Selection of intense rainfall events based on intensity thresholds and lightning data in Switzerland

This paper presents a method to identify intense warm season storms with convective character based on intensity thresholds and the presence of lightning, and analyzes their statistical properties. Long records of precipitation and lightning data at 4 stations and 10 min resolution in different climatological regions in Switzerland are used. Our premise is that thunderstorms associated with lightning generate bursts of high rainfall intensity. We divided all recorded storms into those accompanied by lightning and those without lightning and found the threshold I ∗ that separates intense events based on peak 10 min intensity Ip ≥ I ∗ for a chosen misclassification rate α. The performance and robustness of the selection method was tested by investigating the inter-annual variability of I ∗ and its relation to the frequency of lightning strikes. The probability distributions of the main storm properties (rainfall depth R, event durationD, average storm intensityIa and peak 10 min intensity Ip) for the intense storm subsets show that the event average and peak intensities are significantly different between the stations. Nonparametric correlations between the main storm properties were estimated for intense storms and all storms including stratiform rain. The differences in the correlations between storm subsets are greater than those between stations, which indicates that care must be exercised not to mix events of different origin when they are sampled for multivariate analysis, for example, copula fitting to rainfall data.


Introduction
Rainfall is one of main causes of natural hazards in hydrological systems (e.g.Vörösmarty et al., 2013;Wilhelmi and Morrs, 2013;Kyselý et al., 2013), and is inseparably related to flood risk (e.g.Hlavcova et al., 2005;Blöschl, 2008;Borga Correspondence to: Peter Molnar (molnar@ifu.baug.ethz.ch) et al., 2011;Winsemius et al., 2013).Storms associated with high rainfall intensities often lead to severe flooding in catchments and urban areas, accelerated hillslope and channel ero-35 sion, triggering of landslides, mud and debris flows (e.g.Llasat, 2009;Brunetti et al., 2010;Badoux et al., 2012).The severity of storms and their damage potential is dependent not only on the total rainfall amount but also on other storm characteristics such as peak rainfall intensity and event dura-40 tion.The statistical properties of these storm characteristics and their inter-dependencies are fundamental descriptors of intense storms for both practical (design) and scientific purposes.That is the general aim of this paper.
The first question we address is how to select intense 45 storms out of all events based on some common property.A good example is the distinction of precipitation into convective and stratiform rain which captures two different cloud dynamical processes (e.g.Houze, 1997).Convective rain is associated with higher rainfall intensities and is commonly 50 identified from ground synoptic observations of clouds, state of weather and/or rain intensity fluctuations (e.g.Pešice et al., 2003;Chvila et al., 2005;Tremblay, 2005;Llasat et al., 2007;Berg et al., 2013;Ruiz-Leo et al., 2013;Rulfová and Kyselý, 2013).Weather radar provides additional informa-55 tion, e.g.updraft velocities and cloud water content, which aid the identification of convective rain (e.g.Rigo and Llasat, 2004;Llasat et al., 2005;Rigo and Llasat, 2007).Although formal convective-stratiform separation techniques (see comparison in Lang et al., 2003) are inherently limited because 60 they do not directly quantify the origin and type of convection (Zimmer et al., 2011) they have one thing in common -that a locally high rainfall intensity is expected in events where convection is present even if embedded in stratiform systems.It is this extremity of rainfall intensity during a storm which we would like to capture in our selection of intense events, regardless of their exact meteorological origin and genesis.
The second question we address in this paper is how to define a threshold intensity for the identification of intense storms.This choice is rather arbitrary and depends on the time resolution used, i.e. the duration for which the intensity exceeds the threshold (e.g.Llasat, 2001).There are very few objective methods to estimate the threshold intensity, one notable exception is Tremblay (2005) who partitions rain into stratiform and convective components with the latter being anomalies in the relation between total precipitation and precipitation intensity (e.g.Ruiz-Leo et al., 2013).This method is based on a global classification from raingauges and provides time-dependent but not station-specific thresholds.It was developed for 6-h aggregations which are not always sufficient to capture short-term variations in rain intensity during convective storms.
One of the main goals in this paper is to develop a stationbased method to choose the intensity threshold based on long records of lightning strike data and precipitation measured at a high resolution (10 min).The premise is that summer thunderstorms with convective activity are commonly associated with lightning (e.g.Tapia et al., 1998;Rivas Soriano et al., 2001;Rivas Soriano and De Pablo, 2003;Barnolas et al., 2008;Tadesse and Anagnostou, 2009;Yair et al., 2010;Koutroulis et al., 2012).A number of recent studies have demonstrated the connection between intra-cloud and cloudto-ground lightning and convective precipitation using rainfall yield, i.e. the amount of rain per lightning flash, as a measure (e.g.Petersen and Rutledge, 1998;Grungle and Krider, 2006;Price and Federmesser, 2006;Pineda et al., 2007;Kochtubajda et al., 2013).Tapia et al. (1998) developed a model to predict the space-time distribution of rainfall as a function of rainfall yield and lightning strikes.However, the predictive strength of lightning count as an explanatory variable on a regional scale in these studies is not very strong.
The approach we take instead is that the simple presence of lightning in the vicinity of a raingauge is indicating the existence of convective activity.We find the rainfall intensity threshold that gives us an acceptable separation error between events accompanied by lightning and those that are not, and study their statistical properties.Although we use accurate station-based lightning data, the same approach applies to lightning data from long-range lightning detection ground networks or satellites (e.g.Defer et al., 2005;Price and Federmesser, 2006;Kotroni and Lagouvardos, 2008).Our study complements other recent efforts focused on using lightning observations to better understand and predict convective storms and their consequences, e.g.flash floods (Price et al., 2011a,b;Koutroulis et al., 2012).
The main contributions of this paper are threefold: (a) We provide a new method to identify intense warm season rainfall events as those in which the maximum intensity exceeds a threshold by using lightning strike data as an explanatory variable.We analyze the performance and robustness of this method by investigating the relations between the estimated threshold and lightning statistics.(b) We show the probability distributions of main storm properties for the intense storm subset identified by our method and discuss the dif-125 ferences between the studied stations.(c) We present nonparametric correlation coefficients between the main storm properties for the intense storm subset and contrast these with the correlations for all storms to highlight the particular interdependencies (or lack thereof) in intense storm properties.

