3.1.

GCOM-C/SGLI Algorithm

The cloud properties algorithm used for the GCOM-C/SGLI is named the Comprehensive Analysis System for Cloud Optical Measurement (CAPCOM). The algorithm is an enhanced version of the solar reflectance method algorithm initially used for the NOAA AVHRR data analysis.¹⁷ It is adjusted to the specific spectral channel characteristics and response functions of the GCOM-C/SGLI. During the prelaunch phase of this satellite mission, areal tests are conducted on various algorithms (among which, the cloud properties’ algorithm), to evaluate their performance. Based on these tests, the capacity of the algorithms to produce quality data that fulfill the satellite target accuracies is assessed. The input data for the implementation of this algorithm are radiances and geometry data of the ADEOS-II/GLI satellite (precursor of the GCOM-C/SGLI). The theoretical basis of the algorithm² and other related information are summarized below. The algorithm uses radiances of three channels: a nonabsorption band (1.05 μm), an absorption band (2.21 μm preferentially used here compared with the 1.65 μm also available), and a thermal band (10.8 μm). It simultaneously retrieves the COT at 0.5-μm wavelength and the CLER using the combination of 1.05 and 2.21-μm wavelengths, as these bands primarily depend on the COT and the effective particle radius, respectively. The other input data needed for the retrievals are a cloud detection mask, ancillary data consisting of the surface albedo at the 1.05 μm and $2.21/1.65\mu \text{m}$ bands, the surface temperature, the atmospheric temperature, water vapor, and height profiles (from the Japan Meteorological Agency re-analysis data). The retrieval process starts with the determination of the CTT. This parameter is used to locate the position of the cloud prior to the application of atmospheric and surface corrections (using ancillary data from thermal and water vapor profiles and surface albedo) on nonabsorption and absorption channels. These correction data are then used for the final retrieval of the other parameters, among which are the COT, CLER, and CLWP. One of the specificities of this algorithm is the use for the entire retrieval process of three bands belonging to the same sensor component, i.e., the SGLI-IRS (this is supposed to help to avoid errors related to pixel-pixel mis-registration between channels of different sensor components), in contrast to MODIS where the channels used belong to at least two sensor components. As in most of the algorithms, such as those of the GCOM-C/SGLI or MODIS, geophysical properties (COT and CLER) retrievals are typically solved by comparing the measured reflectances with entries in a lookup table (LUT) and by searching for the combination of COT and CLER that gives the best fit.⁷ The first property to be retrieved in the GCOM-C/SGLI algorithm is the CTT. It is obtained through the infrared window method using atmospherically corrected brightness temperatures at 10.8 μm. For all the cloud properties examined in this article, four LUTs are used under a water cloud scheme for the retrievals: cloud-reflected radiance in bands 1.05 and 1.65 μm, transmissivities and reflectivity in bands 1.05 and 2.2 μm, respectively, and band 10.8-μm transmissivity. The grid system of the LUTs is composed of the equivalent water vapor amount of above cloud, the cloud layer, and below cloud, and then the cloud top height and cloud geometrical thickness, the satellite zenith, solar zenith, relative azimuth angles, and the simulated COT and CLER. To build these LUTs simulating the satellite signal, an efficient radiative transfer scheme^18–20 is used. A log-normal size distribution, function of the particle radius with a mode radius related to the effective particle radius as $r_e=r_0{e}^{2.5{\sigma }^2}$, a log-standard deviation of the size distribution for marine stratocumulus clouds ($\sigma =0.35$), and a Lambert surface (underlying surface) are assumed for the computations. Based on the radiative transfer theory for parallel-plane layers with an underlying Lambert surface, the unexpected radiation components, such as the solar radiation reflected by the ground surface and the thermal radiation emitted from the cloud layer and the ground surface, are removed from the satellite-received radiance in order to decouple the radiation component reflected by the cloud layer. The solar radiation reflected by the ground is removed through the use of the ground albedo data. When the removed radiation significantly dominates over the signal, the algorithm’s main iteration may not converge. This mainly occurs at optically thin clouds. When that is the case, the analysis does not go further and is consequently cancelled. Multiple reflections between the ground surface and the upper layer are taken into consideration, though their effect may be negligible, especially for optically thin clouds and ground surfaces with low reflectance. But, with optically thick clouds and large values of ground albedo, the effect is relatively large at visible wavelengths.

