2.1 Lightning Data

We use lightning data from the Vaisala Global Lightning Dataset, GLD360 (Said et al., 2010; Said & Murphy, 2016). It contains time, location, peak current and type (CG or intracloud (IC)). The detection efficiency (DE) and location accuracy (LA) for CG flashes in the USA is evaluated to ∼75%–85% relative to the National Lightning Detection Network, which has a CG flash DE > 95% (Mallick et al., 2014). The median LA for CG strokes is 1.8 km. The accuracy in the Adriatic Sea is assumed to be the same, since an earlier test in Europe showed better DE and LA than those found in the USA (Poelman et al., 2013). We also use lightning data from The Earth Networks Total Lightning Network (ENTLN) (Lapierre et al., 2019; Zhu et al., 2022). Its CG flash DE and LA in Florida, USA is evaluated to be >98% and 91 m, respectively (Zhu et al., 2022). These parameters vary with network sensor density. For some lightning events, that will be discussed in Section 4.4.2, ENTNL provided the lightning waveforms which were interpreted manually.

2.2 Broadband Electric-Field Sensor

The vertical broadband electric field from 1 kHz to 5 MHz was measured with a dipole whip antenna installed in the center of France (46.1°N, 2.8°E), 882–1,330 km from the storm location (labeled BB in Figure 1). The system is the same as the one presented in Farges and Blanc (2011). It measures the variation of the electric field and triggers if the variation exceeds 2 V/m, storing 30 ms of data from 6 ms before the trigger at a sampling frequency of 12.5 MHz. The resolution is 14 bits and the dynamic range is ±200 V/m. We use the measurements to characterize lightning and attenuation of MF radio transmitter signals.

2.3 ELF Receiver

The current moment waveform (CMW) and CMC were obtained from measurements of an ELF receiver system in the Bieszczady mountains in Poland (49.2°N, 22.5°E ∼850 km from the storm and labeled ELF in Figure 1). It measures the magnetic field component with two antennas aligned in the geographic north-south and east-west directions in the frequency range 0.02 Hz–1.1 kHz. The receiver features a Bessel anti-aliasing filter with a bandwidth of 900 Hz. The sampling frequency is 3 kHz. The CMW and the CMC were reconstructed using the method of Mlynarczyk et al. (2015) that accounts for the dependence with the frequency of the signal attenuation and the propagation velocity in the ELF range.

2.4 VLF Receiver

A Sudden Ionospheric Disturbances (SID) monitor measures perturbations to narrow-band VLF signals from powerful transmitters used for communication with submarines. The monitor used in this study is operated by the Slovak Organization for Space Activities and placed in Bojnice, Slovakia (48.8°N, 18.6°E, labeled SID in Figure 1). It can record up to 16 VLF transmitters simultaneously with a sampling frequency of 2 Hz. We use the NSY transmitter operated by the Naval Computer and Telecommunications Station in Sicily (37.1°N, 14.4°E), which broadcasts at 45.9 kHz with 250 kW. The signal propagates from Sicily to Slovakia (∼1,345 km) and reflects multiple times at the surface of the Earth and the bottom of the ionosphere. The propagation great circle path (GCP) of the NSY signal crosses directly the thunderstorm location (see Figure 1).

The NSY-SID link is the only VLF link we could identify that passes the thunderstorm area. The analog output of the SID monitor is digitized at a relatively modest rate of 2 Hz because it is designed to detect solar flare effects on the D-region ionosphere. Although the sampling rate is too low for the unambiguous identification of LOREs and early/fast, early/slow events, we use the data to identify such events in lack of alternative VLF systems with higher temporal resolution. Ideally, the identification of these events requires accuracy of a few ms to determine their onset time, 20 ms to identify fast events, and 200 ms to distinguish between fast and slow events. It means we cannot distinguish between early/fast and early/slow, but lump the two into “early” events. We then identify perturbations as LORES or “early” if they correlate with CG strokes within the 0.5 s uncertainty of the onset time and otherwise have the same pulse characteristics as such events. As we are not aware of alternative event types that resemble “early” observations, we suggest their identification is quite robust. We note that although the link passes close to the active thunderstorm cells, LOREs and early events may be missed because of VLF mode coupling, the properties of the Earth-ionosphere waveguide, or properties of the disturbed region (Haldoupis et al., 2010).

