Chapter 19 . Age Estimation of Marine Fish
379
The separation of training, test and production sets provides a test of the ability of
the trained networks to correctly classify new samples. These results will therefore
provide a more accurate indication of the accuracy that could be expected if the
method was implemented for ongoing production age estimation.
This study has also confirmed the value of three improvements suggested by
Robertson and Morison (1999). Firstly, it has showed that the DFT can effectively
capture the information from a transect across an otolith image. This robust and
widely used transformation greatly reduces the number of data inputs and hence
the complexity of the neural network model that needs to be constructed. The
PNN models were fitted with less than half the number of inputs used in the pilot
study, but were still able to include signal data from five transects rather than one,
and incorporate auxiliary data as weB. Lagardere and Troadec (1997) used a
Fourier transformation for estimating the number of daily increments on the
otoliths of larval sole Solea solea, but applied it to a demodulated signal. Further
refinements of this approach are possible, so that the information in a greater
amount of the image can be incorporated into future models.
Secondly, this study has shown that the biological data (fish length, otolith
weight, and date of capture) can be incorporated into neural networks and
contribute to their ability to estimate the age of samples. The relative importance
of these auxiliary data sources to each model can be assessed from the magnitude
of the individual scaling factors the trained network assigns to each input variable
. For this study the most important variables varied for each species: for A.
butcheri 9 of the 10 inputs with the highest scaling factors were Fourier transform
harmonics; for P. auratus and M. novaezelandiae otolith weight, transect lengths,
fish length and Fourier transform harmonics were aB given high values. However,
there was no clear point of demarcation of input variables with high or low scaling
factors. The nature of the differences between the species were unexpected,
especiaBy since the otoliths of A. butcheri and P. auratus are similar in shape and
clarity whereas those of M. novaezelandiae are more complex and difficult to
interpret.
Thirdly, this implementation of neural networks has confirmed that PNNs are
an appropriate network type for the problem of age estimation. The relative
performance of PNNs compared with other types of neural networks was not
explored here, and would be worthwhile, but the scaling factors they produce for
each network input are one feature of PNNs that shows there utility for fish age
estimation. They provide an ability to differentiate important inputs from those
that contribute little discriminating power in a manner that is analogous to
sensitivity tests. Sensitivity tests have been used to explore the relative importance
of the data sources and the strength of their contribution to the response variable
of interest (Walter et al. 2001). It does not add greatly to the complexity of the
model to incorporate aB the readily available data types into a model.
Nevertheless, it would be useful to have a formal procedure for the identification
of the minimum number of inputs needed to obtain satisfactory model predictions.
The PNNs showed an improvement over the results of the back-propagation
models used in the pilot study (Robertson and Morison 1999) in producing
acceptable results under more rigorous training and testing procedures. Results for
M. novaezelandiae in particular were superior, as the earlier neural network
379
The separation of training, test and production sets provides a test of the ability of
the trained networks to correctly classify new samples. These results will therefore
provide a more accurate indication of the accuracy that could be expected if the
method was implemented for ongoing production age estimation.
This study has also confirmed the value of three improvements suggested by
Robertson and Morison (1999). Firstly, it has showed that the DFT can effectively
capture the information from a transect across an otolith image. This robust and
widely used transformation greatly reduces the number of data inputs and hence
the complexity of the neural network model that needs to be constructed. The
PNN models were fitted with less than half the number of inputs used in the pilot
study, but were still able to include signal data from five transects rather than one,
and incorporate auxiliary data as weB. Lagardere and Troadec (1997) used a
Fourier transformation for estimating the number of daily increments on the
otoliths of larval sole Solea solea, but applied it to a demodulated signal. Further
refinements of this approach are possible, so that the information in a greater
amount of the image can be incorporated into future models.
Secondly, this study has shown that the biological data (fish length, otolith
weight, and date of capture) can be incorporated into neural networks and
contribute to their ability to estimate the age of samples. The relative importance
of these auxiliary data sources to each model can be assessed from the magnitude
of the individual scaling factors the trained network assigns to each input variable
. For this study the most important variables varied for each species: for A.
butcheri 9 of the 10 inputs with the highest scaling factors were Fourier transform
harmonics; for P. auratus and M. novaezelandiae otolith weight, transect lengths,
fish length and Fourier transform harmonics were aB given high values. However,
there was no clear point of demarcation of input variables with high or low scaling
factors. The nature of the differences between the species were unexpected,
especiaBy since the otoliths of A. butcheri and P. auratus are similar in shape and
clarity whereas those of M. novaezelandiae are more complex and difficult to
interpret.
Thirdly, this implementation of neural networks has confirmed that PNNs are
an appropriate network type for the problem of age estimation. The relative
performance of PNNs compared with other types of neural networks was not
explored here, and would be worthwhile, but the scaling factors they produce for
each network input are one feature of PNNs that shows there utility for fish age
estimation. They provide an ability to differentiate important inputs from those
that contribute little discriminating power in a manner that is analogous to
sensitivity tests. Sensitivity tests have been used to explore the relative importance
of the data sources and the strength of their contribution to the response variable
of interest (Walter et al. 2001). It does not add greatly to the complexity of the
model to incorporate aB the readily available data types into a model.
Nevertheless, it would be useful to have a formal procedure for the identification
of the minimum number of inputs needed to obtain satisfactory model predictions.
The PNNs showed an improvement over the results of the back-propagation
models used in the pilot study (Robertson and Morison 1999) in producing
acceptable results under more rigorous training and testing procedures. Results for
M. novaezelandiae in particular were superior, as the earlier neural network