On the Role of Continuing Currents in Lightning‐Induced Fire Ignition

Lightning flashes are an important source of wildfires worldwide, contributing to the emission of trace gases to the atmosphere. Based on experiments and field observations, continuing currents in lightning have since a long time been proposed to play a significant role in the ignition of wildfires. However, simultaneous detections of optical and radio signals from fire‐igniting lightning confirming the role of continuing currents in igniting wildfires are rare. In this work, we first analyze the optical signal of the lightning‐ignited wildfires reported by the Geostationary Lightning Mapper over the Contiguous United States (CONUS) during the summer of 2018, and we then analyze the optical and the Extremely Low Frequency signal of a confirmed fire‐igniting lightning flash in the Swiss Alps. Despite data uncertainties, we found that the probability of ignition of a lightning flash with Continuing Current (CC) lasting more than 10 ms is higher than that of cloud‐to‐ground lightning in CONUS. Finally, we confirm the existence of a long CC (lasting about 400 ms) associated with a long‐lasting optical signal (lasting between 2 and 4 s) of a video‐recorded fire‐igniting lightning flash.

The Geostationary Lightning Mapper (GLM) can provide continuous measurements of the optical signal emitted by lightning in the Americas, as well as in some parts of the Pacific Ocean and the Atlantic Ocean.Fairman and Bitzer (2022) have recently developed a new method to classify lightning flashes with CC lasting more than 10 ms derived from the 777 nm optical characteristics of flashes reported by GLM.They reported that in 2018, 13.3% of all flashes detected by GLM in CONUS with no degraded flags (see below) contained continuing currents.Note that Degraded flashes are those that are composed of events that occur out of order temporally, consist of more than 101 groups, or exceed a duration of 2,998 ms (Fairman & Bitzer, 2022).Fairman and Bitzer (2022) reported that during the summer months and over land, the ratio of lightning flashes with continuing currents to total lightning flashes reaches its minimum at nearly 10%.Additionally, Pérez-Invernón et al. (2023) searched for lightning candidates with CC for wildfires in the Contiguous United States (CONUS) by using a spatio-temporal threshold, reporting that more than 10% of lightning-induced wildfires can be associated with lightning with CC.Apart from GLM, the Lightning Mapping Imager aboard the Feng-Yun-4 satellite (FY-4) since 2018 (Yang et al., 2017), and the Meteosat Third Generation geostationary satellites of the EUropean organization for the exploitation of METeorological SATellites equipped with a lightning imager launched on 13 December 2022 (Stuhlmann et al., 2005), will also provide continuous geostationary observations of lightning that can serve to develop a global climatology of lightning flashes with CC and to monitor their relationships with fire ignitions.
In this work, we investigate the role of CC in the ignition of lightning-induced wildfires by using lightning data from GLM, fire data provided by the U.S. Department of Agriculture (Short, 2021;Wright et al., 2011) during 2018, and vegetation data from the Moderate Resolution Imaging Spectroradiometer (MODIS) imagery and Forest Inventory and Analysis (FIA) plot data (Ruefenacht et al., 2008).We extend the analysis of Pérez-Invernón et al. (2023) by comparing the lightning candidates with and without CC and by introducing an estimation of the total number of cloud-to-ground lightning candidates based on the climatology provided by Medici et al. (2017),

