372
S.G. Robertson . A.K. Morison
Neural networks may side-step this problem. We showed that a backpropagation neural network, using single transects across otolith images, had the
potential to reproduce the age estimates of an experienced reader with acceptable
precision for two of the three species tested (Robertson and Morison 1999). The
method worked better than any previous methods and for a greater range of age
classes.
Using neural networks to automate the fish age estimation is a fundamentally
different approach to those used previously. Past attempts have mosdy tried to
duplicate the outcome by duplicating the process used by human readers: identify
each valid annual increment and then count them. Using neural networks aims to
duplicate the outcome without identifying the process. This pragmatic approach
focuses on the quality of the age estimates without putting effort into developing
algorithms to identify increments in images. It recognizes that the process used by
human readers is enormously complex. Experienced otolith readers use a
combination of features of each otolith to assist them in making age estimates.
They consider the shape of the otolith, the width and spacing of increments and
how this changes from the otolith centre (the primordium) to the edge, the
consistency or completeness of increments around the image, the sharpness of the
transition from opaque to translucent zones. Even for fish of the same age these
patterns have significant individual variation, and may vary between species, with
sex and maturity, with fish size, and among different sub-populations, and from
year to year. In interpreting this variation readers employ their experience from
having viewed many thousands otolith of the same species. The way this is done is
difficult to codify into strict decision rules. This is further complicated in that
weight given to different aspects may vary among readers, without necessarily
compromising the accuracy of the estimated ages. The results of the initial trial of
a neural network were encouraging but the sampie size was very small (40
individuals) and the testing method did not use a completely independent sampie
set to evaluate network performance. We were also aware that other network types
and additional data inputs may substantially increase the effectiveness of the
networks. A more rigorous test of the approach was needed.
19.4
The Application of a Probabilistic Neural Network to Fish
Age Estimation
The same three species were used in these experiments as in the pilot study:
Pagrus auratus (Sparidae) (987 sampies), Acanthopagrus butcheri (Sparidae)
(913 sampies) and Macruronus novaezelandiae (Merlucciidae) (1531 sampies).
Age estimates for these sampies had previously been made using the procedures
and protocols described by Morison et al. (1998a). Thin-sections of sagittal
otoliths viewed with transmitted light were used for all species, and the number of
opaque increments used to estimate the age. This method of age estimation has
previously been validated for A. butcheri by Morison et al. (1998b), for P. auratus
by Francis et al. (1992), and for M. novaezelandiae by Kalish et al. (1997) and
S.G. Robertson . A.K. Morison
Neural networks may side-step this problem. We showed that a backpropagation neural network, using single transects across otolith images, had the
potential to reproduce the age estimates of an experienced reader with acceptable
precision for two of the three species tested (Robertson and Morison 1999). The
method worked better than any previous methods and for a greater range of age
classes.
Using neural networks to automate the fish age estimation is a fundamentally
different approach to those used previously. Past attempts have mosdy tried to
duplicate the outcome by duplicating the process used by human readers: identify
each valid annual increment and then count them. Using neural networks aims to
duplicate the outcome without identifying the process. This pragmatic approach
focuses on the quality of the age estimates without putting effort into developing
algorithms to identify increments in images. It recognizes that the process used by
human readers is enormously complex. Experienced otolith readers use a
combination of features of each otolith to assist them in making age estimates.
They consider the shape of the otolith, the width and spacing of increments and
how this changes from the otolith centre (the primordium) to the edge, the
consistency or completeness of increments around the image, the sharpness of the
transition from opaque to translucent zones. Even for fish of the same age these
patterns have significant individual variation, and may vary between species, with
sex and maturity, with fish size, and among different sub-populations, and from
year to year. In interpreting this variation readers employ their experience from
having viewed many thousands otolith of the same species. The way this is done is
difficult to codify into strict decision rules. This is further complicated in that
weight given to different aspects may vary among readers, without necessarily
compromising the accuracy of the estimated ages. The results of the initial trial of
a neural network were encouraging but the sampie size was very small (40
individuals) and the testing method did not use a completely independent sampie
set to evaluate network performance. We were also aware that other network types
and additional data inputs may substantially increase the effectiveness of the
networks. A more rigorous test of the approach was needed.
19.4
The Application of a Probabilistic Neural Network to Fish
Age Estimation
The same three species were used in these experiments as in the pilot study:
Pagrus auratus (Sparidae) (987 sampies), Acanthopagrus butcheri (Sparidae)
(913 sampies) and Macruronus novaezelandiae (Merlucciidae) (1531 sampies).
Age estimates for these sampies had previously been made using the procedures
and protocols described by Morison et al. (1998a). Thin-sections of sagittal
otoliths viewed with transmitted light were used for all species, and the number of
opaque increments used to estimate the age. This method of age estimation has
previously been validated for A. butcheri by Morison et al. (1998b), for P. auratus
by Francis et al. (1992), and for M. novaezelandiae by Kalish et al. (1997) and