Classification and regression trees to predict the species of an unidentified Puntius specimen

Abstract

Out of 82 freshwater fish species in Sri Lanka, the Genus Puntius represents 16 species (19.5%). However, ambiguities in taxonomic identification of different Puntius species remain as a well known research area. Hence, in this study Classification and Regression Trees (CART) and Random Forests analysis were carried out to identify and differentiate among Puntius species using their morphometric, meristic and coded variables. Total of 316 specimens representing eight Sri Lankan Puntius species were collected at four different altitude ranges from five major river basins in Sri Lanka. Fifteen meristic characters, four coded variables and twenty three morphometric characters were recorded from each specimen. In the case of combining meristic and coded variables, the correct classification rate for model was 98% and the value of the Kappa statistic which is a chance-corrected measure of prediction was 0.982. In random forests analysis, the classification error rates based on the out of bag samples, averaged over many bootstrap samples, provide an unbiased estimate of prediction error for combination of meristic and coded variables was 0.95%. The overall correct classification rate of CART model to predict the species of an unidentified Puntius specimen using its morphometric measurements was 80% and the value of the Kappa statistic was 0.769. The corresponding unbiased estimate of prediction error from random forests for morphometric data was 14.24%. In this study eight Puntius species were considered and Number o f transverse scales (tr) and Total length, (TL) were the most important meristic and morphometric variable respectively for differentiating among those species.