Lightning Data
We used lightning data over CONUS between 15 May 2018 and 31 August 2018 provided by the GLM, aboard the Geostationary Operational Environmental Satellite-16 (GOES-16) (Goodman et al., 2013) since 2017 and aboard the GOES-17 since 2018.The GLM is a Charge-Coupled Device imager with narrow spectral band filters centered on the atomic oxygen line 777.4 nm that are usually associated with lightning (Kieu et al., 2021;Orville, 1968;Pérez-Invernón, Gordillo-Vázquez, et al., 2022).The GLM can provide continuous measurements of the optical signal emitted by lightning in the Americas, some parts of the Pacific Ocean and the Atlantic Ocean.The GLM pixel sizes exhibit a range of approximately 8 km across the majority of its field of view (FOV), which gradually increases to around 14 km at its edges (Goodman et al., 2013).The clustering algorithm groups events (pixels illuminated during 2 ms) into groups and considers groups of a single flash if they fall within 330 ms and less than 16.5 km of each other, and using the centroid of the events coordinates as the location of the groups (Goodman et al., 2013;Mach et al., 2007;Peterson, 2019).Groups detected by GLM can be associated with the strokes detected by ground-based networks.The algorithm that cluster groups into flashes estimates the coordinates of each flash as the centroid of the groups weighted by their energy.The optical signature of lightning measured by GLM is usually scattered and reflected by clouds above the lightning channel (Montanyà et al., 2021;Pérez-Invernón, Gordillo-Vázquez, et al., 2022), causing the illumination of a large area of clouds.Flashes with CC observed by GLM and confirmed with ELF sensors (Montanyà et al., 2021(Montanyà et al., , 2022) ) showed that the illuminated area can extend further than 0.5° from the location of the grounding-point of the flash.Nevertheless, the activity of the groups is more frequent and intense near the point where the lightning strikes the ground (Montanyà et al., 2021;Pérez-Invernón, Gordillo-Vázquez, et al., 2022), suggesting that taking the position of the flash as the centroid of the groups weighted by their energy is appropriate to search for lightning candidates for wildfires when dealing with lightning with CC.On the other hand, using the position of individual groups to determine the best candidate group has the disadvantage that the vast majority of pulses detected by GLM corresponds to Intra-Cloud (IC) pulses.
The GLM product does not provide information about the type of the lightning flashes, i.e, IC or Cloud to Ground (CG).However, Ringhausen et al. (2021) have demonstrated that a classification method can be used to distinguish between IC and CG flashes based on the optical characteristics of lightning flashes detected by GLM.In addition, Peterson and Mach (2022) have developed a method to acquire height information from GLM data.We used the method developed by Fairman and Bitzer (2022) to classify lightning flashes with CC lasting more than 10 ms from GLM measurements based on the characteristics of the measured optical signals.For each flash we identified the maximum sample of time-contiguous groups and estimated the maximum distance between two groups in the sample, the maximum group footprint, the maximum group optical energy, the maximum number of time-contiguous groups, the median group optical energy and the total group optical energy.Then, we applied the logistic regression model of Fairman and Bitzer (2022, Table 1) by assuming that flashes with CC had probabilities at or above a 0.33 threshold, which implies a probability of detection of lightning with CC of 78% and a false alarm rate of 6%.Furthermore, we used lightning data from the Earth Networks Global Lightning Network (ENGLN), the lightning detection network (LINET) of Nowcast GmbH and the ELF sensor located at the Eskdalemuir Observatory in Scotland (Musur & Beggan, 2019) to analyze a single lightning-ignited wildfire in the Swiss Alps.The ENGLN is a global network composed of VLF/LF sensors that provide the position, time of occurrence, polarity, and peak current of lightning strokes with, at the time of these events, a median location uncertainty of 215 m as observed in Florida U.S.A. (Zhu et al., 2015(Zhu et al., , 2022)).In turn, LINET is a lightning network with VLF and LF sensors located over four continents (including 65 sensors in Europe) monitoring the occurrence of lightning strokes with an average location uncertainty of 150 m (Betz et al., 2009).

Fire Data
Fire data, including the cause of the wildfires, are reported by the U.S. Department of Agriculture (Short, 2021;Wright et al., 2011).The coordinates of the wildfires are reported with a minimum spatial accuracy of 1-square mile (i.e., 2.6 km −2 ), together with the date and time of detection.As in Pérez-Invernón et al. ( 2023), we have extracted 5,860 LIW taking place over CONUS between 1 June 2018 and 31 August 2018.
The investigated lightning-ignited fire in the Alps was extracted from the forest fire database of Switzerland "Swissfire" (Moris et al., 2020;Pezzatti et al., 2019).In particular, this event took place on 21 July 2021 in the Valais (Switzerland) 615,307 m E and 126,417 m N CH 1,903 coordinates (7.64°E and 46.29°N).The flash was recorded at a frequency rate of 30 frames per second by the Complementary Metal Oxide Semiconductor (CMOS) camera of a Samsung Galaxy S8 Plus smartphone between 21 and 22 hr (UTC-0).According to the video (Berckum et al., 2023), the fire-igniting lightning flash was followed by a second flash that took place less than a second after its onset.

