The ATOLL observatory (50.6114° N, 3.1406° E; 19 m above ground level), shown in Fig. 1, operates within the ACTRIS framework (https://www.actris.eu/, last access: 8 October 2026) and is located in Villeneuve d'Ascq, a suburban area situated a few kilometers from the city of Lille, the central core of a densely populated metropolitan region. The surrounding environment is characterized by a mix of emission sources including traffic, residential heating, agricultural practices, industrial activities, and contributions from maritime transport (Chen et al., 2022). Measurements reported at this site (Velazquez-Garcia et al., 2023) highlight aerosol properties consistent with those typically observed in urban conditions, as defined within the GAW classification scheme (Laj et al., 2020; Rose et al., 2021). In addition to local influences, the site occasionally experiences the impact of air masses transported over long distances, bringing in particles associated with volcanic activity or Saharan dust events (Mortier et al., 2013; Bovchaliuk et al., 2016; Boichu et al., 2019).

Satellite observations are used to retrieve optical properties of clouds. On the one hand, instruments on polar orbiting satellites have a better spatial resolution, e.g., about 1 km spatial resolution for cloud products from MODIS instruments for example (Platnick et al., 2003), covering the Earth surface in one day, so that only one measurement per day can be co-located with ground-based observations. On the other hand, instruments on geostationary satellites generally provide lower spatial resolution but offer a fixed viewing geometry and high temporal sampling, enabling frequent observations of the same region and facilitating the temporal co-location of satellite retrievals with ground-based measurements. The Spinning Enhanced Visible and Infrared Imager (SEVIRI), on board the geostationary METEOSAT Second Generation satellites, has a temporal resolution of 15 min and measures in four visible and near-infrared channels and eight infrared channels, with a spatial resolution of 1 km for the High-Resolution Visible (HRV) channel (broadband, 0.4–1 µm) and 3 km for the others (Schmetz et al., 2002). At the latitude and longitude of the ATOLL station, the horizontal spatial resolution of SEVIRI pixels is about 5 × 5 km².

The CLoud property dAtAset using SEVIRI – edition 3 (CLAAS-3, Benas et al., 2023) retrieves cloud properties based on measurements in the visible and near infrared. Cloud detection is performed using a naive Bayesian approximation. Cloud observations from the space-based Lidar CALIOP and SEVIRI reflectances were used to train the algorithm. Cloud top pressure is retrieved from a multilayer perceptron neural network trained with data from CALIOP and SEVIRI measurements in the infrared channels (3.7, 8.5, 11, and 12 µm). Cloud thermodynamic phase is inferred from an adapted algorithm described by Pavolonis et al. (2005) based on a decision tree measurement from infrared channels (3.9, 6.3, 8.7, 10.8, 12, and 13.4 µm). The algorithm retrieves six different types of clouds: liquid, supercooled, opaque ice, cirrus, overlap and overshooting. Only liquid clouds are studied here, so only the first two categories are considered. Cloud optical thickness (τ) and cloud droplet effective radius (re) are determined using a look-up table following the method of Nakajima and King (1990) via a visible channel at 0.6 µm and a shortwave infrared channel at 3.9 µm. The LWP is retrieved from the cloud optical depth and re assuming a vertically homogeneous water content, following the method from Stephens (1978). Cloud droplet number concentration (nd) is also retrieved from the cloud optical depth and re considering an idealized stratiform boundary-layer cloud following adiabaticity assumptions, with a fixed dispersion of the assumed cloud droplet size distribution (Bennartz and Rausch, 2017).

A comprehensive evaluation of CLAAS-3 is given in Benas et al. (2023), and we report here a brief summary. A comparison with CALIPSO on a 1.5° × 1.5° grid for 2013 reveals differences in liquid cloud fraction of up to 30 %, but lower than 10 % over the ATOLL station. The difference is even smaller if upper clouds with τ less than 0.1 are excluded from CALIOP measurements. Note that CALIPSO serves as the reference in Benas et al. (2023), and that the reported bias encompasses all cloud types – not exclusively low-level clouds – so type-specific biases may vary due to the sensitivity of the instruments (Stubenrauch et al., 2024). Cloud optical depth, re and nd were compared with AMSR2, MODIS, and in situ measurements. Coefficients of determination of 0.53 and 0.58 have been retrieved between τ and re respectively from CLAAS-3 using the 3.9 µm channel and from MODIS measurements using 3.7 µm (Benas et al., 2023). Similarly, a correlation coefficient of 0.71 was retrieved between nd from CLAAS-3 and MODIS measurements.

