Chapter 19 . Age Estimation of Marine Fish
371
19.3
Approaches to Automation in Fish Age Estimation
Williams and Bedford (1974) remarked that "otolith reading remains ... as much
an art as a science" and the subjective nature of the interpretation of the
increments has driven the search for more objective methods. In some instances
these have exploited the correlation between otolith growth and fish by developing
linear regressions relating fish age and various measures of otolith size or weight
(Boehlert 1985; Fletcher 1995). The statistical approach using correlates of fish
age has not been widely adopted. Its use for the age estimation of short-lived
species such as pilchards (Sardinops neopilchardus) (Fletcher 1995) owes as
much to the great difficulties experienced with the interpretation of the highly
variable increment patterns on the otoliths of these species, as to the efficacy of
the technique. Spatial and temporal variation in these relationships also create
problems for the routine application of such methods (Worthington et al. 1995).
The processing power of modem computers and image analysis software has
stimulated much interest in their ability to automate the increment recognition and
age estimation process. At one level these have required or allowed interaction
with the user to assist in the identification of increments (Macy 1995; Cailliet
1996). More advanced methods have sought algorithms that identified increments
as minima and maxima in digital profiles of brightness level (Troadec 1991;
Welleman and Storbeck 1995; Lagardere and Troadec 1997; Morison and
Robertson 1997; Takashima et al. 2000).
One problem with these approaches is that they assume that one growth
increment is formed each year. This is usually true, but it is also common that
such increments are comprised of two or more distinct sub-annual increments,
called false increments. The ability to consistently distinguish annually formed
increments is the valuable skill that an experienced reader brings to the process.
Recognition of peaks and troughs in a brightness profile is statistically simple, but
discerning which should be counted and wh ich excluded is not. These methods are
also only reliable on sampies with very clear increments, but such sampies are
usually in the minority.
We have tried the 'find-the-peaks-and-count-them' approach for these species
in a pilot study comparing different automation methods and found them to be
satisfactory only in a very narrow range of circumstances (Morison and Robertson
1997). The signals are inherently noisy and the change in slope method for
identifying peaks (e.g. Takashima et al. 2000) requires specifying a particular
search bandwidth. As increments usually become narrower towards the otolith
margins, this bandwidth must be progressively reduced to work equally re!iably on
the broad inner increments, when growth is fast, and on the narrow outer
increments, formed when growth slows. Signal processing techniques have also
been applied to these data involving scaling, demodulation and spectral analysis,
to extract the number of cycles as an age estimate (Troadec 1991; Lagardere and
Troadec 1997; Troadec et al. 2000). However, in either case the approach requires
a rigid algorithm that reflects the expected reduction in increment spacing. The
expected pattern has first to be identified and numerically coded and therefore the
methods do not cope weIl with variations from the standard pattern.
371
19.3
Approaches to Automation in Fish Age Estimation
Williams and Bedford (1974) remarked that "otolith reading remains ... as much
an art as a science" and the subjective nature of the interpretation of the
increments has driven the search for more objective methods. In some instances
these have exploited the correlation between otolith growth and fish by developing
linear regressions relating fish age and various measures of otolith size or weight
(Boehlert 1985; Fletcher 1995). The statistical approach using correlates of fish
age has not been widely adopted. Its use for the age estimation of short-lived
species such as pilchards (Sardinops neopilchardus) (Fletcher 1995) owes as
much to the great difficulties experienced with the interpretation of the highly
variable increment patterns on the otoliths of these species, as to the efficacy of
the technique. Spatial and temporal variation in these relationships also create
problems for the routine application of such methods (Worthington et al. 1995).
The processing power of modem computers and image analysis software has
stimulated much interest in their ability to automate the increment recognition and
age estimation process. At one level these have required or allowed interaction
with the user to assist in the identification of increments (Macy 1995; Cailliet
1996). More advanced methods have sought algorithms that identified increments
as minima and maxima in digital profiles of brightness level (Troadec 1991;
Welleman and Storbeck 1995; Lagardere and Troadec 1997; Morison and
Robertson 1997; Takashima et al. 2000).
One problem with these approaches is that they assume that one growth
increment is formed each year. This is usually true, but it is also common that
such increments are comprised of two or more distinct sub-annual increments,
called false increments. The ability to consistently distinguish annually formed
increments is the valuable skill that an experienced reader brings to the process.
Recognition of peaks and troughs in a brightness profile is statistically simple, but
discerning which should be counted and wh ich excluded is not. These methods are
also only reliable on sampies with very clear increments, but such sampies are
usually in the minority.
We have tried the 'find-the-peaks-and-count-them' approach for these species
in a pilot study comparing different automation methods and found them to be
satisfactory only in a very narrow range of circumstances (Morison and Robertson
1997). The signals are inherently noisy and the change in slope method for
identifying peaks (e.g. Takashima et al. 2000) requires specifying a particular
search bandwidth. As increments usually become narrower towards the otolith
margins, this bandwidth must be progressively reduced to work equally re!iably on
the broad inner increments, when growth is fast, and on the narrow outer
increments, formed when growth slows. Signal processing techniques have also
been applied to these data involving scaling, demodulation and spectral analysis,
to extract the number of cycles as an age estimate (Troadec 1991; Lagardere and
Troadec 1997; Troadec et al. 2000). However, in either case the approach requires
a rigid algorithm that reflects the expected reduction in increment spacing. The
expected pattern has first to be identified and numerically coded and therefore the
methods do not cope weIl with variations from the standard pattern.