Mercury in the Free Troposphere and Bidirectional Atmosphere-Vegetation Exchanges – Insights from Maïdo Mountain Observatory in the Southern Hemisphere Tropics

Correspondence to


Introduction
Atmospheric mercury (Hg) observations are unequally distributed over the globe in several ways. On the one hand, many more sampling sites exist in the Northern Hemisphere (NH) than in the Southern Hemisphere (SH) (Sprovieri et al., 2016). On the other hand, most Hg observations in either hemisphere are made at ground level and within 50 the boundary layer. In this lowermost layer of the atmosphere, Hg concentrations are importantly affected by atmosphere-surface interactions such as local emissions and dry deposition. However, much of the long-range transport of Hg, which leads to its global distribution, does not occur within the relatively shallow boundary layer, but in the free troposphere where winds tend to be strongest and transport tends to be fastest (Travnikov, 2011).
The fate of Hg in the free troposphere, detached from direct surface influences, depends strongly on chemical 55 transformations (Travnikov, 2011) because divalent oxidized mercury (Hg II ), water-soluble and readily incorporated into water droplets and adsorbed onto particles, is removed from the atmosphere much more quickly than poorly soluble elemental mercury (Hg 0 ) (Schroeder and Munthe, 1998;Ariya et al., 2015;Lindberg et al., 2007).
Hg redox chemistry in the atmosphere is still subject to considerable uncertainties, and the debate on the dominant 60 Hg oxidants in the atmosphere has been ongoing for more than two decades (Lin, 2011;Ariya et al., 2008;Lindberg et al., 2007;Dibble et al., 2020;Calvert and Lindberg, 2005;Lindqvist and Rodhe, 1985). In many studies, it was assumed that atmospheric ozone (O3) and hydroxyl radicals (OH) act as dominant Hg oxidants (Lin, 2011).
However, thermodynamic considerations and quantum chemistry calculations showed that the homogeneous and direct oxidation of Hg 0 to Hg II via either O3 or OH is likely insignificant in the real atmosphere (Calvert and al., 1998) led to the increasing consideration of halogens, mainly bromine (Br) radicals, as important Hg oxidants.
Indeed, the widely-used mercury simulation of the GEOS-Chem chemical transport model (Selin et al., 2007) employed in recent versions a two-step Br-initiated pathway as the main Hg 0 oxidation pathway (Horowitz et al., 2017;Feinberg et al., 2022). This reaction scheme was recently updated by Shah et al. (2021), who somewhat 70 reconciled earlier studies by introducing, among others, a two-step OH-initiated oxidation pathway alongside the abovementioned Br-induced pathway, and by introducing O3 as a second-stage oxidant for both the Br-and OHinitiated oxidation pathways.
Despite these important new developments, Hg redox chemistry remains insufficiently constrained by observations. As Hg concentrations in the free troposphere are less sensitive to direct surface-atmosphere 75 interactions than in the boundary layer, it could be argued that Hg observations in the free troposphere are especially valuable for constraining Hg redox chemistry. However, such observations are rare, especially in the SH. Apart from aircraft campaigns, mountain observatories are currently the only practical way of measuring Hg in the free troposphere (Travnikov, 2011). While these observatories are still subject to surface influences and often exhibit a complex variability in the origin of sampled air masses due to orographic and thermal flows (Forrer 80 et al., 2000), they are usually able to sample air from the lower free troposphere (LFT) with regularity, especially during stable atmospheric conditions and at night (Reidmiller et al., 2010;Kleissl et al., 2007;Hahn et al., 1992; Baray et al., 2013;Collaud Coen et al., 2011). Atmospheric Hg observations from NH mountain sites such as Pic du Midi in France (Fu et al., 2016a, c;Marusczak et al., 2017), Mauna Loa in Hawaii (Carbone et al., 2016;Luippold et al., 2020), Jungfraujoch in Switzerland (Denzler et al., 2017), Mount Bachelor (Weisspenzias et al., 85 2007;Swartzendruber et al., 2006) in the US, Storm Peak (Obrist et al., 2008;Faïn et al., 2009)

in the US, and
Lulin in Taiwan (Nguyen et al., 2019(Nguyen et al., , 2022Sheu et al., 2010), among others, have provided important insights into transport and chemistry of atmospheric Hg. Meanwhile, to our best knowledge, mountain-top observations of atmospheric Hg in the SH have until now only been reported from the Chacaltaya observatory in the tropical Bolivian Andes (Koenig et al., 2021).