From our filtered dataset, over the ATOLL station, the uncertainty ranges from 0.56 to 17.67 µm for re and from 3.6 to 3276.6 cm⁻³ for nd, with 90 % of the dataset showing an uncertainty lower than 3.95 µm and 452.7 cm⁻³ for re and nd, respectively. Lastly, it is important to acknowledge that several hypotheses and a priori assumptions in the retrieval methods could introduce uncertainties. These include (Meirink et al., 2022): (i) the assumption of plane-parallel clouds, (ii) passive satellite sensors measuring an averaged radiance of both cloudy and cloud-free areas within a pixel, (iii) surface brightness, (iv) the measured signal originating from an integrated profile, (v) LWP retrievals assuming a vertically uniform re, while nd retrievals assume an increasing re with height, which is more representative of stratocumulus clouds, (vi) aerosols above clouds not being considered in the retrievals, and (vii) re and τ being retrieved from look-up tables using the doubling adding Koninklijk Nederlands Meteorologisch Instituut (KNMI) method (Wolters et al., 2006), assuming liquid clouds between 1 and 2 km in height. It is also worth noting that SEVIRI requires approximately 12 min to scan the Earth's full disk from south to north, meaning that the temporal resolution, in addition to the coarse spatial resolution, may introduce further uncertainty. This effect, however, may be partially mitigated by the temporal averaging of ground-based measurements over the 15 min CLAAS-3 resolution. Spatial collocation between CLAAS-3 and the ground-based station was performed using a nearest-neighbour approach, i.e., the CLAAS-3 pixel closest to the ATOLL station location was selected, without any spatial interpolation.

Our analysis focuses exclusively on aerosol measurements within the BLH, and we restrict our analysis to cases where the clouds are embedded within it. To identify such conditions, we use ground-based MWRP observations to determine the height of the boundary layer (e.g. Sinclair et al., 2022). Cloud base and top heights are inferred from MWRP-derived LWC profiles, with a threshold of 0.01 g m⁻³ for cloud detection. To isolate shallow boundary-layer clouds, we applied a set of complementary filtering criteria. Only clouds with a LWP below 200 g m⁻² (Pruppacher and Klett, 2010), a cloud base located above 100 m and below the estimated boundary-layer height, and a cloud geometric thickness smaller than 1 km were retained. These criteria are applied jointly to select shallow clouds while minimizing the influence of fog, free-tropospheric clouds and deep convective systems. Applying this criterion reduces the dataset to 19 % of the original dataset consisting of valid data in which all instruments (SMPS, Nephelometer, MWRP and SEVIRI) simultaneously provide valid data.

Atmospheric dynamics play a critical role in cloud formation and evolution (Klein and Hartmann, 1993). To isolate the influence of aerosols on cloud microphysics, we restrict our analysis to meteorologically stable conditions (Coopman et al., 2018). Atmospheric stability was assessed using temperature and humidity profiles derived from ground-based MWRP observations, combined with surface measurements to compute the equivalent potential temperature (θe) at both the surface and the top of the cloud layer. The vertical gradient of θe (dθe/dz) – calculated as the difference between these two levels – served as an indicator of moist stability. Only cases with positive gradients, indicative of thermodynamically stable stratification with respect to a saturated adiabat, were retained for further analysis. Additionally, to ensure that only shallow clouds were included, we excluded all clouds with a geometric thickness greater than 1 km, effectively filtering out cumulus clouds. This approach minimises the influence of convective mixing and dynamic lifting on cloud development. Moreover, we only retained cases with a boundary layer height (BLH) larger than 500 m to ensure a robust and reliable estimation of the boundary layer structure, avoiding situations where very shallow BLH retrievals are uncertain. Applying these stability filters independently reduced the dataset to approximately 40 % of its original size. Only cases with positive gradients, indicative of thermodynamically stable stratification with respect to a saturated adiabat, were retained for further analysis. By selecting cases that were convectively stable, and additionally excluding all clouds with a geometric thickness greater than 1 km, it is assumed that cumuliform clouds are removed from the sample.

