TROPOMI enables high resolution SO2 flux observations from Mt. Etna, Italy, and beyond

The newly launched imaging spectrometer TROPOMI onboard the Sentinel-5 Precursor satellite provides atmospheric column measurements of sulfur dioxide (SO2) and other gases with a pixel resolution of 3.5 × 7 km2. This permits mapping emission plumes from a vast number of natural and anthropogenic emitters with unprecedented sensitivity, revealing sources which were previously undetectable from space. Novel analysis using back-trajectory modelling of satellite-based SO2 columns allows calculation of SO2 flux time series, which would be of great utility and scientific interest if applied globally. Volcanic SO2 emission time series reflect magma dynamics and are used for risk assessment and calculation of the global volcanic CO2 gas flux. TROPOMI data make this flux time series reconstruction approach possible with unprecedented spatiotemporal resolution, but these new data must be tested and validated against ground-based observations. Mt. Etna (Italy) emits SO2 with fluxes ranging typically between 500 and 5000 t/day, measured automatically by the largest network of scanning UV spectrometers in the world, providing the ideal test-bed for this validation. A comparison of three SO2 flux datasets, TROPOMI (one month), ground-network (one month), and ground-traverse (two days) shows acceptable to excellent agreement for most days. The result demonstrates that reliable, nearly real-time, high temporal resolution SO2 flux time series from TROPOMI measurements are possible for Etna and, by extension, other volcanic and anthropogenic sources globally. This suggests that global automated real-time measurements of large numbers of degassing volcanoes world-wide are now possible, revolutionizing the quantity and quality of magmatic degassing data available and insights into volcanic processes to the volcanological community.

Scientific RepoRts | (2019) 9:957 | https://doi.org/10.1038/s41598-018-37807-w as detailed in Material and Methods. Westerly winds dominated and caused southeastward dispersing, elongated plumes ( Fig. 2a,b), until 4-7 August when major portions of the plume dispersed towards North West and North of Mt. Etna, an area not covered by the FLAME scanners (Fig. 1c). This explains the absence of FLAME data for days after 4 August (Fig. 3a). The SO 2 flux averaged over all data of the full month for TROPOMI data is 2.83 ± 1.66 kt/day and agrees with the average for FLAME (2.39 ± 1.09 kt/day). The daily averages of SO 2 fluxes from FLAME and TROPOMI agree within the uncertainty for 25 out of 31 days (81%), but strong intra-day variability and errors mean the degree of linear correlation is relatively low (Pearson's correlation coefficient 0.31). It is particularly encouraging that the trending between the data is compatible (Fig. 3a), such as the periodicity of ~7 days seen in both datasets, and the slight decrease in SO 2 flux after 26 July. Also evident is an unusually low SO 2 degassing strength between 5 and 7 August that both approaches picked up (Fig. 3a). The mean difference between fluxes derived from TROPOMI and FLAME is −0.18 kt/day (Fig. 3b). The absolute flux difference is 0.99 kt/day on average (mean absolute error, MAE, Fig. 3b).
Although for some days the plume shape indicated a relatively complex wind field, such as 12 July (Fig. 2b), the agreement with FLAME is fair. This is an important result as this indicates that computational less expensive, quasi real time flux retrievals may be achieved with reasonable confidence using the back-trajectory analysis performed by PlumeTraj.
On a 30 min time scale (Fig. 4), SO 2 fluxes from TROPOMI data are compatible with the FLAME time series for 20 days (or 65%). Although confirmed by the FLAME result multiple times (6,7,8,26,28 July), the occasional symmetric, bell-like shape of the flux time series (e.g. 6 July, Fig. 4) may partly be caused by geometrical spreading of the plume and the associated dispersion of SO 2 , diluting SO 2 concentrations per pixel below the detection limit of TROPOMI as a function of dispersion time. Upon converting TROPOMI VCDs to fluxes, PlumeTraj contributes to the total uncertainty of the fluxes due to erroneous back-propagation of pixels to the vent due to wind-field data errors, introducing greater uncertainty in plume height and injection time. Depending on the day and time considered, the corresponding share was between 1% and 97% of the total error, the rest being contributed by uncertainties in TROPOMI VCDs (equation 1 in Material and Methods). As a result, there may be lags between the TROPOMI and the FLAME time series. This appears to be the case, for instance, for 7 July with a lag of ~30 minutes (Fig. 4).
