Retrieval of Outgoing Longwave Radiation from COMS Narrowband Infrared Imagery

Outgoing longwave radiation (OLR) at the top of the atmosphere represents the total outgoing radiation flux emitted from the Earth's surface and atmosphere in the infrared (IR) wavelength. OLR varies with changes in atmospheric thermal conditions, water vapor, and cloudiness owing to different emission temperatures. Accordingly, OLR has been used to study various topics in weather and climate studies, such as atmospheric water vapor distribution (Sinha and Harries, 1997) and large-scale ocean-atmosphere coupled processes and climatic effects (Madden and Julian, 1971; Lau and Chan, 1983; Kiehl et al., 1998; Jeong et al., 2005). Furthermore, as OLR is strongly linked to the radiative-convective process for clouds, it has been commonly used to investigate the cloud feedback mechanism (e.g., Lau et al., 1996; Lindzen et al., 2001; Hartmann and Michelsen, 2002; Ho et al., 2002; Choi and Ho, 2006; Lindzen and Choi, 2009, 2011; Cho et al., 2012).

Global observations of OLR have been obtained from satellite instruments in the form of either broadband or narrowband (of bandwidth within 1 μm) radiometers. Unlike narrowband radiometers, broadband radiometers provide near-direct observations of the terrestrial emission of OLR by measuring radiances over the whole IR spectral range. Most broadband radiometers have been loaded onto polar orbiting platforms for the sake of obtaining global coverage, and their observations are therefore limited to specific local times. This type of broadband radiometer includes the Earth Radiation Budget Experiment (ERBE; Barkstrom et al., 1989) from 1985-1999 and the Clouds and Earth's Radiant Energy System (CERES; Wielicki et al., 1996) from 2000 to the present day. Note that, on the geostationary orbit, only the European satellite Meteosat-8 has a broadband radiometer, the Geostationary Earth Radiation Budget (Harries et al., 2005). This exceptional layout has allowed measurements of the hourly variation in radiation flux, but the areal coverage is limited to Europe and Africa.

Narrowband radiometers can be used to observe a confined portion of the IR spectral range within a bandwidth of about 1 μm. Their observations can be converted to OLR for a broad spectral range (usually 2-100 μm) through a pre-calculated relation between narrowband radiances and OLR (Ellingson and Ferrado, 1983). Because the OLR data from the narrowband radiometers loaded on a geostationary platform have high temporal resolution (usually one hour), the investigation of rapid atmospheric processes (diurnal cycle, convective and radiative processes, and the interaction among them), which are of vital importance to both the hydrological cycle and energy budget, is possible (e.g., Ellingson and Ba, 2003; Ho et al., 2008; Park et al., 2011; Taylor, 2012; Cho et al., 2012). A disadvantage of such OLR retrievals may be lower accuracy than near-direct (i.e., broadband) OLR measurements. Therefore, it is important to establish the accuracy of OLR from geostationary satellite narrowband observations, which is the ultimate goal of this study.

Several approaches to selecting the best channel(s) have been suggested for retrieving OLR. A well-known method is estimation of OLR using a single IR window channel (centered at 12.0 μm; OLR_12.0 algorithms), as a window channel radiance is sensitive to the thermal emission from the lowest radiating surface, i.e., the Earth's surface or cloud top (Ellingson and Ferrado, 1983). Together with a single window channel, (Schmetz and Liu, 1988) included a water vapor channel (e.g., 6.7 μm) to consider the effect of water vapor in the troposphere (OLR_6.7+10.8 algorithm). On the other hand, (Inoue and Ackerman, 2002) used the channel difference between two adjacent window channels (e.g., 10.8 and 12.0 μm) for better accuracy over regions with low-level clouds or no clouds (OLR_10.8+12.0 algorithm).

These methods form the basis of the current operational OLR retrieval algorithms for several international weather satellites: OLR_12.0 for the National Oceanic and Atmospheric Administration satellites, and OLR_6.7+10.8 for the European Meteosat series of satellites. In addition, we suggest an OLR retrieval algorithm using all three channels (6.7, 10.8, and 12.0 μm) (OLR _All algorithm). These algorithms, however, have not yet been compared all together with observations from the same satellite, so any discrepancy in the accuracy of the algorithms still remains unknown.

The first Korean geostationary satellite——Communication, Ocean and Meteorological Satellite (COMS)——was launched in June 2010 and has been releasing OLR as one of its operational products (Choi et al., 2007). Because the COMS full field-of-view (FOV) includes a vast domain [roughly (50^°S-50^°N, 70^°-170^°E)——the region within 70^° of the satellite zenith angle in Fig. 1], the advantage of OLR retrievals from COMS observations is the capture of the hourly variation of OLR in Asian and Australian monsoons, intraseasonal oscillations such as the Madden-Julian Oscillation, tropical cyclones, etc. A great challenge of this study is to establish consistent accuracies of the OLR retrievals over the full COMS FOV; the domain includes a variety of atmospheric profiles (tropics and midlatitudes), surface types (ocean, forest, and desert), cloud conditions, and even different seasons in the Northern and Southern hemispheres. The primary purpose of this study is to evaluate the COMS OLRs (OLR_12.0, OLR_6.7+10.8, OLR_10.8+12.0, OLR _All) with reference data, i.e., CERES OLR from the Terra satellite.

