The Deep Space Climate Observatory (DSCOVR) satellite was launched on 11 February 2015 and in June 2015 began making observations of both the Earth and the Sun from its Lissajous orbit about the Earth's L1 Lagrangian point, a gravity-neutral position 1.5 million km from the Earth. The DSCOVR payload includes two Earth-observing instruments, namely the Earth Polychromatic Imaging Camera (EPIC) and the National Institute of Standards and Technology (NIST) Advanced Radiometer (NISTAR). At the L1 point, DSCOVR will remain near the Sun–Earth line, viewing the entire sunlit half of the Earth continuously at near-backscatter directions with scattering angles ranging from roughly 164 to 176∘. NISTAR observes the Earth as a single pixel and provides measurements of the solar radiation reflected by the Earth, as well as the Earth's total radiant power, in four broadband channels. EPIC, with its 2048 × 2048 CCD (charge-coupled device) array, provides higher-spatial-resolution details of the daytime Earth that can be used to interpret the single-pixel NISTAR data. Note that at the L1 point the CCD array resolution yields roughly 8 km pixel sampling at the Earth's surface at nadir, though aggregating the native radiances to coarser resolution (e.g., 1024 × 1024) remains a possibility.

Numerous geophysical products will be derived from EPIC observations at 10 spectral channels ranging from the ultraviolet (UV) (317.5 nm) to the near infrared (NIR) (779.5 nm) (Yang et al., 2013). These spectral channels are sensitive to various atmospheric and surface components, and can provide information on ozone, aerosol, cloud, and vegetation properties. The EPIC cloud products will include cloud masking, cloud optical thickness (COT), and cloud effective height. The EPIC cloud mask will adopt a threshold method utilizing information from multiple spectral channels; cloud height will be derived from the oxygen (O2) A- and B-band observations (Yang et al., 2013). While the cloud mask and cloud height products are critical inputs for COT retrievals, this paper will focus solely on the EPIC cloud phase and COT algorithm and its potential uncertainties.

The EPIC COT product will be produced with the same core algorithms as those used by the operational MODerate-resolution Imaging Spectroradiometer (MODIS) cloud optical and microphysical property product (MOD06) (Platnick et al., 2003, 2015) that is now part of a shared-core cloud retrieval algorithm suite that has been applied to other spaceborne and airborne imagers such as the Visible Infrared Imaging Radiometer Suite (VIIRS) on board the polar-orbiting Suomi-NPP (National Polar-orbiting Partnership; Platnick et al., 2013), the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) on board EUMETSAT's geostationary Meteosat (e.g., Hamman et al., 2014), and the Enhanced MODIS Airborne Simulator (eMAS) flown on NASA's ER-2 high-altitude research aircraft (e.g., King et al., 2010), among others. MOD06 provides cloud top pressure, temperature, and height retrievals from the infrared (IR) window and CO2-slicing techniques (Menzel et al., 2008; Baum et al., 2012), as well as simultaneous retrievals of COT and cloud effective particle radius (CER) using the well-known bi-spectral Nakajima and King (1990) method and cloud phase retrievals using a variety of IR brightness temperature, cloud top temperature, dual-phase CER, and spectral tests (Baum et al., 2012; Marchant et al., 2016). The MOD06 two-channel COT-CER retrieval approach couples a non-absorbing visible (VIS), NIR, or shortwave infrared (SWIR) spectral channel sensitive to COT with an absorbing SWIR or mid-wave infrared (MWIR) spectral channel sensitive to CER. However, because the EPIC spectral channels do not extend to wavelengths longer than the NIR, COT retrievals will be performed using a single-channel approach similar to that of the International Satellite Cloud Climatology Project (ISCCP) (Rossow and Schiffer, 1999) and the Multi-angle Imaging SpectroRadiometer (MISR) mission (Marchand et al., 2010), both of which assume appropriate values for CER.

A critical component of the COT retrieval process is determining the radiatively appropriate cloud thermodynamic phase. For EPIC, cloud thermodynamic phase will be inferred by imposing thresholds on cloud temperature. Due to the lack of thermal IR spectral channels, specifically those used for cloud altitude retrievals from either IR window or CO2-slicing techniques, EPIC cloud temperature will instead be converted from O2 A-band cloud effective height retrievals utilizing atmospheric profiles provided by the Goddard Earth Observing System Model, Version 5 (GEOS-5) (Rienecker et al., 2011). Oxygen, like CO2, is a well-mixed gas, and reflectance measurements within the O2 absorption region have been previously exploited for estimates of cloud altitude (e.g., Heidinger and Stephens, 2000; Rozanov and Kokhanovsky, 2004; Kokhanovsky et al., 2009). EPIC is equipped with two pairs of O2 A-band (779.5 and 764 nm) and B-band (680 and 687.75 nm) reference and absorption channels that will be used to retrieve cloud effective height (Yang et al., 2013), which can then be converted to temperature and used to infer the cloud thermodynamic phase.