Vegetation Data and Ecological Regions
The Lightning Ignition Efficiency (LIE) is defined as the number of wildfires ignited per lightning (Pineda et al., 2022;Podur et al., 2003).The spatial distribution of the ratio of lightning with CC to total lightning is not homogeneous across CONUS (Fairman & Bitzer, 2022).Therefore, comparing the LIE of total flashes with the LIE of flashes with CC at a continental scale may be misleading.We calculated the LIE for all flashes, CG flashes, and flashes with CC separately over different forest types and ecological regions to analyze the role of continuing currents in the ignition of wildfires.Distinguishing between forest types may help to reduce the effect of the type of vegetation in the probability of ignition by lightning, while distinguishing between ecological regions may serve us to partially remove the contribution of climate (Pérez-Invernón, Huntrieser, & Moris, 2022).Therefore, we used a 250 m resolution map of the United States forest types (Ruefenacht et al., 2008) and a map of ecological regions to investigate the LIE in the main forest types of CONUS (Podur et al., 2003).The forest type map was generated from MODIS imagery and FIA plot data, and classifies the forest area into discrete classes based on the dominant tree species (Ruefenacht et al., 2008).In turn, the map of ecological regions follows a holistic approach based on the major components of ecosystems, including air, water, land and biota (Commission for Environmental Cooperation, 1997(Revised 2006)).We assigned a forest type and an ecological region to each fire using their coordinates.To accurately account for the minimal spatial uncertainty introduced by GLM pixels (∼8 km and increasing to ∼14 km at the edge of the full-disk view (Marchand et al., 2019)), we assigned to each flash the predominant forest type and ecological region within a 4 km radius from its coordinates.

Search for GLM Lightning Candidates
Establishing a one-by-one relationship between LIW and lightning flashes is complex because the detection efficiency of GLM in CONUS ranges between 43% and 90% (Marchand et al., 2019) depending on the solar zenith angle and the charge structure of the thunderstorm.However, there is no indication that the variability of the DE of GLM could affect the total number of lightning with and without continuing currents differently (Fairman & Bitzer, 2022).In addition, the official GLM product does not provide information about the type of lightning, that is, IC or CG, and the position and time of the flashes, the groups, and the fires can include errors that can influence the matching between them.For these reasons, we used the proximity index A proposed by Larjavaara et al. (2005) to search for the most probable lightning candidates and to find the parent thunderstorm of each fire, where D is the distance between the reported fire location and the lightning flash or group, and T is the time between fire ignition and detection, also known as holdover time (Moris et al., 2023;Wotton & Martell, 2005).
Parameters (T max ) and (D max ) correspond to the maximum holdover time and distance between a fire and a lightning discharge to consider the latter as the potential cause of ignition.A distance of D max = 10 km is often applied in the literature to account for possible large location errors in fire and lightning data reported by lightning location systems (Larjavaara et al., 2005;Moris et al., 2020;Pérez-Invernón et al., 2021;Schultz et al., 2019;Wotton & Martell, 2005).However, a threshold D max = 10 km could be considered too small for flashes coordinates reported by GLM, as the uncertainty of the position of flashes reported by GLM is larger than in the case of lightning location systems.Methods to match space-based optical signal of lightning and lightning location systems usually set a maximum distance of 25 km for the flash level (Bitzer et al., 2016;Pérez-Invernón, Huntrieser, Erbertseder, et al., 2022;Rudlosky et al., 2017).Therefore, we use D max values ranging between 10 and 25 km in order to estimate the sensitivity of our results.Using T max = 14 days is a conservative approach that allows us to include a representative sample of LIW by discarding fires with long (>14 days) holdover durations that might have been erroneously labeled as natural fires (Schultz et al., 2019).Thus, we did not include in subsequent analysis fires for which no lightning discharges were detected within the proposed spatio-temporal windows.
As an alternative method to assess the robustness of the results, we have searched for lightning candidates by using the groups detected by GLM.For each fire, we calculated the A-index for all groups within a 10 km radius and up to 14 days before the detection.Subsequently, we selected lightning flashes as candidates if they included any of the candidate groups.This approach may include numerous groups generated by IC leader activity, but it allows us to evaluate the limitations of our study.
We explored the role of continuing currents in lightning-induced fire ignitions by using the proximity index A of all the possible flash candidates.The GLM detects total lightning (CG and IC), while only CG lightning can ignite a fire.Therefore, we cannot assume that the lightning candidate with maximum proximity index is the only possible candidate.Instead, we search for all the possible candidates for each fire with a proximity index A greater than 0. We then calculated, for each single fire, the distribution of values of the proximity index A of all the flashes candidates.We considered that the fire may have been ignited by a flash with CC if there is at least one CC flash with a proximity index above a given percentile of the index A values.In particular, we used different minimum percentiles ranging between the 0th and the 100th percentile to test the robustness of the results.