A schematic flow chart outlining the scope of the study is shown in Fig. 2. This study includes pre-calculation (Fig. 2a) and the retrieval process (Fig. 2b) of COMS OLR, as well as its validation (Fig. 2c). In the following section, COMS and CERES observations used in this study are briefly explained and the validation methodology concerning how the data from the two satellites were collocated is described (interpolation and homogeneous tests in Fig. 2c). To retrieve the OLR from the COMS observations, it is essential to calculate various coefficients through regression analyses based on radiative transfer model (RTM) calculations (Fig. 2b), which are described in section 3. Validation results (statistics as results of Fig. 2c) are presented in section 4. Finally, a summary and conclusions are given in section 5.

Figure 1.  The full field of view of the COMS satellite. Land areas are colored black, and oceans are in gray. The area where the COMS satellite zenith angle is larger than 70^° is marked in red.

Figure 2.  Flow chart presenting the scope of the current study. This paper includes (a) pre-calculation for COMS OLR algorithm, (b) OLR retrieval from COMS observations, and (c) validation of OLR with CERES. The pre-calculation and OLR retrieval are described in section 3. Validation results are shown in section 4.

This study uses hourly radiances from three different channels from COMS for the period of April 2011. The data are obtained from the Korean National Meteorological Satellite Center. The spatial resolution of the data is 5 km. Among the five channels of COMS centered at 0.67, 3.7, 6.7, 10.8, and 12.0 μm, the OLR algorithms use radiances at 6.7, 10.8, and 12.0 μm channels. This study includes the full COMS FOV as a validation domain (Fig. 1); the analyses of the OLR retrievals during the entire month may be sufficient for the statistical evaluation.

The CERES OLR retrievals used as reference data in this study are from the ERBE-like product named ES8 from the CERES instrument onboard Terra (CER_ES8_Terra-FM1_Edition3). In particular, this study uses hourly instantaneous FOV scanner longwave radiation fluxes at the top of the atmosphere. Two instruments, CERES flight models 1 and 2 (FM1 and FM2, respectively) are onboard the Terra satellite, and FM1 data have been analyzed. The CERES passes a region at 1030 local time, in sun-synchronous polar orbit (with an altitude of 705 km). The spatial resolution of CERES OLR is approximately 20 km at its nadir and its root-mean-square error (RMSE) is known to be 2.7 W m^-2 (Young et al., 1998). More detailed information on the CERES data can be found online at http://eosweb.larc.nasa.gov.

Figure 3.  Schematic representation of the OLR retrieval algorithms. The input data include radiance observations (L) in the three channels and satellite zenith angle (θ). The procedure consists of two main conversions: (i) radiance to irradiance, and (ii) irradiance to OLR.

This study assumes that a comparison based on a 1^°× 1^° grid reasonably reduces sampling error caused by the different footprint sizes between COMS and CERES (5 km ×5 km for COMS and 20 km × 20 km for CERES). Thus, for comparison of the two types of OLRs, we reproduce the collocated COMS and CERES OLRs on the basis of 1^°× 1^° earth reference grids. Here, CERES pixels observed within 15 minutes from the COMS observation time are selected. Other information such as scene identification and the standard deviation of instantaneous OLR are also compiled.

The use of the Terra CERES data for validating COMS algorithms becomes even more complicated by the differences in viewing angles and observation times between the two satellites (Ba et al., 2003). Specifically, these two satellites can observe cloud process forming and dissipating rapidly in a few minutes at different observation times with different viewing angles. To exclude these cases in the validation of this study, we perform a "homogeneous test" on each 1^°× 1^° grid, as in (Ba et al., 2003). Thus, the validation only uses the remapped CERES grids with the standard deviation of the instantaneous CERES OLR less than 2% of the mean value. A 2% homogeneity test corresponds to about 24, 7, and 4% variations about 50% cloud cover for black clouds at 3, 7, and 10 km, respectively. Finally, 40 495 homogeneous grids are selected for validation, and the results are presented in section 4.

Prior to discussing validation results, we describe the OLR retrieval process (section 3.1) and the method for calculating coefficients through regression analyses of RTM simulations (section 3.2). In addition, the intrinsic accuracy of the algorithm is estimated based on analyses of RTM simulations.

A simple flow chart for OLR retrieval is shown in Fig. 2b, but Fig. 3 illustrates more detailed OLR retrieval procedures. There are two main conversion procedures: (1) radiance-to-irradiance at a given spectral band; and (2) irradiance-to-OLR (i.e., narrowband-to-broadband). First, the irradiance (F in W m^-2 μm^-1) must be obtained by integrating the radiance within a finite solid angle (L in W m^-2 sr^-1 μm^-1) over all directions. For isotropic emission, this integration can be performed using the simple relation F=πL. However, the Earth's radiance at a certain altitude is not isotropic in direction. Thus, the L-F relation must be corrected to a function of viewing geometry; that is, the satellite zenith angle (θ) for the geostationary orbit (Otterman et al., 1997). The dependence of F on θ can be approximated as a limb darkening function (Abel and Gruber, 1979; Schmetz and Liu, 1988), \begin{eqnarray} F&=&[k_{1}+k_{2}({ sec}\theta-1)+k_{3}({ sec}\theta-1)^{2}]L(\theta)+k_{4}+\nonumber\\ &&k_{5}({ sec}\theta-1)+k_{6}({ sec}\theta-1)^{2} , (1)\end{eqnarray} where k₁,k₂,…,k₆ are coefficients determined by RTM simulations (Table 1).

