Chapter 19
Age Estimation of Fish Using a Probabilistic
Neural Network
S.G. Robertson . A.K. Morison
19.1
Introduction
Age composition data provide fundamental insights into fish biology and stock
productivity and allow the estimation of the basic parameters for describing
growth, mortality rates and recruitment. Much time and money is spent on the
collection and preparation of sampIes, and skilled technicians labour for many
hours at microscopes, counting increments in the prepared structures. It is
estimated that over 1 million fish were aged worldwide in 1999, mostly using
scales and otoliths (Campana and Thorrold 2001). However, the process is
somewhat subjective and there is much interest in automating the process and
making estimates more reliable. To date none of the tested methods have been
successful. A pilot study by Robertson and Morison (1999) first suggested that
neural networks may provide the way forward for this previously intractable
problem. In this paper we firstly give abrief ac count of traditiona1 approaches to
age estimation. We then describe the previous attempts to develop automatie or
computer-aided methods and the problems they have encountered. Finally we
describe the results of arecent application of a probabilistic neural network to the
process of age estimation in fish and discuss the strengths of this novel approach.
19.2
Traditional Methods of Age Estimation
Fisheries scientists have been estimating the age and growth of fish by various
means since it was first realised that modes in length-frequency distributions could
be used to estimate the age of discrete cohorts. Since then methods of estimating
fish age have proliferated but usually involve counting of growth increments on
hard parts (scales, vertebrae, fin rays, fin spines, eye lenses and otoliths). On
otoliths (usually the structure of choice for all but the youngest of fish), these
increments are composed of alternating opaque and translucent zones. The
Age Estimation of Fish Using a Probabilistic
Neural Network
S.G. Robertson . A.K. Morison
19.1
Introduction
Age composition data provide fundamental insights into fish biology and stock
productivity and allow the estimation of the basic parameters for describing
growth, mortality rates and recruitment. Much time and money is spent on the
collection and preparation of sampIes, and skilled technicians labour for many
hours at microscopes, counting increments in the prepared structures. It is
estimated that over 1 million fish were aged worldwide in 1999, mostly using
scales and otoliths (Campana and Thorrold 2001). However, the process is
somewhat subjective and there is much interest in automating the process and
making estimates more reliable. To date none of the tested methods have been
successful. A pilot study by Robertson and Morison (1999) first suggested that
neural networks may provide the way forward for this previously intractable
problem. In this paper we firstly give abrief ac count of traditiona1 approaches to
age estimation. We then describe the previous attempts to develop automatie or
computer-aided methods and the problems they have encountered. Finally we
describe the results of arecent application of a probabilistic neural network to the
process of age estimation in fish and discuss the strengths of this novel approach.
19.2
Traditional Methods of Age Estimation
Fisheries scientists have been estimating the age and growth of fish by various
means since it was first realised that modes in length-frequency distributions could
be used to estimate the age of discrete cohorts. Since then methods of estimating
fish age have proliferated but usually involve counting of growth increments on
hard parts (scales, vertebrae, fin rays, fin spines, eye lenses and otoliths). On
otoliths (usually the structure of choice for all but the youngest of fish), these
increments are composed of alternating opaque and translucent zones. The