Lightning Candidate for the Recorded Lightning-Ignited Wildfire
A fire-igniting lightning flash has been recorded on 21 July 2021 between 21 and 22 hr (UTC-0) in the Valais (Switzerland) by a camera operating at 30 frames per second.Based on this video, we searched for the candidate stroke manually by using as reference the two flashes recorded in the video.The exact timing of the fire-igniting lightning stroke is needed in order to investigate its ELF signature.We used lightning measurements from lightning location networks to establish the timing of the stroke.We inspected all the ENGLN and LINET strokes reported between 21 and 22 hr (UTC-0) within 10 km distance around the position of the fire and selected the pair of CG strokes with the same time between strokes as the two flashes recorded in the video.We calculated the Charge Moment Change (CMC) of the discharge from the ELF signal by following Füllekrug (2000), Füllekrug and Constable (2000), and Füllekrug et al. (2006).With this purpose, we first calculate the transfer function as where l is the lightning flash length, assumed to be 7 km.ω is the frequency of the radio wave, ϑ is the angular distance between the flash and the sensor, h 1 (ω) is the conduction boundary (50 km), and   1  (cos ) are the associated Legendre polynomials of degree n and order m = 1 on a spheroidal earth with radius a = 6,371 km.The complex modal frequency ω n is calculated according to Füllekrug (2000) and by considering the ionospheric reflection height at 100 km.The lightning current I(ω) is then calculated as where  ( ) is the recorded magnetic field.Finally, the CMC is calculated by integration of I(ω) and assuming 7 km flash length.

Lightning Ignition Efficiency Estimate
We calculate the LIE of total lightning, CG lightning and lightning with CC in different forest types and ecological regions.The LIE of total lightning is estimated as the ratio of the total number of lightning-induced wildfires to total flashes.The LIE of CG lightning in each forest type and ecological region is estimated from total lightning by using the mapped climatological IC fraction reported by Medici et al. (2017, Figure 6) over CONUS during more than 2 decades.We calculated the average IC fraction and the standard deviation in each forest type and ecological region.We used the total number of flashes reported by GLM and the average IC fraction ± the standard deviation of the IC fraction to calculate the total number of CG lightning and an uncertainty measure.Finally, the LIE of lightning with CC is calculated by assuming that a wildfire was ignited by a flash with CC only if there is at least one lightning candidate with CC with a proximity index A above the 95th or the 99th percentile by using a holdover of T max = 14 days and a spatial matching constraint D max = 10 km.We selected the 95th and the 99th percentiles as thresholds for identifying wildfires ignited by flashes with CC to reduce the possibility that flashes with CC were within 10-25 km and 14 days from the fire by pure chance when the number of lightning candidates is high (Section 3.1.1).We did not include in our analysis forest types and ecological regions where the total number of fires is too low to be significant.In particular, we limited our analysis to forest types and ecological regions containing at least 2% of the reported total number of natural fires.