The next step is conversion from narrowband irradiance (F) to broadband OLR. At this stage, four equations are employed, \begin{eqnarray} { OLR}_{12.0}&=&a_{0}+a_{1}F_{12.0}+a_{2}F_{12.0}^{2} ,(2)\\[0.5mm] { OLR}_{10.8+12.0}&=&b_{0}+b_{1}F_{10.8}+b_{2}(F_{10.8}-F_{12.0}) ,(3)\\[0.5mm] { OLR}_{6.7+10.8}&=&c_{0}+c_{11}F_{6.7}+c_{12}F_{6.7}^{2}+c_{13}F_{6.7}^{3}+c_{21}F_{10.8}\nonumber\\[0.5mm] &&+c_{22}F_{10.8}^{2}+c_{23}F_{10.8}^{3} ,(4)\\[0.5mm] { OLR}_{ All}&=&d_{0}+d_{11}F_{6.7}+d_{12}F_{6.7}^{2}+d_{13}F_{6.7}^{3}+d_{21}F_{10.8}+\nonumber\\[0.5mm] &&d_{22}F_{10.8}^{2}+d_{23}F_{10.8}^{3}+d_{31} F_{12.0}+d_{32}F_{12.0}^{2}+\nonumber\\[0.5mm] &&d_{33}F_{12.0}^{3} , (5)\end{eqnarray} where F_6.7, F_10.8, and F_12.0 are the irradiances at 6.7, 10.8, and 12.0 μm, respectively; and a, b, c, and d are coefficients determined by RTM simulations (Table 2). As shown in the equations, four OLR values are estimated using different methods: OLR_12.0 is determined from a second-order polynomial function of a 12.0 μm channel (Eq. 2, Ohring et al., 1984); OLR_10.8+12.0 is estimated from a first-order polynomial function of F_10.8 and F_12.0 (Eq. 3, Inoue and Ackerman, 2002); OLR_6.7+10.8 is calculated from a third-order polynomial function of 6.7 and 10.8 μm (Eq. 4, Schmetz and Liu, 1988); and OLR _All is determined from a third-order polynomial function of all three IR channels [Eq. (5)].

The different method/function may be based on the physical relationship between F and OLR, which is likely dependent on the selected domain. Over the range of terrestrial temperatures, the flux on the window channel is essentially linear with total black body irradiance for the same temperature (Minnis et al., 1995). The domain of (Inoue and Ackerman, 2002) is the subtropical eastern Pacific, where extreme high or low OLRs (usually observed over desert areas or tropical deep continental convection) are absent. Thus, a linear regression function of infrared channel and the window split channel [Eq. (3)] may be somewhat acceptable. For an extremely low IR temperature, the contribution of the IR spectral interval to the total blackbody irradiance asymptotically approaches zero while the longwave flux remains substantially greater than zero (Minnis et al., 1995). Thus, the role of the second-order polynomial function [Eq. (2)] is to permit this asymptotic behavior. The third-order polynomial regression function [Eq. (4)] seems more flexible in applying suitable F-OLR relations both over an extreme low and high OLR ranges. This is because the function has an inflection point, enabling the algorithm to have decreased/increased slope of the function below/above the point (not shown).

As shown in Fig. 2a, the coefficients in the above equations are determined by regression analyses of F on L [for the ks in Eq. (1)] and of OLR on F [for the as, bs, cs, and ds in Eqs. (2-5)]. The L, F, and OLR values are obtained from theoretical RTM calculations (e.g., Ellingson and Ferrado, 1983; Schmetz and Liu, 1988; Ba et al., 2003). In this study, the discrete ordinates radiative transfer (DISORT) model, "Santa Barbara DISORT Atmospheric Radiative Transfer" (SBDART; Ricchiazzi et al., 1998), is used. The RTM simulations are designed to apply spectral response functions to COMS channels.

Various cloud and atmospheric conditions are assumed in the calculation of radiances and irradiances for the COMS channels (L_6.7(θ), L_10.8(θ), L_12.0(θ), F_6.7, F_10.8, and F_12.0), and OLR values. These conditions are five cloud optical thicknesses (8, 16, 32, 64, 128), eight cloud top heights (2, 4, 6, 8, 10, 12, 14, 16 km), and seven typical atmospheric profiles (i.e., US62 standard, tropics, midlatitude summer, midlatitude winter, sub-arctic summer, sub-arctic winter, and desert). The vertical profiles of environmental pressure, temperature, specific humidity, and ozone mass concentration are obtained from (McClatchey et al., 1972) in the SBDART RTM package, except for the desert condition.

