that the non-linear terms of the model equations are computed without aliasing. The forecast at an 8 km resolution was run on a domain with 240 × 216 grid points, which included a band of 11 points along the northern and eastern boundaries, with unphysical terrain created for the biperiodization (Fig. 1a). The dynamical computations were performed using semi-implicit semi-Lagrangian discretisation (Robert, 1982) to solve the hydrostatic dynamics and finite difference method on 37 levels of hybrid pressure type eta coordinates (Simmons and Burridge, 1981) in the vertical. The operational physics package at the time used prognostic TKE, cloud water and an ice, rain and snow and diagnostic scheme for deep convection. The prognostic equations for condensates were solved using the barycentric approach (Catry et al., 2007).

Upon numerous case studies of severe weather events (e.g. Tudor and Ivatek-Šahdan, 2010), an additional operational forecast run was established in July 2011, which used ALADIN with non-hydrostatic dynamics and a complete set of physics parameterisations, including the convection scheme. The high 2 km-resolution forecast using ALADIN model with non-hydrostatic dynamics (Bénard et al., 2010) with the physics package that included the convection scheme was running operationally during the HyMeX SOP1 campaign (Fig. 1b). The convection scheme used in the high-resolution model is modular multiscale microphysics and a transport (3MT) scheme for precipitation and clouds (Gerard and Geleyn, 2005; Gerard, 2007; Gerard et al., 2009).

Both runs used SSTs from the initial file of the global model ARPEGE forecast. Additional details of the model characteristics can be found in Table 1.

The influence of different meteorological characteristics and physical processes that produced HPEs over the Adriatic TA and Dinaric Alps are briefly analysed and summarized. Previous research on the occurrence of HPEs in the wider Adriatic region (e.g. Doswell et al., 1998; Romero et al., 1998; Vrhovec et al., 2001; Kozarić and Ivančan-Picek, 2006; Horvath et al., 2006; Mastrangelo et al., 2011; Mikuš et al., 2012) highlighted cyclonic activity in the western Mediterranean and Adriatic as a triggering mechanism for a range of extreme weather phenomena, including HPEs. The positions of cyclones that appear in the Adriatic Sea basin strongly influence the climate and weather conditions in the area (Horvath et al., 2008).

During SOP1, several upper-level troughs entered the western Mediterranean and induced cyclogenesis over the Gulf of Genoa, Tyrrhenian Sea and Adriatic Sea. Figure 3 shows the mean sea level pressures and low-level horizontal winds for IOP4, IOP9, IOP13, IOP16, IOP18 and IOP19. Although most of the events were related to cyclone activity in the region, some events were not characterized by a cyclone moving over the area. In the following text, we summarize the analyses of selected characteristic IOPs that affected the Adriatic area. Large-scale conditions similar to those found in the IOPs helped generate mesoscale and local processes, leading to quite different precipitation patterns.

The performances of the operational precipitation forecasts with the ALADIN model at 8 km and ALADIN model at 2 km grid spacing during SOP1 were assessed by comparing the forecasts with the measurements from the Croatian surface observation network. The model results were compared with 24 h accumulated precipitation measured by the rain gauges. Before the calculation of the verification score results for ALADIN 2 km, the model was upscaled to an ALADIN 8 km grid to avoid double penalty errors and make a more direct comparison. The precipitation amount from the ALADIN (8 km and upscaled 2 km) model was obtained from the nearest model point with respect to the observation location. Contingency tables (Tables 3 and 4) were evaluated with three categories defined according to the amount of 24 h accumulated precipitation and classified as dry, medium and strong. An event was defined as dry if the 24 h accumulated precipitation on the rain gauge station was less than or equal to 0.2 mm/24 h. The border between the medium and strong categories was defined as the 95th percentile of the measured 24 h accumulated precipitation (50.42 mm/24 h) during the SOP1 period, but with the dry events excluded.

