Of the nine stations composing the Paris in-situ greenhouse gas monitoring network in 2025, three are equipped with FTIR spectrometers for measuring total columns. The first (JUS) has been operating since 2014, while the most recent (GNS) began in July 2022 (see Table 1). SAC started in between, in December 2021. These three stations provide a time series of more than 10 years, including two years and a half common to all stations.

Figure 1 shows the location of the three FTIR sites, aligned along the prevailing wind axis (approximately NE–SW) characteristic of the Paris region and much of Western Europe. The polar plot displays the distribution of wind directions at Saclay (100 m above ground level) calculated over the 2014–2024 period, which is representative of regional winds. These measurements are used to define upwind–downwind gradients between stations, reflecting surface-atmosphere fluxes across the intervening urban area. This approach, known as Differential Column Measurements (DCM), is described in Chen et al. (2016) and Dietrich et al. (2021). Only gradients associated with the selected wind sectors (green and blue in Fig. 1) are analysed in Sect. 5.

The EM27/SUN spectrometers at all three sites are automated using dedicated enclosures designed at LSCE. Each enclosure is equipped with a direct sunlight sensor and a rain sensor, which trigger the automatic opening of the hood and start-up of the CamTracker (Gisi et al., 2011) and OPUS (Bruker Users Manual, Bruker, 2026) measurement software. For the JUS site, the IFS 125 HR measurement automation is designed and developed by MONARIS and is effective since June 2024.

In addition, each system includes a calibrated Vaisala PTU300 probe for continuous pressure measurements, ensuring accurate data processing (see Sect. 2.2). A travelling EM27/SUN is also used for side-by-side comparisons (Sect. 2.3) and scale transfer.

Data processing, from raw interferograms of the EM27/SUN spectrometers to final total column estimates, is performed with PROFFAST v2.3 and PROFFASTpylot v1.2, developed and maintained at Karlsruhe Institute of Technology (KIT). More details can be found in Herkommer et al. (2024) and Feld et al. (2024). At LSCE, these tools run on a dedicated server, complemented by an overlay system designed to automate and parallelize processing for about a dozen EM27/SUN instruments. This workflow enables quasi-near-real-time processing, allowing rapid identification of instrumental issues and maximizing data completeness. Between versions 2.3 and 2.4 of PROFFAST, the XCO₂ estimates differ, on average, by -0.03 ppm with some dispersion around this mean value: 90 % of the XCO₂ differences are between -0.08 and 0.03 ppm. There are a few outliers (up to 0.4 ppm) but we have not been able to identify how these outliers differ from the typical case. The XCO₂ differences (between the two versions) of PROFFAST are correlated between the two ICOS stations so that the dispersion in term of ΔXCO₂ is lower than the dispersion on XCO₂. Indeed, 90 % of the ΔXCO₂ changes (between two versions of PROFFAST) are between -0.03 and 0.03 ppm. These differences are relatively small in comparison to the other sources of errors. They are therefore neglected in the following.

PROFFAST was originally developed to incorporate a priori atmospheric profiles from the GGG2014 and then the GGG2020 model (Laughner et al., 2023), used by the TCCON community with the GGG/GFit retrieval. The LSCE overlay software also allows retrievals using CAMS (Copernicus Atmospheric Monitoring System) forecast atmospheric profiles as a priori atmospheric profiles (Agustí-Panareda et al., 2014; Tang et al., 2018). Thus, for each measurement, two parallel datasets are produced: one using GGG2020 model and another using CAMS forecast. Differences between the two are analysed in Sect. 3.5.

As part of the PAUL project (Pilot Application in Urban Landscapes, also know as ICOS-Cities), LSCE's automated chain was validated by cross-processing a raw dataset with the MUCCnet group from the Technical University of Munich (TUM). Comparison of final datasets (both using GGG2014 a priori) showed no significant differences between the two groups.

Within the NUBICOS (New Users for a Better ICOS, a project that aims to strengthen ICOS procedures and improve value generation from observations) Horizon-Infra project, total columns retrieved from EM27/SUN are also integrated into the ICOS database, along with surface pressure records. This integration enables use of ICOS tools for quality control and visualization, while also ensuring long-term traceability (Hazan et al., 2016).