Lightning Candidates and Continuing Currents
We found a total number of 1,033,318 and 6,085,706 flashes candidates for 5,860 fires labeled as LIW by using D max values of 10 and 25 km, respectively, and T max = 14 days.In particular, we found at least one lightning candidate for 5,585 (95%) and 5,731 (98%) fires by using D max values of 10 and 25 km and the flashes coordinates, respectively.In turn, we found 10,683,999 groups candidates by using D max values of 10 km and T max = 14 days.We found at least one group candidate for 5,634 (96%) fires by using D max = 10 km and the groups coordinates.
Figure 1 shows the holdover time (a, b) and the distance (c) between each fire and the most probable (maximum A index) lightning flash candidate by using D max = 10 km.As reported by Pérez-Invernón, Huntrieser, and Moris (2022) and Schultz et al. (2019) in CONUS, we obtained that most wildfires (62%) are detected within less than 24 hr after the most probable lightning candidate.Furthermore, we obtained a significant number of lightning candidates (40%) more than 2 km away from the reported location of the fire.This proportion is considerably reduced to 26% when we select the distance between wildfires and flashes as the distance to the closest group detected by GLM (Figure 1d).
Table 1 shows the total number of LIW with at least one flash with CC candidate above a given percentile by using the coordinates of flashes ("Flash-based search") and the coordinates of groups ("Group-based search").A maximum holdover of 14 days has been used, while the maximum distance imposed has been 10 km in both cases.Alternatively, a maximum distance of 25 km has also been used for the "Flash-based approach" case, since the position of the flashes has greater uncertainty than that of the groups.The minimum 0th percentile indicates that, for a fire, there is at least one flash with CC candidate within the sample of all candidate flashes, while the maximum 100th percentile indicates that the flash candidate with the maximum A index value is actually a flash with CC.On average, 185 lightning flashes per fire are assigned as possible candidates by using D max = 10 km, T max = 14 days, and the flashes coordinates, while this quantity descends to 10 and 2.5 when only lightning flashes with a proximity index A above the 95th and the 99th percentiles are selected, respectively.During the summer season in CONUS, the climatological ratio of CC flashes to total lightning is typically less than 10%.By setting the 95th and the 99th percentiles as the thresholds for identifying wildfires ignited by flashes with CC, we can ensure that the average number of flash candidates with CC per fire would be less than 1 if there is no correlation between CC and wildfires.These thresholds are chosen based on the assumption that a low ratio of CC flashes to total lightning reduces the likelihood of misclassifying wildfires as having been ignited by flashes with CC.The percentage of LIW produced by flashes with CC ranges between 12% for the 100th percentile and D max = 10 km and 99% for the 0th percentile and D max = 25 km by using the flashes coordinates.Finally, the percentage of LIW produced by flashes with CC ranges between 14% and 86% by using the groups coordinates.

Lightning Ignition Efficiency
Tables 2 and 3 show the total number of LIW, LIW associated with lightning with CC lasting more than 10 ms, total flashes, flashes with CC and LIE in different forest types and ecological regions by following the "Flash-based search," respectively.Wildfires are classified as ignited by a CC flash if there is at least one CC flash candidate with an A index about the 95th and the 99th percentiles, respectively.Lightning-induced wildfires are heterogeneously distributed over different forest types and ecological regions.Most of them took place over non-forest areas and pinyon juniper woodland, while the ecological regions with the highest occurrence of LIW were the Western Cordillera and the Cold Deserts.
The percentage of lightning-induced wildfires produced by flashes with CC are significantly higher than the ratio of flashes with CC to total flashes in all the analyzed forest types and ecological regions.Second, the total number of fires caused by lightning with CC is highly influenced by the percentile (e.g., 95th and 99th) of the A index value chosen as threshold.As a result, we encountered significant uncertainty in estimating the LIE of flashes with CC.Third, the uncertainty of the LIE of CG flashes is a consequence of the high variability (standard deviation) of the CG fraction over particular forest types and ecological regions.Fourth, total LIE, CG LIE, and LIE of flashes with CC show a relatively high variability over different forest types and ecological regions, while the variation in the percentage of flashes with CC is lower.Finally, when data from all the regions and forest types of the CONUS are pooled together, the global LIEs of lightning with CC using percentiles (95th and 99th) are greater than the average LIE of CG lightning, although within the interval of possible CG LIE defined by the standard deviation of the CG fraction.
Figure 2 shows a comparison between the LIE of CG flashes and the LIE of CC flashes in different forest types and ecological regions.We included error bars to take into consideration the effect of the standard deviation of the

Note.
LIW is the total number of Lightning-Ignited Wildfires (LIW).LIW (CC) 99th and LIW (CC) 95th are the total number of LIW produced by lightning with CC according to the two criteria described in the text.The CG fraction is given together with the standard deviation over each forest type.The total, CG and CC lightning-ignition efficiency (LIE) for each forest type was calculated as the ratio of total Lightning-Ignited Wildfires (LIW) to total lightning, CG flashes and flashes with CC over the same ecological region, respectively.Only forest types with ≥2% of the total lightning-ignited wildfires are included here.

