Phytoplankton Classification Using a Simplified Convolutional Neural Network and EDGE Enhancement

Phytoplankton is the most fundamental component of marine ecosystems and its distribution and abundance can provide early warning of natural disasters. However, the small intra class variations, significant underwater dynamics, and complex environment of plankton make it difficult to capture crucial subtle features in plankton classification tasks. This work presents a simplified convolutional neural network model based on convolutional residual connections, which can accomplish the fine-grained classification task of plankton with high classification accuracy. To depress the interference from the background noise an edge-enhanced strategy is presented as a preprocessing of the convolutional neural networks, which could filter out the background information and help in extracting discriminative and informative regions of the image and enhance the useful feature representation. The effectiveness of the proposed method was studied and verified on five classes of phytoplankton from a public dataset WHOI. Two harmful algae, Gymnodinium and Phaeocystis globosa, were included in the dataset. The classification accuracy by the proposed method was 97.66% which was slightly better than the Transformer-based network structure. Moreover, the proposed model has a simple structure with less than half of the model parameters of the Swin Transformer, which could significantly reduce the computational costs and improve the processing speed. The proposed method balances accuracy and efficiency that can better learn the characteristics of discriminative features and is suitable for onsite phytoplankton monitoring.