Automated recognition of deep-sea benthic megafauna in polymetallic nodule mining areas based on deep learning
The exploitation of deep-sea polymetallic nodules has attracted global attention. To mitigate its impact on deep-sea ecosystems, accurate identification of benthic megafauna is essential for developing science-based mining strategies. Deep learning has emerged as an promising approach in biological...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954125003280 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The exploitation of deep-sea polymetallic nodules has attracted global attention. To mitigate its impact on deep-sea ecosystems, accurate identification of benthic megafauna is essential for developing science-based mining strategies. Deep learning has emerged as an promising approach in biological identification, offering significantly improved efficiency and reduced reliance on taxonomic expertise. However, challenges such as similar species characteristics, occlusion, and species abundance imbalance hinder accurate recognition of deep-sea benthic megafauna. To address these issues, the present study proposes Benthic Megafauna-YOLO (BM-YOLO). Its backbone integrates deformable convolutions, attention mechanisms, and ResNet structures to improve feature extraction and reduce background interference. Additionally, a transformer-enhanced PAFPN incorporates global context, improving occluded target recognition. To address the challenge of low-abundance species, we introduce a varifocal head that enhances the focus of the model on rare species. To address the lack of relevant datasets, we constructed the Pacific Ocean Benthic Megafauna dataset using images collected during the China Oceanic 45 and 50 surveys. Notably, BM-YOLO achieved AP50 scores exceeding 90 % and 95 % for six and four species, respectively. Furthermore, the model achieved the mAP50 and mAP50:95 values of 85.9 % and 56.2 %, respectively, improving on YOLOv8s by 2.6 % and 4.6 % and YOLOv11s by 0.7 % and 1.5 %. On the public underwater dataset DUO, BM-YOLO achieved mAP50 and mAP50:95 values of 82.4 % and 62.7 %, respectively, outperforming YOLOv10s by 2.0 % and 2.9 %. These results highlight the potential of BM-YOLO in automated deep-sea species identification. We also identified key factors influencing species recognition accuracy through experimental analysis. |
|---|---|
| ISSN: | 1574-9541 |