The desert profile is set to have the same pressure and ozone profiles as the tropics profile, but 20% of the specific humidity (in g cm^-2) and 110% of the temperature (in K). Note that this desert profile setting is artificially obtained from the tropical sounding. Therefore, the desert profile set in this study may not be representative of all desert areas over the tropics, midlatitudes, and those with wet and dry soil moisture levels. However, the current setting is reasonable in simulating extremely high OLR values that are actually observed over a desert area by the CERES sensor. Without this profile setting, the previous set of RTM simulations that only considered the desert surface condition were unable to simulate such high OLR values. Thus, the previous algorithm based on these simulations caused serious biases in all OLR algorithms over the desert area. Although the current atmospheric setting is somewhat artificial, it is reasonable in retrieving reasonable OLR values for all desert areas (except for day/clear scenes for some OLR retrieval methods).

The minimal cloud optical thickness for the radiative calculation is 8 because the radiative transfer simulations show a nearly one-to-one correspondence between L and F for thick clouds only (coefficient of determination R²=99.6%). For optically thin clouds (optical thickness <8), the simulated L-F correlation is much lower (R²=91.5%). Therefore, greater uncertainty in the OLR retrievals may be inevitable for thinner clouds.

For estimating OLR from COMS observations (Fig. 2b), it is difficult to consider all different surface types over the full COMS FOV (Fig. 1), and classification of surface types is not carried out in this study. Accordingly, all RTM simulations (Fig. 2a) are designed to have an "ocean" surface type, and the ocean's spectral emission function is applied. Here, an assumption is that F-OLR (L-F) relationships from ocean surface simulations are almost consistent with the F-OLR relationships over other surfaces (e.g., land, desert). The coefficients from ocean RT simulations are then applicable for estimating OLR over all surface types. Using the coefficients, it is inevitable that the COMS OLR retrievals may have lower accuracy over land than over the ocean, which is confirmed in section 4.

The number of streams used in the DISORT model is 20 for L-to-F conversions and 4 for F-to-OLR conversions. The spectral resolution is 0.1 cm^-1 for L and F calculations and 0.005 μm for OLR calculations within the range of 4 to 100 μm. The simulations assume a plane-parallel atmosphere, and every radiative quantity is defined at an altitude of 100 km with zero solar zenith angle and zero relative azimuth angle. Nine satellite zenith angles (θ) are used for L(θ) calculations, from 0 to 80^° at 10^° intervals.

In accordance with most operational OLR algorithms (e.g., Ohring et al., 1984), this study derives a single set of regression coefficients for each F-to-OLR conversion [i.e., as, bs, cs, and ds in Eqs. (2-5)]. The coefficients obtained in this study are listed in Table 2. That is, the current algorithms use uniform coefficients for all pixels, similar to those of (Ohring et al., 1984) and (Ba et al., 2003), whereas (Inoue and Ackerman, 2002) constructed multiple sets of coefficients for different cloud types or altitudes. As described later, the current validation show that reasonable accuracy of OLR retrievals is possible even without separation of cloud types.

In the radiative transfer simulations, the desert profile is responsible for very different simulations from other profiles, such that the determination of the coefficients in Eqs. (2-5) is subject to the inclusion of the desert simulations. To clarify the effect of the desert profile, Fig. 4 depicts the scatter plots of RTM-simulated irradiances for (a) F_6.7, (b) F_10.8, and (c) F_12.0 versus RTM-simulated OLR. Because of the high surface temperature and dry atmosphere, the F and OLR in desert conditions (red open circles) can have much higher OLR values, which cannot be obtained under the other conditions (black crosses). To examine whether the inclusion of the desert simulation is necessary for calculating the coefficients, we attempt to derive a new set of F-to-OLR coefficients (a, b, c, and ds) by excluding the desert cases. The coefficients obtained without the desert simulation are found to cause large errors in all OLR algorithms (not shown), as they cannot cover actual OLR value ranges (>300 W m^-2) from the COMS FOV, which includes the Australian desert area. The simulated OLR (or radiances) should cover a realistic OLR range in developing such regression models. Therefore, we have included the desert simulations for the final calculation of the coefficients.

Figure 4.  Scatter plots of irradiances (F) and OLRs from radiative transfer simulations. Red open circles represent simulations for desert environments, and the black crosses are all the other simulations.

Figure 5.  Scatter plots of (a) OLR_12.0, (b) OLR_10.8+12.0, (c) OLR_6.7+10.8, and (d) OLR _All against CERES OLR for 40 494 homogeneous 1^°× 1^° grids during the period 1-30 April 2011. The dotted line presents 1:1 correspondence; the dashed line is the linear regression line.

Note that, among the three irradiances (F_6.7, F_10.8, and F_12.0 in Fig. 4), F_12.0 has the closest linear relationship with OLR, as seen by the smallest spread in the F_12.0-OLR plot (Fig. 4c). On the other hand, the F_6.7-OLR relationship (Fig. 4a) exhibits a large degree of scattering, particularly for F_6.7>3 W m^-2 μm^-1, because-unlike the window channels——the 6.7 μm channel is fairly sensitive to column-integrated humidity due to the significant water vapor absorption in the upper troposphere (between approximately 200 and 500 hPa) (Choi et al., 2007). Insofar as F and OLR values decrease with increasing cloud top height, it is natural that the spread is narrow for F_6.7<2 W m^-2 μm^-1 because there should be little water vapor above high clouds.