2.5 Optical Observations

The optical observations are performed by a TLE observatory installed in Rustrel, France (43.94°N, 5.48°E) at 1,025 m altitude. It is mounted on a building made available by Laboratoire Souterrain à Bas Bruit (LSBB). The camera is a Watec 1/2” monochrome CCD camera (WAT-902H) with a 16 mm lens that gives 23° horizontal and 17° vertical field of view (FOV). It takes 50 de-interlaced fields per second, corresponding to a time resolution of 20 ms. The images are time referenced after synchronization with an NTP server, giving an absolute time uncertainty below 5 ms. The camera is mounted on a QuickSet motorized Pan-Tilt unit allowing for active and automatic tracking of thunderstorms. The night of the observations, the camera tracked two thunderstorms automatically, selecting a new pointing direction every 15 min. The camera, therefore, pointed away from the storm during parts of the night (see Section 4.1 for the gaps). The analysis excludes strokes that occurred in these gaps. From 02:45 UTC the camera stayed in the same position for the rest of the night.

2.5.1 Methods and Error Estimations

Figure 2 shows an image of an elve observed at 03:57:56.578 UTC on 10 December 2020. The rest of the elves are shown in Figures S2–S64 of Supporting Information S1. From the images, we estimated their altitude and relative brightness. Following Van der Velde and Montanya (2016), we calculated the altitude by combining the elevation of the elve centers with the location of the CG stroke reported by GLD360. The elevation angle was retrieved by matching the background stars to the star catalog in the software ‘Cartes du ciel’. The calculations are done assuming a spherical Earth with radius of 6,370 km and a camera altitude of 1,025 m. Before the image analysis, the background for each elve, determined as the mean value of two de-interlaced fields preceding the elve, was subtracted. The elve centers were estimated manually with the cursor. Since the elves are faint and diffuse, their centers can be difficult to determine, which introduces an error in the elevation angle. Repeated estimates suggest that the uncertainty in this naked-eye method is 0.1°, which corresponds to an altitude uncertainty of ±1.4 km at 800 km distance to the elves. The uncertainty in the location of the parent strokes (median value 1.8 km (Said & Murphy, 2016)) introduces an uncertainty of around 0.3 km at 800 km distance to the storm. We estimated the relative brightness of the elves from the sum of all pixels in the background-subtracted elve images after dividing by the width of the elve expressed in number of pixels. The brightest elve was used as a reference.

4.1 Elves

During the night of the storm, the camera in Rustrel detected 63 elves above the Adriatic Sea. One is shown in Figure 2, and the rest in Figures S2–S64 of Supporting Information S1. We note that the large distance, and thereby the low elevation angle, leads to high atmospheric absorption of photons, which reduces the detection efficiency of elves by the camera. In Chang et al. (2010), the ISUAL FUV photometer on the FORMOSAT-2 satellite appeared to detect 16 times more elves than the imager, also suggesting that a large number of events may be below a sensor detection threshold. The lowest altitude at the storm that the camera could observe was 50 km (at 750 km distance) and 80 km (at 950 km distance). These values are calculated using great circle geometry assuming a spherical Earth with radius 6,370 km, a camera altitude at 1,025 m and based on an estimation in “Cartes du Ciel” that the horizon is at around 0.4° of elevation. Thus, we cannot rule out other TLEs, such as sprites and halos, occurring below these altitudes. The camera observed the storm from 20:00 to 05:35 UTC with gaps totaling 1 hour and 45 min (22:00-22:30; 23:30-00:00; 00:30-01:15 UTC).

