An Operational Global Near-Real-Time High-Resolution Seamless Sea Surface Temperature Products From Satellite-Based Thermal Infrared Measurements

A machine learning-driven algorithm was developed to generate a near-real-time (NRT), high-resolution (1-km), all-sky global SST product using the thermal infrared measurements of MODIS Terra. The algorithm can generate the NRT SST under all-sky conditions, while most current operational algorithms only derive NRT SST for pixels without clouds. Meanwhile, the previous algorithms typically constructed seamless SST products by integrating data from multiple satellites and in situ SST sources with a high computational cost, which are not suitable NRT applications. The current algorithm uses a single sensor dataset to deliver NRT seamless SST. The model has outstanding accuracy in retrieving SST with a high sample-based (leave one year out) cross-validation (CV) (LOYOCV)
$R^{2}$
of 1.0 (1.0) and a small RMSE of 0.39 °C (0.43 °C), respectively. Over the different regions, the model performs very well without showing any spatial dependence. The developed algorithm retrieves the SST with higher accuracy than the nonlinear SST algorithm, the most used infrared-based SST operational algorithm, and provides SST covering the entire satellite image despite clouds. It is noted that the SST, satellite zenith angle (SZA), and water vapor do not have a discernible impact on the errors of the SST retrievals. The proposed model demonstrates exceptional promise in generating operational high-resolution, all-sky NRT SST retrievals using satellite-based thermal infrared sensors and is not limited to MODIS.