90
To fill data gaps in the SH in general and the SH LFT more specifically, we continuously measured from September 2017 to May 2018 (9 months) atmospheric Hg at Maïdo mountain observatory (2160 masl), a highaltitude regional GAW station on Réunion Island (21.1°S, 55.4°E) in the tropical Indian Ocean (Baray et al., 2013).
Atmospheric Hg was sampled in the form of gaseous elemental mercury (GEM; atmospheric Hg 0 ) with a high time resolution (every 15 minutes), and in the form of reactive mercury (RM; atmospheric Hg II ) with a lower time 95 resolution (integrated over ~6-14 days). Our measurement period overlapped with the OCTAVE project ("Oxygenated Compounds in the Tropical Atmosphere: Variability and Exchanges", http://octave.aeronomie.be), dedicated to the study of oxygenated organic compounds in tropical regions (Rocco et al., 2020). This is further complemented by regular and continuous observations of atmospheric trace gases such as carbon dioxide (CO2), methane (CH4), carbon monoxide (CO), and ozone (O3) at the observatory (Baray et al., 2013;Zhou et al., 2018;100 Duflot et al., 2019).
Here, we 1) give an overview of the 9 months of continuous GEM and RM observations at Maïdo observatory in the tropical Indian Ocean. We 2) derive and discuss a time series of GEM in the LFT and estimate RM in the LFT.
Finally, with the help of ancillary data and the FLEXPART-AROME lagrangian dispersion model, we 3) explore possible drivers for remarkably pronounced GEM diurnal cycles at Maïdo and address the potential role of GEM 105 photo-reemission from the island's vegetated surface.

110
Maïdo observatory (21.0792°S, 55.38°E) lies at an altitude of 2160 masl on the remote Réunion Island in the tropical Indian Ocean, about 700 km east of Madagascar (see Figure 1a). Réunion is a relatively small island (~2512 km 2 ) with a complex orography due to its volcanic origin (Gillot and Nativel, 1989). The volcano "Piton de la Fournaise" to the east of the island is still very active, with on average one eruption every 7 -10 months (Stieltjes and Moutou, 1989;Villeneuve and Bachèlery, 2006 The local atmospheric circulation of Réunion Island is complex and a result of both orographic and thermal flows (Lesouëf et al., 2011;Foucart et al., 2018). Orographic flows are induced by the island's rugged relief, which represents an environmental obstacle to the trade winds (east-south-easterlies on average). Upon encountering Réunion Island, these air masses can either rise (orographic lifting regime) or they can flow around the island, in 125 which case an overturning loop can be observed in the northwest of the island (Foucart et al., 2018). Thermal flows, which are driven by differential heating and cooling of the islands' surface, follow an important diurnal cyclicity (see Figure 1b). Nocturnal radiative surface cooling creates a cold downwash (or katabatic wind) along the slopes of the island (Baray et al., 2013), which usually leads to cloudless nights in the mountain regions. After sunrise, radiative heating typically generates a sea breeze circulation and upslope winds (or anabatic winds), which 130 are accompanied by cumulus clouds (Lesouëf et al., 2011).
Maïdo observatory samples mostly air from the LFT during nighttime due to the abovementioned katabatic winds that develop after sunset and mostly manifest as easterlies ( Figure 1). During the daytime, the observatory is importantly influenced by the planetary boundary layer (PBL) of the island as well as the marine boundary layer (MBL) of the surrounding ocean, brought to the observatory by a sea breeze circulation and anabatic winds, which 135 usually manifest as westerlies (Lesouëf et al., 2011). When the daytime sea breeze weakens on the west coast, moist air masses can also originate from the nearby Cirque de Mafate to the northeast, or get advected from the windward (eastern) side of the island by strong south-easterly trade winds (Lesouëf et al., 2011(Lesouëf et al., , 2013Tulet et al., 2017).
While atmospheric transport to Réunion Island on the mesoscale is dominated by south-easterly trade winds 140 (Foucart et al., 2018, see also supplementary Sect. S1), transport pathways can change under the influence of tropical cyclones developing over the South-West Indian Ocean (Tulet et al., 2021;Pohl et al., 2016), mostly from  The numbers give wind speed in m s -1 . The color scale is normalized, so that "1" corresponds to the most frequent 160 combination of wind direction and wind speed.

GEM and RM observations
Between September 2017 and June 2018, gaseous elemental mercury (GEM) was continuously measured from the instrumented platform at Maïdo Observatory using two Tekran® 2537 Model 2537A analyzers (Tekran Inc., observatory experienced less than 5% and 1% shifts in the manual injection checking and mass-flow meter calibration, respectively. The room in which the instrument was placed was permanently air-conditioned to 22 -23°C, thus allowing it to be permanently 2 °C to 3 °C above the outside temperatures, whatever the season, to avoid condensation in the sampling line. Raw GEM time series were quality controlled according to the guidelines proposed within the GMOS network, using a dedicated software developed at the Institute of Environmental

