Short abstract Satellite retrievals of sea surface skin temperature (SSTskin) have become necessary for many near-real time applications. The missions of the two MODISs have provided continuous measurements for more than twenty years and have played a significant role in generating time series of quantitative estimates of SSTskin. This study used four machine learning approaches: eXtreme Gradient Boosting (XGBoost), support vector machines (SVM), random forests (RF), and artificial neural networks (ANN), to develop improved atmospheric correction algorithms for satellite-derived SSTskin in the Caribbean region. A set of satellite and in-situ measurements, including SST, the atmospheric state and surface radiation, taken on research cruises, from surface moorings and drifting buoys was used to train the machine learning models. Finally, the reliability and shortcomings of various machine learning methods were assessed through comparisons with SSTskin derived from shipboard and other in situ measurements. Overall comparisons show encouraging results: the biases of various machine learning approaches vary between -0.076 K to 0.013 K; with the XGBoost showing the best correlation in a statistical analysis of in situ SST measurements. This study contributes to improving our understanding of the key environmental properties and will reduce uncertainty in earth science and climate research.