3.2.

MODIS Algorithm

In contrast with the GCOM-C/SGLI cloud algorithm, where all the retrievals are made from a single program, the MODIS retrievals are obtained through two algorithms: the optical thickness, effective particle radius, and thermodynamic phase algorithm (MOD05) for the COT and CLER retrievals,⁷ and the cloud top properties and thermodynamic phase algorithm²¹ for the retrievals of the CTT, the cloud top pressure (CTP), etc. Below is a summary description of these algorithms, as extracted from their respective algorithm theoretical-basis documents (ATBD).

4.1.

Cloud Top Temperature

In cloud properties algorithms, such as the GCOM-C/SGLI or MODIS algorithm, the vertical position of the cloud along the troposphere is important for the determination of the magnitude of the atmospheric corrections to be applied on the initial TOA radiances. The determination of the position of the cloud is essentially based on the CTT data. As seen earlier, the CTT of the GCOM-C/SGLI is obtained based on the infrared window method, while that of MODIS is either derived from the ${\mathrm{CO}}_2$ slicing (high and middle clouds) or from the infrared window method (low clouds). Figure 2 shows the CTT retrieved from cloud scenes of May 23, 2003, in North America (between Canada and the USA) for the GCOM-C/SGLI and the MODIS algorithms. This figure shows that in cold cloud areas, the GCOM-C/SGLI has higher CTT values than Terra-MODIS (e.g., the relatively large and nearly vertical cloud band at the right of the images where MODIS shows a larger extension of temperatures of the range $<240\mathrm{K}$ than GCOM-C/SGLI). At warm cloud areas, MODIS clouds tend to be warmer than those of the GCOM-C/SGLI (e.g., above 50°N at the left of the scenes). Figure 3 represents the scatter plot [Fig. 3(a)] of the CTT between MODIS and GCOM-C/SGLI, and then the separate frequency histogram distributions [Fig. 3(b)] of the datasets of the satellites. These graphs confirm the tendencies noticed with the CTT comparison images, i.e., the GCOM-C/SGLI is systematically overestimated (compared with Terra-MODIS) in cold to relatively warm clouds. This positive bias of GCOM-C/SGLI CTT gets larger as we move toward mid-thermal clouds, i.e., at MODIS CTT near 245 to 250 K. A slightly negative bias occurs at warm clouds as the MODIS CTT tends to have higher CTT than that of the GCOM-C/SGLI. The frequency distribution histogram related shows that MODIS has a higher frequency of cold clouds ($<240\mathrm{K}$) and very warm clouds ($>275\mathrm{K}$) than the GCOM-C/SGLI, while the latter shows gradually abundant mid-thermal to warm clouds.

Fig. 2

(a) Cloud top temperature (CTT) retrieved from North America by GCOM-C/SGLI SGLI (May 23, 2003, 18:25 UTC) and (b) Terra-MODIS (May 23, 2003, 18:10 UTC).

Fig. 3

Scatter CTT diagram corresponding to Fig. 2 and showing the $1:1$ relation between the GCOM-C/SGLI and the Terra-MODIS (a); then, the frequency data distribution histogram of these individual datasets (b).

4.2.

