Deep learning for automated fish detection in underwater images: a tool for sustainable marine ecosystem monitoring

Abstract

Deep learning has emerged as a powerful tool for automated object detection, offering unprecedented speed and accuracy in analyzing complex visual data. In the context of marine ecosystem monitoring, convolutional neural networks (CNNs), particularly YOLO-based architectures, have demonstrated remarkable efficiency in detecting and classifying fish species in underwater imagery. Traditional fish identification methods rely on manual annotation, which is both time-consuming and prone to inconsistencies. By implementing a semi-automated labeling approach, where human experts refine AI-generated predictions, the annotation process can be streamlined while ensuring taxonomic precision. A key aspect of this research is the creation of a comprehensive training guide that optimizes the model’s performance by detailing best practices in dataset preparation, annotation techniques, hyperparameter tuning, and augmentation strategies. Using a dataset derived from the OBSEA marine observatory, results indicate that the YOLO extra-large model, trained with a small learning rate and high-resolution images, achieves optimal performance in fish identification. The findings underscore the potential of AI-assisted methodologies in ecological research, offering a scalable and efficient alternative to manual annotation for sustainable marine biodiversity monitoring.

This work has been supported by various funding sources and research initiatives. We acknowledge the financial support from grants 2023 INV-2 00044 (position codes 200044TC31 and 200044TC6). Additionally, this research has been funded by the European Commission’s HORIZON-INFRA-2021-SERV-01 program under the iMagine project (grant agreement 101058625). We also recognize the use of the EGI infrastructure with dedicated support from EGI-IFCA-STACK, which contributed to the computational resources required for this study. Furthermore, the researchers wish to acknowledge the support of the Associated Unit Tecnoterra, composed of members from UPC and ICM-CSIC, for their valuable collaboration in this work.

Peer Reviewed

Postprint (published version)