Semi-Supervised Deep Learning Models for Automatic Identification of Benthic Fauna

This paper explores the automatic classification of benthic fauna, specifically copepoda, which play a crucial role in aquatic ecosystems. Given the challenge of limited labeled data, we investigate the use of semi-supervised learning (SSL) to improve classification performance. We evaluate multiple deep learning models, including ResNet and DenseNet architectures, under both supervised and semi-supervised frameworks. DenseNet201 achieved the best performance, with an accuracy of 91.81% when trained on a small labeled set and 100.0% when trained on a larger dataset. Remarkably, SSL reached 100.0% accuracy using only the small labeled set, demonstrating its effectiveness in leveraging unlabeled data. This highlights a key motivation for employing SSL: improving classification accuracy when labeled data are scarce. By integrating pseudo-labeling our approach significantly enhances classification accuracy. These results highlight the potential of SSL for ecological studies, providing an effective tool for automated species classification while reducing reliance on large labeled datasets. The findings contribute to advancing environmental monitoring techniques and open new opportunities for further research in marine biodiversity analysis.