130
The analysis is conducted on 4 stations in different climatological regions in Switzerland and provides a basis for a follow-up regionalization effort.

Meteorological data and study sites 135
The data used in this study are from stations of the Swiss-MetNet network (MeteoSwiss).These are automatic weather stations recording a range of hydroclimatic data at a 10-min sampling resolution.Our main data are precipitation and air temperature for the period 1981-2012 (32 years) and light-140 ning strikes for the period 1987 to 2005 (19 years) at 62 stations with about 3000 meter altitude difference (Figure 1).
Precipitation is recorded by a heated tipping-bucket raingauge (Lambrecht) with orifice area 200 cm 2 and tip resolution 0.1 mm.The data are quality checked and corrected

Event definition
Our analysis is based on precipitation events which we iden-165 tify separately for the warm (April-September) and cold (October-March) half-years, which we also call seasons in this paper.For the selection of potentially intense rainfall events we limit ourselves to the warm season only and add a constraint that the air temperature during an event T > 4 • C 170 to ensure liquid precipitation and minimize the likelihood of snow or mixed events at high altitude stations.In the determination of individual events we disregarded all intervals with only 1 tip in 10-min (0.1 mm/10-min) because these are often artifacts of condensation and funnel drainage at the end of events and increase the duration of events with minimal increase in rainfall depth.
In practice it is common to define individual rainfall events by rainless intervals exceeding a selected duration called the inter-arrival time (also inter-event time, inter-storm period) t i (Figure 2).Many application-based criteria for identifying rainfall events using a fixed t i have been presented in the literature, with values ranging from 3 min to 24 hours (e.g.Dunkerley, 2008Dunkerley, , 2010)).In our study we have chosen t i = 2 hrs based on the de-correlation time in the rainfall records (see next section for details).
For each event we determined four main event properties: total rainfall depth (R), event duration (D), average rainfall intensity (I a = R/D), and peak (maximum) 10-min intensity during the event (I p ) (Figure 2).Table 1 summarizes the final number of events per warm season for the four stations, which ranged from an average of 84 (0.46 per day) at Scoul to 98 (0.54 per day) at Napf in the pre-Alpine area.

