ML4Sustain/beluga-seeker

Beluga Whale Detection from Satellite Imagery with Point Labels

This repository hosts the model checkpoints for Beluga-Seeker, a project for detecting beluga whales in Arctic satellite imagery using point-label supervision and automated bounding-box generation via Segment Anything Model (SAM).

Model Description

Beluga-Seeker provides YOLOv8s-based object detectors trained on very-high-resolution satellite imagery. Two variants are released:

Checkpoint	Description	Classes
YOLO-SAM.pth	Trained on SAM-generated bounding boxes (recommended)	certain whale, uncertain whale, harp seal
YOLO-Buffer.pth	Trained on point-label buffer bounding boxes	certain whale, uncertain whale, harp seal

Both models are trained with the MMYOLO framework and can be loaded via mmengine.

Paper

Yijie Zheng, Jinxuan Yang, Yu Chen, Yaxuan Wang, Yihang Lu, Guoqing Li.
Beluga Whale Detection from Satellite Imagery with Point Labels.
Accepted for oral presentation at IGARSS 2025.
arXiv:2505.12066

Installation

# Recommended environment
conda create -n belugaSeeker python=3.10
conda activate belugaSeeker
pip install torch==2.0.1 torchvision==0.15.2 --index-url https://download.pytorch.org/whl/cu118

# Install MMYOLO and dependencies
pip install openmim
mim install "mmengine>=0.6.0" "mmcv>=2.0.0rc4,<2.1.0" "mmdet>=3.0.0rc6,<3.1.0" "mmyolo"

Usage

from mmengine.runner import Runner
from mmyolo.utils import register_all_modules

register_all_modules()

# Load checkpoint
cfg = 'configs_beluga/yolov8_s_b24-100e.py'
runner = Runner.from_cfg(cfg)
runner.load_checkpoint('YOLO-SAM.pth')

For a complete inference demo, see inference_demo.ipynb in the source repository.

Model Checkpoints

File	Size	Training Data
YOLO-SAM.pth	~87 MB	SAM-generated bounding boxes
YOLO-Buffer.pth	~87 MB	Point-label buffer boxes

Dataset & Labels

• Classes: certain whale, uncertain whale, harp seal
• Palette: [(250, 165, 30), (255, 255, 0), (0, 255, 0)]
• Labels were originally created as point annotations and converted to bounding boxes using SAM to improve annotation efficiency.

Citation

@inproceedings{zheng2025beluga,
  title={Beluga Whale Detection from Satellite Imagery with Point Labels},
  author={Zheng, Yijie and Yang, Jinxuan and Chen, Yu and Wang, Yaxuan and Lu, Yihang and Li, Guoqing},
  booktitle={IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
  year={2025}
}

Acknowledgements

This work is supported by the 4th IEEE GRSS Student Challenge.

The authors gratefully acknowledge Amou for assistance with ArcGIS Pro and Zori for delivering the data.

License

This project is licensed under the GPL-3.0 License.