Evaluation of INCA precipitation analysis using a very dense rain gauge network in southeast Austria

An accurate estimate of precipitation is essential to improve the reliability of hydrological models and helps for decision-making in agriculture and economy. Merged radar–rain-gauge products provide precipitation estimates at high spatial and temporal resolution. In this study, we assess the ability of the INCA (Integrated Nowcasting through Comprehensive Analysis) precipitation analysis product provided by ZAMG (the Austrian Central Institute for Meteorology and Geodynamics) in detecting and estimating precipitation for 12 years in southeast Austria. The blended radar–rain-gauge 15 INCA precipitation analyses are evaluated using WegenerNet – a very dense rain gauge network with about 1 station per 2 km 2 – as „true precipitation‟. We analyze annual, seasonal, and extreme precipitation of the 1 km × 1 km INCA product and its development from 2007 to 2018. Based on the results, the performance of INCA can be divided into three different periods. From 2007 to 2011, the annual area-mean precipitation in INCA was slightly higher than WegenerNet, except in 2009. However, INCA underestimates precipitation in grid cells farther away from the two ZAMG meteorological stations in 20 the study area (which are used as input for INCA), especially from May to September (“wet season”). From 2012 to 2014, INCA's overestimation of the annual-mean precipitation amount is even higher, with an average of 25 %, but INCA performs better close to the two ZAMG stations. From 2015 onwards, the overestimation is still dominant in most cells but less pronounced than during the second period, with an average of 12.5 %. Regarding precipitation detection, INCA performs better during the wet seasons. Generally, false events in INCA happen less frequently in the cells closer to the ZAMG 25 stations than in other cells. The number of true events, however, is comparably low closer to the ZAMG stations. The difference between INCA and WegenerNet estimates is more noticeable for extremes. We separate individual events using a 1-hour minimum inter-event time (MIT) and demonstrate that INCA underestimates the events‟ peak intensity until 2012 and overestimates this value after mid-2012 in most cases. The overestimation of the peak-intensity is more pronounced during July. In general, the precipitation rate and the number of grid cells with precipitation are higher in INCA. Furthermore, 40 % 30 of the individual events start earlier, and 50 % end later in INCA. Considering four extreme convective short-duration events, there is a time shift in peak intensity detection. The relative differences in the peak intensity in these events can https://doi.org/10.5194/hess-2021-34 Preprint. Discussion started: 15 February 2021 c © Author(s) 2021. CC BY 4.0 License.

The aim of this study is to evaluate INCA precipitation analyses over a period of 12 years, using gridded precipitation fields 65 from the dense WegenerNet weather and climate station network in southeast Austria. The main focus lies on analyzing the ability of INCA to detect and estimate precipitation, and on studying the impact of modifications of INCA algorithms and input data during these 12 years. We analyze annual data, seasonal data, and extremes, using different metrics. Moreover, INCA"s detection skill is studied using categorical metrics. Furthermore, we identify individual events using a simple threshold based on the interval between two consecutive events and compare the events" characteristics in both datasets. 70 Finally, we separately study extreme convective short-duration events and demonstrate four representative examples. The following research questions are addressed and discussed in this study:

1.
How well can INCA detect and estimate precipitation in an area with a moderate topography?

2.
How did the developments in the Austrian radar network affect INCA"s performance?

How reliable are INCA estimates of extremes? 75
This paper is structured as follows. In Sect. 2, we introduce the study area and each dataset's main features; in Sect. 3, the methodology is described. The results based on different time scales and individual events are discussed in Sect. 4, and we conclude in Sect. 5.

