An Improved Estimation of Outgoing Longwave Radiation Based on Geostationary Satellite

Abstract

The Outgoing Longwave Radiation (OLR) is an important satellite-driven variable for understanding the Earth's energy budget balance. The geostationary OLR retrievals require angular and spectral integration using an empirical equation for irradiance flux-to-OLR from a regression analysis, which determines the accuracy of the narrowband satellite-based OLR. We selected homogeneous pixels which is satisfied less temporal-spatial variability of cloud, on three infrared channels (6.7, 10.8, $12.0{\mu}m$) of the first multipurpose geostationary satellite in Korea, namely the Communication, Ocean and Meteorological Satellite/Meteorological Imager (COMS/MI). Multiple regression analysis was performed to retrieve OLR with improved accuracy using selected parameters based on theoretical and physical significance. This algorithm yielded retrieval with higher accuracy than broadband-based OLR retrieval: RMSE of 10.54 to $3.81W\;m^{-2}$, and bias of -8.49 to $-0.07W\;m^{-2}$.

1. Introduction

Quantitative observation of Outgoing Longwave Radiation (OLR) is essential in understanding the energy budget of the earth. Satellite-based OLR data are being used forthis purpose (Tsushima and Manabe, 2013). It is calculated using broadband or narrowband satellite dat. The OLR from broadband satellites produces high quality data by directly observing a broadband range of 3 to 100 μm at the top of the atmosphere (Singh et al., 2007).It is used asreferences data due to its high accuracy. But they do not provide instantaneous TOA OLR radiative fluxes at fixed angles(Loeb et al., 2007). Geostationary narrowband based OLRobservations offerthe significant advantage of an instantaneous response to surface temperature changes(Ba et al., 2003). Geostationary OLRretrieval requires angular and spectral integration using an empirical equation for irradiance flux-to-OLR from a regression analysis, which determines the accuracy of the narrowband satellite-based OLR.

The aim of this study is to develop a new algorithm for OLR using three infrared channels (6.7, 10.8, 12.0 μm) of the first multipurpose geostationary satellite in Korea, namely the Communication, Ocean and Meteorological Satellite/Meteorological Imager (COMS/MI). The COMS/MI obtains OLR products using the Algorithm Theoretical Basis Documents (ATBD) developed for COMS/MI, known as the COMS/MI Meteorological Data Processing System (CMDPS) (NIMR, 2009). The CMDPS-based OLR (hereinafter referred as OLRCMDPS) showed a root mean square error (RMSE) of 17.35 W m^-2 (NMSC, 2012). These results do not satisfy the initial targeted accuracy of OLR by the Korea Meteorological Administration (KMA), which is an RMSE of 10.0 W m^-2 when validated against CERES OLR data (NIMR, 2009).In this paper, we use homogeneous pixels which is satisfied less temporal-spatial variability to improve an empirical equation that is applied to convert irradiance flux to OLRestimations. Once this equation is established, new OLR products can be produced using the COMS/MI infrared channels. The new OLR data are evaluated with the same homogeneous grids selected by the STEM. To evaluate the new OLR products, developed using the STEM method, they are validated with daily-averaged, but coarse resolution, CERES-ES4 OLR data.

2. Data

This study was conducted using data for three yearsfromApril 1, 2011 to March 31, 2014.The spatial range is 10.75°S-67.25°N, 50.75°-205.65°E.The OLR products were retrieved using COMS/MI.We used the WaterVapor(WV, 6.7 μm) and infrared channels(IR1, 10.8 μm;IR2, 12.0 μm)in digital number(DN)format, with 3-hourintervalsto produce OLRCMDPS products by following the procedure provided by the CMDPS algorithm.It collects observations at 15 minute and the spatial resolution 4 km. The validation data is CERES data fromthe paired satellites,Terra andAqua.The data is produced daily and has a spatial resolution of 2.5 degrees. The CERES ORL data used a flux-to-OLR method and is produced based on Earth Radiation Budget Experiment (ERBE) data. This data has been confirmed to have high accuracy (Loeb et al., 2007).

3. Method

1) Data pre-processing

In this study, the COMS/MI DNs were converted to brightness temperature (Tb) by application of the Global Space-based Inter-Calibration System (GSICS) annual coefficients, which are provided by the COMS/MI data provider, namely the NMSCof Korea. The corrected Tb was then converted to radiance (L) using Planck’s law. The next step is the conversion of radiance (W m^-2 μm^-1 sr^-1) to irradiance flux (F) (W m^-2 μm^-1) through an angular integration process. The coefficientsforradiance-irradiance flux conversion were provided by the NMSC.

We used the Spatio-Temporally Equalized Matchup method (STEM) to select homogeneous pixels in two data with differentspatial and temporalresolution. In thisstudy, the criteria related to STEM were the same as that proposed by Kim et al. (2015). These selected pixels were used as validation and training data.

2) Developing an algorithm (F-OLR conversion)

We set the least atmospheric interference (F10.8), differences in infrared and WV channels (F10.8-F6.7 and F12.0-F6.7), and the difference in infrared channels (F12.0-F10.8) asthe parameters ofthisstudy, which are mainly used to calculate OLR using narrowband band data (Costa et al., 2012; Schmetz et al., 1988, Inoue et al., 2002).Astepwise regression analysiswas performed using these parameters to determine constants and coefficients for the irradiance flux-OLR conversion equation, where the results are shown in Table 1. The most marked improvement in the regression analysis results occurred when all three parameters were accounted for, which resulted in the lowest standard error, of approximately 3.86. Note that variations in coefficients had a maximum value of 0.187 for model 3, but even this largest variation was insignificant.

