The study area was Catalonia, a region of ∼ 32 000 km2 located in the north-east Iberian Peninsula and in the north-west of the Mediterranean basin (Fig. 1a). Air flows through the region are modified by local effects, mainly due to two geographical features: the Pyrenees range (∼ 3000 m a.s.l.), creating a natural border to the north, and the Mediterranean Sea to the east (Fig. 1b). The spatial variability of rainfall is high; annual precipitation values range from 300 mm in the south to 1200 mm in the Pyrenees (Martín-Vide et al., 2008). Annual average temperatures also vary greatly due to the considerable altitude range, from an average of 17 ∘C on the southern coast to about 0 ∘C along the Pyrenees. In terms of thunderstorm activity, Catalonia and its surroundings are among the regions with moderate lightning activity within western Europe, with an average lightning flash density of ∼ two flashes km-2 yr-1 (Anderson and Klugmann, 2014; Poelman et al., 2016).

The spatial distribution of lightning in Catalonia is governed by the distance to the Mediterranean and local orographic features like the Pyrenees, leading to flow deviations and, consequently, convergence zones (Campins et al., 1995; Pascual and Callado, 2002; Pineda et al., 2011). Lightning activity in Catalonia peaks during the summer months (JJA). The thunderstorm season gently starts around late April as surface heating becomes the main source of instability and convection. Lightning activity starts moving to the seashore from mid-September, where it becomes dominant in autumn. This change is related to the evolution of the land and sea average temperature, as the sea surface is warmer compared with land from mid-September to April. Regarding the daily cycle of lightning activity in Catalonia, the displayed pattern is clearly related to the solar heating cycle, with an increase ∼ 12:00 UTC and a maximum at 15:00 UTC, followed by a slow decrease (Pineda et al., 2011; Aran et al., 2015). Contrarily, the daily cycle of the Balearic Sea is more homogeneous, with minimum activity at midday and maximum activity in the evening and first night hours. The warm waters of the Mediterranean at the end of the summer and beginning of autumn, coupled with favourable thermodynamic conditions for convection, ensure sustained moisture supply, favouring long-lasting thunderstorms over the Mediterranean Sea (Pineda et al., 2009; Kotroni and Lagouvardos, 2016).

This study relies on 16 years (2005–2020) of lightning data gathered by the Lightning Location System (LLS) of the Servei Meteorològic de Catalunya (SMC) (for details, see Pineda and Montanyà, 2009). Cloud-to-ground (CG) lightning data were used to define a thunderstorm event as a period of 6 h with more than 100 CG strikes recorded in the study area. In doing so, a total of 2351 significant events were identified.

Convective indices like the CAPE (convective available potential energy) index and the convective condensation level (CCL) were retrieved from the Barcelona sounding station (N 41∘ 23′ 4.08′′ E 2∘ 7′ 3.36′′ – WMO code 08190).

Information on the cause of ignition and the date, time and coordinates of the point of ignition of the LIWs that occurred in Catalonia was retrieved from the wildfire database (WFDB), managed by the Forest Protection Agency of the autonomous government of Catalonia (Servei Prevenció Incendis Forestals, SPIF). According to the SPIF-WFDB (GENCAT, 2021), in the period of study (2005–2020), there were 561 wildfires a year in Catalonia, which burned 2732 ha of forest area each year on average. Large fires (> 100 ha) account for almost 90 % of the burned area. Humans cause most fires (i.e. negligence 40 %, deliberate fires 25 % and accidents 11 %), LIWs account for ∼ 11 % and the remaining proportion have an unknown origin (GENCAT, 2021). In addition, there have been 2250 wildfires that originated by lightning since 1986, thus an average of 66 LIWs per year. Most of them occur during summer, the months from June to September, encompassing 90 % of the LIWs. Regarding the period of the present study (2005-2020), there were 935 LIWs, averaging 59 per year. In the analysed period, these LIWs burned a total of 2150 ha of forest area (134 ha yr-1). Although the average per LIW is 2.3 ha burned, and 96 % of the LIWs burned 1 ha or less, there are also large fires caused by lightning, like the Tivissa LIW in 2014, with 490 ha burned.

NCEP/NCAR Reanalysis data (Kalnay et al., 1996) at 2.5∘ resolution were used to classify SWPs on a 6 h period basis. The gridded area covered was 30 to 70∘ N and 30∘ W to 30∘ E. Variables chosen in this study were selected to discriminate, as far as possible, convective and non-convective events. Firstly, the vertical totals index, defined by VT = T850–T500, was chosen as it accounts for the static stability (Miller, 1972). Secondly, relative humidity at 700 and at 850 hPa were selected since they consider the water content, which is necessary for assessing thunderstorms associated with potential instability (Haklander and Van Delden, 2003). We used the normalised anomalies of variables, subtracting from each grid point value the temporal average of the grid points separately for each day and each 6 h and corrected daily anomalies at each grid point by the square root of the cosine of the latitude.

A reliable lightning–wildfire association is first needed to properly describe the natural fire regime. As in Soler et al. (2021) and Pineda et al. (2022), each LIW recorded in the SPIF-WDB was crossed with the SMC-LLS lightning database to designate the most probable candidate (MPC) as the fire starter. The MPC is designated through a “proximity index” (Larjavaara et al., 2005). The proximity index (A) is a combined probability that a lightning with a temporal distance and spatial distance to ignition fire could have caused the wildfire. See Pineda et al. (2022) for details. The MPC is the lightning stroke with the highest score in A. In total, we matched 870 LIWs using A (93 %). The selection of a fire starter allows a date and time of ignition to be assigned to each LIW.