In this study, the expected EPIC COT retrieval uncertainties due to the required assumption of a fixed CER value will be estimated, in addition to the sensitivity of cloud thermodynamic phase retrievals to temperature thresholds. Moreover, the impact of EPIC's coarse spatial resolution, compared to the higher-resolution polar-orbiting instruments such as MODIS (1 km resolution at nadir), on both cloud phase and COT retrievals and statistics will be explored. Proxy EPIC single-channel COT retrievals are obtained by running MOD06, assuming a fixed phase-dependent CER, on a subset of Aqua MODIS granules selected for its angular proximity to the backscatter region; expected COT errors are found via direct comparison with MOD06-like two-channel COT-CER retrievals. Cloud phase sensitivity is determined via adjustments to the cloud temperature threshold tests used for phase assignment. It should be noted that the error estimates provided here are intended to demonstrate whether a single-channel COT retrieval will be feasible for EPIC. Furthermore, this study does not consider potential errors from sources such as cloud masking and cloud height retrievals, nor does it include 3-D radiative transfer effects and sub-pixel cloud heterogeneity, though these error sources are expected to be substantial due to the large EPIC field of view (FOV) (e.g., Davis et al., 1997).

The planned operational EPIC COT product will leverage a shared-core cloud retrieval algorithm suite that includes MOD06 (Platnick et al., 2003). The shared-core concept provides consistency in retrieval methodology, ancillary data sets, etc., across multiple spaceborne and airborne instrument platforms. Numerous enhancements to this retrieval suite were made for the recent MOD06 Collection 6 (C6) reprocessing effort (Platnick et al., 2015) and will thus be included in the EPIC retrievals. C6 enhancements relevant to EPIC include new bulk ice cloud radiative models (Yang et al., 2013), improved characterization of surface reflectance for ocean (Cox and Munk, 1954a, b) and land (Schaaf et al., 2011), new cloud radiative transfer (RT) look-up tables (LUTs), and improved handling of ancillary data sets.

In addition, the C6 cloud optical property retrieval algorithm now processes and reports retrievals of pixels that are identified as being only partially cloud covered and are thus likely to deviate from the 1-D homogeneous plane-parallel assumption. These “partly cloudy” (PCL) pixels, retrievals of which are reported separate from the “overcast” pixels, are expected to be poor retrieval candidates and are generally associated with higher retrieval uncertainty and, for MODIS, increased rates of retrieval failure (i.e., the reflectance measurements lie outside the defined LUT solution space) (Cho et al., 2015). Identification of these PCL pixels is accomplished via a clear-sky restoral (CSR) algorithm (e.g., Zhang and Platnick, 2011; Pincus et al., 2012) that for MODIS uses sub-pixel cloud reflectance variability and cloud edge detection tests. For EPIC the cloud edge detection test will be applied, though the application of the sub-pixel spatial variability test depends on the availability of full-CCD-resolution observations in at least one spectral channel. Because the EPIC pixel size is much larger than that of polar-orbiting imagers such as MODIS and VIIRS, PCL pixels are expected to represent a larger fraction of the cloudy pixel population (Dey et al., 2008). Thus identifying these pixels is of increased importance, specifically as a means of assessing the quality of the optical property retrievals.

