The impact of measurement uncertainty and spatial variability on the accuracy of skin and subsurface regression-based sea surface temperature algorithms

An ongoing limitation of common regression-based infrared (IR) satellite sea surface temperature (SST) algorithms has been the lack of sufficient in situ skin temperature measurements for derivation of the algorithm coefficients. Since IR brightness temperatures respond to the skin temperature, use of the more numerous subsurface observations to tune the algorithms introduces uncertainty into the resulting SST products. Coincident in situ skin and subsurface SST measurements from three years of cruises are used to derive parallel skin and subsurface multichannel SST (MCSST)-type regression algorithms to determine the extent to which improved accuracy can be obtained using the skin measurements. Through use of only coincident measurements, the advantage offered by the greater volume of available subsurface observations is eliminated. Surprisingly, we find no accuracy improvement using skin SST algorithms relative to algorithms derived from the research-grade ship-borne subsurface temperature measurements used in our analysis. However, better accuracy was found relative to algorithms derived from subsurface observations whose accuracy was degraded to that of buoys. The results are robust with regard to satellite resolution, collocation criteria, geographical regions, and time of day.The accuracy differences are found to be generally consistent with the effects of: (1) increased measurement uncertainty of radiometric measurements relative to research-grade subsurface observations, and (2) differences in spatial variability between the skin SST and temperature-at-depth. The subsurface algorithms are regenerated after degrading the subsurface measurements by adding increasing levels of Gaussian white noise to determine the amplitude of the additional variability required to ensure equal accuracy between the skin and subsurface products. The required supplemental noise ranges between 0.10 and 0.17. K for all data combined and generally decreases with tighter collocation windows and higher-resolution satellite observations. Variogram analysis and filtering of the in situ measurements suggest that differences in measurement uncertainty between the infrared radiometers and the subsurface sensors can explain 0.07-0.10. K of the required noise, while differences in spatial variability with depth can account for up to 0.07-0.10. K of the residual noise. A key consequence is that spatial averages of the skin temperature over satellite footprints of 2. km or more, while potentially biased in the mean, may exhibit less variance relative to point samples of the subsurface temperature than to the actual radiometric skin temperature.