Prototyping GOES-R ABI Hybrid SST Retrieval Algorithm with MSG/SEVIRI: Combining Regression and Radiative Transfer Model Approaches

Retrieval of sea surface temperature (SST) from measurements of top-of-the-atmosphere radiances in IR atmospheric transparency windows is a well established remote sensing technique. SST products are (or will be) routinely generated from polar (NOAA and MetOp AVHRR, EOS MODIS, NPOESS VIIRS) and geostationary (GOES, MSG/SEVIRI, GOES-R ABI) sensors. SST retrievals will be an important part of the future geostationary mission GOES-R. The Advanced Clear-Sky Processor for Oceans (ACSPO) was developed at NESDIS initially for processing AVHRR measurements and has been adopted for prototyping SST from GOES-R ABI using MSG/SEVIRI data as a proxy. The GOES-R SST approach is to implement two SST algorithms, regression and physical (or radiative-transfer-model-based), within one processing system and combine them together into a superior hybrid SST algorithm. This presentation describes the current status of the SST retrieval algorithm for GOES-R ABI.

Regression algorithms have been employed for more than 20 years, due to their simplicity and computational efficiency. In the regression technique, variations in the atmospheric transmission are corrected by establishing an average relationship between the SST and observed radiances on a global scale. Presently, two regression algorithms are used in AVHRR heritage systems. The Nonlinear SST (NLSST) algorithm employs only two longwave split-window bands centered at 11 and 12 μm. This algorithm can be used during both day and night. The Multi-Channel SST (MCSST) algorithm additionally employs a mid-IR band at 3.7 μm. This band can only be used for SST at night, due to solar contamination during daytime. In this study, the mid-IR band is not employed in SEVIRI analyses, for day-night continuity and pending resolution of the radiative transfer model (RTM) anomaly currently present in this band.

Recent developments of the fast and sufficiently accurate RTMs, and the availability of global analysis SSTs (such as Reynolds, OSTIA, and RTG) and Numerical Weather Prediction (NWP) information on upper-air fields, have given rise to attempts to more comprehensively account for local variations in atmospheric transmission in SST retrievals. In particular, the Optimal Estimation (OE) technique was recently applied to SST retrieval from 11 and 12 μm channels of MetOp AVHRR and MSG SEVIRI (Merchant et al, 2008, 2009). It has been found, however, that the OE technique underestimates the spatial and temporal SST variations, due to forcing the solution to the low-space and time-resolution first guess SST fields (which currently, for instance, do not resolve the diurnal cycle).

First-guess top-of-atmosphere BTs in ACSPO are simulated online using the Community Radiative Transfer Model (CRTM) with first-guess SST (daily Reynolds) and upper-air fields (from the National Centers for Environmental Prediction Global Forecast System, NCEP/GFS). Both regression (split-window NLSST) and physical (OE) SST algorithms have been implemented with MSG/SEVIRI data within one data stream. Additionally, a “hybrid” algorithm was also tested, which also takes a form of the split-window NLSST regression equation but applies it to SST and BT increments (i.e., deviations of in-situ SSTs and SEVIRI BTs from their first-guess estimates). According to our sensitivity studies, the “hybrid” algorithm accounts for local variations in the atmospheric transmission more accurately than the conventional regression. On the other hand, it does not require forcing the SST estimate to the first guess and, hence, better reproduces spatial and temporal SST contrasts than does the OE technique. The calculation of regression coefficients in the case of incremental regression was performed with an additional constraint, which requires that variability in the retrieved hybrid SST preserves the contrasts in the observed in situ SST. This additional constraint is required because errors in BT increments are comparable with the signal (variations in BT increments).

The results of sensitivity studies of the Regression, OE and Hybrid algorithms will be presented, including comparisons of SST retrievals from SEVIRI data with reference SST fields and with in situ measurements. Sensitivity of the “Physical” and “Hybrid” SST estimates to the selection of the first guess SST reference field is also discussed.