185
Geosciences in 2012 (see https://gmos.aeris.data.fr, last access: 25 May 2022). During this automated procedure, the raw dataset is compared against potential flags corresponding to more than 40 criteria that specifically refer to all operation phases related to the calculation of Hg concentrations and calibration (D' Amore et al., 2015), thus marking GEM readings as "valid", "warning", or "invalid". The quality-insured and quality-controlled dataset was then generated from the site manager under consideration of the previously flagged dataset, field notes,
RM was sampled with a significantly lower frequency than GEM, with collection times varying from 5.8 to 14.1 days to allow for the concentration of enough RM mass on the dedicated 0.45 µm PES filter. It has been shown that PES membranes can collect RM (gaseous oxidized mercury + particle-bound mercury) quantitatively, similar to cation exchange membranes (Dunham-Cheatham et al., 2020;Gustin et al., 2015). After collection in the field,

195
PES filters were separately stored in Petri dishes, inside double-zipper bags, and kept frozen at -20 °C until analysis. After shipment to the laboratory, each PES filter was placed in a PTFE beaker. PES filters were digested in 16 mL of ERi 2.5% inverse aqua regia (solution of 97.5% vol H2O, 1.7% vol HNO3 [78% vol] and 0.8% vol HCl [83% vol]). The beakers were closed and placed on a heating plate (120 °C) for 12 hours before analysis, which was done with a Brooks Rand Model III cold vapor atomic fluorescence spectrometry detector (CVAFS).

200
The method detection limit was 5 pg Hg (Marusczak et al., 2017) and the average LOD was around 35 pg for the entire sampling and analytical process. Standard Reference Material was NIST 3133 and ORMC-5 from NRC.
The standard measurement procedure included the CVAFS calibration (0 -50 pg; NIST 3133) and a quality control

Volatile organic compound (VOC) observations
We make use of VOC observations to characterize the origin of sampled air masses and the extent to which they 210 were impacted by surface influences. This data was generated in the framework of the OCTAVE project (http://octave.aeronomie.be; last accessed 25 May 2022), which aimed to better understand the transport and role of VOCs in tropical regions (Verreyken et al., 2019(Verreyken et al., , 2020(Verreyken et al., , 2021Rocco et al., 2020).

Meteorological observations
We use observations on meteorological parameters (Temperature, relative humidity, wind speed, wind direction, and incoming solar radiation) as continuously taken at the Piton-Maïdo meteorological station, at ~2150 masl and around ~ 1 km from Maïdo Observatory. This meteorological station is permanently checked and validated by the

230
French national meteorological service (Meteo France) and consists of a pyrocontrol probe, a Vaisala HMP110 humidity probe, an ultrasonic GILL WS2 sensor, and a K&Z CM5 pyranometer. Specific humidity was calculated from temperature and relative humidity observations but assuming a constant atmospheric pressure corresponding to the sampling altitude (2150 masl → 780 hPa).

Detection of Atmospheric Composition Change), and ACTRIS (Aerosols, Clouds and Trace gasses Research
Infrastructure) atmospheric measurement site, and a regional GAW (Global Atmospheric Watch) station (WMO region I, Africa). In this framework, the atmospheric observatory continuously houses a suite of both in-situ and 240 remote sensing instruments from which a list of continuous measurements can be found online (https://osur.univreunion.fr/observations/osu-r-stations/opar/, last access: 26 May 2022). Among them, greenhouse gases (CO, CO2, CH4) were specifically useful in the context of the present study.
Greenhouse gas measurements were performed with a PICARRO G2401 analyzer and following standardized ICOS protocols for measurement, processing, calibration, and quality control, described in detail elsewhere (Hazan 245 et al., 2016;Laurent, 2017;Heiskanen et al., 2022;Yver-Kwok et al., 2021). The analyzer was calibrated once a month with four cylinders of reference gases prepared and calibrated by the Flask and Calibration Laboratory (FCL) of ICOS in accordance with the WMO reference scales (CO2-X2019, CH4-X2004A, CO-X2014A). Two other reference gases were used for the quality control of the measurements at a rate of a daily injection for the short-term target gas, and one injection per month for the long-term target gas.

250
Ozone (O3) was measured with a UV photometric Thermo Scientific 49i analyzer with a detection limit of 1 ppb and a time resolution of 1 min. This instrument operates on the principle that O3 molecules absorb UV light at a wavelength of 254 nm, and that the degree to which the UV light is absorbed is directly related to the O3 concentration as described by the Beer-Lambert Law (Swinehart, 1962). Calibration was carried out every three months with an O3 generator, validated by the French air quality monitoring agency.

FLEXPART-AROME transport modelling and Source-Receptor-Relationships (SRRs)
Here we use the model results obtained by Verreyken et al. (2021) concerning mesoscale air mass transport to Maïdo. Briefly, mesoscale transport to Maïdo was estimated with the help of FLEXPART-AROME (Verreyken et al., 2019;Brioude et al., 2013), which feeds the FLEXPART lagrangian particle dispersion model (Stohl et al., 260 2005;Pisso et al., 2019) with meteorological input from the high resolution (horizontal: 2.5 x 2.5 km 2 ) AROME regional climatological model (Seity et al., 2011). AROME is used by Meteo France as the operational mesoscale numerical weather prediction model for the Indian Ocean. magl (500 m thickness), and two additional layers above (10 000 magl and 24 000 magl). More detailed information on this model run can be found in Verreyken et al. (2021).