Cloud Optical Thickness

After the position of the cloud along the troposphere has been determined and the necessary atmospheric corrections on the nonabsorption and absorption signals are made, the COT is retrieved by matching the corrected cloud reflectances with the LUT reflectances. The present version of the GCOM-C/SGLI cloud algorithm determines the COT LUT in the range 0 to 70, while that of Terra-MODIS is in the range 0 to 100. Consequently, the quantitative comparison between the two algorithms will be limited to COT values equal or below 70. Figure 4 shows the COT retrievals from similar scenes as those used for Fig. 2. This figure shows that, in low COT areas, the GCOM-C/SGLI has thicker clouds than MODIS (e.g., left of the scenes’ areas and below 50°N latitude showing larger extension of clouds in the 0 to 10 COT range in the MODIS scene than that of GCOM-C/SGLI). This is also true in relatively thick cloud areas (e.g., the vertical cloud band at the right of the scenes shows a wider spread of large COT in the GCOM-C/SGLI scene than that of MODIS). Figure 5 is the $1:1$ scatter plot between the two algorithms [Fig. 5(a)] and the corresponding frequency distribution histogram of the individual datasets [Fig. 5(b)]. The scatter plot shows a good alignment of the data along the $1:1$ line between the two algorithms. However, the dispersion around this line tends to increase with the increase in COT. In general, there is an apparent overestimation of the COT at thin and relatively thick clouds of GCOM-C/SGLI against MODIS (also, there is a higher frequency of thin and very thick clouds in MODIS than GCOM-C/SGLI), while the latter seems to slightly dominate in very thick cloud areas ($\mathrm{COT}>50$).

Fig. 4

Same as Fig. 2 but for the cloud optical thickness (COT).

Fig. 5

Scatter diagram of the COT corresponding to images in Fig. 4 and showing the $1:1$ relation between the GCOM-C/SGLI and the Terra-MODIS (a) data; then, the individual datasets frequency distribution histogram (b).

4.3.

Cloud Particle Effective Radius

The CLER is simultaneously retrieved with the COT. This parameter is mostly a function of the absorption channels (1.6, 2.1, or 3.75 μm). It reflects the size of the cloud particles. Its upper limit in the present version of the GCOM-C/SGLI algorithm is 62 μm, while that of MODIS is 100 μm. The comparison between the two satellites will then be limited to CLER values below 62 μm. Figure 6 shows this comparison. The visual matching is mostly satisfactory with nearly similar variations of CLER observed between the images. However, in small particle-dominated areas, the GCOM-C/SGLI tends to show lower CLER than Terra-MODIS (e.g., below 50°N latitude and the corner between 105°W and 110°W longitude, and then 52° latitude and around 105°W longitude). At large CLER areas, differences between the two algorithms’ retrievals are hardly apparent. The good match between the retrieval images is also reflected by the graphs of Fig. 7, where the scatter plot [Fig. 7(a)] between the two retrieval datasets and the frequency distribution histograms [Fig. 7(b)] are presented. In the scatter plot, a good alignment of the two datasets along their $1:1$ line is clearly visible and only few data show large deviations from that line. The distribution histogram shows that in the lowest range data ($<10\mu \text{m}$), the GCOM-C/SGLI has a larger data frequency than the MODIS. At mid-ranges, MODIS slightly dominates and at high ranges the frequency seems to be nearly similar. Around the $1:1$ line of the scatter plot between the two satellites, some of the GCOM-C/SGLI data show a constant value ($\mathrm{CLER}=9.0$). This happens because of the difficulties of the iteration method to properly derive the CLER of these clouds; therefore, the radius is approximated to a fixed value of 9.0.

Fig. 6

Same as Fig. 2 but for the cloud particle effective radius (CLER).

Fig. 7

Scatter diagram of the CLER corresponding to Fig. 6 and showing the $1:1$ relation between the GCOM-C/SGLI and the Terra-MODIS (a); then, the related individual datasets distribution histogram (b).

4.4.