WegenerNet 80
The WegenerNet network is a dense climate station network located in the Feldbach region in southeast Austria (see Fig. 1).
The network includes 155 ground stations, almost uniformly spread over an area of about 22 km × 16 km (i.e., about one station per 2 km 2 ) provided by the Wegener Centre for Climate and Global Change, University of Graz, Austria (Kirchengast et al., 2014;Fuchsberger et al., 2020b). The highest altitude in this region is 609 m above Mean Sea Level (MSL), located in the Southern part. The altitude decreases northward to the valley of the river Raab (see Fig. 1). The Feldbach region is 85 affected by both Mediterranean and continental climates. Most of the precipitation occurs from May until September (here defined as the "wet season"), when monthly average precipitation is approximately twice as high as during the "dry season" from October to April (O and Foelsche, 2019). Considering that the average number of days with fresh snow in this area is less than 15 days during 1971 to 2000 (Prettenthaler et al., 2010), and has been decreasing over time, snowfall is relatively unimportant in this area.  descriptions of the quality control system, we refer to Kabas et al. (2011), Kirchengast et al. (2014), Scheidl (2014), and Fuchsberger et al. (2020b). In this study, we used WegenerNet gridded data from WegenerNet"s level 2 data (Fuchsberger et al., 2020a), generated with the inverse-distance-squared weighting method based on quality-controlled station data, and provided on a 200 m grid. WegenerNet data have been validated against data from operational weather stations (O et al., 2018) and is shown to have a reliable performance in terms of magnitude, frequency, and the exact location of extreme 100 events . The data have been used as a reference in multiple validation studies, a selection of which is addressed below. Kann et al. (2015) used WegenerNet data to validate 6 months of INCA data (see Sect. 2.2) and O et al. (2017) used WegenerNet data as a reference to evaluate satellite data from the Global Precipitation Measurements (GPM) mission. Based on half-hourly Integrated Multi-satellite Retrievals for GPM (IMERG) rainfall estimates for the period of April-October in 105 2014 and 2015, the results indicate that all IMERG products perform better in estimating moderate rainfall (0.3 to 3 mm per 30 min) than light and heavy rainfall events. In general, IMERG Early and IMERG Late overestimate low rain rates and all three IMERG runs tend to underestimate heavy rain. In another study, Lasser et al. (2019) (2019) analyzed the spatial variability of heavy rainfall events using WegenerNet gridded data. In addition, they described the dependency of area-mean rainfall on the number of gauges and temporal resolution during heavy events.
The study showed that from May to September, the spatial variability in rainfall is higher than from October to April, due to a higher proportion of convective events. Based on their results, the high density and the regular distribution of WegenerNet 115 stations generate spatially homogeneous gridded rainfall fields. A complete up-to-date list of WegenerNet-related literature can be found at https://wegcenter.uni-graz.at/en/wegenernet/publications/ (last access: 16 January 2021).

INCA
The INCA precipitation analysis provides data on a 1 km × 1 km spatial grid with 15 min temporal resolution, using a combination of rain gauge data, weather radar estimates, and high-resolution topography . The 120 following data are used as input for generating the INCA precipitation analysis product:  The topography, based on digital elevation data provided by the United States Geological Survey (USGS).
 Precipitation data from about 250 semi-automatic ground stations (Teilautomatische Wetterstationen, TAWES), operated by the ZAMG, with an average interstation distance of 18 km, all of them located in Austria and two of them in the study area. Note that more stations have been added to the INCA analysis algorithm during the study 125 period.
 Precipitation data from the Austrian hydrographic service (AHYD), which were added over time.
 Radar data from five Austrian C-band radars, supplemented by data from weather radars of neighboring countries.
Starting from 2011, four of the Austrian radars were replaced by new ones (see Table B1 in Appendix B for more details). 130 INCA data are generated based on a Lambert conformal conic projection as a coordinate system, with reference latitudes 46° N and 49° N, and a central reference point at 47°30" N, 13°20" E.
Steps taken to produce INCA precipitation analyses are described by Haiden et al. (2011). A brief summary of this process is given below: 1. Interpolation of station data using inverse-distance-squared weighting. Note that there are two ZAMG stations in 135 the study area, namely: Feldbach station (11298) and Bad Gleichenberg (11244), which measure precipitation with 1 min temporal resolution. These two stations were added to this step in September 2011.
2. Climatological scaling of the radar data using a climatological scaling factor to partially correct topographic shielding. The scaling factor is the ratio between the multi-year, 3-monthly accumulated precipitation from the station data and the corresponding accumulated radar precipitation data. 140 3. Rescaling radar data using the latest observations: based on the comparison of observations and radar fields at the station locations, the fields from the last step are rescaled again. The rescaling is a weighted average of the ratio between the data from the radar and the nearest rain gauge, where the weight decreases with increasing distance, increasing difference in climatological scaling, and decreasing rain at the station. INCA is equal to the interpolated station field of step (1). Between the stations, the weight of radar information increases. In the areas where the radar return is weak due to orographic shielding, the analysis reduces to station interpolation, considering elevation effects.
Related to this study, it should be noted that the closest radars to the study area are Zirbitzkogel (approx. 100 km) and 150 Rauchenwarth (approx. 140 km) (see Table B1 Appendix B). Considering these distances and the mountains between the study area and radars, the minimum detection height by the radar network in the study area is about 2000 m above the ground, leading to detection and estimation errors. Based on Kann et al. (2015), the ground clutter correction is the only correction of radar data. Hence, some errors such as bright band, signal attenuation, scan strategy, radar miscalibration, radome wetting, and errors due to non-meteorological echoes may still exist in INCA precipitation products. 155 Kann et al. (2015) used WegenerNet station data to evaluate 5 min INCA analysis data (rapid-INCA) for wet season (April-September) of 2011 and 4 different heavy precipitation events. The study showed a general underestimation in rapid-INCA during the wet season. The rapid-INCA also underestimated the average precipitation rate in three out of 4 events. They also showed the roles of rain gauges and radars in rapid-INCA analysis performance.

