© P. Coïsson et al., Published by EDP Sciences 2025

1 Introduction

Since the end of 2013, the three satellites of the European Space Agency (ESA) Earth Explorer Swarm mission (Friis-Christensen et al., 2006) measure the magnetic field of the Earth from polar Low Earth Orbits (LEO) at altitudes around 450 km and 510 km, with an inclination of 87.75° and 87.35°, respectively. The local time of their ascending nodes drifts by −5.4 and −5.1 min every day. This study focuses on data from one of the Swarm instruments, the Absolute Scalar Magnetometer (ASM), (Léger et al., 2015), the nominal role of which is to provide 1 Hz absolute scalar values for calibration of the Vector Field Magnetometer (VFM) fluxgate vector magnetometer that completes the magnetometry payload of the mission and provides the nominal vector data. The ASM instrument exploits the Zeeman effect on the energy levels of the electrons of helium atoms to precisely and rapidly measure the absolute magnitude of the ambient magnetic field. Internally, it applies a cascade of filters to produce its nominal 1 Hz data. The possibility of producing data in burst mode at 250 Hz sampling was further included, initially with the objective of acquiring test data for performance assessment during the in-orbit commissioning phase of the satellites. A second experimental vector mode was also implemented. This second mode takes advantage of a set of orthogonal coils placed around the helium cell to impose three orthogonal modulations, each with its own distinct and well-controlled frequency, close to 10 Hz. A specific processing then allows the recovery of the three orthogonal components of the field every second, in addition to the nominally requested absolute scalar data (Gravrand et al., 2001; Leger et al., 2009). These two experimental operation modes are mutually exclusive. As the experimental vector mode quickly proved to provide vector data of interest for both nominal vector data quality monitoring and global field modelling (Hulot et al., 2015; Vigneron et al., 2015, 2021), this mode became the default operation mode on all satellites, and the vector mode data are now distributed to the scientific community (Hulot et al., 2025a, 2025b). A few burst mode campaigns lasting from some hours to two days were nevertheless conducted between December 2013 and February 2014 on each Swarm satellite: A, B, and C. Although each satellite embarks two ASM for redundancy, the redundant ASM unit on Swarm C did not survive launch, and its only remaining ASM failed on 5 November 2014. The cause of this failure was later identified as being a microcircuit permanently damaged by a heavy ion (Fratter et al., 2016). As the scientific value of the data acquired during the initial burst mode campaigns became clear, it was decided to acquire new data in this mode. Since August 2019, monthly campaigns of one week each have been conducted, initially on only one satellite at a time. Later, from July 2021, monthly one-week campaigns have been run on both Swarm A and B. The vast majority of the more than a hundred campaigns conducted so far have been successful, and new campaigns are now being run regularly.

The main scientific objective of these campaigns is to take advantage of the burst mode data to study electromagnetic phenomena in the Extremely Low Frequencies (ELF), between 10 and 125 Hz. Quite a few natural and artificial signals have already been observed in these data (Emsley et al., 2025). However, the main motivation for running regular campaigns is to investigate the numerous lightning-generated whistlers. These are short signals, typically lasting less than 1 s, that result from the propagation through the ionosphere of the short electromagnetic impulse produced by a lightning strike. The dispersive nature of the ionosphere produces a separation of the frequency components of the signal, with the highest frequencies travelling faster. As a result, a whistler detected from space is perceived as a lowering tone. When translated into sound in the audible part of the spectrum, these signals are perceived as a whistle, hence the whistler name (e.g., Eckersley, 1935; Storey, 1953; or the review Al’pert, 1980).

Whistler signals detected in the ASM burst data were first investigated in the context of the ESA-supported feasibility project entitled “Investigating lightning generated ELF whistlers to improve ionospheric models (ILGEW)” (Coïsson et al., 2021). This study demonstrated the unique capabilities of the Swarm mission in providing high-quality data in the ELF, a frequency band rarely monitored by other space missions. Swarm satellites also have the characteristics of being in near-polar orbits, slowly drifting in local time, enabling the possibility to cover the full range of geographic locations and local times. As lightning activity on the surface of the planet strongly depends on the position of continental masses and on local time, sources of whistler signals are not homogeneously distributed in space and time. They tend to concentrate in clusters, where active storms produce a large number of strong lightning. Each Swarm satellite samples the whole range of latitudes at a specific local time sector on any given day, but by acquiring data month after month, good coverage of all local times could be achieved. In that context, we developed a new scientific data product for the Swarm mission, containing the list and characteristics of the whistler signals detected during burst mode campaigns. Its code name in the Swarm database is WHIxEVT_2_, and whistlers data are organised in daily files containing processed ASM data and whistler information.

