Taxonomic classification of phytoplankton with multivariate optical computing, part I: design and theoretical performance of multivariate optical elements.

Phytoplankton are single-celled, photosynthetic algae and cyanobacteria found in all aquatic environments. Differential pigmentation between phytoplankton taxa allows use of fluorescence excitation spectroscopy for discrimination and classification. For this work, we applied multivariate optical computing (MOC) to emulate linear discriminant vectors of phytoplankton fluorescence excitation spectra by using a simple filter-fluorometer arrangement. We grew nutrient-replete cultures of three differently pigmented species: the coccolithophore Emiliania huxleyi, the diatom Thalassiosira pseudonana, and the cyanobacterium Synechococcus sp. Linear discriminant analysis (LDA) was used to determine a suitable set of linear discriminant functions for classification of these species over an optimal wavelength range. Multivariate optical elements (MOEs) were then designed to predict the linear discriminant scores for the same calibration spectra. The theoretical performance specifications of these MOEs are described.

Taxonomic classification of phytoplankton with multivariate optical computing, part II: design and experimental protocol of a shipboard fluorescence imaging photometer.

Differential pigmentation between phytoplankton allows use of fluorescence excitation spectroscopy for the discrimination and classification of different taxa. Here, we describe the design and performance of a fluorescence imaging photometer that exploits taxonomic differences for discrimination and classification. The fluorescence imaging photometer works by illuminating individual phytoplankton cells through an asynchronous spinning filter wheel, which produces bar code-like streaks in a fluorescence image. A filter position is covered with an opaque filter to create a reference dark position in the filter wheel rotation that is used to match each fluorescence streak with the corresponding filter. Fluorescence intensities of the imaged streaks are then analyzed for the purpose of spectral analysis, which allows taxonomic classification of the organism that produced the streaks. The theoretical performance and signal-to-noise ratio (SNR) specifications of these MOEs are described in Part I of this series. This report describes optical layout, flow cell design, magnification, depth of field, constraints on filter wheel and flow velocities, procedures for blank subtraction and flat-field correction, the measurement scheme of the instrument, and measurement of SNR as a measurement of filter wheel frequency. This is followed by an analysis of the sources of variance in measurements made by the photometer on the coccolithophore Emiliania huxleyi. We conclude that the SNR of E. huxleyi measurements is not limited by the sensitivity or noise attributes of the measurement system, but by dynamics in the fluorescence efficiency of the E. huxleyi cells. Even so, the minimum SNR requirements given in Part I for the instrument are met.

Taxonomic classification of phytoplankton with multivariate optical computing, part III: demonstration.

We describe the automatic analysis of fluorescence tracks of phytoplankton recorded with a fluorescence imaging photometer. The optical components and construction of the photometer were described in Part I and Part II of this series in this issue. An algorithm first isolates tracks corresponding to a single phytoplankter transit in the nominal focal plane of a flow cell. Then, the fluorescence streaks in the track that correspond to individual optical elements on the filter wheel are identified. The fluorescence intensity of each streak is integrated and used to calculate ratios. This approach was tested using 853 fluorescence measurements of the coccolithophore Emiliania huxleyi and the diatom Thalassiosira pseudonana. Average intensity ratios for the two classes closely follow those predicted in Part I of this series, with a distribution of ratios in each class that is consistent with the signal-to-noise ratio calculations in Part II for single cells. No overlap of the two class ratios was observed, yielding perfect classification.