Note.
LIW is the total number of Lightning-Ignited Wildfires (LIW).LIW (CC) 99th and LIW (CC) 95th are the total number of LIW produced by lightning with CC according to the two criteria described in the text.The CG fraction is given together with the standard deviation over each ecological region.The total, CG and CC lightning-ignition efficiency (LIE) for each ecological region was calculated as the ratio of total Lightning-Ignited Wildfires (LIW) to total lightning, CG flashes and flashes with CC over the same ecological region, respectively.Only ecological regions with ≥2% of the total lightning-ignited wildfires are included here.Table 3 LIE in Ecological Regions of CONUS LIE of CG flashes.When using the 95th percentile threshold, the LIE of CC is higher than the LIE of CG flashes in the majority (but not all) of forest types and ecological regions (Figures 2a and 2b).However, this tendency is not clear when using the 99th percentile threshold (Figures 2c and 2d).If the error bars include the value 0, this indicated that the LIE of CC may not be higher (or lower) than the LIE of CG flashes for those forest types and ecological regions.In turn, we obtained high uncertainty in the comparison of the LIE of CC flashes and the LIE of CG lightning as a consequence of the high standard deviation of the CG fraction.

Optical and ELF Signature of a Lightning-Ignited Wildfire in the Valais
In this section, we analyzed the CC of a fire-igniting lightning flash recorded by a camera in the Valais (Switzerland) on 21 July 2021 between 21 and 22 hr (UTC-0) at 30 frames per second.Figure 3 shows four selected frames from the video.The advancing leader can be seen in the frame 124 before attachment to ground, while the frame 156 shows the most luminous phase of the return stroke after attachment.A second lightning flash behind the fire-igniting flash can be seen in the frame 169, about 0.5 s after the most luminous phase of the first flash.Finally, frame 200 shows the fire-igniting flash during the decay of its luminosity.The total duration of the flash luminosity was about 4 s, a long duration that could be influenced by the potential afterglow of pixels of the camera.
Table 4 shows all the strokes candidates for the fire-igniting flash detected by ENGLN and LINET between 21 and 22 hr and within 10 km of the position of the fire.The strokes labeled as * and # were simultaneously detected by both lightning location systems as CG, while the difference between them coincide with the timing between the two flashes detected in the video (lower that 0.5 s).Therefore, we concluded that the most probable stroke candidate for the recorded lightning-ignited wildfire in Valais is the positive CG stroke labeled as * and detected by ENGLN and LINET at about 21:23:26.21hr.The magnetic field reported by the ELF sensor associated with the stroke candidate is plotted in Figure 4 together with the temporal evolution of the luminosity.The luminosity is extracted from a single pixel located in the channel of the flash that produces the fire.This pixel has coordinates (200, 160), and it is positioned at half the height of the channel and radially centered.Because the exact absolute timing of the video is not known, the temporal evolution of the luminosity has been manually synchronized with the ELF signal.The temporal evolution of the magnetic field indicates the existence of a CC lasting about 400 ms, while the duration of the luminosity is significantly larger (between 2 and 4 s).In turn, there are several coincidences between the peaks of the magnetic and the optical signals.On the other hand, we obtained a CMC of 188 C km.In summary, the analysis of the optical and the ELF radio signal of this event confirms that the recorded fire-igniting flash had a long CC.

Lightning Candidates
A daily cycle emerges in the distribution of the holdover (Figure 1b), probably, as a consequence of the diurnal cycle of temperature and relative humidity (Pérez-Invernón, Huntrieser, & Moris, 2022;Pineda et al., 2014;Pineda & Rigo, 2017).Contrary to Pérez-Invernón, Huntrieser, and Moris (2022) and Schultz et al. (2019), we obtained a significant number of lightning candidates (40%) more than 2 km away from the reported location of the fire (Figure 1c).The high frequency of distances above 2 km may be due to the difference between the point where the lightning strikes the ground and the flash location calculated as the position of the centroid of the groups weighted by their energy.The frequency of distances above 2 km is significantly reduced from 40% to 26% when we consider the distance between a fire and a flash as the distance to the closest group reported by GLM.