To ensure consistency between ground-based (MWRP) and satellite (CLAAS-3) liquid water path (LWP) retrievals, we applied a 60 % relative difference threshold as an empirical filter. This threshold was selected pragmatically to retain more than 50 % of the available cases while reducing observations for which the two independent measurements disagree substantially, likely because they sample different cloud scenes (e.g., due to horizontal displacement or vertically heterogeneous cloud structures). When applied independently, this filter removes approximately 45 % of the original dataset.

Aerosol properties measured at the surface are used as a proxy for the aerosol population interacting with the cloud base. This assumption is made for practical reasons, as direct observations at cloud base are not available. When the boundary layer is more stable, it is reasonable to anticipate a difference between aerosol concentrations measured at the surface and those at the cloud base. However, since cloud susceptibilities to aerosol burden is quantified in terms of relative changes, the metric is relatively insensitive to absolute differences in aerosol concentration: changes in surface aerosol concentrations are assumed to scale proportionately with changes at cloud base.

CCN are not directly measured at ATOLL. Fine aerosols (with diameters Dp less than 100 nm) are generally inefficient both in acting as CCN and in scattering light (Dusek et al., 2006). Figure 2 compares the aerosol number concentration (CN) and σsp data between 2020 and 2024 when the dataset is constrained or not for CCN. The highest aerosol concentrations were typically observed during the summer months, coinciding with the peak frequency of new particle formation events (Crumeyrolle et al., 2023). During these periods, CN concentrations often reached values larger than 10 000 cm⁻³. A threshold for the effective radius (greater than 39 nm) calculated from the SMPS size distribution measurements was used to remove cases when aerosols are too small to serve as CCN (Garrett et al., 2004). Using this threshold allows the maximisation of the correlation coefficient (from 0.13 to 0.35) between CN and σsp and the retention of 20 % of the initial database (Table 2).

When applied individually to the dataset, low cloud altitude, atmospheric stability, similar scenes and the CCN criteria retained 19 %, 40 %, 45 % and 20 % of the total cases, respectively (see Table 2). However, when applied simultaneously, only 3 % of the original dataset (201 out of 8019 cases) remained. Assuming the four criteria were statistically independent, we would expect less than 1 % of the dataset to satisfy all four criteria simultaneously. The much larger observed overlap (3 %) indicates that those criteria are not fully independent and may overlap more than expected, meaning some cases simultaneously satisfy multiple conditions more often than random chance would predict. While this stringent filtering reduces the dataset to only 3 % of the original cases, it improves the physical consistency of the selected cases and ensures that retained cases are representative of the targeted regime. Although this drastic reduction in sample size is a limitation of the method, it is the necessary trade-off for obtaining a physically consistent dataset. These selected data points represent situations with minimal uncertainty in cloud-aerosol attribution and meet all requirements in terms of temporal and spatial collocation, cloud type, aerosols and boundary layer conditions. Therefore, despite their limited number, these cases provide a robust and representative basis for our analysis.

Figure 3 presents a climatology of re, nd, CN, σsp,525nm, and dθe/dz, alongside the number of data points retained after applying the constraints. Following the filtering process, 31 % of all data points are observed during winter, while only 13 % are observed during summer. This disparity primarily stems from the stability filter: since dθ/dz is typically higher in winter, more data points meet the criteria during this season. For the cloud parameters, no distinct seasonal trends emerge; instead, they exhibit relative constancy throughout the year. Indeed, re and nd unconstrained climatologies (Fig. S1 in the Supplement) are similar to the constrained ones (Fig. 3). While the year 2023 accounts for a larger fraction of the filtered dataset (∼ 44 %), most retrieved parameters do not show strong variations across years, with the exception of a slight increase in droplet size in 2023, whose origin remains unclear (cf. Fig. S2).

