MERIS is a programmable imaging spectrometer [1]. It is one of ten core instruments on the polar orbiter ENVISAT (Environmental Satellite, launched on 1 March 2002) flying at 800 km in a sun-synchronous orbit with an equator crossing time of 10:00 AM, descending node, and 98.5° inclination. MERIS consists of 5 identical push-broom imaging spectrometers operating in the solar spectral range between 390 and 1,040 nm, arranged in a fan shape configuration which covers a total field of view of 68.5° and spans a swath width of around 1,150 km. The spectral dispersion is achieved by mapping the entrance slit of a grating spectrometer onto a CCD array. The integration time, instrument optics and CCD array resolution are adjusted such that MERIS has a spatial resolution of 260 m × 300 m and a spectral sampling of 1.25 nm. The instrument electronic data rate provides 15 channels, which are programmable by ground command in width and in position. In the regular operation mode the spatial resolution is reduced by a factor of 4 along and across track (“reduced resolution” mode). In the “full resolution” mode, the full spatial resolution is transmitted. The central wavelengths of the spectral channels, placed between 400 and 900 nm in the operational mode, vary slightly across the field of view of MERIS. This so-called “spectral smile” is caused by curvature of the image of the slit formed in the focal plane array [2].

EPIC is part of the payload of NASA’s DSCOVR satellite. The satellite was built for sun and earth observations from the L1 Lagrangian point with the possibility to constantly observe the sun-lit part of the earth. EPIC is a multispectral imager, featuring 10 spectral channels, namely four channels in the UV spectral range at 317.5, 325, 340, and 388 nm and 6 channels in the visible and near infrared region at 443, 551, 680, 687.75, 764 and 779.9 nm. The spatial resolution of EPIC will be 10 × 10 km². The EPIC oxygen A band channel will thus be somewhat different from the MERIS channel: The central wavelength of the MERIS channel varies between 760.5 nm and 762.5 nm. It has a width of roughly 3.5 nm, compared to the 1 nm wide channel of EPIC, located at 764 nm. Besides, due to the fact that DSCOVR will be placed in the L1 point, it will perform the measurements under different illumination conditions, as the sun will always be in the back of the instrument, resulting in an observed scattering angle close to 180 °.

The O₂ A method for the remote sensing of CTP used in this work is based on the analysis of the absorption of solar radiation by oxygen between 760 nm and 775 nm by relating the strength of absorption to the transmitted air mass: the transmission decreases as the transmitted absorber mass increases. A wide field of remote sensing applications uses this differential absorption technique for the estimation of masses, e.g., the estimation of atmospheric water vapour or trace gases. The presence of clouds significantly alters the path lengths of reflected photons, with high clouds leading to shorter path lengths and high transmission, and low clouds leading to longer path lengths and low transmission (e.g., [3–9]). The transmission can not be measured directly; it is estimated by the ratio r of the measured radiance in an absorption channel and one spectrally close absorption-free window channel: r = L O 2 A / L Window

Clouds are formed by a number of liquid droplets or ice particles suspended in a volume of air. Due to the non-solid character of clouds, the majority of the photons detected at the satellite stem from scattering from particles within the cloud rather than from reflection at the cloud-top. The resulting enhancement of the photon path lengths due to in-cloud multiple scattering, as compared to the case of a solid reflector, is a function of the geometrical thickness of the cloud [10], or, more precisely, of the vertical profile of extinction, which is unknown in general. If the oxygen absorption inside the cloud is not accounted for, the derived cloud-top height will always be too low with the magnitude of underestimation being a function of the vertical extinction profile and the observation geometry [10]. To avoid these systematic biases, the window radiance L_Window can be used to estimate the cloud optical thickness, which is to some extent correlated to the cloud’s vertical extent and thus to the photon path inside the cloud [5]. In doing so, it is essential to assume a realistic vertical extinction profile and cloud geometrical thickness for the forward simulation of the oxygen A band channel ratio, in order to ensure that the retrieved cloud-top height is free of bias on average. In this respect, the technique used in this framework is different from other cloud height retrievals based on oxygen absorption (e.g., [7,8,11]), which neglect oxygen absorption and consequently retrieve a pressure level that corresponds to a middle-of-cloud pressure (e.g., [12]).