The optical characteristics were studied for all but a few elves: In four cases, no parent stroke was reported by GLD360 that allowed the estimation of their distance, three were too faint to define their shape, and the brightness of four could not be determined because the moon was behind the elve in the video field. Thus, out of 63 elves, 56 were used for the brightness and altitude calculations.

The altitudes of the elves ranged from 80 to 90 km (see Figure S1 in Supporting Information S1), which is within the range and variability of the results in Van der Velde and Montanyà (2016). All elves below 83 km occurred in the first hour of observations. We attribute this to the storm itself and not to changes in the ionosphere based on results from the NASA international reference ionosphere model (IRI) that appear to be constant. There is no clear trend between altitude and local time for the rest of the night, as seen from the altitude versus time plot in Figure S1 of Supporting Information S1. The relative brightness varies down to ∼17% of the brightest elve. The brightness is correlated with the peak current and the Power Spectral Density (PSD) of the parent stroke in the band of the broadband receiver (see below), a parameter discussed in later paragraphs.

4.2 Lightning Broadband Waveforms

During the periods the camera observed the storm, GLD360 detected 234 CG strokes with absolute peak current values above 200 kA within the camera's FOV. This means that 63 strokes produced elves and 175 strokes did not produce observable elves (four elves did not have matching strokes). To understand why some produce elves and others do not, we analyzed the waveform of the vertical electric field from the lightning strokes, measured by the broadband receiver in France. For each stroke, we determined the maximum amplitude of the ground wave (E_GW), the rise-time of the electric field pulse, defined as the time from 50% to 90% of E_GW, and the fall time from the maximum to the background. The statistics, presented in Table S2 of Supporting Information S1, show no apparent difference between strokes with and without elves. Comparing the average of the normalized electric field waveforms of strokes with and without elves (Figure S66 in Supporting Information S1) shows that the ground wave and the first sky wave on average have similar amplitudes, and that the second sky wave and the following ones (third and fourth) are stronger with elves.

In Blaes et al. (2016), the peak current of the CG stroke is found to be a decisive parameter for elve generation, with a probability reaching 50% at 88 kA. Other authors found thresholds from 38 to 130 kA (Chen et al., 2014; Van der Velde & Montanyà, 2016). In our case, the formation threshold of elves is around 200 kA. The variability of reported thresholds is likely an effect of the sensitivity of the optical instrument used (camera or photometer, e.g.,), or uncertainty in the estimation of the peak currents reported by the detection networks. However, a question remains why not all high-current strokes generate bright elves, as noted in Kolmašová et al. (2021).

To explore this question further, we computed the electric field wave power, E², which is the frequency integral of the electric field PSD over the whole antenna bandwidth (1 kHz–5 MHz) using the method of Ripoll et al. (2021). It is computed over 1.5 ms from the arrival time of the ground wave, to include the ground wave and all the sky waves. Figure 5 shows that the CG strokes producing observable elves have on average 3.23 larger power than those that did not. We also see from Table S2 in Supporting Information S1 that the elve producing strokes in this study have three orders of magnitude (1593X) larger power than the typical flashes analyzed in Ripoll et al. (2021) using the same electric field sensor. This suggests that the generation of elves depends on the complete electromagnetic energy release of the stroke.

