Comparing lightning observations of the ground-based EUCLID network and the space-based ISS-LIS

The Lightning Imaging Sensor (LIS) on the International Space Station (ISS) detects lightning from space by capturing the optical scattered light emitted from the top of the clouds. On the other hand, the ground-based European Cooperation for Lightning Detection (EUCLID) makes use of the low-frequency electromagnetic signals generated by lightning discharges to locate those accordingly. The objective of this work is to quantify the similarities and contrasts 10 between the latter two distinct lightning detection technologies by comparing the EUCLID cloud-to-ground strokes and intracloud pulses to the ISS-LIS groups, in addition to the correlation at the flash level. The analysis is based on the observations made during March 01, 2017 and March 31, 2019 within the EUCLID network and limited to 54° north. A Bayesian approach is adopted to determine the relative and absolute detection efficiencies (DE) of each system. It is found that the EUCLID relative and absolute flash DE improves by approximately 10 % towards the center of the EUCLID 15 network up to a value of 50.3 % and 69.4 %, respectively, compared to the averaged value over the full domain, inherent to the network geometry and sensor technology. On the other hand, the relative and absolute ISS-LIS flash DE over the full domain is 49 % and 68.9 %, respectively, and is somewhat higher than the values obtained in the centre of the EUCLID network. The behavior of the relative DE of each system in terms of the flash characteristics of the other reveals that the greater the value the more likely the other system detects the flash. For instance, when the ISS-LIS flash duration is smaller 20 or equal to 200 ms, the EUCLID relative flash DE drops below 50 %, whereas this increases up to 80 % for ISS-LIS flashes with a duration longer than 750 ms. Finally, the distribution of the diurnal DE indicates higher (lower) ISS-LIS (EUCLID) DE at night, related to an increased ISS-LIS:EUCLID flash ratio at night.


Introduction
Lightning processes in the cloud and from cloud-to-ground involve the formation of channels carrying tens of kiloamperes of 25 electric current with temperatures as high as 30,000 K. Those processes are accompanied by intense radiation in the optical frequency range with the peak power typically being of the order of 10 9 W (Christian et al., 1989). These optical emissions are a result of dissociation, excitation, and subsequent recombination of various atmospheric constituents as a result of the sudden intense heating, and primarily occur at discrete atomic lines. Satellite-based optical imagers operating in the visible and near infrared frequency ranges record these optical emissions. The geolocation is carried out by using geometric 30 projection of the images taken from space. In the seventies of last century different satellite programs started to use various https://doi.org/10.5194/amt-2019- 435 Preprint. Discussion started: 23 January 2020 c Author(s) 2020. CC BY 4.0 License. optical sensors to measure lightning, e.g., Vorpahl J.A. et al. (1970), Sparrow & Ney (1971) and Turman (1978). Due to the limited technology at this time, these satellite-based sensors had location accuracies of the order of hundreds of kilometers (due to the low spatial resolution of the optical imagers) and a detection efficiency of less than 2 %.
In 1995 the OV-1 (MicroLab 1) satellite carrying the optical transient detector (OTD) and in 1997 the TRMM satellite 35 carrying the lightning imaging sensor (LIS) were launched. OV-1 orbited at an altitude of 750 km and TRMM at an altitude of 350 km and changed to 400 km after 2001 (Cecil et al., 2014). Therefore, the latter two satellites had a large field of view of 1300 x 1300 km and 600 x 600 km for OTD and LIS, respectively. Those optical imagers measure the signals emitted at 777.4 nm wavelength, associated with dissociation of molecular oxygen into atomic oxygen due to intense heating produced by lightning processes. Data from such sensors typically consist of the time of occurrence of lightning event, latitude, and 40 longitude. The radiance (brightness) for each pixel is also available, but interpretation of these measurements is complicated because the optical properties of the path between the emission and the measurement point vary. Since no relationship exists between the peak optical radiance measured by such sensors and peak currents of lightning events, estimated peak current and polarity of lightning events are not reported by satellite-based lightning sensors. OTD and LIS have a location accuracy of about ten to a few tens of kilometers and a temporal resolution of a few milliseconds with better than 100 ms temporal 45 accuracy, e.g., Boccippio et al. (2000). They detect emissions from both cloud and cloud-to-ground discharges but they cannot distinguish between them. Total flash detection efficiency for LIS during the day and night is estimated to be about 70 % and 88 % respectively and about 38 % and 52 % respectively for OTD, see Boccippio et al. (2002) and Cecil et al. (2014).