Jussieu total column measurements from IFS 125 HR instrument are processed using the retrieval software GGG2020 package (GGG2020 a priori profiles, GFit retrieval code, post-processing, see Laughner et al., 2024) and distributed through the TCCON archive (https://tccondata.org, last access: March 2026). This involves using the GFit retrieval algorithm to calculate columns with profiles generated by GGG2020 (unlike EM27/SUN data processing, which uses PROFFAST with GGG2020 a priori profiles).

The completeness of solar FTIR column time series strongly depends on weather conditions. The automated enclosures maximize data acquisition by exploiting short clear-sky intervals, reducing the need for manual operation.

Data availability is evaluated monthly as: 1A=100⋅NMesDays/NAllDays where NMesDays is the number of days with valid measurement data and NAllDays is the total number of days in the month.

Data availability at the three stations, along with clear-sky fraction (CSF) from the nearest Météo France site, are shown in Fig. S1 in the Supplement. As expected, availability is highly seasonal, from ∼ 10 %–20 % in winter to 70 %–80 % in summer, occasionally exceeding 85 % under favourable conditions. The gap at Gonesse between November 2022 and January 2023 is due to instrument failure.

At Jussieu, the situation differs: before automation of the high-resolution spectrometer in May 2024, it was operated manually, leading to lower data availability – typically 5 %–15 % in winter and ∼ 40 % in summer. Additional interruptions occurred due to technical failures in winter 2022–2023, spring 2023, and summer 2024.

Total-column measurements in the Paris region are used to compute spatial gradients (see Sect. 5). Network consistency is therefore essential. We regularly perform side-by-side comparisons at each site using a travelling/mobile EM27/SUN as a common reference. These exercises ensure network-wide consistency and help detect potential instrumental drifts, which can be corrected by applying adjustment factors.

Figure 2 reports daily means and uncertainties of XCO₂, XCH₄, and XCO for each instrument relative to the mobile reference EM27/SUN. Let δ denote the difference between a fixed station and the mobile reference. For each comparison day, we first smooth minute data with a 20 min rolling average, then compute the mean σ and standard deviation of the differences (σ). The uncertainty is given by: 2Unc=σN with:

σ, the standard deviation

The travelling instrument serves as a reference frame, but shall not be considered as absolute truth. Over the full period, XCO₂ side-by-side comparisons indicate good stability at both stations, with a bias lower than about 0.15 ppm between 2023 and 2024. We also find very good consistency for methane (<1 ppb) and carbon monoxide (∼ 1–2 ppb). This supports using both stations without additional corrections when computing regional gradients. A side-by-side comparison with the IFS 125 HR was carried out in October 2023 in Jussieu. These measurements show a bias of 0.5 ppm of XCO₂ on the comparison day (with an uncertainty of around 0.1 ppm), as well as 1.9 ppb (uncertainty around 0.5 ppb) of XCH₄ and 4 ppb (uncertainty around 1 ppb) of XCO.

The main source of observed short-term variability in EM27/SUN XCO₂ is instrumental noise. We quantify this using minute data smoothed with a 20 min rolling average. For each day we compute the standard deviation of the residuals to that smoothed series.

Figure 3 illustrates a sunny-day example at Gonesse: raw XCO₂ (blue), 20 min smoothing average (red), and residuals (bottom panel). For that day, the noise is estimated to be 0.12 ppm. Applying the same method to all instruments and their full records yields stable noise levels of ∼ 0.12 ppm (XCO₂), 0.6 ppb (XCH₄), and 0.4 ppb (XCO), in agreement with Gisi et al. (2012). We find no significant differences among EM27/SUN instruments. These values are shown by black dotted lines in Fig. S2. The Jussieu IFS 125 HR usually alternates between NDACC and TCCON measurements. Since we only use TCCON measurements, days with more than 200 spectra are fewer than in Saclay and Gonesse. On those days, using the exact same method than for EM27/SUNs, we estimate 0.55 ppm for XCO₂, 2.9 ppb for XCH₄, and 0.7 ppb for XCO; these estimates are indicated as a red dotted line in Fig. S2. For XCO, we exclude the EM27/SUN instrument #88 and report values representative of #103 and #179 only, because #88 exhibits an abnormally high average XCO noise (4.9, up to 10 ppb on some days), due to a known technical issue unresolved despite Bruker investigations.

