Chapter 18 . Identification of Marine Microalgae
363
automatically determined using the orthogonal least squares learning algorithm to
select an optimal sub set from a large pool of candidate HLNs (Chen et al. 1991).
Following training, the test data set was used to assess the network performance
in terms of proportion of correct identification, and to determine which species
were consistently misidentified. Each network was replicated five times and the
results averaged.
18.2.6
Hardware and Software
All RBF networks were simulated in software using AimsNet
(http://www.cf.ac.uklbiosilstaff/wilkins/aimsnet). a software package developed
by one of the authors (M.F.W.) under the AIMS project (CEC grant no. MAS3CT97-0080) running on a 500MHz PC under WindowsNT.
18.3
Results
Using the pulse shape data alone, six species (Synechococcus sp., Cryptomonas
calceiformis, Cryptomonas maculata, Cryptomonas reticulata, Tetraselmis
tetrathele and Hemiselmis brunnescens) were all discriminated correctly with
greater than 50% accuracy; however over the whole test data set only 28.6% of
data patterns were correctly recognised (Table 18.1). The corresponding figure
using summary statistics alone was 67.2%; this rose slightly to 69.2% when the
pulse shape and summary statistic information were combined. The addition of
pulse shape information generally improved the discrimination between species
from the same genus, e.g. Tetraselmis (4 species), Chrysochromulina (3 species),
Cryptomonas (3 species) and Dunaliella (2 species). For example, the amount of
T. striata misidentified as T. tetrathele was reduced by a factor of two; plotting the
pulse shapes for these species shows that while the SSC, FLO and FLR signals are
similar, the FSC signal for T. tetrathele shows a pronounced asymmetry which can
be exploited to aid discrimination (Fig. 18.3). For the three Cryptomonas species,
the FSC signal for C. maculata shows two distinct peaks that are significantly
eioser together than for the other two species (Fig. 18.4).
363
automatically determined using the orthogonal least squares learning algorithm to
select an optimal sub set from a large pool of candidate HLNs (Chen et al. 1991).
Following training, the test data set was used to assess the network performance
in terms of proportion of correct identification, and to determine which species
were consistently misidentified. Each network was replicated five times and the
results averaged.
18.2.6
Hardware and Software
All RBF networks were simulated in software using AimsNet
(http://www.cf.ac.uklbiosilstaff/wilkins/aimsnet). a software package developed
by one of the authors (M.F.W.) under the AIMS project (CEC grant no. MAS3CT97-0080) running on a 500MHz PC under WindowsNT.
18.3
Results
Using the pulse shape data alone, six species (Synechococcus sp., Cryptomonas
calceiformis, Cryptomonas maculata, Cryptomonas reticulata, Tetraselmis
tetrathele and Hemiselmis brunnescens) were all discriminated correctly with
greater than 50% accuracy; however over the whole test data set only 28.6% of
data patterns were correctly recognised (Table 18.1). The corresponding figure
using summary statistics alone was 67.2%; this rose slightly to 69.2% when the
pulse shape and summary statistic information were combined. The addition of
pulse shape information generally improved the discrimination between species
from the same genus, e.g. Tetraselmis (4 species), Chrysochromulina (3 species),
Cryptomonas (3 species) and Dunaliella (2 species). For example, the amount of
T. striata misidentified as T. tetrathele was reduced by a factor of two; plotting the
pulse shapes for these species shows that while the SSC, FLO and FLR signals are
similar, the FSC signal for T. tetrathele shows a pronounced asymmetry which can
be exploited to aid discrimination (Fig. 18.3). For the three Cryptomonas species,
the FSC signal for C. maculata shows two distinct peaks that are significantly
eioser together than for the other two species (Fig. 18.4).