POSTER Session 4

THE UNDER-SAMPLING PROBLEM ASSOCIATED WITH EMPIRICAL ALGORITHMS

Empirical algorithms inherently grapple with under-sampling issues. Whether explicit such as a polynomial fit or implicit such as a machine learning model, they rely solely on existing data (hereon the ‘database’) and hence suffer from potentially unconstrained uncertainties when confronted with waters exhibiting characteristics not represented in this database. For example, water of previously unmeasured reflectance spectra. Furthermore, akin to many inverse problems, empirical inversions may not be unique, meaning different spectra in the training data could be associated with the same empirical retrieval, or very similar spectra could be linked to different retrievals. This challenge is compounded by uncertainties in both spectra and dataset. In this study we showcase the under-sampling problem by comparing satellite spectra to those in match-up datasets utilized for training. We suggest approaches and metrics to identify cases that deviate from the training dataset, facilitating their flagging to warn users about their departure from the training and evaluation data distribution. It is crucial that satellite products are generated with an uncertainty measure or a flag specifying how or whether the observations fit the model.

Emmanuel Boss, University of Maine, USA, [email protected], https://orcid.org/0000-0002-8334-9595

Robert Frouin, Scripps Institute of Oceanography, USA, [email protected]

Charlotte Begouen Demeaux, University of Maine, USA, [email protected]

Jing Tan, Scripps Institute of Oceanography, USA, [email protected]

Patrick Gray, University of Maine, USA, [email protected]

B.B. Cael, National Oceanographic Center, USA, [email protected]

Code of Conduct
Privacy Policy