270
Source-Receptor-Relationships (SRRs) describe the sensitivity of observed concentrations to surface emissions from regions of interest and are thus a practical tool to estimate quantitatively to what degree observed concentrations at Maïdo observatory are impacted by emissions from the island's surface. We calculated SRRs by dividing FLEXPART-AROME-derived air mass residence times by a constant minimal boundary layer height, as described in Seibert and Frank (2004). More details can be found in supplementary Sect. S6.

Data treatment and statistical tools
All data analysis was performed with R 3.6.0 (R Core Team, 2019) and using the "tidyverse" collection of R packages (Wickham et al., 2019). Most visualizations were done with the R package "ggplot2" (Wickham, 2016), WebPlotDigitizer (Rohatgi, 2021).
We worked with hourly averages whenever possible. Hourly averages were defined as follows: For GEM, CO, CO2, CH4, O3, and VOCs, the hourly average at hour "h" corresponds to all data taken between "h" and "h+1", e.g. the hourly average for 9 AM corresponds to all observations between 9:00 and 9:59.

285
Relative humidity (RH) and temperature (T) measurements as obtained from Meteo France are reported at the round hour only. To make this as congruent as possible with the abovementioned hourly averaging of trace gases and VOCs, we assigned the average of the two reported values at "h" and "h+1" to measurements taken at hour "h". Continuing with the above example, the hourly average of RH (or T) for 9 AM corresponds to the average of the two RH (or T) observations reported at 9:00 and 10:00.

290
The used FLEXPART dispersion runs (see section 2.6) were launched instantaneously at every hour on the hour (e.g. 8:00, 9:00, 10:00). Here, we assigned to in-situ measurements taken between "h" and "h+1" the FLEXPART dispersion run arriving at "h+1", e.g., we assigned to the hourly average of 9 AM, which represents all data taken between 9:00 and 9:59:59, the backward dispersion run launched at 10:00. The same applies for all FLEXPARTderived SRRs. Higher GEM in SON might be related to the increased occurrence of biomass burning events in the SH from August to November (Edwards et al., 2006). In fact, it has 310 been shown that biomass-burning-influenced air masses can get transported to Maïdo observatory, especially those originating from Africa and Madagascar (Verreyken et al., 2020).
GEM showed a marked diurnal variation, with a minimum at night (23:00 to 5:59 local time (LT); mean: 0.78 ng m -3 ; SD: 0.11 ng m -3 ), rising concentrations after dawn, and a peak around noon (from 12:00-13:59 LT; mean: 0.95 ng m -3 ; SD: 0.08 ng m -3 ), after which concentrations decrease again (Figure 2b). This contrasts with the 315 relatively weak diurnal variation (diurnal range < 0.05 ng m -3 ) at marine or coastal sites in the SH such as Amsterdam Island and Cape Point (Slemr et al., 2020;Angot et al., 2014;Slemr et al., 2015). On the other hand, Maïdo GEM diurnal variation with its midday peak is comparable to the summertime diurnal GEM variation at Dumont d'Urville and Concordia Station in Antarctica, where diurnal GEM variability has been attributed to photo-reemission of GEM from the snowpack (Song et al., 2018;Angot et al., 2016a, b). The possible relationship 320 between diurnal GEM cycles at Maïdo and photo-reemission is explored in section 3.3.

Reactive Mercury (RM)
Mean RM at Maïdo is quite low (10.6 pg m -3 ; SD: 5.9; 35 samples in total) compared to reported mean RM concentrations at other mountain observatories (range: ~20 -133 pg m -3 ; Fu et al., 2016a;Marusczak et al., 2016;325 Nguyen et al., 2021;Swartzendruber et al., 2006;Luippold et al., 2020;Faïn et al., 2009). It must be noted that all these other observatories lie in the NH, and most of them measured RM (in the form of GOM and PBM separately) with a Tekran® speciation unit. The potassium chloride (KCl) denuder used in the Tekran® speciation unit has been proven to not collect all RM species quantitatively, causing RM measurements in ambient air to be biased low (Gustin et al., 2019(Gustin et al., , 2021McClure et al., 2014;Lyman et al., 2010). Denuder-based RM observations at Pic 330 du Midi mountain observatory in France were later corrected upwards by 30% (mean from 40 pg m -3 to 52 pg m -

3
) after comparison to a PES-membrane-based measurement protocol near-identical to ours (Marusczak et al., 2017), which has been shown to sample RM more quantitatively (Gustin et al., 2021;Dunham-Cheatham et al., 2020). The difference between RM concentrations at Maïdo and these other mountain observatories may thus be even larger if the likely low bias of the earlier denuder-based measurements is accounted for.