Cloud Liquid Water Path

The CLWP, expressed in $\mathrm{g}/{\mathrm{m}}^2$, represents the total amount of liquid water along a cloud column. It is obtained by combining the COT and CLER through the equation:

Figure 8 presents the CLWP scatter plot between the GCOM-C/SGLI and the MODIS and the individual CLWP datasets’ frequency distribution histograms. The figure shows how many of the deviations of the individual parameters of COT and CLER cancel out when combined to form the CLWP. This leads to a very good correlation between the satellites’ retrievals. Also, the frequency distribution of the two datasets is very much even along the CLWP ranges.

Fig. 8

The cloud liquid water path (CLWP), expressed in ${\mathrm{g}/\mathrm{m}}^2$, shows how many of the deviations noticed with the individual COT, and CLER parameters cancel each other to give a very good CLWP relationship between the two satellites.

6.1.

Discrepancies Between the Retrievals

In the previous sections, the results of the GCOM-C/SGLI cloud algorithm performance tests were presented. In these tests, cloud retrievals from this new satellite cloud algorithm were evaluated through a comparison of these retrievals with MODIS cloud products. Also, the capacity of the algorithm to produce data that fulfill the accuracy targets of the GCOM-C/SGLI mission was assessed. The first parameter analyzed was the CTT. It showed large RMSE, particularly acute around mid-thermal clouds. This was the result of the GCOM-C/SGLI CTT systematic overestimation (compared with MODIS) in virtually all representative areas of the globe. The COT showed good alignment and nearly equal dispersion around the $1:1$ line, but RMSE% is slightly beyond the satellite mission accuracy target. The CLER showed better results than the COT as the accuracy targets were met in most of the cases. However, mid-CLER ranges seemed generally overestimated. This contributed to increase the average CLER values of the GCOM-C/SGLI against MODIS. To understand all these differences, we will examine the effects of the uncertainties in the cloud type interpretation, the cloud thermal properties, the solar and satellite geometries, etc., on these parameters.

First, the effects of the uncertainty in the cloud-type interpretation on the cloud retrievals are examined. For this purpose, a simple cloud-type classification, using the split-window channels’ differences,²³ i.e., the brightness temperature difference between the 11 and 12 μm (BT11—BT12) channels against the brightness temperature at the 11 μm (BT11), is used. The parameters involved in this classification (BT11 and BT12) are more independent in this article than the optical parameters used in other classifications. Four types of clouds are initially identified in this split-window channels’ classification: cirrus (Ci), dense cirrus (D-Ci), cumulonimbus (Cb), cumulus (Cu), and others or mixture of remaining clouds (Mix). For the purpose of our article, the cirrus clouds will be divided in two: cold cirrus or C-Ci ($<250\mathrm{K}$) and warm cirrus or W-Ci ($\mathrm{BT}11>250\mathrm{K}$). Figure 10 shows the RMSE and bias at each cloud type for all the examined parameters. The CTT RMSE and bias are large not only in the Ci, but also in the Mix clouds, while the cumulonimbus (Cb) and the dense cirrus (D-Ci) clouds show the lowest RMSE (even below 1.5 or 3.0 K, which are, respectively, the satellite mission accuracy target at sea level and above) and bias. Between the Ci clouds, W-Ci clouds show the largest RMSE and bias. The Cu clouds have either higher or close RMSE and bias than Cb clouds. The trends noticed with the CTT are nearly similar with those of the COT, CLER, and CLWP, i.e., the lowest RMSE% are found with Cb and D-Ci, and the largest are seen with the W-Ci, then gradually lower values with the C-Ci and Cu. The areal comparison between the two algorithms showed that the RMSE or RMSE% for all parameters had the largest values consistently above forest areas. One major explanation of this situation is that most of the forest areas are around the equator, and this region has annually the highest frequency of cirrus clouds. These clouds have shown large discrepancies between the two examined satellites’ datasets.

Fig. 10

RMSE (K) or RMSE (%) and bias or bias (%) change with the cloud type.