4.3 ELF Measurements

We analyzed ELF measurements of the electromagnetic signals from a subset of the lightning strokes by the sensor in the Bieszczady Mountains in Poland. The selection included high-current strokes with and without elves, and strokes with elves associated with LOREs and not. The selection was made to use events that occurred at different times throughout the night. We calculated the current moment (CM) and the impulse charge moment change (iCMC) (Cummer & Lyons, 2004) defined as the total CMC during the first 2 ms of the lightning stroke according to the method of Mlynarczyk et al. (2015). The iCMC is likely more relevant for elves than the total CMC since elves are generated within the first milliseconds. The results and other relevant stroke parameters are presented in Table 1. The average iCMC is −131 ± 16.4 C km for strokes that generate elves and −41.2 ± 13.4 C km for strokes without observable elves. The average maximum CM is −121 ± 19.2 kA km and −39.25 ± 11.1 kA km for the two types, respectively. Thus, in this sample, the average iCMC is 3.1 times larger and the average maximum CM is 3.2 times larger for strokes that generate elves. Larger values can result from larger currents, longer channel lengths, or both. Since GLD360 data only provide the maximum current obtained in the VLF range, one cannot expect full correlation with the iCMC or CM, as pointed out in Lu et al. (2012). Large iCMC values are closely related to sprite and halo production. However, since the elves, for which we calculated the iCMC, were all caused by −CG lightning and the common threshold for sprite productions is around 600 C km (Cummer & Lyons, 2005) and possibly higher for negative sprites (e.g., Qin et al., 2013), we do not find it likely that there were elve-sprite pairs in these cases. In addition, the camera would most likely observe halos or sprites, had they been there. One must think more, however, as to what the physical significance of the large CMs and iCMCs is, in relation to lightning that causes bright elves. This would require modeling and could be the topic of a new paper. Their relation to LOREs is discussed in a following paragraph.

4.4 VLF Transmitter Signal Perturbations

The amplitude of the NSY VLF signal from the night of 9–10 December is shown in Figure 6 on different temporal scales. The variation on scales larger than ∼30 min are not related to thunderstorm activity, but to other ionospheric processes (e.g., Hunsucker (1982); Kumar et al. (2017) and references therein) because the signal from the same transmitter during a night without storms shows similar variations. However, amplitude perturbations on shorter scales are multiple, and many correlate in time with lightning activity detected by GLD360. These are identified using the criterion that the perturbation amplitude must be greater than 0.25 dB relative to the average amplitude (in dB) of the preceding 10 s. This condition corresponds to a threshold of 4σ, where σ is the average standard deviation for the night. The value is close to the typical value at 0.2 dB of Inan, Pasko, and Bell (1996). In addition, we require a lightning stroke to be detected by GLD360, or an elve by the camera, within 0.5 s before the perturbation. The 0.5 s limit corresponds to the temporal resolution of the receiver. The algorithm used to identify candidate events is used on a 3-point moving average of the signal, shown in Figure 6 in red. All candidate events are manually validated and categorized.

The time resolution of the VLF receiver allows for classification in LOREs and early events, but not distinction between early/fast and early/slow events (Section 2.4), and is sufficient to exclude most lightning-induced electron precipitation events that have onsets 0.3–1.6 s relative to their causative stroke (Burgess, 1993; Peter & Inan, 2007). The two types are identified using a 20-point moving average (corresponding to 10 s), and are shown in Figures 6 and 7. Early events appear as sudden increases in the signal with a recovery (decrease of signal) that starts within 10 s after the peak. These events, therefore, appear as a peak with no plateau on the top in the smoothed signal. In our signal, we categorize LOREs as step-like perturbations that can be either positive or negative and that do not show recovery within the first 20 s. It means there will be either a plateau after the step or the amplitude keeps decreasing/increasing for at least 20 s (see the example in Figure 7a). We identify 68 early events (all with positive amplitude) and 18 LOREs (14 negative and 4 positive). The events are marked in Figure 6 and the main characteristics of the lightning strokes related to the two types of perturbations are given in Table 2. There are also signal perturbations related to high peak current lightning or even elves that have the shape of negative amplitude LOREs except that the onset is significantly slower at 1–3 s. One example is seen in Figure 6b at 22:46:53 UTC. For brevity, such events are not investigated further in this work.

Because of the considerable variation of the background signal and the high number of lightning-induced perturbations, it is hard to determine a recovery time for the individual perturbations. However, for all the early events, the recovery time looks shorter than 3 min (most are ∼ 1 min), which is consistent with the typical recovery time of these events of 10–100 s (Inan et al., 2010; Inan, Pasko, & Bell, 1996). The LOREs may not recover before other variations mask them. However, they appear longer than the early events. In some cases, the perturbation is hard to categorize (positive LORE or early event) because the LORE step or the shape of an early event is unclear due to the varying background signal. Another complication is that perturbations can overlap. Thus, a few events can be miscategorized.

