Validation of the Sentinel-5 Precursor TROPOMI cloud data with Cloudnet, Aura OMI O 2 - O 2 , MODIS and Suomi-NPP VIIRS

. Accurate knowledge of cloud properties is essential to the measurement of atmospheric composition from space. In this work we assess the quality of the cloud data derived from Copernicus Sentinel-5 Precursor (S5P) TROPOMI radiance measurements: cloud top height and cloud optical thickness (retrieved with the S5P OCRA/ROCINN_CAL algorithm), cloud height (S5P OCRA/ROCINN_CRB and S5P FRESCO) and radiometric cloud fraction (all three algorithms). The analysis combines: (i) the examination of cloud maps for artiﬁcial geographical patterns, (ii) the comparison to other satellite cloud 5 data (MODIS, NPP-VIIRS and OMI O 2 - O 2 ), and (iii) ground-based validation with respect to correlative observations (2018-04-30 to 2020-02-27) from the CLOUDNET network of ceilometers, lidars and radars. Peculiar geographical patterns were identiﬁed, and will be mitigated in future releases of the cloud data products. Zonal mean latitudinal variation of S5P cloud properties are similar to that of other satellite data. S5P OCRA/ROCINN_CAL agrees well with NPP VIIRS cloud top height and cloud optical thickness, and with CLOUDNET cloud top height, especially for the low (mostly liquid) clouds. For the 10 high clouds, S5P OCRA/ROCINN_CAL cloud top height is below the cloud top height of VIIRS and of CLOUDNET, while its cloud optical thickness is higher than that of VIIRS. S5P OCRA/ROCINN_CRB and S5P FRESCO cloud height are well below the CLOUDNET cloud mean height for the low clouds, but match on an average better with the CLOUDNET cloud mean height for the higher clouds. As opposed to S5P OCRA/ROCINN_CRB and S5P FRESCO, S5P OCRA/ROCINN_CAL is well able to match the lowest CTH mode of the CLOUDNET observations. ceilometer and radiometer Measurement The radar, ceilometer and microwave radiometer data for the Summit overall programmatic and logistical additional


Introduction
Since decades the global distribution of atmospheric constituents has been monitored by ultraviolet/visible/near-infrared (UV/VIS/NIR) spectrometers measuring at the nadir of a satellite the radiance scattered by the Earth's atmosphere and reflected by its surface. tional for TROPOMI on Sentinel-5 Precursor , the OCRA/ROCINN cloud retrieval scheme will also be used operationally for the upcoming UVN instrument on Sentinel-4, the first mission of its kind for a geostationary view of air quality over Europe.
FRESCO (Fast Retrieval Scheme for Clouds from the O2 A-band) is a fast algorithm to retrieve cloud fraction and cloud height by fitting the spectral reflectance inside and outside the O2 A-band at 760 nm by a Lambertian cloud model. The 5 FRESCO retrieval method has been applied to GOME, SCIAMACHY, GOME-2 and TROPOMI. The method with its refinements over the years, like inclusion of Rayleigh scattering and directional surface albedo, has been described by Koelemeijer et al. (2001), Wang et al. (2008Wang et al. ( , 2016. FRESCO data are mainly used to correct for the cloud effect in trace gas retrievals, and to filter clouds in trace gas and aerosol retrievals. FRESCO-S (FRESCO for Sentinel) has been implemented in the L2 processor of TROPOMI as a support product for KNMI and SRON level 2 products. Due to the increase in the spectral resolution 10 in the TROPOMI instrument, the different spectral grid for each viewing direction, and small wavelength shifts introduced by inhomogeneous illumination of the spectral slit due to spatial variation of the brightness of the scene, some changes were introduced in the FRESCO-S algorithm. The spectral resolution of the reflectance database was increased to allow for interpolation of the database to the wavelengths of the observation, rather than the reverse which is used in FRESCO. Each viewing direction has its own reflectance database, to adjust to the different nominal wavelength grids and the variation of the instrument spectral 15 response function.
The OMI instrument covers the UV and visible wavelength range (270-500 nm). This means that the oxygen A-band that is used by FRESCO is not available in the spectral range measured by OMI, and an alternative cloud retrieval algorithm is required. For OMI a cloud retrieval algorithm was developed that uses the O2-O2 collision induced absorption feature at 477 nm. This is done by using a cloud model that is very similar to the model used in FRESCO (Acarreta et al., 2004;Veefkind et al., 20 2016). The sensitivity of the O 2 -O 2 cloud retrieval algorithm differs from FRESCO, because of the different wavelength range with a generally much lower surface albedo, especially over vegetated land, and a reduced sensitivity for (very) high clouds due to the reduced absorption at low pressures due to the density-squared nature of the absorption feature itself. Otherwise, both FRESCO and O 2 -O 2 Cloud are expected to retrieve a height around the mid-level of the cloud Stammes et al., 2008). 25 The OCRA/ROCINN_CAL, S5P OCRA/ROCINN_CRB and S5P FRESCO cloud properties are input to several other S5P products: total and tropospheric ozone column, ozone profile, stratospheric, tropospheric and total NO 2 column, tropospheric HCHO column, total SO 2 column, aerosol layer height and CH 4 column (Fig. 1). Hence, given the central role of the S5P cloud products, their validation is key. In this work, a comprehensive validation is performed using ground-based data from CLOUDNET as well as cloud data from other satellites: NPP-VIIRS, OMI and MODIS. 30 Section 2 gives an overview of the different cloud data products discussed in this work, the cloud properties, the S5P mission requirements, and establishes terminology. In section 3 the different satellite and ground-based data sets are described in more detail. Section 4.1.1 discusses peculiar geographical patterns that can occur in the current version of the S5P cloud products, and how these are improved in to-be-released versions. The latitudinal variation of zonal means of cloud fraction and of cloud height of different satellite cloud products is compared in Section 4.2. Section 4.3 compares specifically cloud top height and CLOUD product files. At the time of submission of this work, S5P O3 profile is not yet operational. Note that regarding to S5P CH4, NPP-VIIRS is used as main cloud mask, while S5P FRESCO is merely a backup. cloud optical thickness of S5P OCRA/ROCINN_CAL with those of NPP VIIRS (not the official production release but a prototype one). Cloud height of the S5P products is compared with ground-based CLOUDNET data in Section 4.4; here also a link is made to the OMI OMCLDO2 -CLOUDNET comparison. Finally, conclusions are given in Section 5.
2 Overview of cloud data products, properties, mission requirements and terminology As we discuss here several cloud data products, with related but not always identical properties, some terminology conventions 5 are in order. Table 1 contains an overview of properties discussed in this work (either as subject for validation or as important influence quantity) and the corresponding abbreviation and mathematical symbol. Table 2 contains an overview of cloud data products and main cloud properties.
The S5P CLOUD product files contain actually two subproducts: S5P OCRA/ROCINN_CAL and S5P OCRA/ROCINN_CRB. For both subproducts, a radiometric cloud fraction 1 (RCF; f rc ) is first obtained by the algorithm OCRA. The ROCINN_CAL 10 1 The term 'effective cloud fraction' is sometimes also used in the literature (e.g., Stammes et al., 2008). Note that radiometric cloud fraction has to be clearly distinguished from the 'cloud radiance fraction' found in e.g., the S5P NO 2 data product and which is a different quantity. a. These refer to the position of the optical centroid. See e.g., Stammes et al. (2008). This depends on the optical thickness of the cloud.
b. The relation between sRCF and RCF is frc,a = frcAc/a. In this work always a = 0.8 is taken.
c. Calculated as the mean of the positions of the cloudy grid cells in a vertical cloud profile. This does not depend on cloud optical thickness.
d. GCF is not derived from S5P measurements.
algorithm retrieves a cloud top height (CTH; h ct ) and a cloud optical thickness (COT; τ c ). The ROCINN_CRB algorithm on the other hand retrieves a cloud height (CH; h c ) (which refers to the optical centroid of the cloud rather than to the cloud top) and a cloud albedo (CA; A c ). Note that a RCF is related to, but different from, a geometrical cloud fraction (GCF; f gc ) as provided by e.g., NPP/VIIRS and MODIS. In general RCF ≤ GCF; an example is a scene that is fully cloud covered (f gc = 1) with an optically thin cloud (f rc < 1) .

