Chapter 18
Identification of Marine Microalgae by Neural
Network Analysis of Simple Descriptors of Flow
Cytometric Pulse Shapes
M.F. Wilkins . L. Boddy . G.ß.J. Dubelaar
18.1
Introduction
Phytoplankton playa pivotal role in marine ecosystems - collectively fuelling the
food web, sometimes forming nuisance blooms, and implicated in climate control.
They are sensitive bioindicators in marine ecosystems. Thus, knowledge of
species composition, distribution and abundance in the worlds oceans is essential.
Traditionally such data have been obtained by microscopic analysis in the
laboratory, but this is laborious and time-consuming, abundance estimates are
uncertain due to limitations on the number of cells that can be counted, and
analysis is often performed a long time after sampling. Analytical flow cytometry
(AFC) is a valuable research tool in marine science (ßurkill and Mantoura 1990;
Jonker et al. 1995) that negates many of these problems. AFC measures various
light scatter, diffraction and fluorescence parameters on individual cells, at rates of
about 10 3 cells sec-I, providing signatures which can allow taxa to be
discriminated.
The vast quantities of non-normally distributed, multivariate data that AFC
generates are difficult to analyse by multivariate statistical methods, but artificial
neural networks (ANNs; Lippmann 1987; Hush and Horne 1993; Fu 1994; Haykin
1994) have been successfully employed. These typically consist of a three-layered
structure of simple data processing elements or nodes (corresponding to the neural
cells of their biological counterparts) connected by weighted connections. The
input layer contains one node for each input parameter, while the output layer
contains one node corresponding to each of the potential categories to wh ich the
input pattern may belong. The network is trained to recognise the different
categories of data via a learning procedure, during which the network is repeatedly
presented with labelIed examples of each data category and the internal weighted
connections between nodes are modified to produce a network output that more
nearly reflects the correct identification. Following training an unknown data
pattern can be presented to the network and the network output indicates the most
Identification of Marine Microalgae by Neural
Network Analysis of Simple Descriptors of Flow
Cytometric Pulse Shapes
M.F. Wilkins . L. Boddy . G.ß.J. Dubelaar
18.1
Introduction
Phytoplankton playa pivotal role in marine ecosystems - collectively fuelling the
food web, sometimes forming nuisance blooms, and implicated in climate control.
They are sensitive bioindicators in marine ecosystems. Thus, knowledge of
species composition, distribution and abundance in the worlds oceans is essential.
Traditionally such data have been obtained by microscopic analysis in the
laboratory, but this is laborious and time-consuming, abundance estimates are
uncertain due to limitations on the number of cells that can be counted, and
analysis is often performed a long time after sampling. Analytical flow cytometry
(AFC) is a valuable research tool in marine science (ßurkill and Mantoura 1990;
Jonker et al. 1995) that negates many of these problems. AFC measures various
light scatter, diffraction and fluorescence parameters on individual cells, at rates of
about 10 3 cells sec-I, providing signatures which can allow taxa to be
discriminated.
The vast quantities of non-normally distributed, multivariate data that AFC
generates are difficult to analyse by multivariate statistical methods, but artificial
neural networks (ANNs; Lippmann 1987; Hush and Horne 1993; Fu 1994; Haykin
1994) have been successfully employed. These typically consist of a three-layered
structure of simple data processing elements or nodes (corresponding to the neural
cells of their biological counterparts) connected by weighted connections. The
input layer contains one node for each input parameter, while the output layer
contains one node corresponding to each of the potential categories to wh ich the
input pattern may belong. The network is trained to recognise the different
categories of data via a learning procedure, during which the network is repeatedly
presented with labelIed examples of each data category and the internal weighted
connections between nodes are modified to produce a network output that more
nearly reflects the correct identification. Following training an unknown data
pattern can be presented to the network and the network output indicates the most