Since it concerns the precision rather than accuracy, even a relatively small error contribution from back-propagation analysis does not imply that its impact on the agreement is small. Errors in back-propagation mainly result from uncertainties in meteorological data, due to limited grid resolution, temporal resolution and the accuracy of the meteorological input data of PlumeTraj (assimilated empirical data versus modeled forecast data). Inaccurate meteorological data may lead to drastic decrease in accuracy of the flux time series with respect to the reference data (FLAME in this case). On the other hand, a high-resolution wind field data set increases computational costs significantly. We tested different meteorological data sets of different grid resolution ( Fig. 5): Global Forecast System with quarter degree resolution (GFS 0.25°), Global Data Assimilation System with 0.5° resolution (GDAS 0.5°) and European Centre for Medium-Range Weather Forecasts with 0.75° resolution (ECMWF 0.75°). We found large variations as function of wind field data in the back-propagation result (accuracy) as shown in Fig. 5. For all days the best match with FLAME data was achieved with GDAS 0.5°, which consequently was used for the analysis. The poor match when using GFS 0.25° data is likely due to the forecast nature of the data while GDAS contains measured data. As expected, the other major source of uncertainty impacting the TROPOMI derived fluxes were found to be clouds (Fig. 3c). The higher the cloud fraction associated with clouds overlying the SO 2 plume the more this impedes the successful retrieval of SO 2 concentrations. Clouds underlying the SO 2 plume may actually enhance retrieval precision. In general, for cloud fraction above ~0.4 and when co-located with at least ~20% of the plume area (July 10, 23, 24, 26, August 1, 2, 4) an underestimated SO 2 flux from TROPOMI was observed (Figs 3a,c and 4). For example, the interrupted SO 2 cloud spreading towards West on the 23 July is interrupted by dense cloud cover (Fig. 2c), which in all likelihood caused the drop in SO 2 flux to below 1 kt/day after 6h30 AM ( Fig. 4). Dense cloud cover also appeared on 1 and 2 August, which possibly caused an underestimated SO 2 flux with respect to the FLAME result (Figs 3 and 4). Notwithstanding the extensive cloud cover there appears to be a shift fairly constant in time, indicating coherence between the FLAME and TROPOMI time series (Fig. 4). Despite widely clear sky conditions and the fact the plume was fairly well recovered by PlumeTraj, the match is poor for the 29, 30 and 31 of July. Dense clouds at altitudes between ~2000 and ~3500 m occurred between 27 and 31 July, but only within a radius of ~15 km around the vents (Figs 1a,b and 3). Cloud cover over the vent affects all pixels and thus may lead to a significantly underestimated SO 2 load, contributing to an underestimated flux. so 2 fluxes from TROPOMI versus driving traverse. Car-based traverses of the plume by UV spectrometer have been carried out during two days (Fig. 1c). The overpass of the Sentinel-5 swath around noon and the necessity of performing the traverse after sunrise limited the number of traverses per day to about 5 or 6. The weather conditions during measurements were typically cloudy, with periods of rain and fog. The cloud was low lying, with the summit of Etna and the plume mostly above the clouds (Fig. 1d,e).
The results are shown in Fig. 6a,b. As with TROPOMI post-processing, the plume speed uncertainty retrieved from the optical flow analysis is prone to a fairly large error propagating into a fairly large flux uncertainty of ~25% (Material and Methods). For both days the time series are temporally coherent. The six traverses on 25 September yielded fluxes between 1.5 and 1.9 kt/day, in agreement with TROPOMI derived fluxes, except for the fluxes of the second and third traverse (between ~6h50 and 7h30 AM) which are substantially higher than the fluxes from TROPOMI data (around 0.6 kt/day, Fig. 6a). FLAME unfortunately yielded data only after 9h45 that day (Fig. 6a), which seem to follow the same trend as fluxes from TROPOMI data and the traverse measurement. Most of the meteorological cloud cover was lying below the plume, so SO 2 overestimation from multiple scattering within the plume (light dilution) is unlikely for both FLAME and the driving traverse, but an underestimation is possible. This makes underestimated fluxes from TROPOMI even more likely.