Inter-arrival time
An objective approach to choose t i is to ensure that subsequent events are independent.Independence can be evaluated on a conceptual basis or statistically.The conceptual basis may for example be a model of the interception process which would identify independent events as those where t i exceeds the time scale of evaporation from a fully saturated canopy (e.g.Lloyd, 1990;Zeng et al., 2000), or a time period during which soil wetness decreases sufficiently not to impact the subsequent flood runoff (e.g.Bracken et al., 2008).The statistical basis may for example be that t i is sufficiently long for the autocorrelation in rainfall to become statistically insignificant (Grace and Eagleson, 1967), or the probability distribution of t i closely exponential, like that of independent events in a Poisson process (Restrepo-Posada and Eagleson, 1982).Selecting long t i may prove useful for ensuring the identification of statistically independent events, but in turn, it has a negative effect on storm properties by creating extensive intra-event gaps, leading to a strong bias in storm duration and average intensity (e.g.Restrepo-Posada and Eagleson, 1982;Dunkerley, 2010).
We chose a combination of conceptual and statistical considerations in identifying t i , bearing in mind that we want to preserve the extremal properties of intense rain events.Conceptually the values of t i were a-priori restricted to the order of the usual time scale of summer thunderstorms, and statistical independence of the events was ensured by analyzing the autocorrelogram of the sequences of 10-min rainfall.The lag at which the autocorrelation coefficient became statistically insignificantly different from zero for a chosen significance level indicates the inter-arrival time at which the storm events may be regarded as independent.
The autocorrelogram, i.e. the Pearson correlation coefficient as a function of time lag was determined for each station and each warm season separately.The autocorrelation functions for individual seasons were then averaged and are shown as dots in Figure 3.The statistical significance was 230 tested at the significance level 10%.The de-correlation time ranged between 2 and 3 hr, and this was similar for the majority of the 62 SwissMetNet stations.Some mountain stations, such as Scuol in Figure 3, show slightly longer de-correlation times.From this result we conclude that t i = 2 hrs gave us an 235 appropriate de-correlation time in the rainfall records.Moreover it also matches the order of lifespan of thunderstorms in general.

Lightning strikes
Lightning-induced electric charge was measured locally at Lightning data are reported as the number of strikes within a certain range from the station at a 10-min time resolution.The sensitivity of the antenna allows an identification of close (<∼ 3 km) and distant (between ∼ 3 − 30 km) light-250 ning.Separation between cloud-to-ground and intra-cloud lightning is not possible.Figure 4 shows the most interesting statistics of lightning data at the Geneve station.Because this data has to our knowledge not yet been reported in the literature we highlight the following main features which were 255 common among all stations.
There are clearly many more distant strikes than close ones.However, their ratio remains rather stable across years.The occurrence and actual sum of strikes in the cold (CHY) and warm (WHY) half-years are dramatically different, with 260 more than 90% of the strikes in the warm season indicating the connection between convectivity, ice-build up in cloud systems, electrification, and thunderstorm activity.The monthly distribution of lightning shows that the three most active months are June-July-August.The hourly distribution 265 of lightning further indicates that lightning has a clear diurnal pattern with maxima between 15:00 and 18:00 hours UTC in the afternoon when convective activity is highest.There were some small differences at the Napf, Scuol and Lugano stations but the general tendencies were identical.Even if we do not take into account the actual count of lightning strikes, because this is a highly uncertain and site-dependent quantity, and only the presence or absence of lightning, the seasonal and diurnal statistics do not change appreciably.
Lightning data sheets like that in Figure 4 for Geneve are 275 available from the authors for all 62 stations of the Swiss-MetNet network.
The method to select intense warm season storms out of our set of all storms is based on identifying events during which the peak 10-min intensity exceeds a threshold I p ≥ I * .This subset of intense storms can be assigned a convectivity index for every event j as the fraction of the event total rainfall depth that falls with intensities greater than I * (Llasat, 2001), All events belonging to the set of intense events then have 0 < β ≤ 1 and those that are not considered intense β = 0.
Based on the findings that the occurrence of intense summer rainfall is indeed well correlated with lightning activity, especially over land (e.g.Petersen and Rutledge, 1998;Rivas Soriano et al., 2001;Rivas Soriano and De Pablo, 2003;Pineda et al., 2007;Barnolas et al., 2008;Tadesse and Anagnostou, 2009;Yair et al., 2010;Koutroulis et al., 2012), we propose to calibrate the threshold parameter I * at every station as a function of the presence or absence of lightning.In 295 Section 3.1 we first analyze the storm properties for all events with and without lightning, and in Section 3.2 we then develop a statistical approach to estimate I * based on lightning data.

