Evaluation of Modeled Diurnal Warming Estimates for Use in Producing SST Analyses

The production of daily, multi-sensor sea surface temperature (SST) analyses involves the blending of observations from sensors with different sampling times and different effective measurement depths. Creating a single SST value representative of a specific depth, like the foundation temperature, requires compensation for variability associated with processes like diurnal warming. A physical model simulating diurnal warming at arbitrary times and depths has been integrated into the NOAA NESDIS GOES-POES Blended SST analysis to evaluate the potential impact of explicit compensation on the product accuracy. The model, based on the Kantha-Clayson one-dimensional turbulence closure model with have effects has been modified for forcing with data obtained from global numerical weather prediction and wave models. Thorough evaluation is necessary, however, particularly in instances of large diurnal warming which are of the greatest importance for application to SST analysis generation.

Here, we evaluate the model performance over four seasons using observations of diurnal warming derived from operational geosynchronous satellite SST retrievals. The model is forced with 6-hourly data from the NOAA Global Forecast System (GFS) and Wave Watch III models and the simulated diurnal warming amplitudes are compared with SST estimates derived from the Meteosat-11, GOES-16, and Himawari-8 satellites. Multiple model configurations and "tunable" parameters are evaluated to identify the best achievable performance. Results from direct point-to-point comparisons and derived distributions of diurnal warming amplitudes provide a recommended model configuration and demonstrate that the model can yield realistic predictions with uncertainty levels sufficient for application to SST analyses. The identified model configuration is also shown to produce accurate estimates of diurnal warming observed from multiple research cruises, lending additional confidence in the model performance.