Using the retrieved injection times from PlumeTraj to constrain the timing of cloud cover on 25 September results in significant cloud cover only occurring between ~6 and ~8h30 AM, with cloud fractions between 0.4 to Etna, while FLAME only provides data after sunrise. Therefore, only those TROPOMI derived fluxes after 5 AM UTC and FLAME derived fluxes before 12h30 are used to compute the means (except for 13 and 23 July due to lack of FLAME data). Only one FLAME flux value was available for 23 July.   Fig. 3. The numbers in brackets indicate maximum cloud fraction as in Fig. 3, v-only vent area affected). The earliest FLAME recordings began at around 6 AM. For some days (14, 18 and 23 July and 6 August) or parts of the day no FLAME data were available. A running mean over five time steps has been applied to the ~6 minute time resolution FLAME data to make it comparable to the 30 min resolution of the TROPOMI derived fluxes.  . 6c), which is in line with the timing of the mismatch and could explain the underestimated TROPOMI result before ~8h30 AM, namely by masking the lower parts of the SO 2 plume from TROPOMI (Fig. 6a). The time series from TROPOMI appears to be roughly 30 min ahead, a lag which is within the injection time uncertainty from PlumeTraj (Material and Methods). Shifting the traverse time series by −30 min would, however, still give an underestimated TROPOMI result for the cloudy period between ~6 and ~8h30 AM (Fig. 6a).  For 26 September, the ground-based time series is shifted to higher fluxes compared to the TROPOMI time series (Fig. 6b). Both follow a similar trend, which is comparable to the situation on days 1 to 3 August (Fig. 4), were similar overcast conditions dominated. Low lying (~1000 m to ~2500 m altitude) and dense cloud cover on 26 September (cloud fraction 0.6 to 1, Figs 6c and 1e), co-located with the plume, appears to have affected parts of the SO 2 plume by attenuating light, which possibly has caused underestimated SO 2 VCDs from TROPOMI.
In addition, during the traverse on the 26 September a splitting of the plume could be observed. Through analyzing the video footage used for the plume speed retrieval (Material and Methods) a significant wind shear between the plume and the higher meteorological clouds was found, which supports the observed plume splitting. This could have caused a dilution of SO 2 below the pixel detection limit for some pixels and contributed to underestimated fluxes from TROPOMI when compared with ground-based measurements (Fig. 6b).
Cloud cover as a main factor causing the mismatch between the TROPOMI derived fluxes and those from the traverses is further suggested by the fact that fluxes from the driving traverse measurement agree with fluxes from FLAME, and fluxes from FLAME usually agree with fluxes from TROPOMI (Figs. 3 and 4).

Discussion
TROPOMI and FLAME approaches have quite different measurement and processing steps to derive SO 2 flux time series. Both TROPOMI and FLAME theoretically probe the full SO 2 column in their respective field of view and both detect VCDs. However, their fields of view differ, so they probe different air masses with different sensitivities (not to mention differences in optical components and electronics). While the telescope of a FLAME scanner has a field of view of ~24 m at a typical distance to the plume of 3000 m, a TROPOMI pixel images an average concentration over an area of about 7 × 3.5 km 2 . FLAME measures SO 2 columns near the source, whilst the TROPOMI data used here contains SO 2 concentrations up to 12 h after emission, where VCDs and hence signal-to-noise is lower than near the crater and may dilute to concentrations below the detection limit of TROPOMI.
PlumeTraj does not currently account for chemical evolution of the plume. Airborne in-situ measurements in the Mt. Etna plume yielded insignificant conversion to sulfate aerosol in the first 5 h after injection 34 . During clear sky summer days SO 2 to H 2 SO 4 conversion may occur with volume concentration changes at rates of a few %/h 35 , which may make quite modest contributions to underestimated fluxes particularly for clear days and close to noon (e.g. 5,6,11,16,17,19,22 July, Fig. 4).
For cloudy days, cloud heights usually varied between ~1000 m, especially near the vent (Fig. 1e), and ~4500 m further downwind. This was the case especially between 27 and 31 July and as discussed for the driving traverse measurements. Clouds were thus vertically co-located with considerable parts of the plume so that scavenging of SO 2 by cloud droplets 36 would take place. A rough estimation assuming sulfur diffusion rates into droplets between 10 −8 and 10 −4 mol/l/s 37 , a vertical plume thickness of 100 m, mean droplet radius of 5 µm and density of 100/cm 3 for stratocumulus clouds 38 yields a negligible SO 2 VCD loss rate below 10 −10 DU/s. Although ash emissions occurred, particularly after 23 August, these emissions had very localized impacts (~100 m) and did not rise above a few hundred meters from the crater. Uncertainties in TROPOMI VCDs due to ash emission can thus be neglected, including adsorption of SO 2 on ash particle 39 .