5
Like S5P OCRA/ROCINN_CRB, S5P FRESCO retrieves an RCF and a CH referring to the optical center of the cloud, but an essential difference is that for most pixels, the CA is a fixed parameter equal to 0.8. The result is a scaled RCF (sRCF; f rc,0.8 ), where f rc,0.8 = f rc A c /0.8. Also the OMCLDO2 algorithm provides an sRCF with CA fixed at 0.8 . Therefore, when doing comparisons of (s)RCFs of products, we first convert all to an sRCF with CA fixed at 0.8. In the case of S5P OCRA/ROCINN_CRB, the cloud albedo A c is attributed a fill value when the RCF f rc = 0. Therefore, the 10 conversion is done as follows Note that at maximum RCF and CA, sRCF reaches 1.2 rather than 1.
The OMCLDO2 data product contains a cloud pressure, but not a cloud height. Therefore, the cloud pressure is converted to 15 a cloud height using a scale height of h scale =7.668 km, (see Eq. (2)) (2)).
f. CMH is not provided as such in the CLOUDNET product. We calculate it here considering classification labels 1-7 as cloudy grid cells and labels 0 and 8-10 as non-cloudy, following Veefkind et al. (2016).
g. Daily gridded L3 product, based on the L2 MYD06 product.
with p s the surface pressure and h s the surface altitude of the OMCLDO2 pixel. The value of 7.668 km was obtained by fitting to the AFGL Mid latitude summer (MLS) profile (Anderson et al., 1986), which is used as reference profile in the FRESCO algorithm.
The CLOUDNET classification data provides, among else, a CTH, which we use directly, and a classification of the vertical grid cells (classes 0 to 10, distinguishing e.g., clear sky, ice cloud, liquid cloud), which we convert to a vertical profile of cloudy 5 (classes 1 to 7) and non cloudy (classes 0 and 8-10) following Veefkind et al. (2016). The cloud mean height (CMT; h cm ) is then calculated as the mean of the vertical positions of the cloudy grid cells.
More detail about the satellite data products and the cloud properties is provided in Section 3.
The mission requirements applicable to the cloud data product from the atmospheric composition Sentinels were first stated in ESA (2017a, b). Adapting the terminology 2 to be compliant with the international metrology standards VIM (International Mission requirements relate to deviations of the satellite data from an (unknown) true value. But in comparisons with reallife reference data, deviations occur also due to imperfect reference measurements and moreover, because of different tempo- OCRA derives the cloud fraction from UV-VIS reflectances by separating the sensor measurements into two components: a 20 cloud-free background and a remainder expressing the influence of clouds. A color-space approach is used, where broad-band UV-VIS reflectances are translated to blue and green colors. The underlying assumption is that clouds appear white in the color-space, meaning that the spectrum of a cloud is wavelength independent across the UV-VIS wavelength range. The actual radiometric cloud fraction is then determined as the distance between the fully cloudy "white" color and the clear-sky colors taken from the reflectance background composite maps. The cloud-free background consists of global monthly composite maps 25 per color which are currently based on three years of OMI data. With the new S5P CLOUD version 2, these have been replaced based on TROPOMI data and the resolution was increased from 0.2 x 0.4 degrees to 0.1 x 0.1 degrees. In a pre-processing step, scan angle dependencies of the colors are addressed by fitting low-order polynomials to monthly mean reflectance data as a function of color, time, across-track pixel position (i.e. viewing zenith angle) and latitude. Instrumental degradation is currently not addressed in OCRA itself since the updated L1b data will themselves include a degradation correction.

