Chapter 18 . Identification of Marine Microalgae
18.4
Discussion
365
These results demonstrate that the incorporation of pulse shape information
provides a modest increase in discriminating power, and that principal component
analysis can be used to separate out genuine variation between pulse shapes, due
to differences between the organisms under study, from spurious variation arising
from the physics of the data capture process.
The data used in this study all originated from small single-celled organisms, a
consequence of the difficulty of obtaining sufficient training data for larger
organisms due to the limited internal buffer space of the CytoBuoy. The pulse
shape information for larger organisms, e.g. colony and chain forming species,
contains proportionately more discriminatory information (Fig. 18.1). Such
particles can be up to a millimeter in length (2000 sampies).
The recognition performance of 69.2% for 36 species appears low by
comparison with previous studies: for example 72 taxa were identified with 70%
success from FACSort data (Boddy et al. , 2000), while 34 species were
discriminated with over 90% correct recognition from EurOPA data (Wilkins et
al. , 1999). This latter result is due to the higher resolution of the EurOPA
cytometer, wh ich also incorporated two-Iaser excitation and additional diffraction
parameters as an index of particle shape. The CytoBuoy prototype has aresolution
2-3 times lower than the EurOPA, due to compromises in the optics imposed by
its compact design and autonomous mode of operation. The optics of the
CytoBuoy have since been substantially redesigned in order to improve sensitivity
and resolution.
18.5
Conclusions
The use of AFC pulse shape information does improve discrimination of
microalgal taxa, and is likely to be even more useful when species that form
chains are to be discriminated. The use of RBF ANNs was again shown to be a
rapid and useful tool for analysing large sets of high dimensional data.
Acknowledgements
This research, together with the AimsNet software, forms part of AIMS, a project
funded by the Commission of the European Community, CEC grant no. MAS3CT97-0080. We thank all partners for valuable discussion.
18.4
Discussion
365
These results demonstrate that the incorporation of pulse shape information
provides a modest increase in discriminating power, and that principal component
analysis can be used to separate out genuine variation between pulse shapes, due
to differences between the organisms under study, from spurious variation arising
from the physics of the data capture process.
The data used in this study all originated from small single-celled organisms, a
consequence of the difficulty of obtaining sufficient training data for larger
organisms due to the limited internal buffer space of the CytoBuoy. The pulse
shape information for larger organisms, e.g. colony and chain forming species,
contains proportionately more discriminatory information (Fig. 18.1). Such
particles can be up to a millimeter in length (2000 sampies).
The recognition performance of 69.2% for 36 species appears low by
comparison with previous studies: for example 72 taxa were identified with 70%
success from FACSort data (Boddy et al. , 2000), while 34 species were
discriminated with over 90% correct recognition from EurOPA data (Wilkins et
al. , 1999). This latter result is due to the higher resolution of the EurOPA
cytometer, wh ich also incorporated two-Iaser excitation and additional diffraction
parameters as an index of particle shape. The CytoBuoy prototype has aresolution
2-3 times lower than the EurOPA, due to compromises in the optics imposed by
its compact design and autonomous mode of operation. The optics of the
CytoBuoy have since been substantially redesigned in order to improve sensitivity
and resolution.
18.5
Conclusions
The use of AFC pulse shape information does improve discrimination of
microalgal taxa, and is likely to be even more useful when species that form
chains are to be discriminated. The use of RBF ANNs was again shown to be a
rapid and useful tool for analysing large sets of high dimensional data.
Acknowledgements
This research, together with the AimsNet software, forms part of AIMS, a project
funded by the Commission of the European Community, CEC grant no. MAS3CT97-0080. We thank all partners for valuable discussion.