RECOGNITION OF FRESHWATER MACROINVERTEBRATE TAXA BY IMAGE ANALYSIS AND ARTIFICIAL NEURAL NETWORKS

Routine taxonomic identification is a limitation factor in the study of macroinvertebrates communities, a key group of freshwater ecosystems. Traditionally, macroinvertebrates has been identified through examination under stereoscopic microscope, an activity that requires high technical expertise and a considerable amount of time. In this paper we present the first automatic taxonomic identification of freshwater macroinvertebrate taxa, achieved through a novel image processing program developed with MATLAB^®. The program works in a completely automated fashion once it has been trained, with no user intervention. A set of morphological and texture parameters are calculated through image analysis and processed by a hierarchical set of partitioned artificial neural networks (ANNs) in order to identify the taxon to which presented specimens belong. Classification performance is estimated by 10-fold stratified cross-validation. Specimens of 10 macroinvertebrate taxa of varying taxonomic hierarchy were isolated and identified from field samples. Digital images of specimens were acquired with a flatbed scanner, yielding a database of 1042 images. Overall recognition performance was > 74% for all taxa, with most values in the 80-90% range, and a highest value of 99.48%. The followed scheme of image processing and hierarchical partitioned ANNs analysis proved to be effective for this particular challenge of pattern recognition, yielding a global classification performance of 87.83% and being able to distinguish between species of the same genus.