The relationships between the nd and re and the aerosol scattering coefficient (σsp,525nm) are illustrated in Figs. 4 and 5, respectively, for two ranges of LWP derived from the MWRP observations: from 20 to 100 and from 100 to 200 g m⁻². We observe that nd increases with σsp,525nm for both LWP ranges. For low LWP, nd values rise from 58 cm⁻³ at a scattering coefficient around 8.0 Mm⁻¹ to 205 cm⁻³ around 180 Mm⁻¹. For larger LWP values, nd values rise from 96 cm⁻³ at a scattering coefficient around 8 Mm⁻¹ to 269 cm⁻³ at 170.0 Mm⁻¹. This observation aligns with previous studies that have explored the relationship between aerosol scattering properties and droplet formation. Under the ACTRIS dry-sampling conditions, higher scattering coefficients reflect increased aerosol loading and/or changes in aerosol size distribution. For example, Garrett et al. (2004) demonstrated that higher scattering coefficients are often associated with greater number of droplets. Similarly, MacDonald et al. (2020) and Motos et al. (2023) reported that enhanced scattering can result from either increased aerosol number concentrations, a shift toward larger particle sizes, or a combination of both, which contribute to increased droplet nucleation rates.

The cloud droplet concentration susceptibility to aerosol scattering coefficient, or nucleation efficiency (Sn), is defined as: 3Sn=dln⁡nddln⁡σsp To assess the uncertainty associated with the linear regression slope in log-log space, we calculated the standard error of the slope based on the residual variance and the spread of the independent variable, and derived a 95 % confidence interval. This interval provides a statistically robust range within which the true slope is expected to lie, thus offering a quantitative estimate of regression uncertainty. Over the period of interest, we find that the cloud droplet concentration susceptibility (Sn) is ranging from 0.40 ± 0.30 for LWP between 20 and 100 g m⁻² to 0.33 ± 0.27 for LWP between 100 and 200 g m⁻². When defined as the ratio of cloud droplet number concentration to CCN number concentration, the activation fraction is physically bounded by unity, as the number of activated droplets cannot exceed the available CCN. However, in this study, aerosol loading is characterized by the aerosol scattering coefficient rather than by CCN number concentration. As an optical proxy for aerosol loading, the scattering coefficient does not represent a particle number concentration and therefore does not impose an upper bound of unity on the resulting ratio. Consequently, values greater than 1 are physically possible within the formulation adopted here and should not be interpreted as indicative of an error in the calculation.

Often, observations above 150 g m⁻² are excluded from consideration to avoid precipitating clouds (McComiskey et al., 2009). As these cases do not impose any constraint on the fit, either upward or downward, we choose to keep such cases due to the low number of data points but those observations are highlighted (red dots) in Figs. 4 and 5. Generally, as expected, an increased aerosol loading leads to higher cloud droplet concentrations.

These results are comparable in magnitude to those reported in the literature (Fig. 5). At least 9 studies reported in Fig. 5 do not provide LWP range using a combination of either airborne measurement, satellite and simulation (Nakajima et al., 2001; Terai et al., 2012; Lu et al., 2008, 2007; Quaas et al., 2006, 2009; Sekiguchi et al., 2003; Zheng et al., 2011, Schmidt et al., 2015). They reported Sn values from 0.078 to 0.787. Sarna and Russchenberg (2016), using ground-based remote sensing retrievals of droplet number concentration and integrated attenuated backscatter, found Sn values between 0.78 and 1.59 for two case studies with LWP between 30 and 90 g m⁻². Furthermore, Painemal and Zuidema (2013) reported a Sn of 0.92 in a range of LWP centered at 55 g m⁻² using onboard in-situ instruments. In a larger statistical analysis, McComiskey et al. (2009) reported substantially lower Sn values, ranging from 0.51 to 0.64 for LWP between 107 and 143 g m⁻², using MWRP-derived nd and aerosol light scattering from a nephelometer.