It is important to note that, similar to VHF lightning mapping systems, optical imagers are able to map the full spatial extent 50 of flashes, although with poorer temporal and spatial resolution, and hence may be viewed as lightning mapping systems.
Since these optical imagers on low earth orbiting satellites observe a given location on earth's surface for a limited time; typically around 90 seconds to a few minutes, they can only take snapshots of thunderstorms and cannot monitor them as they develop and evolve.
Generally, for all applications of lightning data it is important to know the performance of the employed lightning location 55 system (LLS). The performance characteristics of lightning locating systems are determined by their ability to geolocate lightning events with high location accuracy (LA), high detection efficiency (DE), with low false detections and to report various other features of the lightning discharge correctly. Different methods or a combination of methods may be used to validate the performance characteristics of different types of lightning locating systemssee Nag et al. (2015). To get information about performance variations over large spatial regions of ground based LLS, data of those systems were 60 compared to data from TRMM-LIS. During the last years several papers provided additional insights in the performance of ground based networks with such an analysis, e.g., the WWLLN (Rudlosky & Shea, 2013), the NLDN (Zhang et al., 2016), the ENTLN (Rudlosky, 2015), the ATDnet (Enno et al., 2018) and the GLD360 . One have to keep in mind that during the last years the performance of the ground based networks improved significantly and therefore the analyses of data between 2008 and 2014 may not provide information about the current LLS performance.
In April 2013 it was decided that a LIS, built as the flight spare for the TRMM satellite, should be put to the International Space Station (ISS). The data of this sensor, called ISS-LIS, was analyzed in Erdmann et al. (2019) for the time period March 2017 to March 2018. They compared ISS-LIS data to the low-frequency LLS of Météorage and the lightning mapping array SAETTA  over Corsica.
In this paper the performance of EUCLID (EUropean Cooperation for LIghtning Detection), a ground based LLS, will be 70 evaluated relative to the ISS-LIS data. This work is timely, given that the Meteosat Third Generation (MTG), which has a lightning imager (LI) on board, is going to be launched in 2 years.

EUCLID
Starting in 2001 the European Cooperation for Lightning Detection (EUCLID) geolocates cloud-to-ground (CG) strokes and 75 intracloud (IC) pulses through a combination of time-of-arrival (TOA) and direction finding (DF) techniques. The EUCLID cooperation is special in the sense that it combines raw sensor data in real-time of independent lightning location systemseither managed by National Meteorological Services (NMS) or by private companies -within a single processor. This is possible since all of the sensors operate in the same low-frequency (LF) range and are from the same manufacturer, i.e., Vaisala. The central processor of EUCLID adopts individually calibrated sensor gains and sensitivities to account for any 80 local sensor site conditions. Those values can differ from the ones used by the local LLS provider due to the implicit higher redundancy in EUCLID as a result of the participation of additional sensors located outside the national borders in a neighboring country. Hence, it assures that the resulting lightning data are of high and nearly homogeneous quality throughout Europe. The performance of EUCLID has been frequently tested over the years in terms of its LA, DE and peak current estimation. Those performances have been determined either from direct lightning measurements at the Gaisberg 85 Tower (GBT) (Diendorfer et al., 2009), Peiβenberg tower in Germany (Heidler & Schulz, 2016) and Säntis tower in Switzerland (Romero et al., 2011;Azadifar et al., 2016) as well as from video and E-field records collected in different regions within Europe (Poelman et al., 2013;Schulz et al., 2016). The current LA is in the order of 100 m based on the location error directly measured at the GBT and based on video and E-field recordings within the majority of the network.