Figure 10 presents the 24 h accumulated precipitation histograms from both the models and rain gauges during the entire SOP1 period and during the specific days corresponding to the 8 IOPs indicated in Table 1. The measurements show that a large percentage of the events were dry (64.7 %) during the entire SOP1 period. The value corresponding to the 95th percentile (50.4 mm) is indicated on the graph, and it appears to be a reasonable threshold for the heavy precipitation events that we want to verify. As expected, the histogram for only the IOP days (8 IOP cases) shows that the number of dry events was reduced (18.1 %), and the relative frequency of events shifted towards events with higher amounts of precipitation.

Although the ALADIN 8 km model distribution was in rather good agreement with the rain gauge measurements during the entire SOP1 period, with the exception of the most intensive rain, the model distribution for the IOP days only shows that the model tended to underestimate the frequencies of the weak and strong precipitation events, whereas it overestimated the frequency of moderate precipitation events. For ALADIN 2 km SOP1 and IOP days only, the histograms show similar results; the model tended to underestimate moderate precipitation, whereas at the same time it tended to overestimate strong precipitation. A comparison of the two models shows that the ALADIN 2 km model better agreed with the measurements, especially for very weak and strong precipitation.

The verification measures (Wilks, 2006) calculated from the comparison of the 24 h accumulated precipitation from the rain gauges and model for the three categories and different periods are summarized in Tables 3 and 4. The indices used here are defined in the Appendix. Because most of the measures are base rate (BR) sensitive and can be safely used only to compare two models for the same event, the polychoric correlation coefficient (PCC; Juras and Pasarić, 2006) was calculated as an additional measure because PCC does not depend on BR or frequency bias (FBIAS). For both ALADIN models, PCC showed rather high levels of association between the observations and forecast for the entire SOP1, whereas it had a smaller value for only the IOP days.

For both models, the smallest value of PCC was for IOP 9, because they overestimated the number of strong precipitation events, especially ALADIN 2 km, which can be seen from the much higher FBIAS than that from the ALADIN 8 km model. Comparing the performances of the two ALADIN models, it can be observed that ALADIN 2 km had higher levels of association between the observations and forecasts for IOP13 and IOP19 compared to ALADIN 8 km. For IOP13, ALADIN 2 km was relatively more accurate in all three categories, which can be seen from the higher values of the critical success index (CSI). For IOP19, the FBIAS values show that ALADIN 2 km overestimated the frequency of strong precipitation, but at the same time it was relatively more accurate for the other two categories (higher CSI). For the dry category, ALADIN 2 km had better scores for almost all the selected cases (higher CSI; FBIAS closer to 1). For medium precipitation, ALADIN 8 km had better scores, except for IOP13 and IOP19. For the strong category, the scores show that ALADIN 2 km tended to overestimate the frequency of strong events, whereas ALADIN 8 km tended to underestimate the frequency of strong events, with the sole exception of IOP19, for which both models overestimated the number of strong precipitation events (especially ALADIN 2 km).

The spatial distribution of the daily rainfall amounts for the IOP2 rain episode indicates that the largest amounts fell over the city of Rijeka (220 mm at the Rijeka meteorological station, which is located 120 m above sea level) and the surrounding mainland hilly slopes and mountainous hinterland. According to the rainfall data recorded by ombrograph at the Rijeka meteorological station, a more detailed insight into the temporal rainfall distribution during the short-term interval of this heavy rainfall event is possible (Fig. 12). The rainfall episode that occurred during the 6 h period between 6 pm and midnight was most intense between 9 and 11 pm. The maximum 20, 30, 40, 50, 60 and 120 min rainfall totals, which would have been the most intense part of the rainfall episode, have not been recorded at the Rijeka station since the beginning of measurements in 1958 (Table 5). The rainfall intervals of 20, 30 and 40 min were especially intense and could be expected once in a more than a thousand, a few hundred and a hundred years, respectively, and correspond to an extraordinarily rare event as computed over the period 1958–2011 (Patarčić et al., 2014). The maximum amounts that fell in the 2 and 4 h intervals could be expected every forty and fifty years, respectively.

