Variations on a Theme: The Classification of Benthic Macroinvertebrates
Acta Universitatis Tamperensis No. 1777


By Henry Joutsijoki
December 2012
Tampere University Press
Distributed By Coronet Books
ISBN: 9789514489525
152 pages
$82.50 Paper Original

This thesis focused on the classification of benthic macroinvertebrates by using machine learning methods. Special emphasis was placed on multi-class extensions of Support Vector Machines (SVMs). Benthic macroinvertebrates are used in biomonitoring due to their properties to react to changes in water quality. The use of benthic macroinvertebrates in biomonitoring requires a large number of collected samples. Traditionally benthic macroinvertebrates are separated and identified manually one by one from samples collected by biologists. This, however, is a time-consuming and expensive approach.

By the automation of the identification process time and money would be saved and more extensive biomonitoring would be possible. The aim of the thesis was to examine what classification method would be the most appropriate for automated benthic macroinvertebrate classification. Two datasets were used in the thesis. One dataset contained benthic macroinvertebrate images from eight taxonomic groups and the other images from 50 species of benthic macroinvertebrates. The thesis produced several novel results. Firstly, a new tie situation resolving strategy was introduced when one-vs-one SVM together with majority voting method was used. Secondly, a novel approach to parameter selection for SVMs was proposed. Thirdly, a new approach to class division problem in Half-Against-Half SVMs was developed by applying Scatter method. Lastly, a new classification method called Directed Acyclic Graph k-Nearest Neighbour was introduced. In this thesis altogether four multi-class extensions of support vector machines and 12 other classification methods were used. SVMs were tested with seven kernel functions, and several feature sets were used in the tests. SVMs were very suitable for the benthic macroinvertebrate classification. With the smaller dataset one-vs-one method achieved over 97% accuracy and half-against-half support vector machine achieved around 96% accuracy. Eleven classification methods other than multi-class support vector machines were tested with the smaller dataset. Of these methods the best ones were Quadratic Discriminant Analysis, Multi-Layer Perceptron and Radial Basis Function network. These methods attained around 94% accuracy. The larger dataset was tested with two classification methods. The accuracies achieved with these methods were around 80%. According to the classification results support vector machines are suitable for automated benthic macroinvertebrate classification when a proper feature set, kernel function and optimal parameter values are found.

Return to Coronet Books main page