Operational viability of stock-separation using shape indices derived from the otolith morphometric outline. An example using sprat and herring.

We present a study concerning the viability of stock-separation of two highly mobile Clupeids (sprat Sprattus sprattus and herring Clupea harengus) using otolith morphometrics. Our analysis focuses on three stock discrimination problems with the aim of reassigning individuals to their source populations using a series of experiments undertaken using a popular suite of machine learning software known as the Waikato Environment for Knowledge Analysis. Within this framework 14 feature sets derived using 3 static transformations that encode combinations of size and shape were explored. By applying 9 state-of-the-art classification algorithms and using leave-one-out cross validation, we assessed how well each of the systems generalise to an independent data set. We performed 1260 cross-validated experiments on the data set, partitioned by age and using ground truth data provided by expert readers. To assess the saliency of size and shape features within each age partition, half of the feature sets include size indices whilst the remainder encode only shape.
We found that for juvenile fish feature sets that encode only shape perform well, but those that retain size indices hold slightly more potential for classification. As specimens increase in age discrimination based on morphometric features becomes ineffective; however as fish approach maturity transforms that encode shape appear to be more robust to aging than those based on size. Using significance tests we show that for juveniles, our results with respect to transforms are significant but those with respect to classification algorithms are not. This study suggests that current operational methods of stock discrimination of Clupeids based on measurements of early incremental growth are likely to be effective, with little benefit in supplementing this information with shape indices derived from morphological outlines.