Computerized identification of pollen grains by texture analysis

Automation of palynology could lead to many advances: rapid results, larger data sets, objectivity, fine resolution sampling and possibly finer determinations. To test the feasibility of automation, SEM photographs of six modern pollen taxa were used. The images were digitised and samples of exine texture covering approximately 1 10 of the total pollen area were extracted from the digital images. Texture analysis was applied to 192 samples obtained in this way. First, a co-occurrence matrix of grey levels was established for each sample. Then texture measures were calculated and used as input to a classification programme. With a leave-one-out strategy and a variable selection procedure, the proportion of pollen grains correctly identified rose to 94.3%. The procedure required c.10 seconds of processing on a VAX computer for each grain. With faster computers and programs, this could be cut to 1 second.