30
ROCINN is a machine learning algorithm for retrieving two additional cloud parameters from the measured NIR radiances around the O2 A-band; the fitting window covers the full spectral range from 758 to 771 nm. The forward problem refers to the simulation of sun-normalized radiances for different cloud configurations using the VLIDORT Radiative Transfer Model (RTM) (Spurr, 2006). A significant set of simulated radiances, which satisfies the conditions of the smart sampling , is used for the training of the operational Neural Network (NN). The replacement of the exact RTM by a NN, which is a well tested approximation for complex operational algorithms like ROCINN, is in particular beneficial for gaining computational efficiency. The cloud top height and cloud albedo/optical thickness are the cloud parameters which can be 5 retrieved simultaneously using Tikhonov regularization technique. Two cloud models are handled in ROCINN: (i) Cloudsas-Reflecting-Boundaries (CRB) which considers the cloud as a Lambertian reflector and (ii) Clouds-as-Layers (CAL) which considers the cloud as a homogeneous cluster of scattering liquid water spherical particles using Mie theory. The CAL cloud base height is not a retrieved quantity but it is fixed by assuming a constant cloud geometrical thickness of 1 km. Other complementary information about the surface properties have been initially estimated from a climatology. In the upcoming version 10 2 of the CLOUD processor, the geometry-dependent surface properties are retrieved directly from TROPOMI measurements within the ROCINN fitting window using the GE_LER algorithm (Loyola et al., 2020).
The nominal operational processing produces an offline (OFFL) data product, with the processor release that is active at the time. There have been a few reprocessing campaigns to produce a consistent data-set. The resulting RPRO files can be combined with OFFL data for longer time series. For the CLOUD product consolidated data produced with processor version 15 1.1.7 is available from 2018-04-30 (start of phase E2). Note that NRTI CLOUD uses the same processor as RPRO and OFFL, and therefore has nearly the same output (for the same processor version). For full technical details, the reader is referred to the Product Readme File (PRF), Product User Manual (PUM) and Algorithm Theoretical Basis Document (ATBD), all available at https://sentinels.copernicus.eu/web/sentinel/technical-guides/sentinel-5p/products-algorithms.

20
FRESCO-S models a cloud as a Lambertian reflector, similar to S5P OCRA/ROCINN_CRB. FRESCO-S retrieves the information on cloud pressure p and effective cloud fraction c from the reflectance in and around the O 2 A band. FRESCO uses three about 1 nm wide wavelength windows, namely 758 -759 nm (continuum, no absorption), 760 -761 nm (strong absorption), and 765 -766 nm (moderate absorption). So both retrieved parameters p and c are consistently retrieved from the same spectral region. As opposed to S5P OCRA/ROCINN_CRB, where cloud albedo is retrieved, in FRESCO-S, the cloud albedo is assumed 25 to be 0.8, except when this assumption would lead to a cloud fraction larger than 1. In those cases the cloud fraction is set to 1, and the cloud albedo is fitted instead, but only if the cloud height is well separated from the surface. The FRESCO-S algorithm uses a surface albedo climatology based on GOME-2 (Tilstra et al., 2017). Due to the difference in overpass time between GOME-2 (in the morning) and Sentinel 5 precursor (in the afternoon), and the large discrepancy in the spatial resolution of both instruments the surface albedo climatology is currently considered as one of the larger sources of error for the FRESCO-S 30 algorithm. To compensate, some adjustments are made to suppress negative effective cloud fraction due to a climatological surface albedo value that is higher than reality. Also note that for scenes with a surface albedo higher than the assumed cloud albedo, the cloud parameters are less reliable.
The version of FRESCO validated here is 1.3, with the same time range as for the S5P OCRA/ROCINN product. The FRESCO processing in all these versions is identical, as the changes only applied to other products. As this is a non-released support product, no official documentation is currently available for FRESCO-S. Information about earlier FRESCO algorithms can be found in Koelemeijer et al. (2001); Wang and Stammes (2014). Note that significant updates to the algorithm were introduced for Tropomi/Sentinel 5 precursor, to deal with the higher spectral resolution and the effects of spectral shifts on the 5 input spectrum due to inhomogeneous slit illumination caused by partial cloud cover. Where in previous instruments (GOME, Sciamachy, GOME-2) the reflectance could be spectrally interpolated to the wavelength grid of the look-up table, for Tropomi we need to interpolate the look-up table to the wavelength grid of the observations. To make this possible, the look-up table is stored with a four-fold spectral oversampling so that spline interpolation can be used for this step.
From previous validation efforts we know that FRESCO retrieves a height near the optical extinction weighted mean height 10 of the cloud, at least for scenes with a significant cloud cover. For scenes with clouds -or aerosols, the algorithm does not distinguish between the two -close to the surface, a height that is closer to the surface will be retrieved, in fact in many cases the actual surface height will be found. This defect can be remedied by using a wider window with low to moderate absorption in the O 2 A-band. Instead of 765 -766 nm, a 5 nm wide window 765 -770 nm will increase the sensitivity to low clouds. This new look-up table will be used for FRESCO versions 1.4 and later. Level-2 (L2) product was developed by NASA  to ensure continuity for the long-term records of Moderate Resolution Imaging Spectroradiometer (MODIS) and VIIRS heritages. Note that the VIIRS data used in this work are not part of NASA VIIRS production release files and potential differences cannot be ruled out. Within the CLDPROP algorithm, the cloud top properties are derived from NOAA's operational algorithms, the so-called Clouds from AVHRR Extended (CLAVR-25 x) processing system, in which the algorithm is based primarily on IR spectral channels, with the additional information of shortwave infrared (SWIR) channels. In particular, the cloud top height is derived from the AWG (Algorithm Working Group) Cloud Height Algorithm (ACHA) (Heidinger and Li, 2019;Heidinger et al., 2019). Moreover, the cloud optical and microphysical property product inherits the MOD06 cloud optical/microphysical property retrieval algorithm from Platnick et al. (2017). The cloud optical thickness COT is retrieved simultaneously with the cloud effective radius CER based on a two 30 channel retrieval introduced in Nakajima and King (1990). Primarily the COT information is derived from the reflectance in a non-absorbing VIS, near infrared (NIR), or SWIR spectral channel which depends on the surface type. The CER information is provided by the reflectance in an absorbing SWIR or mid-wave infrared (MWIR).