As was done for the nd analysis, re is plotted as a function of the aerosol scattering coefficient (σsp,525nm) as shown in Fig. 6, showing that re decreases as the scattering coefficient increases: considering the linear fit function, for a scattering coefficient of 8 Mm⁻¹, the median re is 12 µm, whereas for 180 Mm⁻¹ it is 8.4 µm for low LWP. This trend indicates that increased aerosol scattering is associated with smaller cloud droplet sizes. For large LWP values, the median re is 11.3 µm when the scattering coefficient is around 8 Mm⁻¹, whereas for 170 Mm⁻¹ it is about 8.8 µm. The Sre, relating the droplet effective radius susceptibility to the aerosol burden, is defined as: 4Sre=-dln⁡redln⁡σsp Since re depends on LWP, Eq. (4) should ideally be evaluated for a fixed LWP value as suggested by Feingold (2003). Over Lille, the Sre lies between the reported values with 0.12 ± 0.09 and 0.08 ± 0.09 for low and high LWP ranges, respectively. The uncertainties were assessed similarly to those for Sn using a 95 % confidence interval. This observation is consistent with established research on Sre, which suggests that higher aerosol concentrations lead to a greater number of smaller droplets due to increased competition for available water vapor (McComiskey et al., 2009; Feingold, 2003).

Figure 7 presents previously reported values of Sre derived from a combination of ground-based observations (in situ and remote sensing), airborne measurements, satellite retrievals, and/or numerical simulations. The results from this study (red symbol) are comparable in magnitude to those reported in the literature. Higher Sre values (0.15–0.33) are generally found in studies where the LWP range was not constrained (Feingold et al., 2001; Twohy et al., 2005; Lu et al., 2008; Costantino and Bréon, 2013). Feingold et al. (2001) and Twohy et al. (2005) reported the highest retrieved Sre values, ranging from 0.12–0.38 and 0.307–0.332, respectively. The elevated Sre values reported by Feingold et al. (2001) may be attributed to smoke-affected conditions over Brazil, whereas those reported by Twohy et al. (2005) may reflect the limited sample size (only 16 observations) and the use of aerosol number concentration measurements rather than aerosol optical properties. For low LWP conditions (below 100 g m⁻²), Sarna and Russchenberg (2016) reported Sre values ranging from 0.00 to 0.09, in good agreement with our findings. In contrast, Garrett et al. (2004) retrieved a higher Sre value of 0.17 over the Arctic, suggesting that aerosol-cloud interactions may vary depending on the environmental conditions in which the observations are made. For higher LWP conditions (above 150 g m⁻²), reported Sre values typically range between 0.02 and 0.17 (Feingold, 2003; Kim et al., 2008), which encompasses the values obtained in the present study. Interestingly, based on cloud-radar estimates of re and Raman lidar aerosol extinction, Feingold (2003) derived Sre values averaged at 0.1 ± 0.04 from seven cases (ranging from 0.02 to 0.16) observed at the DOE ARM Southern Great Plains site in Oklahoma. Using the same methodology at the same site, Kim et al. (2008) reported comparable Sre values (0.04–0.17) from a three-year dataset of continental stratus clouds. This close agreement suggests that, when the retrieval methodology and observational setting are consistent, the number of observations alone may not substantially influence the retrieved Sre values.

For both nd and re, the susceptibility uncertainties are larger than the susceptibility values themselves. To evaluate the robustness of our results, we performed a sensitivity analysis by randomly excluding 25 % of the dataset 1000 times and recalculating Sn and Sre for each realization. The mean retrieved Sre values were 0.12 and 0.08 for the low and high LWP ranges, respectively, while the corresponding mean Sn values were 0.41 and 0.33 (see Fig. S3). These results indicate that our analysis is robust to random reductions in the dataset, despite its limited size. Increasing the number of data points, by including additional ACTRIS sites, for example, may substantially reduce uncertainties or permit increasing the number of LWP bins. Zheng et al. (2020) were able to estimate cloud effective radius susceptibilities for 6 LWP bins using ground-based measurements alone, and show that susceptibility tends to be largest for smaller LWP values.