The Improvement of FY-3B/VIRR SST Algorithm and Its Accuracy

The evolution of sea surface temperature (SST) algorithms is introduced and a set of SST regression formalisms are given. Some improvements are made based on operational SST algorithm from FY-3B meteorological satellite visible and infrared radiometer (VIRR) data. On matching algorithm, quality controlled in situ data from the in situ quality monitor (iQUAM) is used to improve the input data precision of regression. Fields of matchup database (MDB) are enlarged to provide the convenience for error analysis. Pixels with "confident clear" flag in FY-3B/VIRR cloud mask (CLM) products are matched up to form gross matchups, and then tightly filtered by some tests to form tight matchups, which make the sample selection more reasonable. On regression algorithm, based on least-square regression used for the early operational SST product, the robust regression is developed, and its performance is tested by NOAA-19/AVHRR MDB of 2010. It shows that the precision of SST is increased by 21% in daytime with split-window non-linear SST (NL) algorithm and 30% in nighttime with triple-window MC (TC) algorithm. On retrieval algorithm, the spatial uniformity test and climate reference test are introduced, the unidentified cloud (especially at night) is excluded and the SST retrieval precision is improved.A set of SST regression formalisms are tested based on NOAA-19/AVHRR 2010 MDB. It shows NL is the best algorithm for daytime while TC is the best algorithm for nighttime, which is accordant with NESDIS/STAR. The monthly MDB is created from FY-3B/VIRR measurements paired with coincident SST measurements from buoys data.The same regression analysis method is also used on FY-3B/VIRR MDB. Comparing three daytime SST algorithms and five nighttime SST algorithms, the best algorithm to retrieve FY-3B/VIRR SST is NL both in daytime and nighttime. It shows for FY-3B/VIRR nighttime TC, the contribution of 3.7 μm band is smaller than split-window bands, and the calibration of 3.7 μm band has stripe phenomenon. A three-month MDB from October to December in 2012 is used to derive coefficients. An independent MDB from January to March in 2013 is used to access the accuracy of the best NL algorithm for FY-3B/VIRR. Based on matchup analyses, the root mean square error (RMSE) between FY-3B/VIRR SST and in situ SST is 0.41℃ (NL_D) and 0.43℃ (NL_N). Compare with Daily Optimum Interpolation SST (OISST), the RMSE of FY-3B/VIRR SST is 1.45℃ (NL_D) and 1.5℃ (NL_N). When the absolute difference between FY-3B/VIRR SST and OISST is within 2℃, the RMSE is 0.82℃ (NL_D) and 0.84℃ (NL_N).