Storm properties and lightning
All storms in our dataset were divided into those that were accompanied by lightning and those that were not.In this step we did not consider the actual count of lightning strikes but only the presence of lightning, similar to Katsanos et al. (2007) who used a fixed 6-h accumulation step for stations in the Mediterranean.We chose to include distant strikes as well as close strikes because we are defining a property of the events which is dependent on a larger scale propensity of the atmosphere around the station to generate extreme rainfall.In associating lightning with storm rainfall we also considered strikes that occurred up to 20 min before the actual rainfall event began, which captures commonly observed time lags in pre-storm lightning-rainfall correlations (e.g.Grungle and Krider, 2006;Koutroulis et al., 2012).
We found that 31% (Scuol), 39% (Napf), 40% (Geneve), and 62% (Lugano) of the events were associated with lightning (Table 1).In this statistic Lugano was an exception and most stations in the SwissMetNet network had between 30-40% of storms with lightning on the average.Naturally stations at higher elevations and/or in valleys surrounded by mountain peaks have a higher proportion simply because lightning strikes are more frequent there.Histograms of the four main rainfall event properties (R, D, I a , I p ) were compared for the subsets of events with and without lightning for the period 1987-2005.The differences in the distributions were objectively assessed with statistical tests (Kolmogorov-Smirnov). Figure 5 shows the results for Geneve.The results were consistent for the other stations and show that rainfall duration D and total rainfall depth R are not statistically significantly related to the occurrence 330 of lightning, except for very short durations and low rainfall totals.However, average rainfall intensity I a and especially peak rainfall intensity I p were consistently different for events accompanied by lighting.
The main difference in I p was that events accompanied 335 by lightning had higher peak intensities in general and the largest peak intensities were in fact observed only in storms with lightning.Also very low peak intensities were much more prevalent in storms without lightning.This result was true for all analyzed stations.We conclude that the short term 340 peak 10-min intensity I p during an event is indeed the most useful variable to separate intense summer storm events of a mostly convective character from our event dataset.

Calibrating the intensity threshold
The threshold intensity I * is an unknown parameter in our 345 approach which selects intense storms of a mostly convective character out of all storms as those for which I p ≥ I * .This is our model under the assumption that lightning data is indicative of convective activity.We can then use the two subsets of storms with and without lightning for the calibra-350 tion of the intensity threshold.For any I * we can compute the sensitivity, i.e., the fraction of events accompanied by lightning which are identified as intense storms (true positives), and specificity, i.e., the fraction of events not accompanied by lightning which are identified as not intense storms 355 (true negatives).Plotting sensitivity versus specificity for a range of I * (see Figure 6 for Geneve) gives us a complete indication of how good our model is because the departure of the data points from the diagonal indicates the power of the model to make a reliable selection (e.g.Begueria, 2006).

360
We choose an acceptable misclassification error α of no lightning events as convective (false positives, Type I error) to make a decision for I * as the threshold where specificity = 1 − α.This decision is based on the argument that when lightning occurs in the vicinity of the gauge but we do not 365 measure high intensity rain this could be caused by sampling the edge of a passing convective cell.While when lightning does not occur but we do measure high intensity rain this is indeed an error in our model which we would like to minimize.This error contains all the physical processes that may take α = 0.1, and the corresponding I * values are between 5.8 and 9.2 mm/h at the four stations (Table 1).