MODIS
Moderate Resolution Imaging Spectroradiometer (MODIS) is on board of Terra and Aqua satellite in descending mode passing the equator in the morning from north to south and ascending mode passing from the equator in the afternoon from south to north, respectively. MODIS has 36 spectral bands ranging in wavelengths from 0.4 µm to 14.4 µm and data products are retrieved in three different spatial resolutions of 250 m, 500 m, and 1 km. The comparison with TROPOMI can be done for the 5 ascending MODIS/Aqua and only in a daily basis using the level-3 MODIS gridded atmosphere daily global joint MYD08_D3 product (Platnick et al., 2015). It contains daily 1 x 1 degree grid average values of atmospheric parameters among others also cloud properties. Cloud-top temperature, height, effective emissivity, phase and cloud fraction are produced using infrared channels with 1-km-pixel resolution and stored in the level-2 MODIS cloud data product file MYD06_L2, which is one of the four level-2 MODIS atmosphere products used for the level-3 MODIS atmosphere daily global parameters.

Aura OMI OMCLDO2
The OMI OMCLDO2 product (Veefkind et al., 2009(Veefkind et al., , 2016 is retrieved from level-1B VIS channel from the Dutch-Finnish UV-Vis nadir viewing spectrometer OMI (Ozone Monitoring Instrument) on NASA's EOS-Aura polar satellite. The nominal footprint of the OMI ground pixels is 24 × 13 km 2 (across × along track) at nadir to 165 × 13km 2 at the edges of the 2600 km swath, and the ascending node local time is 13:42 hrs. The OMCLDO2 cloud model is similar to the cloud model used in the 15 FRESCO algorithm, with the difference that it is based on the O 2 -O 2 collision-induced absorption rather than O2 absorption lines, as the O 2 A-band is not observed by OMI. As a result it is more sensitive to lower altitudes. Similar to FRESCO, it fixes the cloud albedo at 0.8, retrieves a radiometric cloud fraction and cloud height, and is sensitive not only to cloud but also to aerosol. The retrieved cloud height is more representative of cloud midheight rather than a cloud top height .

20
Following Veefkind et al. (2016) we include here a comparison of OMCLDO2 with CLOUDNET data, to judge how this is different from the S5P comparisons with CLOUDNET, using the same comparison settings.

CLOUDNET
Europe operates a network of ground-based cloud-profiling active remote sensing stations as part of the Aerosol, Clouds and 25 Trace Gas Infrastructure Network (ACTRIS). These stations operate vertically-pointing cloud radar and lidar/ceilometer and use the Cloudnet processing scheme (Illingworth et al., 2007) for the continuous evaluation of cloud profile properties. The Cloudnet scheme combines the cloud radar and lidar measurements at a temporal resolution of 30 s and a vertical resolution of 30 m to create a target categorization product which diagnoses the presence or absence in each pixel of: aerosol, insects, drizzle, rain, liquid cloud droplets, supercooled liquid droplets, ice cloud particles, melting ice cloud particles. Note that multiple targets 30 can be diagnosed within a single pixel. The Cloudnet processing scheme was also applied to similar cloud-profiling measurements from the US Department of Energy Atmospheric Radiation Measurement (ARM) sites. The sites included in this validation dataset are given in Table 3, dis-10 played in Fig. 2, and Cloudnet products are freely available for download from the Cloudnet database (http://cloudnet.fmi.fi/).