Figures 8a and 8c show the location of the lightning strokes related to the different types of perturbation and Figure 8b a histogram of the minimum distance from the stroke to the VLF path for LOREs and early events.

4.5 MF Radio Wave Attenuation

As mentioned in the introduction, Farges et al. (2007) conducted a statistical analysis of MF attenuation related to more than 4,000 nighttime CG strokes with peak currents from 10 kA to more than 100 kA. They used nine radio links between 900 and 1,557 kHz with the transmitters located between 380 and 970 km from the receiver. They found that the temporal variation of the MF attenuation matches the typical rise time and duration of optical pulses of elves and that MF-attenuation can be detected for lightning to a distance from the ray-path that compares to the radii of elves (Farges et al., 2007). This suggests a fast process in elve regions, which led Farges et al. (2007) to conclude that attenuation was caused by electronic heating of the lower ionosphere by the EMP.

Here we explore the impact of CG strokes with an absolute peak current value greater than 200 kA, some of which triggered observable elves. Our objective is to know how the elve presence contributes to the occurrence of the MF perturbations. During the storm of 9–10 December 2020, 59 out of 234 CG strokes with peak current above 200 kA were followed by the observation of an elve. For this study, we use exactly the same method of data analysis as Farges et al. (2007), whose main steps are recalled in the Supporting Information for easier reading. The new results presented here can thus be fully compared to those of Farges et al. (2007).

We identified seven MF radio transmitters from www.mwlist.org, where the signals to the receiver pass over the storm. More information about the MF transmitters is provided in Table S1 of Supporting Information S1. An example of an MF perturbation when an elve occurs is shown in Figure 9. The top panel shows the broadband waveform corresponding to the −CG stroke that was associated with an elve at 03:57:56.578 UTC, the middle panel shows the corresponding narrow-band signal amplitudes of four MF transmitters at 540 kHz, 576 kHz, 630 kHz, and 891 kHz, and the bottom panel shows the GCPs of the signals TR path. The attenuation is most pronounced at 540 kHz, smaller for the three other frequencies and absent in the remaining three links. From the map, we see that the four links impacted pass close to the center of the elve. Other examples showing the seven MF narrowband signals, disturbed or not, are given in the Supporting Information (Figures S67-73 in Supporting Information S1).

The same technique used by Farges et al. (2007) was systematically employed here to calculate the parameters describing the perturbation. The peak attenuation, onset time, rise time and duration of the events were systematically measured after each CG stroke for the seven MF radio links, whether or not the MF perturbation was observed. The median, the mean and the standard deviation of these parameters are given for each transmitter in Table S3 of Supporting Information S1. These statistical values are also calculated, not taking into account the transmitter frequency, and are shown in Table 4 for the 59 events with elves and the 175 without and compared to the results from Farges et al. (2007).

Lightning affects MF-links in 86% of the 59 cases with elves, in line with Farges et al. (2007) for flashes over 60 kA, and in 53% of 175 cases without elves. We found that up to 5 MF-links can be disturbed with elves, but most often between 1 and 3. It is less likely that multiple links are perturbed simultaneously without elves. Perturbations are then observed more frequently with elves, and their presence seems to reflect an increase the size of the disruptive region. The number of MF-perturbations appears to depend on the local time of strokes without elves, whereas this trend is absent with elves (Figure S74 in Supporting Information S1).