Results
We present the results by first discussing the calibrated intensity thresholds I * with regard to some previous studies.
We then analyze the performance of the method to select intense warm season rainfall events, and finally present the basic statistics of the main storm properties for this set of events and their cross-correlations.

Comparison of intensity thresholds
The calibrated thresholds to identify intense warm season storms with convective character at the four studied stations (between 5.8 and 9.2 mm/h) are lower than what is reported in most climatological and engineering hydrology literature for extreme convective rain.For example, Dutton and Dougherty (1979) and Watson et al. (1982) present a threshold of 48-50 mm/h for considering rainfall to be all of convective origin.Llasat (2001) considers that convective rain events exceed a peak 5-min intensity of 35 mm/h, and a threshold for average hourly intensities for heavy rainfall defined by the Spanish National Meteorological Institute is 15 mm/h.Earlier studies, like Wussow (1922) developed an empirical formula for Germany which gives 7.1 mm for strong rainfall of 10-min duration, i.e., 43 mm/h, and Reinhold (1940) classified heavy rainfall when intensities for 10-min durations exceeded 40 mm/h.Velikanov (1964) describes a classification where the threshold for 10-min intensive rainfall is given as 3.8 mm, i.e., 23 mm/h.This classification was also used in former Czechoslovakia by Dub and Němec (1969).Dunkerley (2010) reports that in the context of erosion an event was only significant by Wischmeier and Smith (1978) if either the rain depth was larger than 12.5 mm or the rain rate exceeded 6.5 mm in 15 min, i.e., 26 mm/h (Kinnell, 2003).
A direct comparison of our thresholds with the above (and other) studies is not possible for several reasons.First is that these studies are generally focused at identifying very heavy convective rainfall events only, while we are interested in selecting a larger set of storms which are of convective character and not only the most extreme of them.Second is that many of these studies use different temporal resolutions of the rainfall data which strongly affects the threshold intensities (e.g.Llasat, 2001), and naturally the results vary considerably from one region to another based on the local climate.In our work it is the misclassification rate α which determines I * .If we are stricter on the acceptable Type I error in our model (lower α) then the threshold I * increases.However, this goes at the expense of increasing Type II error, i.e. our model will have a lower sensitivity with respect to identifying events accompanied by lightning as convective (see Figure 6).

Testing the robustness of the method
Because lightning data which are key in our method to estimate I * may vary strongly in time and space and be nonhomogeneous, a question to be answered is how robust the 430 results are with respect to lightning variability.We examine two aspects: the inter-annual variability of I * and its relation to the frequency of lightning strikes.
Inter-annual variability in I * estimated on a bi-annual basis is shown together with the estimate for the entire 18-year 435 record (until 2004) with reliable lightning data for all four stations in the left panel in Figure 7.The result is that there is no evident trend in the estimate of I * , although the local station inter-annual variability in lightning strikes, as seen for instance in Lugano, may lead to larger variability in the 440 estimate.We also plot the corresponding bi-annual estimate of sensitivity in the right panel in Figure 7.If lightning was not a predictor of intense events in our model, the sensitivity would be equal to the chosen misclassification rate α = 0.1.The results show that the sensitivity at all stations is in fact 445 much greater than α and is in the range ∼ 0.4 on the average.There is some inter-annual variability in model performance, however, it is not unacceptable.In fact in Lugano the performance is practically constant, which is a good indication of the robustness of the connections between lightning and 450 intense rain over time at this station.
The second aspect to test is whether I * is related to the frequency of lightning strikes and its variability from year to year.We would like to clarify if it is possible that the frequency of storms with lightning changes in time and thereby 455 affects the estimate of I * .The number of storms with and without lightning estimated on a bi-annual basis are shown in the left panel in Figure 8.The proportion of storms with lightning remains reasonably consistent between years at all stations.More importantly, I * is clearly independent of the 460 frequency of storms with lightning shown in the right panel in Figure 8, which allows us to conclude that we have a robust method to extract intense storms from our dataset based on peak 10-min intensity alone, which at the same time coincide well with those which are accompanied by lightning.

