Krieger, Jonathan D. , Sprunt, Susan V , Farler, David Miguel , Munson, Richard D. .
Automated identification of fern taxa using neural net analysis of scale characters.
The taxonomy of the scaly polypods (Pleopeltis s.l., Polypodiaceae) depends largely on characters of the scales which cover the leaves and rhizome. Scale distribution, shape, color, color pattern, and margin ornamentation vary, and scales are frequently dimorphic between the rachis and lamina. Up to seven subspecies or varieties are recognized within Pleopeltis polypodioides: aciuluaris, burchellii, ecklonii, knoblochianum, michauxiana, minus, and polypodioides. Morphometric approaches to quantifying differences among these subspecies are intractable— collecting linear measurements or outline data for individual scales is so tedious that, even if informative, it is unlikely to be replicated for identification—and won't capture much of the putatively taxonomically informative data, e.g., distribution, color, and color pattern.
Neural net analysis can capture all of these types of variation, through photographs of the leaf abaxial surface. The DAISY plastic self-organizing map (PSOM) neural net is a scalable system which has demonstrated consistently high classification rates (better than expert taxonomists and morphometric approaches) for systems as diverse as butterflies and planktonic foraminifera. We performed a series of DAISY neural net analyses to assess which types of variation had the highest taxonomic information content using six of the subspecies of P. polypodioides.
Our results suggest that color is not important to discrimination, and that both scale dimorphy and scale distribution are key to making correct identification. The widespread subspecies P. polypodioides polypodioides had the highest correct identification rates. Nevertheless, identification rates were, overall, much lower than seen in other neural net analyses, perhaps due to ongoing hybridization. In this instance, neural net analysis may be of little utility in automating rote identification; however, it does provide a quantitative index to the ability to distinguish these taxonomic units based on a specific set of characters.
Log in to add this item to your schedule
1 - the Natural History Museum, Palaeontology, Cromwell Road, London, SW7 5BD, United Kingdom
2 - Miami University, Department of Botany, Oxford, Ohio, 45056, USA
neural net analysis
Presentation Type: Oral Paper:Papers for Sections
Date: Monday, July 28th, 2008
Time: 4:00 PM