On 12 September 2012, a sequence of convective events hit the north-eastern part of Italy and, in particular, the eastern part of the Veneto region and the plain of the Friuli Venezia Giulia region. During that day, at least two of the events could be classified as supercells, and the first one was also associated with heavy hail (Manzato et al., 2015; Miglietta et al., 2016). After a few hours, a third storm system that resembled a squall line, although of limited dimensions, swept over the area.

EUMETSAT was conducting its first experimental 2.5 min rapid scan with the MSG-3 satellite, and data are available from early morning until 09:00 UTC of the IOP2 day. Unfortunately, the MSG-3 satellite (renamed Meteosat-10) experimental rapid scan data, which have intervals of 2.5 min, are available only until 09:00 UTC on 12 September 2012.

The nearby areas of Istria and Rijeka first received rain in the early afternoon, which soon stopped before the torrential rain in the evening between 21:00 and 23:00 UTC. This rain was connected to a third storm over Italy (as discussed in Manzato et al., 2015), which was an elongated storm moving along the coast of the northern Adriatic. Convection developed over the northern Adriatic, and warm and moist advection produced intensive precipitation triggered by the orography inland.

Satellite data show that cumulonimbus clouds formed (Fig. 13). This intensive rainfall band reached Trieste and Slovenia according to the radar data (not presented) and merged with the rainfall band that formed above Trieste at 18:00 UTC. Another rainfall band formed above the Istria peninsula at 19:30 UTC. Intensive rainfall spread to Rijeka and persisted there for several hours. During that time, other rainfall bands formed and moved over Rijeka, intensifying the precipitation and prolonging the period of high precipitation intensity.

According to the hourly amounts, the largest precipitation intensity occurred from 21:00 to 22:00 UTC (85.3 mm/h), with 20.6 and 51.7 mm h-1 in the previous and following hour (Fig. 12).

Sounding data measured at Zadar–Zemunik, which is located approximately 150 km south-south-east of the area where the largest rainfall was recorded, can provide information on the vertical structure of the troposphere. Although the thermodynamic profile characteristics are not completely representative of the preconvective environment over the study area, this is the only available sounding data for the eastern Adriatic. The soundings featured a low-level moist atmospheric layer from the surface to approximately 850 hPa that was connected with SE jugo wind, confirming that there was a suitable environment for strong convective activity (not presented). The winds strengthened throughout the troposphere, and the highest intensity was observed at 400 hPa.

During SOP1, DHMZ made the operational forecast available from the ALADIN operational forecast model at 8 km and non-hydrostatic 2 km horizontal resolutions (Sect. 2.2). The two versions of the ALADIN model are compared here, and the comparison shows the capability for forecasting intense convective activity in the area.

The short-range forecasts well reproduced the large-scale and mesoscale features responsible for the event (Fig. 11). The low-level wind field was dominated by two low-level jet streams (LLJs) and caused the appearance of the low-level wind convergence over the northern Adriatic that was associated with the main Genoa cyclone (Fig. 11b). In this case, the performance of the model was rather successful in comparison with the ECMWF reanalysis (not presented). One south-westerly LLJ was elongated from Italy towards the mid-Adriatic and carried warm and humid Mediterranean air to the Adriatic Sea. Another north-easterly LLJ (bora wind) was modified and intensified by the pressure gradient across the southern flank of the Alps (Fig. 11a). This convergence was responsible for the convective triggering in the late afternoon. Although the mesoscale characteristics were correctly reproduced, the location and timing of the precipitation were not well predicted. The intensive precipitation event was predicted by both models, with precipitation close to or exceeding 100 mm/24 h inland of Rijeka (Fig. 4), but the amount of precipitation was underestimated for the city of Rijeka, which lies on the coastline in all operational models, possibly due to an absence of the cold pool that formed after the showers in the early afternoon or the low-level wind from north-east that started earlier than in the model forecast.

