Analyzing changes in sediment meiofauna communities using the image analysis software ZooImage

Wehere propose a novelmethod of automatic classification of higher taxa frombenthicmeiofaunal communitiesusing the image analysis software ZooImage. Meiofauna was extracted from sediment at five sites at differentdepths in the Gullmar Fjord on the Swedish west-coast, and digitalized through scanning. The resulting imageswere analyzed with the image analysis software, by comparing them to a reference image library of meiofaunalgroups of higher taxa, as well as non-faunal groups consisting of different types of debris. The accuracy of theanalyses was tested using ZooImage internal cross-validation method, and by comparing digitalized samplesfrom the different sites with manually classified samples. The internal validation accuracy (82–93%)was comparableto prior published studies on zooplankton. The largest errors in classification were within the non-faunalgroups and accuracy for classification of faunal groups was as high as 93%. Misclassification within the faunalgroups was mainly due to either too few images of some classes in the training set, or to when fauna couldnot be sufficiently separated from debris in the images used for the library, causing interference in the learningalgorithm. Comparison with manual classification confirmed the errors revealed in the internal cross-validation.Statistical analysis revealed differences in the meiofauna communities between different depths as well as atemporal chang