Compared to the observations in Farges et al. (2007), the attenuation we observe is consistent with the ones predicted for CG peak currents higher than 125 kA. The onset times are shorter, the rise times (for elve cases) are similar, and the average duration is slightly shorter. According to Farges et al. (2007), such differences are to be expected. They note that the temporal parameters depend on the radio frequency and the relative locations of the stroke, transmitter, and receiver. The transmitter frequencies we use are lower (531–909 kHz vs. 900–1,557 kHz), and the links are longer (2540–5,650 km vs. 383–971 km). They find that the flash location relative to the link may affect the onset time up to a factor of ∼3. In addition, the CG strokes we consider are over the ocean, and those of Farges et al. (2007) are over land, which may affect the EMP in the lower ionosphere. We conclude, then, that the very brief MF radio emission perturbations are present even when the storm is between 882 and 1,330 km from the receiver. We also find from Table 4 that there is little difference in the parameters (except for MF attenuation occurrence) for events with elves and those with no elves. The lack of difference is perhaps because no-elves events do have elves that are undetected by the camera system. We have, indeed, cases without elves with E² of the same order of magnitude as with elves.

5.1 Elves, LOREs, and MF Attenuations

Our observations support the understanding that elves are associated with LOREs (Haldoupis et al., 2013), while the early VLF perturbations are distinct from LOREs and have other origins. However, 78% of the elves occur without LOREs and in the following we try to answer what determines the generation of a LORE.

As seen from Table 1, the presence of LOREs are not mirrored in the parameters derived from the ELF data. For instance, the CM is not significantly smaller for cases without LOREs. We cannot, then, search for an explanation in the ELF data. Turning to the VLF data, Marshall and Inan (2010) present a finite-difference, frequency domain model of narrow-band VLF transmitter signal propagation in the Earth-ionosphere wave-guide. They place an electron density perturbation at the ionospheric boundary at 85 km somewhere along the propagation path and calculate the changes in the signal properties at a receiver. They show that the amplitude perturbation can be both positive and negative, that it is largest if the disturbance is where the amplitude at the ground is low, such as an interference null, and can be suppressed entirely if it falls at an interference null at the reflection altitude. The amplitude also depends on parameters such as the path length, ionospheric and ground properties, and the signal frequency, however, the nighttime electron density fluctuations have a minor influence. In line with their model, we find both positive and negative perturbations in the amplitude and no dependence in local time of the occurrence of LOREs (see Figure 6a).

In Figure 8a, we show the location given by GLD360 of the CG strokes that produced elves and LOREs (early events are also shown), and in panel c a zoom of a region close to the VLF path. As noted earlier, Figure 8a shows that all negative LOREs come from the same storm and that the positive LOREs come from a different one. In that sense, the observations support the model results of Marshall and Inan (2010) and observations in NaitAmor et al. (2013) that for a given TR-link the relative location of the lightning/disturbance and the VLF path is important for the sign of LOREs. The possible physical explanation for this is that the LORE signature is due to forward scattering or reflection from an area of perturbed conductivity to the receiver (Haldoupis et al., 2013; Rodger, 2003). The scattered or reflected signal beats with the direct signal to the receiver and, depending on their phase difference, there will be either constructive (positive LORE) or destructive (negative LORE) interference.

However, although many of the elves without LOREs are at a greater distance and different location relative to the VLF signal path, such as those of the cell at ∼42.8°N, ∼14.5°E, some are very close (Figure 8c). Since in this study, the ground and ionospheric conditions and the TR characteristics are the same for all elves, it suggests that other parameters related to the lightning and elves can affect the detection of LOREs. For example, we see from Figure 10a that elves with LOREs are brighter and produced by strokes of higher power and spectral energy, E². This observation implies that the stroke energy should be large enough to create appreciable ionization before we can observe a LORE. Such a result could be anticipated based on the correlation between CG stroke peak current and LORE probability from Haldoupis et al. (2013) and that higher peak current strokes produce brighter elves (Barrington-Leigh & Inan, 1999) and larger electron density changes (Marshall et al., 2010). But observations showing the relationship between stroke peak current, broadband energy, elve brightness and LOREs has not been shown before.

