Automatic Estuarine Fish Species Classification System Based on Deep Learning Techniques

Abstract

Fish classification (FC) is crucial in various domains, including fishery management and ecological research. Traditional FC methods rely mainly on morphological criteria such as body shape and patterns. Although these methods are useful, they require expert knowledge and are prone to subjective interpretation. Recent advances in technology and the availability of datasets have allowed deep learning (DL) techniques to be used in fish species classification. These methods automatically extract relevant features from fish images and categorize them into species groupings. Traditional DL models, however, have difficulties capturing long-range dependencies and require fixed input sizes, making them less adaptive when working with images with varying proportions. The Vision Transformer (ViT) addresses these constraints by utilizing the transformer model’s self-attention mechanisms. So, in this study, a ViT is used to solve the FC problem. The performance of ViT is assessed against pre-trained models, VGG16, VGG19, DenseNet121, ResNet50v2, InceptionV3, InceptionResNetV2, and Xception. The experiments make use of a curated Estuarine Fish species dataset (EFD). In this study, ViT outperformed state-of-the-art literature by achieving 99.04% and 100% accuracy without and with augmentation, respectively. The presented research is tailored to the task of recognizing estuarine fish species that are useful in the aquaculture domain. Additionally, our research aligns with the objectives of Sustainable Development Goals (SDGs) 2 and 14. This emphasises the broader societal and environmental implications of our work, emphasizing its potential to positively impact food security and aquaculture ecosystem sustainability.

Keywords

• Aquaculture
• deep learning
• estuarine fish dataset
• fish classification
• vision transformers