If we only compare the BTD of the clouds, it appears that the CTT, COT, and the CLER RMSE tend to increase with the increase of the BTD, at the upper end of which are the Ci clouds. The variation of the brightness temperature with respect to the wavenumber for the clouds is mainly due to multiple scattering and absorption processes inside the clouds.²⁴ Because the emission is the dominant process in these thermal infrared wavelengths, the cloud absorption interpretation between the two algorithms may differ. To understand these processes, we will now examine the interpretation of the contribution of the cloud emission to the discrepancies between the algorithms. In the infrared wavelength, the radiance perceived by a sensor is a function of the radiation emitted from the target and the radiation emitted and absorbed by the intervening atmosphere, integrated over the response function of the sensor.¹⁵ The atmosphere radiation is treated by the atmospheric correction while the radiation of the target (cloud) is part of the cloud correction scheme. At a specific wavelength, the energy emitted by clouds is controlled by the average temperature of the cloud surface and the cloud emission properties.²¹ The cloud emission correction factor is mostly related to the cloud emissivity. When the emissivity at the 11 μm reaches 1, the clouds become black, and when it is lower, the clouds are considered as gray bodies. The change in the cloud effective emissivity, i.e., in cloud emissivity and/or cloud fraction, may have an important effect on the TOA radiance.²¹ For a given observed radiance, if the emissivity is overestimated, the cloud top pressure is also overestimated, putting it too low in the atmosphere.²¹ Figure 11 shows the comparison between the retrieved images of the CTT by the two algorithms and the cloud emissivity, and the cloud fraction maps obtained from MODIS. Areas where the GCOM-C/SGLI CTT is strongly higher than that of MODIS correspond to low emissivity areas, as shown by the red circles on the CTT and emissivity maps. There is no visible change in the cloud fraction accompanying the differences in CTT between the two satellites. A quantitative illustration of the relationship between the CTT discrepancies and the emissivity is shown in Fig. 12. The histogram of the emissivity variation with RMSE and bias shows a gradual decrease of the RMSE and bias with the increase of the emissivity: low cloud emissivity areas tend to correspond to the major areas of CTT discrepancies. The difference between the two algorithms is minimal (even lower than the GCOM-C target accuracy at sea level) for emissivity higher than 0.9, while at lower emissivity, the two algorithms gradually diverge. Based on these observations, we can separate the CTT values in two groups: large emissivity ($>0.9$) and low emissivity ($<0.9$) CTTs. The scatter plots between the two algorithms clearly show that low emissivity clouds are at the origin of the major differences between the algorithms. The large emissivity clouds match very well (strong correlation) and show very few differences between the two algorithms. The cloud emission method used by the GCOM-C/SGLI may, therefore, insufficiently correct the cloud brightness temperature data used for the CTT. It was also noticed that low emissivity clouds tend to show stronger discrepancies (though less pronounced than in the case of the CTT) than high emissivity clouds for the COT, CLER, and CLWP parameters.

Fig. 11

CTT maps of GCOM-C/SGLI (a) and Terra-MODIS (b); then, the corresponding cloud emissivity (c) and cloud fraction (d) from Terra-MODIS. The red circles represent areas where the GCOM-C/SGLI CTT is higher than that of Terra-MODIS and the emissivity changes related to those areas.

Fig. 12

Cloud emissivity histogram variation with the CTT RMSE (a), the distribution of data used in histogram (b), and scatter plot of the CTT between the two algorithms for emissivity $ϵ$ more and less than 0.9 (c and d, respectively).

The effects of the uncertainties of the solar and satellite geometries on the cloud properties discrepancies did show consistent patterns. These effects are therefore considered marginal. Besides the factors examined already, the CTT, COT, CLER, and CLWP RMSEs may be influenced by other factors as well. Among these factors we may cite, the uncertainties in the calibration of the absorption or nonabsorption channels, the cloud detection, the atmospheric correction, the ground surface reflection (based on the surface albedo) correction, the water vapor, and temperature profiles (ancillary data). These various factors are examined in the following sections through sensitivity tests analyses.

6.2.

