Atmospheric Correction of Hyperspectral Data Over Coastal Waters Based on Machine Learning Models

Standard Atmospheric Correction algorithms that predict water-leaving radiance, while working well for the open-ocean using multispectral data, can be inaccurate or computationally demanding for coastal and optically-complex waters, where the phytoplankton signal might be masked or modified by the presence of other substances. Here, different Machine Learning models are presented, trained, and evaluated using simulated hyperspectral ocean color data of top-of-the-atmosphere radiance from coastal waters to predict water-leaving radiance and other ocean color variables directly, such as chlorophyll concentration. High accuracy of up to 99% for some of the variables is achieved when trained and evaluated on simulated data.