Comparing the Performance of Classification Methods Used in Fisheries

Classification methods are an important tool used evaluating spatial structure in fisheries, especially when using otolith chemistry as a natural tag. Different classification methods have been used in the literature to increase classification success, including parametric methods such as linear and quadratic discriminant-function analysis (LDFA, QDFA) and computer-numeric approaches such as neural networks and random forests, among others. Very few studies have addressed, by direct comparison, the strengths and weakness of these methods. A recent  study showed that computer-numeric approaches resulted in better classification than parametric methods, but this study relied on un-normalized data. Theoretically, parametric methods should provide at least equal, if not superior accuracy when the assumptions of the methods are met. We used simulations on the Iris data set used by Fisher to develop multivariate statistics techniques and on menhaden otolith chemistry obtained in the field to directly compare these methods. We show that parametric methods such as LDFA and QDFA consistently outperform computer-numeric methods. We will discuss the constraints on methods such as the normality assumption and how non-normality degrades parametric performance. Although we compare data from otolith chemistry as our example, the issues we raise are applicable to any study that seeks to increase classification accuracy in fisheries.