Sensitivity Tests

Measured and calculated radiances differ because of the cumulative effects of instrument noise, spectral response function errors, inadequate knowledge of the atmospheric and surface state, radiative model approximations, etc.⁷ In this section, we will try to examine the effect of perturbations of some of the cloud external assumptions on the retrievals.

6.2.2.

Sensitivity to channel 1.05-$\mathsf{\mu }\mathsf{m}$ radiance uncertainty

The 1.05-μm channel is the nonabsorption channel used by the GCOM-C/SGLI algorithm for the COT retrieval. The equivalent channel used by MODIS is channel 0.65 μm for land, 0.86 μm for ocean, or 1.240 μm for snow/ice. A 5% increase in the radiance of the 1.05-μm channel increases the RMSE% by 18% and 0.7% for COT and CLER, respectively. A 5% decrease in the radiance of the same channel leads to an RMSE% decrease of 1.8% and 1.6% for COT and CLER, respectively. Because the nonabsorption channels used by both satellites’ algorithms for the COT retrieval are different, it was worth examining the impact of these differences on the RMSE%. A simple way to do that is to compare, for example, the difference between the reflectance of similar channels with those of MODIS, i.e., 0.65 μm (land) or 0.86 μm (ocean) existing in the GCOM-C/SGLI and the 1.05-μm reflectance effectively used in the latter satellite algorithm. Figure 13 shows how these differences affect the RMSE%. As the reflectance difference increases, the COT RMSE% also increases. When there is almost no or very little difference between the two channels, the COT RMSE% is lower than the GCOM-C/SGLI target accuracy ($<50\%$). At larger reflectance differences, the RMSE% is also large and may be as high as 40%. Differences in the sensitivity of the 0.65 μm or 0.86 and 1.05-μm channels appear to explain some of the discrepancies between the COTs of the two algorithms.

Fig. 13

Differences between reflectance (R) values of the nonabsorption channels of the GCOM-C/SGLI (0.65 μm or 0.865 and 1.05 μm), i.e., R0.65–R1.05 for land and R0.865–R1.05 for ocean, against the GCOM-C/SGLI COT RMSE% and bias%. The 0.65 and 0.865 μm are equivalent to the channels used by Terra-MODIS in the COT retrievals. The land area is Siberia (a) and the ocean area is the South Atlantic (b).

6.2.7.

Sensitivity to the cloud detection uncertainty

The pixels used in the implementation of the GCOM-C/SGLI cloud properties algorithms are selected following a cloud screening or cloud mask algorithm named CLAUDIA.²⁵ The cloud mask allows the distinction of clouds following a clear sky probability scale ranging from 0 (100% cloud probability) to 1 (100% clear sky probability). Intermediate values have either a higher probability of clear sky (above 0.5) or a higher probability of cloud (below 0.5). All the values except “1” (100% clear sky probability) are used in the cloud properties determination. The comparison of these retrievals with those of MODIS shows that the cloud detection affects both the COT and the CLER. Figure 14 shows cloud retrievals above the Amazon forest, one of the forest areas that often showed large RMSE% of COT and CLER for all clouds, and 100% or less cloud probability. It appears clearly that the low-probability cloud pixels’ bad alignment along the $1:1$ line between the algorithms contribute to the increase of the RMSE% of both the COT and the CLER and the deterioration of the correlation between the datasets. No clear difference was noticed in the case of the CTT. The strongly overestimated CLER pixels which often blur the correlation between the two algorithms clearly belong to the low-probability clouds. When these clouds are removed, there is a good alignment along the $1:1$ line between the algorithms. The high RMSE% seen in COT and CLER are strongly due to the differences in the interpretation of the radiation of the low-probability clouds between the two algorithms.

Fig. 14

Amazon forest clouds: comparison between GCOM-C/SGLI and Terra-MODIS for all cloud pixels (a), 100% cloud probability pixels (b), and less than 100% cloud probability pixels (c) for the COT, CLER, and CTT.