Improving Atmospheric Correction Algorithms for Sea Surface Skin Temperature Retrievals from Moderate-Resolution Imaging Spectroradiometer Using Machine Learning Methods

Satellite-retrieved sea-surface skin temperature (SSTskin) is essential for many Near-Real-Time studies. This study aimed to assess the potential to improve the accuracy of satellite-based SSTskin retrieval in the Caribbean region by using atmospheric correction algorithms based on four readily available machine learning (ML) approaches: eXtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Random Forest (RF), and the Artificial Neural Network (ANN). The ML models were trained on an extensive dataset comprising in situ SST measurements and atmospheric state parameters obtained from satellite products, reanalyzed datasets, research cruises, surface moorings, and drifting buoys. The benefits and shortcomings of various ML methods were assessed through comparisons with withheld in situ measurements. The results demonstrate that the ML-based algorithms achieve promising accuracy, with mean biases within 0.07 K when compared with the buoy data and ranging from -0.107 K to 0.179 K relative to the ship-derived SSTskin data. Notably, both XGBoost and RF stand out for their superior correlation and efficacy in the statistical results of validation. The improved SSTskin derived using the ML-based algorithms could enhance our understanding of vital oceanic and atmospheric characteristics and have the potential to reduce uncertainty in oceanographic, meteorological, and climate research.