Lightning-related SWPs were obtained using an objective methodology developed by Aran et al. (2011). Firstly, a principal component analysis (PCA) was used to reduce data dimensions (Richman, 1986). The PCA is not only designed as a data reduction technique, but also as a method which ensures that only the fundamental variation modes of the data are considered for the clustering process. The PCA in S mode was applied separately to VT, with relative humidity at 700 and at 850 hPa. In this mode, variables are grid points and days are observations. The scree test was used to determine the number of components involved (Cattell, 1966), and orthogonal varimax rotation was applied to minimise the dependence between the common principal components (Richman, 1986).

Secondly, cluster analysis (CA) was used to determine the main SWPs associated with lightning activity in Catalonia. CA was applied to the score matrix formed by scores for each variable retained in the PCA analysis. The non-hierarchical K-means method (McQueen, 1967) was used as the clustering algorithm. To determine the number of clusters, the elbow method (Thorndike, 1953 cited by Zhang et al., 2016) was applied to a hierarchical Ward cluster (Ward, 1963), which gives the number of initial groups for K means (Aran et al., 2011).

Discriminant analysis (DA) was applied to the score matrix retained in the PCA, as independent variables, and a categorical dependent variable, i.e. the cluster label from CA. The DA was used to (i) obtain the discriminant functions from a stepwise selection criterion, Wilks' lambda criterion (e.g. Diab et al., 1991); (ii) validate the SWPs to determine whether CA is an effective tool for predicting the category membership; (iii) use the discriminant functions to reclassify any borderline cases if necessary (Sioutas and Flocas, 2003; Michailidou et al., 2009); and (iv) also to classify and/or predict future events in Catalonia.

i.

Short-wave trough–thermal low (SWT-TL). This is the most frequent SWP (28 % of the 2351 6 h events). It is characterised by the presence of an Iberian thermal low and a short-wave trough at 500 hPa. Thunderstorms under this synoptic pattern mostly affect the Pyrenees and the Pre-Pyrenees area, during summer months and between 12:00 and 18:00 UTC. Southerly flows provide moisture from the Mediterranean and the Pyrenees the orographic lift.

Considering the scale of dynamics and convection type, the 10 SWPs can be grouped into three broad categories. Thermal dynamics categories (SWT-TL and SWT-NF) contribute 50 % of the total SWPs, mesoscale dynamics (SWT-PC, SWT-IL and SWT-IB) 32 % and synoptic dynamics (IrL, SWT-WML, COL, DB-EF and SL-NF) the remaining 16 %.

All in all, results show that SWPs conducive to thunderstorm activity in Catalonia are mainly related to atmospheric convection caused by a trough oriented from north to south at the 500 hPa level (up to 90 % of the total, namely all categories with SWT). This general configuration leads to thunderstorm development in the region of study, often coupled with severe weather, as stated by other authors (e.g. Aran et al., 2007; Mateo et al., 2008). Additionally, regional differences are modulated by the predominant surface flow.

Once the SWPs supporting thunderstorm activity in Catalonia have been described, we now focus on the SWPs that favour lightning-caused wildfires (summarised in Table 1). Most wildfires belong to SWT-TL (51.4 %). The other SWP with a relevant number of LIWs is SWT-NF (23.7 %) and SWT-PC (13.3 %). Together, these three categories account for 88.4 % of the LIWs. Compared to the distribution for regular thunderstorm events, the frequency of the category SWT-TL is almost doubled in the case of LIWs. Moreover, in 67 % of the SWT-TL episodes, a LIW took place. At the other end, categories like COL, SWT-IB and SL-NF are irrelevant to LIW production.

Convective indices like the CAPE and the CCL derived from the Barcelona radiosonde were used to analyse seven representative episodes, showing multiple LIWs, corresponding to the three SWPs that account for the majority of the LIWs. These three SWPs, which occurred in July and August, are characterised by the arrival of a short-wave trough at 500 hPa. Under these conditions, we have seen that thunderstorms that trigger LIWs during the warmer hours of the day generally develop under relative high CAPE values. Still, a high CAPE is not exclusive of the LIW episodes but to thunderstorm development. Contrarily, the CCL index is more representative, since in five of the seven cases analysed it is above 1000 m and even around 2000 or higher in three cases (Fig. 3 shows an example).

Such conditions would also be conducive to the type of storm described in Pérez-Invernón et al. (2021), who characterised fire-producing thunderstorms by clouds with a high base and by weak updrafts between 300 and 450 hPa pressure levels. Isolated afternoon storms are more likely to generate a fire, compared to storms associated with frontal systems. In fact, small thunderstorm cells produce scattered, low-intensity precipitation, decreasing the likelihood that precipitation would either extinguish the fire or provide enough moisture to inhibit ignition (Larjavaara et al., 2005; Hall, 2008; Soler et al., 2021).

Table 1 shows that in about half of the 6 h episodes, there were multiple LIWs, sometimes more than five. The three main categories linked to the LIWs had a similar average of ∼ two LIWs per 6 h period. From an operational perspective, simultaneous fires impede the coordination of the fire extinction brigades. So, especially in these circumstances, prediction is particularly important. In addition, lightning ignitions can occur in remote locations, and therefore these complex LIW episodes may overwhelm fire brigades, resulting in longer response times and more difficult firefighting campaigns (Podur et al., 2003; Wotton and Martell, 2005; Morin et al., 2015).