Geographical patterns
Global cloud maps of S5P OCRA/ROCINN and of S5P FRESCO were analyzed. Some peculiar patterns were discovered, described here below.
15 Their appearance is not fully deterministic and is mainly related to the clear-sky background reflectance maps and scan-angle dependency correction that are both using OMI data in S5P CLOUD version 1. These OMI-based auxiliary data are functions of several parameters, e.g. time, wavelength, latitude, viewing zenith angle etc. The patterns listed below are not a general 5 issue seen at all times and geolocations but rarely appear only for some combinations of (time, geometry, geolocation). With the update to CLOUD version 2, these OMI-based auxiliary data are replaced based on the TROPOMI data themselves and the effects listed below are largely reduced.
The following patterns may appear in S5P CLOUD version 1: an enhanced radiometric cloud fraction and cloud height mainly at the east edge of the swath at some months at some

S5P FRESCO
An issue in S5P FRESCO is that, at low radiometric cloud fraction, there is a tendency to retrieve a cloud height equal to the surface altitude. As an example, we discuss here a cloud-aerosol event captured by TROPOMI over China at 2019-02-23 ( Fig. 4 and Fig. S3). The aerosol is observed by the S5P Absorbing Aerosol Index (AAI) product (Fig. 4, top left), and attributed a height of 300 to 500 m above the surface by the S5P Aerosol Layer Height (ALH) product (Fig. S3, bottom right). A low 5 RCF cloud (RCF~0.3) of approximately the same shape is perceived by S5P FRESCO (Fig. 4, top right). Although a cloud is detected by S5P FRESCO, it is attributed zero offset from the surface (Fig. 4, bottom left).
A new version of S5P FRESCO, with a more wide fit window ('FRESCO-A wide') has been tested and is scheduled for release (S5P FRESCO version 1.4) in the near future. For this new product, the sensitivity to low clouds in the low atmosphere is improved. Fig. 4, bottom right, shows that FRESCO-A wide places the cloud at 300-500 m above the surface. Note the 10 different horizontal extent of the cloud as perceived by FRESCO wide.
Errors in the cloud height can have an important impact on the retrieval of tropospheric NO 2 columns by TROPOMI. The improvement in NO 2 column retrieval by using FRESCO-A wide instead of the operational FRESCO version is discussed in more detail by Eskes et al. (2020). 14 https://doi.org/10.5194/amt-2020-122 Preprint. Discussion started: 29 June 2020 c Author(s) 2020. CC BY 4.0 License.

Comparison of zonal means between cloud products
In this section zonal mean comparisons are presented for the different cloud products.   days (see Fig. S4). There is a good correspondence between the three products between approximately -60 • and +40 • latitude. FRESCO and OMCLDO2, but, again as expected, the geometrical cloud fraction is larger than the scaled radiometric fractions of the other cloud products. Note that at the extreme latitudes, the latitudinal variation of MODIS GCF is rather comparable to that of S5P OCRA/ROCINN than to that of S5P FRESCO.  (Fig. S4). The consistently higher MODIS CTH compared to S5P OCRA/ROCINN_CAL CTH is also observed when comparing 1 month of data (Fig. S5). These results are consistent with Schuessler et al. (2014) showing that the ROCINN_CAL model retrieves higher clouds than the ROCINN_CRB model, and also consistent with the results from Loyola et al. (2010) showing higher clouds from infrared 5 sounders compared to ROCINN.

Data selection and processing
For the current study, six days of NASA VIIRS data have been provided to DLR for the initial validation of TROPOMI OCRA/ROCINN_CAL product Version 1. As it was stated earlier, the original NASA VIIRS pixels have a resolution of 750 10 m and a pixel-by-pixel comparison is only possible if the cloud quantities from both instruments are brought to the same grid.
The re-gridding of the original NASA data brought them to the TROPOMI footprints as it is explained by the S5P-NPP Cloud Processor ATBD (Siddans, 2016). The datasets have been filtered according to several criteria to ensure that the comparison is meaningful. Only data with a OCRA/ROCINN_CAL qa_value (this is the quality indicator for the TROPOMI cloud product) above 0.5 were selected. The S5P CLOUD snow/ice flag was used to exclude data over such high reflective surfaces because 15 the cloud retrievals are particularly challenging in these conditions. Furthermore, only pixels with a VIIRS geometrical cloud fraction, which originates from a cloud mask, above 0.9 contributed to the comparison. This filtering was considered necessary in order to mitigate artefacts of the re-gridding process especially at the cloud boundaries and at scattered small-scale clouds.
Finally, only pixels which obey to the threshold criteria of CTH<15 km and 1<COT<150 were used for the validation exercise.
Those thresholds have been set because the S5P OCRA/ROCINN algorithm in CLOUD version 1 can retrieve clouds up to 20 a maximum CTH of 15 km and with an optical thickness not lower than 1 while the re-gridded VIIRS COT has a maximum of 150. After the aforementioned filtering and harmonization process of the two datasets, the total number of valid pixels for comparison exceeded the number of 30.000.000.
The COT from VIIRS is not comparable directly to the S5P COT because VIIRS has a geometric cloud fraction and not a radiometric one. For this reason we have introduced an effective COT which is the product of COT with the RCF. For the 25 optically thin clouds, the radiometric cloud fraction is underestimated and this results to an overestimated COT. By taking the product COT x RCF, any overestimation of the COT is compensated by the corresponding underestimation of the RCF.
The NASA VIIRS COT is retrieved simultaneously with the effective particle size from the reflection function at 0.75 µm and 2.16 µm. The reflection function at 0.75 µm is in principle sensitive to the COT and the reflection function at 2.16 µm is sensitive to the effective radius (Nakajima and King, 1990). S5P OCRA/ROCINN_CAL COT is retrieved at the continuum of   The daily distribution and statistical characteristics do not seem to vary significantly for the several days, as it can be seen 5 from the box plots of Fig. S7. In particular, as far as the COT is concerned, the distribution for TROPOMI is much wider than the one of VIIRS (with a standard deviation of TROPOMI being 22.8 compared to 12.7 for VIIRS) and the median for TROPOMI is about 12 while for VIIRS it is about 7. The first quartile Q1 for both instruments is 3, but the third quartile Q3