Table 1. Statistics from the stepwise multiple regression analysis to convert irradiance flux to OLR using selected parameters

The reserved training dataset of even number of days were selected within its homogeneity. We performed a stepwise regression analysis of the three channels to generate equation (1). The OLR equation includes F10.8, F10.8-F12.0, and F12.0-F6.7 as variables, and results in an R² of 0.99:

\(\begin{aligned} O L R=& \alpha_{0}+\alpha_{1} \times F_{10.8}+\alpha_{2} \times\left(F_{10.8}-F_{12.0}\right)+\\ & \alpha_{3} \times\left(F_{12.0}-F_{6.7}\right) \end{aligned}\)       (1)

where OLR denotes the retrieved OLR, and the coefficients determined fromthe regression analysis for irradiance flux-OLRconversion, α₀, α₁, α₂, and α₃ were 73.68, 15.40, -16.58, and -7.76, respectively.

4. Result

1) Quantitative analysis

We produced scatter plots of conventional OLRCMDPS data and the new OLR retrieval results compared to CERES OLR for all 3 years for which data were retrieved. For data that satisfied the F10.8 standard deviation (σ) ≤ 1.0Wm^-2 μm^-1 condition (Kim et al., 2015), existing OLRCMDPS and new OLR retrieval results were compared to CERES OLR for all 3 years. The homogeneous data distributions in Fig. 3(a) and (b) show distinguishable distributions The RMSE for the retrieved data relative to the reference was improved from 10.5432 to 3.81 W m^-2, and the bias improved from -8.49 to -0.07 W m^-2. The new OLR was clear improvement in fitting with the new algorithm at high OLR values of 250 to 300 W m^-2 betterthan OLRCMDPS (Kim et al., 2015).In addition to the quantitative improvements in RMSE and bias, the problem ofthe shifted distribution at high OLR was avoided and the data exhibit a better fit to the 1:1 line. The well-adjusted high valuesfor OLR are significant, as the OLR data are densely populated in the range 250-300 W m^-2 , as shown in Fig. 1.

Fig. 1. Scatter plots of (a) CERES outgoing longwave radiation (OLR) and OLRCMDPS retrievals (Kim et al., 2015) and (b) CERES OLR and new OLR retrievals for all 3 years for which data were retrieved.

2) Time series analysis

We selected points by surface types at 165°E, 12.5°N (Fig. 2(a-b), ocean), 104.15°E, 41.4°N (Fig. 2(c-d), desert and barren sample data for time series analysis and further evaluation of the new OLR algorithm. A full year from December 2012 to November 2013 was investigated on a seasonal basis by dividing the data into DJF&MAM (December to May), and JJA&SON (June to November). We compared OLRCMDPS and the new OLR retrievals with CERES-ES4 OLR reference data from selected points for 1 year; the higher the homogeneity, the better the performance of the newly developed OLR algorithm. In both figures, the straight black line represents the reference OLR from CERES, the dotted gray line corresponds to OLRCMDPS, and the blue dash line denotes the new OLRretrievals.The red crossesrepresent homogenous conditions that satisfy the criterion F10.8 σ ≤ 1.0 W m^-2 μm^-1. The time series analysis shows that the new algorithm improved the accuracy under both homogeneous and inhomogeneous conditions.

Fig. 2. Time series analysis at two points from December 2012 to November 2013. (a)–(b) the point represents an ocean, and (c)-(d) the point represents a desert and barren.

3) Scene analysis (Qualitative analysis)

Scenes were also analyzed for comparison of OLRCMDPSwiththenewOLRretrievals.Toinvestigate the performance of OLRCMDPS and the new OLR in comparison to the CERES data, subtraction of the reference OLR from the retrievals highlights the differences in mid-latitude regions, as shown in Fig. 3(a) and (b). Overall, the new OLR data exhibit fewer differencesthan the OLRCMDPS retrievals; however, differences of over 50Wm^-2 can be seen in some pixels at higher latitudes, which remain a challenge for the new COMS/MI OLR algorithm.

 

Fig. 3. The difference in COMS/MI retrievals by (a) CMDPS and (b) the new algorithm, relative to reference OLR data on Jan. 8, 2013 in the extended Northern Hemisphere. The projection is adjusted from GOES to WGS84.

5. Conclusions

We have developed an algorithm that agrees better withCERES OLRthan the conventional OLRCMDPS product in terms of scene analysis and time series analysis; it also exhibits improved accuracy, as demonstrated by reductions in RMSE and bias when comparing the full 3 years’ worth of data. The homogeneity method was used forretrieving OLRand assessing the performance of the algorithm. However, the thresholds used for the standard deviation were determined froman empirical analysis using the current data, which may need to be more flexible when dealing with other data. This indicates that the radiative flux anomaly can result in deterioration of the performance of the current OLR algorithm. Therefore, updates are necessary in the case of meteorological eventsthat may influence OLR forcing, such as volcanic eruptions (Minnis et al., 1993), which will be the subject of a future publication. The Geo-Kompsat 2A, the COMS/ MI’s follow-up satellite, will be operational. This satellite observes the same area and contains more.