Implications of single-stage deep learning networks in real-time zooplankton identification

Abstract

Zooplankton are key ecological components of the marine food web. Currently, laboratory-based methods of zooplankton identification are manual, time-consuming, prone to human error and require expert taxonomists. Therefore, alternative methods are needed. In this study, we describe, implement and compare the performance of six state-of-the-art single-stage deep learning models for automated zooplankton identification. The highest prediction accuracy achieved is 99.50%. The fastest detection speed is 285 images per second, making the models suitable for real-time zooplankton classification. We validate the predictions of the generated models on unseen images. The results demonstrate the capabilities of the latest deep learning models in zooplankton identification.