Rethinking Prior Approaches – Bayesian Neural Networks for Information Retrieval from Ocean Color

The first ocean color images were captured by CZCS in 1978, and since then, proxies of ocean biogeochemical process have been derived from them using, primarily, empirical or semi-analytical methods. These approaches are, in simplistic terms, a two-part process: atmospheric correction, in which the contribution of the atmosphere to the top of atmosphere radiance signal is ‘removed’ using models and observation-based assumptions, followed by application of the ocean color algorithm to the remaining water-leaving signal. These methods provide reasonably accurate estimates of biogeochemical proxies such as chlorophyll-a concentration and inherent optical properties (IOPs) across a wide variety of water types. However, accurate estimates over waters where atmospheric correction is challenging, e.g., in the coastal ocean or inland water bodies, remains elusive. We will present a successful machine learning approach that bypasses conventional atmospheric correction and uses Bayesian neural networks (NNs) to estimate chlorophyll-a or the phytoplankton absorption coefficient (an IOP) directly from top of atmosphere reflectance. In an extension of this approach, we also investigate the utility of adding polarimetric information to these Bayesian NNs using a variety of ocean polarimeter datasets collected in situ and from aircraft campaigns. These investigations are of particular relevance to the upcoming NASA PACE (Plankton, Aerosol, Cloud, ocean Ecosystem) mission whose payload includes polarimeters and an ocean color instrument, and whose central scientific goals will require powerful state-of-the-art data analysis approaches.