In this subsection, we attempt to estimate the intrinsic accuracy of each OLR algorithm. The intrinsic accuracy is defined as the accuracy induced by the algorithm equation (corresponding to F-to-OLR conversion as well as L-to-F conversion). This intrinsic accuracy can be obtained by comparison between (1) OLR retrieved from each algorithm with RTM-simulated F, and (2) an RTM-simulated OLR. Here, the errors in the L-to-F conversions are assumed to be small due to a near one-to-one relation between L and F due to the proper consideration of the viewing geometry [Eq. (1)]. However, it should also be noted that the L-F relation could be uncertain for optically thin clouds (as shown in the previous section) and for satellite zenith angles larger than 70^° (Ellingson and Ferrado, 1983).

Table 3 shows the intrinsic accuracies of COMS OLR in terms of RMSEs and maximum errors. Overall, the single channel algorithm OLR_12.0 has the lowest accuracy: 9.07 W m^-2 RMSE and 26.63 W m^-2 maximum error. The additional use of the 10.8 μm channel does not seem to have any significant impact on the accuracy. The RMSE of OLR_10.8+12.0 is 8.34 W m^-2 and the maximum error is 26.09 W m^-2; the change of RMSE is only 0.6 W m^-2, and there is almost no change in maximum error compared to OLR_12.0. Finally, it is concluded that the intrinsic accuracies of the OLR_6.7+10.8 and OLR _All algorithms are better than the other two methods (2.22 W m^-2 RMSE and 5.69 W m^-2 maximum error for OLR_6.7+10.8; 1.92 W m^-2 RMSE and 5.47 W m^-2 maximum error for OLR _All). The actual accuracy of COMS OLR retrievals may also be affected by many other factors, such as channel calibration and instrument stability, which is examined in the next section.

Here, we show comparison results between COMS OLR and CERES OLR as reference data. Figure 5 displays the scatter plots of the comparisons of (a) OLR_12.0, (b) OLR_10.8+12.0, (c) OLR_6.7+10.8, and (d) OLR _All with CERES OLR for the period 1-30 April 2011. The dots in each graph represent the 40 494 homogeneous grids. For OLR_12.0,

OLR_10.8+12.0, OLR_6.7+10.8, and OLR _All, the regression slopes are 0.86, 0.96, 0.93, and 0.95, respectively; the mean differences (biases) from CERES OLR are 3.06, 1.01, -5.56, and -5.80 W m^-2, respectively; and the RMSEs are 8.64, 7.97, 6.08, and 5.58 W m^-2, respectively. Most of the RMSEs originate from large spreads in the intermediate OLR values (200 W m^-2< CERES OLR<300 W m^-2 in Fig. 5) for all four OLRs. Such errors may be caused by short-term variations of broken clouds and/or of cloud top heights (Ba et al., 2003).

The accuracies of all the OLRs are within the range of 5.5 to 8.7 W m^-2 (i.e., 2.0%-3.2% of the mean CERES OLR, 274 W m^-2). The values in the present study are slightly smaller than those of (Ohring et al., 1984) (∼ 11 W m^-2) and comparable with those of (Ba et al., 2003) (∼ 7 W m^-2). It should be noted that the area of the present study covers most of the full geostationary satellite FOV, whereas (Ba et al., 2003) covers continental United States and surrounding oceans. Taking the various atmospheric and surface conditions included in the COMS coverage into account, the accuracy in our OLR retrievals seems appropriate for drawing general conclusions.

It is also notable that the single-channel algorithm OLR_12.0 tends to overestimate OLR values (particularly, CERES OLR >300 W m^-2 in Fig. 5a). The bias of OLR_12.0 is 3.06 W m^-2, and the \(\Delta OLR_12.0/\Delta OLR_ CERES\) (∆ OLR _CERES/∆ OLR_12.0) regression slope is 1.16 (0.86). (Ohring et al., 1984) states that positive bias exists in the single-window channel OLR algorithm because samples of atmospheric soundings in RT simulations may not be representative of real atmospheric profiles. Based on Fig. 2, we determine that the F₁₂-OLR regression has a higher slope for desert cases (red circles) than for the other cases (black crosses). That is, the greater relative contribution of desert cases in simulations (one of seven atmospheric soundings in RTM simulations, as in section 3.2) than in real observational cases (2062 desert grids of 40 495 total grids for COMS validation, Table 4) may cause the coefficients in Eq. (2) to have a higher regression slope, which will finally lead to positive biases in OLR_12.0 retrieval. This kind of tendency also occurs in the other three algorithms, but to a lesser degree (Figs. 5b-d). On the other hand, the RMSE of OLR_10.8+12.0 is smaller than the RMSE of OLR_12.0. In addition, the bias of OLR_10.8+12.0 is the smallest among all OLRs. These improved statistics imply that the radiance difference between the split window channels (10.8 and 12.0 μm) may well correct the effects of atmospheric water vapor in clear-sky conditions where most of the high OLR values are present (Inoue and Ackerman, 2002). This improvement is important because, on average, clear-sky scenes constitute about 38% of the entire domain (Table 5).