Properties of selected storms
Because the subset of intense summer storms should mostly contain intensities associated with convective events, we use the convectivity index defined by Llasat (2001) in Eq. (1) to quantify the strength of convection on an event basis.The 470 cumulative probability distributions of β in Figure 9 show an almost uniform shape at all stations.As expected, the highest convectivity overall was observed in Lugano in southern Switzerland.Llasat (2001) identified events with β > 0.8 as strongly convective in Spain, which would correspond to 475 18% (Napf and Scuol), 20% (Geneve) and 30% (Lugano) of the events in our dataset.
The cumulative distributions of the main storm properties of the events are shown in Figure 10 and some basic statistics are listed in Table 1.The exceedance probability is shown to accentuate the behaviour of extremes.The stations show a very similar distribution of storm duration, except for Lugano which has the longest events.Storm duration is not a variable that is very strongly affected by the different regional climates.However, the other variables are statistically significantly different between the stations.Average event intensity is higher in Lugano than in the other stations, leading to the highest event rainfall depths.Also peak 10-min intensity during an event is highest in Lugano, followed by Napf and Geneve.Extremes are particularly affected.For example an event from our subset with a probability of exceedance of 20%, that is an event that occurs about 5-6 times a year, has a peak 10-min intensity of about 40 mm/h in Lugano, 25 mm/h in Napf and 20 mm/h in Geneve.
One of our main aims was to quantify the correlations between main storm properties for the intense storm subset and contrast these with the correlations for all storms, to highlight the particular inter-dependencies (or lack thereof) in intense storm properties.We chose all storms instead of the subset of less intense storms since we would like to show the differences between our selection and a traditional choice where all storms are used regardless of their meteorological origin.To quantify the correlations we use the Spearman rank correlation coefficient because we intend to use the relations between storm properties for statistical multivariate analysis with the method of copulas.Copula parameters depend on the correlations between the variables, which are generally quantified with a non-parametric correlation like Spearman (or alternatively Kendall) in the literature (e.g.Genest and Favre, 2007;Serinaldi, 2008;Vandenberghe et al., 2010).
The Spearman rank correlation coefficient estimated for all 6 pairs of variables and all stations is shown in Figure 11.The results show that the intense summer storm events have practically no correlation between peak intensity I p and the duration of the event D, while when all storms are considered including stratiform rain this correlation was significantly positive.Another interesting observation is that I p is less strongly correlated to total rainfall depth R in our intense storm subset than in all events.In other words, in events with intensities associated usually with convection, a high intensity over a short period of time does not necessarily lead to a high overall event total, while in events with stratiform rain rates that is generally the case.As expected, average storm intensity I a is negatively correlated with duration D in the convective intense event set but not for all storms.Some of the correlation coefficients in the event sets are statistically significant (Figure 11) and they suggest that a multivariate approach, for instance on the basis of copulas, is advised for probabilistic modelling of extreme storm properties (e.g.Kao and Govindaraju, 2007;Serinaldi, 2008;Bardossy and Pegram, 2009;Vandenberghe et al., 2010).Overall the differences in the correlation coefficients between sta-tions are much smaller than between the convective event set and the entire set of events, which means that it is of first order importance to select wisely the sample of intense events 535 for multivariate analysis.