The DE for negative CG strokes/flashes reaches 70/96 % based on GBT data and is determined to be 84/98 % using video 90 and E-field records. On the other hand, the DE for positive events is greater than 84 % and 87 % for strokes and flashes, respectively . Finally, IC DE has been validated during the HyMeX experiment (Ducrocy et al., 2013;Defer et al., 2015) in the south of France Pédeboy et al., 2014). For this purpose, EUCLID observations were matched to the observations made by the Lightning Mapping Array "HyLMA". It is found that the DE of isolated IC flashes, i.e., pure IC flashes without any CG stroke in it, has a large variation ranging from 10 % up to 67 % from one 95 thunderstorm to another. This variability is mainly attributed to differences in the vertical extent of the IC flash and to the flash rates during a storm. Regarding the peak current estimates, EUCLID tends to overestimate those slightly with respect to https://doi.org/10.5194/amt-2019- 435 Preprint. Discussion started: 23 January 2020 c Author(s) 2020. CC BY 4.0 License. 4 the currents measured at the GBT with a median error of 4 %. More information regarding the performance and observations by the EUCLID network are found in Schulz et al. (2016) and Poelman et al. (2016). swath width of about 650 km of the Earth's surface. Hence, due to the continuous movement of the ISS with an orbital speed of approximately 7 km/s, lightning observations over a specific region lasts no longer than 90 seconds per overpass. When a lightning discharge occurs, the optical signal scatters throughout the cloud. In almost all of the cases, except in the unlikely case the cloud is extremely optically thick, this results in an extended area being light up on the top of the cloud when viewed from space. At the moment a pixel on the CCD array receives this optical pulse, the signal is compared to the 115 dynamically changing detection background threshold. Once this threshold is exceeded, the processor identifies this illuminated pixel as a LIS event. It is important to note that a LIS event has no counterpart when compared to the observations made by a ground-based LLS such as EUCLID. However, the collection of LIS events from adjacent pixels during the same 2 ms frame integration time, defined as a LIS group, is comparable with either a CG stroke or cloud pulse.

ISS-LIS
Note that the ISS-LIS group location is the radiance-weighted centroid of all the events within the respective group (Mach et 120 al., 2007). In its turn, groups are clustered within a flash when the spatial and time criteria of 5.5 km and 330 ms, respectively, are met. In contrast to EUCLID, LIS is not able to distinguish between CG and IC lightning. Nevertheless, Boccippio et al. (2002) estimated an upper bound for the TRMM-LIS total flash DE of 88 ± 9 %.
In this work, we make use of the non-quality controlled ISS-LIS dataset (Publication date: 2019-08-19, Version 1, Processing level 2) made available by the NASA Global Hydrology Resource Center DAAC. This includes information on 125 geolocated and time-tagged lightning events, orbit statistics and metadata. For more in depth information on the LIS instrument, the interested reader is referred to Christian et al. (1989), Blakeslee et al. (2014) and Blakeslee & Koshak (2016).

3 Methodology
In this paper EUCLID and ISS-LIS lightning observations are correlated using data in between March 01, 2017 until March 31, 2019, as observed within the EUCLID domain and limited to 54° north. ISS-LIS detects optically bright discharges, such 130 as return strokes and in-cloud discharges inducing a rapid change in the electric field (Goodman et al., 1988). Those rapid changes in the electric field are exactly the features detected by EUCLID. Hence, the fundamental unit of ISS-LIS, i.e., groups, and EUCLID, i.e., CG strokes and IC pulses, largely corresponds to the same physical process and is therefore directly comparable . Additionally, the comparison will be performed as well on the artificial derived flash level. 135 The approach taken in this work has been applied and described in detail in Rubinstein (1994) and , in which a probabilistic method is used to estimate the relative and absolute detection efficiencies of both systems under investigation. The concepts are briefly defined below. Neither EUCLID nor ISS-LIS observe all of the lightning activity that actually occurred at a given moment in time. Hence, let S be the set of all occurred lightning discharges and A and B the set of discharges detected by ISS-LIS and EUCLID, respectively, as illustrated in Figure 1. Note that it is possible that both of 140 the systems contain some false alarm detections and therefore fall outside S. However, those false alarms constitute roughly 1 % of the total amount of events detected by EUCLID (Poelman et al., 2017), whereas the false event rate requirement for LIS is set to be less than 5 %. Hence, the latter has a minor influence on the final outcome. The system dependent relative detection efficiencies can be expressed as: with n A and n B the amount of discharges detected by system A and B, respectively, and n A ⋂n B the intersection, containing discharges detected by both systems. Thus P(A|B) represents the conditional probability that LLS A detects a discharge relative to LLS B, and vice versa in case of P(B|A). In addition, the true detection efficiency, for example of system A, reads 150 as: However, the actual amount of occurred discharges n S is not known a priori. Therefore, the estimated absolute detection efficiencies (ignoring false detections) can be calculated in the following way: 155 https://doi.org/10.5194/amt-2019- 435 Preprint. Discussion started: 23 January 2020 c Author(s) 2020. CC BY 4.0 License.
