SARMSSD

SARMSSD (SAR Multi-Scale Ship Detection) Dataset

This repository provides the implementation of deep learning-based ship detection in SAR imagery using the YOLO11 model, associated with our research paper "Impact of Data Enhancement Methods on Ship Detection Using YOLO11 in SAR Imagery" which we present in Machine Intelligence for GeoAnalytics and Remote Sensing (MIGARS 2025) International Conference. The study evaluates the effects of image denoising and data augmentation techniques on detection performance using a newly constructed dataset, SARMSSD, which combines three open-source SAR datasets (SSDD, HRSID, and SRSDD).

Accesss our paper --> https://ieeexplore.ieee.org/document/11231948

Dataset Details

• Number of Total Images: 7,430
• Number of Total Ship Instances: 22,195
• Spatial Resolution: Ranges from 0.5m to 15m
• Image Resolution: Varies between 300x300, 500x500, 800x800 and 1024x1024 pixels
• Annotation Format: YOLO
• Annotation Style: HBB (Horizontal Bounding Box)

Download --> SARMSSD Dataset

Impact of Data Enhancement Methods on Ship Detection Using YOLO11 in SAR Imagery

Ship detection in synthetic aperture radar (SAR) imagery is critical for maritime surveillance, security, and environmental monitoring. However, SAR images often contain speckle noise and complex backgrounds, affecting detection accuracy.

This project explores:

Data Collection and Processing

The SARMSSD dataset was constracted by merging the following publicly available SAR datasets:

• SSDD: 1,160 SAR images (obtained from RadarSat-2, TerraSAR-X and Sentinel-1 satellites) with 2,358 ship instances, having resolutions from 300x300 to 500x500 pixels and spatial resolutions ranging from 1m to 15m. This dataset has the PascalVOC annotation format, it has been converted to YOLO annotation format with voc2yolo.py
• HRSID: 5,604 SAR images (obtained from Sentinel-1B, TerraSAR-X, and TanDEM-X satellites) with 16,951 ship instances, featuring a fixed resolution of 800x800 pixels and spatial resolutions between 0.5m and 3m. This dataset has the COCO annotation format, it has been converted to YOLO annotation format with coco2yolo.py
• SRSDD: 666 SAR images (obtained from Chinese Gaofen-3 satellite) with 2,886 ship instances, all at 1024x1024 pixels with a spatial resolution of 1m. This dataset has the DOTA Oriented Bounding Box(OBB) annotation format, it has been converted to YOLO annotation format with obb2hbb.py

Noise Reduction Techniques

To improve the quality of SAR images and reduce noise in the SARMSSD dataset, three enhancement techniques have been applied: Median filtering with 5x5 window median.py, BM3D method with sigma_psd set to 0.1 apply_bm3d.py, and a deep learning-based SAR-CAM method implemented in test_sar-cam_sarmssd.ipynb with utilizing our trained weight SAR-CAM_best.pth. These image processing methods yielded three enhanced variants of the original dataset which you can access them below:

Download --> SARMSSD-Median Dataset

Download --> SARMSSD-BM3D Dataset

Download --> SARMSSD-SAR-CAM Dataset

YOLO11 Deep Learninig Experiments

Training and evaluating the 4 SARMSSD datasets using YOLO11 algorithm & Comparing performance results.

Model	Dataset	size (pixels)	Precision	Recall	mAPtest 0.50	mAPtest 0.50:0.95	params (M)
YOLO11x	SARMSSD	640	0.935	0.829	0.909	0.764	56.9
YOLO11x	SARMSSD-Median	640	0.938	0.824	0.911	0.718	56.9
YOLO11x	SARMSSD-BM3D	640	0.936	0.812	0.902	0.737	56.9
YOLO11x	SARMSSD-SAR-CAM	640	0.950	0.815	0.908	0.770	56.9

Data Augmentation Experiments:

Mosaic, scale, and flip augmentations aplplied through SARMSSD-SAR-CAM dataset trainings to further improve ship detection accuracy.

Model	Augmentation	size (pixels)	Precision	Recall	mAPtest 0.50	mAPtest 0.50:0.95	params (M)
YOLO11x	scale	640	0.924	0.881	0.942	0.739	56.9
YOLO11x	flip(lr) left-right	640	0.944	0.839	0.919	0.772	56.9
YOLO11x	mosaic	640	0.951	0.881	0.945	0.788	56.9
YOLO11x	mosaic, flip(lr)	640	0.922	0.888	0.940	0.754	56.9
YOLO11x	scale, flip(lr)	640	0.921	0.869	0.939	0.711	56.9
YOLO11x	mosaic, scale	640	0.945	0.882	0.950	0.755	56.9
YOLO11x	mosaic, scale & flip(lr)	640	0.929	0.885	0.948	0.733	56.9

All experiments were conducted using the YOLO11-x architecture (56.9M parameters) and optimized with SGD (learning rate: 0.01, weight decay: 0.0005) for 100 epochs with a batch size of 16.

Citation

If you make use of this research, please cite our paper.

@INPROCEEDINGS{11231948,
  author={Büyükkanber, Furkan and Yanalak, Mustafa and Musaoğlu, Nebiye},
  booktitle={2025 International Conference on Machine Intelligence for GeoAnalytics and Remote Sensing (MIGARS)}, 
  title={Impact of Data Enhancement Methods on Ship Detection Using YOLO11 in SAR Imagery}, 
  year={2025},
  volume={},
  number={},
  pages={1-4},
  keywords={Deep learning;Noise reduction;Data augmentation;Radar polarimetry;Data models;Reliability;Marine vehicles;Remote sensing;Machine intelligence;Image denoising;deep learning;ship detection;SAR dataset;image denoising;SAR-CAM;SARMSSD;YOLO11},
  doi={10.1109/MIGARS67156.2025.11231948}}

Contact

For further information or any question, you can use the issues (https://github.com/buyukkanber/SARMSSD/issues) tab.

---