A carregar...
Publicação
Automated detection and measurement of commercial fish species using deep learning object detection models : a comparative study across Portuguese ports
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.62 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Manual measurement remains a conventional and widely used method for sampling fish length in fisheries monitoring and research, but it is time-consuming. This thesis contributes to the development of an automated image-based system using object detection models to identify and measure multiple fish species landed by commercial fisheries across varied conditions in ports. The study focuses on four species of commercial relevance, Pagellus acarne, Trisopterus luscus, Merluccius merluccius, and Dicentrarchus labrax, sampled across seven Portuguese ports. Two deep learning-based object detection models were tested: a previously developed model trained only using data from one port (Model Sesimbra 2020), and a new model trained using data from all ports (Model 2025). In each case, both single-species and multi-species versions were evaluated using standardised metrics, including mAP50 and mAP50–95 for detection accuracy, and mean difference, Pearson correlation, and Bland-Altman analysis for length estimation accuracy. The single-species Model Sesimbra 2020 showed variable performance both within and outside its training environment, with some species-port combinations showing large measurement errors even at the training location. When testing the single-species versions of this model, length measurement errors ranged from -1.54 mm to +9.29 mm, with higher measurement errors observed for larger individuals. The multi-species version of the Model Sesimbra 2020 reduced measurement errors between manual and model derived estimates compared to the single-species models. The classification accuracy of the multi-species Model 2025 reached 94% in D. labrax and was lowest in T. luscus, at 89%, with inter-species misclassification rates below 0.6% for all species pairs. The findings demonstrate that geographically diverse training data can substantially increase model generalisation, with improved detection accuracy, species classification, and length measurement across variable conditions. While false positive detections (detecting fish where none are present) remain a challenge, especially in T. luscus, performance across species and ports improved substantially in the 2025 model. The results support the use of automated systems in fisheries monitoring to expand data coverage and increase efficiency in stock assessment programs.
Descrição
Tese de Mestrado, Ecologia Marinha, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Deep learning Computer vision Neural networks Object detection Fisheries monitoring