Results
is higher in TROPOMI than in VIIRS (i.e., 24 and 15, respectively). For the CTH, the distribution for TROPOMI is narrower than the one for VIIRS with a standard deviation of 2.9 km and 4.0 km for TROPOMI and VIIRS, respectively. The median for 10 TROPOMI is 2.5 km and for VIIRS it is 4 km. Similarly to the COT, the first quartile Q1 is for both sensors the same around 1.8 km. However, the third quartile is for TROPOMI at 6 km and for VIIRS at 9 km.
The general features of the statistical measures can be drawn from Fig. 8, which depict the histograms of the CTH and COT, respectively. The complete dataset including both surface types (land and ocean/water) for the 6 days is used. First of all, one can see that both instruments capture the same CTH mode at around 1.8 km. This mode is mainly dominant over the ocean (see 15 Figure S8 in the supplementary information) and it refers to the low-level marine stratocumulus clouds. Differences at the tails for the CTH distributions are present. TROPOMI seems to underestimate the high level clouds with CTH larger than 8 km. The mean TROPOMI CTH is lower than the one from VIIRS (3.8 km and 5.4 km, respectively). The observed negative bias is in the order of 1.6 km. As will be seen in Section 4.4, a negative bias is also observed from the Cloudnet comparison. Therefore this is a general outcome of the CTH validation using independent sensors. Regarding the COT, TROPOMI seems to overestimate 20 this cloud parameter meaning that the clouds appear optically thicker than in VIIRS. The S5P ROCINN overestimation is consistent with the GOME ROCINN results from (Loyola et al., 2010). The mean COT from VIIRS was found at 11.4 whereas the mean COT of TROPOMI is found at 19.2 leading to a positive bias of 7.9. The positive bias in the COT and the negative bias in the CTH has been also initially seen from the zonal means (see Figure 9 and Figure 10). The explanation for those biases are mainly related to the fact that S5P OCRA/ROCINN_CAL assumes liquid-water clouds and their properties stem from the Mie scattering theory. The ice clouds are not parameterized with the current S5P ROCINN_CAL algorithm in CLOUD version 5 1 but they will be included in future versions. While water clouds are assumed to be composed entirely by spherical droplets, ice clouds consist of a variety of habits (i.e., mixtures of randomly-oriented hexagonal plates and columns, two dimensional bullet rosettes and aggregated). The direct impact of the cloud microphysics to the retrieved COT is discussed in Zeng et al. (2012).
The similarities of the two datasets can be well summarized through the Taylor diagram (Taylor, 2001) which is shown 10 in Figure S9. The correlation coefficients for the cloud parameters CTH and COT are shown based on the surface type. The CTH is highly correlated for both surface types with the correlation coefficient r being 0.86 and 0.74 over water and land, respectively. Similarly, the COT appears with a higher correlation coefficient r=0.66 over water in comparison to 0.48 over land. The low correlation coefficients for the COT over land might be due to non-realistic surface albedo values (extracted from a climatology). Usually, over land the surface albedo might change more rapidly than over water. From the Taylor diagram 15 much more information can be extracted. The CTHs for land and water, which are lying in the inner area of the dashed arc, imply that the corresponding dataset has a lower standard deviation and indicates that the pattern variations are of a decreased amplitude. The COTs for land and water, which are lying in the outer area, imply that the pattern variations are higher than expected. Moreover, the CTH appears with lower Root-Mean-Square (RMS) errors than the COT.
The agreement between NASA VIIRS and S5P OCRA/ROCINN_CAL seems to be much better for the low-level clouds 20 which usually consist of liquid water particles. The main question is how well the two sensors agree for the several cloud types. For identifying in which type of clouds the differences are larger we follow the ISCCP (International Satellite Cloud Climatology Project) classification (Schiffer and Rossow, 1983). The scheme depicted in Fig. S10 classifies the clouds based on their CTH and COT combinations. Cumulus, stratocumulus and stratus are the low-level clouds, the altocumulus, altostratus and nimbostratus are the mid-level clouds and finally, the cirrus, cirrostratus and deep convective are classified as high-level clouds. The biases for all low-, mid-and high-level clouds are summarized in Table S1. Cumulus and stratocumulus (low-level clouds) over water appear 50% of the time with small negative CTH biases of a few hundred meters. The largest COT bias among the low-level clouds appears for the stratus type over water, but these type of clouds are not so frequent. From the mid-level clouds (see Table S1) the altocumuli show a low CTH bias, but the other two types (among which the most frequent 5 is the altostratus), have a negative bias of about 1.5 km. Altostratus and nimbostratus over water appear with a high positive bias in the COT. Finally, from the high-level clouds, the cirrostratus and cirrus, which appear with a frequency higher than 80%, show high biases in both CTH and COT. All in all, the agreement between VIIRS and TROPOMI cloud properties is certainly best for low level clouds and worse for high-level clouds.