Since the number of detections in S, n S , is larger than the unique set of combined discharges in A and B, the estimated 160 absolute DE is an upper limit for the true detection efficiency. However, to precisely calculate the above detection efficiencies only those EUCLID discharges that occurred within the ISS-LIS fov, n B , need to be taken into account. To this end, the corner points of two consecutive ISS-LIS fovs, separated by roughly 35 seconds, are extrapolated to every second to increase accuracy. Then, for each second the EUCLID detections are extracted within the respective fov. As an example, the ISS-LIS groups and EUCLID CG strokes and IC pulses are plotted on top of the ISS-LIS fov in Fig. 2. This is the biggest 165 difference compared to the future MTG-LI observations from a geostationary orbit. Next, the individual EUCLID CG strokes and IC pulses are correlated in time and space with the ISS-LIS groups in order to retrieve the amount of overlapping detections. A match is found when the temporal and spatial criteria of 10 ms and 20 km, respectively, are fulfilled. Those particular criteria have been used in similar inter-comparison studies such as Franklin (2013), Zhang et al. (2016;. Note that only one LIS group can be matched to a single EUCLID discharge and vice versa. At the flash 170 level, matching is somewhat more complicated due to the fact that EUCLID and ISS-LIS have their own specific flash clustering algorithms. For the flash analysis EUCLID strokes/pulses are matched to ISS-LIS groups using a larger temporal (100 ms) and spatial (30 km) criterion to account for the fact a flash can consist out of different discharges spread over some time interval. Subsequently, the strokes/pulses and groups are traced back to the respective flash it belongs to. Thus a matched flash can have one or multiple matched discharges or groups. Note that since the flash grouping algorithms between 175 EUCLID and ISS-LIS are different, the matched flash count n A ⋂n B , used in Eq. 4 and 5, is slightly different depending on the network, even though the matched count of EUCLID discharges and ISS-LIS is the same.

EUCLID stroke/pulse and ISS-LIS group level
In Fig. 3 the spatial distribution of the IC:CG ratio, observed by EUCLID, is plotted at the level of the IC pulses and CG 180 strokes, as well as at the flash level. Only data within the EUCLID domain as indicated by the dashed polygon and cut off at 54° north to account for the ISS-LIS latitudinal coverage are used for quantitative analysis in this work. The geographic spread does not reflect the actual IC:CG occurrence within Europe, but mainly highlights areas where EUCLID is more capable detecting IC activity over others due to sensor technology. Not surprisingly, the highest IC:CG ratios are found in regions where the baseline between LS700x sensors is small and drops off towards the south and east of the domain where 185 https://doi.org/10.5194/amt-2019-435 Preprint. Discussion started: 23 January 2020 c Author(s) 2020. CC BY 4.0 License. mainly IMPACT sensors were installed during the period of investigation. Additionally, during this time period, in the south of Italy and in Spain significant communication problems deteriorated the results. The mean IC:CG stroke (flash) ratio over the entire region is 2.6 (1.9) and increases to 4.1 (2.8) within the rectangle highlighted in white. The rectangle highlighted in white in Fig. 3 will be referred to as the centre of the EUCLID network throughout the paper. The mean IC:CG flash ratio in the centre of the network is comparable to the values observed by the U.S. National Lightning Detection Network (NLDN) 190 in various parts throughout the contiguous United States as presented by Medici et al. (2017). Since EUCLID observes most of the IC pulses in the centre of the network, in the remainder of the paper results will be presented for the full domain, as well as for the centre where indicated.
The distance offset Δd in one kilometre intervals between matched EUCLID detections and ISS-LIS groups is indicated in Similar as with the distance offset, the timing differences Δt, calculated here as t ISS-LIS -t EUCLID , can be calculated between matched discharges detected by both systems. The distribution of the time offset between matched ISS-LIS groups and EUCLID pulses is indicated in Figure 5. A negative (positive) value indicates that the ISS-LIS group occurred earlier (later) than the EUCLID match. It is found that the distribution peaks sharply around ±1 ms, with a mean (median) time offset 205 of -1.3 ms (-0.6 ms). Thus, on average an ISS-LIS group occurs first. Nonetheless, the majority of the timing differences fall inside the ISS-LIS timing accuracy set by the frame integration time of 2 ms. Unlike for the distance difference, the time offset does not differ within the EUCLID domain.
