Palynology (the study of pollen) is one of the most important techniques in palaeoclimatic reconstruction [ 1 ]. Many of the problems involved in preparation have been overcome but the analysis of core samples still requires time consuming manual identification by highly trained palynologists. Commercial rates are in the order of US$50-200 per sample (1995 prices)( ibid. ). The approaches used are generally statistical in nature, results for a section of the slide are taken and assumed to be representative of the entire slide. Further problems are caused by the subjective nature of the pollen identification.
Various attempts have been made to automate the process of pollen identification. Vezey and Skvarla [ 2 ] used SEM images to automate the measurement of surface features. Langford [ 3 ] and Langford et al. [ 4 ] [ 5 ] used texture analysis on the exine surfaces, again using images obtained by SEM. Using Haralick's [ 6 ] texture measures in conjunction with Laws masks they differentiated between 6 different pollen types with 98.4% accuracy. SEM work is however expensive and slow, and initial attempts to use optical microscopy were unsuccessful, Witte [ 7 ]. Stillman and Flenley [ 1 ] report work by Treloar in 1994 which achieved over 90% accuracy with 12 reference types. However, both Langford and Treloar used small example sets (10-20 examples per taxon) and recognised that this led to unreliable classifiers, since the number of features being used for identification is far in excess of the available samples.
As an initial step towards producing a useful classification system the current work is investigating the detection of pollen grains in images obtained from an optical microscope. Two very different approaches are compared in this paper. The first is a model based approach, using the commonly found ``double edge'' of certain types of pollen. A ``snake'' [ 8 ] is used to detect the presence (or otherwise) of this edge. The second approach uses a shape extracting neural network, the Paradise network [ 9 ], which classifies objects according to their similarity with respect to certain features (e.g. edges, texture).
An outline of each method is given and then each is tested using a large set of images extracted from samples taken from a reference slide containing Hawthorn pollen grains and a contaminant grain of unknown origin.
Mr I France