Both OLR_6.7+10.8 and OLR _All are from the algorithm using a 6.7 μm water vapor channel. Interestingly, OLR_6.7+10.8 and OLR _All show smaller RMSEs than the other two OLRs. This smaller RMSE may be because F_6.7 provides adequate information for resolving the tropospheric water vapor absorption, which cannot be captured in window-channel algorithms. Despite employing all three irradiances, we note that OLR _All does not show a decisive superiority to OLR_6.7+10.8, which indicates that the additional inclusion of F_12.0 does not improve the accuracy of OLR_6.7+10.8. This may be because the combination of F_6.7 and F_10.8 has enough degrees of freedom to resolve variations in OLR. (Ellingson et al., 1989) showed that the channels for OLR retrieval from the High Resolution Infrared Radiation Sounder, centered at 13.3, 6.7, 8.2, and 14.4 μm, explained 96.4%, 1.8%, 1.1%, and 0.5% of the OLR variations, respectively. This confirms that in addition to surface or cloud top information from one window channel, the upper tropospheric humidity represented by the 6.7 μ m channel is crucial for increasing the accuracy of OLR retrievals. Specifically, a combination of one window channel and one water vapor channel can capture more than 98% of the OLR variability.

The RMSE of OLR_6.7+10.8 (6.08 W m^-2 in Fig. 5c) is larger than its algorithm intrinsic accuracy (2.22 W m^-2 in Table 3) estimated from RTM simulations. In contrast, the RMSEs of OLR_12.0 and OLR_10.8+12.0 are smaller than their algorithm intrinsic accuracies (Table 3). This may be due to differences in the relative contribution of environmental conditions between the COMS FOV (i.e., a larger contribution of the wide Pacific Ocean than the Australian desert) and the RTM environments (i.e., equal contributions of tropical ocean and desert simulations). Thus, the desert case would lower the accuracy in the 12 μm channel more for the RTM simulation than for COMS observations.

Figure 6.  The number of each scene type for the 40 494 homogeneous grids (bars), the averages of CERES OLR (symbols), and the standard deviations of CERES OLR (vertical lines).

To better understand the major causes of the decreasing accuracy in the OLR retrievals, dependence on various conditions is tested. Following the original classification of the ERBE scene (Diekmann and Smith, 1989) available in the CERES data, a cloudiness condition of a 1^°× 1^° grid is classified in terms of cloud fraction: clear (less than 5% cloudy); partly cloudy (5%-50% cloudy); mostly cloudy (50%-95% cloudy); and overcast (more than 95% cloudy). Figure 6 shows overall statistics of the CERES observations for the 40 494 homogeneous 1^°× 1^° grids (the sample numbers are indicated by the bar graphs, and the OLR values are indicated by the plot graphs). Categories with less than 30 grids are not shown. As the COMS FOV covers a large proportion of the western Pacific, most of the grids are classified as ocean scenes. The number of snow scenes is small because the high-latitude regions are barely observed from geostationary orbit. Based on the analysis of CERES OLR distributions in Fig. 6, the OLR decreases with increasing cloudiness due to the colder temperatures of cloud tops compared to the Earth's surface.

On the whole, OLR₁₂ has a positive bias by day, but no significant bias by night (Table 4). The positive daytime biases of OLR_12.0 over most surface types except snow tend to be larger over clear regions than cloudy regions. For instance, in the case of ocean surfaces in the daytime, 4.76 W m-2 for the clear region is larger than 3.84 W m^-2 for the cloudy region (Table 4). The larger positive bias over the clear region than the cloudy region is associated with overestimation of high OLR and underestimation of low OLR, as explained in the previous subsection. Such large positive biases in the single-channel algorithm can be corrected by the additional use of the 10.8 μm channel; OLR_10.8+12.0 shows reduced positive biases for clear conditions, especially over ocean surfaces, although the overall RMSEs are not greatly improved.

Figure 7.  The same as Fig. 3, but for 1920 desert grids. Red indicates daytime, blue indicates nighttime.

As expected from Figs. 5c and 5d (the accuracy for all cases), OLR_6.7+10.8 and OLR _All have smaller RMSEs than OLR_10.2 and OLR_10.8+12.0, regardless of surface type (Table 4). For the ocean surface type, the RMSEs of OLR_6.7+10.8 are 4.27 to 5.69 W m^-2, and the RMSEs of OLR _ALL are 4.07-4.75 W m^-2. These values are about 1-2 W m^-2 smaller than the RMSEs of the other two methods, OLR_10.2 and OLR_10.8+12.0. Also, the biases of OLR_6.7+10.8 and OLR _All are negative for all surface types in both clear and cloudy conditions (Table 4). However, these biases are systematic and can simply be reduced by an adjustment of coefficients.

The case of desert conditions in the daytime requires more discussion because the differences in OLR between COMS and CERES for that case are exceptionally large (Table 4). As seen in Fig. 4 (the RTM simulations), OLR is quite sensitive to spectral Fs. A small bias in F would lead to a large bias in OLR in the present simplified regression approach for extremely dry atmosphere and hot surface (during the daytime) conditions. Moreover, the biases of OLR in the daytime are much larger than at nighttime over the desert. This is also clearly shown in Fig. 7, which presents the scatter plots of COMS versus CERES OLRs for the desert in daytime (red) and at nighttime (blue). Although daytime positive biases exist for all four OLRs, the biases of OLR₁₂ and OLR_10.8+12.0 (26.94 and 16.74 for desert clear-sky scenes in Table 4) are much larger than those of the other two retrievals. In contrast, the nighttime negative bias is noticeable only for OLR_6.7+10.8.