EM27/SUN and IFS 125 HR record the solar spectrum after atmospheric transmission. Column retrievals are sensitive to the vertical distribution of absorbers because absorption varies with pressure and temperature. This vertical weighting is captured by the averaging kernel (AK), which depends on the species, the spectral windows, solar zenith angle (SZA), and instrument characteristics.

Figure S3 shows typical AKs for XCO₂, XCH₄, and XCO. For CO₂ and CH₄, EM27/SUN sensitivity is close to unity across most of the atmosphere at low SZA. At high SZA the sensitivity is enhanced in the lower troposphere. For CO, the sensitivity varies less with SZA and remains close to unity up to ∼ 200 hPa. Consequently, retrieved columns exhibit SZA-dependent behaviour (Sect. 3.4).

At Paris' latitude in summer, SZA spans ∼ 25–90°, offering a test for systematic SZA-related biases, especially at low elevations where path length and Earth curvature matter. For each day with sufficient data, we define the mean XCO₂ over the 10:00–14:00 local time as a reference (SZA < 50°). We then compute differences between minute-scale XCO₂ and that reference. We only consider days with noon SZA < 50° (spring–summer) and assume no large geophysical trend between midday and early/late hours.

Figure 4 reports mean differences versus SZA (bins of 1°) from 1 July 2022 to 31 December 2024 for instruments #88 and #103 (in Saclay and Gonesse); the shaded band shows the standard deviation in each SZA bin. Standard deviations increase with SZA, reflecting both increasing time from noon and genuine variability. Typical spreads are a few tenths of a ppm (XCO₂), ∼ 1 ppb (XCH₄), and a few ppb (XCO), confirming high precision.

We detect no systematic drift up to ∼ 60°. Beyond ∼ 65–70°, a drift emerges that grows with SZA and differs by gas: for XCO₂, it becomes noticeable above ∼ 75° (> 1 ppm) and reaches up to 3.5 ppm for 80-85°; for XCH₄, it remains smaller (<2 ppb up to 80°, up to 5 ppb between 80–85°); for XCO, it is negligible. The two instruments show consistent drifts, suggesting a primarily SZA-driven effect that is linked to poorly modelled radiative processes, with minor instrument-specific modulation (Tu, 2019).

PROFFAST uses a priori vertical profiles for CO₂, CH₄, CO, and H₂O, along with temperature and pressure, and retrieves columns by scaling the profiles to minimize the spectrum residuals. Because the a priori shapes differ, retrievals can be profile-dependent. We have compared the retrievals obtained with two different a priori from the GGG2020 and CAMS models, using the following equation: 3XGAS=0.2095.ColumnGASColumnO2 with:

XGAS, the measured GAS total column [in ppm or ppb]

For both stations, the data points show a linear correlation (rSAC=-0.76 et rGNS=-0.75), indicating a systematic sensitivity of the retrievals to the a priori profile. The slope of the best linear fit is small (-0.11 at Saclay and -0.10 at Gonesse) and consistent between the two sites. We conclude that the EM27/SUN estimates using the PROFFAST retrieval tool shows some sensitivity to the a priori profile, which is independent of the station. The negative slope is physically consistent: most CAMS–GGG2020 differences occur in the lower troposphere, where GGG2020 (climatology) lacks synoptic variability. In those layers, AK > 1 (slightly), so (1 -AK) < 0; thus, increasing the a priori can decrease the retrieved column (see Eq. 4b).

Since the a priori profiles used for the estimates are very similar for the two sites (at a given time), this sensitivity has a negligible impact on the measured station-to-station gradient.