The operational forecast set-up of the ALADIN 2 km-resolution run overestimated the rainfall above the mountains (at least when compared to the 3B41 products from the TRMM data server), but it was consequently closer to the extreme amounts measured in the Rijeka area (Fig. 14). Although the 3B41 product is an estimate of precipitation intensity, which also suffers from errors, the rain over the southern Velebit mountains was overestimated, but was correct for the mountains inland of Rijeka. In the hours of peak precipitation intensity in Rijeka, the satellite measurement data-derived precipitation (TRMM 3B41RT product available from NASA's Giovanni web service) was also considerably lower than that measured in situ.

The high-resolution, non-hydrostatic operational forecast showed upward motions along the coastal mountains of Croatia that were associated with the convergence line and the rain band over the sea (Fig. 15). The wave of upward motion moved from the Po valley eastward and reached Rijeka area 1 h after the recorded maximum intensity in precipitation, and the model might, therefore, have been slightly later than the real weather events. A permanent wave formed over the southern Velebit (and several other mountains) and persisted throughout the night. The wave was responsible for triggering the precipitation, and its intensity was probably overestimated. Apparently, small but tall topographic obstacles can trigger too much precipitation; this issue must still be solved.

Figure 16 presents a scatter plot of the 24 h accumulated precipitation from rain gauges over Croatia and the forecast values from the ALADIN model taken from the nearest grid points for IOP2. The ALADIN 8 km model underestimated precipitation and forecasted up to 92 mm/24 h of rainfall, whereas the measurements reached 220 mm/24 h. Much better results were obtained from the ALADIN 2 km model; the values predicted by the model reached 200 mm/24 h. A location error is also evident for both models, especially for the area where the most intense precipitation occurred (Istria peninsula; red dots), but it was smaller for the ALADIN 2 km model. The medium precipitation amounts were better forecast than the strong precipitation amounts but were still slightly overestimated for the ALADIN 8 km model, and much more spread can be seen for the ALADIN 2 km model, with both overestimation and underestimation, but with better results for the Istria peninsula. From Tables 3 and 4, it can be observed that ALADIN 2 km was relatively more accurate (higher CSI) for the dry and strong categories, but not for the medium category, than ALADIN 8 km. FBIAS was better for ALADIN 2 km for the medium category in addition to the dry and strong categories compared to the ALADIN 8 km results.

Because the lack of model skill when simulating HPEs could be partially attributed to imperfect initial conditions, we performed several NWP experiments to assess the impact of data assimilation on the IOP2 forecast accuracy. Observations used in the operational data assimilation system include ground station observations (2 m temperature, 2 m relative humidity, pressure), radio soundings (temperature, humidity, wind components), aircraft-based observations (temperature, wind components), wind components derived from a cloud motion detection process based on the measurements of geostationary satellites and brightness temperature from geostationary and polar satellites.

A comparison of the measurements with an operational forecast and simulations without data assimilation is shown in Fig. 17. The rain gauges showed that an elongated area of stronger precipitation along the Croatia–Slovenia border was present, and that pattern was better forecasted by the operational run that incorporated data assimilation. In addition, higher amounts of the medium rain category over the Istria peninsula were found in the operational run, which better accorded with measurements. This can also be seen in Fig. 16, where for the run with data assimilation the points are less scattered, and more points with higher values of precipitation over Istria are present. The maximum recorded around the town of Rijeka was not adequately represented by either model.

The verification measures (Table 3) show that the simulation with data assimilation produced slightly better results. The scores for the entirety of Croatia show that the strong precipitation category results were improved for the operational run (CSI = 0.28) compared to the run without data assimilation (CSI = 0.23). In addition, PCC showed that the model and observations for the run with data assimilation were better associated. The impact of data assimilation was rather small, but it yielded an improvement in the 24 h precipitation forecast. It should be considered that for the selected case, better results were obtained with the higher-resolution model and that the data assimilated in the operational ALADIN 8 km model was mainly synoptic. Thus, implementing data assimilation in the higher-resolution model and adding additional high-resolution temporal and/or spatial data to the data assimilation system are apparently good ways of further enhancing operational forecasts.