Machine learning and multiscale methods in the identification of bivalve larvae

We describe a novel application of support vector machines and multiscale texture and color invariants to a problem in biological oceanography: the identification of 6 species of bivalve larvae. Our data consists of polarized color images of scallop and other bivalve larvae (between 2 and 17 days old) collected from the ocean by a shipboard optical imaging system of our design. Larvae of scallops, clams, and oysters are small (100 microns) with few distinguishing features when observed under standard light microscopy. However, the use of polarized light with a full wave retardation plate produces a vivid color, birefringence pattern. The patterns display very subtle differences between species, often not discernable to human observers. We show that a soft-margin support vector machine with Gaussian RBF kernel is a good discriminator on a feature set extracted from Gabor wavelet transforms and color distribution angles of each image. By constraining the Gabor center frequencies to be low, the resulting system can attain classification accuracy in excess of 90% for vertically oriented images, and in excess of 80% for randomly oriented images.