Bias due to aerosol is not explicitly treated in the data processing 25 . Aerosol extinction is highly variable and a function of aerosol properties (e.g. size distribution). However, the cloud detection algorithm is expected to overestimate the cloud fraction in the presence of aerosols (if at similar heights), so that at least the non-absorbing part of aerosol extinction is accounted for.
While the spatiotemporal resolution of the meteorological data yields satisfactory agreement overall, much of the uncertainty of the back-trajectory analysis may indeed be attributed to the spatial and/or temporal wind field resolution being too low for some cases. An underestimation occurs when pixels which are clearly containing SO 2 do not back-propagate to the vent, due to the limited temporal resolution of the wind field; this appears to be evident for 25 September (Fig. 6a). This particularly applies for more complicated plumes with small-scale vertical or horizontal wind shear, which includes the days 10 July and 2, 3, 4 August.
An overestimated (or underestimated) plume height by PlumeTraj may cause underestimated (or overestimated) fluxes. Figure 2a shows an example for 13 July. A considerable number of plume-pixels at the eastern flank of the plume are associated with the maximum modeled plume height of 7 km, which seems too high for quiescent degassing and thus could be erroneous, contributing to underestimated SO 2 loads and thus fluxes.
An overestimation in flux may occur due to pixels that are visually not part of the plume, but are back-propagated to the vent (eastern plume flank, Fig. 2a; northern plume flank Fig. 2b). All three error sources (wrong trajectory, cloud cover, missing pixels) may contribute simultaneously, which, based on the PlumeTraj results and cloud fraction data, was likely the case for 27 July and 3 and 5 August.
Pixels near the swath edge have lower signal-to-noise since the TROPOMI detector binning is modified for these pixels to keep pixel sizes reasonable. This caused a relatively low signal-to-noise for 9, 20 and 25 July. The intra-diurnal trends are compatible for 20 and 25 July but the noise may have contributed to overestimated SO 2 fluxes compared with the FLAME result.
Meteorological input data quality is an important issue also for other satellite remote sensing problems, in particular the spatial resolution of the wind field. Back-trajectory modeling was used to constrain CO 2 enhancements detected by the Orbiting Carbon Observatory 2 (OCO-2) over Yasur volcano (Vanuatu) 40 using wind field data with 0.5° grid resolution. The plume pixels were correctly back-propagated to the vent, even though the diameter of the Yasur vent is only ~200 m compared with ~1300 m at Mt. Etna.
The uncertainty of all three approaches (TROPOMI, FLAME, driving traverse) depends largely on how well the wind field can be constrained. The flux time series from FLAME is thus by no means an absolute reference for the TROPOMI result and prone to uncertainties that may be much larger than the −22% and +36% assumed here (Material and Methods). This is mainly due to an empirical relationship used between plume height Scientific RepoRts | (2019) 9:957 | https://doi.org/10.1038/s41598-018-37807-w and wind speed 29 . This fact motivated the driving traverse measurements, since VCDs can be retrieved without knowledge of the plume height. The result from the driving traverse is in line with the result from FLAME. This validates fluxes retrieved from the FLAME system, 12 years after its installation. Clearly, a month of data is insufficient to study how the agreement between TROPOMI and FLAME varies over the course of a year and with the seasons, which is planned to be done once a year of data becomes available. To some extent, however, this has been assessed by the measurements in September where cloudy conditions dominated. Clouds are expected to have the largest impact on the flux retrieval from TROPOMI for the fall and winter months.
For validation purposes perhaps more significant than agreement of absolute fluxes at a given time is the substantial degree of temporal coherence between the flux time series derived from TROPOMI and FLAME, which is clearly the case, as it appears even for dense cloud cover (e.g. 1 August, 26 September).
In conclusion, after comparing SO 2 fluxes from TROPOMI with fluxes from ground-based measurements we find that the agreement of the monthly average is excellent, the agreement between the daily averages is fairly good and the intra-diurnal agreement on a 30-min timescale is satisfactory. This indicates that using TROPOMI satellite images of SO 2 columns and back-trajectory analysis with a fairly coarse wind field can deliver reliable fluxes of volcanic SO 2 (and potentially other gases such as CO) in nearly real-time with high spatiotemporal resolution. This allows studying quiescent degassing processes of a tremendous number of volcanoes globally, providing new insights into the inner working of volcanoes, their risks, and their impact on atmospheric chemistry and physics. This leads to the main conclusion that the future allows for global automated real-time measurements of large numbers of degassing volcanoes world-wide, revolutionizing the quantity and quality of magmatic degassing data available to the volcanological community.