Data Preparation
Precipitation data from 2007 to 2018 are used in this study. After transforming WegenerNet gridded data to the Lambert conformal conic projection, we used the conservative remapping scheme (Jones, 1999) to generate 1 km gridded fields (see Fig. B1 in Appendix B). The conservative remapping scheme is based on preserving water flux and has been widely used as a remapping scheme for precipitation observations and climate model outputs (e.g., O"Gorman, 2012, Nikulin et al., 2012, 165 Sillmann et al., 2013, Prein and Gobiet, 2017, Tapiador et al., 2020, Fallah et al., 2020. We aggregated the 5 min WegenerNet data to 15 min to have the same temporal resolution in both datasets. As an example, Fig. 2   Although WegenerNet gridded data are generated from quality-controlled station data (Kirchengast et al., 2014), some errors may still remain in the data. The Mahalanobis Distance methodan approach to find multivariate outliers (Ben-Gal, 2005)was implemented to detect these possible errors in WegenerNet gridded data. A brief explanation of this approach can be found in Appendix A. We only found a few grid cells that had to be treated as outliers for parts of the measurement period 175 (see Table A1 Appendix A): Station 44 is peculiar because it is located on top of a tall building, at 55 m above ground, and therefore suffers from stronger wind-induced undercatch. Station 145 is another outlier, which has been detected by the WegenerNet team in 2015, which led to a replacement of its rain gauge. Since we found only few outliers, it can be concluded that the operational quality control system of the WegenerNet can filter outliers reasonably well.

Approaches 180
We use the annual area-mean precipitation and compare it in both datasets to gain a general overview of INCA developments and find possible systematic errors. The annual difference between the two datasets in each pixel is also calculated to portrait the spatial pattern and the possible relation with ZAMG stations' distance. This is supported by a Since the Austrian weather radars overlook events that happen lower than about 2000 m above the ground in this region (due 190 to radar beam blockage by the surrounding mountains) and the radar data are only corrected for ground clutter, some detection errors such as missing events and classifying non-precipitation phenomena as precipitation can occur. We implement three indices: Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI) to address these errors in INCA. We calculate POD and FAR indices for each grid cell, classifying them into seven different precipitation intensities from 0.1 to 5 mm per 15 min. We then illustrate them for each pixel in order to analyze possible 195 dependencies on the distance from the two ZAMG stations.
One way to address extreme precipitation is to consider the higher part of the intensity distribution (e.g., 99th quantile) as a threshold and calculate the average intensity of all time steps with higher intensities (the highest 1 %). The benefit of using this approach is that it includes changes in all events above the threshold (Haylock and Nicholls, 2000). We consider all events with equal or more than 0.1 mm per 15 min that happened in both datasets simultaneously and then calculate the 200 average intensity of the highest 1 % for each dataset.
Since event-based analysis is important for some hydrological studies, such as soil erosion and runoff generation, we separate events in each dataset and evaluate INCA"s ability from this perspective. We use a simple threshold based on the minimum dry period between two consecutive events (Minimum Inter-event Time MIT) to separate individual events. This approach has been used in different studies (e.g., Brown et al., (1985); Haile et al., (2011)) to identify individual events. A 205 wide range of MIT values has been selected in the literature. Choosing different MIT values leads to different characteristics of derived events. There should be a compromise between the independency of events and intra-event variability (Dunkerley, 2008). Since the study is affected by both convective and large-scale systems, we choose the MIT value to be 1 hour and the minimum precipitation to be 0.1 mm per 15 min at any pixel. After separating events based on these criteria for each dataset, we analyze the characteristics of these events, such as event duration, accumulated precipitation, area-average intensity, peak 210 intensity ,and the average number of wet cells, both in INCA and in WegenerNet. Note that the accumulated precipitation is the total amount of area-average precipitation during an event (mm), area average intensity is the total precipitation divided by the duration of an event (mm h -1 ), peak intensity is the maximum intensity at a pixel during an event (mm per 15 min), and the average number of wet cells is the average number of cells that have more than 0.1 mm per 15 min of precipitation during an event. 215 To study extreme convective short-duration events (ECSDEs) based on the events in the previous paragraph, we define an ECSDE by having three characteristics: 1. the area average intensity is more than the 95 th quantile of WegenerNet intensities, 2. the duration is less than four hours, and 3. a coverage of less than 2/3 of the study area. We evaluate INCA"s performance in these events with a focus on the spatial characteristics of 4 ECSDEs in both datasets, along with the area-mean and peak intensities. 220 https://doi.org/10.5194/hess-2021-34 Preprint. Discussion started: 15 February 2021 c Author(s) 2021. CC BY 4.0 License.