The retrieval algorithm used in this work is based on radiative transfer simulations using the Matrix Operator Model (MOMO, [13–15]). The simulations were used to derive coefficients of a multi-dimensional non-linear regression that relates the measured radiance to CTP. For the radiative transfer simulations, a representative cloud extinction was assumed for every cloud type, in order to avoid the above mentioned bias due to in-cloud multiple scattering. The input to the algorithm is the window radiance at 754 nm, the ratio of the radiance of the absorption channel at 762 nm and the window radiance, the tabulated surface reflectance, the central wavelength of the absorption channel and the observation geometry. In a validation study using airborne LIDAR measurements, it could be shown that the cloud-top pressure derived from MERIS measurements is accurate to 30 hPa for low Stratocumulus clouds [16]. Due to the lack of adequate validation sources, the accuracy of the retrieval is hard to assess for high clouds. It is expected that the measurements are hardly sensitive to optically thin clouds, such as Cirrus. Optically thick, high clouds show a larger variety with respect the vertical extent and extinction profile, resulting in an increased uncertainty of the estimated path length inside the cloud and the retrieved cloud-top pressure. Further details of the MERIS cloud-top pressure algorithm can be found in [17]. As mentioned in Section 1.2, the characteristics of MERIS and EPIC are somewhat different in terms of observation geometry and the spectral location and width of the used channels: The window channel of EPIC is located at 779 nm, at the longwave end of the O₂ A band, whereas we use the window radiance at 754 nm. Since neither the gaseous absorption nor the cloud properties vary significantly between the two channels, we do not expect any influence on the general results of this study. The absorption channel of MERIS is spectrally broader than that of EPIC and there is a spectral distance of a few nanometers between them. As shown by [17], the central wavelength as well as the spectral width of the oxygen absorption channel have some effect on both the sensitivity of the measurements to CTP as well as to the susceptibility to certain error sources, mainly uncertainties in spectral calibration. However, the general, qualitative conclusions to be drawn from this study, concerning the effect of spatial resolution on the derived cloud-top pressure, are regarded as unaffected by these differences between the instrument design concepts. The same holds for the different measurement geometries of EPIC and MERIS.

The overall comparison of CT P_MERIS and CT P_EPIC is shown in Figure 2(a). The root mean square deviation between both datasets is 43 hPa with a correlation coefficient of 0.97. CT P_EPIC is 5 hPa lower than CT P_MERIS on average. The scatter is mainly due to cases with a low subscale cloud fraction whereas the subscale variability of cloud height has a weaker influence, as will be shown below.

The unknown subscale cloud fraction clearly has a large influence on the derived CT P. Globally, 35% of all EPIC pixels were found to be fully cloudy, 50% had a cloud fraction ≥ 0.75. About 20% of EPIC pixels had a cloud fraction of less than 0.1. Figure 2(b) shows the expected decrease in accuracy of CT P_EPIC with decreasing cloud fraction. The sign of the deviation mainly is a function of cloud height and surface reflectance, and, to a lesser extent, the subgrid variability of CTP. The spatial resolution of the used MERIS reduced resolution data is 1 km × 1 km. Since the characteristic size of cloud structures is thus below MERIS resolution, it can be expected that similar errors as found for CT P_EPIC arise for MERIS in presence of small-scale broken clouds. This does not affect the validity of the above findings.

The measured oxygen A band ratio is related to the average photon path length L in the atmosphere. In clear sky cases, L is a function of air mass, the surface reflectance and the vertical distribution of aerosols in the atmosphere and corresponds to an effective reflecting layer located somewhere in the middle atmosphere. Consequently, in cases of broken clouds, the clear-sky contribution on average results in a negative deviation ΔCT P_EPIC = CT P_EPIC − CT P_MERIS for low clouds and a positive deviation for high clouds. Figure 2(a) exhibits the expected deviations. Figure 2(c) shows ΔCT P_EPIC as a function of cloud height for cases with a subscale cloud fraction less than 0.3. Again, the effect of shifting broken clouds towards the middle atmosphere is clearly visible. A corresponding modification of the global histogram of CTP is apparent in Figure 2(f).

