National Oceanic and Atmospheric Administration (NOAA) satellites provide repetitive daily global coverage of the Earth. Since the early 1980s, the National Environmental Satellite, Data, and Information Service (NESDIS) has been operationally generating Sea Surface Temperature (SST) products from the Advanced Very High Resolution Radiometers (AVHRR), which are onboard several NOAA platforms. Globally, AVHRR data are merged with in situ SSTs in a space and time window of 4 hr and 25 km to create monthly match-up files. Early in each satellite's mission, these match-up files are used to calibrate the SST algorithms (i.e., calculate coefficients in the SST regression equations) and then to routinely validate the SST products throughout the operational lifetime of the platform.

The primary objectives of this paper are to describe the heritage NESDIS AVHRR SST Calibration/Validation (Cal/Val) system and to document the latest validation results from NOAA -16, -17, and -18 AVHRRs from 2003 until the present. In situ SSTs are strongly contaminated due to instrument malfunction or erroneous data acquisition and relay. Quality control of the match-up dataset is critically important for ensuring proper calibration and validation of the satellite SST products. Robust identification of outliers is particularly crucial. A method for deriving robust location (mean) and scale parameters (RMSD) for outlier removal is adapted, which makes use of the entire distribution with a data-specific weighted function. The proposed scale parameter is based on the L moments, rather than conventionally used central moments, for characterizing the distribution of the match-up data about the average value. The L moment deviation (L2) is subsequently employed to exclude extreme values. The L2 approach proved to perform significantly better than the conventional approach without much loss of data.

Validation results show a typical global monthly bias within ±0.2°C and RMSD of about 0.4°C - 0.6°C. We also check the robustness and seasonal stability of the derived MC/NLSST coefficients and RMSD, and estimate their sensitivity to the quality control of the dataset. We conclude the presentation by identifying potential improvements to the heritage AVHRR SST Cal/Val system. The lessons learned from the heritage SST Cal/Val are directly applicable to the future SST products to be derived from GOES-R and NPOESS.