The identification of whistler signals in long time series is a delicate task, because these signals occur randomly in the data. Indeed, lightning strikes occur somewhere on the Earth’s surface at any given second, but only a limited number of strikes produce detectable whistler signals (e.g., Collier et al., 2006, 2009; Jacobson et al., 2016). For an operational processing, it was necessary to implement an efficient procedure. An increasing number of processing algorithms for detecting specific signals in time series have been successfully developed using artificial neural network techniques. For our goal of whistler detection, we took our inspiration from the algorithm for detections of Ultra Low Frequency (ULF) events in Swarm data developed by the Observatory of Athens (Balasis et al., 2019), who introduced us to this analysis technique. It used an Adaptive Network-based Fuzzy Inference System (ANFIS) (Jang, 1993), and the details of our implementation for whistler detection are provided in Section 3.

Several other approaches specifically developed for whistler detection have been successfully applied to various space missions, mostly based on the analysis of the spectral content: e.g., Detection of Electro-Magnetic Emissions Transmitted from Earthquake Regions (DEMETER) (Elie et al.,1999; Parrot et al., 2019), Akebono (Dharma et al., 2014), Communication/Navigation Outage Forecast System (C/NOFS) (Jacobson et al., 2016), China Seismo-Electromagnetic Satellite (CSES) (Yuan et al., 2021, 2022; Recchiuti et al., 2025) or Arase (Suarjaya et al., 2024).

Most of these studies focused on the analysis of whistler signals in the Very Low Frequency (VLF) domain, above the lower hybrid frequency, for which the propagation properties in the ionosphere are slightly different from those in the ELF. In many cases, there is the possibility of identifying the lightning strike that originated a specific whistler, mostly by comparing the temporality of these events (e.g Jacobson et al., 2016) or by including the calculation of the propagation path and time through the ionosphere before reaching the satellite (Chum et al., 2006). These studies make use of lightning identification from ground VLF measurements, provided by the World Wide Lightning Location Network (WWLLN) (Hutchins et al., 2012). We conducted some preliminary studies on the identification of whistler-lightning pairs, using WWLLN in the VLF, and ELF ground data from the World ELF Radiolocation Array (WERA) network (Mlynarczyk et al., 2017), but did not perform a systematic analysis of the whole Swarm data set. This analysis nevertheless showed that in most cases, whistlers are detected within 2,000 km distance from the lightning strike, and occasionally up to several thousands kilometres away (Coïsson et al., 2021).

When propagating through the ionosphere between its base and the Swarm satellite, the whistler signal gets dispersed due to the properties of the medium. The relation between a whistler frequency and its arrival time was initially identified by Eckersley (1935) in the frequency range between 400 Hz and 4 kHz, and can be mathematically represented as:

$$D=\sqrt{f\left(t\right)}(t-t_0),$$(1)

where D is the whistler dispersion, f(t) the instantaneous frequency, t the time and t ₀ the time when the lightning signal entered the ionosphere. Recently, Jenner et al. (2024) demonstrated that, in the ELF range, this dispersive effect is mainly proportional to the square root of the local plasma density and that its effect at first order cumulates along the propagation path in a relatively simple way. The dispersion of the whistler detected by Swarm can thus be exploited to measure the amount of plasma crossed by the whistler signal. This quantity, measured in the ELF, has been named Total Root Electron Content (TREC). This new concept shows that, with the right assumptions, lightning-generated ELF whistlers can be used to passively sound the ionosphere below LEO. We present here the ASM burst acquisition campaigns, the processing algorithms used to identify and characterise whistlers, and the statistics of ELF whistlers occurrences.

2 Campaigns of ASM burst mode acquisition

After the initial burst mode campaigns of 2014, this mode was not activated for several years. Test burst campaigns were run in 2018 for Swarm A and in early 2019 for Swarm B, confirming the scientific value of such data and the possibility of acquiring more data in this acquisition mode. Some adjustments in the housekeeping data collected onboard and in the ground processing of the nominal Level 1b magnetic data were subsequently implemented to allow operation of burst mode campaigns on a more regular basis. One week of continuous burst mode acquisition at a time was then identified as suitable to keep good enough control of the relative time provided by the internal clock of the instrument. Indeed, since the experimental burst mode was not originally planned to run over long periods of time, a complete continuity of data collection throughout the acquisition campaign is needed to maintain accurate time control. If data packets are lost during acquisition, reliable correction of the timestamp of subsequent data becomes challenging, rendering that data difficult, if at all possible, to exploit. This was found to usually not happen if the campaigns last one week or less. So far, such a problem has only occurred on a couple of occasions.