Since the EPIC channel set is limited compared to that of MODIS, there are important differences between the EPIC retrieval algorithm and its MOD06 counterpart. These include the use of a single channel, 680 nm, for COT retrievals with the assumption of fixed phase-dependent CER values due to EPIC's lack of absorbing spectral channels in the SWIR or MWIR that are sensitive to cloud particle size; note that, as is common practice, the EPIC COT retrievals, as well as all COT retrievals referred to here, are referenced to the visible 0.66 µm wavelength. Since COT can be dependent on CER, in particular where the bi-spectral COT-CER retrieval solution space is non-orthogonal such as at small COT, single-channel COT retrievals are thus prone to errors larger than those of the two-channel COT-CER retrievals. The fixed CER values are derived from global MODIS retrieval statistics (e.g., King et al., 2013) and are assumed to be 12 and 30 µm for liquid and ice phase clouds, respectively. The lack of key SWIR or MWIR channels, for which liquid and ice absorption are known to differ (Kou et al., 1993), as well as IR channels sensitive to cloud temperature/altitude, will also limit the available information content for determining cloud thermodynamic phase, i.e., liquid water, ice, or undetermined (i.e., ambiguous). Nevertheless, cloud phase can still be inferred by applying a dual threshold on cloud temperature derived from the O2 A-band effective cloud height retrievals.

All ancillary data sets for EPIC COT retrievals will be identical to what are used by MOD06, with appropriate substitutes when necessary. These include gap-filled land and snow/ice surface spectral albedos derived from MODIS (Schaaf et al., 2011), Near-real-time Ice and Snow Extent (NISE) (Nolin et al., 1998) data from the National Snow and Ice Data Center, and NCEP sea ice concentration (Grumbine, 1996, and references therein). For the operational EPIC retrievals, profiles of atmospheric temperature, pressure, water vapor, and ozone, as well as surface wind velocity, will be obtained from NASA's GEOS-5 model (Rienecker et al., 2011); for the present investigation, however, these parameters will be obtained from the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS) 6 h “Final Run” archive product (Derber et al., 1991) that is used in the current C6 MOD06 retrievals. The forward-calculated COT retrieval LUTs will be generated under assumptions and cloud models identical to those of MOD06 using the 1-D plane-parallel DIScrete-Ordinates Radiative Transfer (DISORT) method (Stamnes et al., 1988) ignoring atmospheric gaseous absorption, which is corrected for during the retrieval process; note that above-cloud Rayleigh scattering, expected to be non-negligible at 680 nm, will also be accounted for during the retrieval process on a pixel-by-pixel basis using the iterative technique of Wang and King (1997), similar to MOD06. For the forward-calculated cloud retrieval LUTs, the liquid cloud droplet size distribution is assumed to be gamma distributed with effective variance of 0.1, with band-averaged scattering properties obtained by integrating spectral Mie calculations over the appropriate spectral response functions. Likewise, the ice crystal size distribution is also assumed to be gamma distributed with effective variance of 0.1, with band-averaged scattering properties integrated over the spectral database of Yang et al. (2013); the ice clouds are assumed to be composed only of severely roughened compact aggregates of eight solid hexagonal columns (hereafter referred to as severely roughened aggregated columns), an assumption shown to provide better retrieval closure between solar-, IR-, and lidar-based retrievals of cirrus COT (Holz et al., 2015).

To investigate the expected EPIC cloud thermodynamic phase sensitivity, 1 month of Aqua MODIS daytime granules is selected for the analysis, namely April 2005. A subset of these granules that has angular coverage within the backward scattering angle region (164 to 176∘) to be viewed by EPIC is selected for the COT retrieval error analysis. The use of 1 month of data ensures that a variety of liquid and ice phase cloudy scenes are included, with sufficient sampling over both land and ocean surfaces. Figure 1 shows the geographic distribution of the counts of April 2005 Aqua MODIS pixels having scattering angles in the backscatter region (Θ > 164∘).

COT retrieval errors resulting from the fixed CER assumption are characterized using two versions of the MOD06 algorithm applied to the Aqua MODIS backscatter granule subset, specifically a baseline version providing the full two-channel COT-CER retrievals and an EPIC proxy version assuming the fixed CER values of 12 and 30 µm for liquid and ice phase clouds, respectively. To simulate the EPIC 680 nm channel, both MOD06 versions use the 660 nm MODIS wavelength channel for COT retrievals.

For this study, cloud phase is determined by applying thresholds to the MOD06 1 km cloud top temperature (CTT) product; i.e., CTT ≤ 240 K is assigned ice phase, CTT ≥ 260 K is assigned liquid phase, and 240 K < CTT < 260 K is assigned undetermined, i.e., ambiguous, phase; these thresholds are derived from a global analysis of the MODIS cloud top temperature and CALIOP (Cloud-Aerosol Lidar with Orthogonal Polarization) cloud phase products (Marchant et al., 2016). The error in cloud phase derived from this simple threshold method is determined by comparing with the operational MOD06 results for all April 2005 daytime granules (i.e., no backscatter angle filtering is applied). The sensitivity of this method to threshold selection is also investigated by perturbing the above-defined 1 km CTT thresholds by discrete temperature changes of increasing magnitude (e.g., ±1∘, ±2∘, etc.) and examining the resulting cloud phase fraction changes; each perturbation is applied simultaneously to both thresholds. All COT and phase retrievals are at the native 1 km MODIS resolution, with the exception of the retrievals performed for the spatial resolution sensitivity analysis (Sect. 4.3), and pixel-level cloud masking information is obtained from the operational MODIS cloud mask product (MOD35) (Ackerman et al., 2010).

