Automated seaweed species classification using deep learning and large language models
Résumé
Seaweeds are foundational components of marine ecosystems and hold significant value in global aquaculture. Accurate and efficient identification of seaweed species is critical for biodiversity monitoring, ecological research, and sustainable management. This study evaluates a dual-approach framework for automated seaweed classification. First, we implemented a standard convolutional neural network (CNN) based on the EfficientNet-B0 architecture, utilizing transfer learning from ImageNet. The model was trained on an augmented dataset of 3,440 images derived from an original collection of 800 field images (stratified 80/10/10 split) covering 43 seaweed species. The CNN achieved a baseline classification accuracy of 89%. Second, we present a proof-of-concept study using a vision-language model (VLM), specifically Claude 3.5 Sonnet, to explore semantic reasoning in taxonomic classification. The VLM achieved a raw accuracy of 70%, which was further improved to an effective accuracy of 92% on a selected subset through a human-in-the-loop (HITL) validation system. While the CNN provides a robust and rapid classification tool, the VLM-HITL approach offers an interpretable semantic alternative for validating challenging specimens. This work contributes to the development of scalable, automated biodiversity monitoring tools for marine science.
Citer ce document
Accès au document
Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter
Voir l'article sur le site de la revueAuteur(s)
Statistiques
Consultations : 1
Téléchargements : 0