UC San Diego

This paper improves on the accuracy of other published machine learning results for quantifying plankton samples. The contributions of this work are: (1) Clarifying the number of expertly labeled images required for machine learning results. (2) Providing guidance as to what algorithms provide the best performance, and how to tune them. (3) Leveraging an ensemble of models to achieve recall rates beyond any single algorithm. (4) Investigating the applicability of abstaining. (5) Using size fractionation to learn more efficiently. (6) Analysis of efficacy of simple geometric features for plankton identification.