At the instrument level, time synchronisation is obtained via the onboard time distribution service based on the Global Positioning System (GPS), providing a pulse-per-second (PPS) signal. The internal clock of the ASM is controlled by a quartz oscillator. When the ASM is operated in vector mode, the PPS is used to adjust the timestamp of the data every second. When the acquisition is switched to burst mode, this PPS adjustment is not performed, and packets of 250 data samples are recorded for each second, counted using the ASM internal oscillator. However, the instantaneous frequency of the oscillator is monitored and available in the ASM housekeeping data. The burst mode data are produced at the slightly faster rate of 250.007 Hz on average, implying that the time interval between two consecutive values is close to 3.9999 ms. During burst mode campaigns lasting several days, this produces an increasing temporal gap between the start of data packets and the round second. For the post-processing of these data, it thus appeared necessary to implement a timestamp correction and achieve a more precise temporal resolution to date burst mode data, capable of recording higher precision than the 1 ms resolution available in the Common Data Format (CDF) CDF_EPOCH time format, which is used for the timestamps of the other Swarm scientific data (Space Physics Data Facility, 2023). To achieve this, the timestamp calculation during the burst mode data processing takes into account deviations relative to the internal frequency of the ASM, as measured at each PPS signal from the GPS. This approach minimises cumulative precision errors. Two timestamps are maintained in the data files: One provides timing accuracy down to the millisecond, and the other records the fractional component of the same timestamp, rounded to the nearest second, providing timing accuracy up to the nanosecond (Chauvet & Hulot, 2024).

ASM burst data that have been processed to identify, characterise and distribute whistler events are summarised in Figure 1, covering the period from May 2019 to December 2024. As already stated, ASM burst data keep being collected every month. Acquisition campaigns have been planned to pursue a complete coverage of geospatial conditions. We were able to meet the first goal of acquiring data for all 24 local time (LT) sectors in 2022 and now the local time coverage has been reached for all 4 seasons (December to February; March to May; June to August; September to November), with a minimum of at least 4 days of acquisition in a local time–season bin. Some configurations have been surveyed on several occasions, and the maximum reaches 14 days in the data set presented here. The local times close to the solar terminator (6 and 18 LT) are those surveyed the largest number of days; sunset corresponds to the temporal sector with the most numerous whistler events, as illustrated in Section 5. It is worth pointing out that the local time drift of the Swarm A and B satellites is slightly different, and that their local time separation changes slowly, following their orbital planes’ position. When we started the ASM burst campaign, the orbital planes were progressively getting closer, and coplanarity occurred in October 2021. The satellites were counter-rotating, sampling the same local time sectors, and getting very close to each other at some point along their orbit. We took advantage of this unique configuration to acquire ASM burst data simultaneously on both A and B satellites during the monthly campaigns conducted between October 2021 and May 2022. Later, we conducted many campaigns, including one day of overlap between the acquisitions of the two satellites. All of these campaigns were conducted to study the occurrences of common whistler detections originating from the same lightning strike. More recently, as the longitudinal separation between the satellites’ orbital planes increased, campaigns have again been conducted for each satellite on separate weeks. From a solar point of view, 2019 was a year of low solar activity, thus the acquisition campaigns have covered so far the entire rising phase of the solar cycle and the current maximum.

Figure 1 Amount of days of ASM burst mode acquisition, cumulated between May 2019 and December 2024, as a function of local time at the location of Swarm satellites, for each four seasons.

3 Algorithms for whistler identification and characterisation

Whistlers were initially identified by visual inspection of ASM burst data spectrograms of the 2014 campaigns that lasted one or two consecutive days. At first, whistler identification and characterisation were done entirely manually, with an operator visually screening the data in batches of 1 minute length: 1,440 diagrams had to be inspected for each day, while events were expected in about 10% of them. This approach was not suitable for routinely processing new ASM burst mode campaigns. To reduce workload and speed up data analysis, it was decided to implement new algorithms for automatic detection of candidate whistler events. The only input data needed for whistler identification are the Swarm ASM burst Level 1b daily data files: The ASMxBUR_1B product in the Swarm database. Details of these data files are available in Chauvet & Hulot (2024).

The process of screening burst data to detect and characterise whistler signals is accomplished in several steps, summarised in Figure 2. It involves a neural network processing and manual operations. This diagram also includes the characterisation step and the optional procedures used for adding newly processed data to the database to retrain the neural network algorithm we implemented. The algorithms for automatically identifying time intervals with whistler-like signals and the subsequent manual characterisation of whistlers are detailed here.

1. Starting from the daily ASM burst Level 1b files, a neural network algorithm of type ANFIS analyses the time series in 3 s time windows to classify them as candidate whistlers or non-candidates. Two outputs are produced: The list of candidate events and the list of minutes where candidate events occur.

