Lightning climatology for the eastern Alpine region on the kilometer scale with daily resolution

Lightning affects many fields of research and aspects of everyday life. Cloud-to-ground lightning strikes may damage equipment and structures such as wind turbines [1, 17] and power lines [4], start fires [6, 22] and injure or kill people [9, 23]. It produces NO\({}_{x}\), which in turn affects the concentration of greenhouse gases [19]. While lightning during cold season occurs with an unusually strong wind field [18], it is closely connected to strong convection in warm season, which adds flash floods, large hail and damaging winds as further hazards. Reliable climatologies of lightning thus aid in assessing of all these risks and the understanding of processes associated with strong convection.

Lightning location systems (LLS) measure lightning discharges continuously in both space and time unlike any other atmospheric measurement system. Detection efficienies often exceed 90% with location accuracies well below 1 km [e.g. 20]. Lightning climatologies are usually based on a “cell-count” method, where an area and a period span each cell. However, at any given point, lightning is a rare event with typically only a few discharges per square kilometer over the whole year. [5] treated lightning as a Poisson-process to provide a theoretical limit for a 90% confidence interval of the flash density estimate with the cell-count method and found that approximately 80 flashes in a cell over the whole observation period are needed for this interval width to be \(\pm\,20\%\). Previous climatologies therefore required an aggregation to large cells, typically about 10–100 \(\text{km}^{2}\) times 1–12 months. We will use the same spatio-temporal definition of “cell” throughout the paper but carry it to a much finer resolution.

Two promising approaches have been used to achieve spatial cell sizes of 1 \(\text{km}^{2}\) and temporal sizes of 1 month or shorter. The first one [2, 11] uses the location uncertainty instead of the precise locations of flashes, which corresponds to performing a kernel density estimation around the location of lightning discharge under an assumed independence between the discharges. The second one [27] exploits the similarity in seasonal, regional and altitudinal characteristics between cells using the statistical method of generalized additive models [GAMs, 34] to achieve even smaller cell sizes of 1 \(\text{km}^{2}\) times 1 day and additionally extract the functional dependence of lightning on these underlying characteristics. Their study was intended as a proof of concept for the suitability of GAMs to achieve high-resolution lightning climatologies and thus limited to a small region (the state of Carinthia in Austria) and basic effects of season, elevation and region.

We will use the GAM method to compute a lightning climatology for the entire eastern Alpine region and the surrounding lowlands and include a more comprehensive set of shared characteristics between cells. To our knowledge it is the first lightning climatology at a cell resolution of 1 \(\text{km}^{2}\) times 1 day for such a large region. Other climatologies for (parts of) this region [e.g. 7, 31, 33] have a considerably coarser—either in space and/or in time—cell resolution.

The climatology of lightning days is based on lightning detection data from the Austrian Lightning Detection & Information System [ALDIS, 26], complemented by data from the digital elevation model TanDEM-X [24] to account for topographic effects.

Each of the 45.2 million spatio-temporal cells—1 square kilometer hexagons spatially times 1 day temporally for each day from April through September—is one data point for our statistical climatology. Each cell contains the number of observed thunderstorm days over the 11 years for this specific day of the year and hexagonal area. These numbers of observed thunderstorm days \(\{0,1,2,\dots,11\}\) can be seen as realizations of Bernoulli’s urn problem with replacement. For each cell, 11 trials are conducted, where each trial has two possible outcomes: lightning and no lightning, respectively.

The outcome of these trials follows a binomial distribution, which is determined by a probability \(\pi\) that lightning occurs. In order to estimate the probability \(\pi\) of the underlying process, we utilize a GAM [34] that incoporates data not only for single data points but leverage information of data points with similar settings, e.g. same region, similar time of the year, similar topographic conditions, which are represented by a choice of explanatory variables. For these explanatory variables non-linear transformations are learned from the data. GAMs have proven their abilities for lightning climatologies in complex terrains [27]. The GAM used for this study is set up as follows,

On the left hand side is the logit-transformed probability \(\pi\). The logit maps \(\pi\) from the probability scale \(\left]0,1\right[\) to a real number \(\left]-\infty,\infty\right[\). On the right hand side is the additive predictor that combines the intercept \(\beta_{0}\) of the regression and multiple, potentially non-linear terms \(f_{\star}\).

With the GAM all these functions \(f_{\star}\) are modelled using regression splines such as cubic regression splines and thin plate regression splines. The spline bases for functions with two (\(f_{2}\)) or three (\(f_{3}\)) covariates are set up using the tensor product of the univariate spline bases. For the technical details on splines within GAMs the reader is referred to [34].

