Enumeration, measurement, and identification of net zooplankton samples using the ZOOSCAN digital imaging system

Identifying and counting zooplankton are labour-intensive and time-consuming processes that are still performed manually. However, a new system, known as ZOOSCAN, has been designed for counting zooplankton net samples. We describe image-processing and the results of (semi)-automatic identification of taxa with various machine-learning methods. Each scan contains between 1500 and 2000 individuals <0.5 min. We used two training sets of about 1000 objects each divided into 8 (simplified) and 29 groups (detailed), respectively. The new discriminant vector forest algorithm, which is one of the most efficient methods, discriminates between the organisms in the detailed training set with all accuracy of 75% at a speed of 2000 items per second. A supplementary algorithm tags objects that the method classified with low accuracy (suspect items), such that they could be checked by taxonomists. This complementary and interactive semi -automatic process combines both computer speed and the ability to detect variations in proportions and grey levels with the human skills to discriminate animals on the basis of small details, such as presence/absence or number of appendages. After this checking process, total accuracy increases to between 80% and 85%. We discuss the potential of the system as a standard for identification, enumeration. and size frequency distribution of net-collected zooplankton. (C) 2004 Published by Elsevier Ltd on behalf of International Council for the Exploration of the Sea.