2. The output of the previous step is used to manually characterise the whistlers, True Positive (TP) events of the ANFIS algorithm, to add and characterise False Negative (FN) events, and to reject False Positive (FP) events, producing daily files of all identified and characterised whistlers.

3. Optionally, to improve the performance of the ANFIS algorithm, the list of candidate whistler events can be compared with the list of characterised whistlers to increment a collection of TP, FP and FN events used to retrain the ANFIS. The last retraining was performed in March 2021.

Figure 2 Diagram of the processing steps to identify and characterise whistlers from Swarm ASM burst data. Detection is done automatically, characterisation manually.

3.1 Preprocessing algorithms

The analysis of the ASM burst data is performed one day at a time. The magnetic field measured by the ASM, F [nT], contains the total magnetic field, which is strongly dominated by the Earth’s main field, with values exceeding 20,000 nT, while whistler amplitudes are of the order of 0.1 nT. It was chosen to construct features for ANFIS analysis using parameters derived from the F time series and from the corresponding Amplitude Spectral Density (ASD) of F, in 3 s time windows. An overlap of 0.5 s is maintained between consecutive time windows, in order to reduce the probability of cutting whistler signals occurring near the limit of these windows, since for most cases whistlers last less than 1 s.

It is first necessary to remove large-scale trends: A polynomial of degree 4 is fitted using one minute of ASM burst Level 1b data centred on the 3 s time window and then subtracted from the total field to compute residuals F _R :

$$F_R\left(t\right)=F\left(t\right)-\sum _{k=0}^4{a}_k{t}^k,$$(2)

where a _k are fitted coefficients of the polynomial and t is time. Time derivatives of the residuals are also calculated:

$$\frac{d{F}_R\left(t_i\right)}{dt}=\frac{F_R\left(t_{i+1}\right)-F_R\left(t_i\right)}{t_{i+1}-t_i} .$$(3)

Working with time derivatives enhances the whistler signal with respect to the background noise. These residuals and their time derivatives are used to compute a first group of input features for the ANFIS analysis, see the first 10 rows of Table 1.

The time series of F _R is then used to calculate the evolution of the ASD using a Fast Fourier Transform (FFT) of order 32, over time windows of 0.13 s and keeping a 75% overlap between consecutive time windows. The timestamp provided for each computed FFT is the time in the centre of the time window. A total of 14 frequencies between 16 and 120 Hz are stored in the ASD array used for the feature calculation. The two lowest frequencies of the FFT have been discarded, since they are affected by instrumental noise; see Léger et al. (2015) for details. The remaining features for the ANFIS analysis are computed using these ASD values; see the last 5 rows of Table 1.

Additional information is extracted from the ASM burst Level 1b data from the flags that are associated to the time series (Chauvet & Hulot, 2024): If flag heater and flag magnetic are both active at the same time in the time window, meaning that an anomalous perturbation is possible on the magnetic field data, that time window is ignored for identifying candidate whistlers.

3.2 Automatic detection of candidate whistlers

To train ANFIS to detect whistlers, a data set of labelled time windows was constructed. It was necessary to include cases of time windows that contain whistlers and cases of time windows that do not contain whistlers. The initial data set that we used for training consisted of a selection of 768 whistlers identified in the ASM burst data acquired between January and February 2014, July 2018, and between January and September 2019. 2,315 non-whistler cases randomly extracted completed this data set, to maintain roughly a proportion of 25% whistler events and 75% negative events. It was chosen to include only whistlers with a complete spectral signature at frequencies between 16 and 120 Hz. Indeed, many of the whistlers found in the ASM data present either interruptions or truncations. To avoid selecting too many false positives, we decided to limit the inclusion of events that could be similar to ambient noise. Before starting training the system, a feature evaluation process was conducted to test several possible features. Individually, these features did not need to allow immediate identification of whistlers, because the training process on the data set would produce relations and dependencies between the various input features and the output labels. An initial set of test features was chosen following the work of Balasis et al. (2019) and other features were defined taking into account the characteristics of the whistler signals. To evaluate the usefulness of each feature, we examined the distributions of each feature separately for whistler and non-whistler events and retained those that showed clear separation between the two classes, shown in Table 1. With this selection, the dimensionality of the problem reduced from 751, corresponding to the number of ASM burst values in the 3 s time window, to 15, the number of selected features. These features are now routinely computed in the whistler identification process.