Material and Methods
Flux retrieval from TROPOMI satellite data using PlumeTraj. The spectral radiances acquired by passive satellite remote sensing instruments are strongly dependent on plume height, but lack altitude information. Thus, the L2 data products used for this study provide images containing SO 2 vertical column densities (VCD) and assume a box of fixed height (altitude level) within which the SO 2 is located. In total, there are three different altitude levels (lower troposphere: surface to 1 km, mid-troposphere: 6.5 to 7.5 km (Fig. 1b) and upper troposphere/lower stratosphere: 14.5 to 15.5 km), corresponding to three images. To retrieve accurate volcanic plume SO 2 masses from these images the actual volcanic plume height and vertical extension have to be retrieved first. To that end a pixel-by-pixel numerical procedure called PlumeTraj is used, which is fully detailed in the references 32,33 .
PlumeTraj integrates the Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT 41 ) using custom-built routines written in the Python programming language. The scheme selects those pixels, which are associated with trajectories emerging from the location of the volcanic vent (plume pixels). For each of these pixels PlumeTraj computes the height (above sea level, asl) at which the SO 2 is located at satellite measurement time instant (plume height), the height (asl) at which the prevailing atmospheric winds starts to disperse the gas into the atmosphere (injection height) and the time when the SO 2 reaches the injection height (injection time, Fig. 2). The retrieved plume height is used to correct the VCDs of each pixel of the satellite images. From the three quantities and the corrected VCDs the SO 2 load is computed for each plume pixel. To compute a flux time series used in the comparison, time bins of 30 minutes length are defined. Pixels with injection times matching a given bin time are combined. The integrated SO 2 load (sum of SO 2 mass of all pixels per bin) divided by the bin length yields the flux for a given time bin. The uncertainty of the SO 2 flux then results from the error in SO 2 mass and injection time. Assuming they are uncorrelated, the relative error of the flux for a given time instant (bin) is computed as where F is the SO 2 flux computed by PlumeTraj, m is the integrated SO 2 mass and ∆ = − + t t t i i 1 the length of a time bin between two subsequent injection times and t i and t i+1 with an associated relative uncertainty calculated as and σ t i are the mean standard deviations of the injection times (in minutes) of two subsequent time instants and quantify the uncertainty of the start and end time of a given bin i. The relative uncertainty of the SO 2 mass δm is directly proportional to the error in VCD and evaluated as where n is the number of pixels per bin, VCD c is the corrected VCD (mean over n) using the plume heights h and σ h the random error (mean over n) of the plume height from PlumeTraj alone, which typically varies between 0 m and 2 km for a given plume. σ h quantifies the uncertainty of VCD introduced by PlumeTraj during the correction of VCDs due to uncertainty in the plume height. σ r is the random error and σ s the systematic error (both mean over n) of TROPOMI data contained in the L2 data product file, accounting for twenty sources of error 25 . For instance, σ r includes shot noise, and uncertainty due to cloud fraction, σ s includes the error introduced by the radiative transfer forward model and uncertainties of spectroscopic parameters such as SO 2 absorption cross section and a contribution from plume height uncertainty, which may lead to a slight overestimation of the total uncertainty in equation 1. σ h and σ t i are estimated from a sensitivity analysis of the distance of approach of the trajectories 32 and are a measure of the accuracy of the back-trajectory analysis performed by PlumeTraj.
Scientific RepoRts | (2019) 9:957 | https://doi.org/10.1038/s41598-018-37807-w TROPOMI data were regridded onto a 0.07° × 0.07° grid (Fig. 2). For a given satellite image (one per day), the back-trajectory modeling was carried out between 1500 and 7000 m asl and from the time of acquisition (around noon UTC) up to 12 h backwards in time. After that, SO 2 is usually largely diluted below TROPOMI detection limit. To speed up the analysis, noisy pixels were excluded from the analysis by applying a threshold VCD of 0.4 DU to the input satellite data (Fig. 2).
Apart from satellite data PlumeTraj needs a meteorological dataset (including the wind field) as input. Here, data from the Global Data Assimilation System (GDAS) with 0.5° resolution were used.