Comparison Metrics
Error metrics have been widely used in many different disciplines, such as hydrology and hydrometeorology, to quantify model result accuracies or compare observations and forecasts. Each metric has its own strengths and weaknesses and quantifies a different aspect of the model. Hence, using multiple metrics for the comparison is more desirable (Jackson et al., 2019). We use four common metrics: Bias, Relative Difference (RD), Root Mean Square Error (RMSE), and Correlation 225 Coefficient (CC), for comparing INCA precipitation estimates with WegenerNet. These indices are computed according to Eq. (1) to (4) in Table 1 below. Where RR and WR indicate the precipitation estimates by INCA and by WegenerNet, respectively; i is the index of the time 230 step, N is the number of time steps and the top bar shows the average over time.
Additionally, in order to evaluate INCA"s ability to detect precipitation, we analyzed three categorical indices: Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI), based on the equations in Table 2    WegenerNet except for 2007, 2009, and 2010. Based on these results, we can conclude that using the merged radar-gauge data leads to more accurate estimates than individual ZAMG station data, even at the ZAMG station location. Furthermore, and similar to Fig. 3b, we show the annual precipitation in the cell where station Bad Gleichenberg is located (Fig. B2 Appendix B). Comparing to station Feldbach, the difference between INCA and WegenerNet is higher, which could be due to the weight of this station in step 1 (see Sect. 2.2). 260   Note that we also calculate RMSE for the annual precipitation (see Fig. B3 Appendix B). There is no noticeable spatial pattern in the first period, and the minimum and maximum of area-mean annual RMSE error are 45.0 mm and 57.5 mm, respectively. This error is considerably higher in the second period with a minimum and maximum of 191.0 mm and 278.0 mm, respectively. In this period, RMSE is lower close to the Feldbach station (just like the relative difference in Fig.  290 4). Similar to the relative difference, the area-mean annual RMSE decreases in the 2015-2018 period. For this period, the minimum RMSE is 92.6 mm in 2015 and the maximum RMSE is 124.1 mm in 2018.
As mentioned before, more stations were added to the INCA algorithm during the years. To study possible changes on INCA estimates after adding these stations, the raw radar data used in INCA should be included in future analyses.   We also calculated CSI for each cell (Fig. B4 Appendix B). Since CSI is a combination of POD and FAR, its behavior is a combination of POD and FAR; i.e., CSI performs better during the wet season. CSI values are higher in the third period (2015-2018) for the dry seasons. Since the weight of the radar estimate becomes higher with increasing distance from the 315 ZAMG stations in the INCA algorithm (see Sect. 2.2), we conclude that the radar detected some precipitation events, which were not observed by ground stations. Since the radar sees precipitation in the study area only beyond about 2000 m above the ground and some errors are not corrected (see Sect. 2.2), "false events" can be due to events that do not reach the ground due to evaporation or due to non-precipitating phenomena. The latter error can explain higher FAR values in the cells with longer distances from ZAMG stations.  INCA grid cells, similar to annual precipitation. This overestimation is higher in the wet seasons and grid cells farther away from the two ZAMG stations. We can conclude that the new radar settings (cf. Sect. 2.2) tend to overestimate precipitation and that the INCA algorithm works reasonably well closer to the two ZAMG stations. However, the overestimation farther 335 away from those stations is still considerably large. The higher relative difference in wet seasons is an indication of difficulties in the radar network to estimate intense rainfall events. Considering RMSE, the pattern is similar to the relative difference (not shown).