From the computational point of view, the ANFIS algorithm is composed of 5 different layers, separated into the premise, responsible for the decoding of the input features, and the consequence, in charge of computing, normalising, and encoding the output labels. The first layer of the ANFIS processing is the fuzzifier: All the input features are normalised in the range from 0 to 1 using input membership functions such as Gaussian functions. There is the possibility of setting a different function for each input feature. Next, some logic rules are defined to correlate the input features and the output. A literal example of this process can be summarised: if feature i and feature j then output k. It is possible to combine more than two features and use the logical operator or instead of and. Many different rules can be set, and ANFIS processes weighted linear combinations between features and outputs of the various rules. Lastly, the defuzzifier layer returns a unique final output between 0 and 1 representing the probability that the input time window contains a whistler, which is rounded, using 0.5 as a threshold, to produce the labelling choice of candidate whistler or non-whistler. During the training phase, the ANFIS searches for the best membership functions: Linear combinations of the input features and weighting coefficients that minimise the root mean square error (RMSE) of the difference between correct outputs from the training data set and the network outputs. These membership functions are calculated in the hidden layers of the ANFIS algorithm. The training strategy involves an initial partition of the available database of labelled time windows into three groups: Training, validation, and test. We randomly selected 70% of events for the training set, 20% for the validation set, and 10% for the test set. During the training phase, only the training data set is used to optimise the coefficients of ANFIS through an iterative minimisation process. At each iteration, the Root Mean Square Error (RMSE) values are computed for both the training and the validation data sets. The purpose is to check that both these RMSE values decrease simultaneously at every iteration of the optimisation. If the training RMSE decreases while the validation RMSE increases, the ANFIS is starting to over-fit, and the optimisation process should be stopped. At the end of the optimisation, the final coefficients are those that minimise the RMSE over the validation data set. After training, the performance of the trained network is evaluated on the test data set. When the first network was trained, there was still a limited amount of ASM burst data. As new acquisition campaigns were acquired in 2019 and 2020, we retrained the network by including cases from the new data. We added the new whistler events detected by ANFIS (TP), along with all other whistlers missed by the algorithm that were identified manually afterwards (FN). We also added non-whistler events, comprising 2/3 manually labelled false positive (FP), and 1/3 of randomly selected events. The latter were introduced to train the algorithm against any kind of signal present in the ASM data. The ratio between false positives and random events was chosen arbitrarily.

The confusion matrix of the ANFIS used for processing ASM burst data and identifying candidate whistler events is shown in Figure 3. From the training of ANFIS, we can expect that 86% of whistlers are correctly detected. The last retraining was realised in March 2021, and since then the production of candidate whistler events uses the same network. In the following manual validation step detailed in the next Section 3.3, false positive events are discarded and undetected whistlers are added to the final outputs. Some false positive examples are provided in Figure A.1 of the online Appendix A.

Figure 3 Confusion matrix of the trained ANFIS detection step.

3.3 Validation and characterisation of whistlers

The operational use of the ANFIS network output involves the validation step carried out by a human operator. Therefore, the presence of false positive cases in the selected data is mitigated by their immediate rejection in the validation step. Some false negative events can also be reintegrated during this process. This is facilitated by visual inspection of the ASD of one entire minute of ASM burst data. The operator systematically checks also the minute preceding and following a burst candidate event to search for false negative cases of undetected whistlers to include. This operation is done simultaneously with the subsequent characterisation of each detected whistler. The measured parameters are: Whistler dispersion D, initial time of ionospheric propagation t ₀, and whistler intensity. The dispersion of Eckersley’s curve (1) is adjusted to match the shape of the signal visible on the ASD plot. Various dispersion values are tested, varying in steps of 0.1 $\sqrt{\mathrm{s}}$. The resulting dispersion curve is also shifted in time, to find the highest correlation between the ASD and the dispersion curve (1). Estimates of t ₀ and its uncertainty are obtained from the analysis of the points at the intersection between the dispersion curve (1) and the tiles composing the ASD plot. Using the central time and central frequency of each tile, each intersection point leads to slightly different origin times. Hence, t ₀ has been defined as the average of these times. Similarly, t ₀ uncertainty has been defined as the highest time range between these times. Lastly, the intensity I [pT²/Hz] of a whistler is deduced as the sum of the squared values of each tile retained as a part of the whistler. It is obtained from:

$$I=\sum _{j=1}^N{p}_j^2,$$(4)

where N is the total number of selected tiles and p _j [pT/$\sqrt{\mathrm{Hz}}$] the ASD value of the j-th tile composing the whistler. See Figure 4, where the white dots follow Eckersley’s dispersion curve (1).

Figure 4 Example of whistler event as provided in the Swarm Level 2 data product. Top panel, residual magnetic field intensity after removing the large-scale trends with a degree 4 polynomial. Bottom panel: ASD of the signal, with superposed white dots indicating the tiles retained for whistler characterisation following Eckersley dispersion.

