Accurate automatic quantification of taxa-specific plankton abundance using dual classification with correction

Optical imaging samplers are becoming widely used in plankton ecology, but image analysis methods have lagged behind image acquisition rates. Automated methods for analysis and recognition of plankton images have been developed, which are capable of real-time processing of incoming image data into major taxonomic groups. The limited accuracy of these methods can require significant manual post-processing to correct the automatically generated results, in order to obtain accurate estimates of plankton abundance patterns. We present here a dual-classification method in which each plankton image is first identified using a shaped-based feature set and a neural network classifier, and then a second time using a texture-based feature set and a support vector machine classifier. The plankton image is considered to belong to a given taxon only if the 2 identifications agree; otherwise it is labeled as unknown. This dual-classification method greatly reduces the false positive rate, and thus gives better abundance estimation in regions of low relative abundance. A confusion matrix is computed from a set of training images in order to determine the detection and false positives rates. These rates are used to correct abundances estimated from the automatic classification results. Aside from the manual sorting required to generate the initial training set of images, this dual-classification method is fully automatic and does not require sub- sequent manual correction of automatically sorted images. The resulting abundances agree closely with those obtained using manually sorted results. A set of images from a Video Plankton Recorder was used to evaluate this method and compare it with previously reported single-classifier results for major taxa.