335
RM at Maïdo showed a more significant seasonal variation than GEM, with 14.9 pg m -3 (SD: 6.5) for September   That said, we observed that nighttime observations at Maïdo are not always representative of the LFT, which is consistent with previous research at Maïdo observatory (Guilpart et al., 2017). Specifically, when humidity remained high at night (specific humidity > ~10 g kg -1 , RH > ~90%), nighttime observations at Maïdo most likely 365 remained importantly BL influenced. This appears to occur notably during cyclonic storms, during which the diurnal variability of GEM and specific humidity nearly disappears (see supplementary Sect. S2).
To estimate GEM in the LFT, we thus selected nighttime data and additionally removed all GEM observations sampled in air masses with a specific humidity higher than a seasonally variable threshold. As threshold, we used the monthly median nighttime specific humidity at coordinates for Maïdo and 800 hPa (~1950 masl) as obtained . This difference suggests that the seasonal variation of Maïdo LFT GEM is likely driven by other processes than TGM (GEM + RM) seasonality at Chacaltaya, which has been mainly attributed to biomass burning, vegetation uptake, and interhemispheric exchange in the upper troposphere (Koenig et al., 2021(Koenig et al., , 2022. Based on the same filtering procedure as for LFT GEM, we estimated LFT CO, CH4, and O3 (see  Ocean, is mainly driven by oxidation loss through reaction with OH (Khalil and Rasmussen, 1983). This could indicate that OH-related Hg redox chemistry (see introduction) plays an important role in driving LFT GEM seasonality.

Maïdo LFT GEM seasonality in the context of atmospheric Hg models
GEOS-Chem modeling results predict an important seasonality of atmospheric Hg in both the NH and the SH (Horowitz et al., 2017;Shah et al., 2021;Feinberg et al., 2022). In contrast, while such a Hg seasonality has been observed in the NH, only a weak Hg seasonality has been observed at SH background sites such as Amsterdam Island and Cape Point (Slemr et al., 2015(Slemr et al., , 2020. It has been proposed that atmospheric Hg seasonality is less 415 pronounced at SH monitoring stations as a consequence of the lower land cover in the SH (19%) compared to the NH (39%) and consequently a lesser importance of vegetation GEM uptake as a driver for Hg seasonality (Feinberg et al., 2022;Jiskra et al., 2018;Zhang and Zhang, 2022).
Adding another piece to this puzzle, we observed an important seasonal variation of GEM at Maïdo, but only after isolating the LFT from BL influences (Figure 3, section 3.2.1). This suggests that Hg seasonality, seemingly weak 420 in the SH BL, might be much more pronounced in the SH free troposphere.
To illustrate how this LFT seasonal variation compares to modeled Hg seasonality in the SH, and how it contrasts with the reported lack of observed Hg seasonality in the BL at other SH background sites, we digitized the GEOS-Chem modeling results from Horowitz et al. (2017) and Shah et al. (2021), which differ importantly in the used atmospheric Hg redox chemistry schemes (see introduction). As can be seen in Figure 4  The goal of this comparison is thus not to evaluate which chemistry scheme performs better, but to illustrate the added benefit from relatively continuous Hg observations in the free troposphere. Modeled seasonality of atmospheric Hg is evidently sensitive to the used chemistry scheme (Figure 4). In consequence, if Hg seasonality is observed in the free troposphere, as in the case of Maïdo, it could be used to evaluate and improve our 440 understanding of atmospheric Hg chemistry. Inversely, the observed discrepancies between modeled and observed seasonalities in the SH MBL may indicate a model misrepresentation of more BL-specific processes, such as surface emissions, deposition velocities, or more BL-specific Hg chemistry. give the mean ± 1 standard deviation.

An upper bound on RM concentrations in the LFT
While LFT and BL influences on Maïdo vary diurnally (see site description in section 2.1), our RM observations correspond to integration times of ~6-14 days. In consequence, RM concentrations reported here correspond to a mix of air masses from the LFT, the PBL, and the MBL. Given the low time resolution of our RM observations, it was not possible to isolate RM in the LFT by only considering dry nighttime air masses, as we did for hourly 455 sampled GEM (see section 3.2.1).
We hence used a different approach to constrain RM in the LFT: We combined FLEXPART-AROME results with ERA5 reanalysis data (Hersbach et al., 2020) to estimate, on a seasonal basis, the fraction of sampled air masses coming from the PBL, the MBL, the cloud-free LFT, and from clouds (see supplementary Sect. S5). We then estimated RM in the (cloud-free) LFT with the mixing equation below (Equation 1). Clouds were addressed 460 explicitly because of the elevated water solubility of atmospheric RM, which can be efficiently scavenged into cloud droplets and rain (Nair et al., 2013).