The altitude of atmospheric perturbation may also affect the triggering of LOREs. Assuming that the main perturbation occurs where elves are generated, we see from Figure 10b that their altitudes are relatively high, from 86 to 90 km. This suggests that the altitude range of effective reflection and therefore the VLF frequency is important for the detection of LOREs. Still VLF propagation in the Earth-ionosphere waveguide is a complex physical process that depends non-linearly on several uncontrolled physical parameters and VLF transmitter signal characteristics such as irregular changes in the Earth-ionosphere waveguide, the scattering process itself, the phase difference at the receiver between the direct and the scattered signal, and so on. The interested reader can refer to Marshall and Inan (2010); Haldoupis et al. (2010); Gordillo-Vázquez et al. (2016) for more information. The present study cannot explain all the mechanisms behind the selective process of VLF event detection. This would require data from multiple sensitive VLF links cutting through the storm region in combination with modeling.

We next turn to the MF attenuations. Whereas changes in the D-region electron density cause LOREs, MF attenuation is related to the heating of electrons. Figure 10 show no simple relationship between MF attenuation and elve brightness, altitude, or lightning power. For example, lightning strokes with power 22–56 (V/m)² generated elves with and without MF attenuations. The location of the disturbance relative to the signal path does, of course, play a role. However, in the region of Figure 8c, all but one observed elve created MF attenuations, which is the same proportion as overall.

Characterization and modeling of EMPs and their effect on the upper atmosphere is complicated, as is the modeling of wave propagation in the Earth-ionosphere waveguide and scattering and absorption of wave energy in D-region perturbations (Gordillo-Vázquez et al., 2016; Haldoupis et al., 2010; Marshall & Inan, 2010). We conclude that our dataset does not allow for understanding why some strokes trigger LOREs or MF attenuations and others do not, and that this requires further modeling and observations.

5.2 Early Events

Sprites are almost always associated with early VLF events, but early events may occur without the detection of sprites or any other type of luminous emissions, and it is not understood why (Haldoupis et al., 2010). However, it is clear that impulsive conductivity changes must be generated at ∼70–80 km altitude with luminous emission that are either too dim to be detected (Pasko, 2010), or absent.

In the case of the storm we report on, no sprites or other luminous events were detected in relation to the early VLF events, although the camera horizon was below 70 km for 74% of the early events. Sprites are usually much brighter than elves when observed by video imaging cameras and should have triggered the camera. Nevertheless, one could argue that a storm that generates many elves and no sprites is uncommon and that the camera system, therefore, must have missed the sprites. However during the winter in Europe over sea in low CAPE conditions, storm cells do not develop large stratiform regions with positive charge reservoirs favorable for the +CG strokes that may trigger sprites. This was the case for the presented storm, where the storm cells were less than 100 km across. In addition, previous studies note that elves occur in Europe mainly over maritime thunderstorms, peaking from November to January, suggesting that the cold season thunderstorm charge configuration favors strokes with large EMPs (Arnone et al., 2020; Van der Velde & Montanyà, 2016). For these reasons it appears that the absence of detected sprites is not caused by problems with the camera system detection algorithm.

We find it more likely that optical emissions were too dim to be detected by the camera (dimmer than elves). One suggestion is that sprite streamers are stimulated, but do not reach luminosities sufficient to be detected as sprites (Haldoupis et al., 2010; Pasko, 2010). However data from two independent lightning detection systems agree that 12 of the 68 early events are caused by IC pulses. GLD finds 49 CG strokes, of these 21 +CG and 28 −CG. We are then left with strong indications that both IC pulses and CG strokes of both polarities generate early events. Since IC pulses and −CG strokes rarely generate sprites, it appears unlikely that undetected streamers cause the perturbations, unless, of course, these are quite common but rarely detected.

