Classification of species in the genus Trapa by morphometric fruiting features|
Dmitry KISLOV, Ekaterina BERESTENKO
Botanical Garden-Institute FEB RAS, Vladivostok, Russia
The paper discusses new approaches for solving the taxonomic problem
of Trapa species, that occur in the territory of Primorsky krai (Russian Far East).
At the present time, the problem of classification of Trapa species is still open, mostly,
because of high variability of its morphological features. Therefore, deriving
significant methods of discriminating between a specific set of objects is very
difficult. It is generally accepted that Trapa fruit morphological features are
used for Trapa species discrimination. In this work, attempts for improving existing
classification schemes of Trapa species are based on two positions. First of these:
to classify properly, one has to use most informative morphometric features of
fruits, controlling feature extraction process by some cross-validation technique.
The second is to follow treatments that are statistically valid. As a result of our research,
we built a classification system of Trapa species. The system uses
up to 27 morphmetric features of Trapa fruits belonging to 8 species growing
at the territory of Primorsky krai. Computations are based on classical and modern
pattern recognition techniques (linear classification, knn-classification, consensus
clustering, analysis of variance) and implemented in Python. It is shown,
that Trapa species could be determined by its common morphometric features
with high accuracy. For instance, misclassification errors for linear classification
algorithm computed by leave-one-out cross-validation method are the following: T.
incisa – 1.3 %, T. nedoluzhkoi – 0 %, T. maximowiczii – 2.5 %. The server of
the classification system is built over the Django Web Framework and includes
additional Python modules contributed by authors of the work. As generalization
of features extraction procedure, a hierarchical classification tree was formed. It is
presented as ensemble of nested dichotomies leading to specific Trapa species.
© 2012 Organizing Committee