Conditions leading to marked GEM diurnal cycles
As described above, GEM at Maïdo exhibits a marked diurnal variation, with a minimum at night when the 490 observatory samples mostly air from the LFT, rising concentrations after dawn, and a peak around noon, after which concentrations decrease again (Figure 2b).
To investigate possible drivers for this diurnal variation and to determine which conditions affect the amplitude of GEM diurnal cycles, we grouped days into 1) days with a strong GEM diurnal variation (Group 1: difference between noon and nighttime > 0.16 ng m -3 ; 55 days in total), and 2) days with a weak diurnal GEM variation 495 (Group 2: difference between noon and nighttime < 0.08 ng m -3 ; 21 days in total). Days belonging to neither of these two groups were excluded from this analysis.
As can be seen in Figure  We find large differences in meteorological conditions corresponding to the two groups. While marked GEM diurnal cycles (Group 1) are associated with mostly sunny days and dry nights, atypically weak diurnal cycles 505 (Group 2) are associated with cloudy days and humid nights ( Figure 5). This suggests that Maïdo receives predominantly LFT air at night for days assigned to Group 1, while this is not the case for days assigned to Group 2.
For data belonging to Group 1, mean diurnal GEM variation anticorrelates significantly (Pearson: r = -0.98, p << 0.01) with the diurnal variation of carbon dioxide (CO2; see Figure 5), a long-lived (Archer et al., 2009) greenhouse 510 gas importantly taken up by vegetation, especially during daytime (Black and Clanton, 1973). Indeed, diurnal CO2 variation at Maïdo is mostly driven by vegetation gas exchange (Callewaert et al., 2022). Group 1 mean diurnal GEM variation also correlates significantly with the diurnal variation of isoprene (r = 0.95, p << 0.01), a shortlived (~1 hour during daytime) VOC mostly emitted by terrestrial vegetation under sunlight and heat stress (Guenther et al., 1993;Pacifico et al., 2009). At Maïdo, observed isoprene mainly originates from the vegetated 515 mountain slopes and the densely vegetated "Cirque de Mafate" (see site description in section 2.1) relatively close to the observatory (Verreyken et al., 2021). These strong relationships suggest that diurnal cycles of GEM may be related to similar drivers and regions of influence as the diurnal cycles of isoprene and CO2. In other words, GEM diurnal variation appears to be linked to the island's vegetated surface under sunlight.
Finally, Group 1 GEM diurnal variation correlates (r = 0.87, p << 0.01) with that of dimethyl sulfide (DMS), a 520 relatively short-lived (~1 -2 days; Chen et al., 2018;Kloster et al., 2006) VOC that is mostly emitted by marine phytoplankton (Stefels et al., 2007) and is frequently used as a tracer for marine air masses. Even though DMS concentrations at Maïdo are relatively low, the diurnal variation of Maïdo DMS is most likely related to marine influences (Verreyken et al., 2021). It is noteworthy that GEM concentrations increase immediately after sunrise (~6 LT -7 LT) and quickly decrease in the afternoon, while DMS concentrations start rising later in the day (at 525 around 9 LT), and do not decline until the evening (~18 LT, Figure 5). This suggests that diurnal GEM variation is not in phase with the diurnal variation of MBL influences.

Constraining the role of mixing processes
Above, we showed that very pronounced diurnal GEM cycles (Group 1) are related to the sampling of LFT air at 540 night and to a strong diurnal variation of DMS, which suggests an important diurnal variation in the fraction of air coming from the MBL.
Considering the reported importance of mixing processes at Maïdo (Lesouëf et al., 2013;Guilpart et al., 2017), we first explore the hypothesis that GEM diurnal cycles are purely driven by a diurnal variation in the sampled mix between LFT and MBL air masses. In this baseline hypothesis, we assume that there are no influences from the 545 island's surface, even though the abovementioned similarities between GEM, CO2, and isoprene diurnal variations suggest otherwise (see Figure 5).
To test this hypothesis, we built a two-box mixing model where we assume that Maïdo GEM depends only on GEM concentrations in the MBL, GEM concentrations in the LFT, as well as the diurnal variation in the mixing between LFT and MBL air (see supplementary Sect. S6). We estimated the latter with the help of FLEXPART-550 AROME (see supplementary Sect. S5). For simplicity, we only focus on those days that show a strong diurnal GEM variation (Group 1 in Figure 5; 55 days in total, of which FLEXPART output is available for 40), which generally correspond to sunny days with few clouds and low nighttime humidity. This choice allows us to minimize the difficult-to-parametrize effect of clouds and provides confidence that air masses sampled at night come indeed predominantly from the LFT (see section 3.2.1).