Another mechanism proposed relates to the halo. Halos are brief (few ms) unstructured layers of emissions at around 70–80 km altitude, which is at the upper end of sprite altitudes and the lower end of elves. They are centered above the causative lightning and extend 30–80 km horizontally (Barrington-Leigh & Inan, 2001; Wescott et al., 2001). It is suggested they are generated by the quasi-electrostatic and induction fields in the D-region ionosphere (Ren et al., 2019) following both polarities of CG lightning (Williams et al., 2007). They have to our knowledge not been observed for isolated IC discharges but are not well investigated because they generally appear dim in standard frame rate cameras, which makes them hard to detect (Inan et al., 2010). It is thought that halos represent increased ionisation, and that the field, therefore, is above the threshold field in an extended region mapped by optical emissions and beyond (Moore et al., 2003).

The electric field related to halos are thought to cause ionization, increasing the electron density and conductivity in a region where the field exceeds the threshold field, and to decrease the electron density in a layer below through electron attachment to neutral particles, driven by sub-break-down fields. This process increases the vertical conductivity gradient, with reduced absorption below and increased absorption above. The perturbation is proposed to lead to early VLF events, where their decay of 10–100 s is considered consistent with the lifetime of free electrons in the enhanced-density region at that altitude (Moore et al., 2003). It is further suggested that a larger diffuse region of perturbation on the scales of a few wavelengths of VLF transmitter signals better explain their scattering than the 100 m - 1 km scales of sprite structures (Inan et al., 2010).

A third proposed triggering mechanism for the observed early events are regions of electron density changes without optical emissions. Marshall et al. (2008) showed that density changes in the lower ionosphere by electron losses through dissociative attachment to molecular oxygen can create measurable amplitude changes in VLF transmitter signals that travel through the disturbed region. The energy required for attachment (3.7 eV) is lower than that of N₂ optical emissions often seen in sprites and elves (7.5 eV) and N₂ and O₂ ionization (15.6 eV) (Haldoupis et al., 2006; Neubert & Chanrion, 2013). This implies that attachment can occur without optical emission and ionization, explaining why we and other studies (e.g., Marshall et al., 2006) report early events without associated optical emissions, and also why optical emissions without early events are rare (Haldoupis et al., 2004, 2010). The timescales of recovery for density changes in the lower ionosphere controlled by attachment-detachment processes is in the order of 100 s (Pasko & Inan, 1994), consistent with the recovery times of the early events. As discussed in Marshall et al. (2008), a consequence of this hypothesis is that early VLF events caused by attachment-depleted regions would mostly have positive perturbation amplitudes due to less VLF signal absorption in the reduced density region. This scenario is consistent with our results as well as results from other previous studies (e.g., Haldoupis et al., 2004; Inan et al., 1993; Inan, Sampson, & Taranenko, 1996; Marshall et al., 2006).

Marshall et al. (2008) show that the field can be attributed to the EMP from successive IC lightning discharges. This is also suggested by observation reported in Johnson and Inan (2000) and Haldoupis et al. (2006) note that densely clustered EMPs of IC activity can explain the slower onset of early/slow events that do not match the timescales of return strokes. Model results from Gordillo-Vazquez et al. (2016) also showed that EMP effects on the ionosphere below 79 km by weak peak currents may produce attachment depleted regions that might be capable of generating early events. The cases we observe are related to positive and negative CG strokes, and IC pulses, and likely, both the EMP and the QE fields play a part. For example, the three cases summarized in Table 3 have ∼500 ms sequences of flashes that reach relatively large CMCs that are well above the threshold for sprite production suggested in Cummer and Lyons (2005) and Yair et al. (2009) but likely increase too slowly for sprites or haloes to be observed (Hiraki & Fukunishi, 2006; Pasko et al., 1997). The observed sequences of flashes are similar to that of the sequential IC flashes assumed by Marshall et al. (2008) necessary to cause a cumulative effect of the attachment process so that an early fast event becomes strong enough to become detectable at the VLF receiver. Figure S65 in Supporting Information S1 shows the CM and CMC associated to the three early events in Table 3.

In summary, the observations of early events lead us to suggest that these observations may relate to regions of reduced conductivity caused by an electron attachment/detachment process at ∼70 km altitude, or by electron enhancements associated with TLEs that are too dim to be detected by the camera.