Role of Flashes With CC in Igniting Lightning-Induced Wildfires
The role of continuing currents in LIW was previously suggested by field and laboratory observations (Feng et al., 2019;McEachron & Hagenguth, 1942;H. Zhang et al., 2021).However, quantifying this role at a larger scale was not possible until recently due to the low detection efficiency of lightning location systems for continuing currents and the uncertainty related to the identification of the lightning causing the reported wildfires.Using continuous optical measurements from geostationary orbit by GLM over CONUS, Pérez-Invernón et al. (2023) suggested that continuing currents could play a significant role in the ignition of wildfires.In this work, we extended the analysis of Pérez-Invernón et al. ( 2023) by calculating the LIE of total, CG and lightning with CC in different forest types and ecological regions.We used the method developed by Fairman and Bitzer (2022) to identify continuing currents from optical measurements of GLM lasting more than 10 ms and the wildfire data reported by Short (2021) and Wright et al. (2011) between June and August 2018 over CONUS.
We found that the global LIE of flashes with CC lasting more than 10 ms is higher than the averaged LIE of CG flashes over CONUS, indicating that models of lightning-induced wildfire risk could benefit from including information on flashes with CC observed from space.However, the results on LIEs vary greatly with the threshold of percentile used to select lightning candidates (Figure 2).In addition, we found that the LIE varies  (Berckum et al., 2023).The two detected lightning channels are labeled with the symbols * and #, respectively.
importantly with the forest type and the ecological region (Tables 2 and 3), which agrees with previous studies (e.g., Pérez-Invernón, Huntrieser, and Moris, 2022;Pineda et al., 2022;Schultz et al., 2019).More data are needed to confirm our results, such as larger samples of LIW and meteorological variables across different regions.D. Fuquay et al. (1972) reported simultaneous observations of 11 fire-igniting lightning by optical cameras and ELF radio sensors, confirming the presence of continuing currents.Here, we combined the optical signal of a lightning discharge starting a wildfire in the Alps from a ground-based camera with lightning measurements provided by two lightning location systems (ENGLN and LINET) and an ELF radio sensor.The duration of both the optical and the ELF radio signature suggests the presence of a long-lasting CC in the lightning flash that ignited the wildfire, followed by possible M-components produced by the interaction between an in-cloud advancing leader and the main cloud-to-ground channel (Tran & Rakov, 2019).The luminosity indicates that the channel does not turn off after 400 ms, suggesting that the ELF radio sensor would miss the decay of the CC.The consecutive peaks in ELF and the luminosity could possibly be attributed to M-components (Cai et al., 2022;Rakov et al., 1995) that would transfer more charge than the CC (Y.Zhang et al., 2016), which would allow for the ELF radio sensor to again detect it, as reported by Lapierre et al. (2014) for a flash with CC in New Mexico.Systematic differences in the total duration of the luminosity and the ELF radio signal of flashes with CC has been previously reported by several authors (e.g., Kohlmann et al., 2022;Saba et al., 2006).For example, Kohlmann et al. (2022) reported that the duration of the CC detected by the ELF radio sensor are between 10% and 69% shorter than in the videos.They suggested a number of reasons that could explain these differences, including a possible overdub of the E-field change, transfer of charge produced by the leader or a chemical afterglow producing optical emissions.The obtained CMC (188 C km) was higher than in typical lightning discharges and is in agreement with the CMC reported by D. Fuquay et al. (1972) in western Montana forests.In conclusion, the analysis of this single case study supports our results at a large spatial scale obtained from the analysis of GLM lightning measurements over the CONUS, suggesting that flashes with CC are more likely to cause wildfires.Note.The labels * and # indicate the strokes candidates for the lightning channels showed in Figure 3.

Table 4 Strokes Candidates (Intra-Cloud [IC] or Cloud to Ground [CG]) for the Fire-Igniting Lightning Recorded in Valais
Figure 4. Temporal evolution of the luminosity of a pixel located at the center of the lightning channel of the video showed in Figure 3 (orange line) and the magnetic field Extremely Low Frequency signature (blue line).The units of the magnetic field are pT, while the luminosity is plotted in arbitrary units in the same y-axis as the magnetic field.