Surface pressure is required by PROFFAST's radiative transfer and column scaling. Between 29 August 2024 and 31 January 2025, the Gonesse pressure sensor suffered a mechanical glitch (0 to 3 hPa bias, with short-term variability). We therefore re-processed the affected spectra using an alternate station 2 km away, altitude-corrected, which allows quantifying retrieval sensitivity to surface-pressure error. The altitude's correction being mostly constant, it allows us to correct the pressure data with no additional uncertainty.

We diagnose consistency with XAIR, derived from O₂ absorption and expected to be near 1 (Frey et al., 2019): 5XAIR=0.2095mair⋅ColumnO2Psurfg-ColumnH2O⋅mH2O Where:

0.2095, the mole fraction of O₂ in the atmosphere

In practice, XAIR deviates slightly due to calibration/ILS (Instrumental Line Shape), surface-pressure or altitude errors, local atmospheric variability (e.g., clouds), O₂ spectroscopy uncertainties, and time stamp error on the recorded spectra. An empirical tolerance of ±4⋅10-3 around 1 is typically accepted; larger deviations lead to rejection of the measurements. Note that 4⋅10-3 in XAIR corresponds to ∼ 0.37 ppm in XCO₂ according to the results shown in Table 2.

Differences between columns before/after pressure correction are linearly proportional to the applied pressure offset (r≥0.99 for all species). The sensitivities are summarized in Table 2.

These values closely match Tu (2019). To keep systematic XCO₂ errors below 0.05 ppm, surface-pressure biases must therefore be lower than 0.3 hPa.

The subsections above identify the dominant sources of error and sensitivities for EM27/SUN and IFS 125 HR total-column retrievals in dense urban and peri-urban environments. We group them as:

Random error: measurement noise (short-term, no reproducible temporal pattern).

For EM27/SUN, we estimate global errors (from the calibration and noise contributions) of ∼ 0.2 ppm for XCO₂, ∼ 1.2 ppb for XCH₄, and ∼ 2 ppb for XCO on raw EM27/SUN columns and ∼ 0.7 ppm for XCO₂, ∼ 3.5 ppb for XCH₄, and ∼ 4.1 ppb for XCO on IFS 125 HR columns.

The time series exhibit variability at multiple timescales - weeks, months, and years. Figure 6 shows XCO₂ time series at Saclay, Gonesse, and Jussieu between 1 July 2022 and 31 December 2024. Colors indicate individual instruments, while the black solid line represents a fit to the seasonal cycle and growth rate, based on the CCGvu algorithm (Thoning and Tans, 1989; Komhyr et al., 1989). The black dotted line is a simple linear fit to the growth rate over the two-year period.

The observed growth rate in Saclay (3.2 ppm yr⁻¹) is consistent with that from surface stations, and with background Northern Hemisphere stations such as Mauna Loa in 2023–2024 (3.34 ppm yr⁻¹). The seasonal cycle, amplitude 7–8 ppm, reflects primarily biospheric activity, though seasonality in anthropogenic emissions may also contribute. Methane (see Fig. S4) and carbon monoxide (see Fig. S5), whose sources are less seasonal and are not assimilated by vegetation photosynthesis, show smaller seasonal cycles relative to short-term variability.

At synoptic scale, variations occur on spatial scales much larger than the inter-station distances considered here. Therefore, the three sites show consistent seasonal cycles and linear growth rates. Compared with surface in-situ records (∼ 20 ppm amplitude in Paris), the column amplitude is much smaller due to vertical mixing. It is important to note that seasonal and long-term signals are very similar at the regional scale (Doc et al., 2024), and therefore cancel each other out when calculating inter-station gradients, so that they do not affect plume detection.

EM27/SUN instruments acquire spectra at 1 min resolution. At such scales, most variability is noise (see Sect. 3.2), and smoothing over 20 min is appropriate. Nonetheless, large short-term variations occasionally occur, over a period of a few hours. Due to their amplitude, typical duration of variation, and correspondence with wind directions connecting the measuring stations, these appear to be most likely geophysical (urban plumes).

