ORYCTOS vol. 7, 2008 129 Analysis of spheniscid humerus and tarsometatarsus morphological variability using DAISY automated image recognition. Stig A. Walsh 1,2 , Norman MacLeod 1 and Mark O’Neill 3 1 The Natural History Museum, Department of Palaeontology, Cromwell Road, London, UK, SW7 5BD. s.walsh@nhm.ac.uk 2 School of Earth and Environmental Sciences, Portsmouth University, Burnaby Road, Portsmouth, UK, PO1 3QL. 3 Centre for Neuroecology, Henry Wellcome Building, University of Newcastle upon Tyne, Newcastle upon Tyne, UK, NE1 7RU. ABSTRACT - Despite a long history of research, relationships within fossil and extant Sphenisciformes remain unclear. This is largely because most fossil species were described on the basis of either the tarsometatarsus or humerus. Neither of these elements is particularly phylogenetically informative, and the extent of intraspeciic morphological variation also remains unknown. Herein we investigate a new approach – the use of artiicial neural-net (ANN) technology – to determine whether either of these elements can be reliably used to identify extant species. The DAISY ANN system was able to recognise most species from either tarsometatarsal or humerus morphology, but its success rate improved when the species training sets were combined into generic groups, indicating the need for larger image libraries. Our preliminary results suggest that these elements can allow reliable identiications for most taxa, but that the tarsometatarsus is on the whole a better element for this purpose. These results also demonstrate the potential for artiicial neural-net technology to address problems in avian taxonomy. Keywords: Spheniscidae, morphology, intraspeciic variability, automated identiication, DAISY. L’analyse de la variabilité morphologique d’humérus et de tarsometatarsus de Spheniscidae em- ployant DAISY a automatisé l’identification d’image - En dépit d’une longue histoire de recherche, les rapports entre les Sphenisciformes fossiles et actuels demeurent peu clairs. C’est en grande partie parce que la plupart des espèces fossiles ont été décrites sur la base du tarsométatarse ou de l’humérus. Ni l’un ni l’autre de ces éléments n’est particulièrement instructif du point de vue de la phylogénie, et l’ampleur de la variation morphologique intraspéciique demeure également inconnue. Nous avons essayé une nouvelle approche - l’utilisation de la technologie de réseau neural artiiciel (ANN) - pour étudier si ces éléments peuvent être employés de façon iable pour identiier des espèces actuelles. Le programme DAISY ANN a pu identiier la plupart des espèces, mais son taux de succès est meilleur quand les espèces formant des ensembles ont été combinées dans des groupes génériques, indiquant le besoin de plus grandes bibliothèques d’images. Nos résultats préliminaires suggèrent que ces éléments peuvent permettre des identiications iables pour la plupart des taxons, mais que le tarsométatarse est dans l’ensemble un meilleur élément à cette in. Ces résultats prometteurs montrent le potentiel de la technologie ANN pour l’étude des problèmes de la taxonomie avienne. Mots clés: Spheniscidae, morphologie, variabilité intraspéciique, identiication automatisée, DAISY. INTRODUCTION With a fossil record extending at least as far back as the early Palaeocene (Slack et al., 2006), Sphenisciformes (penguins) has a fossil record longer than many extant avian clades. The irst fossil penguin remains were described by Huxley in 1859 and, by now, one would expect this group’s early evolution, systematics and historical diversity pat- tern to be well understood. This is not the case. Two main impediments to fossil spheniscid research exist. The irst of these is that, while their remains are not uncommon in some strata, their stratigraphic and geographic distribution is somewhat disjunct (Fordyce & Jones, 1990). New discover- ies, particularly in New Zealand and South America (e.g., Slack et al., 2006; Walsh & Suarez, 2006; Acosta Hospita- leche et al., 2007), may in time help to resolve this aspect of uncertainty. A second and potentially more serious problem is that the majority of remains come from nearshore, relatively high-energy sequences, where skulls and articulated or asso-