Table 5 summarizes the dependence of OLR accuracies on various environmental conditions. In all four algorithms, the RMSE is significantly smaller over the ocean than over land. The ocean-land contrast is largest for OLR_12.0; the RMSEs are 7.20 W m^-2 for the ocean and 11.99 W m^-2 for the land. The biases for OLR_12.0 are also smaller over the ocean than over land (2.28 W m^-2 for the ocean, 5.68 W m^-2 for land), as are the biases for OLR_10.8+12.0 (-0.54 W m^-2 for the ocean, 6.05 W m^-2 for land). The large difference between over the land and ocean may be attributed to improper specifications of surface emissivity. It is well known that the complicated structure of land surface emissivity is difficult to reflect in several narrowband radiances and could lead to inherent OLR errors (Schmetz and Liu, 1988; Inoue and Ackerman, 2002; Ba et al., 2003).

For clear-sky conditions, the RMSEs are slightly larger than those for cloudy sky conditions; for example, the RMSE of OLR_6.7+10.8 is 6.87 W m^-2 for clear skies and 5.52 W m-2 for cloudy skies (Table 5). Specifically, the accuracy of OLR for cloudy skies is quite high over the full COMS FOV, despite the highly variable properties of clouds. This supports the feasibility of the current OLR algorithm, which does not incorporate complex cloud properties as input variables. Table 5 also shows that the accuracy of OLR is better at nighttime than in the daytime. Compared to the CERES OLR, the nighttime OLR errors are smaller by 2.51 W m^-2 for OLR_12.0, 2.44 W m^-2 for OLR_10.8+12.0, and 1.51 W m^-2 for OLR_6.7+10.8. The day/night dependence may be associated with the warm/cold temperatures; that is, the errors tend to be larger for a high OLR range (see Fig. 5), which is observed more frequently in the daytime. Note that the same explanation can be applied to the larger errors over the land surface than the ocean surface, or the larger errors with clear-sky scenes compared to cloudy skies.

Several previous studies (e.g., Otterman et al., 1997) have pointed out the strong link between L-to-F conversion error and satellite zenith angle. Therefore, we examine the OLR errors and biases depending on satellite zenith angle in Table 6. Specifically, with an increase in satellite zenith angle from 0^°-30^° to 60^°-90^°, the negative biases of the two OLRs using the 6.7 μm water vapor channel increase from 0.00 to -8.56 W m^-2 for OLR_6.7+10.8 and from -1.39 to 8.68 to W m^-2 for OLR _All. Therefore, the negative biases of OLR_6.7+10.8 and OLR _All (shown in Table 4) are attributed to the water vapor channel for higher satellite zenith angles (>60^°). In contrast, for OLR_10.2 and OLR_10.8+12.0, the positive bias increases when approaching the nadir point (from -0.12 to 9.30 W m^-2 for OLR_12.0 and from -0.65 to 3.41 W m^-2 for OLR_10.8+12.0). The bias characteristics of OLRs depending on the satellite zenith angles are related to the regional distributions of the bias of OLR (Fig. 8). A larger bias at the nadir point implies that the limb darkening effect of the atmosphere still remains significant in the irradiance calculation and is not eliminated through the estimated relation between L(θ) and F [Eq. (1)].

This study has attempted to evaluate four OLR retrieval algorithms (OLR_12.0, OLR_10.8+12.0, OLR_6.7+10.8, and OLR _All) with different combinations of the three best channels (10.8, 12.0, and 6.7 μm) on a geostationary satellite. The OLR retrieval algorithms were constructed on the basis of the coefficients (Tables 1 and 2) of calculated regression analyses of RTM simulations of L, F, and OLR (Section 3). The RTM simulations were performed for various surface, atmospheric, and cloud conditions, given the spectral characteristics of COMS. Using COMS radiance observations, the OLRs from the four different methods were retrieved for the period of 1-30 April 2011 (Fig. 2b). The retrieved OLRs were collocated with the CERES OLR on 1^°× 1^° grids, and validation was performed only for "homogeneous grids" as in (Ba et al., 2003).

It was hypothesized to be a challenging task to have consistent and appropriate accuracy of COMS OLR over the full FOV [roughly (50^°S-50^°N, 70^°-170^°E), as shown in Fig. 1] due to diversities in surface conditions, atmospheric profiles, seasons, etc. The comparison between COMS OLR and the collocated CERES OLR over a broad domain shows good agreement within RMSEs of 5.5 to 8.7 W m^-2 (Fig. 3). The errors in this study are smaller than the 11 W m^-2 RMSE reported by (Ohring et al., 1984), but comparable to the 7 W m-2 RMSE reported by (Ba et al., 2003). The error between GOES OLR and CERES OLR in (Ba et al., 2003) is from validation over the American continent and surrounding oceans. Compared to this previous study, the current validation domain includes a much larger ocean area, which could contribute to higher OLR accuracy. However, we also have a larger satellite zenith area, which would lower the OLR accuracy. As a result, our validation results show smaller or similar accuracies compared to (Ba et al., 2003). Therefore, it is believed that COMS OLR retrievals are in reasonable agreement with CERES OLR.

Figure 8.  Bias maps for (a) OLR_12.0, (b) OLR_10.8+12.0, (c) OLR_6.7+10.8, and (d) OLR _All against CERES OLR. The values are averaged over a 5^°× 5^° grid, and the gray color indicates missing values due to a lack of reference data.