4 ELF Whistler data set

Whistlers detected in the ASM burst data are distributed in daily files as a Level 2 data product of the Swarm mission (Coïsson et al., 2023), through the ESA Swarm data servers. Whistlers occur randomly in Swarm data; therefore, there is a variable number of events recorded during one day. Moreover, the duration of a whistler signal is variable and depends on its dispersion. The provided list of whistler events includes three seconds of filtered ASM burst data centred on the arrival time of the 120 Hz component of the whistler, taken as the start time of the whistler signal in Coordinated Universal Time (UTC). The corresponding ASD that was used for whistler analysis is also included. An example of whistler data is shown in Figure 4. Some parameters characterising each whistler are provided: The Eckersley dispersion D [$\sqrt{\mathrm{s}}$], the estimated starting time of whistler propagation inside the ionosphere t ₀ [s] with respect to the whistler reference time and its estimated uncertainty [s], and the whistler intensity I [pT²/Hz] in the frequency range between 16 and 120 Hz. Additional parameters are: The position of the Swarm satellite at the whistler start time (latitude [°], longitude [°], radius [m]) and the local solar time [h] at this point. The general uncertainty on whistler dispersion has been estimated to be 0.4 $\sqrt{\mathrm{s}}$.

Many different situations are present in this data set. Whistlers often occur in rapid sequences and all are characterised individually. As an example, Figure 5 shows the characterisation of the second of the two whistlers separated by 0.5 s seen in the figure. In this particular case, there is no superposition of both whistlers, but for cases when whistlers have a large dispersion, the low-frequency components of the first one occur at the same time as the high-frequency component of the second one. For example, several whistlers can be seen in Figure 6, where the strongest whistler of our data set is presented. It occurred close to the geographic equator over Africa on the evening of 16 February 2022, at a time when intense whistler activity was taking place. In such cases, the rapid sequence of whistlers may prevent the possibility of characterising each of them. The background noise level of the ASM data also limits the detection of faint whistlers in the ELF range. In the example of Figure 6, three whistlers were characterised, starting at 19:42:04.1523, 19:42:04.7923 and 19:42:05.1763, but the noise level does not allow to individually identify other whistlers. Important to finally note is that the whistler data set only includes manually checked whistlers within the minute in which ANFIS detected an event, as well as within the preceding and following one minutes. Thus, very isolated events not detected by the ANFIS algorithm may be missed.

Figure 5 Example of whistlers occurring 0.5 s one after the other, following the same representation as the one of Figure 4.

Figure 6 Strongest whistler detected during ASM burst mode campaigns, on 16 February 2022.

5 Distribution of whistlers in space and time

In the period between May 2019 and December 2024 a total of 116,793 whistler events have been characterised in the ELF in Swarm ASM burst data. They are detected more frequently when the satellites pass close to regions where thunderstorm activity is strong. This pattern is clearly identified on the map of Figure 7. It shows the positions where whistlers were detected, indicating the total number of events that occurred in an area of 2° × 2° in latitude and longitude. We note that there is a marked asymmetry between the north and the south hemisphere and that whistlers are more frequent on continents or large archipelagoes than over large oceans. It appears that whistlers are predominantly detected at low latitudes, apart from a narrow band following the magnetic equator, where there is a significant drop in detections. One of the reasons for this drop can be found in the propagation properties through the ionosphere that tend to bend the wave vector toward the magnetic equator (see Jenner et al., 2024), preventing whistlers from propagating up to the heights of Swarm satellites, where the direction of the ambient magnetic field is close to horizontal.

Figure 7 Spatial distribution of the whistlers detected between May 2019 and December 2024.

There are three areas where the number of detections is particularly high: Above Central America, above the Congo basin, and above Southeast Asia. These regions roughly correspond to some of the places where intense lightning activity takes place; see Figure 8 showing the distribution of WWLLN lightning with energies above 1,000 J. The ELF electromagnetic pulse generated by a lightning strike can propagate hundreds of kilometres along the surface of the Earth, producing whistlers into the ionosphere along its propagation. This two-component propagation enables the detection of whistlers even thousands of kilometres away from the strike location. This can explain the fairly homogeneous distributions of whistlers over the oceans. There are other clear geographical differences between the locations of whistler detections and the locations of the strongest lightning strikes: The African continent presents a relatively small number of very energetic strikes compared to the number of whistlers detected over this region. In contrast, the large number of strong lightning strikes in South America produces a relatively smaller number of whistlers. A possible explanation for this behaviour can be related to the configuration of the magnetic field: In the longitudinal section between Africa and Asia, the magnetic equator is in the northern hemisphere, while the Amazon basin at the same latitude is below the magnetic equator. Seasonal maps are provided in Figures B.1, B.2, B.3 and B.4 of the online Appendix B showing a predominance of events in the summer hemisphere and more homogeneous distributions during the equinoxes.

