Palynological data are used in a wide range of applications, but the tasks of classification and counting of pollen grains are highly skilled and laborious. The development of an automated system for pollen identification and classification would be of great benefit. Previous attempts at computer classification have taken approaches that have been intrinsically difficult to develop into fully automated systems that could operate largely independently of a human operator. We describe a new approach to the problem based on improving the quality of the image processing rather than the data collected using images collected with an optical microscope. Two sets of experiments are described, demonstrating the ability of the system firstly, to differentiate between pollen and detritus, and secondly, to classify different pollen types correctly. The results of these tests, in which the pollen images were acquired using an automated system, are encouraging and demonstrate that even using relatively low spatial resolution we can reliably differentiate between three taxa of pollen grains. Based upon the experience that we have gained we describe the characteristics required of the next generation of automated pollen identification and classification systems. (C) 2000 Elsevier Science Ltd.