This setup leads to approximately 100 regression coefficients of the GAM. These coefficients are estimated via penalized maximum likelihood. Here the amount of smoothing of the potentially non-linear functions \(f_{\star}\) is determined by generalized cross-validation as implemented in the mgcv extension for R. Estimating such a flexible regression model for the large data set on hand with 45 192 046 data points is feasible with a fitting algorithm for giga data [35, implemented in the mgcv extension for R].

The resulting probabilities \(\pi\) can be converted into expected number of thunderstorm days within \(k\) years at a specific location. Since one can assume that—for a specific day of the year and location—the occurrence of lightning from one to the next year is independent from each other, the expected number of thunderstorm days within \(k\) can be computed by the product,

Lightning in the eastern Alps is rare locally. Only 5.47% of the 45.2 million spatio-temporal cells of 1 \(\text{km}^{2}\) times 1 day in the 182 days from months April through September had lightning occurrence in at least one of the eleven years 2010–2020. Fig. 2 shows that the maximum probability for lightning to strike within a grid box of 1 square kilometer on a particular day barely exceeds 2%. The maps for day 15 of each month demonstrate that the probability of lightning waxes from spring through the end of July and then wanes into autumn. In spring (May 15, top) the likelihood for lightning strikes north of the Alps is regionally uniform below about 0.5%. The probability along the northern Alpine rim slightly exceeds these values. In the Alps, valleys and mountain ranges stand out (cf. Fig. 1, top) with probabilities in valleys considerably higher than over mountain ranges, which are still mostly snow-covered at this time of the year. By mid-June these differences blur and valleys are almost no longer distinguishable from mountain ranges. By now, lightning in the higher terrain north of the Alps strikes more readily than in the surrounding flat lands. Differences south of the Alps are even starker. The flat land around location I (Udine) in Fig. 1 has the highest probability in the whole study region. This is exceptional as lower terrain everywhere else has lower probabilities than higher terrain in its vicinity. By mid-July lightning activity favors higher terrain. Lightning probability again clearly differs between valleys and moutain ranges, however now reversed to mid-May. The probability in valleys is considerably lower than over the mountains. Overall, the strongest activity has shifted to the south of the Alpine crest and probabilities even exceed the ones in the former hotspot in the flat land in the vicinity of location I (Udine). One month later in mid-August the overall pattern is the same but probabilities are lower—with the exception of location I (Udine) where the probabilities remain relatively high even into mid-September, by which time lightning activity is low everywhere, most pronouncedly so north of the Alps.

Contrary to the results from the GAM method, results from the cell-count method in Fig. 2b are noisy and lack details despite averaging the results over an 11-day window centered on that particular day. Grid boxes with probabilities of more than 10% occur next to boxes with 0%, which necessitated a different color scale in the figure. Only the major features of lower lightning probabilities/frequencies over the high mountains early in the lightning season, the shift into the high mountains and south of the Alpine crest, and the hotspot near location I (Udine, Fig. 1) are visible, admittedly more easily so when one has previously seen the results from the GAM method. Sample sizes in the grid boxes are simply far too small to obtain reliable frequencies of lightning occurrence, which is a rare event of only about 1% per day in a 1 square-kilometer grid box. The uncertainty of the cell-count method can be determined using the modified Wilson method [3] to estimate the width of the 95% confidence interval for a binomial distribution (lightning yes/no) for a sample size of 11—the number of years in our data set. The width is a staggering 27% for an event with a probability of 1%. Even increasing the sample size by a factor of 11 to \(11\times 11=121\) using a time-window of 11 days instead of 1 day, as is done in Fig. 2b, still yields a width of 4.9 percentage points. Therefore finding a lot of noise in the results from cell-count method comes as no surprise.

The power of the GAM method and its ability to produce detailed and smooth results even at such high spatial and temporal resolution lies in its harnessing of information from grid boxes that share seasonal, topographical and regional characteristics. Extracting the functional form of these characteristics in turn leads to increased understanding of the contributors to the overall spatio-temporal distribution of the rare event “lightning occurrence”. Fig. 3 shows the effects of each of the additive terms in Eq. 1 as they contribute to \(\textrm{logit}(\pi)\). Seasonally (Fig. 3a), the probability of lightning for the whole region increases rapidly in spring and slowly during mid-summer, then tapers off in autumn.

