A Machine Learning-Based SST Retrieval from Thermal Infrared Observations of INSAT-3D Imager: Improvement Over Regression-Based NLSST Algorithm

Abstract

Sea surface temperature (SST) is one of the key Essential Climate Variables for studying and monitoring Earth’s climate, besides playing an important role in physical oceanographic processes and as a boundary condition in the numerical prediction models. Understanding these processes requires the availability of accurate and consistent SST products over the global ocean, which can be fulfilled by retrieving them from satellite-based observations. Therefore, the present study exploits a supervised machine learning technique, Deep Neural Network (DNN), for the retrieval of SST using thermal infrared (TIR) split-window observations from Imager onboard India’s geostationary satellite, INSAT-3D, which was launched in 2013. A matchup dataset is prepared to train and test the DNN, comprising the collocated brightness temperatures of TIR channels of INSAT-3D Imager with the in-situ SST measurements for 2017–2020. The DNN-based algorithm exhibits a similar statistics with reference to the in-situ SST for both training and testing datasets. It is further assessed on independent INSAT-3D observations for May 2021- February 2022 to demonstrate its robustness. Moreover, the performance of the DNN is also compared to the widely used regression-based non-linear SST (NLSST) algorithm, which is presently operational for INSAT-3D. Validation against the in-situ SST shows an improvement of about 37.5% in the accuracy of SST retrieved using DNN (RMSE = 0.5 K) over the NLSST (RMSE = 0.8 K) algorithms for INSAT-3D Imager.