555
To represent LFT GEM concentrations in the two-box mixing model, we use the LFT GEM time series derived in section 3.2.1 (see also Figure 3). MBL GEM concentrations at Réunion Island are more difficult to constrain from Maïdo observations alone, as pure MBL air is rarely sampled. We thus assumed MBL GEM concentrations to be within 0.9 -1.1 ng m -3 for the two-box mixing model, based on previously reported GEM concentrations in the SH MBL (Slemr et al., 2015(Slemr et al., , 2020. Considering the reported absence of clear GEM diurnal variation and the 560 weak seasonality in the MBL of the SH (Slemr et al., 2015), we also assumed MBL concentrations to be constant throughout the day and year.
While the two-box mixing model suggests an important influence of mixing processes on GEM diurnal variation, we find that the modeled and observed diurnal GEM variations do not agree in timing ( Figure 6). A variation in mixing between LFT and MBL air as the sole driver of GEM diurnal variation cannot explain the early morning 565 rise in GEM and would place the GEM peak in the afternoon, around 2 -4 hours later than observed. It must be said that the modeled diurnal GEM variation in the two-box mixing model depends directly on the FLEXPART-AROME-based estimate of the mixing between LFT air and MBL air, which could conceivably be biased. To exclude such a bias, we compared the estimated LFT/MBL mixing to the observed diurnal DMS variation as a proxy for marine influences, finding that diurnally varying MBL influences appear to be captured adequately (see 570 supplementary Sect. S5).
We thus consider that mixing processes between MBL and LFT air, even though likely an important contributor, cannot adequately explain GEM diurnal variability alone. In the following section, we explore the potential role of the island's vegetated surface under sunlight. are shown. Shaded areas give the arithmetic mean ± 2 times the standard error.

Radiation-driven surface emissions as potential driver for GEM diurnal variation
We found that diurnal GEM variation at Maïdo depends on solar radiation, and that marked diurnal cycles of GEM 585 at Maïdo relate importantly to the diurnal cycles of isoprene and CO2, which have been attributed to the island's vegetated surface. This suggests that GEM diurnal cycles are related to vegetated surfaces under sunlight.
Previous work has shown that solar radiation, especially in the ultraviolet range, can cause photo-reemission of Hg from surfaces, such as snow, and lead to diurnal GEM cycles characterized by a midday peak (Song et al., 2018;Angot et al., 2016a). While terrestrial vegetation is globally a net sink of atmospheric mercury (Zhou and 590 Obrist, 2021), fluxes between vegetation and the atmosphere are bidirectional (Agnan et al., 2016;Luo et al., 2016), and re-emission also occurs from vegetated surfaces and soils (Yuan et al., 2019;Yu et al., 2020;Converse et al., 2010;Osterwalder et al., 2017). Not only RM deposited on leaf surfaces can get photo-reduced and reemitted as GEM, but also Hg from within the leaf tissue (Yuan et al., 2019).
Considering all this, we propose the hypothesis that Maïdo GEM diurnal variation is, in addition to mixing between 595 LFT and MBL air, driven by net daytime photo re-emission of GEM from the island's vegetated surfaces (i.e. vegetation + soil), especially from the vegetated mountain slopes close to the observatory. We addressed this hypothesis by including a term expressing the (net) photo-reemission of GEM from vegetated surfaces into the mixing model from the previous section. We parametrized the impact of photo-reemission on observed GEM as the product of 1) the FLEXPART-AROME-derived SRRs between Maïdo observatory and 600 vegetated surfaces, 2) total solar radiation as measured at Maïdo, and 3) a constant radiation-dependent surface emission term (see supplementary Sect. S6 for details). As in the previous section concerning the influence of mixing processes, and for the same reasons, we only focus on those days that exhibit a strong diurnal cycle (Group 1 in Figure 5). While we assume direct proportionality between net GEM reemission and solar radiation (see supplementary Sect. S6), the magnitude of this relationship (in other words: the slope) is not known. We estimate 605 this slope in an inverse modeling approach, i.e., we determine the most likely slope by computing the model for a wide range of values and evaluating the root mean square error (RMSE) between modeled and observed mean diurnal GEM variation.
We find that (Group 1) diurnal GEM cycles could be well explained (RMSE = 0.022 ng m -3 ) by net surface GEM emission of 0.032 ng m -2 h -1 per watt (more precisely W m -2 ) of incoming total solar radiation (Figure 7),