Figure 8 Distribution of lightning strikes with energies above 1,000 J on days when at least one of the Swarm satellites was acquiring ASM burst data.

The number of whistlers detected during 24 h is highly variable, often changing by a factor of two, or even three, from one day to the next. For this reason, it was decided to acquire data during campaigns of one week duration, to increase the statistical significance of the observations. Looking at the median number of whistlers detected during a 7-days campaign according to the local times at the satellite position, it is possible to analyse trends in whistler activity. Polar satellites in near-Sun-synchronous orbit sample 2 specific local times during the ascending and the descending portions of their orbit. A strong asymmetry in the number of whistlers in these local times has always been observed in Swarm data. In contrast, the altitude difference of ~50 km between Swarm A and Swarm B does not appear to have a significant impact on either the number of whistlers detected or the amplitudes of those whistlers. As already mentioned, we have been progressively able to obtain data from all 24 h of the day, combining the measurements of both satellites that drift by one hour every 11 days. The statistics of the average number of whistlers per day, computed separately for each season, are shown in Figure 9. A very strong and coherent diurnal variation is observed, with a minimum of just a few detections around noon and averages around 15 during the daytime hours between 8 and 13 LT. Later in the afternoon, there is a progressive increase in events that reaches 100 at 17 LT. A strong maximum follows between 18 and 19 LT, with a large seasonal variability, reaching 230 to 430 whistlers per day. Night presents strong whistler activity, with large seasonal differences, 200–300 events per night. During the last hours of the night, the whistler activity decreases, with a sharp drop at dawn, when the whistler activity decreases by a factor of 2 every hour between 6 and 8 LT. This diurnal behaviour is partially similar to the distribution of lightning strikes, see Figure 10. Lightning occurrence is also minimal around noon and maximal during the night, but there are noticeable differences: There is just a factor of 3 between the minimum and the maximum number of lightning strikes, while for whistlers it varies between 3 and 10. Another noticeable difference is the afternoon maximum between 14 and 16 LT in lightning activity that is not seen in whistler data. This discrepancy suggests that the penetration of the lightning signal into the ionosphere is strongly affected by changes in the ionosphere during the day. Specifically, the lower ionospheric layers, the E and D regions, control the propagation of the lightning signal into the ground-ionosphere waveguide and the capability of the lightning signal to penetrate into the ionosphere.

Figure 9 Distribution of detected whistlers as a function of local time at the location of Swarm satellites for each four seasons.

Figure 10 Distribution of WWLLN lightning strikes as a function of local time during the days when ASM burst campaigns have been acquired.

The occurrence of whistlers appears to be partially correlated with the solar cycle. In the years 2023 and 2024, during the solar maximum, the night-time median number of whistlers is generally larger than in 2020, which was at solar minimum. Figure 11 shows the statistics for a binning in 8 LT sectors of the median number of whistler detected during the 7 days of each ASM burst campaign in the corresponding local times of the ascending and descending part of the orbit. These median values are compared to the daily sunspot number, which displays a large day-to-day variability. The tendency of increasing whistler occurrences with increasing solar activity is particularly clear during the night, while the day always presents a small number of events, and no particular tendency can be observed. We checked that these tendencies are robust independently of the season in the year, see also Figures C.1, C.2, C.3 and C.4 of the online Appendix C. The largest variability of whistler detections is seen around sunset, at any level of solar activity, but the highest values are obtained during the years of solar maximum. During the night, the lower layers of the ionosphere almost disappear at around 100 km altitude, and the tendency with solar activity is to maintain a more developed ionosphere throughout the night. Swarm data seem to indicate that these conditions are more favourable for ELF penetration and whistler detection.

Figure 11 Median over each burst mode campaign of the number of whistlers detected in a local time sector per day, compared with the daily sunspot numbers. Each panel contains data sampled in a local time sector of 3 h.

Most of the acquisition campaigns were carried out during geomagnetic quiet conditions, and no significant impact of geomagnetic activity was observed on whistler occurrences. As an example, on 10 and 11 May 2024, the last two days of an ASM burst campaign on Swarm A, when an extreme geomagnetic storm occurred (Hayakawa et al., 2025), which only resulted in a higher level of FP detections by ANFIS but no significant change in the number of characterised whistlers. See also Figure D.1 of the online Appendix D.

6 Conclusions

ASM burst data have been acquired for one week every month by two Swarm satellites for more than 5 years, and new acquisition campaigns will be acquired during the rest of the mission. We were able to process these data to form a catalogue of events that now contains more than 100,000 events. Whistler signals in the ELF often have a very small amplitude, thus it is not possible to clearly identify all of them, whether using manual or automatic processing. However, this data set covers the whole globe at any local time, while most of the other space missions cover either a limited range of local times (e.g., DEMETER and CSES) or latitudes (e.g., C/NOFS).

