Retrieval of oceanic chlorophyll concentration from GOES-R Advanced Baseline Imager using deep learning

Abstract

Data of chlorophyll (Chl-a) concentration obtained from polar-orbiting ocean color satellite sensors are subject to major spatial and temporal gaps in data coverage owing to the intrinsically low sampling frequency that coincides with adverse conditions such as cloud obstruction. Such gaps can be minimized with geostationary satellites owing to their higher scanning frequency over a given region, allowing more sampling opportunities as clouds move. However, geostationary ocean color missions have not been available for most regions of the globe. The Advanced Baseline Imager (ABI) aboard the Geostationary Operational Environmental Satellite R series (GOES-R) is a potential alternative for the U.S. owing to improved signal-to-noise ratio compared with its predecessors. Nonetheless, retrieving [Chl-a] from ABI measurements is challenging; the number, placement, and width of its spectral bands are suboptimal and an “ocean-grade” atmospheric correction scheme is lacking. Here we describe the construction of a deep learning-based algorithm to derive [Chl-a] from GOES-16 ABI L1B radiance and ancillary data without an explicit atmospheric correction algorithm. A large (∼22 million examples) training dataset of paired ABI observations and [Chl-a] estimates derived from the Visible Infrared Imaging Radiometer Suite (VIIRS) on the Suomi National Polar-orbiting Partnership (SNPP) were used to train, test and validate a 10-layer deep neural network model to retrieve [Chl-a] from ABI data. Preliminary results show that without atmospheric correction but with the aid of deep learning, it is possible to detect open ocean features with strong contrast such as across eddies and fronts, but there are still significant issues in coastal waters and occasionally in the open oceans. In general, the ABI-derived [Chl-a] agrees reasonably well with VIIRS-derived values in the open ocean during 3–5 h of the day. The location of oceanographic features such as fronts and eddies in the ABI-derived [Chl-a] maps are also coincident with those obtained from traditional ocean color and sea surface temperature retrievals. The findings of this study demonstrate that deep learning is a powerful tool for capturing intricate patterns that are too subtle for traditional algorithms developed by humans, Furthermore, the utilization of GOES-R ABI has the potential to complement polar-orbiting satellite data and mitigate data gaps.