Discussion of selection method and extremes
The traditional approach to the analysis of extreme rainfall is the depth-duration-frequency (DDF) curve applied to annual maxima or peaks over a threshold (POT), where for a 540 given duration, the most extreme rainfall depths (intensities) in a year are selected and their probability of occurrence is estimated.Estimation methods for the analysis of annual and POT extremes are well known and reported in the literature (e.g.Katz et al., 2002) as well as the mathematical frame-545 work for studying DDF curves (e.g.Koutsoyiannis et al., 1998).Estimation uncertainties are related to fitting an extreme value distribution to the selected data (mostly GEVtype) and functional forms for the DDF relation parameters, and these may be quantified from station as well as weather 550 radar data (e.g.Overeem et al., 2008Overeem et al., , 2009)).The result is an estimated rainfall intensity for a given duration and return period.The threshold intensity comes from the minimum observed annual extreme in the record (annual maxima), or the choice of the average number of events per year (POT).Its 555 basis is purely statistical -it is record period and durationdependent.
Our approach is different in two main ways: (1) We are looking at rainfall events as independent outcomes of atmospheric water vapour fluxes and precipitation formation pro-560 cesses.We do not fix a-priori a given rainfall duration in the analysis (like in DDF curves) but we study several event properties and their statistics.(2) We select events which have a common physical origin and genesis, i.e. the presence of convection, insofar as it is reflected in a high peak 10-565 min intensity during the event.In this way we select many events, which are both short and long, with different rainfall totals, for the analysis, in contrast to only extreme intensities for DDF curves.Our threshold I * is calibrated with lightning data, which are independent measurements.To us this 570 threshold has a physical meaning, as a short-term rain intensity which was exceeded in most convective rainfall events at a given station.
The consequence is that our threshold I * is not in fact directly comparable with POT or any other intense rainfall 575 thresholds which are purpose-driven.Most studies define subjective thresholds based on hydrological risk, monitoring issues, or meteorological considerations.Some of these are discussed in Section 4.1.At this point we can say that despite the fact that our thresholding is different from the DDF 580 approach, the statistics of extremes in our selection of intense storms agree reasonably well with the HADES (Hydrological Atlas of Switzerland) estimates from DDF analysis, where the 1 hr duration extremes with a return period 2.33 yr ranged between 18-34 mm/h at our four study sites (Jensen et al., 1997), while our selection gave peak 10-min rainfall intensities with a non-exceedance probability p = 0.8 (i.e., the largest 5-6 events per year) between 20-40 mm/h.In a subsequent study we will explore the full spatial distribution of event properties across all 62 stations in Switzerland.

Conclusions
Properties of intense storms, such as total rainfall depth, storm duration, average intensity and peak intensity are important descriptors of the rainfall regime at a site.For the statistical analysis of rainfall events and extremes it is often important that the storms in the analyzed sample have a common meteorological origin, for instance convective rain, stratiform rain, hail, snowfall, etc.A similar concept may also be found in catchment hydrology where on the basis of meteorological conditions and catchment state different flood types are defined (Merz and Blöschl, 2003) and statistically analyzed (e.g.Gaál et al., 2012).In this paper we took this common property of intense warm season storms to be the short term (10 min) peak in rainfall intensity.It is this extremity of rainfall intensity during a storm which we would like to capture in our selection of intense events.We estimated the threshold intensity with the help of a unique lightning dataset, thereby capturing the convective character of these intense events.
The main conclusions of the paper are: (1) We developed a new method to identify intense warm season rainfall events as those in which the peak 10-min intensity exceeds a threshold I p ≥ I * by using lightning data as an external variable for the presence of convection.We divided all storms into those accompanied by lightning and those without lightning and found the threshold I * that separates intense events for a chosen misclassification rate α.Altogether 31% (Scuol), 39% (Napf), 40% (Geneve), and 62% (Lugano) of the events were associated with lightning at our studied stations and the calibrated I * values are between 5.8 and 9.2 mm/h at the four stations (Table 1).
(2) The performance and robustness of the selection method was tested by investigating the inter-annual variability of I * and its relation to the frequency of lightning strikes.We found that there is no evident trend in the estimate of I * and the performance of the model in time.I * is independent of the absolute frequency of storms with lightning at all stations, which allows us to conclude that our method to extract intense storms, which at the same time coincide well with those which are accompanied by lightning and therefore have a convective character, is robust.
(3) We analyzed the probability distributions of the main storm properties (rainfall depth R, event duration D, average intensity I a and peak 10-min intensity I p ) for the intense storm subset and found that while storm duration did not vary significantly between stations, the other properties did.Average event and peak intensity was much higher in Lugano in southern Switzerland, than in the other stations.The strength of convection measured by the convectivity index (Llasat, 2001) β was also highest in Lugano where 30% 640 of the events would be considered strongly convective.
(4) We presented non-parametric correlation coefficients between the main storm properties for the intense storm subset and compared them with the correlations for all storms to highlight the particular inter-dependencies in intense storm 645 properties.The results show that the intense warm season storm events have practically no correlation between peak intensity I p and event duration D, I p is less strongly correlated to total rainfall depth R, and average storm intensity I a is more strongly negatively correlated with duration D in 650 our intense event set than for all events, including those of a stratiform nature.The differences in the correlations between storm subsets were significant and greater than those between stations, which indicates that care must be exercised not to mix events of different origin when they are sampled 655 for multivariate analysis, e.g.copula fitting to rainfall data.
The analysis in this paper was reported for 4 representative stations in different climatological regions in Switzerland to illustrate the selection methodology.In a follow-up effort we will extend the analysis to all 62 stations of the SwissMetNet 660 network and develop regionalization approaches which will allow the estimation of spatially-distributed I * in Switzerland to be used where lightning data are not available.Schüepp and Gensler (1980).The four selected stations in this paper are marked by red triangles: GVE -Geneve-Cointrin, NAP -Napf, SCU -Scuol, LUG -Lugano.Fig. 2. A schematic to illustrate the identification of independent rainfall events (storms).Symbols: D -event duration, R -total rainfall depth, Ia -average rainfall intensity, Ip -peak 10-min intensity, ti -inter-arrival time.annual sums of close and distant lightning.The remaining figures show the distribution of all the observed lightning (regardless whether close or distant) from an annual (row 2), seasonal (row 3) and daily (row 4) perspective.Left panels show the actual number of lightning strikes, right panels show the sums of the lightning occurrence (binary per 10 min interval).WHY (CHY) stands for warm (cold) half-year.Time is in UTC, which is -2 h LT (summer) and -1 h LT (winter).

