Semi-automated image analysis for the identification of bivalve larvae from a Cape Cod estuary

Machine-learning methods for identifying planktonic organisms are becoming well-established. Although similar morphologies among species make traditional image identification methods difficult for larval bivalves, species-specific shell birefringence patterns under polarized light permit identification by color and texture-based features. This approach uses cross-polarized images of bivalve larvae, extracts Gabor and color angle features from each image, and classifies images using a Support Vector Machine. We adapted this method, which was established on hatchery-reared larvae, to identify bivalve larvae from a series of field samples from a Cape Cod estuary in 2009. This method had 98% identification accuracy for four hatchery-reared species. We used a multiplex polymerase chain reaction (PCR) method to confirm field identifications and to compare accuracies to the software classifications. Image classification of larvae collected in the field had lower accuracies than both the classification of hatchery species and PCR-based identification due to error in visually classifying unknown larvae and variability in larval images from the field. A six-species field training set had the best correspondence to our visual classifications with 75% overall agreement and individual species agreements from 63% to 88%. Larval abundance estimates for a time-series of field samples showed good correspondence with visual methods after correction. Overall, this approach represents a cost- and time-saving alternative to molecular-based identifications and can produce sufficient results to address long-term abundance and transport-based questions on a species-specific level, a rarity in studies of bivalve larvae.