Artificial neural networks as potential classification tools for dinoflagellate cyst images: A case using the self-organizing map clustering algorithm

Abstract

Automated palynological analysis has been previously proposed but proved difficult to achieve. Here, the first instance of dinoflagellate cyst (dinocyst) image clustering from palynological samples based on morphological and textural image analysis (IA) features is presented. Dinocyst-dominated example images (including acritarchs and other algae) were acquired from Cretaceous, Paleogene and Holocene samples. IA features that cover a broad array of measurements are used, including morphological, Fourier and textural descriptors, as well as geometric moments and color. To determine clusters, unsupervised self- organized maps (a genre of artificial neural networks) were used. Self-organized maps (SOMs) can determine their own classification of data in cases where no knowledge of the true class of each data point exists. Using a SOM five major clusters can be identified including clusters of freshwater alga, proximate dinocysts, proximate dinocysts with long horns, proximochorate dinocysts and chorate dinocysts. Minor variations can also be identified based on red, green and blue color, textural variations, and dinocyst process length in the form of morphological (shape) descriptors. These major and minor clusters demonstrate the open-ended capability of the system to be a wider palynological identification tool. The advantages of automation in palynology are articulated and the hurdles yet to overcome discussed. These advantages include freeing humans from ‘routine identification’ so that more emphasis can be placed upon distinguishing the rarer (‘unknown’) particles and placing them in a descriptive context. Challenges yet to conquer include distinguishing between images of closely related morphologies and/or textures. It is envisaged that technological and computational developments will quickly facilitate such further developments.