ESTIMATION OF OCEANIC PARTICULATE ORGANIC CARBON WITH MACHINE LEARNING

• Published in: ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
• Main Authors: R. Sauzède, J. E. Johnson, H. Claustre, G. Camps-Valls, A. B. Ruescas
• Format: Article in Journal/Newspaper
• Language: English
• Published: Copernicus Publications 2020
• Subjects: Technology; T; Engineering (General). Civil engineering (General); TA1-2040; Applied optics. Photonics; TA1501-1820; North Atlantic
• Online Access: https://doi.org/10.5194/isprs-annals-V-2-2020-949-2020; https://doaj.org/article/7fbf854757b04e86a31f27638c455dfe

Understanding and quantifying ocean carbon sinks of the planet is of paramount relevance in the current scenario of global change. Particulate organic carbon (POC) is a key biogeochemical parameter that helps us characterize export processes of the ocean. Ocean color observations enable the estimation of bio-optical proxies of POC (i.e. particulate backscattering coefficient, b bp ) in the surface layer of the ocean quasi-synoptically. In parallel, the Argo program distributes vertical profiles of the physical properties with a global coverage and a high spatio-temporal resolution. Merging satellite ocean color and Argo data using a neural networkbased method has already shown strong potential to infer the vertical distribution of bio-optical properties at global scale with high space-time resolution. This method is trained and validated using a database of concurrent vertical profiles of temperature, salinity, and bio-optical properties, i.e. b bp , collected by Biogeochemical-Argo (BGC-Argo) floats, matched up with satellite ocean color products. The present study aims at improving this method by 1) using a larger dataset from BGC-Argo network since 2016 for training, 2) using additional inputs such as altimetry data, which provide significant information on mesoscale processes impacting the vertical distribution of b bp , 3) improving the vertical resolution of estimation, and 4) examining the potential of alternative machine learning-based techniques. As a first attempt with the new data, we used some feature-specific preprocessing routines followed by a Multi-Output Random Forest algorithm on two regions with different ocean dynamics: North Atlantic and Subtropical Gyres. The statistics and the b bp profiles obtained from the validation floats show promising results and suggest this direction is worth investigating even further at global scale.