Ground-based SO 2 flux measurements with the FLAME network. The FLAME network consists of nine ultraviolet scanning spectrometers spaced ~7 km apart and installed at a mean altitude of ~900 m asl on the flanks of Mt Etna (Fig. 1c). Each station scans the sky over 156°, intersecting the plume at a mean distance of ~14 km from the summit craters 29 . The network produces measurements on 86% of days per year. At each scanner, diffuse sky radiation is received through a filter for visible light (HOYA U330) and then reflected by a 45° plane mirror into a telescope of 8 mrad instrumental field-of-view (FOV). The beams are then focused into fiber optics cables connected to S2000 Ocean Optics spectrometers with UV output in the 295 to 375 nm spectral range. The system is collecting data daily for almost 9 h depending on the season and acquiring a complete scan (105 spectra including a dark spectrum) in ~5 min. Open-path ultraviolet spectra are converted to SO 2 column amounts on site applying the DOAS technique and using a modeled clear-sky spectrum 29 . Inverted data are transmitted to INGV Catania where SO 2 emission rates are computed by multiplying by the wind speed 29 .
There is a trade-off between number of scanners and accuracy of plume height, which is needed to compute SO 2 column densities as accurately as possible. Errors due to plume-height uncertainties arising from sparse scanner density is somewhat masked by both the large natural variability in the SO 2 flux emitted from Etna and the inherent error in wind speed. Uncertainties in SO 2 flux by stationary automatic scanning array arise from several sources, such as: (i) DOAS retrieval using a modelled reference spectrum (12%) 29 ; (ii) multiple scattering effects (±10%); (iii) plume speed (±10 up to ±20%); (iv) height of the plume (±10 to ±20%) due mainly to wind speed error; (v) and evaluation of the flux using the slant column amounts (15%) 42 . Errors in geometric corrections are negligible. The sum of the individual errors, as square root of the sum of the squares of the individual errors, yields a total uncertainty between −22 and +36%, which is adopted here.
Since PlumeTraj calculates fluxes at the vent location, all FLAME-based flux time series have been shifted in time by -d/v where d is the mean distance between vent and flame network and v is the wind speed.
Ground driving traverse. Ground-based traverse measurements were performed on two days along roads as close to the volcano as possible (Fig. 1c). These follow a similar method to the stationary scanners, but instead of measuring the plume cross-sections by scanning across the sky, a zenith pointing spectrometer is moved across the plume, preferably close to orthogonally to the plume direction. For these measurements a collimating telescope (mounted to a car) collected the diffuse sky light and focused it into a fiber optic cable which was connected to an Ocean Optics USB2000 + spectrometer inside the car. These spectra were analyzed in real-time using the iFit method 43 to retrieve the SO 2 vertical column amounts (VCA). The position of each spectrum is recorded using GPS. The VCA from each spectrum is multiplied by the distance travelled, corrected for a non-orthogonal direction of travel with respect to the plume direction, which is defined as the vector from the crater to the center-of-mass of the plume. These are then summed over the traverse to give the SO 2 plume cross-section, which is converted to a flux by multiplying by the wind speed, as with the FLAME network. The spectra were acquired at a frequency of ~0.2 Hz, varying the integration time and number of spectra that are averaged to keep the temporal resolution constant while maximizing the signal-to-noise in the spectrum. Traverses were performed with a 30-45 minutes period, both in a clockwise and anti-clockwise direction around Etna to avoid errors from changes in the plume azimuth during the traverses.
Although the temporal resolution of the traverses is much lower than the FLAME network, the spectra are measured pointing at zenith so some of the sources of error are removed (specifically from the plume height and use of slant columns). Errors from the retrieval scheme, multiple scattering and the wind speed still remain. To help constrain the wind speed (and uncertainty), images from the INGV camera permanent network were used to measure the plume speed during the traverses. The camera was located at Montagnola (Fig. 1c,d, latitude: 37.719° N, longitude: 15.0036°E, ~2600 m asl), above the level of the cloud. The footage is first masked to select the plume, after which the Farnebäck optical flow algorithm is applied to produce flow maps of the plume. Corrections are applied to take the plume azimuth and pixel size into account. The uncertainty in the wind speed was typically 20-25%, depending on the absolute speed and uncertainty (typically errors are much larger for slower wind speeds). This gives a total uncertainty of approximately ±25%. As with FLAME, the time series have been shifted in time to correct for the distance between vent location and flame network.