Precipitation detection 295
Furthermore, we calculate the temporal correlation coefficient between INCA and WegenerNet (Fig. B5 Appendix B). Based on these results, the correlation is noticeably lower in the wet seasons. We interpret this as a consequence of a higher 340 percentage of convective events, which are harder to capture. Similar to the relative difference, INCA performs better in the cells close to the ZAMG stations in the wet seasons.

Extreme precipitation
In this section, we compare extreme events in INCA and WegenerNet based on the different seasons for the three periods.
Note that the 99th quantile was calculated for time steps with precipitation equal to or more than 0.1 mm per 15 min, which 350 happened in both datasets (see Sect. 3.2). Figure 8 shows the mean values of all time steps exceeding quantile 99 in each pixel for both datasets and the relative differences in wet and dry seasons. underestimated it in the other parts of the study area, similar to the mean seasonal values (see Fig. 7). Compared to the mean seasonal values, underestimation is larger in extreme precipitation, especially in the dry season. The maximum and minimum differences in this period are 26 % and -45 %, respectively. It is worth mentioning that the spatio-temporal evolution of extremes is not particularly well captured by INCA (cf. Fig. 2).
Between 2012 and 2014, an overestimation of INCA in the corresponding cells of ZAMG stations is also noticeable, 360 particularly in the wet season. In contrast to annual and mean-seasonal values, there is no relationship between distance from ZAMG stations and the relative difference. In the dry season, INCA shows less underestimation than in the first period. The maximum and minimum differences in this period are 101 % and -36 %, respectively. For wet seasons in the 2015-2018 period, the behavior was relatively similar to the second period with a decrease in overestimation. In this period, the maximum and minimum differences are 44 % and -29 %, respectively. Based on these results, the overestimation of INCA is 365 larger in extremes, especially in the wet seasons.

Event-based evaluation
In this section, we consider individual events, and based on the criteria we described in Sect. 3.2, we identified 4699 separate events in INCA and 5116 in WegenerNet over 12 years. The number of events in different seasons is shown in Fig. 9. Similar to Sect. 4.2, the number of detected events in INCA is lower in the dry seasons. Note that the number of dry-season-370 events can be slightly biased in WegenerNet, due to snowfall events, which can get recorded twice: Once when the snow is measured by heated rain gauges, and again when the snow melts at the unheated gauges. In general, snowfall events in the region are rare (cf. Sect. 2.1); thus we do not expect them to have significant influence in metrics other than the number of events. Another effect on the number of dry season events is that the radars tend to miss precipitation more often in winter due to beam blockage by surrounding mountains (cf. Sect. 2.2), especially for low lying clouds which are often present in the 375 dry season.  Table 3 describes the statistics of separate events in both INCA and WegenerNet. Note that the accumulated precipitation 380 and precipitation rate are based on the area-mean value in each time step. The peak intensity, however, is the maximum value that happened in one cell during an event.
https://doi.org/10.5194/hess-2021-34 Preprint. Discussion started: 15 February 2021 c Author(s) 2021. CC BY 4.0 License.  Table 3, the average accumulated precipitation and the precipitation rate measured by INCA are higher than WegenerNet. Similarly, the average number of wet cells is higher in INCA, which can affect the accumulated precipitation.
The difference between the average number of wet cells in INCA and WegenerNet is higher in the dry season. This could be due to a slightly lower effective resolution of INCA in the study area, where the radar beam of the nearest radar is already comparatively wide. Whereas the average duration of events only differs by 3 min, the difference increases significantly for 390 longer events.
To check a possible time shift in INCA, we consider events that fulfill the following conditions: 1. The absolute difference between the starting time of an event in INCA and WegenerNet is less than 1 hour, and 2. the absolute difference between To separate the errors associated with time shift from errors related to intensity, we focus only on those events that happened at the same time in both datasets and we found 2949 events. Similar to the results in Table 3, the accumulated precipitation is 400 higher in INCA and the bias value is 0.14 mm 15 min -1 . Although INCA overestimates accumulated precipitation most of the time, the peak intensity is slightly higher in WegenerNet except during July. The overall bias for peak intensity is -0.04 mm 15 min -1 .
We also studied the time of the peak intensity in both datasets and found that the peak happens during the first half of the event duration. Also, in the majority of events, the peak intensity in INCA happens slightly later (approximately 5 min) than 405 in WegenerNet.
The monthly average of the peak intensity is shown in Fig. 10. INCA generally underestimates the precipitation peaks in the first period but generally overestimates them from mid-2012 onwards. This behavior is likely due to the change of the radar https://doi.org/10.5194/hess-2021-34 Preprint. Discussion started: 15 February 2021 c Author(s) 2021. CC BY 4.0 License. network in 2012. There is a noticeable peak-intensity overestimation in mid-2013. Contrary to the mean precipitation (see 410 Fig. 3a and 4), the differences in peak-intensity between INCA and WegenerNet decreased significantly in 2018. Comparing these results with those by (Kann et al., 2015), we see that the 15 min INCA precipitation analysis performs better than 5 min rapid INCA. The time shift in detection was also observed in rapid-INCA, caused by radar data.