The elevation effect (Fig. 3b) varies seasonally. Note that the effect is logarithmic (cf. Eq. 1) but the ordinate axis has been labeled in meters to ease interpretation. This effect shows a reduced likelihood for lightning at the highest elevations early in the season, which is reversed in the second half of June. The sign reverses last at the very highest elevations, where snow melts last. Snow-covered ground greatly reduces sensible heat flux and thus the chances for convection to occur. The seasonal elevation effect also changes sign in elevations typical of intra-alpine valleys. It is increased in spring, when the valleys are already free of snow, and reduced in summer and early autumn. The elevation effect is positive from mid-May onwards for the lowest elevations, found mostly at the southern edge of the study region.

The regional effect in Fig. 3c accounts for adjustments needed to add to the effects of season and elevation. At first glance, the patterns do not seem to vary seasonally but a closer look reveals a shift of enhanced and reduced regions from spring into summer and fall. In spring the lightning probability in some parts of the lowlands and rims on either side of the Alps and the high mountains in southeastern Switzerland is enhanced. By mid-July lightning activity along the whole southern side of the Alps is enhanced as well as in the westernmost and easternmost parts of the lowlands north of the Alps. By mid-September, at the end of convective season in most of the region, the northwestern rim and lowlands and especially the southeastern rim and lowlands have enhanced lightning probabilities.

Protruding topography such as peaks increases the likelihood of lightning. Again, the effect is logarithmic but elevation difference in Fig. 3d is given in units of meters. Here the difference between the maximum elevation at 90 m horizontal resolution to the average elevation in the one-square-kilometer grid box serves as proxy for how jagged the terrain is. Overall, where that difference exceeds about 100 m, lightning likelihood increases somewhat, which is in alignment with findings from lightning research [e.g. 7, 11, 21].

Time series of lightning probability at specific locations in the eastern Alps in Fig. 4 exemplify the sum of the seasonally varying regional and elevation effects and the roughness effect from Fig. 3 and fill in the times between the regional maps of Fig. 2. These locations are shown in Fig. 1. They represent the north side and the northern rim of the Alps (A–E) both in flatlands and high terrain. The two exposed peaks Saenits (A) and Gaisberg (C) are included since measurements from the instrumented towers there have featured prominently in numerous papers. Locations F and G straddle the Alpine crest, and H–J exemplify the southern rim of the Alps. Locations I and H in the Alpine lightning hotspot represent flatland and adjacent mountain range, respectively.

The lightning season starts about three weeks later at the highest mountains, exemplified by locations F (Central Alps N) and G (Central Alps S), than at other locations (cf. Fig. 1 and Table 1 for their coordinates). The difference in probability between the two locations F (Central Alps N) and G (Central Alps S) stems from their difference in elevation. Similar peak probabilities are reached by locations north of the Alps at A (Saentis) and D in the Bohemian Forest, although they are considerably lower. Their seasonal cycles are shifted by about one week, with D (Bohemian Forest) starting out one week earlier and A (Saentis) lasting one week longer. Lightning season at location C (Gaisberg), just south of D (Bohemian Forest) at the northern rim of the Alps starts about at the same time but peak probabilities are about one third lower. Both A (Saentis) and C (Gaisberg) are locations with a long history of direct lightning measurements from instrumented towers. Although both are at the northern rim of the Alps, seasonally varying regional differences (Fig. 3c) and an elevation difference of 1.1 km result in different shapes and peaks of their seasonal lightning probability distributions. Flat terrain north of the Alps, exemplified by locations B (Munich) and E (North of Vienna), experiences lightning less frequently than the mountainous counterparts A (Saentis), C (Gaisberg) and D (Bohemian Forest). They also reach a first peak in probability earlier and then increase slightly to a second peak at the end of July. Locations in the flat lands south of the Alpine crest (E – North of Vienna – and I – Udine) show a similar bimodal distribution. Location I (Udine), which is situated in the flatlands south of the Alps has the most abrupt increase of lightning probability of all 10 locations shown. Despite its low elevation, its peak probabilities exceed the ones of the highest mountains near the Alpine crest. It is located in the hotspot so clearly visible in the regional probability maps in Fig. 2. However, probabilities at location H (Range NW of Udine) at the first mountain range north of I (Udine) are even higher.