The important conclusion of this study is the determination of differences in the four OLR methods using the same geostationary satellite (Fig. 5). Overall, the errors of OLR _All and OLR_10.8+6.7 (5.58 to 6.08 W m^-2 RMSE) are smaller than those of OLR_10.8+12.0 and OLR_12.0 (7.97 to 8.64 W m^-2 RMSE), implying that the water vapor channel has a significant role in OLR retrievals. We also assume that the use of the third-order polynomial function of these two methods allow the algorithms to have more flexible IR-OLR relationships that are suitable both over extremely low and high OLR ranges. Although some negative biases of about 5 W m^-2 exist for OLR_10.8+12.0 and for OLR _All, they can be removed by proper adjustment of the offset coefficients in the algorithms.

In all four methods, the errors are about 2-3 W m^-2 smaller over the ocean than over land, and the biases are also smaller over the ocean (-0.54 W m^-2 over the ocean vs. 6.05 W m^-2 over land for OLR_10.8+12.0, Table 5). Because the current OLR algorithms do not consider the variation of surface emissivity over land, the accuracy of the OLR retrievals is better over the ocean than over land. In addition, larger errors for COMS OLR retrievals in the daytime compared to nighttime are evident (Table 5) because the errors tend to be larger for higher OLR values, as in Fig. 5. Our analyses have also revealed a slightly larger error for clear skies than for cloudy skies: e.g., the RMSE of OLR_6.7+10.8 is 6.87 W m^-2 for clear skies and 5.52 W m^-2 for cloudy skies (Table 5). These accuracies in OLR for cloudy skies seem to be good enough over the full COMS FOV despite the large variation in the cloud properties (e.g., cloud top height or thickness). Although cirrus cloud cases are known to be problematic for estimating OLR, the dependence of the accuracy on cloud type was not investigated in the current study.

COMS OLR is highly biased relative to CERES OLR for deserts in the daytime, which is one of the challenging cases for obtaining high OLR accuracy, as discussed in (Ellingson et al., 1994). It is believed that such biases may be later improved by more realistic desert atmospheric profiles in RTM simulations, and also by consideration of each pixel's surface types in OLR algorithms. Our analyses also revealed positive biases of COMS OLR over the area with high satellite zenith angles. The biases could be attributable to incomplete elimination of the limb darkening effect by L-F conversion in the RTM calculation. In addition, because the CERES has viewing angles ranging only from 0^° to 51^° for FM1, different satellite viewing angles of the COMS and CERES can be another factor causing positive biases of COMS from higher viewing angles. It is possible that the dependence of the accuracy on various environmental factors can be conditionally improved by constructing an algorithm that uses different coefficients [in Eqs. (1-5)] for each individual scene type. However, this attempt to improve the accuracy may cause problems, such as computing time delay or degradation of the consistency and spatial homogeneity of retrievals.

Three of the OLR algorithms, i.e., OLR₁₂, OLR_10.8+6.7, and OLR_10.8+12.0, have been implemented in the operational COMS Meteorological Data Processing System, along with other retrieval algorithms for meteorological variables (e.g., Oh et al., 2006; Choi et al., 2007; Park et al., 2007; Yoon et al., 2007; Kim et al., 2008; Lee et al., 2011; Park et al., 2012). This study shows that the accuracies of the three OLR algorithms appear to be higher than, or similar to, other OLR algorithms; further improvement of the OLR retrievals may be possible by adjusting coefficients in the OLR algorithms based on the validation results with the CERES OLR. The adjusted algorithms based on the validation results yield smaller RMSEs: 7.35, 7.91, 5.71 and 5.42 W m^-2 for OLR_12.0, OLR_10.8+12.0, OLR_6.7+10.8 and OLR _All, respectively.

The validation results of this study are particularly valuable for comparing the COMS OLR and CERES OLR over the full geostationary satellite FOV, whereas previous evaluation studies are limited to specific regions. Therefore, the following rules are suggested to determine the optimal algorithm(s) for obtaining reasonable accuracies over the full geostationary satellite FOV.

(1) If the systematic biases are corrected, the OLR_10.8+6.7 and OLR _ALL algorithm will be suggested over a vast domain including various surface and cloud conditions due to the consistently good accuracies. Considering computational efficiencies, the OLR_10.8+6.7 method using the two channels only may be optimal due to spending less computational resources than the OLR _ALL.

(2) The present results also imply that the OLR retrieval algorithms could be specifically selected for different domains. In the case of retrieving OLRs over ocean regions, as in (Inoue and Ackerman, 2002), the OLR_10.8+12.0 method is a reasonable choice since the RMSE in ORL_10.8+12.0 is not so different from those in OLR_6.7+10.8 and OLR _ALL for ocean conditions. For a desert surface, the application of OLR_12.0 or OLR_10.8+12.0 methods is not recommended due to more severe errors for clear and daytime scenes.

The current validation results confirm a successful COMS OLR retrieval and can also provide important guidelines for designing OLR algorithms for next-generation satellites. Furthermore, the present results could be informative for assimilation of satellite-observed OLR in numerical models, or for scientific applications focusing on atmospheric processes.