10
In this section we discuss the comparison of S5P OCRA/ROCINN_CAL CTH, S5P OCRA/ROCINN_CRB CH and S5P FRESCO CH with ground-based CLOUDNET data. Moreover, we compare also OMI OMCLDO2 with CLOUDNET, using the same methodology. Aura/OMI has a similar overpass time as S5P/TROPOMI. Like S5P CLOUD and FRESCO, OM-CLDO2 provides rather effective cloud heights which are used as input in the retrieval of atmospheric gases. By comparing the S5P products with the CLOUDNET data on one hand, and OMCLDO2 with CLOUDNET data on the other hand, one 15 learns better how the effective cloud heights of these different products relate to the (vertically resolved) lidar/radar cloud observations of CLOUDNET, and where they are different. It also allows to make the connection with the work of Veefkind et al. (2016). It should be noted that other OMI cloud products could have been taken here for comparison, like the OMCLDRR which is based on the Fraunhofer filling signatures (346-354 nm) to derive effective cloud fraction and cloud optical centroid pressure (Joiner and Vasilkov, 2006), but this is beyond the scope of the current work.
Satellite -CLOUDNET comparison pairs are established as follows.
-Where applicable, the satellite RCF is converted to an sRCF, at CA=0.8.
-Satellite pixels are selected only if they cover the CLOUDNET site, have a qa value>50% (S5P) or no error flag (OM-25 CLDO2), and have sRCF > 0.05.
-All CLOUDNET measurements within ±600 s of a S5P or OMCLDO2 overpass are considered. From these, the CLOUDNET cloud occurrence fraction (COF), mean CLOUDNET CTH h ct and CMH h cm , and standard deviation of CLOUDNET CTH σ(h ct ) and CMH σ(h cm ) are calculated.
Despite these filter criteria, some comparison error due to co-location mismatch will persist. Another, more fundamental, problem is that the cloud (top) heights obtained from the radar-lidar based CLOUDNET data on one hand, and the more 'effective' cloud (top) heights of the S5P and OMI products are not fully comparable. For example, while the CLOUDNET CMH is obtained as a simple mean of all 'cloudy' grid positions (regardless of the local optical thickness), the cloud (top) heights returned by the S5P and OMI cloud products do depend on optical thickness. Moreover, none of the S5P or OMI cloud 5 products take into account the possibility that a cloud can be multi-layered. Following conclusions can be drawn: -Co-location pairs with a low satellite RCF are more scattered, in line with the higher cloud height uncertainty.

15
-OMCLDO2 vs CLOUDNET. There are far fewer co-locations with CLOUDNET available for OMCLDO2 than for the S5P cloud products, showing a clear advantage for S5P. For the lowest (CLOUDNET CTH 2 km; mostly liquid) clouds, OMCLDO2 CH corresponds to the CLOUDNET CTH. At CLOUDNET CTH 3 km, OMCLDO2 CH rather corresponds to the CLOUDNET CMH (e.g., Juelich, Palaiseau, Mace Head) or is below the CLOUDNET CMH (e.g., Munich, Schneefernerhaus, Leipzig). The latter pattern was also seen at Cabauw by Veefkind et al. (2016) (their figure   20 10). At Juelich, both CLOUDNET CTH and OMCLDO2 CH distributions (Fig. 10) have a low-altitude local mode at a well-matching ≈2 km. Also, for several other sites, the low-altitude mode agrees within 20% (Fig. 11), but there are also exceptions (Lindenberg, Norunda, Ny Alesund).

25
The low-altitude local CTH modes of CLOUDNET and S5P OCRA/ROCINN_CAL are reasonably well matched: in most cases they agree within 20% (Fig. 11). At several sites (e.g., Graciosa island, Juelich, Chilbolton) a higher altitude CTH mode is also captured by S5P OCRA/ROCINN_CAL, but shifted towards lower altitude compared to CLOUDNET CTH. Mean and median S5P OCRA/ROCINN_CAL CTH are lower than those of CLOUDNET, mainly due to the CTH mismatch of the higher altitude clouds which have an ice component.  Local conditions can impact the comparison: -The CLOUDNET station at Schneefernerhaus is located at a mountain, at 2.7 km, while the surface altitude attributed to 5 the relatively coarse satellite pixels is generally lower. This causes a mismatch in cloud height for low altitude clouds, where the cloud height observed by the satellite can be below the station altitude.
-Ny-Alesund can be affected by snow-ice conditions and as a consequence a high surface albedo, which is a challenge for satellite cloud retrievals. Different satellite products can be affected in different ways. OMCLDO2 is characterized by many zero-offset cloud heights with a high sRCF. For S5P OCRA/ROCINN_CAL, S5P OCRA/ROCINN_CRB and 10 S5P FRESCO, one notices a significant number of low-sRCF data points where the retrieved cloud height is significantly overestimating the CLOUDNET CTH.
-Summit is covered by permanent ice. OMCLDO2 overestimates the CLOUDNET CTH. For most data points of S5P OCRA/ROCINN_CAL, S5P OCRA/ROCINN_CRB and S5P FRESCO, zero offset clouds are obtained.
Figs. 12 and 13 present boxplot comparisons between the S5P cloud products cloud (top) height and CLOUDNET height, 15 with indications of the mission requirements on bias and uncertainty; the latter is compared here with a robust dispersion estimator: 0.5 of the central 68 interpercentile interval (0.5 IP68), which amounts to one standard deviation in the ideal case of a Gaussian error distribution. It should be noted that apart from satellite error, several other components contribute to the bias and dispersion: measurement error in the ground-based data, temporal and spatial co-location mismatch, and the fact that the effective cloud heights from satellite, and those from CLOUDNET, are not fully comparable. In particular regarding the calculated dispersion vs the stated uncertainty requirement, it must be clear that this can only serve as a partial quality test. If the dispersion is lower than the uncertainty threshold, one can be confident that the satellite uncertainty is within the threshold.