Lightning location systems (LLS) are unique among atmospheric measurement systems in that they provide continuous measurements in both time and space—two-dimensional or three-dimensional depending on LLS type. Since they measure events that are rare and infrequent at any given location, translating these measurements into high-resolution climatologies is difficult. Most climatologies so far have used the cell-count method of counting how many flashes occurred within a particular time-space cell, which necessitates making the spatial and/or temporal dimensions of such a cell large in order to achieve a sufficient signal to noise ratio. Frequently cell sizes of 10–100 square kilometers by 1–12 months have been used [e.g. 7, 20, 25]. [2, 11, 27] demonstrated how the cell-count method is unsuitable for smaller cell sizes corroborating the theoretical considerations in [5], who used the Poisson distribution to derive the minimum number of flashes in a cell required for reaching a specified accuracy. The Poisson distribution is the limiting case of the binomial distribution, which describes the binary event of the occurrence of lightning (yes/no). We therefore use the binomial distribution to compute the 95% confidence interval for a lightning probability of 1%. It has a width of staggering 27% for a sample size of 11 for a particular day in our 11-year data set. Increasing the time dimension of a cell from 1 day to averaging over 11 days to obtain a sample size of 121 still gives an uncertainty of 4.9%, five times larger than what we need to detect. Fig. 2b illustrates the high noise level the cell-count method produces for high-resolution cells of 1 \(\text{km}^{2}\) times 1 day. Fig. 2a, on the other hand, produced with the generalized additive model (GAM) method, is devoid of noise and yet provides intricate spatial details for any given day in the lightning season (of which the map shows five). Similarly spatially detailed climatological maps can be achieved using a bivariate Gaussian error distribution for each lightning discharge instead of a precise location [2, 11]. The advantage of the GAM method is its ability to harness regional, elevational and seasonal characteristics that cells share that need not be immediately “next” to another and thus make a high combined spatio-temporal resolution possible. To our knowledge, this paper presents the first daily-resolved kilometer-scale lightning climatology over a large region. GAMs have another advantage: the resulting functional forms of characteristics shared among cells. Fig. 3 yields insights into the processes contributing to the final climatological result.

The purely seasonal effect (Fig. 3a) shows a rapid increase of lightning probability in late spring but a more gradual tapering off in late summer and early fall, which was previously also found for a smaller region in the eastern Alps [27]. The peak, however, is less sinusoidal but more plateau-like, probably due to the larger region where different subdomains peak at different times (cf. Fig. 4).

Lightning probability also varies with elevation (Fig. 3b). [30] and [27] found an overall increase with elevation. This is due to thermally and mechanically forced lifting along higher topography [e.g. 12]. Daytime winds from the plains to the mountains [“Alpine pumping” 15, 16] lead to updrafts and additional amounts of moisture and heat that provide lifting and destabilization of the atmosphere conducive to the formation of convection [e.g. 8]. Valley and slope winds have similar effects on smaller scales. Mountains can also modify larger-scale flow to create favorable conditions for convection [e.g. 10]. Since both thermally and mechanically driven flows vary seasonally it is highly advantageous that the GAM method makes it possible to implement a seasonally varying elevation effect. It finds an increased probability in spring at elevations in lowlands and intra-Alpine valleys compared to higher elevations (above approx. 1200 m msl), which are still (partly) snow covered. This pattern reverses later and high elevations have a higher probability when the snow has melted. The most pronounced increase on the southern side of the Alps (vicinity of location I – Udine) with maxima in June and late August/early September is found at the very lowest elevations.

This regional hotspot was already known from earlier cell-count climatolgies [e.g. 7, 25, 31] and is possibly linked to the propagation and moisture transport of the daytime seabreeze front from the Adriatic Sea [14, 32]. With GAMs we could additionally identify how regional differences vary with season—in extension to [27]—when different elevations are already accounted for.

The approach of using similar characteristics to compute high-resolution yet non-noisy lightning climatologies with GAMs pioneered by [27] can be expanded by including further characteristics beyond region, elevation and time of year. Snow cover could be explicitly included since it suppresses convection. In our analysis it is implicitly contained in the seasonally varying elevation effect where lightning season at the highest elevations with the longest snow cover duration has a delayed start. Further characteristics of earth’s surface such as slope angle and exposure, soil moisture, amount and type of vegetation cover, and land/water could be included as well as the presence of tall structures, which can trigger lightning [21]. The GAM approach can also use further additive terms with output from numerical weather prediction models in order to provide forecasts of the probability of thunderstorm occurrence [28].

Our GAM models the probability whether lightning occurs on a given specific day of the season and location. These probabilities can be intuitively converted to expected number of thunderstorm days as shown by the alternative scales in Figs. 2 and 4. Flash densities could also be estimated with a GAM, though this would require a different set up: The ALDIS data have to be aggregated to counts of cloud-to-ground discharges in each cell. These counts would serve as target variable in the GAM and thus a parametric count data distribution has to be applied for the estimation. As several parametric count data distributions have been proposed, e.g. Poisson [27] or negative binomial [29], further reseach is needed to find the distribution that fits best the data on hand. Another challenge is handling the data size as the count data distributions require a different optimization approach for the class of distributional regression models [13].

The results in this study were achieved using R, a software environment for statistical computing and graphics. The add-on packages mgcv were used for building the statistical model.