New efficient data analysis techniques based on artificial neural networks are constantly being developed. Some have already been applied in the context of Swarm, with similar aims of detecting oscillating signals, but in other frequency domains (Antonopoulou et al., 2022). With the purpose of improving the overall performances of our current algorithms, we recently started exploring other kind of artificial neural networks, taking also advantage of the now much larger data set of whistlers available. Initial results are promising, but additional significant work is needed to obtain an improvement that could be included in the Swarm whistler processing chain to increase its reliability and reduce manual processing time. Strong diurnal variability of whistler occurrences has been detected, partially following the diurnal distribution of lightning strikes, as detected by WWLLN. Not all strong lightning strikes occurring close to the satellite location produce detectable whistlers in the ELF. More studies are needed to fully understand the penetration conditions of ELF signals from the neutral atmosphere to the ionosphere. Data acquired from solar minimum to solar maximum during the rising phase of the current solar cycle showed an increase in the number of whistler events during the night. New data will be acquired during the upcoming solar cycle declining phase to continue investigating whistler occurrences and their possible relation with solar activity.

Acknowledgments

The ASM instrument was developed by CEA‐Léti and provided as Customer Furnished Instrument by CNES. The authors acknowledge the ESA Swarm team for the support provided in acquiring ASM burst data and for the successful operations of the mission. The authors wish to thank the World Wide Lightning Location Network (http://wwlln.net), a collaboration among over 50 universities and institutions, for providing the lightning location data used in this paper. The editor thanks two anonymous reviewers for their assistance in evaluating this paper.

Funding

The development of the WHIxEVT_2_ production chain was started during the project “Investigating Lightning Generated ELF Whistlers to improve ionospheric models” (ILGEW), ESA contract No. 4000126708/19/NL/IA and implemented under ESTEC contract No. 4000109587/13/1-NB (DTU sub‐contract SW‐CO‐DTU‐GS‐010). This project was financially supported by the Centre National d’Etudes Spatiales (CNES) as an application of the Swarm mission, through the project “Suivi et exploitation de la mission Swarm”.

Conflicts of interest

The authors declare no conflict of interest.

Data availability statement

Swarm whistler data can be accessed from https://swarm-diss.eo.esa.int/#swarm/Level2daily/Entire_mission_data/WHI/EVT. Sunspot data are available from WDC-SILSO, Royal Observatory of Belgium, Brussels, DOI: https://doi.org/10.24414/qnza-ac80. Lightning data are available at a nominal cost from http://wwlln.net.

Supplementary material

Appendix A: Example of false positive events.

Appendix B: Seasonal statistics of whistler occurrences.

Appendix C: Whistler occurrences during the ascending phase of the solar cycle.

Appendix D: Whistlers during extreme geomagnetic storm.

References

All Tables

All Figures

	Figure 1 Amount of days of ASM burst mode acquisition, cumulated between May 2019 and December 2024, as a function of local time at the location of Swarm satellites, for each four seasons.
In the text

	Figure 2 Diagram of the processing steps to identify and characterise whistlers from Swarm ASM burst data. Detection is done automatically, characterisation manually.
In the text

	Figure 3 Confusion matrix of the trained ANFIS detection step.
In the text

	Figure 4 Example of whistler event as provided in the Swarm Level 2 data product. Top panel, residual magnetic field intensity after removing the large-scale trends with a degree 4 polynomial. Bottom panel: ASD of the signal, with superposed white dots indicating the tiles retained for whistler characterisation following Eckersley dispersion.
In the text

	Figure 5 Example of whistlers occurring 0.5 s one after the other, following the same representation as the one of Figure 4.
In the text

	Figure 6 Strongest whistler detected during ASM burst mode campaigns, on 16 February 2022.
In the text

	Figure 7 Spatial distribution of the whistlers detected between May 2019 and December 2024.
In the text

	Figure 8 Distribution of lightning strikes with energies above 1,000 J on days when at least one of the Swarm satellites was acquiring ASM burst data.
In the text

	Figure 9 Distribution of detected whistlers as a function of local time at the location of Swarm satellites for each four seasons.
In the text

	Figure 10 Distribution of WWLLN lightning strikes as a function of local time during the days when ASM burst campaigns have been acquired.
In the text

	Figure 11 Median over each burst mode campaign of the number of whistlers detected in a local time sector per day, compared with the daily sunspot numbers. Each panel contains data sampled in a local time sector of 3 h.
In the text