145
at the hourly resolution.The 10-min precipitation data have been analysed for their scaling-based properties (Molnar and Burlando, 2008) and short-term oscillations (Paschalis et al., 2012) and we are confident in their good quality.Air temperature is a standard measurement by a ventilated thermometer 150 at 2 m height.Four of the 62 stations are chosen as examples of different climatological regions in this paper (Figure 1): Geneve-Cointrin (GVE, 420 m a.s.l.) in the lowland region north of the Alps, Napf (NAP, 1404 m a.s.l.) in the Swiss pre-155 Alpine region in central Switzerland, Scuol (SCU, 1304 m a.s.l.) in the Graubünden region in the East of Switzerland where substantial blocking of precipitation is observed, and Lugano (LUG, 273 m a.s.l.) in the Tessin region to the south of the main Alpine divide which is significantly affected by 160 Mediterranean weather patterns and orographic forcing, especially in the summer and fall seasons (Panziera and Germann, 2010).
240 each SwissMetNet station by a ground antenna from 1987 to 2005 (19 years).After 2005 the data are obtained under the Europe-wide EUCLID network (http://www.euclid.org)and are extrapolated to the stations.To ensure homogeneity of the records we only consider the earlier period of local 245 station measurements in this paper. 465

Fig. 3 .Fig. 4 .
Fig. 3. Autocorrelograms for the selected 4 stations.Solid lines represent the autocorrelation coefficients for the individual warm half-years (abbr.WHY) with lags of 10 mins.The dots show the average of the autocorrelation functions over all years.The dashed horizontal straight line is the critical correlation for the significance level 10%.

Fig. 5 .
Fig. 5. Empirical probability density functions (PDFs) for the main storm properties: total rainfall depth R (top left), storm duration D (top right), peak intensity Ip (bottom left) and average storm intensity Ia (bottom right), differentiated by lightning, for station Geneve -Cointrin.

Fig. 6 .
Fig. 6.Estimation of the threshold peak intensity I * on the basis of sensitivity (left) and specificity (center) at station Geneve -Cointrin.The plot on the right shows the inter-relation of sensitivity and specificity at the same station.The numbers within the plot indicate the corresponding value of peak intensity in mm/h.

Fig. 7 .Fig. 8 .
Fig. 7. Threshold peak intensity I * estimated on a bi-annual basis (left) and the sensitivity corresponding to I * on a bi-annual basis (right) at the studied stations.The dashed horizontal line on the right panels denotes the chosen misclassification error α = 0.1.