Limits of the Study
Here, we report a higher probability of wildfire ignition by flashes with CC lasting more than 10 ms than by total and CG lightning.However, several aspects limit the interpretation of our findings.For instance, the uncertainty related to one-to-one relationship between wildfires and lightning flashes remains high.We used a proximity index that combines the holdover time and distance between reported lightning and fires to search for the most probable lightning candidates.The large amount of lightning flashes detected by GLM with positive values of the proximity index (see Table 1) introduces a substantial uncertainty in the determination of the most probable lightning candidate.This uncertainty is enhanced by the difficulties in distinguishing between IC and CG flashes from space (Ringhausen et al., 2021).In addition, the misclassification of continuing currents may introduce more uncertainty, given that we cannot rule out the possibility that some IC lightning producing long-lasting optical emissions are also classified as CG lightning with CC (Fairman & Bitzer, 2022).
On the other hand, under conservative scenarios for selecting lightning candidates (=10 km and 95-100th percentiles of the proximity index A), we found that 12%-40% of all the LIW may have been caused by a lightning event with CC.Therefore, our results suggest that a high proportion of wildfires may actually be caused by lightning without CC given that these non-CC lightning are simply much more common.Consequently, we cannot exclude the possibility that most of the lightning-induced wildfires are ignited by lightning discharges without a CC phase.
The high uncertainty levels related to both the selection of candidate lightning as well as the fraction of CG flashes complicate the comparison between values of LIE CC and LIE CG.First, the percentiles used to decide how many fires were produced by lightning with CC (e.g., 95th or 99th) have a strong influence on LIE CC (Table 2).On the other side, the spatial variability in the climatology of CG fraction results generally in high values of standard deviation of CG fraction, and consequently in broad ranges of potential values of LIE CG (Table 2).Future research could apply better estimations on the fraction of CG flashes, for example, with the support of auxiliary lightning data from the specific period of study (e.g., summer 2018 in this research).
Finally, the prolonged duration of the lightning-induced wildfire in the Alps suggests that the CC duration could last for several hundred milliseconds (LCC lightning), whereas our current classification of lightning with CC from GLM measurements includes all flashes with a CC lasting more than 10 milliseconds.This difference in duration may be of relevance for estimating the LIE of lightning with CC.Hence, developing a new classification of LCC lightning based on GLM measurements could be valuable for studying the impact of CC duration on ignition probability.

Conclusions
The main conclusions of this work are: 1.The probability of wildfire ignition by CG lightning is larger for flashes with CC.In particular, we found that approximately 12%-14% of lightning-induced wildfires are attributed to flashes with CC as the most probable lightning candidate, whereas the overall occurrence of flashes with CC among all lightning strikes is approximately 10%. 2. The significant variation in the LIE of lightning with CC across different forest types and ecological regions suggests that other factors, such as wildland fuel characteristics and fire-related meteorological conditions, exert a substantial influence.3. The detailed analysis of one single lightning-ignited fire in the Alps by combining optical, VLF, and ELF measurements confirms the presence of a LCC in the lightning discharge that started the fire.4. The holdover time and spatial distance between wildfires and lightning candidates identified from the GLM groups show remarkable similarities from the stroke and flash data typically observed in lightning networks.This novel finding holds significant value for ecological studies.
Our results on the role of CC lightning in the ignition of wildfires indicates that parameterizations of flashes with CC in atmospheric models (Pérez-Invernón, Huntrieser, Jöckel, & Gordillo-Vázquez, 2022) together with information on forest types and their fuel characteristics may improve the forecasting of lightning-induced wildfires.More data on wildfire and lightning detection across the field of view of GLM aboard the GOES-17 satellite will provide new information about the role of flashes with CC in the Americas.Similarly, the MTG-LI will PÉREZ-INVERNÓN ET AL.

Figure 1 .
Figure 1.Holdover time in days (a), holdover time in hours for the first 100 hr (b) and distance between each fire and the most probable (maximum A index) lightning flash candidate (c) and group candidate (d) using D max = 10 km.

Figure 2 .
Figure 2. Differences (in %) between the Lightning Ignition Efficiency (LIE) of Cloud to Ground (CG) lightning and the LIE of flashes with Continuing Current (CC) in different forest types and ecological regions.The first and the second row show the comparison of the LIEs by assuming that lightning-induced wildfires are ignited by a flash with CC if there is at least one lightning candidate with CC with a proximity index A above the 95th percentile (a, b) and the 99th percentile (c, d), respectively.The colormap indicates the total number of lightning-induced wildfires in each forest type or ecological region.The error bar (gray dashed lines) represents the results by assuming the Intra-Cloud (IC) fraction ± σ.

Table 1
Total Number of Lightning-Ignited Wildfires With At Least One Flash With Continuing Current Candidate Lasting More Than 10 ms Above a Given Percentile of Proximity Index A Among the Lightning Candidates, by Using Two Different Values of D max Note.The numbers in parenthesis indicate the percentage of LIW that could have been caused by flashes with CC, relative to the 5,585, 5,731, and 5,634 analyzed wildfires with flashes candidates.

Table 2
LIE in Forest Types of CONUS