A notable event occurred on 9 February 2023, when XCO₂ at Gonesse increased by ∼ 6 ppm in 2 h. Figure 7 shows the measured and WRF-Chem simulated (see Sect. 5.2.1–5.2.2) XCO₂ time series at Saclay, Gonesse, and Jussieu. Between 11:00 and 13:00, XCO₂ at Gonesse rose from 423 to 429 ppm, then dropped by 4 ppm within 30 min, followed by a 2 ppm rebound. At Saclay, only a modest increase (∼ 1 ppm) was observed during that period.

WRF-Chem simulations reproduce some features: at Gonesse, a rise of ∼ 2 ppm is simulated between 04:00 and 10:00, followed by stabilization; Saclay shows no enhancement. Simulated maxima are lower and shifted in time compared with the observations, suggesting model limitations.

The lower panels of Fig. 7 show simulated CO₂ maps between 04:00 and 13:00. They reveal nighttime accumulation over Paris under calm conditions, followed by north-north-eastward transport. At 10:00, Gonesse lies in the plume core (426 ppm), while Saclay remains in background air. As winds shift, the plume moves westward, explaining the rapid decrease observed at Gonesse.

This episode, remarkable for its amplitude, is consistent with a CO₂ plume from Paris. However, the discrepancy in amplitude (7 ppm observed vs. 3 ppm simulated) and timing highlights the difficulty of comparing measurements with models near large sources. Errors in transport fields can dominate apparent source mismatches. Therefore, caution is required when interpreting model–measurement differences solely as inventory biases (Bréon et al., 2015).

As indicated in Sect. 1, we use hourly averages of wind speed and direction measured at Saclay (100 m above ground level in an open area) as representative of the regional wind conditions. EM27/SUN instruments acquire one spectrum per minute, which we then aggregate into hourly averages (to correspond to the time step of simulated data). We used the measured gradients calculated between the different pairs of stations (SAC-GNS, SAC-JUS and JUS-GNS), and compare them to the simulated gradients in Sect. 5.2.

Figure 10 (top panels) shows measured gradients as a function of wind speed (radius) and upwind wind direction (angle) for the three station pairs.

For SAC–GNS (left-hand panel), the gradient depends strongly on wind direction: south-westerlies generally yield negative values (plume toward Gonesse), while north-easterlies yield positive values (plume toward Saclay). Typical magnitudes are around +1 ppm when flow is aligned with the SW–NE axis. These are of the same order of magnitude as those observed during the preliminary campaign in 2015 (Vogel et al., 2019) or what can be observed in other large urban areas with population characteristics similar to Paris (Berlin, Hase et al., 2015; Zhao et al., 2019, Munich, Dietrich et al., 2021; Zhao et al., 2023 or Mexico, Che et al., 2024). On the other hand, we would theoretically expect a sharp dependency on wind speed, which is not clearly observed here in measured and simulated gradients.

The Jussieu station is located in the city center, and measurements there exhibit higher noise than at Saclay and Gonesse (see Sect. 3.2), which complicates the detection of small gradients. For SAC-JUS (middle panel), when wind speed < 3 m s⁻¹ the gradient is systematically negative (down to -1.5 ppm), regardless of wind direction. This reflects CO₂ accumulation above central Paris (Jussieu) relative to suburban Saclay. For wind speed > 3 m s⁻¹ and NE winds, the gradient becomes positive (up to +0.75 ppm), consistent with plume transport toward Saclay. The amplitude is roughly half that of SAC-GNS under comparable conditions. For south-westerlies, the SAC-JUS gradient varies between -1 and +1 ppm with no clear pattern, suggesting a weaker, locally variable column signal for this pair.

For JUS-GNS (right-hand panel), the behaviour is similar. At low wind (<2–3 m s⁻¹), gradients are mostly positive (∼ +1 ppm), consistent with a stronger urban load over Jussieu than over Gonesse. At higher wind speeds, south-westerlies yield negative gradients (down to -1.5 ppm), indicating plume export toward Gonesse; north-easterlies show more variable behaviour. Under NE winds, positive gradients tend to be slightly stronger than the SAC-JUS case under SW winds, consistent with Gonesse being more urbanized than Saclay, with higher emissions.