The estimated peak current of matched EUCLID CG strokes (solid line) and IC pulses (dashed line) are correlated with the ISS-LIS group radiance in Figure 6. Note that positive discharges with peak currents smaller than 10 kA are likely to be 210 misclassified as CG strokes because those are more likely to be of intracloud nature Wacker & Orville, 1999a;b;Jerauld et al., 2005;Orville et al., 2002;Biagi et al., 2007). Hence, positive CG strokes below 10 kA are all categorized as IC pulses and therefore no data for positive CG below 10 kA exist. Additionally, the largest positive IC pulse in the EUCLID dataset has an estimated peak current of 28.8 kA, limiting the positive IC pulse curve in the plot. In general, higher peak current signals observed by EUCLID correspond with higher ISS-LIS group radiances. At larger absolute peak 215 current values, i.e., |I p | ≥ 20 kA, the correlation becomes more variable. However, the latter is an artefact of the sample size, as indicated by the grey curve in the plot.

Flash level 230
The spatial distribution of the EUCLID and ISS-LIS flash counts is indicated in the upper plots of Figure  baselines.. Additionally, note that in Zhang et al. (2016), the NLDN observations used were restricted to the areas where the NLDN detection efficiency is highest. Hence, P(EUCLID) of 69.4 % in this work should be compared to the 58.2 % in Zhang et al. (2016). From Figure 7 it is found that P(ISS-LIS) is highest outside the center of the EUCLID network, while it 255 is the opposite in case of P(EUCLID). Note that, contrary to what is found in this study, the absolute DE for ISS-LIS should be uniform over the entire region, since it is highly unusual to expect a geographic dependence. It is believed that the spatial dependence is related to the limit of the Bayesian algorithm using only two networks. Probably making use of additional networks would make it disappear.
Average characteristics of all ISS-LIS flashes, those observed (matched) and not observed (unmatched) by EUCLID are 260 listed in Table 1, while Figure 8 provides relative detection efficiency as a function of those characteristics. Note that the ISS-LIS flash characteristics as in Table 1 do not differ much throughout the EUCLID domain. Therefore those latter values are averages over the full domain. From Table 1, it is found that the duration of ISS-LIS flashes is longer by a factor of 1.5 for those that have a match with a EUCLID flash compared to those not observed by EUCLID. Additionally, Fig. 8a demonstrates

Conclusions 300
There exists a multitude of different technologies to detect and locate the electrical activity in thunderstorms, whether on a local, continental or global scale. This leads to various sets of observations of the same phenomenon. Hence, depending on the requirements, e.g., spatial accuracy and/or extent, the user can favor one system over the other. Nevertheless, it is important to investigate the similarities and differences among different systems. Incited by the forthcoming launch of the Meteosat Third Generation geostationary satellites, with onboard the Lightning Imager, this study aims at comparing for the 305 first time over a large area in Europe the lightning observations from the ground-based EUCLID network to the optical signals detected by the space-based ISS-LIS. The analysis is based on the lightning activity recorded during March 01, 2017 and March 31, 2019 within the EUCLID network and limited to 54° north. In this study the EUCLID cloud-to-ground strokes and intracloud pulses are compared to the ISS-LIS groups, in addition to the correlation at the level of the flashes of both systems. Besides measuring the temporal and spatial differences between matched observations a Bayesian approach is 310 adopted to determine the relative and absolute detection efficiencies (DE) of each system. It is found that the matched EUCLID strokes/pulses and ISS-LIS groups are separated by a median distance of 4.8 km, corresponding to approximately two ISS-LIS pixels in the CCD imager. A negative median time difference, t ISS-LIS -t EUCLID , between matched discharges of 0.6 ms is well within the time accuracy of ISS-LIS and indicates that on average the ISS-LIS group occurs first. On the other hand, higher peak current signals observed by EUCLID correspond with higher ISS-LIS 315 group radiances. The ISS-LIS group relative DE is 34.3 % and drops slightly to 33.9 % in the centre of the EUCLID network. The latter values are in contrast to the much lower EUCLID stroke/pulse relative DE values of 8.1 % and 11.3 % over the full domain and in the centre of the network, respectively. This is related to the higher amount of ISS-LIS groups       *Multiplicity here means the number of strokes, pulses or the sum of both in a pure CG, IC, or hybrid flash, respectively.