Note that for the operational EPIC COT retrieval the above thresholds for cloud thermodynamic phase determination will necessarily be adjusted. As previously mentioned, the EPIC cloud temperature will be derived from the O2 A- and B-band effective cloud height retrievals. Unlike MOD06, for which cloud top pressure is determined using information from either the IR window or CO2 channels (Menzel et al., 2008; Baum et al., 2012) that are more sensitive to the top of the cloud, the EPIC effective cloud height is more sensitive to the middle of the cloud (Yang et al., 2013). The operational temperature thresholds for EPIC cloud thermodynamic phase will be derived by collocating EPIC and MODIS observations, and using the MOD06 phase results as the baseline once the EPIC cloud height product is available. We note again that this paper focuses on the uncertainties in COT retrievals resulting from a fixed CER assumption and the uncertainties in cloud thermodynamic phase determination using only cloud temperature thresholds; uncertainties in cloud temperature due to effective height retrieval errors are beyond the scope of this paper and are left for future investigations.

At the L1 point, the DSCOVR platform will continuously observe the entire sunlit half of the Earth at near-backscatter directions, providing a unique perspective of the Earth's atmosphere, clouds, and aerosols. The DSCOVR cloud products will be derived from EPIC observations in the ultraviolet and visible spectra, with cloud optical property retrieval heritage from ISCCP and the operational MODIS cloud products (MOD06). Because a given location at the Earth's surface will be observed multiple times each day, this data set is expected to provide insight into the cloud diurnal cycle. In addition, because EPIC is equipped with a 2048 × 2048 CCD array, the DSCOVR cloud products are expected to provide sub-pixel cloud information that can be used for interpretation of broadband solar reflectance data from NISTAR, which observes the Earth as a single pixel. Moreover, the DSCOVR cloud products can be complementary to other EPIC-based data sets such as UV aerosol and trace gas retrievals.

Due to the lack of CER-sensitive channels, EPIC COT retrievals will adopt a single visible channel technique and assume appropriate phase-dependent values for CER. The uncertainties resulting from this assumption are investigated using a subset of 1 month of Aqua MODIS granules from April 2005 for which MODIS observed the same scattering angle space as that observed by EPIC, i.e., the backscatter angles between 164 and 176∘. Results show that for ice clouds the errors in the single-channel retrieval are minimal (< 2 %), while for liquid clouds the error can reach 10 % or higher for thin clouds (COT < 2) at large scattering angles (Θ > 174∘).

EPIC will only have cloud temperature information from its O2 A- and B-bands; hence for this study, cloud thermodynamic phase is determined using cloud temperature thresholds only. Compared to the operational MOD06 cloud optical property phase results that use a suite of tests based on IR cloud top and spectral SWIR-based CER retrievals (Marchant et al., 2016), cloud phase determined from cloud temperature thresholds only is expected to yield more pixels with undetermined, or ambiguous, phase results. An investigation of the sensitivity of cloud phase determination to temperature thresholds shows that a 10 K difference in thresholds results in absolute phase fraction changes of up to 14 %, though it should be noted that this sensitivity necessarily assumes that the thresholds are appropriately defined with respect to the distribution of cloud phase versus cloud temperature.

Finally, we note that potential uncertainties in the EPIC cloud mask and cloud temperature retrievals are not considered in this study. Furthermore, the COT retrieval approach assumes that the pixels are overcast and homogeneous, consistent with the plane-parallel forward radiative transfer assumption. Errors resulting from sub-pixel cloud heterogeneity and 3-D radiative transfer effects are quite beyond the scope of this investigation and are in practice difficult to quantify. They nevertheless can be expected to be sizeable given the relatively large size of the EPIC field of view. Cloud masking errors, i.e., erroneously determining a given FOV is cloudy or clear, are likewise difficult to quantify, much less define, and are thus ignored here.