5
If, on the other hand, the dispersion is higher than the uncertainty threshold, it is by no means a proof that the satellite data exceeds the uncertainty threshold. OCRA/ROCINN_CAL CTH and CLOUDNET CMH. In agreement with the above results, S5P OCRA/ROCINN_CAL CTH minus CLOUDNET CTH is characterized by a negative bias, bordering to or exceeding the 20% bias requirement. However, 10 if one compares S5P OCRA/ROCINN_CAL CTH with the CLOUDNET CMH, the bias requirement is fulfilled in most cases.
Dispersions are calculated here using a robust estimator: 0.5 of the central 68 interpercentile interval (0.5 IP68) which amounts to one standard deviation in the ideal case of a Gaussian error distribution. The dispersion exceeds the 0.5 km uncertainty threshold in almost all cases (but see the note on the calculated bias and dispersion vs. bias and uncertainty thresholds above). Again, the dispersion exceeds the 0.5 km uncertainty threshold in almost all cases (but see the note on the calculated bias and dispersion vs. bias and uncertainty thresholds above).

Discussion and conclusions
The TROPOMI cloud products S5P OCRA/ROCINN_CAL, S5P OCRA/ROCINN_CRB and S5P FRESCO are validated in 20 this work, using independent satellite data (NPP VIIRS, MODIS and OMI OMCLDO2) and ground-based CLOUDNET data.
The following conclusions are obtained.
-S5P OCRA/ROCINN RCF and C(T)H can exhibit enhanced values at the east swath edge and a N-S gradient in the cloud albedo. This is expected to improve with the to-be-released S5P CLOUD version 2.
-S5P FRESCO has a tendency to retrieve at low cloud fraction a cloud height equal to the surface altitude. This is expected 25 to improve with the to-be-released S5P FRESCO 1.4 version.
-In the comparison of zonal means between different satellite products, similar latitudinal variations in cloud fraction, cloud (top) height and cloud optical thickness are obtained, sometimes with offsets. Radiometric cloud fractions, scaled to a fixed cloud albedo, between S5P OCRA/ROCINN, S5P FRESCO and OMCLDO2 agree well, except at extreme latitude where S5P FRESCO diverges.

30
-CTH and COT of S5P OCRA/ROCINN_CAL and NPP VIIRS CTH agree best for low-level liquid clouds, while they disagree for the high-level clouds. A similar conclusion is reached when comparing S5P OCRA/ROCINN_CAL CTH  with ground-based CLOUDNET CTH. In a future release of S5P CLOUD, liquid clouds and ice clouds will be retrieved separately. Then an improved agreement with NPP VIIRS and CLOUDNET can be expected.
-The different S5P cloud products, and OMCLDO2, track different vertical portions of a cloud (as observed by CLOUD-NET). For low clouds, OMCLDO2 CH corresponds to the CLOUDNET CTH, while S5P OCRA/ROCINN_CRB CH and S5P FRESCO CH are below the CLOUDNET CMH. For higher clouds, OMCLDO2 CH is sometimes at, but also 5 sometimes below, the CLOUDNET CMH. This is in line with expectations: there is a reduced sensitivity for high clouds due to the reduced absorption at low pressures due to the density-squared nature of the absorption feature. On the other hand, S5P OCRA/ROCINN_CRB CH and S5P FRESCO CH rather follow the CLOUDNET CMH for high clouds. S5P OCRA/ROCINN_CAL CTH is mostly somewhere between the CLOUDNET CMH and the CLOUDNET CTH.
-As opposed to ROCINN_CRB and FRESCO (both based on a Lambertian model), ROCINN_CAL (based on Mie scatter-10 ing cloud model), is well able to match the lowest CTH mode of the CLOUDNET observations. At several CLOUDNET sites, ROCINN_CAL also observes a second high mode, but shifted towards smaller CTH compared to the CLOUD-NET CTH. Furthermore, S5P OCRA/ROCINN_CAL CTH has far less a tendency to retrieve a cloud height equal to the surface altitude.
Several improvements in the cloud products are foreseen in upcoming version releases. These new data versions should be 15 validated with the same system as used in the current paper, allowing the necessary independent assessment of the S5p data product evolution.
Author contributions. SC and AA carried out the global validation analysis, with support from RL and MS. JCL, DH, AK and TV contributed input and advise at all stages of the analysis. AA, DL, RL and FR developed the OCRA/ROCINN retrieval algorithms and the corresponding operational UPAS processors for TROPOMI, GOME-2 and GOME at DLR. MS, PS and PW developed the TROPOMI and OMI cloud data processors at KNMI. AMF and EOC post-processed Cloudnet data tailored to S5p validation and contributed ground-based scientific expertise. All authors revised and commented on the manuscript.
Competing interests. The authors declare that they have no conflict of interest. Rasson for post-processing of the network and satellite data and for their dedication to the S5p operational validation.
This work contains modified Copernicus Sentinel-5 Precursor data (2018-2020) processed by DLR and KNMI, and post-processed by BIRA-IASB. We acknowledge the ACTRIS RI and the ACTRIS-2 project (European Commission contract H2020-INFRAIA, grant no. 654109) for providing the ground-based data from the Cloudnet sites in this study, which was produced by the Finnish Meteorological Institute, and is available for download from http://cloudnet.fmi.fi/. The cloud radar, ceilometer and microwave radiometer data for the ARM sites used in this study (Graciosa and Villa Yacanto) were obtained from the Atmospheric Radiation Measurement (ARM) user facility, managed by the Office of Biological and Environmental Research for the U.S. Department of Energy Office of Science. The cloud radar, ceilometer and microwave radiometer data for the Summit Station were obtained from NOAA; overall programmatic and logistical support for was provided by the US National Science Foundation, with additional instrumental support provided by the NOAA Earth System Research