Conclusions 475
The evaluation of precipitation estimates helps to improve the understanding of errors and uncertainties from different sources (e.g., systematic errors, random errors, and spatio-temporal dependency). In this study, we evaluated INCA considerably overestimates precipitation by up to 60 %. However, this overestimation is less pronounced close to the station Feldbach. Starting in 2015, this spatial pattern continues but with a lower overestimation compared to the second period. We 485 conclude that this overestimation is a result of systematic errors from newly installed radars. This overestimation was partly removed in the INCA algorithm using reference gauges.
We used categorical metrics to study the ability of INCA in detecting precipitation. Generally, the number of false events is smaller in the cells closer to ground stations operated by the ZAMG (which are used as input for INCA), especially in the wet season. Surprisingly, the number of true events close to the ZAMG stations is comparably smaller too. This could be due 490 to the INCA algorithm for removing false precipitation events that unintentionally removes some light precipitation events.
We evaluated extremes during these three periods, and for wet and dry seasons. In the first period, INCA overestimates precipitation in cells close to the ZAMG stations and underestimates it in other cells, especially during the wet season. This overestimation is more noticeable, and dominates most of the study area in the wet season from 2012 until 2014. However, for the dry seasons in this period, underestimation by INCA is dominant. For the third period, the pattern for the wet seasons 495 is similar, but INCA tends to overestimate extremes during the dry seasons.
We also considered individual events in both datasets and analyzed their characteristics. Based on these results, INCA tends to underestimate the peak precipitation intensity of the events until mid-2012 and overestimates it afterward. The largest overestimation of the peak-intensity happened in July. Generally, the precipitation rate is higher in INCA, and there is a time shift for event detection in INCA. Based on our results, INCA starts to detect precipitation earlier than WegenerNet in 40 % 500 of the events and ends detecting it later than WegenerNet in 50 %. Considering four examples of extreme short-duration convective events, there is a time-shift in detecting the peak intensity in INCA. In these events, the peak intensity bias is considerably larger than in all events. In general, INCA has been improving in detecting and estimating precipitation.
However, there are errors due to radar estimates and the algorithm for merging radar and rain gauges, which can negatively affect the INCA analysis product. In addition, it is shown that gauges are crucial for correcting some errors due to radar 505 estimates. Careful consideration must be taken when using merged rain-gauge-radar products, especially in extreme events.

Suggestions for future studies:
For future studies, it is suggested to include the raw radar data used in INCA in the analysis, to separate errors due to radar estimates. Also, we suggest using the results of this study to consider high-impact events and analyze the effects of INCA uncertainties on risk management. In addition, the relation between wind speed and precipitation estimates in both gauges 510 and radars needs to be considered separately.

540
Note that only events with equal or more than 0.1 mm per 15 min were considered. The black circles indicate the two ZAMG stations in the study area. Research, the University of Graz, the state of Styria (which also included European Union regional development funds), and 555 the city of Graz; detailed information can be found online (http://www.wegcenter.at/wegenernet, last access: 16 Jan. 2021).