Weakly Supervised Segmentation for Coral Mapping via UAV Imagery
A research paper on arXiv (2508.18958) presents a multi-scale weakly supervised semantic segmentation (WSSS) framework for coral habitat mapping. The method uses fine-scale, multi-label predictions from underwater imagery combined with broad-coverage aerial data from Unmanned Aerial Vehicles (UAVs). Point-level classifications are converted into coarse supervision masks to train a segmentation model on UAV orthophotos, with a second training step using the model's refined predictions to improve spatial accuracy without extra annotations. The approach enables large-area segmentation of coral morphotypes and demonstrates flexibility in integrating new classes.
Key facts
- arXiv paper 2508.18958
- Multi-scale weakly supervised semantic segmentation (WSSS) framework
- Combines underwater imagery and UAV aerial data
- Point-level classifications converted to coarse supervision masks
- Second training step uses model's refined predictions
- Demonstrated on coral reef imagery
- Enables large-area segmentation of coral morphotypes
- Flexibility in integrating new classes
Entities
Institutions
- arXiv