Oral Session 10
Friday, October 11
10:10–11:10
10:10–10:30 | ABSTRACT 1011
ELUCIDATING THE MICRO(NS) FROM THE MACRO: THE ROLE OF MACHINE LEARNING IN PREDICTING PHYTOPLANKTON COMMUNITY COMPOSITION FROM OCEAN COLOR
For decades, we have grappled with the challenge of elucidating phytoplankton community composition (PCC) from measurements of ocean color. The Imaging Flow CyotoBot (IFCB), provides measurements of phytoplankton abundance and taxonomic identification – a departure from traditional phytoplankton marker pigment approaches. We are investigating the relationships between IFCB data and ocean color through various machine learning (ML) approaches. Our candidate IFCB model training datasets are still relatively sparse. Therefore, we utilized a synthetic dataset generated from the NOBM-OASIM (NASA Ocean Biogeochemical Model-Ocean-Atmosphere Spectral Irradiance Model) to model hyperspectral remote sensing reflectance associated with different phytoplankton taxonomic groups that occur in various relative abundances. The global simulation was subsetted to a region in the NW Atlantic to yield 7807 data points. After running a suite of candidate ML models, we converged on a random forest model characterized by 100 trees, and which utilized the Gini impurity criterion. The results were as follows: cyanophytes were correctly classified as the dominant group 80% of the time; chlorophytes were correctly predicted 86% of the time, and the model is most accurate for diatoms with successful prediction 92% of the time. We are further developing the ML approaches to include predictions of relative PCC abundances using Bayesian additive regression trees (BART), multivariate linear regression, and Gaussian processes, all of which have their own distinct advantages in terms of resistance to overfitting and suitability for sparse data. These efforts represent cutting edge advances in how we infer ecological information from hyperspectral ocean color data.
Susanne Craig, NASA Goddard Space Flight Center, and University of Maryland Baltimore County (UMBC) GESTAR II, USA, [email protected], https://orcid.org/ 0000-0002-8963-0951
Erdem Karaköylü, Private Consultant, USA, [email protected]
Ian Carroll, Goddard Space Flight Center, and University of Baltimore County GESTAR II, USA, [email protected]
Cecile Rousseaux, Goddard Space Flight Center, USA, [email protected]
Jeremy Werdell, Goddard Space Flight Center, USA, [email protected]
Questions?
Contact Jenny Ramarui,
Conference Coordinator,
at [email protected]
or (1) 301-251-7708