615
S6). To assess uncertainties, we recalculated in a Monte Carlo approach (Janssen, 2013;Metropolis and Ulam, 1949) the average net daytime emission flux for a wide range of parameters, obtaining a 95% confidence interval of 8 -22 ng m -2 h -1 (see supplementary Sect. S7). It should be noted that our flux estimate corresponds to the full ecosystem, i.e. the sum of vegetation and soil fluxes, with their relative contributions being unknown. In addition, our estimate assumes that GEM concentrations at Maïdo are not strongly affected by anthropogenic emissions 620 downslope, similar to what has been reported for CO and CO2 (mean anthropogenic contribution at noon < 7 ppbv and < 0.2 ppm, respectively; Callewaert et al., 2022).
To our knowledge, no previous studies have derived differentiated day-and nighttime GEM fluxes for mostly pristine tropical mountain forests or shrublands comparable to those on Réunion Island. Observed daytime Hg emission fluxes from background sites in other terrestrial environments (i.e. predominantly low-altitude and extra-625 tropical) were generally below ~ 3.5 ng m -2 h -1 (median ~ 0.8 ng m -2 h -1 ), significantly lower than found here (Agnan et al., 2016). Mean daytime fluxes of up to ~6 ng m -2 h -1 (depending on the season) have been reported for temperate mountain meadows in the US (Converse et al., 2010) and Tibet (Sun et al., 2020). Daytime fluxes above ~10 ng m -2 h -1 , similar to what we derived here, were observed in some tropical environments, e.g. for an open field soil in Amazonia (Almeida et al., 2009), and a naturally preserved but anthropogenically influenced (TGM > 630 5 ng m -3 ) forest soil (soil Hg: ~0.13 mg Hg kg soil −1 ) in tropical China (Fu et al., 2012). It is well-reported that daytime emissions from terrestrial surfaces correlate with soil Hg concentrations and solar radiation (Agnan et al., 2016).  (Zhou and Obrist, 2021), whose opening varies as a function of daytime, heat stress, and evaporative loss 645 (Roelfsema and Hedrich, 2005). While stomata tend to open most widely during the day, they often remain at least partly open during the night, especially in the case of tropical vegetation . This could allow for nighttime stomatal GEM uptake, in addition to possible non-stomatal uptake pathways (Converse et al., 2010). In fact, significant nighttime GEM uptake by vegetation has been reported before (Kurz et al., 2020;Fu et al., 2016b;Jiskra et al., 2018;Yu et al., 2018). Adding to GEM dry deposition, katabatic winds could bring comparatively

Conclusions & perspectives
We presented 9 months of observations of GEM (sampled every 15 minutes) and RM (integrated over ~ 6 -14 Such a marked seasonal variation stands in contrast to the reported weak Hg seasonality at background sites in the SH marine boundary layer (MBL) but is congruent with the significant and photochemistry-dependent Hg seasonality reported in modeling studies. This suggests that Hg observations in the SH LFT may be particularly 675 useful to constrain Hg chemistry and that continuous Hg monitoring at SH mountain sites could prove highly beneficial for the community.
Mean RM at Maïdo, sampled on polyether sulfone (PES) membranes, was ~10.6 pg m -3 , significantly lower than reported RM concentrations for mountain sites in the NH. While we estimate that RM in the (cloud-free) LFT at Maïdo may be about twice as high (~20 pg/m 3 ), we were limited by the low time resolution of RM observations 680 (sample integration time of ~ 6 -14 days). This limitation did not allow us to capture the likely significant diurnal RM variation at Maïdo and to rigorously separate the LFT from boundary layer influences. In future studies on mountain observatories, it would be advisable to measure RM with a higher time resolution that permits resolving diurnal variations, or at least capturing day-and nighttime differences in RM concentrations.
GEM at Maïdo exhibits marked diurnal cycling, with a nighttime minimum and a noon maximum. GEM diurnal 685 cycling is significantly more pronounced on sunny days than on cloudy days and disappears altogether during large-scale cyclonic storms. Marked GEM diurnal cycles are significantly correlated with isoprene, emitted from vegetation under sunlight, and anti-correlated with CO2, which is taken up by vegetation during photosynthesis.
This suggests net GEM emission from the island's mostly vegetated surface during the day. GEM diurnal cycles at Maïdo could be well explained by significant GEM photo-reemission from the island's vegetated surfaces (i.e.

690
vegetation + soil) during daylight hours, combined with important MBL influences in the afternoon. Through inverse modeling, we estimated that vegetated surfaces around and downslope of Maïdo could emit on average 13.5 ng m -2 h -1 (95% CI: 8 -22 ng m -2 h -1 ) of GEM during daylight hours. To maintain this net daytime emission, significant net nighttime Hg deposition from the atmosphere to these vegetated surfaces is to be expected, likely through deposition of both GEM and RM. While these results are subject to considerable uncertainties, they 695 suggest important and diurnally variable bidirectional fluxes between vegetated surfaces and the atmosphere, at least for tropical evergreen mountain forests and shrublands such as those found on Réunion Island. Future measurement campaigns on Réunion Island could explore Hg concentrations and isotopic signatures below the forest canopy, in soil, in rainfall, litterfall, and throughfall, to rigorously constrain these bidirectional Hg fluxes and to further investigate the potentially important role of nighttime GEM uptake by vegetation.

Acknowledgements
This publication is part of the GMOS-Train project that has received funding from the European Union's Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No. 860497.

740
Maïdo GEM data were collected via instruments coordinated by the IGE-PTICHA technical platform dedicated to atmospheric chemistry field instrumentation. The authors acknowledge the AERIS data infrastructure for providing access to the GEM data in this study. We