As mentioned above, in clear sky cases, the surface reflectance has a strong influence on the average photon path length L. In this work, in order to simplify the data processing, the surface reflectance is approximated by the average window reflectance at top of atmosphere in the clear fraction of MERIS pixels contributing to each synthetic EPIC pixel. In doing so, the errors caused by neglecting the small atmospheric contribution to the top-of-atmosphere reflectance are accepted, instead of relying on a reflectance database which would be applicable only over land anyway. For cases of broken clouds (subscale cloud fraction less than 0.3), Figure 2(d) shows that the resulting deviation ΔCT P tends to be negative for dark surfaces, since the contribution of the surface to the measured signal is weak and L is mainly determined by scattering in the atmosphere. For brighter surfaces there is a weak tendency for a positive bias of CT P_EPIC. The oscillating pattern of the two-dimensional histogram is caused by the different frequencies of occurrence of reflectance values around the oxygen A band, exhibiting maxima below a reflectance of 0.1 (ocean cases) and above 0.2 (land surfaces).

The remote sensing of spatially heterogeneous targets such as clouds is complicated by a number of error sources, such as the assumption of homogeneous pixels and neglecting the horizontal photon transport by assuming radiatively independent pixels (Independent Pixel Approximation, e.g., [19]). While three-dimensional effects are important for spatially highly resolving measurements (resolution ≤200 m), the subscale heterogeneity is the dominant contributor for spatial resolutions of 1 km and larger [20]. In order to determine the influence of heterogeneous clouds on CT P_EPIC, all pixels with a subscale cloud fraction of 1 were analysed with respect to the standard deviation σ of CT P among the 100 contributing MERIS values. Figure 2(e) shows the resulting ΔCT P as a function of σ, revealing a slight low bias of CT P_EPIC for high values of σ. Only 1%–2% of the analyzed cases exhibit a subscale standard deviation of more than 100 hPa (note the logarithmic color scale), corresponding to an underestimation of CT P by roughly 20 hPa on average. This underestimation of CTP is due to the concavity of the channel ratio r, when plotted against CTP (see Figure 3). The consequence of this is, that (r(CT P₁) + r(CT P₂))/2 > r((CT P₁ + CT P₂)/2), where the left side of the equation represents the EPIC measurement and the right side represents the algorithm assumption of a homogeneous cloud layer. For a specific case, this means that the measured channel ratio for a closed, homogeneous cloud layer at a height of 500 hPa is somewhat smaller than that caused by a split cloud deck, with 50% of the clouds residing at 200 hPa and 50% at 800 hPa. The larger channel ratio in cases of a high subscale variability of CTP will therefore always result in a negative bias in retrieved CTP, as compared to the average CTP in the field fo view.

Figure 4 shows the geographical distribution of errors in monthly averaged CTP (May 2005), as caused by degrading the spatial resolution. The monthly averaged CT P_EPIC was calculated at a 0.5° × 0.5° resolution from all synthetic EPIC pixels including at least one cloudy MERIS pixel. The result in each grid box was compared to the average CTP of all cloudy MERIS pixels. Generally, an underestimation of CTP (overestimation of cloud-top height) is found over the dark ocean. In agreement with the interpretation of Figure 2(d), this can be explained with the atmospheric contribution to the signal in cases of broken clouds over dark surfaces, causing the average photon path length to correspond to a level in the middle atmosphere. This effect is reduced in regions with high occurrence of deep convection and the associated high reaching cloud tops, such as found in the ITCZ, and regions with high EPIC subpixel cloud fractions, such as found in regimes with persisting, homogeneous Stratocumulus clouds (e.g. off the South-African Westcoast). Over bright land surfaces with few clouds, such as the Saharan desert, the opposite effect is observed, namely an overestimation of CTP. Again, the sign of the deviation can be explained with the strong surface contribution to the average photon path length. Continental regions with strong cloud contamination, such as found in the mid-latitudes or the inner tropics, do not show a significant error of CT P_EPIC.